Datasets:

math-ai

/

AutoMathText

like 148

Follow

math-ai 19

https://arxiv.org/abs/2002.11025

New bounds for perfect $k$-hashing

Let $C\subseteq \{1,\ldots,k\}^n$ be such that for any $k$ distinct elements of $C$ there exists a coordinate where they all differ simultaneously. Fredman and Komlós studied upper and lower bounds on the largest cardinality of such a set $C$, in particular proving that as $n\to\infty$, $|C|\leq \exp(n k!/k^{k-1}+o(n))$. Improvements over this result where first derived by different authors for $k=4$. More recently, Guruswami and Riazanov showed that the coefficient $k!/k^{k-1}$ is certainly not tight for any $k>3$, although they could only determine explicit improvements for $k=5,6$. For larger $k$, their method gives numerical values modulo a conjecture on the maxima of certain polynomials.In this paper, we first prove their conjecture, completing the explicit computation of an improvement over the Fredman-Komlós bound for any $k$. Then, we develop a different method which gives substantial improvements for $k=5,6$.

\section{Introduction} For positive integers $k\geq 2$ and $n \geq 1$, consider a subset $C\subset \{1,\ldots, k\}^n$ with the property that for any $k$ distinct elements of $C$ there exists a coordinate where they all differ. We call such a set a \emph{perfect $k$-hash code} of length $n$, or simply $k$-hash for brevity. The name is motivated by the idea that if each coordinate of $C$ is interpreted as a $k$-hash function on a set $U$ of cardinality $|C|$, then any $k$ elements of $U$ are hashed onto $\{1,2,\ldots,k\}$ by at least one function. Determining the largest possible cardinality of such a set $C$ as a function of $k$ and $n$ is a classic combinatorial problem in theoretical computer science. One standard formulation is to study, for fixed $k$, the grow of the largest possible $|C|$ as $n$ goes to infinity. It is known that $|C|$ grows exponentially in $n$. Then one usually defines the \emph{rate} of the code as \footnote{Here and in the whole paper $\log{x}$ is understood to be in base two.} \begin{equation} R=\frac{\log|C|}{n} \end{equation} and asks for bounds on the rate of codes of maximal cardinality as $n\to \infty$. This formulation of the problem can also be cast as a problem, in information theory, of determining the zero-error capacity under list decoding for certain channels (REF). In this paper we consider upper bounds on $R_k$, defined as the limsup, as $n\to\infty$, of the rate of largest $k$-hash codes of length $n$. A simple packing argument (see \cite{Elias1}) shows that for all $k\geq 2$ one has $R_k\leq \log(k/(k-1))$. For $k=3$, the simplest non-trivial case, this evaluates to $\log(3/2) \approx 0.5850$ and is still the best known upper bound to date (the best lower bound is $1/4\log(9/5)\approx 0.212$). For $k\geq 4$, the first important result was derived by Fredman and Koml{\'o}s \cite{FredmanKomlos}, who proved that \begin{equation} \label{eq:fredmankomlosRk} R_k\leq k!/k^{k-1}\,. \end{equation} We also refer to \cite{Korner1}, \cite{korner86}, \cite{Korner2} and \cite{Korner3} where the Fredman-Koml{\'o}s bound (and some generalizations to hypergraphs) has been cast using the language of graph entropy and to \cite{nilli} where a simple probabilistic proof has been presented. Improvements were obtained for $k=4$ in \cite{Arikan1}, \cite{Arikan2} and more recently in \cite{DalaiVenkatJaikumar}, \cite{DalaiVenkatJaikumar2}. The most recent progress we are aware of was obtained in \cite{venkat} where the Fredman-Koml{\'o}s bound is proved to be non-tight for any $k\geq 5$, with an explicit new numerical bound for $k=5,6$. For larger $k$, the authors show that an explicit improvement of the Fredman-Koml{\'o}s bound can be obtained subject to a conjecture on the maxima of certain polynomials. Other recent papers on this topic that deserve to be recalled are \cite{Jaikumar2} where the asymptotic behavior of $R_k$ has been studied and \cite{CostaDalai} where the authors attempt to use the polynomial method to upperbound $R_3$ and they state some limitations of this method. In this paper we make further progress on this problem. We first prove the conjecture formulated in \cite{venkat} and thus complete their proof of explicit new upper bounds on $R_k$ which beat the Fredman-Koml{\'o}s bound for all $k\geq 5$. Our main contribution is then to expand on the idea used in \cite{DalaiVenkatJaikumar2} to derive a further improvement for $k=5,6$. In Section \ref{background} we give a brief summary of the approaches used in \cite{DalaiVenkatJaikumar2} and in \cite{venkat}, upon which we build our contribution. In Section \ref{sec:venkat_conj} prove the conjecture stated in \cite{venkat} and give a numerical evaluation of the ensuing bound for $k>6$. In Section \ref{sec:newbounds} we present our improvement for $k=5,6$. \section{Background} \label{background} The bounds presented in \cite{FredmanKomlos}, \cite{Arikan2}, \cite{DalaiVenkatJaikumar2} and \cite{venkat} can all be derived by starting with the following Lemma on graph covering (see \cite{Jkumar}). \begin{Lemma}[Hansel \cite{hansel}] Let $K_r$ be a complete graph on $r$ vertices and let $G_1,\ldots,G_m$ be bipartite graphs on those same vertices such that $\cup_{i}G_i=K_r $. Let finally $\tau(G_i)$ represent the number of non-isolated vertices in $G_i$. Then \begin{equation} \sum_{i=1}^m \tau(G_i) \geq \log r\,. \end{equation} \end{Lemma} The connection with $k$-hashing comes from the following application. Given a $k$-hash code $C$, fix any $(k-2)$-elements subset $\{x_1,x_2,\ldots, x_{k-2}\}$ in $C$. For any coordinate $i$ let $G_i^{x_1, \ldots, x_{k-2}}$ be the bipartite graph with vertex set $G\setminus\{x_1,x_2,\ldots, x_{k-2}\}$ and edge set \begin{equation} \label{eq:FKgraph} E=\left\{ (v,w) : x_{1,i},x_{2,i},\ldots, x_{k-2,i},v,w \mbox{ are all distinct} \right\}\,. \end{equation} Then, since $C$ is a $k$-hash code, we note that $\bigcup_i G_i^{x_1, \ldots, x_{k-2}}$ is the complete graph on $G\setminus\{x_1,x_2,\ldots, x_{k-2}\}$ and so \begin{equation} \label{eq:hansel_hash} \sum_{i=1}^n \tau(G_i^{x_1, \ldots, x_{k-2}})\geq \log(|C|-k+2)\,. \end{equation} This inequality can be used to prove upper bounds on $|C|$. Since it holds for any choice of $x_1,x_2,\ldots, x_{k-2}$, one can show that the right hand side is small by proving that left hand side cannot be too large for all possible choices of $x_1,x_2,\ldots, x_{k-2}$. One can either use it for some specific choice or take expectation over any random selection. Let $f_{i}$ be probability distribution of the $i$-th coordinate of $C$, that is, $f_{i,a}$ is the fraction of elements of $C$ whose $i$-th coordinate is $a$. Note that the graph in \eqref{eq:FKgraph} is empty if the $x_{1,i},x_{2,i},\ldots, x_{k-2,i}$ are not all distinct. We will say in this case that $x_1,x_2,\ldots,x_{k-2}$ \emph{collide} in coordinate $i$. Then, we have \begin{equation} \tau(G_i^{x_1, \ldots, x_{k-2}})= \begin{cases} 0 \hspace{2cm} x_1,\ldots,x_k \mbox{ collide in coordinate }i\\ \left(\frac{|C|}{|C|-k+2}\right)\left(1-\sum_{j=1}^{k-2}f_{i,x_{ji}}\right) \hspace{0.5cm}\mbox{otherwise} \end{cases} \end{equation} So, one can make the left hand side in \eqref{eq:hansel_hash} small by either taking a set $x_{1}, \ldots,x_{{k-2}}$ which collide in many coordinates, so forcing the corresponding $\tau$'s to zero, or by taking a set which uses ``popular'' values in many coordinates. The Fredman-Koml{\'o}s bound is obtained by taking expectation in $\eqref{eq:hansel_hash}$ over a uniform random extraction of $x_1,x_2,\ldots, x_{k-2}$. By linearity of expectation the computation can be performed over each single coordinate. Denoting with $\mathbb{E}$ the expectation, for large $n$ and $|C|$ \begin{align*} \mathbb{E} [ & \tau(G_i^{x_1, \ldots, x_{k-2}})] \\ & =\left(1+o(1)\right)\sum_{\stackrel{\text{distinct }}{ a_1,\ldots,a_{k-2}}} f_{i,a_1}f_{i,a_2}\cdots f_{i,a_{k-2}}(1-f_{i,a_1}\cdots-f_{i,a_{k-2}} ) \end{align*} where the coefficient $o(1)$ is due to sampling without replacement. One can show that the worst-case $f_i$ is the uniform distribution, which gives \begin{equation} \label{eq:fredmankomlostau} \mathbb{E} [ \tau(G_i^{x_1, \ldots, x_{k-2}})] \leq \frac{k!}{k^{k-1}}\left(1+o(1)\right)\,. \end{equation} The procedures used in \cite{DalaiVenkatJaikumar2} and \cite{venkat} are based on the idea that one can also take $x_1,x_2,\ldots, x_{k-2}$ uniformly from a subset $C'\subset C$ which ensures they collide in all coordinates $i$ in some subset $T \subset \{1,2\ldots, n\}$. Then, if $g_{i,a}$ is the frequency of symbol $a$ in the coordinate $i\notin T$ of $C'$, one has \begin{align} \mathbb{E} [ & \tau(G_i^{x_1, \ldots, x_{k-2}})] \nonumber \\ & =\left(1+o(1)\right)\sum_{\stackrel{\text{distinct }}{ a_1,\ldots,a_{k-2}}} g_{i,a_1}g_{i,a_2}\cdots g_{i,a_{k-2}}(1-f_{i,a_1}\cdots-f_{i,a_{k-2}} )\label{eq:exptaugf} \end{align} The worst case $g$ and $f$ here, if taken independently, give in general a value which exceeds the $k!\slash k^{k-1}$ of \eqref{eq:fredmankomlostau}. In \cite{DalaiVenkatJaikumar2}, for $k=4$, it was shown that one can deal with this by also taking $C'$ randomly from a partition of $C$ (based on the values in positions $i\in T$), thus adding an additional (outer) expectation. In that case $g_i$ is also random and constrained to satisfy $\mathbb{E}[g_i]=f_i$. Using some concavity argument it was shown that under this random selection the bound \eqref{eq:fredmankomlostau} still holds for $i\notin T$, thus gaining on average compared to \cite{FredmanKomlos}. However, for $k>4$ that approach seems infeasible. The idea used in \cite{venkat} is to suppress the random selection of $C'$ and show that one can carefully choose $C'$ so that $x_1, \ldots, x_{k-2}$ collide in a portion of the coordinates large enough to more than compensate the increase in $\mathbb{E}[\tau(G_i^{x_1, \ldots, x_{k-2}})]$ for $i\notin T$ obtained in \eqref{eq:exptaugf} with respect to \eqref{eq:fredmankomlostau}. This leads to a proof that \eqref{eq:fredmankomlosRk} is not tight for all $k>4$. However, explicit numerical improvements were only proved for $k=5,6$, and given for $k>6$ modulo a conjecture on the optimal value of some polynomials. In the next two sections we present our contribution. First we prove the conjecture formulated by the authors in \cite{venkat}, thus completing their proof of the new bounds on $R_k$ for all $k$. Then, we prove stronger results for $k=5,6$. Our idea is based on a symmetrization of \eqref{eq:exptaugf} which allow us to resurrect the random selection of $C'$ in an effective way, replacing the concavity argument of \cite{DalaiVenkatJaikumar2} with new bounds on the maxima of some polynomials. \section{Guruswami-Riazanov bounds} \label{sec:venkat_conj} A crucial role in all bounds discussed in this paper is played by the sum appearing in equation \eqref{eq:exptaugf}. We simplify the notation and set, for general probability vectors $g=(g_1,\ldots,g_k)$ and $f=(f_1,\ldots,f_k)$, \begin{equation} \psi(g,f) =\sum_{\sigma\in S_k} g_{\sigma(1)}g_{\sigma(2)}\cdots g_{\sigma(k-2)}f_{\sigma(k-1)}\,, \end{equation} observing that equation \eqref{eq:exptaugf} can be rewritten as \begin{equation} \label{eq:Expinpsi} \mathbb{E} [ \tau(G_i^{x_1, \ldots, x_{k-2}})] = \left(1+o(1)\right)\psi(g_i,f_i) \end{equation} We can now prove the conjecture stated in \cite{venkat}. \begin{Proposition}[Conjecture 1 \cite{venkat}]\label{conjecture} Under the constraints $f_i\geq \gamma,\forall i$, $\psi(g,f)$ attains a maximum in a point $(g,f)$ with vector $f$ of the form $f=(\gamma,\dots,\gamma,1-(k-1)\gamma)$. \end{Proposition} \proof Since $\psi(g,f)$ is invariant under (identical) permutations on $g$ and $f$, we can study maxima for which $g_k$ is the minimum among the values $g_1,g_2,\dots, g_k$ and show that for those points $f=(\gamma,\dots,\gamma,1-(k-1)\gamma)$. We prove this by considering the components of $f$ one by one. Assume on the contrary that $f_1>\gamma$. Given $\epsilon \leq f_1-\gamma$, set $\tilde{f}=(f_1-\epsilon,f_2\ldots,f_{k-1},f_k+\epsilon)$. Then \begin{align*} \psi(g,\tilde{f}) & = \sum_{\sigma:\sigma(k-1)\neq 1,k} g_{\sigma(1)}g_{\sigma(2)}\dots g_{\sigma(k-2)}f_{\sigma(k-1)}\\ &\qquad + \sum_{\sigma:\sigma(k-1)=1} g_{\sigma(1)}g_{\sigma(2)}\dots g_{\sigma(k-2)}(f_1-\epsilon)\\ &\qquad + \sum_{\sigma:\sigma(k-1)=k} g_{\sigma(1)}g_{\sigma(2)}\dots g_{\sigma(k-2)}(f_k+\epsilon)\\ & = \psi(g,f) - \epsilon\cdot \sum_{\sigma:\sigma(k-1)=1} g_{\sigma(1)}g_{\sigma(2)}\dots g_{\sigma(k-2)}\\ & \qquad + \epsilon \cdot \sum_{\sigma:\sigma(k-1)=k} g_{\sigma(1)}g_{\sigma(2)}\dots g_{\sigma(k-2)}\,. \end{align*} Since we assumed $g_1\geq g_k$, \begin{equation} \sum_{\sigma:\sigma(k-1)=1} g_{\sigma(1)}g_{\sigma(2)}\dots g_{\sigma(k-2)}\leq \sum_{\sigma:\sigma(k-1)=k} g_{\sigma(1)}g_{\sigma(2)}\dots g_{\sigma(k-2)}. \end{equation} and hence $\psi(g,\tilde{f})\geq \psi(g,f)$. By repeating the above procedure for $f_2$, $f_3,\ldots,f_{k-1}$, we find that indeed $f=(\gamma,\dots,\gamma,1-(k-1)\gamma)$ maximizes $\psi(g,f)$ under the considered constraints whenever $g_k$ is the minimum among $g_1,g_2,\dots, g_k$, and in particular for the optimal $g$ sorted in this way.\endproof It terms of $g$, it was already shown in \cite{venkat} that assuming the above result one could show that the maximum value of $\psi(g,f)$, under the constraint that $f_i\geq \gamma,\forall i$, is attained at a point $(g,f)$ with $g$ of the form $(\beta,\beta,\ldots,1-(k-1)\beta)$. Assuming this, it was shown in \cite{venkat} that a new explicit numerical bound can be given on $R_k$ which strictly improves the Fredman-Koml\'os bound for all $k$. Table \ref{tab:numbounds} gives numerical results\footnote{We believe that due to a minor error in the computation, the bound given for $R_5$ in \cite{venkat} is not really the best possible using their method. We report here the optimal.} for the first values of $k$. \begin{table} \centering \begin{tabular}{|c|c|c|c|c|} \hline $k$ & 5 & 6 & 7 & 8 \\ \hline Bound from \cite{FredmanKomlos} & 0.19200 & 0.092593 & 0.04284 & 0.019227\\ \hline Bound from \cite{venkat} & 0.19079 & 0.092279 & 0.04279 & 0.019213\\ \hline \end{tabular} \caption{Numerical values for the bounds on $R_k$ from \cite{FredmanKomlos} and from \cite{venkat} in light of Proposition \ref{conjecture}. All numbers are rounded upwards.} \label{tab:numbounds} \end{table} \section{Better bounds for small $k$} \label{sec:newbounds} In this section we combine insights from both the approaches of \cite{DalaiVenkatJaikumar2} and \cite{venkat}. Instead of looking at one subcode $C'$, as done in \cite{venkat}, we follow the idea in \cite{DalaiVenkatJaikumar2}. We consider a partition $\{C_{\omega}: \omega \in \Omega\}$ of our $k$-hash code $C$ and randomly select a subcode $C_{\omega}$. Then we randomly extract codewords $x_1,\ldots,x_{k-2}$ from $C_{\omega}$ and bound the expected value in \eqref{eq:exptaugf} over both random code and codewords. At this point, we replace the concavity argument of \cite{DalaiVenkatJaikumar2} with a symmetrization trick combined with new bounds on the maxima of certain polynomials. This procedure leads to the following nontrivial improvement on the rates $R_5$ and $R_6$. \begin{Theorem}\label{main} For $k=5,6$ the following bounds hold \begin{itemize} \item $R_5\leq 0.1697$; \item $R_6\leq 0.0875$. \end{itemize} \end{Theorem} \subsection{Proof of Theorem \ref{main}} Here our goal is to find a family of subcodes such that any $k-2$ codewords $x_1,x_2,\dots,x_{k-2}$ of a given subcode $C_{\omega}$ collide in all coordinates of $T=[1,\ell]$ for a carefully chosen value of $\ell$, that is, for any coordinate $t\in T$ there exist $i,j$ such that $x_{i,t}=x_{j,t}$. This will ensure that the coordinates from $T$ contribute $0$ to the LHS of \eqref{eq:hansel_hash}. To do this, we cover all the possible prefixes of length $\ell$; the following lemma can be seen as a special case of the known results on the fractional clique covering number (see \cite{PartialCovering}). \begin{Lemma}\label{covering} For any positive $\epsilon$, for $\ell$ large enough, there exists a partition $\Omega$ of $\{1,2,\dots,k\}^{\ell}$ such that: \begin{enumerate} \item $|\Omega|\leq \left\lfloor{\left(\frac{k}{k-3}+\epsilon\right)^{\ell}}\right\rfloor$. \item For all $\omega\in\Omega$ and $i=1,\ldots,\ell$, the $i$-th projection of $\omega$ has cardinality at most $ k-3$. \end{enumerate} In particular, for any $\omega\in\Omega$, any $k-2$ sequences in $\omega$ collide in all coordinates $i=1,\ldots, \ell$. \end{Lemma} \proof For any $i\in [1,k]$, consider the set $A_i=\{i,i+1,\dots,i+(k-4)\}$ , where the sums are performed modulo $k$ in $[1,k]$. To a string $s=(i_1,\dots,i_{\ell})$ in $[1,k]^{\ell}$ we associate a set $\omega_s=A_{i_1}\times A_{i_2}\times \dots \times A_{i_{\ell}}\subset [1,k]^\ell$. Fix a word $x\in [1,k]^{\ell}$, and choose uniformly at random the string $s$; the probability that $x\not\in \omega_s$ is $1-\left(\frac{k-3}{k}\right)^{\ell}$. Therefore, if we choose randomly $h$ strings $s_1,\dots,s_h$, the probability that $x\not \in (\omega_{s_1}\cup \dots \cup \omega_{s_h})$ is $\left(1-\left(\frac{k-3}{k}\right)^{\ell}\right)^h$. Hence, the expected number of words $x\in [1,k]^{\ell}$ that do not belong to any of the $\omega_{s_1}, \dots, \omega_{s_h}$ is \begin{align*} \mathbb{E}(|\{x\in[1,k]^ \ell:\ x\not \in \omega_{s_1}\cup \dots \cup \omega_{s_h}\}|) = k^{\ell}\left(1-\left(\frac{k-3}{k}\right)^{\ell}\right)^h. \end{align*} If this value is smaller than $1$, then there exists a choice of $s_1,\dots, s_h$ such that that the family $\{\omega_{s_1},\dots, \omega_{s_h}\}$ covers the whole set $[1,k]^{\ell}$. This happens whenever $$k^{\ell}\left(1-\left(\frac{k-3}{k}\right)^{\ell}\right)^h<1 $$ or equivalently $$h>\frac{-\ell \log{k}}{\log\left(1-\left(\frac{k-3}{k}\right)^{\ell}\right)}\,, $$ which holds for $$h>\ell \left(\frac{k}{k-3}\right)^{\ell} \frac{\log{k}}{\log{e}}.$$ For $\ell$ large enough, setting $h=\left\lfloor{\left(\frac{k}{k-3}+\epsilon\right)^{\ell}}\right\rfloor$ we have the desired inequality. Removing possible intersections between the sets $\omega_s$ we obtain a partition of $[1,k]^\ell$ with the desired properties, since condition 2) is satisfied by construction. \endproof Let $\Omega=\{\omega_1,\ldots,\omega_h\}$ be a partition of $[1,k]^\ell$ as derived from Lemma \ref{covering} and consider the family of subcodes $C_{\omega_1},\dots, C_{\omega_h}$ of $C$ defined by $$C_\omega=\{x\in C: (x_1,x_2,\ldots, x_{\ell})\in \omega\}.$$ Clearly, any $k-2$ codewords $x_1,x_2,\dots,x_{k-2}$ of a given subcode $C_{\omega}$ collide in all coordinates of $T=[1,\ell]$. As in \cite{DalaiVenkatJaikumar2}, define a subcode $C_{\omega}$ to be \emph{heavy} if $|C_{\omega}|> n$ and to be \emph{light} otherwise. We can show that, if $\ell$ is not too large, most of the codewords are contained in heavy subcodes. Indeed, if we consider $\ell$ such that $\left(\frac{k}{k-3}+\epsilon\right)^{\ell}\leq 2^{nR-2\log{n}}$, that is $\ell\leq\frac{nR-2\log{n}}{\log\left(\frac{k}{k-3}+\epsilon\right)}$, we have that $$\left|\bigcup_{C_{\omega} \mbox{ is ligth}} C_{\omega}\right|\leq n\left(\frac{k}{k-3}+\epsilon\right)^{\ell}\leq n2^{nR-2\log{n}}= \frac{|C|}{n}.$$ This means that at least a fraction $(1-1/n)$ of the codewords are in heavy subcodes. If we remove from $C$ the light codes, the rate changes by an amount $\frac{1}{n}\log(1-1/n)$, which vanishes as $n$ grows. So, in the following we can assume, without loss of generality, that all the subcodes are heavy. We are finally ready to describe our strategy to pick the codewords $x_1,\dots,x_{k-2}$: first we choose a subcode $C_{\omega}$ with probability $\lambda_{\omega}=|C_{\omega}|/|C|$ and then we pick uniformly at random (and without replacement) $x_1,\dots,x_{k-2}$ from $C_{\omega}$. Since those codewords collide in all the coordinates from the set $T=[1,\ell]$, we obtain in \eqref{eq:hansel_hash}: \begin{align} \log(|C|-k+2)& \leq\mathbb{E}_{\omega\in \Omega}(\mathbb{E}[\sum_{i\in [\ell+1,n]}\tau(G_i^{x_1,x_2,\dots,x_{k-2}})]) \\ &=\sum_{i\in [\ell+1,n]}\mathbb{E}_{\omega\in \Omega}(\mathbb{E}[\tau(G_i^{x_1,x_2,\dots,x_{k-2}})])\label{eq:sumellton}. \end{align} Let again $f_{i}$ be probability distribution of the $i$-th coordinate of $C$, and let $f_{i|\omega}$ be the distribution of the subcode $C_\omega$. Invoking \eqref{eq:Expinpsi} for the expectation over the random choice of $x_1,\ldots,x_{k-2}$, we can write for $i\in [\ell+1,n]$ \begin{align*} \mathbb{E}_{\omega\in \Omega} (\mathbb{E}[\tau(G_i^{x_1,x_2,\dots,x_{k-2}})]) =(1+o(1))\sum_{\omega\in \Omega} \lambda_{\omega} \psi(f_{i|\omega},f_i). \end{align*} Since $f_i=\sum_{\mu \in \Omega} \lambda_{\mu} f_{i|\mu}$ and $\psi$ is linear in its second variable, we have that $$ \mathbb{E}_{\omega\in \Omega}(\mathbb{E}[\tau(G_i^{x_1,x_2,\dots,x_{k-2}})]) =(1+o(1))\sum_{\omega,\mu\in \Omega}\lambda_{\omega}\lambda_{\mu}\psi(f_{i|\omega},f_{i|\mu})\,. $$ We exploit now a simple yet effective trick. Since the sum above is symmetric in $\omega$ and $\mu$, we can write \begin{align} \mathbb{E}_{\omega\in \Omega} (\mathbb{E}[\tau &(G_i^{x_1,x_2,\dots,x_{k-2}})]) \nonumber\\ &=\left(1+o(1)\right)\frac{1}{2}\sum_{\omega,\mu\in \Omega}\lambda_{\omega}\lambda_{\mu}[\psi(f_{i|\omega},f_{i|\mu})+\psi(f_{i|\mu},f_{i|\omega})].\label{simmetrizzata} \end{align} Here, we note that $f_{i|\omega}$ has no relation with $f_{i|\mu}$. Therefore we can just consider the following polynomial function over two generic probability vectors $p=(p_1,p_2,\dots,p_k)$ and $q=(q_1,q_2,\dots,q_k)$ \begin{align} \Psi(p;q)& : =\psi(p,q)+\psi(q,p)\nonumber\\ &=\sum_{\sigma\in S_k} p_{\sigma(1)}p_{\sigma(2)}\dots p_{\sigma(k-2)}q_{\sigma(k-1)}+ q_{\sigma(1)}q_{\sigma(2)}\dots q_{\sigma(k-2)}p_{\sigma(k-1)}.\label{eq:defPsi} \end{align} Because of \eqref{simmetrizzata}, if $M_k$ is the maximum of $\Psi$ over probabilistic vectors $p$ and $q$, equation \eqref{eq:sumellton} says that \begin{align*} \log{|C|} & \leq (1+o(1))\frac{1}{2}(n-\ell)\sum_{\omega,\mu\in \Omega}\lambda_{\omega}\lambda_{\mu}M_k\\ & =(1+o(1))\frac{1}{2}(n-\ell) M_k. \end{align*} Recalling that $|C|=2^{n R}$ and taking $\ell=\left\lfloor{\frac{nR-2\log{n}}{\log\left(\frac{k}{k-3}+\epsilon\right)}}\right\rfloor$, we obtain \begin{align*} R\leq (1+o(1))\left[1-\frac{R-2\log(n)/n}{\log\left(\frac{k}{k-3}+\epsilon\right)}\right]\frac{M_k}{2}. \end{align*} Rearranging the terms, taking $n\to\infty$ first and then $\epsilon\to 0$, we deduce the following proposition. \begin{Proposition}\label{FromMaxToRate} Let $M_k$ be the maximum of $\Psi$ over probabilistic vectors $p=(p_1,p_2,\dots,p_k)$ and $q=(q_1,q_2,\dots,q_k)$. Then we have the following upperbound on $R_k$ $$R_k\leq \left(\frac{2}{M_k}+\frac{1}{\log(k/(k-3))}\right)^{-1}.$$ \end{Proposition} In the next subsection we will prove that $M_5=\frac{15(48+\sqrt{5})}{1936}\approx 0.389226$ and $M_6=24/125$. This implies Theorem \ref{main}. \subsection{Bounds on $\Psi$} The goal of this subsection is to find the maximum of the function $\Psi$ as defined in \eqref{eq:defPsi}. For this purpose we first introduce two lemmas that provide some restrictions on this maximum. \begin{Lemma}\label{lagrange1} Let $\bar{p}=(\bar{p}_1,\dots,\bar{p}_k)$ and $\bar{q}=(\bar{q}_1,\dots,\bar{q}_k)$ be two probabilistic vectors. If $(\bar{p};\bar{q})$ is a maximum for $\Psi$ such that $\bar{p}_1,\bar{p}_2,\bar{q}_1,\bar{q}_2$ are nonzero, then also $(\frac{\bar{p}_1+\bar{p}_2}{2},\frac{\bar{p}_1+\bar{p}_2}{2},\bar{p}_3,\dots,\bar{p}_k;\frac{\bar{q}_1+\bar{q}_2}{2},\frac{\bar{q}_1+\bar{q}_2}{2},\bar{q}_3,\dots,\bar{q}_k)$ is a maximum for $\Psi$. \end{Lemma} \proof If $\bar{P}=(\bar{p};\bar{q})$ is a maximum for $\Psi(p;q)$ under the constraints $p_1+p_2+\dots+p_k=1$ and $q_1+q_2+\dots+q_k=1$, then it is a maximum also under the stronger constraints $p_1+p_2=c_1$, $q_1+q_2=c_2$ where $c_1=\bar{p}_1+\bar{p}_2$, $c_2=\bar{q}_1+\bar{q}_2$, and $p_i=\bar{p}_i,q_i=\bar{q}_i$ for $i\in\{3,4,\dots,k\}$. Because of the Lagrange multiplier method this means that: $$\frac{\partial \Psi}{\partial p_1}\Big|_{\bar{P}}=\frac{\partial \Psi}{\partial p_2}\Big|_{\bar{P}}$$ and $$\frac{\partial \Psi}{\partial q_1}\Big|_{\bar{P}}=\frac{\partial \Psi}{\partial q_2}\Big|_{\bar{P}}\,.$$ It follows that: $$(\bar{p}_1-\bar{p}_2)a+(\bar{q}_1-\bar{q}_2)b=0$$ and $$(\bar{q}_1-\bar{q}_2)d+(\bar{p}_1-\bar{p}_2)c=0$$ where $a=\frac{\partial^2 \Psi}{\partial p_1\partial p_2}\big|_{\bar{P}}$, $b=\frac{\partial^2 \Psi}{\partial p_1\partial q_2}\big|_{\bar{P}}=\frac{\partial^2 \Psi}{\partial q_1\partial p_2}\big|_{\bar{P}}=c$ and $d=\frac{\partial^2 \Psi}{\partial q_1\partial q_2}\big|_{\bar{P}}$. If we set $\bar{p}_1-\bar{p}_2=x$, $\bar{q}_1-\bar{q}_2=y$, the previous equations became: $$\begin{cases} ax+by=0;\\ cx+dy=0. \end{cases}$$ In the case $ad-bc\not=0$ the previous system admits only the solution $x=y=0$ that means $\bar{p}_1=\bar{p}_2$ and $\bar{q}_1=\bar{q}_2$. It is clear that here we have $\bar{p}_1=\frac{\bar{p}_1+\bar{p}_2}{2}=\bar{p}_2$, $\bar{q}_1=\frac{\bar{q}_1+\bar{q}_2}{2}=\bar{q}_2$ and hence the thesis is satisfied. Let us assume $ad-bc=0$. Then there exists a line $L$ of points $P(t)$ such that $P(1)=\bar{P}$, $P(0)=(\frac{\bar{p}_1+\bar{p}_2}{2},\frac{\bar{p}_1+\bar{p}_2}{2},\bar{p}_3,\dots,\bar{p}_k;\frac{\bar{q}_1+\bar{q}_2}{2},\frac{\bar{q}_1+\bar{q}_2}{2},\bar{q}_3,\dots,\bar{q}_k)$ and $$\frac{\partial \Psi}{\partial p_1}\Big|_{P(t)}-\frac{\partial \Psi}{\partial p_2}\Big|_{P(t)}=\frac{\partial \Psi}{\partial q_1}\Big|_{P(t)}-\frac{\partial \Psi}{\partial q_2}\Big|_{P(t)}=0.$$ It follows that $\Psi(P(t))$ is constantly equal to the value of $\Psi$ in $\bar{P}=P(1)$. Since $(\frac{\bar{p}_1+\bar{p}_2}{2},\frac{\bar{p}_1+\bar{p}_2}{2},\bar{p}_3,\dots,\bar{p}_k,\frac{\bar{q}_1+\bar{q}_2}{2},\frac{\bar{q}_1+\bar{q}_2}{2},\bar{q}_3,\dots,\bar{q}_k)$ belongs to the line $L$, this point is also a maximum for $\Psi$. \endproof With essentially the same proof we also obtain the following result. \begin{Lemma}\label{lagrange2} Let $\bar{p}=(\bar{p}_1,\dots,\bar{p}_k)$ and $\bar{q}=(\bar{q}_1,\dots,\bar{q}_k)$ be two probabilistic vectors. If $(\bar{p};\bar{q})$ is a maximum for $\Psi$ such that $\bar{p}_1,\bar{p}_2$ are nonzero while $\bar{q}_1=\bar{q}_2=0$ then also $(\frac{\bar{p}_1+\bar{p}_2}{2},\frac{\bar{p}_1+\bar{p}_2}{2},\bar{p}_3,\dots,\bar{p}_k;0,0,\bar{q}_3,\dots,\bar{q}_k)$ is a maximum for $\Psi$. \end{Lemma} In the next two lemmas, we will provide some further restrictions on the maximum of $\Psi$ using just some combinatorial arguments. \begin{Lemma}\label{LemmaAmmazzaCasi1} We have that: $$\Psi(0,p_2,\dots,p_k;0,q_2,\dots, q_k)\leq \Psi(0,p_2,\dots,p_k;q_2,0,q_3,\dots, q_k).$$ \end{Lemma} \proof Because of the definition, we have that $\Psi(0,p_2,\dots,p_k;0,q_2,\dots, q_k)$ evaluates as $$\sum_{\sigma:\ \sigma(k)=1} p_{\sigma(1)}p_{\sigma(2)}\dots p_{\sigma(k-2)}q_{\sigma(k-1)}+ q_{\sigma(1)}q_{\sigma(2)}\dots q_{\sigma(k-2)}p_{\sigma(k-1)}.$$ Similarly, we have that $\Psi(0,p_2,\dots,p_k;q_2,0,q_3,\dots, q_k)$ equals \begin{align*} \sum_{\sigma:\ \sigma(k)=1} & p_{\sigma(1)}p_{\sigma(2)}\dots p_{\sigma(k-2)}q_{\sigma(k-1)}+ q_{\sigma(1)}q_{\sigma(2)}\dots q_{\sigma(k)}p_{\sigma(k-1)}+\\ & (k-2)p_2q_2\left(\sum_{\sigma\in Sym(3,\dots,k)} p_{\sigma(3)}\dots p_{\sigma(k-1)}+ q_{\sigma(3)}\dots q_{\sigma(k-1)}\right). \end{align*} The claim follows since each term of the last sum is non negative. \endproof The following Lemma is in the same spirit of Proposition \ref{conjecture}. \begin{Lemma}\label{LemmaAmmazzaCasi2} We have that: $$\Psi(p_1,\dots,p_{k-3},0,0,0;q_1,q_2,\dots, q_k)\leq \Psi\left(1,0\dots,0;0,\frac{1}{(k-1)},\dots,\frac{1}{(k-1)}\right).$$ \end{Lemma} \proof We suppose, without loss of generality that $q_1$ is the minimum among the values $q_1,q_2,\dots,q_{k-3}$. Setting $p=(p_1,\dots,p_{k-3},0,0,0)$ and $q=(q_1,\dots,q_k)$, we have \begin{align*} \Psi(p;q)& = \sum_{\sigma:\ \sigma(k-1)\not\in \{1,2\}} q_{\sigma(1)}q_{\sigma(2)}\dots q_{\sigma(k-2)}p_{\sigma(k-1)}\\& +\frac{p_1+p_2}{2}\sum_{\sigma:\ \{\sigma(k-1), \sigma(k)\}=\{1,2\}} q_{\sigma(1)}\dots q_{\sigma(k-2)}\\ &+(p_1q_2+q_1p_2)(k-2)\sum_{\sigma\in Sym(3,\dots,k)} q_{\sigma(3)}\dots q_{\sigma(k-1)}.\end{align*} Similarly, setting $p'=(p_1+p_2,0,p_3,\dots,p_{k-3},0,0,0)$, we have that: \begin{align*} \Psi(p';q)& = \sum_{\sigma:\ \sigma(k-1)\not\in \{1,2\}} q_{\sigma(1)}q_{\sigma(2)}\dots q_{\sigma(k-2)}p_{\sigma(k-1)}\\&+ \frac{p_1+p_2}{2}\sum_{\sigma:\ \{\sigma(k-1), \sigma(k)\}=\{1,2\}} q_{\sigma(1)}\dots q_{\sigma(k-2)}\\& +(p_1+p_2)q_2(k-2)\sum_{\sigma\in Sym(3,\dots,k)} q_{\sigma(3)}\dots q_{\sigma(k-1)}. \end{align*} Since $q_1\leq q_2$ we have that $$ \Psi(p;q)\leq \Psi(p';q)\,. $$ Reiterating the previous procedure, since $q_1$ is the minimum among the values $q_1,\dots,q_{k-3}$, we obtain \begin{equation}\label{maggiorazione}\Psi(p_1,\dots,p_{k-3},0,0,0;q_1,q_2,\dots, q_k)\leq \Psi(1,0,\dots,0,0;q_1,q_2,\dots, q_k).\end{equation} Since $q_1$ does not appear in the value of $\Psi(1,0,\dots,0,0;q_1,q_2,\dots, q_k)$, this is certainly maximized for $q_1=0$. Finally, due to the Muirhead's inequality, we obtain that the RHS of \eqref{maggiorazione} is maximized for $q_2=q_3=\dots=q_k=\frac{1}{k-1}$. \endproof As a consequence of the previous lemmas, $\Psi$ attains a maximum in a point of one of the following types: \begin{itemize} \item[a)] $\left(1,0\dots,0;0,\frac{1}{(k-1)},\dots,\frac{1}{(k-1)}\right)$; \item[b)] $(1/k,\dots,1/k;1/k,\dots,1/k)$; \item[c)] $(0,0,\alpha,\dots,\alpha,\beta,\beta;\gamma,\gamma,\delta,\dots,\delta,0,0)$\\ where $(k-4)\alpha+2\beta=1$ and $2\gamma+(k-4)\delta=1$; \item[d)] $(0,0,\alpha,\dots,\alpha,\beta;\gamma,\gamma,\delta,\dots,\delta,0)$\\ where $(k-3)\alpha+\beta=1$ and $2\gamma+(k-3)\delta=1$; \item[e)] $(0,0,1/(k-2),\dots,1/(k-2);\gamma,\gamma,\delta,\dots,\delta)$\\ where $2\gamma+(k-2)\delta=1$; \item[f)] $(0,\alpha,\dots,\alpha,\beta;\gamma,\delta,\dots,\delta,0)$\\ where $(k-2)\alpha+\beta=1$ and $\gamma+(k-2)\delta=1$; \item[g)] $(0,1/(k-1),\dots,1/(k-1);\gamma,\delta,\dots,\delta)$\\ where $\gamma+(k-1)\delta=1$. \end{itemize} In particular, because of Lemma \ref{LemmaAmmazzaCasi2}, a maximum with three or more $p$-coordinates (resp. $q$-coordinates) equal to zero is also attained in a point of the form $(a)$. Otherwise, there are at most two zero coordinates both for the vector $p$ and for the vector $q$. Due to Lemma \ref{LemmaAmmazzaCasi1}, we can then assume those zeros are in different positions and finally, using Lemma \ref{lagrange1} and \ref{lagrange2}, we obtain the required characterization of the maximum. For $k=5,6$, we have inspected using Mathematica all cases listed above and determined the maximum explicitly. \begin{Theorem}\label{max} \begin{itemize} The following hold: \item for $k=5$, the global maximum of $\Psi$ is $\frac{15(48+\sqrt{5})}{1936}\approx 0.389226$ and is obtained in case $(g)$ with $\delta=1/44(4+\sqrt{5})$ and $\gamma=1-4\delta$; \item for $k=6$, the global maximum of $\Psi$ is $24/125=0.192$, obtained in case $(a)$. \end{itemize} \end{Theorem} Theorem \ref{main} follows immediately from Theorem \ref{max} and Proposition \ref{FromMaxToRate}. \begin{Remark} For $k>6$, the value obtained for $p$ and $q$ as in case $(a)$, which we conjecture to be the true maximum, is too big to improve the known upper bounds on $R_k$. \end{Remark} \section*{Acknowledgements} This research was partially supported by Italian Ministry of Education under Grant PRIN 2015 D72F16000790001. Helpful discussions with Jaikumar Radhakrishnan and Venkatesan Guruswami are gratefully acknowledged.

https://arxiv.org/abs/1502.05030

Spherical sets avoiding a prescribed set of angles

Let $X$ be any subset of the interval $[-1,1]$. A subset $I$ of the unit sphere in $R^n$ will be called \emph{$X$-avoiding} if $<u,v >\notin X$ for any $u,v \in I$. The problem of determining the maximum surface measure of a $\{ 0 \}$-avoiding set was first stated in a 1974 note by Witsenhausen; there the upper bound of $1/n$ times the surface measure of the sphere is derived from a simple averaging argument. A consequence of the Frankl-Wilson theorem is that this fraction decreases exponentially, but until now the $1/3$ upper bound for the case $n=3$ has not moved. We improve this bound to $0.313$ using an approach inspired by Delsarte's linear programming bounds for codes, combined with some combinatorial reasoning. In the second part of the paper, we use harmonic analysis to show that for $n\geq 3$ there always exists an $X$-avoiding set of maximum measure. We also show with an example that a maximiser need not exist when $n=2$.

\section{Introduction} Witsenhausen \cite{witsenhausen74} in 1974 presented the following problem: Let $S^{n-1}$ be the unit sphere in $\mathbb{R}^n$ and suppose $I \subset S^{n-1}$ is a Lebesgue measurable set such that no two vectors in $I$ are orthogonal. What is the largest possible Lebesgue surface measure of $I$? Let $\alpha(n)$ denote the supremum of the measures of such sets $I$, divided by the total measure of $S^{n-1}$. The first upper bounds for $\alpha(n)$ appeared in \cite{witsenhausen74}, where Witsenhausen deduced that $\alpha(n) \leq 1/n$. In \cite{frankl-wilson81} Frankl and Wilson proved their powerful combinatorial result on intersecting set systems, and as an application they gave the first exponentially decreasing upper bound $\alpha(n) \leq (1+o(1))(1.13)^{-n}$. Raigorodskii \cite{raigorodskii99} improved the bound to $(1+o(1))(1.225)^{-n}$ using a refinement of the Frankl-Wilson method. Gil Kalai conjectured in his weblog \cite{kalai09} that an extremal example is to take two opposite caps, each of geodesic radius $\pi/4$; if true, this implies that $\alpha(n) = (\sqrt{2} + o(1))^{-n}$. Besides being of independent interest, the above \emph{Double Cap Conjecture} is also important because, if true, it would imply new lower bounds for the measurable chromatic number of Euclidean space, which we now discuss. Let $c(n)$ be the smallest integer $k$ such that $\mathbb{R}^n$ can be partitioned into sets $X_1, \dots, X_k$, with $\|x-y\|_2 \neq 1$ for each $x,y \in X_i$, $1 \leq i \leq k$. The number $c(n)$ is called the \emph{chromatic number of $\mathbb{R}^n$}, since the sets $X_1,\dots,X_k$ can be thought of as colour classes for a proper colouring of the graph on the vertex set $\mathbb{R}^n$, in which we join two points with an edge when they have distance $1$. Frankl and Wilson \cite[Theorem~3]{frankl-wilson81} showed that $c(n) \geq (1+o(1))(1.2)^n$, proving a conjecture of Erd\H{o}s that $c(n)$ grows exponentially. Raigorodskii in 2000 \cite{raigorodskii00} improved the lower bound to $(1+o(1))(1.239)^n$. Requiring the classes $X_1,\dots,X_k$ to be Lebesgue measurable yields the \emph{measurable chromatic number} $c_m(n)$. Clearly $c_m(n) \geq c(n)$. Remarkably, it is still open if the inequality is strict for at least one $n$, although one can prove better lower bounds on $c_m(n)$. In particular, the exponent in Raigorodskii's bound was recently beaten by Bachoc, Passuello and Thiery \cite{bachoc14} who showed that $c_m(n) \geq (1.268+o(1))^n$. If the Double Cap Conjecture is true, then $c_m(n)\ge (\sqrt{2}+o(1))^n$ because, as it is not hard to show, $c_m(n) \geq 1/\alpha(n)$ for every $n\ge 2$. Note that the best known asymptotic upper bound on $c_m(n)$ (as well as on $c(n)$) is $(3+o(1))^n$, by Larman and Rogers~\cite{larman+rogers:72}. Despite progress on the asymptotics of $\alpha(n)$, the upper bound of $1/3$ for $\alpha(3)$ has not been improved since the original statement of the problem in \cite{witsenhausen74}. Note that the two-cap construction gives $\alpha(3)\ge 1-1/\sqrt{2}=0.2928...$\ . Our first main result is that $\alpha(3) < 0.313$. The proof involves tightening a Delsarte-type linear programming upper bound (cf.\ \cite{delsarte73}, \cite{delsarte77}, \cite{bachoc09}, \cite{oliveira09}) by adding combinatorial constraints. Let $\mathcal{L}$ be the $\sigma$-algebra of Lebesgue surface measurable subsets of $S^{n-1}$, and let $\lambda$ be the surface measure, for simplicity normalised so that $\lambda(S^{n-1}) = 1$. For $X \subset [-1,1]$, a subset $I \subset S^{n-1}$ will be called \emph{$X$-avoiding} if $\langle \xi, \eta \rangle \notin X$ for all $\xi, \eta \in I$, where $\langle \xi, \eta \rangle$ denotes the standard inner product of the vectors $\xi, \eta$. The corresponding extremal problem is to determine \begin{align}\label{eq:ir} \alpha_X(n):=\sup \{ \lambda(I)~:~\text{$I \in \mathcal{L}$, $I$ is $X$-avoiding} \}. \end{align} For example, if $t \in (-1,1)$ and $X=[-1,t)$, then $I \subset S^{n-1}$ is $X$-avoiding if and only if its geodesic diameter is at most $\arccos(t)$. Thus Levy's isodiametric inequality~\cite{levy:pcaf} shows that $\alpha_X$ is given by a spherical cap of the appropriate size. A priori, it is not clear that the value of $\alpha_X(n)$ is actually attained by some measurable $X$-avoiding set $I$ (so Witsenhausen \cite{witsenhausen74} had to use supremum to define $\alpha(n)$). We prove in Theorem~\ref{thm:attainment} that the supremum is attained as a maximum whenever $n \geq 3$. Remarkably, this result holds under no additional assumptions whatsoever on the set $X$. However, in a sense only closed sets $X$ matter: our Theorem~\ref{thm:closureIR} shows that $\alpha_X(n)$ does not change if we replace $X$ by its closure. When $n=2$ the conclusion of Theorem~\ref{thm:attainment} fails; that is, the supremum in \eqref{eq:ir} need not be a maximum: an example is given in Theorem \ref{thm:irCircle}. Besides also answering a natural question, the importance of the attainment result can also be seen through the historical lens: In 1838 Jakob Steiner tried to prove that a circle maximizes the area among all plane figures having some given perimeter. He showed that any non-circle could be improved, but he was not able to rule out the possibility that a sequence of ever improving plane shapes of equal perimeter could have areas approaching some supremum which is not achieved as a maximum. Only 40 years later in 1879 was the proof completed, when Weierstrass showed that a maximizer must indeed exist. The layout of the paper will be as follows. In Section \ref{sec:prelim} we make some general definitions and fix notation. In Section \ref{sec:comb} we prove a simple and general proposition giving combinatorial upper bounds for $\alpha_X(n)$; this is basically a formalisation of the method used by Witsenhausen in \cite{witsenhausen74} to obtain the $\alpha(n) \leq 1/n$ bound. We then apply the proposition to calculate $\alpha_X(2)$ when $|X|=1$. In Section \ref{sec:lp} we deduce linear programming upper bounds for $\alpha(n)$, in the spirit of the Delsarte bounds for binary \cite{delsarte73} and spherical \cite{delsarte77} codes. We then strengthen the linear programming bound in the $n=3$ case in Section \ref{sec:lp+comb} to obtain the first main result. In Section \ref{sec:max} we prove that the supremum $\alpha_X(n)$ is a maximum when $n \geq 3$, and in Section \ref{sec:closure} we show that $\alpha_X(n)$ remains unchanged when $X$ is replaced with its topological closure. In Section \ref{sec:single} we formulate a conjecture generalising the Double Cap Conjecture for the sphere in $\mathbb{R}^3$, in which other forbidden inner products are considered. \section{Preliminaries}\label{sec:prelim} If $u,v \in \mathbb{R}^n$ are two vectors, their standard inner product will be denoted $\langle u, v \rangle$. All vectors will be assumed to be column vectors. The transpose of a matrix $A$ will be denoted $A^t$. We denote by $SO(n)$ the group of $n \times n$ matrices $A$ over $\mathbb{R}$ having determinant $1$, for which $A^t A$ is equal to the identity matrix. We will think of $SO(n)$ as a compact topological group, and we will always assume its Haar measure is normalised so that $SO(n)$ has measure $1$. We denote by $S^{n-1}$ the set of unit vectors in $\mathbb{R}^n$, \[ S^{n-1} = \{ x \in \mathbb{R}^n : \langle x, x \rangle = 1 \}, \] equipped with its usual topology. The Lebesgue measure $\lambda$ on $S^{n-1}$ is always taken to be normalised so that $\lambda(S^{n-1}) = 1$. Recall that the standard surface measure of $S^{n-1}$ is \begin{equation}\label{eq:OmegaN} \omega_n = \frac{2 \pi^{n/2}}{\Gamma(n/2)}, \end{equation} where $\Gamma$ denotes Euler's gamma-function. The Lebesgue $\sigma$-algebra on $S^{n-1}$ will be denoted by $\mathcal{L}$. When $(X, \mathcal{M}, \mu)$ is a measure space and $1 \leq p < \infty$, we use \[ L^p(X) = \left\{ f :~\text{$f$ is an $\mathbb{R}$-valued $\mathcal{M}$-measurable function and $\int |f|^p\,\mathrm{d}\mu < \infty$} \right\}. \] For $f \in L^p(X)$, we define $\|f\|_p := \left( \int |f|^p\,\mathrm{d}\mu \right)^{1/p}$. Identifying two functions when they agree $\mu$-almost everywhere, we make $L^p(X)$ a Banach space with the norm $\| \cdot \|_p$. We will use bold letters (for example $\bm{X}$) for random variables. The expectation of a function $f$ of a random variable $\bm{X}$ will be denoted $\mathbb{E}_{\bm{X}}[f(\bm{X})]$, or just $\mathbb{E}[f(\bm{X})]$. The probability of an event $E$ will be denoted $\mathbb{P}[E]$. When $X$ is a set, we use $\mathbbm{1}_X$ to denote its characteristic function; that is $\mathbbm{1}_X(x) = 1$ if $x \in X$ and $\mathbbm{1}_X(x) = 0$ otherwise. If $G = (V,E)$ is a graph, a set $I$ is called \emph{independent} if $\{u,v\} \notin E$ for any $u,v \in I$. The \emph{independence number} $\alpha(G)$ of $G$ is the cardinality of a largest independent set in $G$. We define $\alpha_X(n)$ as in \eqref{eq:ir}, and for brevity we let $\alpha(n) = \alpha_{\{0\}}(n)$. \section{Combinatorial upper bound}\label{sec:comb} Let us begin by deriving a simple ``combinatorial'' upper bound for the quantity $\alpha_X(n)$. \begin{proposition}\label{pr:comb} Let $n \geq 2$ and $X \subset [-1,1]$. For a finite subset $V \subset S^{n-1}$, we let $H=(V,E)$ be the graph on the vertex set $V$ with edge set defined by putting $\{ \xi, \eta \} \in E$ if and only if $\langle \xi, \eta \rangle \in X$. Then $\alpha_X(n) \leq \alpha(H) / |V|$. \end{proposition} \begin{proof} Let $I \subset S^{n-1}$ be an $X$-avoiding set, and take a uniform $\bm{O}\in SO(n)$. Let the random variable $\bm{Y}$ be the number of $\xi\in V$ with $\bm{O}\xi\in I$. Since $\bm{O}\xi\in S^{n-1}$ is uniformly distributed for every $\xi\in V$, we have by the linearity of expectation that $\mathbb{E}(\bm{Y})=|V|\, \lambda(I)$. On the other hand, $\bm{Y} \le \alpha(H)$ for every outcome $\bm{O}$. Thus $\lambda(I) \leq \alpha(H)/|V|$. \end{proof} We next use Proposition~\ref{pr:comb} to find the largest possible Lebesgue measure of a subset of the unit circle in $\mathbb{R}^2$ in which no two points lie at some fixed forbidden angle. \begin{theorem}\label{thm:irCircle} Let $X = \{x\}$ and put $t = \frac{\arccos{x}}{2\pi}$. If $t$ is rational and $t = p/q$ with $p$ and $q$ coprime integers, then \begin{align*} \alpha_X(2) = \begin{cases} 1/2, &~\text{if $q$ is even,}\\ (q-1)/(2q), &~\text{if $q$ is odd.} \end{cases} \end{align*} In this case $\alpha_X(2)$ is attained as a maximum. If $t$ is irrational then $\alpha_X(2) = 1/2$, but there exists no measurable $X$-avoiding set $I \subset S^1$ with $\lambda(I) = 1/2$. \end{theorem} \begin{proof} Write $\alpha = \alpha_X(2)$, and identify $S^1$ with the interval $[0,1)$ via the map $(\cos x, \sin x) \mapsto x/2\pi$. We regard $[0,1)$ as a group with the operation of addition modulo $1$. Notice that $I \subset [0,1)$ is $X$-avoiding if and only if $I \cap (t+I) = \emptyset$. This implies immediately that $\alpha \leq 1/2$ for all values of~$x$. Now suppose $t = p/q$ with $p$ and $q$ coprime integers, and suppose that $q$ is even. Let $S$ be any open subinterval of $[0,1)$ of length $1/q$, and define $T : [0,1) \to [0,1)$ by $T x = x+t \mod 1$. Using the fact that $p$ and $q$ are coprime, one easily verifies that $I = S \cup T^2 S \cup \cdots \cup T^{q-4} S \cup T^{q-2} S$ has measure $1/2$. Also $S$ is $X$-avoiding since $T S = T S \cup T^3 S \cup \cdots \cup T^{q-3} S \cup T^{q-1} S$ is disjoint from $S$. Therefore $\alpha = 1/2$. Next suppose $q$ is odd. With notation as before, a similar argument shows that $I \cup T^2 I \cup \cdots \cup T^{q-3} I$ is an $X$-avoiding set of measure $(q-1)/(2q)$, and Proposition~\ref{pr:comb} shows that this is largest possible, since the points $x, T x, T^2 x, \dots, T^{q-1} x$ induce a $q$-cycle. Finally suppose that $t$ is irrational. By Dirichlet's approximation theorem there exist infinitely many pairs of coprime integers $p$ and $q$ such that $| t - p/q | < 1/q^2$. For each such pair, let $\varepsilon = \varepsilon(q) = | t - p/q |$. Using an open interval $I$ of length $\frac{1}{q} - \varepsilon$ and applying the same construction as above with $T$ defined by $Tx = x+p/q$, one obtains an $X$-avoiding set of measure at least $((q-1)/2)(1/q - \varepsilon) = 1/2 - o(1)$. Alternatively, the lower bound $\alpha \ge 1/2$ follows from Rohlin's tower theorem (see e.g.\ \cite[Theorem~169]{kalikow+mccutcheon:oet}) applied to the ergodic transformation $Tx = x+t$. Therefore $\alpha = 1/2$. However this supremum can never be attained. Indeed, if $I \subset [0,1)$ is an $X$-avoiding set with $\lambda(I) = 1/2$ and $T$ is defined by $Tx = x+t$, then $I \cap T I = \emptyset$ and $T I \cap T^2 I = \emptyset$. Since $\lambda(I)=1/2$, this implies that $I$ and $T^2 I$ differ only on a nullset, contradicting the ergodicity of the irrational rotation $T^2$. \end{proof} \section{Gegenbauer polynomials and Schoenberg's theorem}\label{sec:gegenbauer} Before proving the first main result, we recall the Gegenbauer polynomials and Schoenberg's theorem from the theory of spherical harmonics. For $\nu > -1/2$, define the \emph{Gegenbauer weight function}\index{Gegenbauer weight function} \[ r_\nu(t) := (1-t^2)^{\nu-1/2},~~ -1 < t < 1. \] To motivate this definition, observe that if we take a uniformly distributed vector $\bm{\xi}\in S^{n-1}$, $n \geq 2$, and project it to any given axis, then the density of the obtained random variable $\bm{X}\in[-1,1]$ is proportional to $r_{(n-2)/2}$, with the coefficient $\left( \int_{-1}^1 r_{(n-2)/2}(x)\,\mathrm{d}x \right)^{-1} =\frac{\omega_{n-1}}{\omega_n}$ where $\omega_n$ is as in (\ref{eq:OmegaN}). (In particular, $\bm{X}$ is uniformly distributed in $[-1,1]$ if $n=3$.) Applying the Gram-Schmidt process to the polynomials $1,t,t^2,\dots$ with respect to the inner product $\langle f, g \rangle_\nu = \int_{-1}^1 f(t) g(t) r_\nu(t)\,\mathrm{d}t$, one obtains the \emph{Gegenbauer polynomials}\index{Gegenbauer polynomials} $C_i^\nu(t)$, $i=0,1,2,\dots$, where $C_i^\nu$ is of degree $i$. For a concise overview of these polynomials, see e.g.\ \cite[Section B.2]{dai+xu:athasb}. Here, we always use the normalisation $C_i^\nu(1)=1$. For a fixed $n \geq 2$, a continuous function $f : [-1,1] \to \mathbb{R}$ is called \emph{positive definite}\index{positive definite!function on $S^{n-1}$} if for every set of distinct points $\xi_1, \dots, \xi_s \in S^{n-1}$, the matrix $(f(\langle \xi_i, \xi_j \rangle))_{i,j=1}^s$ is positive semidefinite. We will need the following result of Schoenberg~\cite{shoenberg:38} (for a modern presentation, see e.g.\ \cite[Theorem~14.3.3]{dai+xu:athasb}). \begin{theorem}[Schoenberg's theorem]\label{thm:schoenberg}\index{Schoenberg's theorem} For $n \geq 2$, a continuous function $f : [-1,1] \to \mathbb{R}$ is positive definite if and only if there exist coefficients $a_i \geq 0$, for $i \geq 0$, such that \[ f(t) = \sum_{i=0}^\infty a_i C_i^{(n-2)/2}(t),~~~~\text{for all}~t \in [-1,1]. \] Moreover, the convergence on the right-hand side is absolute and uniform for every positive definite function $f$. \end{theorem} For a given positive definite function $f$, the coefficients $a_i$ in Theorem~\ref{thm:schoenberg} are unique and can be computed explicitly; a formula is given in \cite[Equation (14.3.3)]{dai+xu:athasb}. We are especially interested in the case $n=3$. Then $\nu = 1/2$, and the first few Gegenbauer polynomials $C_i^{1/2}(x)$ are \begin{gather*} C_0^{1/2}(x) = 1,~~~C_1^{1/2}(x) =x,~~~C_2^{1/2}(x) = \frac{1}{2} \left( 3x^2 - 1\right),\\ C_3^{1/2}(x) = \frac{1}{2}\left( 5x^3 - 3x \right), ~~~C_4^{1/2}(x) = \frac{1}{8}\left( 35x^4 -30x^2+3 \right). \end{gather*} \section{Linear programming relaxation}\label{sec:lp} Schoenberg's theorem allows us to set up a linear program whose value upper bounds $\alpha(n)$ for $n \geq 3$. The same result appears in \cite{bachoc09} and \cite{oliveira09}; we present a self-contained (and slightly simpler) proof for the reader's convenience. In the next section we strengthen the linear program, obtaining a better bound for~$\alpha(3)$. \begin{lemma}\label{lm:pt} Suppose $f,g \in L^2(S^{n-1})$ and define $k : [-1,1] \to \mathbb{R}$ by \begin{align}\label{eq:correlation} k(t) := \mathbb{E}[f(\bm{O} \xi) g(\bm{O} \eta)], \end{align} where the expectation is taken over randomly chosen $\bm{O} \in SO(n)$, and $\xi,\eta \in S^{n-1}$ are any two points satisfying $\langle \xi, \eta \rangle = t$. Then $k(t)$ exists for every $-1 \leq t \leq 1$, and $k$ is continuous. If $f=g$, then $k$ is positive definite. \end{lemma} \begin{proof} The expectation in \eqref{eq:correlation} clearly does not depend on the particular choice of $\xi,\eta \in S^{n-1}$. Fix any point $\xi_0 \in S^{n-1}$ and let $P : [-1,1] \to SO(n)$ be any continuous function satisfying $\langle \xi_0, P(t) \xi_0 \rangle = t$ for each $-1 \leq t \leq 1$. We have \begin{align}\label{eq:correlation1} k(t) = \mathbb{E}[f(\bm{O}\xi_0) g(\bm{O}P(t)\xi_0)]. \end{align} The functions $O \mapsto f(O\xi_0)$ and $O \mapsto g(OP(t)\xi_0)$ on $SO(n)$ belong to $L^2(SO(n))$; being an inner product in $L^2(SO(n))$, the expectation \eqref{eq:correlation1} therefore exists for every $t \in [-1,1]$. We next show that $k$ is continuous. For each $O \in SO(n)$, let $R_O : L^2(SO(n)) \to L^2(SO(n))$ be the \emph{right translation} operator defined by $(R_O F)(O') = F(O' O)$, for $F \in L^2(SO(n))$. For fixed $F$, the map $O \mapsto R_O F$ is continuous from $L^2(SO(n))$ to $L^2(SO(n))$ (see e.g. \cite[Lemma~1.4.2]{deitmar09}). Therefore the function $t \mapsto R_{P(t)} F$ is continuous from $[-1,1]$ to $L^2(SO(n))$. Using $F(O) = g(O\xi_0)$, the continuity of $k$ follows. Now suppose $f=g$; we show that $k$ is positive definite. Let $\xi_1, \dots, \xi_s \in S^{n-1}$. We need to show the $s \times s$ matrix $K = (k(\xi_i, \xi_j))_{i,j=1}^s$ is positive semidefinite. But if $v = (v_1, \dots, v_s)^T \in \mathbb{R}^s$ is any column vector, then \begin{align*} v^T K v &= \sum_{i=1}^s \sum_{j=1}^s \mathbb{E}[f(\bm{O} \xi_i) f(\bm{O} \xi_j)] v_i v_j = \mathbb{E} \left[ \left( \sum_{i=1}^s f(\bm{O} \xi_i) v_i \right)^2 \right] \geq 0. \end{align*} \end{proof} \begin{theorem}\label{thm:lp-upper-bound} $\alpha(n)$ is no more than the value of the following infinite-dimensional linear program. \begin{align}\label{eq:lpweak} \begin{gathered} \max x_0\\ \sum_{i=0}^\infty x_i = 1\\ \sum_{i=0}^\infty x_i C_i^{(n-2)/2}(0) = 0\\ x_i \geq 0,~\text{for all $i = 0,1,2,\dots$}~. \end{gathered} \end{align} \end{theorem} \begin{proof} Let $I \in \mathcal{L}$ be a $\{0\}$-avoiding subset of $S^{n-1}$ with $\lambda(I)>0$. We construct a feasible solution to the linear program \eqref{eq:lpweak} having value $\lambda(I)$. Let $k : [-1,1] \to \mathbb{R}$ be defined as in \eqref{eq:correlation}, with $f=g=\mathbbm{1}_I$. Then $k$ is a positive definite function satisfying $k(1) = \lambda(I)$ and $k(0) = 0$. By Theorem~\ref{thm:schoenberg}, $k$ has an expansion in terms of the Gegenbauer polynomials: \begin{align}\label{eq:gegenbauer} k(t) = \sum_{i=0}^\infty a_i C_i^{(n-2)/2}(t), \end{align} where each $a_i\ge 0$ and the convergence of the right-hand side is uniform on $[-1,1]$. Moreover, for each fixed $\xi_0 \in S^{n-1}$, we have by Fubini's theorem and \eqref{eq:correlation} that \begin{align}\label{eq:gegenbauerInt} \int_{S^{n-1}} k(\langle \xi_0, \eta \rangle) \,\mathrm{d}\eta &= \int_{S^{n-1}} \int_{S^{n-1}} k(\langle \xi, \eta \rangle) \, \mathrm{d} \xi \,\mathrm{d}\eta\\ &= \mathbb{E} \left[ \left( \int_{S^{n-1}} \mathbbm{1}_I(\bm{O} \xi) \, \mathrm{d}\xi \right)^2 \right] = \lambda(I)^2. \end{align} Note that \begin{align*} \int_{S^{n-1}} C_i^{(n-2)/2}(\langle \xi_0, \eta \rangle) \, \mathrm{d}\eta = \frac{\omega_{n-1}}{\omega_n} \int_{-1}^1 C_i^{(n-2)/2}(t) (1-t^2)^{(n-3)/2}\,\mathrm{d}t = 0 \end{align*} whenever $i \geq 1$ by the definition of the Gegenbauer polynomials. Putting \eqref{eq:gegenbauer} and \eqref{eq:gegenbauerInt} together and using that $C_0^{(n-2)/2} \equiv 1$, we conclude that $a_0 = \lambda(I)^2$. Recalling that $C_i^{(n-2)/2}(1) = 1$ for $i \geq 0$, we find that setting $x_i = a_i / \lambda(I)$ for $i=0,1,2,\dots$ gives a feasible solution of value $\lambda(I)$ to the linear program \eqref{eq:lpweak}.\qedhere \end{proof} Unfortunately in the case $n=3$, the value of \eqref{eq:lpweak} is at least $1/3$, which is the same bound obtained when Witsenhausen first stated the problem in \cite{witsenhausen74}. This can be seen from the feasible solution $x_0 = 1/3, x_2=2/3$ and $x_i = 0$ for all $i \neq 0,2$. \section{Adding combinatorial constraints}\label{sec:lp+comb} For each $\xi \in S^{n-1}$ and $-1 < t < 1$, let $\sigma_{\xi,t}$ be the unique probability measure on the Borel subsets of $S^{n-1}$ whose support is equal to the set $$ \xi^t := \{ \eta \in S^{n-1} : \langle \eta, \xi \rangle = t \}, $$ and which is invariant under all rotations fixing $\xi$. Now let $n=3$. As before, let $I \in \mathcal{L}$ be a $\{0\}$-avoiding subset of $S^2$ and define $k : [-1,1] \to \mathbb{R}$ as in \eqref{eq:correlation} with $f=g=\mathbbm{1}_I$; i.e. \[ k(t) = \mathbb{E}[\mathbbm{1}_I(\bm{O}\xi) \mathbbm{1}_I(\bm{O}\eta)], \] where $\xi, \eta \in S^2$ satisfy $\langle \xi, \eta \rangle = t$. Our aim now is to strengthen \eqref{eq:lpweak} for the case $n=3$ by adding combinatorial inequalities coming from Proposition~\ref{pr:comb} applied to the sections of $S^2$ by affine planes. We proceed as follows. Let $p$ and $q$ be coprime integers with $1/4 \leq p/q \leq 1/2$, and let $$ t_{p,q} = \sqrt{\frac{-\cos(2\pi p/q)}{1-\cos(2\pi p/q)}}.$$ Let $\xi \in S^{n-1}$ be arbitrary. If we take two orthogonal unit vectors with endpoints in $\xi^{t_{p,q}}$ and the centre $\xi_0=t_{p,q}\xi$ of this circle, then we get an isosceles triangle with side lengths $(1-t_{p,q}^2)^{1/2}$ and base $\sqrt{2}$; by the Cosine Theorem, the angle at $\xi_0$ is $2\pi p/q$. Let $\xi_0, \eta_0 \in S^{n-1}$ be arbitrary points satisfying $\langle \xi_0, \eta_0 \rangle = t_{p,q}$. By Fubini's theorem we have \begin{align*} k(t_{p,q}) &= \mathbb{E}[\mathbbm{1}_I(\bm{O}\xi_0) \mathbbm{1}_I(\bm{O}\eta_0)] = \int_{\xi_0^{t_{p,q}}} \mathbb{E}[\mathbbm{1}_I(\bm{O}\xi_0) \mathbbm{1}_I(\bm{O}\eta)] \, \mathrm{d}\sigma_{\xi_0,t_{p,q}}(\eta)\\ &= \mathbb{E} \left[ \mathbbm{1}_I(\bm{O}\xi_0) \int_{\xi_0^{t_{p,q}}} \mathbbm{1}_I(\bm{O}\eta) \, \mathrm{d}\sigma_{\xi_0,t_{p,q}}(\eta) \right]. \end{align*} But if $q$ is odd, then $\int_{\xi_0^{t_{p,q}}} \mathbbm{1}_I(O \eta) \, \mathrm{d} \sigma_{\xi_0,t_{p,q}}(\eta) \leq \frac{q-1}{2q}$ for all $O \in SO(n)$ by Proposition~\ref{pr:comb} applied to the circle $(O\xi_0)^{t_{p,q}}\cong S^1$, since the subgraph it induces contains a cycle of length $q$. Therefore $k(t_{p,q}) \leq \lambda(I) \frac{q-1}{2q}$. It follows that the inequalities \begin{align}\label{eq:combIneq} \sum_{i=0}^\infty x_i C_i^{1/2}(t_{p,q}) \leq (q-1)/2q, \end{align} are valid for the relaxation and can be added to \eqref{eq:lpweak}. The same holds for the inequalities $\sum_{i=0}^\infty x_i C_i^{1/2}(-t_{p,q}) \leq (q-1)/2q$. So we have just proved the following result. \begin{theorem} $\alpha(3)$ is no more than the value of the following infinite-dimensional linear program. \begin{align}\label{eq:lp-super-strong} \begin{gathered} \max x_0\\ \sum_{i=0}^\infty x_i = 1\\ \sum_{i=0}^\infty x_i C_i^{1/2}(0) = 0\\ \sum_{i=0}^\infty x_i C_i^{1/2}(\pm t_{p,q}) \leq (q-1)/2q,~~\text{for $q$ odd,~~$p,q$ coprime}\\ x_i \geq 0,~\text{for all $i = 0,1,2,\dots$}\ . \end{gathered} \end{align} \end{theorem} Rather than attempting to find the exact value of the linear program \eqref{eq:lp-super-strong}, the idea will be to discard all but finitely many of the combinatorial constraints, and then to apply the weak duality theorem of linear programming. The dual linear program has only finitely many variables, and any feasible solution gives an upper bound for the value of program \eqref{eq:lp-super-strong}, and therefore also for $\alpha(3)$. At the heart of the proof is the verification of the feasibility of a particular dual solution which we give explicitly. While part of the verification has been carried out by computer in order to deal with the large numbers that appear, it can be done using only rational arithmetic and can therefore be considered rigorous. \begin{theorem} $\alpha(3) < 0.313$. \end{theorem} \begin{proof} Consider the following linear program \begin{align}\label{eq:lpstrong} \max \Big\{ x_0 &: \sum_{i=0}^\infty x_i = 1, \sum_{i=0}^\infty x_i C_i^{1/2}(0) = 0, \sum_{i=0}^\infty x_i C_i^{1/2}(t_{1,3}) \leq 1/3,\\ &\sum_{i=0}^\infty x_i C_i^{1/2}(t_{2,5}) \leq 2/5, \sum_{i=0}^\infty x_i C_i^{1/2}(-t_{2,5}) \leq 2/5,\nonumber\\ &~~~~~~~~~~~x_i \geq 0,~\text{for all $i = 0,1,2,\dots$}\nonumber \Big\}. \end{align} The linear programming dual of \eqref{eq:lpstrong} is the following. \begin{align}\label{eq:lpdual} \begin{gathered} \min ~b_1 + \frac{1}{3}b_{1,3} + \frac{2}{5}b_{2,5} + \frac{2}{5}b_{2,5-}\\ b_1 + b_0 + b_{1,3} + b_{2,5} + b_{2,5-} \geq 1\\ b_1 + C_i^{1/2}(0) b_0 + C_i^{1/2}(t_{1,3}) b_{1,3} + C_i^{1/2}(t_{2,5}) b_{2,5} + C_i^{1/2}(-t_{2,5}) b_{2,5-} \geq 0 ~\text{for $i = 1,2,\dots$}\\ b_1, b_0 \in \mathbb{R}, ~b_{1,3}, b_{2,5}, b_{2,5-} \geq 0. \end{gathered} \end{align} By linear programming duality, any feasible solution for program \eqref{eq:lpdual} gives an upper bound for \eqref{eq:lpstrong}, and therefore also for $\alpha(3)$. So in order to prove the claim $\alpha(3) < 0.313$, it suffices to give a feasible solution to \eqref{eq:lpdual} having objective value no more than $0.313$. Let \begin{align*} b = (b_1, b_0, b_{1,3}, b_{2,5}, b_{2,5-}) = \frac{1}{10^6}(128614, 404413, 36149, 103647, 327177). \end{align*} It is easily verified that $b$ satisfies the first constraint of \eqref{eq:lpdual} and that its objective value less than $0.313$. To verify the infinite family of constraints \begin{align}\label{eq:constraints} b_1 + C_i^{1/2}(0) b_0 + C_i^{1/2}(t_{1,3}) b_{1,3} + C_i^{1/2}(t_{2,5}) b_{2,5} + C_i^{1/2}(-t_{2,5}) b_{2,5-} \geq 0 \end{align} for $i=1,2,\dots$, we apply Theorem~8.21.11 from \cite{szego92} (where $C_i^{\lambda}$ is denoted as $P_i^{(\lambda)}$), which implies \begin{align}\label{eq:gegenbauer-upper-bound} | C_i^{1/2}(\cos{\theta}) | \leq \frac{\sqrt{2}}{\sqrt{\pi} \sqrt{\sin{\theta}}}\, \frac{\Gamma(i+1)}{\Gamma(i+3/2)} + \frac{1}{\sqrt{\pi} 2^{3/2} (\sin{\theta})^{3/2}}\, \frac{\Gamma(i+1)}{\Gamma(i+5/2)} \end{align} for each $0 < \theta < \pi$. Note that $t_{1,3} = 1/\sqrt{3}$ and $t_{2,5}=5^{-1/4}$. When $\theta \in A:= \{ \pi/2, \arccos{t_{1,3}}, \arccos{t_{2,5}}, \arccos{(-t_{2,5})} \}$, we have $\sin{\theta} \in \{ 1, \sqrt{\frac{2}{3}}, \gamma \}$, where $\gamma =\frac{2}{\sqrt{5+\sqrt{5}}}$. The right-hand side of equation \eqref{eq:gegenbauer-upper-bound} is maximized over $\theta\in A$ at $\sin{\theta} = \gamma$ for each fixed $i$, and since the right-hand side is decreasing in $i$, one can verify using rational arithmetic only that it is no greater than $128614 / 871386 = b_1 / (b_0+b_{1,3}+b_{2,5}+b_{2,5-})$ when $i \geq 40$, by evaluating at $i=40$. Therefore, \begin{align*} &b_1 + C_i^{1/2}(0) b_0 + C_i^{1/2}(t_{1,3}) b_{1,3} + C_i^{1/2}(t_{2,5}) b_{2,5} + C_i^{1/2}(-t_{2,5}) b_{2,5-}\\ \geq&~ b_1 - (b_0+b_{1,3}+b_{2,5}+b_{2,5-})\max_{\theta \in A}\{ | C_i^{1/2}(\cos{\theta}) | \}\\ \geq&~0 \end{align*} when $i \geq 40$. It now suffices to check that $b$ satisfies the constraints \eqref{eq:constraints} for $i=0,1,\dots,39$. This can also be accomplished using rational arithmetic only. \end{proof} The rational arithmetic calculations required in the above proof were carried out with \emph{Mathematica}. When verifying the upper bound for the right-hand side of \eqref{eq:gegenbauer-upper-bound}, it is helpful to recall the identity $\Gamma(i+1/2) = (i-1/2)(i-3/2) \cdots (1/2) \sqrt{\pi}$. When verifying the constraints \eqref{eq:constraints} for $i=0,1,\dots,39$, it can be helpful to observe that $t_{1,3}$ and $\pm t_{2,5}$ are roots of the polynomials $x^2 - 1/3$ and $x^4-1/5$ respectively; this can be used to cut down the degree of the polynomials $C_i^{1/2}(x)$ to at most $3$ before evaluating them. The ancillary folder of the \texttt{arxiv.org} version of this paper contains a \emph{Mathematica} notebook that verifies all calculations. The combinatorial inequalities of the form \eqref{eq:combIneq} we chose to include in the strengthened linear program \eqref{eq:lpstrong} were found as follows: Let $L_0$ denote the linear program \eqref{eq:lpweak}. We first find an optimal solution $\sigma_0$ to $L_0$. We then proceed recursively; having defined the linear program $L_{i-1}$ and found an optimal solution $\sigma_{i-1}$, we search through the inequalities \eqref{eq:combIneq} until one is found which is violated by $\sigma_{i-1}$, and we strengthen $L_{i-1}$ with that inequality to produce $L_i$. At each stage, an optimal solution to $L_i$ is found by first solving the dual minimisation problem, and then applying the complementary slackness theorem from linear programming to reduce $L_i$ to a linear programming maximisation problem with just a finite number of variables. Adding more inequalities of the form $\eqref{eq:combIneq}$ appears to give no improvement on the upper bound. Also adding the constraints $\sum_{i=0}^\infty x_i C_i^{1/2}(t) \geq 0$ for $-1\leq t \leq 1$ appears to give no improvement. A small (basically insignificant) improvement can be achieved by allowing the odd cycles to embed into $S^2$ in more general ways, for instance with the points lying on two different latitudes rather than just one. \section{Adjacency operator}\label{sec:adj} Let $n \geq 3$. For $\xi \in S^{n-1}$ and $-1 < t < 1$, we use the notations $\xi^t$ and $\sigma_{\xi,t}$ from Section \ref{sec:lp+comb}. For $f \in L^2(S^{n-1})$ define $A_tf: S^{n-1}\to\I R$ by \begin{align}\label{eq:transOperator} (A_t f)(\xi) := \int_{\xi^t} f(\eta)\,\mathrm{d}\sigma_{\xi,t}(\eta),\quad \xi\in S^{n-1}. \end{align} Here we establish some basic properties of $A_t$ which will be helpful later. The operator $A_t$ can be thought of as an adjacency operator for the graph with vertex set $S^{n-1}$, in which we join two points with an edge when their inner product is $t$. Adjacency operators for infinite graphs are explored in greater detail and generality in \cite{bachoc13}. \begin{lemma}\label{lm:adjacency} For every $t \in (-1,1)$, $A_t$ is a bounded linear operator from $L^2(S^{n-1})$ to $L^2(S^{n-1})$ having operator norm equal to $1$. \end{lemma} \begin{proof} The right-hand side of \eqref{eq:transOperator} involves integration over nullsets of a function $f \in L^2(S^{n-1})$ which is only defined almost everywhere, and so strictly speaking one should argue that \eqref{eq:transOperator} really makes sense. In other words, given a particular representative~$f$ from its $L^2$-equivalence class, we need to check that the integral on the right-hand side of \eqref{eq:transOperator} is defined for almost all $\xi \in S^{n-1}$, and that the $L^2$-equivalence class of $A_t f$ does not depend on the particular choice of representative~$f$. Our main tool will be Minkowski's integral inequality (see e.g.\ \cite[Theorem~6.19]{folland99}). Let $e_n = (0,\dots,0,1)$ be the $n$-th basis vector in $\mathbb{R}^n$ and let \[ S = \{ (x_1, x_2, \dots, x_n) : x_n=0, x_1^2 + \dots + x_{n-1}^2 =1 \} \] be a copy of $S^{n-2}$ inside $\mathbb{R}^n$. Considering $f$ as a particular measurable function (not an $L^2$-equivalence class), we define $F : SO(n) \times S \to \mathbb{R}$ by \[ F(\rho,\eta) =f\left (\rho\left(t e_n + \sqrt{1-t^2}\, \eta\right)\right),\qquad \rho \in SO(n),\ \eta\in S. \] Let us formally check all the hypotheses of Minkowski's integral inequality applied to $F$, where $SO(n)$ is equipped with the Haar measure, and where $S$ is equipped with the normalised Lebesgue measure; this will show that the function $\tilde{F} : SO(n) \to \mathbb{R}$ defined by $\tilde{F}(\rho) = \int_S F(\rho, \eta)\,\mathrm{d}\eta$ belongs to $L^2(SO(n))$. Clearly the function $F$ is measurable. To see that the function $\rho \mapsto F(\rho, \eta)$ belongs to $L^2(SO(n))$ for each fixed $\eta \in S$, simply note that \[ \int_{SO(n)} \left| F(\rho,\eta) \right|^2 \,\mathrm{d}\rho = \int_{SO(n)} \left| f(\rho(te_n + \sqrt{1-t^2}\,\eta)) \right|^2 \,\mathrm{d}\rho = \|f\|_2^2. \] That the function $\eta \mapsto \| F(\cdot, \eta) \|_2$ belongs to $L^1(S)$ then also follows easily (in fact, this function is constant): \[ \int_S \left( \int_{SO(n)} \left| F(\rho,\eta) \right|^2 \,\mathrm{d}\rho \right)^{1/2} \,\mathrm{d}\eta = \int_S \|f\|_2 \,\mathrm{d}\eta = \|f\|_2. \] Minkowski's integral inequality now gives that the function $\eta \mapsto F(\rho, \eta)$ is in $L^1(S)$ for a.e.\ $\rho$, the function $\tilde{F}$ is in $L^2(SO(n))$, and its norm can be bounded as follows: \begin{align} \|\tilde{F}\|_2 &= \left( \int_{SO(n)} \left| \int_S F(\rho,\eta) \,\mathrm{d}\eta \right|^2 \,\mathrm{d}\rho\right)^{1/2}\nonumber\\ &\leq \int_{S} \left( \int_{SO(n)} |F(\rho,\eta)|^2 \,\mathrm{d}\rho \right)^{1/2} \,\mathrm{d}\eta = \|f\|_2.\label{eq:FNormBound} \end{align} Applying \eqref{eq:FNormBound} to $f-g$ where $g$ is a.e.\ equal to $f$, we conclude that the $L^2$-equivalence class of $\tilde{F}$ does not depend on the particular choice of representative $f$ from its equivalence class. Now $(A_t f)(\xi)$ is simply $\tilde{F}(\rho)$, where $\rho \in SO(n)$ can be any rotation such that $\rho e_n = \xi$. This shows that the integral in \eqref{eq:transOperator} makes sense for almost all $\xi \in S^{n-1}$. We have $\|A_t\| \leq 1$ since for any $f \in L^2(S^{n-1})$, \begin{align*} \| A_t f \|_2 = \left( \int_{S^{n-1}} \left| (A_t f)(\xi) \right|^2 \,\mathrm{d}\xi \right)^{1/2} &= \left( \int_{SO(n)} \left| (A_t f)(\rho e_n) \right|^2 \,\mathrm{d}\rho \right)^{1/2}\\ &= \left( \int_{SO(n)} \left| \tilde{F}(\rho) \right|^2 \,\mathrm{d}\rho \right)^{1/2} \leq \| f \|_2, \end{align*} by \eqref{eq:FNormBound}. Finallly, applying $A_t$ to the constant function $1$ shows that $\|A_t\| = 1$.\qedhere \end{proof} \begin{lemma}\label{lm:twoPoint} Let $f$ and $g$ be functions in $L^2(S^{n-1})$, let $\xi, \eta \in S^{n-1}$ be arbitrary points, and write $t = \langle \xi, \eta \rangle$. If $\bm{O} \in SO(n)$ is chosen uniformly at random with respect to the Haar measure on $SO(n)$, then \begin{align}\label{eq:k-t} \int_{S^{n-1}} f(\zeta) (A_t g)(\zeta) \,\mathrm{d}\zeta = \mathbb{E}[ f(\bm{O}\xi) g(\bm{O}\eta) ], \end{align} which is exactly the definition of $k(t)$ from \eqref{eq:correlation}. \end{lemma} \begin{proof} We have \begin{align*} \int_{S^{n-1}} f(\zeta) (A_t g)(\zeta)\,\mathrm{d}\zeta &= \int_{SO(n)} f(O\xi) (A_t g)(O\xi)\,\mathrm{d}O\\ &= \int_{SO(n)} f(O\xi) \int_{(O \xi)^{t}} g(\psi) \,\mathrm{d}\sigma_{O\xi, t}(\psi)\,\mathrm{d}O, \end{align*} If $H$ is the subgroup of all elements in $SO(n)$ which fix $\xi$, then the above integral can be rewritten \begin{align*} \int_{SO(n)} f(O\xi) \int_H g(Oh\eta) \,\mathrm{d}h \,\mathrm{d}O. \end{align*} By Fubini's theorem, this integral is equal to \begin{align*} &\int_H \int_{SO(n)} f(O\xi) g(Oh\eta) \,\mathrm{d}O \,\mathrm{d}h\\ =& \int_H \int_{SO(n)} f(O h^{-1} \xi) g(O\eta) \,\mathrm{d}O \,\mathrm{d}h\\ =& \int_{SO(n)} f(O \xi) g(O\eta) \,\mathrm{d}O, \end{align*} where we use the right-translation invariance of the Haar integral on $SO(n)$ at the first equality, and the second inequality follows by noting that the integrand is constant with respect to $h$. \end{proof} \begin{lemma}\label{lm:SelfAdj} For every $t\in (-1,1)$, the operator $A_t:L^2(S^{n-1})\to L^2(S^{n-1})$ is self-adjoint. \end{lemma} \begin{proof} Fix $\xi,\eta \in S^{n-1}$ that satisfy $\langle \xi, \eta \rangle = t$. Lemma~\ref{lm:twoPoint} implies that for any $f,g\in L^2(S^{n-1})$, we have \[ \langle A_t f, g \rangle = \mathbb{E}_{\bm{O} \in SO(n)}[ f(\bm{O}\xi) g(\bm{O} \eta) ] = \langle f, A_t g \rangle, \] giving the required.\end{proof} \section{Existence of a measurable maximum independent set}\label{sec:max} Let $n \geq 2$ and $X \subset [-1,1]$. From Theorem~\ref{thm:irCircle} we know that the supremum $\alpha_X(n)$ is sometimes attained as a maximum, and sometimes not. It is therefore interesting to ask when a maximizer exists. The main positive result in this direction is Theorem~\ref{thm:attainment}, which says that a largest measurable $X$-avoiding set always exists when $n \geq 3$. Remarkably, this result holds under no additional restrictions (not even Lebesgue measurability) on the set $X$ of forbidden inner products. Before arriving at this theorem, we shall need to establish a number of technical results. For the remainder of this section we suppose $n \geq 3$. For $d \geq 0$, let $\mathrsfs{H}_d^n$ be the vector space of homogeneous polynomials $p(x_1,\dots,x_n)$ of degree $d$ in $n$ variables belonging to the kernel of the Laplace operator; that is \[ \frac{\partial^2 p}{\partial x_1^2} + \cdots + \frac{\partial^2 p}{\partial x_n^2} = 0. \] Note that each $\mathrsfs{H}_d^n$ is finite-dimensional. The restrictions of the elements of $\mathrsfs{H}_d^n$ to the surface of the unit sphere are called the \emph{spherical harmonics}. For fixed $n$, we have $L^2(S^{n-1}) = \oplus_{d=0}^\infty \mathrsfs{H}_d^n$ (\cite[Theorem~2.2.2]{dai+xu:athasb}); that is, each function in $L^2(S^{n-1})$ can be written uniquely as an infinite sum of elements from $\mathrsfs{H}_d^n$, $d=0,1,2,\dots$, with convergence in the $L^2$-norm. Recall the definition \eqref{eq:transOperator} of the adjacency operator from Section \ref{sec:adj}: \[ (A_t f)(\xi) := \int_{\xi^t} f(\eta) \,\mathrm{d}\sigma_{\xi,t}(\eta),\quad f\in L^2(S^{n-1}). \] The next lemma states that each spherical harmonic is an eigenfunction of the operator $A_t$. It extends the Funk-Hecke formula (\cite[Theorem~1.2.9]{dai+xu:athasb}) to the Dirac measures, obtaining the eigenvalues of $A_t$ explicitly. The proof relies on the fact that integral kernel operators $K$ having the form $(Kf)(\xi) = \int f(\zeta) k(\langle \zeta, \xi \rangle) \,\mathrm{d}\zeta$ for some function $k : [-1,1] \to \mathbb{R}$ are diagonalised by the spherical harmonics, and moreover that the eigenvalue of a specific spherical harmonic depends only on its degree. \begin{proposition}\label{pr:adjEigs} Let $t \in (-1,1)$. Then for every spherical harmonic $Y_d$ of degree~$d$, \[ (A_t Y_d)(\xi) = \int_{\xi^t} Y_d(\eta) \,\mathrm{d}\sigma_{\xi,t}(\eta) = \mu_d(t) Y_d(\xi), ~~\xi \in S^{n-1}, \] where $\mu_d(t)$ is the constant \[ \mu_d(t) = C_d^{(n-2)/2}(t) (1-t^2)^{(n-3)/2}. \] \hide{ \[ \mu_d(t) = C_d^{(n-2)/2}(t) (1-t^2)^{(n-3)/2} \bigg/ C_d^{(n-2)/2}(1). \] } \end{proposition} \begin{proof} Let $\,\mathrm{d}s$ be the Lebesgue measure on $[-1,1]$ and let $\{ f_\alpha \}_\alpha$ be a net of functions in $L^1([-1,1])$ such that $\{ f_\alpha \,\mathrm{d}s \}$ converges to the Dirac point mass $\delta_t$ at $t$ in the weak-* topology on the set of Borel measures on $[-1,1]$. By Theorem~1.2.9 in \cite{dai+xu:athasb}, we have \[ \int_{S^{n-1}} Y_d(\eta) f_\alpha(\langle \xi, \eta \rangle) \,\mathrm{d}\eta = \mu_{d,\alpha} Y_d(\xi), \] where \[ \mu_{d,\alpha} = \int_{-1}^1 C_d^{(n-2)/2}(s) (1-s^2)^{(n-3)/2} f_\alpha(s) \,\mathrm{d}s. \] \hide{ \[ \mu_{d,\alpha} = \int_{-1}^1 C_d^{(n-2)/2}(s) (1-s^2)^{(n-3)/2} f_\alpha(s) \,\mathrm{d}s \bigg/ C_d^{(n-2)/2}(1). \] } By taking limits, we finish the proof. \end{proof} The next lemma is a general fact about weakly convergent sequences in a Hilbert space. \begin{lemma}\label{lm:weakCompact} Let $\mathcal{H}$ be a Hilbert space and let $K : \mathcal{H} \to \mathcal{H}$ be a compact operator. Suppose $\{ x_i \}_{i=1}^\infty$ is a sequence in $\mathcal{H}$ converging weakly to $x \in \mathcal{H}$. Then \[ \lim_{i \to \infty} \langle K x_i, x_i \rangle = \langle K x, x \rangle. \] \end{lemma} \begin{proof} Let $C$ be the maximum of $\| x \|$ and $\sup_{i \geq 1} \| x_i \| $, which is finite by the Principle of Uniform Boundedness. Let $\{ K_m \}_{m=1}^\infty$ be a sequence of finite rank operators such that $K_m \to K$ in the operator norm as $m \to \infty$. Clearly \[ \lim_{i \to \infty} \langle K_m x_i, x_i \rangle = \langle K_m x,x \rangle \] for each $m=1,2,\dots$\ . Let $\varepsilon > 0$ be given and choose $m_0$ so that $\| K - K_{m_0} \| < \varepsilon/(3C^2)$. Choosing $i_0$ so that $| \langle K_{m_0} x_i, x_i \rangle - \langle K_{m_0} x, x \rangle | < \varepsilon/3$ whenever $i \geq i_0$, we have \begin{align*} & | \langle K x_i, x_i \rangle - \langle K x, x \rangle | \\ \leq& | \langle K x_i, x_i \rangle - \langle K_{m_0} x_i, x_i \rangle | + | \langle K_{m_0} x_i, x_i \rangle - \langle K_{m_0} x, x \rangle | + | \langle K_{m_0} x, x \rangle - \langle Kx, x \rangle | \\ \leq& \| K - K_{m_0} \| C^2 + \varepsilon/3 + \| K - K_{m_0} \| C^2 \ <\ \varepsilon, \end{align*} and the lemma follows. \end{proof} The next corollary is also a result stated in \cite{bachoc13}. \begin{corollary}\label{cor:compact} If $n \geq 3$ and $t \in (-1,1)$, then $A_t$ is compact. \end{corollary} \begin{proof} The operator $A_t$ is diagonalisable by Proposition~\ref{pr:adjEigs}, since the spherical harmonics form an orthonormal basis for $L^2(S^{n-1})$. It therefore suffices to show that its eigenvalues cluster only at $0$. By Theorem~8.21.8 of \cite{szego92} and Proposition~\ref{pr:adjEigs}, the eigenvalues $\mu_d(t)$ tend to zero as $d \to \infty$. The eigenspace corresponding to the eigenvalue $\mu_d(t)$ is precisely the vector space of spherical harmonics of degree $d$, which is finite dimensional. Therefore $A_t$ is compact. \end{proof} For each $\xi \in S^{n-1}$, let $C_h(\xi)$ be the open spherical cap of height $h$ in $S^{n-1}$ centred at $\xi$. Recall that $C_h(\xi)$ has volume proportional to $\int_{1-h}^1 (1-t^2)^{(n-3)/2} \,\mathrm{d}t$. \begin{lemma}\label{lm:caps} For each $\xi \in S^{n-1}$, we have $\lambda(C_h(\xi)) = \Theta(h^{(n-1)/2})$, and\\ $\lambda(C_{h/2}(\xi)) \geq \lambda(C_{h}(\xi))/2^{(n-1)/2} - o(h^{(n-1)/2})$ as $h \to 0^+$. \end{lemma} \begin{proof} If $f(h) = \int_{1-h}^1 (1-t^2)^{(n-3)/2} \,\mathrm{d}t$, then we have $\frac{df}{dh}(h) = (2h-h^2)^{(n-3)/2}$. Since $f(0) = 0$, the smallest power of $h$ occurring in $f(h)$ is of order $(n-1)/2$. This gives the first result. For the second, note that the coefficient of the lowest order term in $f(h)$ is $2^{(n-1)/2}$ times that of $f(h/2)$. \end{proof} \begin{lemma}\label{lm:density} Suppose $n \geq 3$ and let $I \subset S^{n-1}$ be a Lebesgue measurable set with $\lambda(I) > 0$. Define $k:[-1,1] \to \mathbb{R}$ by \[ k(t) := \int_{S^{n-1}} \mathbbm{1}_I(\zeta) (A_t \mathbbm{1}_I)(\zeta) \,\mathrm{d}\zeta, \] which by Lemma~\ref{lm:twoPoint} is the same as Definition \eqref{eq:correlation} applied with $f=g=\mathbbm{1}_I$. If $\xi_1, \xi_2 \in S^{n-1}$ are Lebesgue density points of $I$, then $k(\langle \xi_1, \xi_2 \rangle) > 0$. \end{lemma} \begin{proof} Let $t = \langle \xi_1, \xi_2 \rangle$. If $t = 1$, then the conclusion holds since $k(1)=\lambda(I)>0$. If $t=-1$, then $\xi_2 = -\xi_1$, and by the Lebesgue density theorem we can choose $h>0$ small enough that $\lambda(C_h(\xi_i) \cap I) > \frac{2}{3} \lambda(C_h(\xi_i))$ for $i=1,2$. By Lemma~\ref{lm:twoPoint} we have \begin{align*} k(-1) &= \mathbb{E}[\mathbbm{1}_I(\bm{O}\xi_1) \mathbbm{1}_I(\bm{O}(-\xi_1))]\\ &\geq \mathbb{E}[\mathbbm{1}_{I \cap C_h(\xi_2) }(\bm{O}\xi_1) \mathbbm{1}_{I \cap C_h(\xi_2) }(\bm{O}(-\xi_1))] \geq \frac{1}{3}\lambda(C_h(\xi_1)). \end{align*} From now on we may therefore assume $-1 < t < 1$. Let $h>0$ be a small number which will be determined later. Suppose $x \in C_h(\xi_1)$. The intersection $x^t \cap C_h(\xi_2)$ is a spherical cap in the $(n-2)$-dimensional sphere $x^t$ having height proportional to $h$; this is because $C_h(\xi_2)$ is the intersection of $S^{n-1}$ with a certain halfspace $H$, and $x^t \cap C_h(\xi_2) = x^t \cap H$. We have $\sigma_{x,t}(x^t \cap C_h(\xi_2)) = \Theta(h^{(n-2)/2})$ by Lemma~\ref{lm:caps}, and it follows that there exists $D>0$ such that $\sigma_{x,t}(x^t \cap C_h(\xi_2)) \leq Dh^{(n-2)/2}$ for sufficiently small $h>0$. If $x \in C_{h/2}(\xi_1)$, then $x^t \cap C_{h/2}(\xi_2) \neq \emptyset$ since $x^t$ is just a rotation of the hyperplane $\xi_1^t$ through an angle equal to the angle between $x$ and $\xi_1$. Therefore $x^t \cap C_h(\xi_2)$ is a spherical cap in $x^t$ having height at least $h/2$. Thus there exists $D' > 0$ such that $\sigma_{x,t}(x^t \cap C_h(\xi_2)) \geq D' h^{(n-2)/2}$ for all $x \in C_{h/2}(\xi_1)$, by Lemma~\ref{lm:caps}. Now choose $h>0$ small enough that $\lambda(C_h(\xi_i) \cap I) \geq (1 - \frac{D'}{2^n D}) \lambda(C_h(\xi_i))$ for $i=1,2$; this is possible by the Lebesgue density theorem since $\xi_1$ and $\xi_2$ are density points. We have by Lemma~\ref{lm:twoPoint} that \[ k(t) = \mathbb{P}[\bm{\eta}_1 \in I, \bm{\eta}_2 \in I], \] if $\bm{\eta}_1$ is chosen uniformly at random from $S^{n-1}$, and if $\bm{\eta}_2$ is chosen uniformly at random from $\bm{\eta}_1^t$. Then \begin{align*} k(t) &\geq \mathbb{P}[\bm{\eta}_1 \in I \cap C_h(\xi_1), \bm{\eta}_2 \in I \cap C_h(\xi_2)]\\ &\geq \mathbb{P}[\bm{\eta}_1 \in C_h(\xi_1), \bm{\eta}_2 \in C_h(\xi_2)] -\mathbb{P}[\bm{\eta}_1 \in C_h(\xi_1) \setminus I, \bm{\eta}_2 \in C_h(\xi_2) ]\\ &~~~~~~~~~~~~~~-\mathbb{P}[\bm{\eta}_1 \in C_h(\xi_1), \bm{\eta}_2 \in C_h(\xi_2) \setminus I]. \end{align*} The first probability is at least \begin{align*} D' h^{(n-2)/2} \lambda(C_{h/2}(\xi_1)) \geq \frac{D'}{2^{(n-1)/2}} h^{(n-2)/2} \lambda(C_h(\xi_1)) - o(h^{(2n-3)/2}) \end{align*} by Lemma~\ref{lm:caps}. The second and third probabilities are each no more than $$ \frac{D'}{2^n D} \lambda(C_h(\xi_1)) D h^{(n-2)/2} = \frac{D'}{2^n} \lambda(C_h(\xi_1)) h^{(n-2)/2} $$ for sufficiently small $h>0$, and therefore by the first part of Lemma~\ref{lm:caps}, \[ k(t) \geq \frac{D'}{2^{(n-1)/2}} \lambda(C_h(\xi_1)) h^{(n-2)/2} - o(h^{(2n-3)/2}) - \frac{D'}{2^{n-1}} \lambda(C_h(\xi_1)) h^{(n-2)/2}, \] and this is strictly positive for sufficiently small $h>0$. \end{proof} We are now in a position to prove the second main result of this paper. \begin{theorem}\label{thm:attainment} Suppose $n \geq 3$ and let $X$ be any subset of $[-1,1]$. Then there exists an $X$-avoiding set $I \in \mathcal{L}$ such that $\lambda(I) = \alpha_X(n)$. \end{theorem} \begin{proof} We may suppose that $1 \not\in X$ for otherwise every $X$-avoiding set is empty and the theorem holds with $I=\emptyset$. Let $\{ I_i \}_{i=1}^\infty$ be a sequence of measurable $X$-avoiding sets such that $\lim_{i \to \infty} \lambda(I_i) = \alpha_X(n)$. Passing to a subsequence if necessary, we may suppose that the sequence $\{ \mathbbm{1}_{I_i} \}$ of characteristic functions converges weakly in $L^2(S^{n-1})$; let $h$ be its limit. Then $0 \leq h \leq 1$ almost everywhere since $0 \leq \mathbbm{1}_{I_i} \leq 1$ for every $i$. Denote by $I'$ the set $h^{-1}((0,1])$, and let $I$ be the set of Lebesgue density points of $I'$. We claim that $I$ is $X$-avoiding. For all $t \in X \setminus \{ -1 \}$, the operator $A_t : L^2(S^{n-1}) \to L^2(S^{n-1})$ is self-adjoint and compact by Lemma~\ref{lm:SelfAdj} and Corollary~\ref{cor:compact}. Since $\langle A_t \mathbbm{1}_{I_i}, \mathbbm{1}_{I_i} \rangle = 0$ for each $i$, Lemma~\ref{lm:weakCompact} implies $\langle A_t h, h \rangle = 0$. Since $h \geq 0$, it follows from the definition of $A_t$ that $\langle A_t \mathbbm{1}_{I'}, \mathbbm{1}_{I'} \rangle = 0$, and therefore also that $\langle A_t \mathbbm{1}_{I}, \mathbbm{1}_{I} \rangle = 0$. But if there exist points $\xi, \eta \in I$ with $t_0 = \langle \xi, \eta \rangle \in X \setminus \{ -1 \}$, then $\langle A_{t_0} \mathbbm{1}_{I}, \mathbbm{1}_{I} \rangle > 0$ by Lemma~\ref{lm:density}. Thus, in order to show that $I$ is $X$-avoiding, it remains to derive a contradiction from assuming that $-1\in X$ and $-\xi,\xi\in I$ for some $\xi\in S^{n-1}$. Since $\xi$ and $-\xi$ are Lebesgue density points of $I$, there is a spherical cap $C$ centred at $\xi$ such that $\lambda(I \cap C) > \frac{2}{3} \lambda(C)$ and $\lambda(I \cap (-C)) > \frac{2}{3} \lambda(C)$. The same applies to $I_i$ for all large $i$ (since a cap is a continuity set). But this contradicts the fact that $I_i$ and its reflection $-I_i$ are disjoint for every $i$. Thus $I$ is $X$-avoiding. Finally, we have \begin{align*} \lambda(I) &= \lambda(I') \geq \langle \mathbbm{1}_{S^{n-1}}, h \rangle = \lim_{i \to \infty} \langle \mathbbm{1}_{S^{n-1}}, \mathbbm{1}_{I_i} \rangle = \lim_{i \to \infty} \lambda(I_i) = \alpha_X(n), \end{align*} whence $\lambda(I) = \alpha_X(n)$ since $\lambda(I) \leq \alpha_X(n)$. \end{proof} \section{Invariance of $\alpha_X(n)$ under taking the closure of $X$}\label{sec:closure} Again let $n \geq 2$ and $X \subset [-1,1]$. We will use $\overline{X}$ to denote the toplogical closure of $X$ in $[-1,1]$. In general it is false that $X$-avoiding sets are $\overline{X}$-avoiding. In spite of this, we have the following result. \begin{theorem}\label{thm:closureIR} Let $X$ be an arbitrary subset of $[-1,1]$. Then $\alpha_X(n) = \alpha_{\overline{X}}(n)$. In particular $\alpha_X(n) = 0$ if $1 \in \overline{X}$. \end{theorem} \begin{proof} Clearly $\alpha_X(n) \geq \alpha_{\overline{X}}(n)$ For the reverse inequality, let $I' \subset S^{n-1}$ be any measurable $X$-avoiding set. Let $I \subset I'$ be the set of Lebesgue density points of $I'$, and define $k:[-1,1] \to \mathbb{R}$ by $k(t) = \int_{S^{n-1}} \mathbbm{1}_I(\zeta) (A_t \mathbbm{1}_I)(\zeta) \,\mathrm{d}\zeta$. Then $k$ is continuous by Lemma~\ref{lm:pt} and Lemma~\ref{lm:twoPoint}, and since $k(t) = 0$ for every $t \in X$, it follows that $k(t) = 0$ for every $t \in \overline{X}$. Lemma~\ref{lm:density} now implies that $I$ is $\overline{X}$-avoiding. The theorem now follows since $I'$ was arbitrary, and $\lambda(I) = \lambda(I')$ by the Lebesgue density theorem. \end{proof} \section{Single forbidden inner product}\label{sec:single} An interesting case to consider is when $|X|=1$, motivated by the fact that $1/\alpha_{\{t\}}(n)$ is a lower bound on the measurable chromatic number of $\I R^n$ for any $t\in(-1,1)$ and this freedom of choosing $t$ may lead to better bounds. Let us restrict ourselves to the special case when $n=3$ (that is, we look at the 2-dimensional sphere). For a range of $t\in [-1,\cos \frac{2\pi}5]$, the best construction that we could find consists of one or two spherical caps as follows. Given $t$, let $h$ be the maximum height of an open spherical cap which is $\{t\}$-avoiding. A simple calculation shows that $h=1-\sqrt{(t+1)/2}$. If $t\le -1/2$, then we just take a single cap $C$ of height $h$, which gives that $\alpha_{\{t\}}(3)\ge h/2$ then. When $-1/2< t\le 0$, we can add another cap $C'$ whose centre is opposite to that of $C$. When $t$ reaches $0$, the caps $C$ and $C'$ have the same height (and we get the two-cap construction from Kalai's conjecture). When $0<t\le \frac{2\pi}5$, we can form a $\{t\}$-avoiding set by taking two caps of the same height $h$. (Note that the last construction cannot be optimal for $t>\frac{2\pi}5$, as then the two caps can be arranged so that a set of positive measure can be added, see the third picture of Figure~\ref{fg:1}.) \begin{figure} \begin{center} \includegraphics[height=4cm]{C3.pdf}\hspace{1cm} \includegraphics[height=4cm]{C4.pdf}\hspace{1cm} \includegraphics[height=4cm]{C5.pdf} \end{center} \caption{$\{t\}$-Avoiding set for $t=-\frac12$, $0$ and $\cos\frac{2\pi}5$} \label{fg:1} \end{figure} Calculations show that the above construction gives the following lower bound (where $h=1-\sqrt{(t+1)/2}$): \begin{equation}\label{eq:lower} \alpha_{\{t\}}(3)\ge \left\{\begin{array}{ll} \frac h2,& -1\le t\le -\frac12,\\ h+t-ht,& -\frac12\le t\le 0,\\ h,& 0\le t\le \cos \frac{2\pi}5. \end{array}\right. \end{equation} We conjecture.that the bounds in (\ref{eq:lower}) are all equalities. In particular, our conjecture states that, for $t\le -1/2$, we can strengthen Levy's isodiametric inequality by forbidding a single inner product $t$ instead of the whole interval $[-1,t]$. As in Section~\ref{sec:lp+comb}, one can write an infinite linear program that gives an upper bound on $\alpha_{\{t\}}(3)$. Although our numerical experiments indicate that the upper bound given by the LP exceeds the lower bound in (\ref{eq:lower}) by at most $0.062$ for all $-1\le t\le 0.3$, we were not able to determine the exact value of $\alpha_{\{t\}}(3)$ for any single $t\in(0,\cos \frac{2\pi}5]$. \section*{Acknowledgements} Both authors acknowledge Anusch Taraz and his research group for their hospitality during the summer of 2013. The first author would like to thank his thesis advisor Frank Vallentin for careful proofreading, and for pointing him to the Witsenhausen problem \cite{witsenhausen74}. \bibliographystyle{alpha}

https://arxiv.org/abs/1408.6982

Optimality conditions for the buckling of a clamped plate

We prove the following uniqueness result for the buckling plate. Assume there exists a smooth domain which minimizes the first buckling eigenvalue for a plate among all smooth domains of given volume. Then the domain must be a ball. The proof uses the second variation for the buckling eigenvalue and an inequality by L. E. Payne to establish this result.

\section{Introduction} We consider the following variational problem. Let $\Omega \subset \mathbb{R}^n$ be a bounded domain and let \begin{eqnarray*} {\cal{R}}(u,\Omega):=\frac{\int\limits_{\Omega}\vert\Delta u\vert^2\:dx}{\int\limits_{\Omega}\vert\nabla u\vert^2\:dx} \end{eqnarray*} for $u\in H^{2,2}_0(\Omega)$. We set ${\cal{R}}(u,\Omega)=\infty$ if the denominator vanishes. We define \begin{eqnarray}\label{eigen} \Lambda(\Omega):=\inf\left\{ {\cal{R}}(u,\Omega)\: :\:u\in \: H^{2,2}_0(\Omega)\right\}. \end{eqnarray} The infimum is attained by the first eigenfunction $u$, which solves the Euler Lagrange equation \begin{eqnarray} \label{eq1} \Delta^2u+\Lambda(\Omega)\Delta u&=&0\quad\hbox{in}\:\Omega\\ \label{eq2} u=\partial_\nu u&=&0\quad\hbox{in}\:\partial\Omega. \end{eqnarray} If we normalize $u$ by $\| \nabla u\|_{L^{2}(\Omega)} =1$, the first eigenfunction is uniquely determined. Otherwise any multiple of $u$ is an eigenfunction as well. The sign of the first eigenfunction may change depending on $\Omega$. \newline \newline The quantity $\Lambda(\Omega)$ is called buckling eigenvalue of $\Omega$. It is well known that there is a discrete spectrum of positive eigenvalues of finite multiplicity and their only accumulation point is $\infty$. The corresponding eigenfunctions form an orthonormal basis of $H^{2,2}_0(\Omega)$. \newline \newline In the sequel, we will assume that $u$ is normalized. If we multiply \eqref{eq1} with $x\cdot\nabla u$ and integrate by parts, we obtain \begin{eqnarray}\label{Rel} \Lambda(\Omega) = \frac{1}{2}\int\limits_{\partial\Omega}\vert\Delta u\vert^2\, x\cdot\nu\:dS. \end{eqnarray} In 1951, G. Polya and G. Szeg\"o formulated the following conjecture (see \cite{PoSz51}). \newline\newline \centerline{{\it{Among all domains $\Omega$ of given volume, the ball minimizes $\Lambda(\Omega).$}}} \newline\newline This conjecture is still open. However, partial results are known. In \cite{Sz50} Szeg\"o proved the conjecture for all smooth domains under the additional assumption that $u > 0$ in $\Omega$. M.S. Ashbaugh and D. Bucur proved that among simply connected domains of prescribed volume there exists an optimal domain \cite{AB03}. In \cite{Wi95} H. Weinberg and B. Willms proved the following uniqueness result for $n=2$. If an optimal plane domain $\Omega$ exists and if $\partial\Omega$ is smooth (at least $C^{2,\alpha}$), then $\Omega$ is a disc. \newline\newline There also exist bounds for $\Lambda(\Omega)$. We only mention Payne's inequality (see \cite{P55}) which states that \begin{eqnarray*} \Lambda(\Omega) \geq \lambda_2(\Omega), \end{eqnarray*} where $\lambda_2$ denotes the second Dirichlet eigenvalue for the Laplacian. Equality holds if and only if $\Omega$ is a ball. \newline\newline In this paper, we assume that there exists an optimal domain $\Omega$, which is smooth and simply connected. We will prove that $\Omega$ must be a ball. Thus we generalize the result of H. Weinberg and B. Willms in \cite{Wi95} to higher dimensions. \newline\newline To consider the second domain variation for $\Lambda(\Omega)$ is motivated by the work of E. Mohr in \cite{Mo75}. He was interested in the clamped plate eigenvalue, where \[ {\cal{R}}(u,\Omega) = \frac{\int\limits_\Omega \vert\Delta u\vert^2\:dx}{\int\limits_\Omega u^2\:dx} \] and $\Omega$ is a smoothly bounded domain in $\mathbb{R}^2$. For the corresponding eigenvalue he computed the second domain variation. The explicit computation of the kernel of the second domain variation then implies that the disc is a unique minimizer among smooth domains of equal volume. \newline\newline Our strategy will be as follows. In Chapter 2 we introduce a smooth family $(\Omega_t)_t$ of perturbations of $\Omega$ of equal volume. We denote by $\Lambda(t) := \Lambda(\Omega_t)$ the corresponding first buckling eigenvalue of $\Omega_t$. As a consequence of the optimality of $\Omega$, the eigenfunction $u$ statisfies the overdetermined boundary value problem \begin{align*} \Delta^2 u + \Lambda(\Omega) \Delta u &=0 \; \mbox{in } \Omega \\ u = \partial_\nu u &=0 \; \mbox{in } \partial\Omega \\ \Delta u &= c_0 \; \mbox{in } \partial\Omega, \mbox{ where } c_0 = \frac{2\Lambda(\Omega)}{\vert\Omega\vert} \text{ by } \eqref{Rel}. \end{align*} This follows from the fact that the first domain variation of $\Lambda(\Omega)$ - computed in Chapter 3 - for an optimal domain necessarily vanishes. \newline In Chapter 4 we compute the second domain variation of $\Lambda(\Omega)$. It turns out that \begin{equation}\label{ddotLambda} \ddot{\Lambda}(0) = \frac{d^2}{dt^2}\Lambda(t) \bigg{|}_{t=0} = 2\int\limits_\Omega\vert\Delta u'\vert^2 -2\Lambda(\Omega)\int\limits_\Omega\vert\nabla u'\vert^2\:dx, \end{equation} where $u'$ is the so called shape derivative of $u$. It solves \begin{eqnarray} \label{shape1} \Delta^2 u' + \Lambda(\Omega) \Delta u' &=&0 \; \mbox{in } \Omega \\ \label{shape2} u' &=&0 \; \mbox{in } \partial\Omega \\ \label{shape3}\partial_\nu u' &=& -c_0 v.\nu \; \mbox{in } \partial\Omega \end{eqnarray} and \begin{eqnarray} \label{shape4}\int\limits_\Omega \nabla u.\nabla u' \: dx =0. \end{eqnarray} The vector field $v$ is the first order approximation of $\Omega_t$ in the sense that for $y \in \Omega_t$ there exists an $x\in\Omega$ such that \begin{eqnarray*} y = x + tv(x) + o(t). \end{eqnarray*} Thus, $\ddot{\Lambda}(0)$ is equal to a quadratic functional in the shape derivative $u'$ which we denote by ${\cal{E}}(u')$ and ${\cal{E}}(u')$ is given by the right hand side of \eqref{ddotLambda}. Since we assume the optimality of $\Omega$, we have ${\cal{E}}(u')\geq 0$. It turns out that the kernel of ${\cal{E}}(u')$ contains the directional derivatives $\partial_1 u,\ldots, \partial_n u$ of $u$. Each directional derivative is a shape derivative, which corresponds to a domain perturbation given by translations. \newline \newline The key idea is to enlarge the class of shape derivatives on which ${\cal{E}}$ is defined. This new class will be denoted by ${\cal{Z}}$ and contains the shape derivatives as a true subset. Nevertheless we can show that ${\cal{E}}$ is still bounded from below and even nonnegative on ${\cal{Z}}$. Moreover $\min_{{\cal{Z}}} {\cal{E}}=0$ since the directional derivatives of $u$ are in ${\cal{Z}}$. This is done in Chapter 5. In Chapter 6 we construct a function $\psi\in{\cal{Z}}$ for which we will show \begin{eqnarray*} 0\leq {\cal{E}}(\psi)\leq \left(\lambda_2(\Omega)-\Lambda(\Omega)\right)\lambda_2(\Omega). \end{eqnarray*} By Payne's inequality we have equality and this proves that the optimal domain is a ball. \newline\newline Some of these results were obtained in the Diplom thesis of the first author \cite{Kn08}. \section{Domain variations}\label{DomVar} Let $\Omega$ be a bounded smooth (at least $C^{2,\alpha}$) and simply connected domain in $\mathbb{R}^n$. We denote by $\nu$ the unit normal vector field on $\partial\Omega$. Let $\delta$ be the distance function to the boundary, i.e. for $x\in\overline{\Omega}$ we have \begin{eqnarray*} \delta(x):=\inf\{\vert x-z\vert\: :\:z\in\,\partial\Omega\}. \end{eqnarray*} Then, for smooth $\partial\Omega$, $\nu:=\nabla\delta$ defines a smooth extension of $\nu$ into a sufficiently small tubular neighbourhood of $\partial\Omega$. With this the following identities hold. \begin{eqnarray}\label{normal} \nu\cdot\nu=1,\qquad \nu\cdot D\nu=0\qquad\hbox{and}\qquad D\nu\cdot\nu=0 \end{eqnarray} on $\partial\Omega$. See e.g. Proposition 5.4.14 in \cite{HePi05} for a proof. \newline \newline Moreover, the mean curvature of $\partial\Omega$ is bounded since $\Omega$ is smooth, i.e. for each $x \in \partial\Omega$ there holds \begin{equation}\label{H_dOmega} \vert H_{\partial\Omega}(x) \vert \leq \max_{\partial\Omega} \vert H_{\partial\Omega}\vert < \infty. \end{equation} We will frequently use integration by parts on $\partial\Omega$. Let $f\in C^1(\partial\Omega)$ and $v\in C^{0,1}(\partial\Omega,\mathbb{R}^n)$. The next formula is often called the {\sl Gauss theorem on surfaces}. \begin{eqnarray}\label{tanpi} \oint_{\partial\Omega}f\:\hbox{div}\:_{\partial\Omega}v\:dS =-\oint_{\partial\Omega}v\cdot\nabla^\tau f\:dS +(n-1)\oint_{\partial\Omega}f(v\cdot\nu)\:H_{\partial\Omega}\:dS, \end{eqnarray} where \begin{align}\label{tgradient} \nabla^{\tau} f =\nabla f -(\nabla f\cdot \nu)\nu \end{align} denotes the tangential gradient of $f$. \newline \newline In this chapter, we describe the class of admissible variations for the domain functional $\Lambda(\Omega)$. For given $t_0>0$ and $t\in(-t_0,t_0)$ let $(\Omega_t)_t$ be a family of perturbations of the domain $\Omega\subset \mathbb{R}^n$ of the form \begin{eqnarray*} \Omega_t=\Phi_{t}(\Omega) \end{eqnarray*} where \begin{eqnarray*} \Phi_{t}\::\:\overline{\Omega}\to\mathbb{R}^n \end{eqnarray*} is a diffeomorphism which is smooth in $t$ and $x$. Thus we may write \begin{eqnarray*} \Omega_t:=\{y=x+tv(x)+\frac{t^2}{2}w(x)+o(t^2):x \in \Omega,\: t\tx{small}\}, \end{eqnarray*} where \begin{eqnarray*} v=(v_1(x),v_2(x),\dots,v_n(x))=\partial_{t}\Phi_{t}(x)\vert_{t=0} \end{eqnarray*} and \begin{eqnarray*} w=(w_1(x),w_2(x),\dots,w_n(x))=\partial^2_{t}\Phi_{t}(x)\vert_{t=0} \end{eqnarray*} are smooth vector fields and where $o(t^2)$ collects terms such that $\frac{o(t^2)}{t^2}\to 0$ as $t\to 0$. For small $t_0$ the sets $\Omega_t$ and $\Omega$ are diffeomorphic. We will frequently use the notation $y:=\Phi_{t}(x)$. Consider the functional \begin{eqnarray* \Lambda(\Omega_t):=\inf\left\{ {\cal{R}}(u,\Omega_t)\: :\:u\in \: H^{2,2}_0(\Omega_t)\right\}, \end{eqnarray*} which only depends on $\Omega_t$. Let $u(t,y)\in H^{2,2}_0(\Omega_t)$ be the minimizer. For short we will write \begin{eqnarray}\label{t-u} \tilde{u}(t):=u(t,y). \end{eqnarray} Then $\tilde{u}(t)$ solves \begin{eqnarray} \label{t-eq1}\Delta^2\tilde{u}(t)+\Lambda(\Omega_t)\Delta\tilde{u}(t) &=&0\quad\hbox{in}\:\Omega_t\\ \label{t-eq2}\tilde{u}(t)=\vert\nabla\tilde{u}(t)\vert&=&0\quad\hbox{in}\:\partial\Omega_t. \end{eqnarray} for each $t\in(-t_0,t_0)$. With this notation we define \begin{eqnarray* \Lambda(t):= {\cal{R}}(\tilde{u}(t),\Omega_t). \end{eqnarray*} Since we assume smoothness of $\Omega$ and $\Phi_t$ the eigenfunction $\tilde{u}$ is also smooth in $t$ and $x$. This has several consequences which we list as remarks. \begin{remark} Since $\partial\Omega_t$ is smooth and since $\tilde{u}(t)=0$ on $\partial\Omega_t$ then necessarily \begin{eqnarray} \label{lapbc1}\Delta \tilde{u}=\partial_{\nu}^2\tilde{u}+(n-1)\partial_{\nu}\tilde{u} \:H_{\partial\Omega_t}\quad\hbox{in}\:\partial\Omega_t, \end{eqnarray} where $H_{\partial\Omega_t}$ denotes the mean curvature of $\partial\Omega_t$. Clearly, if $\tilde{u}=\vert\nabla \tilde{u}\vert =0$ on $\partial\Omega_t$, then necessarily \begin{eqnarray} \label{lapbc2}\Delta \tilde{u}=\partial_{\nu}^2\tilde{u}\quad\hbox{in}\:\partial\Omega_t. \end{eqnarray} \end{remark} \begin{remark} Since \eqref{t-eq2} holds for all $t\in(-t_0,t_0)$, we also have \begin{eqnarray} \label{tdot-eq2}\dot{\tilde{u}}(t)=\vert\nabla\dot{\tilde{u}}(t)\vert=0\quad\hbox{in}\:\partial\Omega_t \end{eqnarray} for all $t\in(-t_0,t_0)$. \end{remark} \begin{remark} Straightforward computations yield \begin{eqnarray*} \dot{\tilde{u}}(t)=\frac{d}{dt}u(t,y)=\partial_{t}u(t,\Phi_{t}(\Phi^{-1}_{t}(y))+\partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y))\cdot\nabla u(t,y) \end{eqnarray*} for all $t\in(-t_0,t_0)$. Let $y\in\partial\Omega_t$. Then \eqref{tdot-eq2} and \eqref{t-eq2} imply \begin{eqnarray} \label{tdot-eq3} 0=\dot{\tilde{u}}(t)=\partial_{t}u(t,y)\quad\hbox{for $y$ in}\:\partial\Omega_t \end{eqnarray} for all $t\in(-t_0,t_0)$. \end{remark} In particular for $t=0$ we compute $\tilde{u}(0)=u(x)$ and \begin{eqnarray*} \dot{\tilde{u}}(0)&=& \partial_{t}u(0,x)+v(x)\cdot Du(0,x)\\ \ddot{\tilde{u}}(0)&=& \partial^2_{t}u(0,x)+2v(x)\cdot D\partial_{t}u(0,x)+w(x)\cdot Du(0,x) +v(x)\cdot D\left(v(x)\cdot Du(0,x)\right). \end{eqnarray*} We will use the notation \begin{eqnarray*} u'(x):=\partial_{t}u(0,x)\qquad\hbox{and}\qquad u''(x):=\partial^2_{t}u(0,x). \end{eqnarray*} Hence, \begin{eqnarray} \label{tildeu1} \dot{\tilde{u}}(0)&=& u'(x)+v(x)\cdot Du(x)\\ \label{tildeu2} \ddot{\tilde{u}}(0)&=& u''(x)+2v(x)\cdot Du'(x)+w(x)\cdot Du(x) +v(x)\cdot D\left(v(x)\cdot Du(x)\right). \end{eqnarray} Note that all these quantities are defined for $x\in\overline{\Omega}$. For $x\in\partial\Omega$ we thus get \begin{eqnarray*} 0 = \dot{\tilde{u}}(0) = u'(x)\qquad \hbox{and} \qquad 0=\nabla\dot{\tilde{u}}(0) =\nabla u'(x)+v(x)\cdot D^2u(x), \end{eqnarray*} where $(v(x)\cdot D^2u(x))_{j}=\sum_{i=1}^{n}v_{i}(x)\partial_{i}\partial_{j}u(x)$ for $j=1,\hdots,n$. Thus, we get the following boundary conditions for $u'$. \begin{eqnarray}\label{uprimebc} u'(x)=0\qquad\hbox{and}\qquad \partial_{\nu}u'(x)=-v(x)\cdot D^2u(x)\cdot\nu(x)\quad \hbox{for}\quad x\in\partial\Omega. \end{eqnarray} Here we used the notation $v(x)\cdot D^2u(x)\cdot\nu(x)=\sum_{i,j=1}^{n}v_{i}(x) \partial_{i}\partial_{j}u(x)\nu_{j}(x)$. \newline \newline Let $\nu_{t}(y)$ be the unit normal vector in $y\in\partial\Omega_t$. We also write this as \begin{eqnarray} \label{normal1}\nu_{t}(y)=\nu(t,\Phi_{t}(x))\qquad \forall\:t\in(-t_0,t_0)\qquad x\in\partial\Omega. \end{eqnarray} Then we have \begin{eqnarray}\label{normal2} \nu'=-\nabla^{\tau}(v\cdot\nu),\qquad\nu\cdot\nu'=0. \end{eqnarray} This follows from direct calculations (see e.g. (5.64) in \cite{HePi05}). \begin{lemma} With the notation from above the following equality holds. \begin{eqnarray} \label{tdot-eq4} \nu_{t}\cdot\nabla(\partial_{t}u(t,y))=-\Delta u(t,y)\:\nu_{t}\cdot \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y))\quad\hbox{for $y$ in}\:\partial\Omega_t \end{eqnarray} for all $t\in(-t_0,t_0)$. Alternatively, we write this for all $t\in(-t_0,t_0)$ and $x\in\partial\Omega$ as \begin{eqnarray} \label{tdot-eq5} \nu(t,\Phi_{t}(x))\cdot\nabla\{\partial_{t}u(t,\Phi_{t}(x))\}=-\Delta u(t,\Phi_{t}(x))\:\nu({t},\Phi_{t}(x))\cdot \partial_{t}\Phi_{t}(x). \end{eqnarray} \end{lemma} {\bf{Proof}} Since $\nabla u(t,\Phi_{t}(x))=0$ for all $\vert t\vert<t_0$ and all $x\in\partial\Omega$ we have \begin{eqnarray*} 0=\frac{d}{dt}\:\nabla u(t,\Phi_{t}(x))=\nabla \partial_{t}u(t,\Phi_{t}(x)) + D^2u(t,\Phi_{t}(x))\cdot \partial_{t}\Phi_{t}(x). \end{eqnarray*} This implies \begin{eqnarray*} 0=\nu_{t}\cdot\nabla(\partial_{t}u(t,y))+\nu_{t}\cdot D^2u(t,y)\cdot \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y))\quad\hbox{for $y$ in}\:\partial\Omega_t \end{eqnarray*} for all $t\in(-t_0,t_0)$. Here we used the notation \begin{eqnarray*} \nu_{t}\cdot D^2u(t,y)\cdot \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y)) = \sum_{ij=1}^{n}\nu_{t,i}\cdot \partial_{i}\partial_{j}u(t,y)\cdot \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y))_{j}. \end{eqnarray*} Since $\nabla\tilde{u}(t)=0$ in $\partial\Omega_t$, we get \begin{eqnarray*} \nu_{t}\cdot D^2u(t,y)\cdot \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y)) = \nu_{t}\cdot D^2u(t,y)\cdot\nu_{t}\:\nu_{t}\cdot \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y)). \end{eqnarray*} Thus, \begin{eqnarray*} \nu_{t}\cdot\nabla(\partial_{t}u(t,y))=-\nu_{t}\cdot D^2u(t,y)\cdot\nu_{t}\:\nu_{t}\cdot \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y))\quad\hbox{for $y$ in}\:\partial\Omega_t. \end{eqnarray*} Formula \eqref{lapbc2} simplifies to \begin{eqnarray*} \nu_{t}\cdot\nabla(\partial_{t}u(t,y))=-\Delta u(t,y)\:\nu_{t}\cdot \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y))\quad\hbox{for $y$ in}\:\partial\Omega_t. \end{eqnarray*} This proves the lemma. \hfill $\square$ \newline \newline The first derivative of $\Lambda(t)$ with respect to the parameter $t$ is called {\sl the first domain variation} and the second derivative is called {\sl the second domain variation}. \newline \newline Our domain variations will be chosen within the class of volume preserving pertubations up to order $2$. Hence, they are chosen such that \begin{eqnarray}\label{volume} {\cal{L}}^n(\Omega_t)={\cal{L}}^n(\Omega) +o(t^2) \end{eqnarray} holds. This puts constraints on the vector fields $v$ and $w$. They were discussed e.g. in \cite{BaWa14}, formula (2.13) and Lemma 1. \begin{lemma}\label{volpres1} Let $v,w\in C^{0,1}(\Omega,\mathbb{R}^n)$ be such that \eqref{volume} holds. Then \begin{align}\label{volume1} \int\limits_{\Omega} \hbox{div}\: v\:dx =0 \end{align} and \begin{eqnarray*} \int\limits_{\Omega}\left((\hbox{div}\: v)^2-Dv : Dv+\hbox{div}\:\:w\right)\:dx=0, \end{eqnarray*} where $Dv:Dv=\sum_{i,j=1}^{n}\partial_{i}v_{j}\:\partial_{j}v_{i}$. The second equality is equivalent to \begin{align}\label{volume2} \int\limits_{\partial\Omega}(v\cdot\nu) \hbox{div}\: v\:dS-\int\limits_{\partial\Omega}v\cdot Dv\cdot\nu\:dS + \int\limits_{\partial\Omega}(w\cdot \nu)\:dS=0. \end{align} \end{lemma} Note that rotations do not satisfy these conditions (see e.g. Remark 1 in \cite{BaWa14}). \section{The first domain variation} We will use the following formula for the computations of the first domain variation of $\Lambda$. It is well known as Reynolds transport theorem and is analyzed in detail in Chapter 5.2.3 in \cite{HePi05}. \begin{theorem}\label{reynolds1} Let $t\in(-t_0,t_0)$ for some $t_0>0$. Let $\Phi_{t}\in C^{0,1}(\mathbb{R}^n)$ be differentiable in $t$ and let $t\to f(t)\in L^1(\mathbb{R}^n)$ be a function which is differentiable in $t$. Moreover, let $f(t)\in W^{1,1}(\mathbb{R}^n)$. Then $t\to I(t):=\int\limits_{\Omega_t}f(t)\:dy$ is differentiable in $t$. Moreover, we have the formula \begin{eqnarray*} \dot{I}(t)=\int\limits_{\Omega_t}\partial_{t}f(t)+\hbox{div}\:\left(f(t)\: \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y))\right)\:dy. \end{eqnarray*} If $\partial\Omega$ is sufficiently smooth (at least Lipschitz continuous), this is equivalent to \begin{eqnarray*} \dot{I}(t)=\int\limits_{\Omega_t}\partial_{t}{f}(t)\:dy+\int\limits_{\partial\Omega_t}f(t)\: \partial_{t}\Phi_{t}(\Phi^{-1}_{t}(y))\cdot\nu(y)\:dS(y). \end{eqnarray*} In particular, for $t=0$ we get \begin{eqnarray*} \dot{I}(0)=\int\limits_{\Omega}\partial_{t} f (t)\vert_{t=0}+\hbox{div}\:\left(f(0)\: v(x)\right)\:dx. \end{eqnarray*} Again, if $\partial\Omega$ is sufficiently smooth, this is equivalent to \begin{eqnarray*} \dot{I}(0)=\int\limits_{\Omega}\partial_{t} f (t)\vert_{t=0}\:dx+\int\limits_{\partial\Omega}f(0)\: v(x)\cdot\nu(x)\:dS(x). \end{eqnarray*} \end{theorem} We apply this formula to $\Lambda(t)=\frac{D(t)}{N(t)}$ where \begin{eqnarray*} D(t):=\int\limits_{\Omega_t}\vert\Delta\tilde{u}(t)\vert^2\:dy\qquad\hbox{and}\qquad N(t):=\int\limits_{\Omega_t}\vert\nabla\tilde{u}(t)\vert^2\:dy \end{eqnarray*} and we assume the normalization \begin{eqnarray}\label{norm} N(t)=\int\limits_{\Omega_t}\vert\nabla\tilde{u}(t)\vert^2\:dy=1\quad\forall\:t\in(-t_0,t_0). \end{eqnarray} We then obtain \begin{eqnarray*} \label{ladot}\dot{\Lambda}(t)&=&2\int\limits_{\Omega_t}\Delta\tilde{u}(t)\:\Delta\partial_{t}{\tilde{u}}(t)\:dy - 2\:\Lambda(t)\:\int\limits_{\Omega_t}\nabla\tilde{u}(t)\cdot\nabla\partial_{t}{\tilde{u}}(t)\:dy\\ &&+ \nonumber\int\limits_{\partial\Omega_t}\vert\Delta\tilde{u}(t)\vert^2\:\partial_{t}\Phi_{t}(\Phi_{t}^{-1}(y))\cdot\nu_{t}(y)\:dS(y), \end{eqnarray*} where $\nu_{t}(y)$ denotes the unit normal vector in $y\in\partial\Omega_t$. We integrate by parts and use \eqref{tdot-eq2}. Then \begin{eqnarray*} \dot{\Lambda}(t)&=& 2\int\limits_{\Omega_t}\left\{\Delta^2\tilde{u}(t)+\Lambda(t)\:\Delta\tilde{u}(t)\right\}\partial_{t}{\tilde{u}}(t)\:dy + 2\int\limits_{\partial\Omega_t}\Delta\tilde{u}(t)\:\partial_{\nu_t}\partial_{t}{\tilde{u}}(t)\:dS(y)\\ &&- 2\int\limits_{\partial\Omega_t}\partial_{\nu_t}\Delta\tilde{u}(t)\partial_{t}{\tilde{u}}(t)\:dS(y) + \int\limits_{\partial\Omega_t}\vert\Delta\tilde{u}(t)\vert^2\:\partial_{t}\Phi_{t}(\Phi_{t}^{-1}(y))\cdot\nu_{t}(y)\:dS(y). \end{eqnarray*} The first integral vanishes since $\tilde{u}(t)$ solves \eqref{t-eq1}. The third integral vanishes since \eqref{tdot-eq3} holds. Finally we use \eqref{tdot-eq4}. This proves the following lemma. \begin{lemma}\label{ladot1} Let $\tilde{u}(t)$ be an eigenfunction (i.e. a solution of \eqref{t-eq1} - \eqref{t-eq2}) and assume \eqref{norm} holds. Let \begin{eqnarray*} \Lambda(t)=\int\limits_{\Omega_t}\vert\Delta\tilde{u}(t)\vert^2\:dy. \end{eqnarray*} Then \begin{eqnarray}\label{ladotform} \dot{\Lambda}(t)&=& - \int\limits_{\partial\Omega_t}\vert\Delta\tilde{u}(t)\vert^2\:\partial_{t}\Phi_{t}(\Phi_{t}^{-1}(y))\cdot\nu_{t}(y)\:dS(y). \end{eqnarray} \end{lemma} \begin{remark}\label{mon} Note that if $\partial_{t}\Phi_{t}(\Phi_{t}^{-1}(y))\cdot\nu_{t}(y)>0$, this implies ${\cal{L}}^n(\Omega_t)>{\cal{L}}^n(\Omega)$ for small $t$. Thus, $\dot{\Lambda}(t)$ is negative in this case. From this we conclude that the first buckling eigenvalue is decreasing under set inclusion. \end{remark} From Lemma \ref{ladot1} we get in particular \begin{eqnarray*} \dot{\Lambda}(0)=-\int\limits_{\partial\Omega}\vert\Delta u\vert^2\:v(x)\cdot\nu(x)\:dS(x). \end{eqnarray*} From Lemma \ref{volpres1} and \eqref{volume1} we deduce $\vert\Delta u\vert=const.$ if $\Omega$ is a critical point of $\Lambda(t)$. Due to formula \eqref{Rel}, this constant is equal to \begin{eqnarray}\label{constant} c_0 := \frac{2\Lambda(0)}{\vert\Omega\vert}. \end{eqnarray} We denote this result as a theorem. \begin{theorem}\label{overdet1} Let $\Omega_t$ be a family of volume preserving perturbations of $\Omega$ as described in Chapter 2. Then $\Omega$ is a critical point of the energy $\Lambda(t)$, i.e. $\dot{\Lambda}(0)=0$, if and only if \begin{eqnarray}\label{overrt} \Delta u=c_0\qquad\hbox{on}\quad\partial\Omega. \end{eqnarray} \end{theorem} In particular, $u$ is a solution of the overdetermined boundary value problem \begin{eqnarray} \label{ceq1} \Delta^2u+\Lambda(\Omega)\Delta u&=&0\quad\hbox{in}\:\Omega\\ \label{ceq2} u=\partial_{\nu}\nabla u&=&0\quad\hbox{in}\:\partial\Omega.\\ \label{ceq3} \Delta u&=&c_0>0\quad\hbox{in}\:\partial\Omega. \end{eqnarray} Note that if we set $U:=\Delta u+\Lambda(\Omega) u$ \eqref{ceq1} - \eqref{ceq3} implies \begin{eqnarray*} \Delta U=0\:\hbox{in}\: \Omega\: \hbox{ and } \: U=c_0\:\hbox{in}\: \partial\Omega. \end{eqnarray*} Hence, \begin{eqnarray}\label{w-eq} U= \Delta u+\Lambda(\Omega) u=c_0\qquad\hbox{in}\:\:\overline{\Omega}. \end{eqnarray} From \cite{Wi95} we know that for $n=2$ this implies that $\Omega$ is a ball. In particular, \begin{eqnarray}\label{w-eq1} \partial_{\nu}\Delta u=0\qquad\hbox{in}\:\:\partial{\Omega}. \end{eqnarray} \section{The second domain variation} Throughout this chapter we assume that $\Omega$ is an optimal domain, i.e. $\dot{\Lambda}(0)=0$ and $\ddot{\Lambda}(0)\geq 0$. This implies that $u$ solves \eqref{ceq1} - \eqref{ceq3} and \eqref{w-eq}. As a consequence \eqref{uprimebc} reads as \begin{eqnarray} \label{upricc} u'(x)=0\qquad\hbox{and}\qquad\partial_{\nu}u'(x)=-c_0\:v(x)\cdot \nu(x)\quad \hbox{for}\quad x\in\partial\Omega. \end{eqnarray} Note that if we differentiate \eqref{t-eq1} - \eqref{t-eq2} in $t=0$ and use the fact that $\dot{\Lambda}(0)=0$, we obtain an equation for $u'$: \begin{eqnarray}\label{uprimeq} \Delta^2 u'(x)+\Lambda(\Omega)\Delta u'(x)=0\qquad\hbox{in}\:\:\Omega. \end{eqnarray} The boundary conditions for $u'$ are given by \eqref{upricc}. Furthermore, the normalization \eqref{norm} implies \begin{eqnarray}\label{graduprime} \int\limits_\Omega \nabla u\cdot\nabla u'\:dx =0. \end{eqnarray} We recall formula \eqref{ladotform}. Before we differentiate with respect to $t$ again we state the following consequence of Reynold's theorem (see e.g. Chapter 5.4.2 in \cite{HePi05}). \begin{theorem}\label{reynolds2} Let $\Omega$ be a bounded smooth domain of class $C^3$. Let $t\in(-t_0,t_0)$ and let $\Phi_{t}\in C^{0,1}(\mathbb{R}^n)$ be differentiable in $t$. Let $t\to g(t)\in L^1(\mathbb{R}^n)$ be a function which is differentiable in $t$. Moreover, let $g(t)\in W^{1,1}(\mathbb{R}^n)$. Then $t\to J(t):=\int\limits_{\partial\Omega_t}g(t)\:dS(y)$ is differentiable in $t$. Moreover, for $t=0$ we have the formula \begin{eqnarray*} \dot{J}(0)=\int\limits_{\partial\Omega}\partial_{t}g(0)+(v(x)\cdot\nu(x))\left\{\partial_{\nu}g(0)+(n-1)g(0)\:H_{\partial\Omega}(x)\right\}\:dS(x), \end{eqnarray*} where $H_{\partial\Omega}$ denotes the mean curvature of $\partial\Omega$ in $x$. \end{theorem} We apply this theorem to \eqref{ladotform}. It is convenient to apply \eqref{tdot-eq4} and to rewrite \eqref{ladotform} as \begin{eqnarray*} \dot{\Lambda}(t)&=& \int\limits_{\partial\Omega_t}\Delta\tilde{u}(t)\:\nu_{t}\cdot\nabla(\partial_{t}u(t,y))\:dS(y). \end{eqnarray*} Let \begin{eqnarray*} g(t):=\Delta\tilde{u}(t)\:\nu_{t}\cdot\nabla(\partial_{t}u(t,y)). \end{eqnarray*} An application of Theorem \ref{reynolds2} yields \begin{eqnarray} \ddot{\Lambda}(0)&=& \label{ddlamb}\int\limits_{\partial\Omega}\Delta u'\:\partial_{\nu}u'\:dS + \int\limits_{\partial\Omega}\Delta u\:\nu'\cdot\nabla u'\:dS + \int\limits_{\partial\Omega}\Delta u\:\partial_{\nu}u''\:dS\\ \nonumber&&+ \int\limits_{\partial\Omega}(v\cdot\nu)\:\partial_{\nu}(\Delta u\:\partial_{\nu}u')\:dS +(n-1) \int\limits_{\partial\Omega}(v\cdot\nu)\:\Delta u\:\partial_{\nu}u'\:H_{\partial\Omega}\:dS. \end{eqnarray} Note that \begin{eqnarray*} \nu_{t}\cdot\nu_{t}=1\:\:\hbox{in}\:\:\partial\Omega_{t}\quad\Longrightarrow\quad\nu\cdot\nu'=0\:\:\hbox{in}\:\:\partial\Omega, \end{eqnarray*} where \begin{eqnarray*} \nu'(x)=\partial_{t}\nu(t,\Phi_{t}(x))\vert_{t=0}\quad \hbox{for} \quad x\in\partial\Omega. \end{eqnarray*} Since \eqref{upricc} implies $\nabla u'=\partial_{\nu}u'\:\nu$, this implies \begin{eqnarray*} \int\limits_{\partial\Omega}\Delta u\:\nu'\cdot\nabla u'\:dS=0. \end{eqnarray*} For the fourth integral we apply \eqref{overrt} and \eqref{w-eq1}. \begin{eqnarray*} \int\limits_{\partial\Omega}(v\cdot\nu)\:\partial_{\nu}(\Delta u\:\partial_{\nu}u')\:dS &=& \int\limits_{\partial\Omega}(v\cdot\nu)\:\partial_{\nu}\Delta u\:\partial_{\nu}u'\:dS + \int\limits_{\partial\Omega}(v\cdot\nu)\:\Delta u\:\partial^2_{\nu}u'\:dS\\ \nonumber&&= 0+c_0\:\int\limits_{\partial\Omega}(v\cdot\nu)\:\partial^2_{\nu}u'\:dS. \end{eqnarray*} With the help of \eqref{upricc} and \eqref{lapbc1} we write \begin{eqnarray*} \partial^2_{\nu}u'=\Delta u'-(n-1)\partial_{\nu}u'\:H_{\partial\Omega}. \end{eqnarray*} Hence, \begin{eqnarray*} \int\limits_{\partial\Omega}(v\cdot\nu)\:\partial_{\nu}(\Delta u\:\partial_{\nu}u')\:dS = c_0\:\int\limits_{\partial\Omega}(v\cdot\nu)\:\Delta u'\:dS - c_0(n-1)\:\int\limits_{\partial\Omega}(v\cdot\nu)\:\partial_{\nu}u'\:H_{\partial\Omega}\:dS. \end{eqnarray*} Our computations yield a first simplification of \eqref{ddlamb}: \begin{align*} \ddot{\Lambda}(0) = \int\limits_{\partial\Omega}\Delta u'\:\partial_\nu u'\:dS + \int\limits_{\partial\Omega}\Delta u\: \partial_\nu u''\:dS + c_0\int\limits_{\partial\Omega}(v\cdot\nu)\:\Delta u'\:dS. \end{align*} In the first integral on the right hand side we use \eqref{upricc} again. Thus, we get \begin{eqnarray}\label{ddlamb1} \ddot{\Lambda}(0)=c_0\:\int\limits_{\partial\Omega}\partial_{\nu}u''\:dS \end{eqnarray} In order to find a lower bound for $\ddot{\Lambda}(0)$, we analyze the integral in \eqref{ddlamb1}. Recall \eqref{tdot-eq5}. We differentiate this equation with respect to $t$ in $t=0$. Then \eqref{w-eq1} and \eqref{overrt} yield \begin{eqnarray*} &&\nu'\cdot\nabla u' + v\cdot D\nu\cdot\nabla u' + \partial_{\nu}u'' + \nu\cdot D^2u'\cdot v =\\ &&- \Delta u'\:(v\cdot\nu) - c_0\:(v\cdot\nu') - c_0v\cdot D\nu\cdot v - c_0\:(w\cdot\nu). \end{eqnarray*} As before, $\nu'\cdot\nabla u'=0$ on $\partial\Omega$. Moreover, by \eqref{upricc} \begin{eqnarray*} v\cdot D\nu\cdot\nabla u'=- c_0v\cdot D\nu\cdot \nu\:(v\cdot\nu)=0, \end{eqnarray*} where the last equality follows from \eqref{normal}. Thus, \begin{eqnarray}\label{lambdaddot2} \ddot{\Lambda}(0)&=& - c_0\:\int\limits_{\partial\Omega}(v\cdot\nu)\:\Delta u'\:dS - c_0\:\int\limits_{\partial\Omega}\nu\cdot D^2u'\cdot v\:dS\\ \nonumber&&- c_0^2\:\int\limits_{\partial\Omega}(v\cdot\nu')\:dS - c_0^2\:\int\limits_{\partial\Omega}v\cdot D\nu\cdot v\:dS - c_0^2\:\int\limits_{\partial\Omega}(w\cdot\nu)\:dS. \end{eqnarray} For the first integral we use \eqref{upricc} and we observe that Gau\ss\ theorem, partial integration and equation \eqref{uprimeq} for $u'$ gives \begin{eqnarray}\label{int1} - c_0\:\int\limits_{\partial\Omega}(v\cdot\nu)\:\Delta u'\:dS = \int\limits_{\partial\Omega}\Delta u'\:\partial_{\nu}u'\:dS = \int\limits_{\Omega}\vert\Delta u'\vert^2\:dx-\Lambda(\Omega)\int\limits_{\Omega}\vert\nabla u'\vert^2\:dx. \end{eqnarray} The second intergal is slightly more involved. We set $v^{\tau}=v-(v\cdot\nu)\nu$. Since $\nabla u'=(\partial_{\nu}u')\nu$ and since \eqref{lapbc1} can be applied to $u'$, we get \begin{eqnarray*} - c_0\:\int\limits_{\partial\Omega}v\cdot D^2u'\cdot\nu \:dS &=& - c_0\:\int\limits_{\partial\Omega}v^{\tau}\cdot D^2u'\cdot \nu\:dS - c_0\:\int\limits_{\partial\Omega}(v\cdot\nu)\left(\Delta u'-(n-1)\partial_{\nu}u'\:H_{\partial\Omega}\right)\:dS\\ &=& - c_0\:\int\limits_{\partial\Omega}v^{\tau}\cdot D\left(\partial_{\nu}u'\:\nu\right)\cdot \nu\:dS - c_0\:\int\limits_{\partial\Omega}(v\cdot\nu)\:\Delta u'\:dS\\ &&- c_0^2\:(n-1)\:\int\limits_{\partial\Omega}(v\cdot\nu)^2\:H_{\partial\Omega}\:dS. \end{eqnarray*} For the last equality we also used \begin{eqnarray*} v^{\tau}\cdot D\nu\cdot \nu=v^{\tau}\cdot D^{\tau}\nu\cdot \nu=0\quad\hbox{in}\:\partial\Omega. \end{eqnarray*} Next we note that with \eqref{upricc} we have \begin{eqnarray*} - c_0\:\int\limits_{\partial\Omega}v^{\tau}\cdot D\left(\partial_{\nu}u'\:\nu\right)\cdot \nu\:dS &=& - c_0\:\int\limits_{\partial\Omega}v^{\tau}\cdot D^{\tau}\left(\partial_{\nu}u'\:\nu\right)\cdot \nu\:dS = c_0^2\:\int\limits_{\partial\Omega}v^{\tau}\cdot D^{\tau}\left((v\cdot\nu)\:\nu\right)\cdot \nu\:dS\\ &=& c_0^2\:\int\limits_{\partial\Omega}v^{\tau}\cdot \nabla^{\tau}(v\cdot\nu)\:dS, \end{eqnarray*} where the last equality uses \eqref{normal}. \newline For the third integral in \eqref{lambdaddot2} we apply formula \eqref{normal2}: \begin{eqnarray*} - c_0^2\:\int\limits_{\partial\Omega}(v\cdot\nu')\:dS = c_0^2\:\int\limits_{\partial\Omega}\:v\cdot\nabla^{\tau}(v\cdot\nu)\:dS = c_0^2\:\int\limits_{\partial\Omega}\:v^{\tau}\cdot\nabla^{\tau}(v\cdot\nu)\:dS. \end{eqnarray*} These computations simplify \eqref{lambdaddot2} and we obtain \begin{eqnarray}\label{lambdaddot3} \ddot{\Lambda}(0)&=& 2\:\int\limits_{\partial\Omega}\partial_{\nu}u'\:\Delta u'\:dS + 2c_0^2\:\int\limits_{\partial\Omega}v^{\tau}\cdot \nabla^{\tau}(v\cdot\nu)\:dS - c_0^2\:(n-1)\:\int\limits_{\partial\Omega}(v\cdot\nu)^2\:H_{\partial\Omega}\:dS\\ && \nonumber- c_0^2\:\int\limits_{\partial\Omega}v\cdot D\nu\cdot v\:dS - c_0^2\:\int\limits_{\partial\Omega}(w\cdot\nu)\:dS. \end{eqnarray} Next we use the volume constraint \eqref{volume2}. \begin{eqnarray*} - c_0^2\:\int\limits_{\partial\Omega}(w\cdot\nu)\:dS &=& c_0^2\:\int\limits_{\partial\Omega}(v\cdot\nu)\:\hbox{div}\: v\:dS - c_0^2\:\int\limits_{\partial\Omega} v\cdot Dv\cdot\nu\:dS\\ &=& c_0^2\:\int\limits_{\partial\Omega}(v\cdot\nu)\:\hbox{div}\:_{\partial\Omega} v\:dS - c_0^2\:\int\limits_{\partial\Omega} v^{\tau}\cdot D^{\tau}v\cdot\nu\:dS. \end{eqnarray*} We integrate by parts in the first integral (see formula \eqref{tanpi} and \eqref{tgradient}). \begin{eqnarray*} - c_0^2\:\int\limits_{\partial\Omega}(w\cdot\nu)\:dS &=& -c_0^2\:\int\limits_{\partial\Omega}v^{\tau}\cdot D^{\tau}(v\cdot\nu)\:dS + c_0^2(n-1)\:\int\limits_{\partial\Omega}(v\cdot\nu)^2\:H_{\partial\Omega}\:dS\\ &&- c_0^2\:\int\limits_{\partial\Omega} v^{\tau}\cdot D^{\tau}v\cdot\nu\:dS. \end{eqnarray*} Thus, \eqref{lambdaddot3} becomes \begin{eqnarray*} \ddot{\Lambda}(0)&=& 2\:\int\limits_{\partial\Omega}\partial_{\nu}u'\:\Delta u'\:dS + c_0^2\:\int\limits_{\partial\Omega}v^{\tau}\cdot \nabla^{\tau}(v\cdot\nu)\:dS - c_0^2\:\int\limits_{\partial\Omega} v^{\tau}\cdot D^{\tau}v\cdot\nu\:dS\\ && \nonumber- c_0^2\:\int\limits_{\partial\Omega}v\cdot D\nu\cdot v\:dS. \end{eqnarray*} An application of \eqref{normal} and \eqref{normal2} yields \begin{eqnarray*} &&v^{\tau}\cdot \nabla^{\tau}(v\cdot\nu) - v^{\tau}\cdot D^{\tau}v\cdot\nu - v\cdot D\nu\cdot v = v^{\tau}\cdot D^{\tau}\nu\cdot v - v\cdot D\nu\cdot v\\ &&\qquad = -(v\cdot\nu)\nu\cdot D\nu\cdot v=0. \end{eqnarray*} Thus, with \eqref{lambdaddot3} we proved the following lemma. \begin{lemma}\label{lambdaddot4} Let $u'$ be the shape derivative of $u$ resulting from a volume preserving perturbation of $\Omega$. Then there holds \[ \frac{\ddot{\Lambda}(0)}{2}= {\cal{E}}(u'), \] where \[ {\cal{E}}(u') = \int\limits_{\Omega}\vert\Delta u'\vert^2\:dx-\Lambda(\Omega)\int\limits_{\Omega}\vert\nabla u'\vert^2\:dx. \] \end{lemma} \section{Minimization of the second domain variation } In this chapter we consider the quadratic functional \begin{eqnarray}\label{E1} {\cal{E}}(\varphi) := \int\limits_{\Omega}\vert \Delta \varphi\vert^2\:dx - \Lambda(\Omega)\int\limits_{\Omega}\vert\nabla \varphi\vert^2\:dx \end{eqnarray} for $\varphi \in H^{1,2}_0\cap H^{2,2}(\Omega)$. It will be convenient to work with an alternative representation of $ {\cal{E}}$. For $\varphi\in H^{1,2}_0\cap H^{2,2}(\Omega)$ there holds \begin{eqnarray*} {\cal{E}}(\varphi) = \int\limits_\Omega \vert D^2\varphi\vert^2 - \Lambda(\Omega)\vert\nabla\varphi\vert^2\:dx + \int\limits_{\partial\Omega}\Delta\varphi\partial_\nu\varphi - \varphi\cdot D^2\varphi\cdot\nu\:dS. \end{eqnarray*} We apply \eqref{lapbc1} and \eqref{normal}. \begin{eqnarray*} \Delta\varphi\partial_\nu\varphi - \varphi\cdot D^2\varphi\cdot\nu &=& \partial^2_\nu\varphi\,\partial_\nu\varphi + (n-1)(\partial_\nu\varphi)^2\,H_{\partial\Omega} - \varphi\cdot D^2\varphi\cdot\nu \\ &=& \nu\cdot D^2\varphi\cdot\nu \,(\nu\cdot\nabla\varphi) + (n-1)(\partial_\nu\varphi)^2\,H_{\partial\Omega} - \varphi\cdot D^2\varphi\cdot\nu \\ &=& (n-1)(\partial_\nu\varphi)^2\,H_{\partial\Omega}. \end{eqnarray*} Consequently, we get \begin{eqnarray}\label{E2} {\cal{E}}(\varphi) = \int\limits_{\Omega}\vert D^2\varphi\vert^2\:dx - \Lambda(\Omega)\int\limits_{\Omega}\vert\nabla \varphi\vert^2\:dx + (n-1)\int\limits_{\partial\Omega}(\partial_\nu \varphi)^2H_{\partial\Omega}\:dS. \end{eqnarray} \begin{remark}\label{lsc} The functional $ {\cal{E}}$ is lower semicontinuous with respect to weak convergence in $H^{1,2}_0\cap H^{2,2}(\Omega)$. \end{remark} Since $\Omega$ is optimal, we know from Lemma \ref{lambdaddot4} that \[ {\cal{E}}(\varphi)\geq 0 \] for all $\varphi$ which are shape derivatives of $u$. Recall that $\varphi$ is a shape derivative, if it solves \eqref{shape1} - \eqref{shape4} for some vector field $v$ in the class described in Chapter 1 (Lemma \ref{volpres1}). \newline \newline The following remark shows a property of shape derivatives we have not yet mentioned. \begin{remark}\label{d_nuprime} Let $\varphi$ be a shape derivative and assume that $\partial_\nu \varphi\equiv0$ in $\partial\Omega$. Then $\varphi \in H^{2,2}_0(\Omega)$ and, since $\varphi$ satisfies equation \eqref{uprimeq}, $\varphi$ is a buckling eigenfunction in $\Omega$. Thus by uniqueness of $u$ we get $\varphi = \alpha u$ for any $\alpha \in\mathbb{R}$. Then formula \eqref{Rel} yields \begin{eqnarray*} \Lambda(\Omega) = \int\limits_{\partial\Omega}\vert\Delta \varphi\vert^2\, x\cdot\nu\:dS = \alpha^2c_o^2\int\limits_{\partial\Omega}x\cdot\nu\:dS = \alpha^2\int\limits_{\partial\Omega}\vert\Delta u\vert^2\, x\cdot\nu\:dS =\alpha^2 \Lambda(\Omega). \end{eqnarray*} Thus, $\alpha^2=1$ and there holds \[ \left\vert \int\limits_\Omega \nabla u\cdot\nabla \varphi \:dx \right\vert =1. \] This is contradictory to \eqref{graduprime} and thus $\partial_{\nu} \varphi$ cannot vanish identically on $\partial\Omega$. \end{remark} This motivates the following definition. \begin{eqnarray*} {\cal{Z}}:= \left\{ \varphi \in H^{1,2}_0\cap H^{2,2}(\Omega) : \int\limits_{\partial\Omega}\partial_\nu\varphi\:dS=0,\, \int\limits_{\partial\Omega}(\partial_\nu\varphi)^2\:dS>0,\, \int\limits_{\Omega}\nabla u\cdot \nabla\varphi\:dx=0 \right\}. \end{eqnarray*} Note that ${\cal{Z}}$ contains elements which are not shape derivatives. Nevertheless we will show that \[ {\cal{E}}\big{|}_{{\cal{Z}}}\geq 0. \] The next lemma ensures that ${\cal{Z}}$ is not empty and that at least for a specific shape derivative $\cal{E}$ is equal to zero. \begin{lemma}\label{E(dku)} For each $1\leq k\leq n$ the directional derivative $\partial_ku$ satisfies $\partial_ku\in \cal{{\cal{Z}}}$. Furthermore, ${\cal{E}}(\partial_ku)=0$. \end{lemma} {\bf{Proof}} Let $1 \leq k\leq n$. Due to \eqref{eq1} and \eqref{eq2} $\partial_ku$ satisfies \begin{eqnarray} \Delta^2\partial_ku+\Lambda(\Omega)\Delta \partial_ku&=&0\quad\hbox{in}\:\Omega \label{d_ku_pde}\\ \partial_k u&=&0\quad\hbox{in}\:\partial\Omega. \notag \end{eqnarray} According to \eqref{lapbc2} there holds $\partial_\nu\partial_ku = c_0\nu_k$ on $\partial\Omega$. Hence, \[ \int\limits_{\partial\Omega}\partial_\nu\partial_ku\:dS = c_0 \int\limits_{\partial\Omega}\nu_k\:dS =0. \] In addition, we find that \[ \int\limits_\Omega \nabla u\cdot\nabla\partial_ku\;dx = \frac{1}{2}\int\limits_{\partial\Omega}\vert\nabla u\vert^2\nu_k\:dS =0. \] Following the idea of Remark \ref{d_nuprime}, we obtain that $\partial_\nu\partial_ku$ does not vanish identically on $\partial\Omega$. Thus, $\partial_ku \in \cal{{\cal{Z}}}$. Moreover, \eqref{w-eq1} and \eqref{d_ku_pde} imply \[ {\cal{E}}(\partial_ku) = \int\limits_\Omega (\Delta^2\partial_ku + \Lambda( \Omega)\Delta\partial_ku)\partial_ku\:dx + \int\limits_{\partial\Omega}\partial_k\Delta u\, \partial_\nu\partial_ku\:dS =0. \] This proves the lemma. \hfill$\square$ \newline \newline Note that each directional derivative of $u$ is a shape derivative resulting from translations of $\Omega$. \begin{theorem} \label{inffinite} The infimum of the functional $\cal{E}$ in ${\cal{Z}}$ is finite. \end{theorem} {\bf{Proof}} We argue by contradiction. Let us assume that $\inf_{\cal{{\cal{Z}}}} {\cal{E}} = -\infty$ and consider a sequence $(\hat{w}_k)_k \subset \cal{{\cal{Z}}}$ such that \[ \lim_{k\to\infty} {\cal{E}}(\hat{w}_k) = -\infty. \] For this sequence there either holds \[ \int\limits_{\partial\Omega}(\partial_\nu\hat{w}_k)^2\:dS \stackrel{k\to\infty}{\longrightarrow}0 \qquad \hbox{or} \qquad \int\limits_{\partial\Omega}(\partial_\nu\hat{w}_k)^2\:dS \stackrel{k\to\infty}{\not\longrightarrow}0. \] If the second case holds true, we normalize the sequence $(\hat{w}_k)_k$ such that $\vert\vert\partial_\nu\hat{w}_k\vert\vert_{L^2(\partial\Omega)}=1$. Hence, in either case, for each $\hat{w}_k$ there holds $\|\partial_\nu\hat{w}_k\|_{L^2(\partial\Omega)}\leq1$. Thus, \eqref{H_dOmega} gives \begin{eqnarray*} \left\vert\, \int\limits_{\partial\Omega} H_{\partial\Omega}\, (\partial_\nu\hat{w}_k)^2\:dS\right\vert \leq \max_{\partial\Omega}\vert H_{\partial\Omega}\vert < \infty. \end{eqnarray*} We use \eqref{E2} and obtain \begin{eqnarray}\label{estE1} {\cal{E}}(\hat{w}_k) \geq -\Lambda(0) \int\limits_\Omega \vert\nabla\hat{w}_k\vert\:dx - (n-1) \max_{\partial\Omega}\vert H_{\partial\Omega}\vert. \end{eqnarray} The assumption $\lim_{k\to \infty}{\cal{E}}(\hat{w_k}) = -\infty$ implies \begin{eqnarray*} \int\limits_\Omega \vert\nabla\hat{w}_k\vert^2\:dx \stackrel{k\to\infty}{\longrightarrow} \infty. \end{eqnarray*} We define \begin{eqnarray*} w_k := \frac{1}{\vert\vert\nabla\hat{w}_k\vert\vert_{L^2(\Omega)}}\hat{w}_k. \end{eqnarray*} Then there holds \begin{eqnarray}\label{w_k} \vert\vert\nabla w_k\vert\vert_{L^2(\Omega)} = 1 \quad \hbox{and}\quad \int\limits_{\partial\Omega}(\partial_\nu w_k)^2\:dS\;\stackrel{k\to\infty}{\longrightarrow}\; 0. \end{eqnarray} Moreover, for each $k\in\mathbb{N}$ estimate \eqref{estE1} implies \[ {\cal{E}}(w_k) \geq -\Lambda(0) - C \] and the infimum of $\cal{E}$ in $M := \{w_k : k\in\mathbb{N}\}$ is finite. Therefore, we can choose a subsequence of $(w_k)_k$, denote by $(w_k)_k$ as well, such that \[ \lim_{k\to\infty} {\cal{E}}(w_k) = \inf_M \cal{E}. \] Now Poincar\'e's inequality and the previous estimates imply \begin{eqnarray*} \vert\vert w_k\vert\vert^2_{H^{2,2}(\Omega)} &=& \int\limits_\Omega \vert D^2w_k\vert^2 + \vert\nabla w_k\vert^2 + w_k^2\:dx \\ &\leq& {\cal{E}}(w_k) + C \int\limits_\Omega\vert\nabla w_k\vert^2\:dx + (n-1)\int\limits_{\partial\Omega}\vert H_{\partial\Omega}\vert(\partial_\nu w_k)^2\:dS \\ &\leq& C. \end{eqnarray*} Thus, the sequence $(w_k)_k$ is uniformly bounded in $H^{2,2}(\Omega)$ and there exists a $w \in H^{2,2}(\Omega)$ such that $(w_k)_k$ weakly converges to $w$. In view of \eqref{w_k}, the limit function $w$ satisfies $ \vert\vert\nabla w\vert\vert_{L^2(\Omega)} = 1 $ and $\partial_\nu w =0$ on $\partial\Omega$. Since $w_k =0$ in $\partial\Omega$ for each $k \in \mathbb{N}$, we conclude that $w \in H^{2,2}_0(\Omega)$. \newline Now let us recall that ${\cal{E}}(\hat{w}_k)$ converges to $-\infty$. Thus there exists a $k_0 \in \mathbb{N}$ such that \begin{eqnarray*} {\cal{E}}(w_k) = \frac{1}{\vert\vert\nabla\hat{w}_k\vert\vert_{L^2(\Omega)}} \,{\cal{E}}(\hat{w}_k) < 0 \end{eqnarray*} for all $k\geq k_0$. Since the functional $\cal{E}$ is lower semicontinous with respect to weak convergence in $H^{2,2}(\Omega)$, we find that ${\cal{E}}(w)<0$. According to the definiton of $\cal{E}$ in \eqref{E1}, this immediately leads to \begin{eqnarray*} \frac{\int\limits_{\Omega}\vert\Delta w\vert^2\:dx}{\int\limits_\Omega\vert\nabla w\vert^2\:dx} < \Lambda(\Omega). \end{eqnarray*} Since $w \in H^{2,2}_0(\Omega)$ this is contradictory to the minimum property of $\Lambda(\Omega)$. \hfill $\square$ \newline \newline As mentioned in the previous proof, a minimizing sequence $(\varphi_k)_k\subset {\cal{Z}}$ satisfies one of the following two conditions \begin{itemize} \item [i)] $ \int\limits_{\partial\Omega}(\partial_\nu \varphi_k)^2\:dS \stackrel{k\to\infty}{\longrightarrow} 0$ \item[ii)]$\int\limits_{\partial\Omega}(\partial_\nu\varphi_k)^2\:dS \stackrel{k\to\infty}{\not\longrightarrow}0$. \end{itemize} In the sequel, we show that the case i) implies that the minimizing sequence $(\varphi_k)_k$ converges to zero. For this purpose, let $(\varphi_k)_k\subset {\cal{Z}}$ be a minimizing sequence which satisfies condition i). From \eqref{E2} we get \begin{eqnarray*} \vert\vert\varphi_k\vert\vert^2_{H^{2,2}(\Omega)} \leq C, \end{eqnarray*} and thus there exists a $\varphi\in H^{2,2}(\Omega)$ such that $\varphi_k$ weakly converges to $\varphi$ in $H^{2,2}(\Omega)$ and ${\cal{E}}(\varphi)= \inf_{\cal{Z}} {\cal{E}}$. Furthermore, condition i) implies $\varphi \in H^{2,2}_0(\Omega)$. From Lemma \ref{E(dku)} we obtain \begin{eqnarray*} \inf_{\cal{Z}}{\cal{E}}={\cal{E}}(\varphi) \leq {\cal{E}}(\partial_lu) =0 \quad \hbox{for any } 1\leq l \leq n. \end{eqnarray*} Hence, \begin{eqnarray*} \frac{\int\limits_\Omega\vert\Delta\varphi\vert^2\:dx}{\int\limits_\Omega \vert\nabla\varphi\vert^2\:dx} \leq \Lambda(\Omega). \end{eqnarray*} Thus $\varphi$ is necessarily an eigenfunction corresponding to $\Lambda(\Omega)$. Since the eigenvalue is simple we have $\varphi = \alpha\,u$ for $\alpha \in \mathbb{R}$. Now let us recall that $\varphi_k \in {\cal{Z}}$ and, therefore, \begin{eqnarray*} \alpha = \alpha\,\int\limits_\Omega \vert\nabla u\vert^2\:dx = \int\limits_\Omega \nabla u\cdot\nabla\varphi\:dx = \lim_{k\to\infty} \int\limits_\Omega \nabla u\cdot\nabla\varphi_k\:dx =0. \end{eqnarray*} Consequently, $\alpha =0$ and $\varphi \equiv 0$ in $\Omega$. Hence $\varphi \notin {\cal{Z}}$. Since we are interested to find minimizers of ${\cal{E}}$ in ${\cal{Z}}$, we restrict ourselves to minimzing sequences which satisfy the condition ii). Thus we consider the functional \begin{eqnarray*} \tilde{{\cal{E}}}(\varphi) := \frac{{\cal{E}}(\varphi)}{\int\limits_{\partial\Omega}(\partial_\nu\varphi)^2\:dS}, \end{eqnarray*} where $\varphi \in {\cal{Z}}$ and we set $\tilde{{\cal{E}}}=\infty$ if $\int\limits_{\partial\Omega}(\partial_\nu\varphi_k)^2\:dS=0$. \begin{remark}\label{gradmise} Suppose $(\varphi_k)_k \subset {\cal{Z}}$ is a minimzing sequence for $\tilde{{\cal{E}}}$ in ${\cal{Z}}$. Then there exists a constant $C>0$ such that $\vert\vert\nabla\varphi_k\vert\vert_{L^2(\Omega)}\leq C$ for every $k\in\mathbb{N}$. This follows by contradiction. Otherwise we may assume that $\vert\vert\nabla\varphi_k\vert\vert_{L^2(\Omega)}$ tends to infinity as $k\to\infty$, we define $\varphi^\ast_k := \vert\vert\nabla\varphi_k\vert\vert_{L^2(\Omega)}^{-1}\varphi_k$. Then $(\varphi^\ast_k)_k$ is uniformly bounded in $H^{2,2}(\Omega)$ and \begin{eqnarray*} \int\limits_{\partial\Omega}(\partial_\nu \varphi^\ast_k)^2\:dS \stackrel{k\to\infty}{\longrightarrow} 0. \end{eqnarray*} Thus, $(\varphi^\ast_k)_k$ converges weakly to a function $\varphi \in H^{2,2}_0(\Omega)$ and for every $1 \leq l\leq n$ there holds \begin{eqnarray*} \inf_{\cal{Z}} \tilde{{\cal{E}}} = \tilde{{\cal{E}}}(\varphi) \leq \tilde{{\cal{E}}}(\partial_lu) =0. \end{eqnarray*} As the previous considerations have shown, this implies $\varphi \equiv 0$ in $\Omega$. Thus, our assumption cannot be true. \end{remark} We now consider a minimizing sequence $(\varphi_k)_k \subset {\cal{Z}}$ which satisfies \begin{eqnarray}\label{normmise} \int\limits_{\partial\Omega}(\partial_\nu\varphi_k)^2\:dS=1 \end{eqnarray} for all $k\in\mathbb{N}$. As before we obtain the inequality \begin{eqnarray*} \vert\vert\varphi_k\vert\vert^2_{H^{2,2}(\Omega)} \leq {\cal{E}}(\varphi_k) + C\int\limits_\Omega\vert \nabla\varphi_k\vert^2\:dx . \end{eqnarray*} Thus, $(\varphi_k)_k$ is uniformly bounded in $H^{2,2}(\Omega)$ and $\varphi_k$ converges weakly to a $\varphi^\ast \in H^{2,2}(\Omega)$. We find that $\varphi^\ast \in {\cal{Z}}$ and ${\cal{E}}(\varphi^\ast) = \inf_{\cal{Z}}E$. In addition, there holds \[ \int\limits_{\partial\Omega}(\partial_\nu\varphi^\ast)^2\:dS = 1. \] Hence, $\varphi^\ast$ minimizes $\tilde{{\cal{E}}}$ in ${\cal{Z}}$. Suppose $\theta \in {\cal{Z}}$, then the minimality of $\varphi^\ast$ implies \[ \frac{d}{dt}\frac{{\cal{E}}(\varphi^\ast +t \theta)}{\int\limits_{\partial\Omega}(\partial_\nu (\varphi^\ast + t \theta))^2\:dS)}\bigg\vert_{t=0} =0 \] and we obtain \[ \int\limits_\Omega [\Delta^2\varphi^\ast + \Lambda(\Omega)\Delta\varphi]\,\theta\:dx - \int\limits_{\partial\Omega}[\Delta\varphi^\ast + \rho\,\partial_\nu\varphi^\ast]\,\partial_\nu\theta\:dS =0. \] Since $\theta\in {\cal{Z}}$ was chosen arbitrary, $\varphi^\ast$ satisfies the Euler-Lagrange equalities \begin{eqnarray*} \Delta^2\varphi^\ast + \Lambda(\Omega)\Delta\varphi^\ast &=& 0 \quad \hbox{in} \quad \Omega \\ \Delta\varphi^\ast + \rho\partial_\nu\varphi^\ast &=& const. \quad \hbox{in} \quad \partial\Omega, \end{eqnarray*} where $\rho := \min_{\cal{Z}}\tilde{{\cal{E}}}$. The following theorem collects the previous results. \begin{theorem} There exists a function $\varphi^\ast \in {\cal{Z}}$ such that $\tilde{{\cal{E}}}(\varphi^\ast)=\min_{\cal{Z}}\tilde{{\cal{E}}}$. Furthermore, any minimizer $\varphi^\ast \in {\cal{Z}}$ satisfies \begin{eqnarray} \label{EL1} \Delta^2\varphi^\ast + \Lambda(\Omega)\Delta\varphi^\ast &=& 0 \quad \hbox{in} \quad \Omega \\ \label{EL2} \Delta\varphi^\ast + \rho\,\partial_\nu\varphi^\ast &=& const. \quad \hbox{in} \quad \partial\Omega\\ \notag\varphi^\ast &=&0 \quad \hbox{in} \quad \partial\Omega, \end{eqnarray} where $\rho := \min_{\cal{Z}}\tilde{{\cal{E}}}$. \end{theorem} The next theorem shows that in fact $\rho =0$. \begin{theorem}\label{minE=0} Suppose $\varphi^{\ast}\in {\cal{Z}}$ is a minimizer of $\tilde{{\cal{E}}}$. Then there holds $\tilde{{\cal{E}}}(\varphi^\ast) =0$. In particular, ${\cal{E}}\geq 0$ in $\cal{Z}$. \end{theorem} {\bf{Proof}} Let $\varphi^\ast \in {\cal{Z}}$ be a minimizer of $\tilde{{\cal{E}}}$. Since $\varphi^\ast$ satisfies equation \eqref{EL1} and $\partial\Omega$ is smooth, $\varphi^\ast$ is a smooth function on $\overline{\Omega}$. Hence, we may define a volume preserving perturbation $\Phi_t$ of $\Omega$ such that \begin{eqnarray*} \partial_\nu u'(x) = \partial_\nu \varphi^\ast(x) \quad \hbox{for} \quad x \in \partial\Omega. \end{eqnarray*} Note that this can be achieved by setting $v = c_0^{-1}\nabla \varphi^\ast$ in $\partial\Omega$. In this way, each minimizer $\varphi^{\ast}$ implies the existence of vector fields $v$ and $w$ in the sense of Section \ref{DomVar}. We define $\psi := u' - \varphi^\ast$, then $\psi \in H^{2,2}_0(\Omega)$ and \begin{eqnarray*} \Delta^2\psi + \Lambda(\Omega)\Delta\psi =0 \quad \hbox{in} \quad \Omega. \end{eqnarray*} The uniqueness of $u$ implies $\psi = \alpha \,u$ for an $\alpha \in \mathbb{R}$. Since $\varphi^\ast\in {\cal{Z}}$, equation \eqref{graduprime} yields \begin{align*} 0 = \int\limits_\Omega \nabla u\cdot\nabla u'\:dx - \int\limits_\Omega \nabla u\cdot\nabla \varphi^\ast\:dx = \int\limits_\Omega \nabla u\cdot\nabla \psi\:dx = \alpha. \end{align*} Consequently, $u' \equiv \varphi^\ast$. Thus $\varphi^\ast$ is a shape derivative. Since $\Omega$ is optimal $\tilde{{\cal{E}}}(\varphi^\ast)\geq 0$. Finally we apply Lemma \ref{E(dku)}. This gives \begin{eqnarray*} 0\leq \tilde{{\cal{E}}}(\varphi^\ast) = \min_{\cal{Z}} \tilde{{\cal{E}}} \leq \tilde{{\cal{E}}}(\partial_ku) =0. \end{eqnarray*} \hfill $\square$ \section{The optimal domain is a ball} We will use an inequaltiy due to L.E. Payne to show that the optimal domain $\Omega$ is a ball. Payne's inequality (see \cite{P55}) states that for each domain $G$ there holds \begin{equation*} \lambda_2(G) \leq \Lambda(G) \end{equation*} and equality only holds if and only if $G$ is a ball. Thereby $\lambda_2$ denotes the second Dirichlet eigenfunction of the Laplacian. In the sequel, we construct a suitable function $\psi \in {\cal{Z}}$ such that the condition ${\cal{E}}(\psi)\geq 0$ (due to Theorem \ref{minE=0}) will imply that the optimal domain $\Omega$ is a ball. For this purpose, we denote by $u_1$ and $u_2$ the first and the second Dirichlet eigenfunction for the Laplacian in $\Omega$. Thus, for $1\leq k\leq 2$ there holds \begin{eqnarray*} \Delta u_k + \lambda_k(\Omega) u_k &=&0 \quad \mbox{in }\Omega \\ u_k &=&0 \quad \mbox{in }\partial\Omega, \end{eqnarray*} where $\lambda_k(\Omega)$ is the $k$-th Dirichlet eigenvalue for the Laplacian in $\Omega$. Note that $0< \lambda_1(\Omega)<\lambda_2(\Omega)$. For the sake of brevity, we will write $ \lambda_k$ instead of $\lambda_k(\Omega)$ and $\Lambda$ instead of $\Lambda(\Omega)$. In addition, we assume $\vert\vert u_k\vert\vert_{L^2(\Omega)}=1$ and \begin{eqnarray*} \int\limits_\Omega u_1u_2\:dx =0. \end{eqnarray*} Without loss of generality, we may assume that \begin{eqnarray*} \int\limits_\Omega u_1\:dx >0 \quad \mbox{and} \quad \int\limits_\Omega u_2\:dx \leq 0. \end{eqnarray*} Consequently, there exists a $t \in (0,1]$ such that \begin{eqnarray}\label{meanval} \int\limits_\Omega (1-t)\:\lambda_1\:u_1 + t\:\lambda_2\: u_2 \:dx =0. \end{eqnarray} This fixes $t$. Next we define \begin{eqnarray*} \psi(x):= (1-t)\:u_1(x) + t\:u_2(x) + c\:u(x)\qquad\hbox{for}\:x\in\overline{\Omega}, \end{eqnarray*} where $u$ is the first buckling eigenfunction in $\Omega$. The constant $c$ is given by \begin{eqnarray*} c := -\frac{1}{\Lambda}\int\limits_\Omega (1-t)\lambda_1\nabla u.\nabla u_1 + t\lambda_2\nabla u.\nabla u_2\: dx. \end{eqnarray*} In a first step we show that $\psi \in {\cal{Z}}$. Note that $\psi \in H^{1,2}_0\cap H^{2,2}(\Omega)$. Moreover the definition of $\psi$, the fact that $\partial_{\nu}u=0$ on $\partial\Omega$, the equations for $u_1$ and $u_2$, and \eqref{meanval} imply \begin{eqnarray*} \int\limits_{\partial\Omega} \partial_\nu \psi \:dS = \int\limits_\Omega (1-t)\:\Delta u_1 + t\:\Delta u_2\:dx = -\int\limits_\Omega (1-t)\lambda_1u_1 + t\lambda_2u_2\:dx =0. \end{eqnarray*} By the unique continuation principle $\partial_\nu \psi $ does not vanish identically in $\partial\Omega$. Thus, to show that $\psi \in {\cal{Z}}$, it remains to prove that \begin{equation}\label{cond_3} \int\limits_\Omega \nabla u.\nabla \psi \:dx =0. \end{equation} We recall that $\Delta u = c_0$ in $\partial\Omega$. Hence \begin{align*} 0 &= \int\limits_\Omega (\Delta^2u + \Lambda \Delta u)\psi \:dx = \int\limits_\Omega\Delta u \Delta \psi \:dx - \Lambda\int\limits_{\Omega}\nabla u.\nabla \psi \:dx \\ &= -\int\limits_\Omega [(1-t)\lambda_1u_1 + t\lambda_2u_2]\Delta u\:dx + c\int\limits_\Omega\vert\Delta u\vert^2\:dx- \Lambda\int\limits_\Omega\nabla u.\nabla \psi \:dx. \end{align*} Since $\vert\vert \nabla u\vert\vert_{L^2(\Omega)}=1$, the second integral is equal to $\Lambda$. Thus, the definition of $c$ implies \eqref{cond_3}. Note that $\psi$ is not a shape derivative since it fails to satisfy \eqref{uprimeq} - unless $t=1$ and $\Omega$ equals a ball. However, $\psi \in {\cal{Z}}$ and, according to Theorem \ref{minE=0}, there holds $\tilde{{\cal{E}}}(\psi)\geq 0$. Consequently, ${\cal{E}}(\psi)\geq 0$. Thus \begin{align*} {\cal{E}}(\psi) &= \int\limits_\Omega \vert\Delta \psi\vert^2-\Lambda \vert\nabla \psi \vert^2\:dx \\ & = (1-t)^2\lambda_1 (\lambda_1-\Lambda) + t^2\lambda_2(\lambda _2- \Lambda) + 2\,c\,c_0 \int\limits_\Omega (1-t)\lambda_1u_1 + t\lambda_2u_2\:dx \\ &\stackrel{\eqref{meanval}}{=} (1-t)^2\lambda_1 (\lambda_1-\Lambda) + t^2\lambda_2(\lambda _2- \Lambda) \geq 0. \end{align*} Since $\lambda_1-\Lambda < 0$ and $\lambda_2-\Lambda\leq 0$, both summands in ${\cal{E}}(\psi)$ have to vanish. Consequently $t=1$ and $\lambda_2(\Omega) = \Lambda(\Omega)$. Payne's inequality implies that $\Omega$ is a ball. This proves the main theorem of the paper. \begin{theorem} Let $\Omega$ be a bounded, smooth and simply connected domain in $\mathbb{R}^n$, which minimizes the first buckling eigenvalue among all bounded, smooth and simply connected domains in $\mathbb{R}^n$ with given measure. Then $\Omega$ is a ball. \end{theorem}

https://arxiv.org/abs/2301.05507

Correlation-Based And-Operations Can Be Copulas: A Proof

In many practical situations, we know the probabilities $a$ and $b$ of two events $A$ and $B$, and we want to estimate the joint probability ${\rm Prob}(A\,\&\,B)$. The algorithm that estimates the joint probability based on the known values $a$ and $b$ is called an and-operation. An important case when such a reconstruction is possible is when we know the correlation between $A$ and $B$; we call the resulting and-operation correlation-based. On the other hand, in statistics, there is a widely used class of and-operations known as copulas. Empirical evidence seems to indicate that the correlation-based and-operation derived inthis https URLis a copula, but until now, no proof of this statement was available. In this paper, we provide such a proof.

\section{Formulation of the problem} \noindent{\bf Correlation-based ``and"-operation.} In many practical situations, we know the probabilities $a$ and $b$ of two events $A$ and $B$, and we need to estimate the joint probability ${\rm Prob}(A\,\&\,B)$. An algorithm $f_\&(a,b)$ that transforms the known values $a$ and $b$ into such an estimate is usually called an {\it and-operation}. One important case when such an estimate is possible is when, in addition to the probabilities $a$ and $b$, we also know the correlation $\rho$ between the corresponding two random events. It is known (see, e.g., \cite{Lucas 1995,Miralles 2022}) that in this case, we can uniquely determine the probability of ${\rm Prob}(A\,\&\,B)$ as $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}.\eqno{(1)}$$ While this formula is true whenever the correlation is known, this formula does not lead to an everywhere defined and-operation. For example, for $a=b=0.1$ and $\rho=-1$, this formula leads to a meaningless negative probability $$0.1\cdot 0.1+(-1)\cdot \sqrt{0.1\cdot 0.9\cdot 0.1\cdot 0.9}=0.01-0.09=-0.08<0.$$ To avoid such meaningless estimates, we need to take into account that the joint probability ${\rm Prob}(A\,\&\,B)$ must satisfy Fr\'echet inequalities (see, e.g., \cite{Frechet 1935}): $$\max(a+b-1,0)\le {\rm Prob}(A\,\&\,B)\le\min(a,b).\eqno{(2)}$$ So, if an expert claims to know the correlation $\rho$ and the estimate for ${\rm Prob}(A\,\&\,B)$ based on this value $\rho$ is smaller than the lower bound $\max(a+b-1,0)$ -- which cannot be -- a reasonable idea is to take the closest possible value of the joint probability, i.e., the value $\max(a+b-1,0)$. Similarly, if the estimate for ${\rm Prob}(A\,\&\,B)$ based on the expert-provided value $\rho$ is larger than the upper bound $\min(a,b)$ -- which also cannot be -- a reasonable idea is to take the closest possible value of the joint probability, i.e., the value $\min(a,b)$. Thus, we arrive at the following and-operation -- which we will call {\it correlation-based and-operation}: $$f_\rho(a,b)=T_{a,b}\left(a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\right),\eqno{(3)}$$ where $$T_{a,b}(c)=\max(a+b-1,0)\mbox{ if } c<\max(a+b-1,0);$$ $$T_{a,b}(c)=c \mbox{ if } \max(a+b-1,0)\le c\le\min(a,b);\mbox{ and}\eqno{(4)}$$ $$T_{a,b}(c)=\min(a,b)\mbox{ if } \min(a,b)<c.$$ \medskip \noindent{\bf Question: is this and-operation a copula?} In probability theory, there is a known class of and-operations known as {\it copulas} (see, e.g., \cite{Nelsen 2007,Schweizer 2011}). These are functions $C(a,b)$ for which, for some random 2-D vector $(X,Y)$, the joint cumulative distribution function $F_{XY}(x,y)\stackrel{\rm def}{=}{\rm Prob}(X\le x\,\&\,Y\le y)$ has the form $F_{XY}(x,y)=C(F_X(x),F_Y(y))$, where $F_X(x)\stackrel{\rm def}{=}{\rm Prob}(X\le x)$ and $F_Y(y)\stackrel{\rm def}{=}{\rm Prob}(Y\le y)$ are known as {\it marginals}. One important aspect of (3)-(4) is that these formulas can be expressed as a copula (2-copula) family as described in \cite{Miralles 2022}, allowing us to operate not only with precise probabilities, but also with interval probabilities and probability boxes. A 2-copula must satisfy the following properties: \begin{enumerate} \item Grounded: $C(0,b)=C(a,0)=0$ \item Uniform margins: $C(a,1)=a;C(1,b)=b$ \item 2-increasing: $C(\overline a,\overline b)+C(\underline a,\underline b)-C(\overline a,\underline b)-C(\underline a,\overline b)\ge 0$ for all $\underline a<\overline a$ and $\underline b<\overline b$ \end{enumerate} It is easy to see that (3)-(4) satisfies the two first properties. In \cite{Miralles 2022} the third property was checked for a dense set of tuples $(\underline a,\overline a,\underline b,\overline b,\rho)$, and for all these tuples, the inequality was satisfied. However, at that moment, we could not prove that the correlation-based and-operation is indeed a 2-copula. In this paper we provide the missing proof. \section{Main result} \noindent{\bf Proposition.} {\it For every $\rho\in[-1,1]$, the correlation and-operation $f_\rho(a,b)$ described by the formulas (3)-(4) is a copula.} \medskip \noindent{\bf Proof.} \medskip \noindent $1^\circ$. It is known that the desired inequality has the following property -- if we represent a box $[\underline a,\overline a]\times[\underline b,\overline b]$ as a union of several sub-boxes, then the left-hand side of the desired inequality is equal to the sum of the left-hand sides corresponding to sub-boxes. Indeed, as one can easily check, there is the following {\it additivity} property: for each box consisting of several sub-boxes, the left-hand side of the inequality (4a) that corresponds to the larger box is equal to the sum of expressions (4a) corresponding to sub-boxes. Thus, if the expressions corresponding to sub-boxes are non-negative, then the expression (4a) corresponding to the larger box is also non-negative. In general, the and-operation described by the formula (4) has three different expressions. So, to prove that the expression (4a) corresponding to this expression is also non-negative, we need to consider cases when at different vertices of the box, we may have different expressions. Good news is that every box whose vertices are described by different expressions can be represented as the union of sub-boxes in which: \begin{itemize} \item either all vertices are described by the same expression \item or two vertices are on the boundary between the areas of different expressions. \end{itemize} This is easy to see visually: the following box, in which the slanted line represents the boundary between the areas \medskip \begin{center} \begin{picture}(150,50) \put(0,0){\line(1,0){150}} \put(0,50){\line(1,0){150}} \put(0,0){\line(0,1){50}} \put(150,0){\line(0,1){50}} \put(50,0){\line(1,1){50}} \end{picture} \end{center} \medskip \noindent can be represented as the union of sub-boxes with the desired property: \medskip \begin{center} \begin{picture}(150,50) \put(0,0){\line(1,0){150}} \put(0,50){\line(1,0){150}} \put(0,0){\line(0,1){50}} \put(150,0){\line(0,1){50}} \put(50,0){\line(1,1){50}} \put(50,0){\line(0,1){50}} \put(100,0){\line(0,1){50}} \end{picture} \end{center} \medskip Thus, to prove that our and-operation is a copula, it is sufficient to consider only boxes of the following type: \begin{itemize} \item boxes for which all four vertices belong to the same area, and \item boxes for which two vertices belong to the boundary between two areas. \end{itemize} The functions $\max(a+b-1,0)$ and $\min(a,b)$ are known to be copulas, so if all four vertices belong to one of these areas, then the desired inequality (4a) is satisfied. So, it is sufficient to consider: \begin{itemize} \item boxes for which all four vertices belong to the new area, in which the and-operation is described by the expression (1); we will consider such boxes in Parts 2--4 of this proof, and \item boxes for which two vertices belong to the boundary between two areas; these boxes will be considered in the following Parts of the proof. \end{itemize} \medskip \noindent $2^\circ$. Let us start by considering boxes for which all four vertices belongs to the area in which the and-operation is described by the formula (1). \medskip It is known \cite{Durante 2010} -- and it is easy to prove by considering infinitesimal differences $\overline x-\underline x$ and $\overline y-\underline y$ -- that for smooth functions, the desired inequality is equivalent to the fact that the partial derivative $$\frac{\partial C}{\partial a}$$ is non-decreasing in $b$, i.e., equivalently, that the mixed derivative is non-negative: $$d\stackrel{\rm def}{=}\frac{\partial^2 C}{\partial a\,\partial b}\ge 0.$$ Thus, to prove that $f_\rho(a,b)$ is a copula, it is sufficient to prove that its mixed derivative is non-negative everywhere where the new formula is applied. Indeed, at the points where the formula (1) is applied, the derivative of $f_\rho(a,b)$ with respect to $a$ has the has the form $$\frac{\partial f_\rho}{\partial a}=b+\rho\cdot \frac{1-2\cdot a}{2\cdot\sqrt{a\cdot (1-a)}}\cdot \sqrt{b\cdot (1-b)},\eqno{(4{\rm b})}$$ and thus, the mixed derivative has the following form: $$d=\frac{\partial}{\partial b}\left(\frac{\partial f_\rho}{\partial a}\right)=1+\rho\cdot \frac{(1-2\cdot a)\cdot (1-2\cdot b)} {4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}.\eqno{(5)}$$ Since the expression (1) does not change if we swap $a$ and $b$, it is sufficient to consider the case when $a\le b$. When $\rho=0$, we get a known copula $f_0(a,b)=a\cdot b$. So, it is sufficient to consider cases when $\rho\ne 0$. This can happen when $\rho>0$ and when $\rho<0$. Let us consider these cases one by one. \medskip \noindent $3^\circ$. Let us first consider the case when $\rho>0$. \medskip In this case, since $a\le b$, we have $\min(a,b)=a$ and thus, the condition (4) takes the form $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le a,\eqno{(6)}$$ i.e., equivalently, $$\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le a-a\cdot b=a\cdot (1-b)\eqno{(7)}$$ and thus, $$\rho\le \frac{a\cdot (1-b)}{\sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}=\frac{\sqrt{a\cdot (1-b)}}{\sqrt{(1-a)\cdot b}}.\eqno{(8)}$$ For all such $\rho$, we need to prove that the expression (5) is non-negative. When both $a$ and $b$ are larger than 0.5 or both are smaller than 0.5, the differences $1-2a$ and $1-2b$ have the same sign and thus, their product is non-negative and the expression (5) is non-negative. So, the only case when we need to check that $d\ge 0$ is when one of the two values $a$ and $b$ is smaller than 0.5 and another one is larger than 0.5. Since $a\le 0.5$, this means that $a<0.5<b$. In this case, the condition $d\ge 0$ takes the form $$1-\rho\cdot \frac{(1-2\cdot a)\cdot (2\cdot b-1)}{4\cdot\sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}\ge 0,\eqno{(9)}$$ i.e., equivalently, $$\rho\cdot \frac{(1-2\cdot a)\cdot (2\cdot b-1)}{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}\le 1,\eqno{(10)}$$ and $$\rho\le\frac{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}{(1-2\cdot a)\cdot (2\cdot b-1)}.\eqno{(11)}$$ So, to prove that we always have $d\ge 0$, we need to prove that every $\rho$ that satisfies the inequality (8) also satisfies the inequality (11). Clearly, if some value $\rho$ satisfies the inequality (11), then every smaller value $\rho$ also satisfies this inequality. Thus, to prove the desired implication, it is sufficient to check that the inequality (11) is satisfied for the largest possible value $\rho$ that satisfies the inequality (8), i.e., for the value $\rho$ which is equal to the right-hand side of the inequality (8). For this $\rho$, the desired inequality (11) takes the form $$\frac{\sqrt{a\cdot (1-b)}}{\sqrt{(1-a)\cdot b}}\le \frac{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}{(1-2\cdot a)\cdot (2\cdot b-1)}.\eqno{(12)}$$ Dividing both sides by $\sqrt{a\cdot (1-b)}$, we get an equivalent inequality $$\frac{1}{\sqrt{(1-a)\cdot b}}\le \frac{4\cdot \sqrt{(1-a)\cdot b}}{(1-2\cdot a)\cdot (2\cdot b-1)}.\eqno{(13)}$$ Multiplying both sides by both denominators, we get the following equivalent inequality: $$(1-2\cdot a)\cdot(2\cdot b-1)\le 4\cdot (1-a)\cdot b.\eqno{(14)}$$ If we open parentheses, this inequality takes the equivalent form $$2\cdot b-4\cdot a\cdot b-1+2\cdot a\le 4\cdot b-4\cdot a\cdot b,\eqno{(15)}$$ i.e., by adding $4\cdot a\cdot b-2\cdot b$ to both sides, the form $$-1+2\cdot a\le 2\cdot b.\eqno{(16)}$$ We are considering the case when $a\le b$ -- since, as we have mentioned earlier, it is sufficient to only consider this case. Thus, the equivalent inequality (12) is also true and hence, for the case when $\rho>0$, we indeed have $d\ge 0$. \medskip \noindent $4^\circ$. To complete the proof, it is now sufficient to consider the case when $\rho<0$. \medskip In this case, if one of the values $a$ and $b$ is smaller than 0.5 and another one is larger than 0.5, then the differences $1-2\cdot a$ and $1-2\cdot b$ have different signs, so the right-hand side of the expression (5) for $d$ is larger than 1 and thus, non-negative. Thus, it is sufficient to consider the cases when: \begin{itemize} \item either both $a$ and $b$ are larger than 0.5 \item or both $a$ and $b$ are smaller than 0.5. \end{itemize} Let us consider these two cases one by one. \medskip \noindent $4.1^\circ$. Let us first consider the case when $a>0.5$ and $b>0.5$. \medskip In this case, $a+b-1>0$, so the inequality (4) takes the form $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\ge a+b-1,\eqno{(17)}$$ i.e., equivalently, that $$|\rho|\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le a\cdot b-a-b+1=(1-a)\cdot (1-b),\eqno{(18)}$$ or that $$|\rho|\le\frac{(1-a)\cdot (1-b)}{\sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}=\frac{\sqrt{(1-a)\cdot(1-b)}}{\sqrt{a\cdot b}}.\eqno{(19)}$$ In this case, the condition $d\ge 0$ that the value (5) is non-negative takes the form $$1-|\rho|\cdot \frac{(2\cdot a-1)\cdot (2\cdot b-1)}{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}},\eqno{(20)}$$ i.e., equivalently, $$|\rho|\cdot \frac{(2\cdot a-1)\cdot (2\cdot b-1)}{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}\le 1\eqno{(21)}$$ and $$|\rho|\le \frac{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}{(2\cdot a-1)\cdot (2\cdot b-1)}.\eqno{(22)}$$ Similarly to the case when $\rho>0$, to check that all values $|\rho|$ satisfying the inequality (19) also satisfies the inequality (22), it is sufficient to check that the largest possible value $|\rho|$ satisfying the inequality (19) satisfies the inequality (22), i.e., that $$\frac{\sqrt{(1-a)\cdot(1-b)}}{\sqrt{a\cdot b}}\le \frac{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}} {(2\cdot a-1)\cdot (2\cdot b-1)}.\eqno{(23)}$$ If we divide both sides by $\sqrt{(1-a)\cdot (1-b)}$, we get the following equivalent inequality $$\frac{1}{\sqrt{a\cdot b}}\le \frac{4\cdot \sqrt{a\cdot b}}{(2\cdot a-1)\cdot (2\cdot b-1)}.\eqno{(24)}$$ Multiplying both sides by both denominators, we get the following equivalent inequality $$(2\cdot a-1)\cdot (2\cdot b-1)\le 4\cdot a\cdot b.\eqno{(25)}$$ Opening parentheses, we get $$4\cdot a\cdot b-2\cdot a-2\cdot b+1\le 4\cdot a\cdot b.\eqno{(26)}$$ Adding $2\cdot a+2\cdot b-4\cdot a\cdot b$ to both sides, we get an equivalent inequality $$1\le 2\cdot a+2\cdot b,\eqno{(27)}$$ which is true since we consider the case when $a+b>1$. So, in this case, we indeed have $d\ge 0$. \medskip \noindent $4.2^\circ$. Let us now consider the case when $a<0.5$ and $b<0.5$. \medskip In this case, $a+b-1<0$, so the inequality (4) takes the form $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\ge 0,\eqno{(28)}$$ i.e., equivalently, that $$|\rho|\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le a\cdot b,\eqno{(29)}$$ or that $$|\rho|\le\frac{a\cdot b}{\sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}=\frac{\sqrt{a\cdot b}}{\sqrt{(1-a)\cdot(1-b)}}.\eqno{(30)}$$ In this case, the condition $d\ge 0$ that the value (5) is non-negative takes the form $$1-|\rho|\cdot \frac{(1-2\cdot a)\cdot (1-2\cdot b)}{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}},\eqno{(31)}$$ i.e., equivalently, $$|\rho|\cdot \frac{(1-2\cdot a)\cdot (1-2\cdot b)}{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}\le 1\eqno{(32)}$$ and $$|\rho|\le \frac{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}{(1-2\cdot a)\cdot (1-2\cdot b)}.\eqno{(33)}$$ Similarly to the cases when $\rho>0$ and when $a+b>1$, to check that all values $|\rho|$ satisfying the inequality (30) also satisfies the inequality (33), it is sufficient to check that the largest possible value $|\rho|$ satisfying the inequality (30) satisfies the inequality (33), i.e., that $$\frac{\sqrt{a\cdot b}}{\sqrt{(1-a)\cdot(1-b)}}\le \frac{4\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}}{(1-2\cdot a)\cdot (1-2\cdot b)}.\eqno{(34)}$$ If we divide both sides by $\sqrt{a\cdot b}$, we get the following equivalent inequality $$\frac{1}{\sqrt{(1-a)\cdot (1-b)}}\le \frac{4\cdot \sqrt{(1-a)\cdot (1-b)}}{(1-2\cdot a)\cdot (1-2\cdot b)}.\eqno{(35)}$$ Multiplying both sides by both denominators, we get the following equivalent inequality $$(1-2\cdot a)\cdot (1-2\cdot b)\le 4\cdot (1-a)\cdot (1-b).\eqno{(36)}$$ Opening parentheses, we get $$1-2\cdot a-2\cdot b+4\cdot a\cdot b\le 4-4\cdot a-4\cdot b+4\cdot a\cdot b.\eqno{(37)}$$ Adding $4\cdot a+4\cdot b-4\cdot a\cdot b-1$ to both sides, we get an equivalent inequality $$2\cdot a+2\cdot b\le 3,\eqno{(38)}$$ which is true since we consider the case when $a+b<1$. So, in this case, we indeed have $d\ge 0$. \medskip In all cases when have $d\ge 0$, thus, the and-operation $f_\rho(a,b)$ is indeed a copula. Thus, for boxes in which all four vertices belong to the area described by the expression (1), the inequality (4a) is always satisfied. \medskip \noindent $5^\circ$. Let us now consider the boxes in which two vertices belong to the boundary between two areas. First, we will consider the case when $\rho>0$ and then, we will consider the case when $\rho<0$. \medskip \noindent $6^\circ$. Let us first consider the case when $\rho>0$. For this case, let us first describe the boundaries between the areas. \medskip \noindent $6.1^\circ$. Let us analyze which of the three areas listed in formula (4) are possible in this case. \medskip When $\rho>0$, we have $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\ge a\cdot b,$$ and since it is known that we always have $a\cdot b\ge \max(a+b-1,0)$, we have $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\ge \max(a+b-1,0).$$ So, for $\rho>0$, we cannot have the first of the three cases described by the formula (4). So, we only have two areas: \begin{itemize} \item the area where the and-operation is described by the formula (1), and \item the area where the and-operation is described by the formula $\min(a,b)$. \end{itemize} \medskip \noindent $6.2^\circ$. Let us describe the two possible areas and the boundary between these two areas. \medskip The first area is characterized by the inequality $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le \min(a,b).\eqno{(39)}$$ Similarly to the previous part of the proof, without losing generality, we can consider the case when $a\le b$. In this case, the inequality (39) describing the first area takes the following form: $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le a.\eqno{(40)}$$ If we subtract $a\cdot b$ from both sides of this inequality, we get the following equivalent inequality: $$\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le a\cdot (1-b).\eqno{(41)}$$ Both sides of this inequality are non-negative, so we can get an equivalent inequality is we square both sides: $$\rho^2\cdot a\cdot (1-a)\cdot b\cdot (1-b)\le a^2\cdot (1-b)^2.\eqno{(42)}$$ The cases when $a$ or $b$ are equal to 0 or 1 can be obtained by taking a limit from the cases when both $a$ and $b$ are located insyed the interval $(0,1)$. For such values, we can divide both side of the inequality by positive numbers $a^2$, $b$, and $1-b$, and get the following equivalent inequality: $$\rho^2\cdot \frac{1-a}{a}\le\frac{1-b}{b},\eqno{(43)}$$ i.e., equivalently, $$\rho^2\cdot \frac{1-a}{a}\le\frac{1}{b}-1.\eqno{(44)}$$ By adding 1 to both sides of this inequality, we get $$\frac{a+\rho^2\cdot (1-a)}{a}\le \frac{1}{b},\eqno{(45)}$$ i.e., equivalently, that $$b\le \frac{a}{a+\rho^2\cdot (1-a)}.\eqno{(46)}$$ This inequality describes the first area, in which the and-operation is described by the formula (1). Thus, the boundary between the two areas is described by the equality $$b=\frac{a}{a+\rho^2\cdot (1-a)}.\eqno{(47)}$$ \medskip \noindent{\it Comment.} One can see that for $a=0$ we get $b=0$, for $a=1$, we get $b=1$. \medskip \noindent $6.3^\circ$. Let us prove that for all $a$, the corresponding boundary value $b$ is greater than or equal to $a$ -- i.e., that for all the points $(a,b)$ on this boundary, we have $a\le b$. \medskip Indeed, for the expression (47), the desired inequality $a\le b$ takes the form $$a\le \frac{a}{a+\rho^2\cdot (1-a)}.\eqno{(48)}$$ If we divide both sides by $a$ and multiply both sides by the denominator of the right-hand side, we get the following equivalent inequality $$a+\rho^2\cdot (1-a)\le 1.\eqno{(49)}$$ If we move all the terms to the right-hand side, we get an equivalent inequality $$0\le 1-a-\rho^2\cdot (1-a)=(1-\rho^2)\cdot (1-a).\eqno{(50)}$$ This inequality is always true, since $\rho^2\le 1$ and $a\le 1$, so indeed, for all boundary points, we have $a\le b$. \medskip \noindent $6.4^\circ$. Let us prove that the boundary describes $b$ as an increasing function of $a$. \medskip By applying, to the equality (47) that describes the boundary, the same transformations that show the equivalent of inequalities (43) and (46), we can conclude that the equality (47) is equivalent to $$\rho^2\cdot \frac{1-a}{a}=\frac{1-b}{b},\eqno{(51)}$$ i.e., to $$\rho^2\cdot \left(\frac{1}{a}-1\right)=\frac{1}{b}-1.\eqno{(52)}$$ The left-hand side is decreasing with respect to $a$, the right-hand side is a decreasing function of $b$. Thus, as $a$ increases, the left-hand side decreases, thus the right-hand side also decreases and hence, the value $b$ increases as well. \medskip \noindent $6.5^\circ$. For $\rho=1$ the condition (46) describing the first area takes the form $b\le a$. Since we have $a\le b$, this means that this condition is only satisfies for $a=b$. For these values, the expression (4a) is equal to $$a\cdot a+\sqrt{a\cdot(1-a)\cdot a\cdot (1-a)}=a^2+a\cdot (1-a)=a^2+a-a^2=$$ $$a=\min(a,b),\eqno{(53)}$$ which means that our and-operation is always equal to $\min(a,b)$. The expression $\min(a,b)$ is known to be a copula. So, we only need to prove the fact that our and-operation is a copula for the case when $\rho<1$. This is the case we will consider from now on. \medskip \noindent $6.6^\circ$. Let us prove that for $\rho<1$, the only boundary points for which $a=b$ are points for which $a=b=0$ and $a=b=1$. \medskip Indeed, as we have mentioned, the points $(0,0)$ and $(1,1)$ are boundary points. Let us prove, by contradiction, that there are no other boundary points for which $a=b$. Indeed, when $a=b$, the equality (52) that describes the boundary takes the form: $$\rho^2\cdot \left(\frac{1}{a}-1\right)=\frac{1}{a}-1.\eqno{(54)}$$ Dividing both sides of this equality by the non-zero right-hand side, we get $\rho^2=1$. This contradicts to the fact that we are considering the case when $\rho<1$ and thus, $\rho^2<1$. This contradiction shows that other boundary points with $a=b$ are not possible. \medskip \noindent $6.7^\circ$. The boundary consists of a curved line that is separate from the line $a=b$ -- except for the endpoints. So, if we limit ourselves to a sub-box $[\varepsilon,1-\varepsilon]\times [\varepsilon,1-\varepsilon]$ for some small $\varepsilon>0$, the boundary line is separated from the line $a=b$ -- there is the smallest distance $\delta>0$ between points of these two lines. So, if we have a box that includes both points with $a\le b$ and with $a\ge b$, we can divide this box into sub-boxes of linear size $<\delta/2$ and thus, make sure that every sub-box that contains boundary points with $a\le b$ cannot contain any points with $a=b$ -- and therefore, only contains points with $a\le b$. So, due to additivity, it is sufficient to prove the inequality (4a) for boxes for which: \begin{itemize} \item two vertices lie on the boundary, and \item we have $a\le b$ for all the points from this sub-box. \end{itemize} This will allow us to prove the inequality (4a) for all sub-boxes of the square $[\varepsilon,1-\varepsilon]\times [\varepsilon,1-\varepsilon]$. We can do it for any $\varepsilon$ and thus, in the limit, get the desired inequality for all sub-boxes of the original square $[0,1]\times [0,1]$ as well. So, suppose that we have a box for which: \begin{itemize} \item two vertices lie on the boundary, and \item we have $a\le b$ for all the points from this box. \end{itemize} Since the boundary describes the increasing function of $a$, the corresponding box has the form \medskip \begin{center} \begin{picture}(50,50) \put(0,0){\line(1,0){50}} \put(0,50){\line(1,0){50}} \put(0,0){\line(0,1){50}} \put(50,0){\line(0,1){50}} \put(0,0){\line(1,1){50}} \end{picture} \end{center} \medskip So, in the corresponding box: \begin{itemize} \item the two vertices $(\underline a,\underline b)$ and $(\overline a,\overline b)$ are on the boundary, \item the vertex $(\overline a,\underline b)$ is in the first area, i.e., for this point, we have the expression (1), and \item the vertex $(\underline a,\overline b)$ is in the second area, i.e., here $C(\underline a,\overline b)=\min(\underline a,\overline b)$. \end{itemize} The desired inequality (4a) has the form $$C(\overline a,\underline b)-C(\underline a,\underline b)\le C(\overline a,\overline b)-C(\underline a,\overline b).\eqno{(55)}$$ The points $(\overline a,\overline b)$ and $(\underline a,\overline b)$ are both in the second area for which $C(a,b)=\min(a,b)$ -- to be more precise, the second of these points is in the boundary, which means it also satisfies the condition $C(a,b)=\min(a,b)$. For all the points from the box, $a\le b$, so we have $$C(\overline a,\overline b)-C(\underline a,\overline b)=\min(\overline a,\overline b)-\min(\underline a,\overline b)=\overline a-\underline a.\eqno{(56)}$$ On the other hand, for the difference in the left-hand side of the formula (55), we have $$C(\overline a,\underline b)-C(\underline a,\underline b)=\int_{\underline a}^{\overline a} \frac{\partial C}{\partial a}\,da.\eqno{(57)}$$ So, if we prove that the partial derivative $\partial C/\partial a$ is always smaller or equal than 1, we would indeed conclude that $$C(\overline a,\underline b)-C(\underline a,\underline b)=\int_{\underline a}^{\overline a} 1\,da=\overline a-\underline a,\eqno{(58)}$$ i.e., exactly, the desired inequality (55). For the points $(\overline a,\underline b)$ and $(\underline a,\underline b)$ -- and the points from the interval connecting these two points -- the expression $C(a,b)$ is described by the formula (1). Thus, the partial derivative of $C(a,b)$ with respect to $a$ is described by the formula (4b). Thus, the inequality $$\frac{\partial C}{\partial a}(a,b)\le 1,\eqno{(59)}$$ takes the form $$b+\rho\cdot \frac{1-2\cdot a}{2\cdot\sqrt{a\cdot (1-a)}}\cdot \sqrt{b\cdot (1-b)}\le 1.\eqno{(60)}$$ Subtracting $b$ from both sides of (60), we get an equivalent inequality $$\rho\cdot \frac{1-2\cdot a}{2\cdot\sqrt{a\cdot (1-a)}}\cdot \sqrt{b\cdot (1-b)}\le 1-b.\eqno{(61)}$$ To separate the variables, we can divide both sides by $\sqrt{b\cdot (1-b)}$, then we get an equivalent inequality $$\rho\cdot \frac{1-2\cdot a}{2\cdot\sqrt{a\cdot (1-a)}}\le \sqrt{\frac{1-b}{b}}.\eqno{(62)}$$ By taking the square root of both sides of the inequality (46), we conclude that: $$\rho\cdot \sqrt{\frac{1-a}{a}}\le \sqrt{\frac{1-b}{b}}.\eqno{(63)}$$ Thus, if we prove that the left-hand side of the inequality (62) is smaller than or equal to the left-hand side of the inequality (63), i.e., that $$\rho\cdot \frac{1-2\cdot a}{2\cdot\sqrt{a\cdot (1-a)}}\le \rho\cdot \sqrt{\frac{1-a}{a}};\eqno{(64)}$$ this will prove the inequality (62) and thus, the desired upper bound (60) on the partial derivative. We can simplify the inequality (64) by dividing both sides by $\rho$ and multiplying both sides by $2\cdot \sqrt{a\cdot (1-a)}$. Then, we get an equivalent inequality $$1-2\cdot a\le 2\cdot (1-a)=2-2\cdot a,\eqno{(65)}$$ which is equivalent to $1\le 2$ and is, thus, always true. Thus, (55) holds, so the inequality (4a) is true for all the boxes in which two vertices are located on the boundary. This completes the proof of the Proposition for the case when $\rho>0$. \medskip \noindent $7^\circ$. Let us now consider the case when $\rho<0$. For this case, let us first describe the boundaries between the areas. \medskip \noindent $7.1^\circ$. Let us analyze which of the three areas listed in formula (4) are possible in this case. \medskip When $\rho<0$, we have $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le a\cdot b,$$ and since it is known that we always have $a\cdot b\le \min(a,b)$, we have $$a\cdot b+\rho\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\le\min(a,b).$$ So, for $\rho<0$, we cannot have the third of the three cases described by the formula (4). So, we only have two areas: \begin{itemize} \item the area where the and-operation is described by the formula (1), and \item the area where the and-operation is described by the formula $$\max(a+b-1,0).$$ \end{itemize} \medskip \noindent $7.2^\circ$. Let us describe the two possible areas and the boundary between these two areas. \medskip The first area is characterized by the inequality $C(a,b)\ge \max(a+b-1,0)$, i.e., equivalently, by two inequalities $$a\cdot b-|\rho|\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\ge 0\eqno{(66)}$$ and $$a\cdot b-|\rho|\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}\ge a+b-1.\eqno{(67)}$$ Let us consider these two inequalities one by one. \medskip \noindent $7.2.1^\circ$. The inequality (66) is equivalent to: $$a\cdot b\ge |\rho|\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}.\eqno{(68)}$$ To separate the variables, let us divide both sides of this inequality by $$a\cdot \sqrt{b\cdot (1-b)},$$ then we get an equivalent inequality $$\sqrt{\frac{b}{1-b}}\ge |\rho|\cdot \sqrt{\frac{1-a}{a}}.\eqno{(69)}$$ Both sides of this inequality are non-negative, thus if we square both sides, we get an equivalent inequality $$\frac{b}{1-b}\ge \rho^2\cdot \frac{1-a}{a}.\eqno{(70)}$$ Reversing both sides, we get an equivalent inequality $$\frac{1-b}{b}\le \frac{a}{\rho^2\cdot (1-a)},\eqno{(71)}$$ i.e., equivalently, $$\frac{1}{b}-1\le \frac{a}{\rho^2\cdot (1-a)}.\eqno{(72)}$$ By adding 1 to both sides, we get $$\frac{1}{b}\le \frac{\rho^2\cdot (1-a)+a}{\rho^2\cdot (1-a)},\eqno{(73)}$$ i.e., equivalently, $$b\ge \frac{\rho^2\cdot (1-a)}{\rho^2\cdot (1-a)+a}.\eqno{(74)}$$ \medskip \noindent $7.2.2^\circ$. The inequality (67) is equivalent to $$a\cdot b-a-b+1\ge |\rho|\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)},\eqno{(75)}$$ i.e., $$(1-a)\cdot (1-b)\ge |\rho|\cdot \sqrt{a\cdot (1-a)\cdot b\cdot (1-b)}.\eqno{(76)}$$ To separate the variables, let us divide both sides by $(1-a)\cdot \sqrt{b\cdot (1-b)}$, then we get an equivalent inequality $$\sqrt{\frac{1-b}{b}}\ge |\rho|\cdot \sqrt{\frac{a}{1-a}}.\eqno{(77)}$$ Both sides of this inequality are non-negative, thus if we square both sides, we get an equivalent inequality $$\frac{1-b}{b}\ge \rho^2\cdot \frac{a}{1-a},\eqno{(78)}$$ i.e., equivalently, $$\frac{1}{b}-1\ge \rho^2\cdot \frac{a}{1-a}.\eqno{(79)}$$ By adding 1 to both sides, we get $$\frac{1}{b}\ge \frac{\rho^2\cdot a + (1-a)}{1-a},\eqno{(80)}$$ i.e., equivalently, $$b\le \frac{1-a}{\rho^2\cdot a+(1-a)}.\eqno{(81)}$$ \medskip \noindent $7.2.3^\circ$. By combining the inequalities (74) and (81), we get the following description of the area in which the and-operation is described by the formula (1): $$\frac{\rho^2\cdot (1-a)}{\rho^2\cdot (1-a)+a}\le b\le \frac{1-a}{\rho^2\cdot a+(1-a)}.\eqno{(82)}$$ Thus, the boundary between the two areas consists of the following two curves: $$b=\frac{\rho^2\cdot (1-a)}{\rho^2\cdot (1-a)+a}\eqno{(83)}$$ and $$b=\frac{1-a}{\rho^2\cdot a+(1-a)}.\eqno{(84)}$$ \medskip \noindent $7.3^\circ$. Let us prove that: \begin{itemize} \item the curve (83) lies in the area where $a+b\le 1$, and \item the curve (84) lies in the area where $a+b\ge 1$. \end{itemize} \medskip \noindent $7.3.1^\circ$. Let us first prove that for each value $b$ described by the formula (83), we have $a+b\le 1$. \medskip We need to prove the inequality $$a+\frac{\rho^2\cdot (1-a)}{\rho^2\cdot (1-a)+a}\le 1.\eqno{(85)}$$ Subtracting $a$ from both sides, we get an equivalent inequality $$\frac{\rho^2\cdot (1-a)}{\rho^2\cdot (1-a)+a}\le 1-a.\eqno{(86)}$$ Dividing both sides by $1-a$ and multiplying both sides by the denominator of the left-hand side, we get the following equivalent inequality: $$\rho^2\le \rho^2\cdot (1-a)+a=\rho^2+(1-\rho^2)\cdot a,\eqno{(87)}$$ which is, of course, always true, since $\rho^2\le 1$ and $a\ge 0$. The statement is proven. \medskip \noindent $7.3.2^\circ$. Let us now prove that for each value $b$ described by the formula (84), we have $a+b\ge 1$. \medskip We need to prove the inequality $$a+\frac{1-a}{\rho^2\cdot a+(1-a)}\ge 1.\eqno{(88)}$$ Subtracting $a$ from both sides, we get an equivalent inequality $$\frac{1-a}{\rho^2\cdot a+(1-a)}\ge 1-a.\eqno{(89)}$$ Dividing both sides by $1-a$ and multiplying both sides by the denominator of the left-hand side, we get the following equivalent inequality: $$1 \ge \rho^2\le a+(1-a)=1-(1-\rho^2)\cdot a,\eqno{(90)}$$ which is, of course, always true. The statement is proven. \medskip \noindent $7.4^\circ$. Similarly to Part 6 of this proof, it is sufficient to prove the inequality (4a) for boxes in which two vertices are on the boundary and for which: \begin{itemize} \item either we have $a+b\le 1$ for all the points from the box, \item or we have $a+b\ge 1$ for all the points from the box. \end{itemize} Let us consider the two parts of the boundary one by one. \medskip \noindent $7.4.1^\circ$. Let us first consider the case when we have $a+b\le 1$ for all the points from the box. In this case, the corresponding part of the boundary is described by the formula (83). By reformulating this expression in the equivalent form $$b=\frac{1}{1+\displaystyle\frac{a}{\rho^2\cdot (1-a)}}= \frac{1} { 1+\displaystyle\frac{1}{\rho^2}\cdot \displaystyle\frac{1}{\displaystyle\frac{1}{a}-1} },\eqno{(91)}$$ we can see that $b$ is a decreasing function of $a$. Thus, the corresponding box has the form \medskip \begin{center} \begin{picture}(50,50) \put(0,0){\line(1,0){50}} \put(0,50){\line(1,0){50}} \put(0,0){\line(0,1){50}} \put(50,0){\line(0,1){50}} \put(0,50){\line(1,-1){50}} \end{picture} \end{center} \medskip So, in the corresponding box: \begin{itemize} \item the two vertices $(\underline a,\overline b)$ and $(\overline a,\underline b)$ are on the boundary, \item the vertex $(\overline a,\overline b)$ is in the first area, i.e., for this point, we have the expression (1), and \item the vertex $(\underline a,\underline b)$ is in the second area, i.e., here $C(\underline a,\overline b)=\max(\underline a+\underline b-1,0)$. \end{itemize} So, for three vertices, we have $C(a,b)=\max(a+b-1,0)$. Since for all the points from the box, we have $a+b\le 1$, this means that for three vertices, we have $C(a,b)=0$. In this case, the inequality (4a) is clearly true. \medskip \noindent $7.4.2^\circ$. Let us now consider the case when we have $a+b\ge 1$ for all the points from the box. In this case, the corresponding part of the boundary is described by the formula (84). By reformulating this expression in the equivalent form $$b=\frac{1}{1+\rho^2\cdot \displaystyle\frac{a}{1-a}}= \frac{1} { 1+\rho^2\cdot \displaystyle\frac{1}{\displaystyle\frac{1}{a}-1} },\eqno{(91)}$$ we can see that $b$ is also a decreasing function of $a$. Thus, the corresponding box has the same form as in the case $a+b\le 1$: \medskip \begin{center} \begin{picture}(50,50) \put(0,0){\line(1,0){50}} \put(0,50){\line(1,0){50}} \put(0,0){\line(0,1){50}} \put(50,0){\line(0,1){50}} \put(0,50){\line(1,-1){50}} \end{picture} \end{center} \medskip So, in the corresponding box: \begin{itemize} \item the two vertices $(\underline a,\overline b)$ and $(\overline a,\underline b)$ are on the boundary, \item the vertex $(\underline a,\underline b)$ is in the first area, i.e., for this point, we have the expression (1), and \item the vertex $(\overline a,\overline b)$ is in the second area, i.e., here $C(\underline a,\overline b)=\max(\underline a+\underline b-1,0)$. \end{itemize} Similarly to Part 6 of the proof, we can show that the desired inequality (4a) is satisfied if we the corresponding partial derivatives is smaller than or equal to 1, i.e., if $$\frac{\partial C}{\partial a}=b-|\rho|\cdot \frac{1-2\cdot a}{2\cdot \sqrt{a\cdot (1-a)}}\cdot \sqrt{b\cdot (1-b)}\le 1.\eqno{(92)}$$ Subtracting $b$ from both sides, we get an equivalent inequality $$-|\rho|\cdot \frac{1-2\cdot a}{2\cdot \sqrt{a\cdot (1-a)}}\cdot \sqrt{b\cdot (1-b)}\le 1-b.\eqno{(93)}$$ We can separate the variable if we divide both sides by $\sqrt{b\cdot (1-b)}$, then we get an equivalent inequality $$-|\rho|\cdot \frac{1-2\cdot a}{2\cdot \sqrt{a\cdot (1-a)}}\le \sqrt{\frac{1-b}{b}}.\eqno{(94)}$$ We know a lower bound on the expression in the right-hand side -- it is provided by the inequality (77). Thus, to prove the inequality (94), it is sufficient to prove that the left-hand side of the formula (94) is smaller than or equal to this lower bound, i.e., that $$-|\rho|\cdot \frac{1-2\cdot a}{2\cdot \sqrt{a\cdot (1-a)}} \le|\rho|\cdot \sqrt{\frac{a}{1-a}}.\eqno{(95)}$$ Let us prove this inequality. Dividing both sides of (95) by $|\rho|$ and multiplying both sides by $2\cdot \sqrt{a\cdot (1-a)}$, we get an equivalent inequality $-(1-2\cdot a)\le 2\cdot a$, i.e., $2\cdot a-1\le 2\cdot a$, which is always true. Thus, the inequality (94) holds, hence the inequality (92) also holds, and therefore, in this case, the inequality (4a) that describes a copula is also true. \medskip \noindent $8^\circ$. We have considered all possible cases, and in all these cases, we have shown that the inequality (4a) -- that defines a copula -- is true. Thus, our and-operation is indeed a copula. The proposition is proven. \section*{Acknowledgments} This research was partly funded by the EPSRC and ESRC CDT in Risk and Uncertainty (EP/L015927/1), established within the Institute for Risk and Uncertainty at the University of Liverpool. This work has been carried out within the framework of the EUROfusion Consortium, funded by the European Union via the Euratom Research and Training Programme (Grant Agreement No 101052200 - EUROfusion). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Commission. Neither the European Union nor the European Commission can be held responsible for them. V.K. was supported in part by the National Science Foundation grants 1623190 (A Model of Change for Preparing a New Generation for Professional Practice in Computer Science), and HRD-1834620 and HRD-2034030 (CAHSI Includes), and by the AT\&T Fellowship in Information Technology. He was also supported by the program of the development of the Scientific-Educational Mathematical Center of Volga Federal District No. 075-02-2020-1478, and by a grant from the Hungarian National Research, Development and Innovation Office (NRDI).

https://arxiv.org/abs/2001.05018

Walking to infinity on gaussian lines

We study analogies between the rational integers on the real line and the Gaussian integers on other lines in the complex plane. This includes a Gaussian analog of Bertrands Postulate, the Chinese Remainder Theorem, and the periodicity of divisibility. We also computationally investigate the distribution of Gaussian primes along these lines and leave the reader with several open problems.

\section{\@startsection {section}{1}{\z@} {-30pt \@plus -1ex \@minus -.2ex} {2.3ex \@plus.2ex} {\normalfont\normalsize\bfseries\boldmath}} \renewcommand\subsection{\@startsection{subsection}{2}{\z@} {-3.25ex\@plus -1ex \@minus -.2ex} {1.5ex \@plus .2ex} {\normalfont\normalsize\bfseries\boldmath}} \renewcommand{\@seccntformat}[1]{\csname the#1\endcsname. } \makeatother \newtheorem{theorem}{Theorem} \newtheorem{lemma}{Lemma} \newtheorem{conjecture}{Conjecture} \newtheorem{proposition}{Proposition} \newtheorem{corollary}{Corollary} \newtheorem*{definition}{Definition} \newtheorem*{remark}{Remark} \newcommand{\mathcal{R}(L)}{\mathcal{R}(L)} \newcommand{\mathbb{N}}{\mathbb{N}} \newcommand{\mathbb{Z}}{\mathbb{Z}} \newcommand{\mathbb{I}}{\mathbb{I}} \newcommand{\mathbb{R}}{\mathbb{R}} \newcommand{\mathbb{Q}}{\mathbb{Q}} \newcommand{{\mathcal{G}\mathcal{P}}(L)}{{\mathcal{G}\mathcal{P}}(L)} \newcommand{{\mathcal{D}}(L)}{{\mathcal{D}}(L)} \newcommand{\alpha}{\alpha} \newcommand{\alpha_0}{\alpha_0} \newcommand{\alpha_n}{\alpha_n} \newcommand{\Z[i]}{\mathbb{Z}[i]} \newcommand{{\text{Re}}}{{\text{Re}}} \newcommand{{\text{Im}}}{{\text{Im}}} \newcommand{\begin{equation}}{\begin{equation}} \newcommand{\end{equation}}{\end{equation}} \newcommand{\begin{eqnarray*}}{\begin{eqnarray*}} \newcommand{\end{eqnarray*}}{\end{eqnarray*}} \begin{document} \begin{center} \uppercase{\bf Walking to infinity along Gaussian lines} \vskip 20pt {\bf Elsa Magness }\\ {\smallit Department of Mathematics, Seattle University, Seattle, WA 98122, USA}\\ {\tt [email protected]}\\ \vskip 10pt {\bf Brian Nugent }\\ {\smallit Department of Mathematics, Seattle University, Seattle, WA 98122, USA}\\ {\tt [email protected]}\\ \vskip 10pt {\bf Leanne Robertson}\\ {\smallit Department of Mathematics, Seattle University, Seattle, WA 98122, USA}\\ {\tt [email protected]}\\ \end{center} \vskip 20pt \vskip 30pt \centerline{\bf Abstract} \noindent We study analogies between the rational integers on the real line and the Gaussian integers on other lines in the complex plane. This includes a Gaussian analog of Bertrand's Postulate, the Chinese Remainder Theorem, and the periodicity of divisibility. We also computationally investigate the distribution of Gaussian primes along these lines and leave the reader with several open problems. \pagestyle{myheadings} \thispagestyle{empty} \baselineskip=12.875pt \vskip 30pt \section{Introduction.}\label{intro} Is it possible to walk from the origin in the complex plane to infinity using steps of bounded length and stepping only on Gaussian primes? Several authors have worked on this intriguing question since it was first posed by Basil Gordon in 1962. Erd\"os conjectured that such a walk to infinity is impossible, but the problem remains unsolved today (see \cite{Gethner} for a discussion of the contradictory references to Erd\"os' role in this problem). In 1970, Jordan and Rabung \cite{JR} showed that steps of length at least 4 would be required, and in 1998, Gethner, Wagon, and Wick~\cite{Gethner} showed that steps of length $\sqrt{26}$ or less are insufficient to reach infinity. In the same paper they showed that it is impossible to walk to infinity on any line in the complex plane by stepping only on Gaussian primes and taking steps of bounded length, and thus established the Gaussian analog of the classical result that there are arbitrarily long sequences of composites on the real line. In 2017, West and Sittinger \cite{West} generalized this result and showed that in any quadratic field of class number 1, it is similarly impossible to walk to infinity along any line using steps of bounded length and stepping only on primes in the ring of integers of the field. Motivated by these results, we further investigate the idea of walking to infinity on lines in the complex plane stepping only on Gaussian integers, and analogies to walking to infinity along the real line. Recall that the ring $\mathbb{Z} [i]$ of {\em Gaussian integers} consists of all complex numbers of the form $\alpha = a+bi$, where $a$ and $b$ are rational integers. Following Gethner et al., we call a line in the complex plane a {\em{Gaussian line}} if it contains two, and hence infinitely many, Gaussian integers. We call a Gaussian line {\em{primitive}} if the integers on the line do not all share a common divisor. With these definitions, we ask what you might discover if instead of wandering freely on Gaussian integers in the complex plane, you walked along a primitive Gaussian line stepping only on Gaussian integers? How different or similar would this stroll to infinity be to that of walking to infinity along the real line stepping only on rational integers? Would you stroll on infinitely many Gaussian primes, or perhaps none at all? Could you observe an analog of Bertrand's postulate on your walk? Would you see a periodicity of divisibility similar to that on the real line? What other properties of the Gaussian integers might you discover? An overview of the paper and our results is as follows. In Section \ref{background}, we establish the background and notation used throughout. In Section \ref{primes}, we investigate the distribution of Gaussian primes on Gaussian lines. We discuss what a theorem of T. Tao says about primes on Gaussian lines and formulate and computationally support an extension of Bertrand's Postulate to these lines. The main questions posed in this section are equivalent to famous open problems about quadratic polynomials representing prime numbers, so we turn to more tractable problems in subsequent sections. In Section~\ref{div}, we prove key divisibility properties of Gaussian integers on Gaussian lines that are important for the rest of the paper. This includes an analogy of the periodicity of divisibility of rational integers on the real line and a characterization of the rational integers and Gaussian primes that divide some Gaussian integer on a given Gaussian line. In Section~\ref{CRT}, we extend the Chinese Remainder Theorem to Gaussian lines and prove a theorem that shows there are always infinitely many Gaussian lines that satisfy any given CRT-type divisibility properties. Finally, in Section \ref{divisor} we return to questions raised in Section \ref{div} and completely characterize the set of Gaussian integers that divide some Gaussian integer on a given Gaussian line. \section{Background and Notation.}\label{background} We begin with some background on Gaussian integers and by establishing the notation concerning Gaussian lines that is used throughout the paper. \medskip The unit group of the Gaussian integers $\mathbb{Z}[i]$ is $\{\pm 1, \pm i\}$, so two Gaussian integers, $\alpha$ and $\beta$, are {\em associates} if and only if $\alpha=\pm \beta$ or $\alpha = \pm i \beta$. The {\em norm} of the Gaussian integer $\alpha=a+bi$ is defined by $N(a+bi)=\alpha\cdot\overline\alpha =a^2+b^2\in\mathbb{Z}$, where the ``bar'' denotes complex conjugation, and its {\em trace} is defined by $Tr(a+bi)=\alpha+\overline\alpha=2a\in\mathbb{Z}$. Unique factorization holds in $\mathbb{Z}[i]$, and this gives the Gaussian integers a well-defined notion of primality. To avoid confusion, we use the terminology {\em rational prime} for a prime in the rational integers $\mathbb{Z}$, and {\em Gaussian prime} for a prime in $\mathbb{Z} [i]$. The Gaussian primes can be classified in terms of the factorization of the rational primes $p\in \mathbb{Z}$ into Gaussian primes as follows: \begin{center} \begin{enumerate} \item If $p=2$, then $p$ ramifies in $\mathbb{Z} [i]$. Specifically, $2=-i(1+i)^2$, so $1+i$ is a Gaussian prime of norm 2. \item If $p \equiv 1 \pmod 4$, then $p = \pi\cdot\overline{\pi}$ splits as a product of two conjugate Gaussian primes of norm $p$ that are not associates in $\mathbb{Z}[i]$. \item If $p \equiv 3 \pmod 4$, then $p$ remains prime in $\mathbb{Z} [i]$ and has norm $p^2$.\\ \end{enumerate} \end{center} Every Gaussian prime is an associate of one of the Gaussian primes described above. If $\pi$ is a Gaussian prime then we say $\pi$ \textit{lies over} $p$ if $\pi$ divides the rational prime $p$. For every Gaussian line $L$, we distinguish two Gaussian integers, $\alpha_0=a+bi$ and $\delta=c+di$, that define $L$ as follows. Let $\alpha_0$ be the Gaussian integer on $L$ of minimum norm, and if there are two such integers, let $\alpha_0 $ be the one with the larger real part. If $L$ is vertical, then take $\delta=i$. Otherwise, let $\alpha_1$ be the Gaussian integer on $L$ closest to $\alpha_0$ (so $N(\alpha_1-\alpha_0)$ is minimal) and with ${\text{Re}}(\alpha_1)>{\text{Re}}(\alpha_0)$. Then take $\delta=\alpha_1-\alpha_0$. Thus $\alpha_0$ is on the line $L$, but $\delta$ is not, provided $\alpha_0\neq 0$. Note that there are only two primitive Gaussian lines with $\alpha_0=0$, namely the real line ${\text{Im}}(z)=0$ and the imaginary line ${\text{Re}}(z)=0$. With $\alpha_0$ and $\delta$ defined in this way, the lemma below describes all Gaussian integers on $L$. This lemma is essentially Lemma 4.2 in \cite{Gethner}, except that we describe the primitive case and specify $\alpha_0$ and $\delta$, since this is convenient for our work. \begin{lemma}\label{notation} Let $L$ be a Gaussian line, and let $\alpha_0=a+bi$ and $\delta=c+di$ be as defined above. Then $c$ and $d$ are relatively prime, $c\geq 0$, and the Gaussian integers on $L$ are exactly the Gaussian integers $\alpha_n$ given by $$ \alpha_n=\alpha_0+\delta n,\ n\in\mathbb{Z}.\label{an}$$ Moreover, $L$ is primitive if and only if $\alpha_0$ and $\delta$ are relatively prime over $\mathbb{Z}[i]$. \end{lemma} \begin{proof} If $L$ is vertical then $\delta=i$ and $\alpha_0=k$ for some $k\in\mathbb{Z}$. Then the Gaussian integers on $L$ are given by $\alpha_n=k+ni, n\in\mathbb{Z}$, $L$ is primitive, and $\alpha_0$ and $\delta$ are relatively prime. Thus, the lemma holds for all vertical Gaussian lines. If $L$ is not vertical, then by our choice of $\delta=c+di$ we have $c> 0$ and $L$ has slope $d/c$. Thus $c$ and $d$ must be relatively prime since otherwise there would be a Gaussian integer on $L$ between $\alpha_0$ and $\alpha_1$, contradicting our choice of $\alpha_1$. Let $\beta$ be a Gaussian integer on $L$. Then $\beta=\alpha_0+r\delta$ for some real number $r$. But, $r=(\beta-\alpha_0)/\delta$ is in the quotient field $\mathbb{Q}(i)$, so $r\in\mathbb{Q}$. Now $r\delta=rc+rdi=\beta-\alpha_0\in\mathbb{Z}[i]$ implies $rc, rd\in\mathbb{Z}$. Since $c$ and $d$ are relatively prime, it follows that $r\in\mathbb{Z}$ as needed. For the second part of the lemma, first suppose $\alpha_0$ and $\delta$ have a common Gaussian prime divisor $\pi$. Then $\pi$ divides $\alpha_0+\delta n$ for all $n\in\mathbb{Z}$, $i.e.$, $\pi$ divides all Gaussian integers $\alpha_n$ on $L$ and $L$ is not primitive. Conversely, if $\alpha_0$ and $\delta$ are relatively prime, then $\alpha_0$ and ${\alpha_1}=\alpha_0+\delta$ are also relative prime, and $L$ must be primitive since it contains at least two Gaussian integers that do not share a common divisor. \end{proof} Throughout this paper, we define a Gaussian line $L$ by its values of $\alpha_0$ and $\delta$ as given in Lemma \ref{notation}. Given these values, we also define a rational integer $\Delta$ associated to $L$ by \begin{equation}\label{D}\Delta=ad-bc.\end{equation} Note that if $\alpha_n=x+yi=\alpha_0+n\delta$, $n\in\mathbb{Z}$, is any other Gaussian integer on $L$, then $x=a+nc$ and $y=b+nd$ and so $\Delta$ can also be computed by $\Delta=xd-yc$. That is, $\Delta$ can easily be computed from the values of $\alpha_n$ and $\delta$, not just from $\alpha_0$ and $\delta$. In Section \ref{div}, we use $\Delta$ to characterize the set of rational integers that divide some Gaussian integer on $L$. Another use of $\Delta$ is given by the following easy lemma. \begin{lemma}\label{delta} Let $L$ be a primitive Gaussian line. Then $\Delta=0$ if and only if $L$ is the real or imaginary line, which holds if and only if $\alpha_0=0$.\end{lemma} \begin{proof} The only part of the lemma that doesn't follow directly from the definitions is the fact that if $\Delta=0$ then $L$ is the real or imaginary line. For this assume $\Delta=0$, so $ad=bc$. Since $c$ and $d$ are relatively prime, it follows that $c\mid a$ and $d\mid b$. Thus, $a=cx$ and $b=dy$ for some $x,y\in\mathbb{Z}$. This gives $cdx=cdy$. We may assume $c$ and $d$ are both nonzero since otherwise $a$ or $b$ is equal to zero and $L$ is the real or imaginary line. Thus, it follows that $x=y$ and $\alpha_0=x\delta$. Hence, $x=0$ since $\alpha_0$ and $\delta$ are relatively prime and $\delta\neq 0$. Therefore, $\alpha_0=0$, and $L$ is either the real or imaginary line. \end{proof} \section{Primes on Gaussian Lines.}\label{primes} One of the first questions we had when we began our study of Gaussian lines was about the distribution of Gaussian primes on these lines. We wondered whether every primitive Gaussian line contains infinitely many Gaussian primes, or if the existence of even one prime is guaranteed. This led us to consider what T.~Tao's~\cite{Tao} beautiful theorem about arbitrarily shaped constellations in the Gaussian primes says about primes on Gaussian lines, and to formulate and computationally support an analog of Bertrand's Postulate to Gaussian lines. \medskip The real and imaginary lines contain infinitely many primes, so it is natural to wonder whether every primitive Gaussian line similarly contains infinitely many Gaussian primes. Finding even one other primitive Gaussian line that contains infinitely many Gaussian primes is a very difficult problem, however, since this is equivalent (by taking norms) to finding a quadratic polynomial that takes on infinitely many rational prime values and no such polynomials are known. For example, determining whether or not there are infinitely many Gaussian primes on the Gaussian line with $\alpha_0=1$ and $\delta=i$ ($i.e.$ Gaussian primes of the form $\alpha_n=1+ni$) is equivalent to determining whether or not there are infinitely many rational primes of the form $1+n^2$, which is Landau's fourth problem given at the 1912 International Congress of Mathematicians and remains open today. In general, it is also not known whether every irreducible quadratic polynomial attains at least one prime value, so similarly we cannot easily decide whether every primitive Gaussian line contains at least one Gaussian prime. Despite the difficulty of finding a Gaussian line that contains infinitely many Gaussian primes, we can apply a result of Iwaniec and Lemke Oliver to prove that infinitely many Gaussian lines contain infinitely many elements that are the product of at most two Gaussian primes. For example, it is a deep theorem of Iwaniec \cite{Iwaniec} that there are infinitely many values of $n$ such that $1+n^2$ is the product of at most two rational primes, from which it is immediate that the vertical Gaussian line defined by $\alpha_0=1$ and $\delta =i$ contains infinitely many elements that are the product of at most two Gaussian primes. Iwaniec notes that his proof generalizes to show that if $G(n)=An^2+Bn+C$ is an irreducible polynomial with $A>0$ and $C$ odd, then there exist infinitely many integers $n$ such that $G(n)$ has at most two rational prime factors. This theorem also follows from a result of Lemke Oliver~\cite{Oliver} generalizing Iwaniec's work. Applied to Gaussian lines, this result yields the following: \begin{theorem}\label{2primes} Let $L$ is a primitive Gaussian line such that $1+i $ does not divide $\alpha_0$. Then $L$ contains infinitely Gaussian integers that are the product of at most two Gaussian primes. \end{theorem} \begin{proof} Let $L$ be a primitive Gaussian line with $\alpha_0=a+bi$, $\delta=c+di$, and $\Delta=ad-bc$ as defined in (\ref{D}). Assume $1+i$ does not divide $\alpha_0$. The norm of an arbitrary Gaussian integer $\alpha_n$ on $L$ can be viewed as a quadratic polynomial $f(n)$ as follows: \begin{align}\label{N} f(n)=N(\alpha_n)=N(\alpha_0+\delta n) &=N(\delta)n^2+Tr(\alpha_0\overline\delta)n+N(\alpha_0)\\ &=(c^2+d^2)n^2+2(ac+bd)n+a^2+b^2\nonumber. \end{align} The discriminant of $f$ is equal to $-4\Delta^2$, which is negative unless $\Delta=0$. Thus, $f$ is irreducible over $\mathbb{Z}$ unless $L$ is the real or imaginary line, by Lemma \ref{delta}. Moreover, the leading coefficient of $f$ is positive and the constant term $N(\alpha_0)$ is odd, since we are assuming $1+i$ does not divide $\alpha_0$. It follows from Iwaniec's theorem discussed above that there are infinitely many $n$ such that $f(n)=N(\alpha_n)$ is a product of at most two rational primes, $i.e.$, $\alpha_n$ is a product of at most two Gaussian primes. \end{proof} Unfortunately Theorem \ref{2primes} does not say anything about the distribution of Gaussian primes on Gaussian lines. For this, we first apply T. Tao's \cite{Tao} astonishing theorem about arbitrarily shaped constellations in the Gaussian primes to Gaussian lines. His theorem says the following: \begin{theorem}[T. Tao \cite{Tao}] Given any distinct Gaussian integers $v_1,\ldots,v_{k}$, there are infinitely many sets $\{\alpha+rv_1,\ldots, \alpha+rv_{k}\}$, with $\alpha\in\mathbb{Z}[i]$ and $r\in\mathbb{Z}\setminus \{0\}$, all of whose elements are Gaussian primes.\end{theorem} By choosing $\delta=c+di\in\mathbb{Z}[i]$ with $\gcd(c,d)=1$ as usual, we can apply Tao's theorem with $v_1=\delta$, $v_2=2\delta,\ldots,v_{k}=k\delta$. The theorem guarantees the existence of infinitely many pairs $(\alpha, r)$ such that all the elements in the set $$P_{\alpha, r}=\{\alpha+r\delta,\alpha+2r\delta,\ldots, \alpha+kr\delta\}$$ are Gaussian primes. For each $\alpha$, there is a primitive Gaussian line $L_\alpha$ with slope $m=d/c$ ($i.e.$, $\delta=c+di$) that passes through all the elements in $P_{\alpha, r}$. Thus, $L_\alpha$ contains $k$ Gaussian primes in arithmetic progression. It is possible that infinitely many of the sets $P_{\alpha, r}$ are actually on the same Gaussian line (that is, infinitely many of the lines $L_\alpha$ have the same $\alpha_0$). In this case, we thus have a Gaussian line that contains infinitely many Gaussian primes. It follows that for a fixed slope $m\in\mathbb{Q}$, either there is a Gaussian line with slope $m$ that contains infinitely many Gaussian primes or, for all $k\geq 1$, there are infinitely many Gaussian lines with slope $m$ that contain $k$ Gaussian primes in arithmetic progression. Considering this for all $m$ and excluding the real and imaginary lines (the case $\alpha_0=0$), gives the following: \begin{corollary} At least one of the following two statements is true: \begin{enumerate} \item There is a Gaussian line with $\alpha_0\neq 0$ that contains infinitely many Gaussian primes. \item For every rational integer $m$ and every positive integer $k$, there are infinitely many distinct Gaussian lines with slope $m$ that contain $k$ Gaussian primes in arithmetic progression. \end{enumerate} \end{corollary} Note that if the first statement in the corollary is true, then by taking norms it is also true that there is a quadratic polynomial that takes on infinitely many prime values. Regarding the second statement, note that it is not possible for a Gaussian line to contain infinitely many Gaussian primes in arithmetic progression. This follows from the result of Gethner et al. \hskip-.05in \cite{Gethner} mentioned earlier that every Gaussian line contains arbitrarily long sequences of consecutive Gaussian composites. We also wondered where to look for primes on Gaussian lines. On the real line, Bertrand's Postulate guarantees the existence of a rational prime between $n$ and $2n$ for every rational integer $n\geq 3$. In other words, there exists a prime between $n$ and the next integer that is divisible by $n$. We wondered if the analogous statement holds on Gaussian lines. If $\alpha_n$ is on a Gaussian line $L$ then to characterize the next Gaussian integer on $L$ divisible by $\alpha_n$, we define a function $\nu:\mathbb{Z}[i]\rightarrow \mathbb{Z}$ by \begin{equation}\label{morm} \nu(x+iy) = \frac{N(x+iy)}{{\rm gcd}(x,y)}. \end{equation} The function $\nu$ is useful because if $\beta\in\mathbb{Z}[i]$ then the smallest positive rational integer divisible $\beta$ is $\nu(\beta)$, and furthermore, $\nu(\beta)$ divides every rational integer that is divisible by $\beta$. For example, if $\beta = 2+6i = 2(1+3i)$ then the smallest positive rational integer divisible by $\beta$ is $2(1+3i)(1-3i)=20=\nu(\beta)$, and $\beta$ divides a rational integer $r$ if and only if $r$ is divisible by 20. With regards to Bertrand's postulate, if $\alpha_n$ is on a Gaussian line $L$ then the next Gaussian integer on $L$ divisible by $\alpha_n$ is $\alpha_{n+\nu(\alpha_n)}=\alpha_{n}+\nu(\alpha_n)\cdot\delta$. Notice that $\nu(r)=r$ for all $r\in \mathbb{Z}$, so Conjecture \ref{bertrand2} below is equivalent to Bertrand's Postulate when $L$ is the real line. We include a second conjecture because $\alpha_{n+N(\alpha_n)}=\alpha_n+ N(\alpha_n)\cdot\delta$ is also divisible by $\alpha_n$ and, as we discuss below, it is more efficient to use the norm when searching for Gaussian primes on lines. Thus, we make the following two conjectures that generalize Bertrand's Postulate. \medskip \begin{conjecture} [Strong Bertrand for Gaussian lines] \label{bertrand2} Let $L$ be a primitive Gaussian line. If $n>1$, then there is always at least one Gaussian prime on $L$ that lies between $\alpha_n$ and $\alpha_{n+\nu(\alpha_n)}$. \end{conjecture} \begin{conjecture} [Weak Bertrand for Gaussian lines] \label{bertrand} Let $L$ be a primitive Gaussian line. If $n>1$, then there is always at least one Gaussian prime on $L$ that lies between $\alpha_n$ and $\alpha_{n+N(\alpha_n)}$. \end{conjecture} \medskip We wrote a program in Sage \cite{Sage} to search for lines $L$ where Conjecture~\ref{bertrand} fails for some Gaussian integer on $L$. We tested well over $10^{10}$ consecutive Gaussian integers on about $\numprint{700,000}$ lines and the conjecture held in every case. About $\numprint{607,000}$ of the lines we checked had $\alpha_0=1$ and $\delta=c+di$, where $c$ and $d$ were relatively prime integers ranging from one to 1,000. Additionally, we checked over 24,000 lines where $c$ and $d$ were random integers between 300 and $10^{18}$. Finally, we checked about 65,000 lines with $\alpha_0\neq 1$. Our algorithm for testing Conjecture \ref{bertrand} relies on the fact that if $\alpha_\ell=\pi$ is a Gaussian prime between $\alpha_n$ and $\alpha_{n+N(\alpha_n)}$ for some $0<n<\ell$, then $\pi$ is also between $\alpha_k$ and $\alpha_{k+N(\alpha_k)}$ whenever $n<k<\ell$. This holds because $N(\alpha_n)<N(\alpha_k)$ whenever $0<n<k$ by our choice of $\alpha_0$ being the element of smallest norm on $L$. The corresponding statement does not hold for $\nu(\alpha_n)$, which is why we focus on Conjecture~\ref{bertrand}. Specifically, for every line $L$ that we tested, we found a sequence of $10^{10}$ Gaussian integers $\alpha_{\ell_i}$, $1\leq i\leq 10^{10}$, on $L$ such that the following three conditions are satisfied: \begin{enumerate} \item Each $\alpha_{\ell_i}$ is a Gaussian prime; \item The first Gaussian prime in the sequence, $\alpha_{\ell_1}$, lies between $\alpha_1$ and $\alpha_{1+N(\alpha_1)}$, $i.e.$, $1<\ell_1<1+N(\alpha_1)$; \item For $i\geq 1$, the Gaussian prime $\alpha_{\ell_{i+1}}$ lies between the previous prime $\alpha_{\ell_i}$ and the Gaussian integer $\alpha_{{\ell_i}+N(\alpha_{\ell_i})}$ on $L$, $i.e.$, $\ell_i<\ell_{i+1}<1+N(\alpha_{\ell_i})$. \end{enumerate} This verifies Conjecture 2 on the line $L$ for all $1<n\leq\ell_{10^{10}}$. If either conjecture is true, then it would follow that there are infinitely many Gaussian primes on every Gaussian line. But proving either conjecture for even one Gaussian line (with $\alpha_0\neq 0$) would give a Gaussian line with infinitely many Gaussian primes, and hence a quadratic polynomial that takes on infinitely many rational prime values. \section{Divisibility on Gaussian Lines.}\label{div} Every second integer on the real line is divisible by 2, every third by 3, every fourth by 4, and so on. We wondered if this basic periodicity property of divisibility extends to Gaussian lines, and furthermore, if there is a simple way to characterize those Gaussian primes that occur as divisors on a particular Gaussian line. In this section we show that the answer to both of these questions is {\em YES}. \medskip Throughout this section, let $L$ be a primitive Gaussian line with $\alpha_0=a+bi$ and $\delta = c+di$ as defined in Section \ref{background}. Then $\alpha_0$ and $\delta$ are relatively prime Gaussian integers, $c$ and $d$ are relatively prime rational integers, and the Gaussian integers on $L$ are exactly the numbers $\alpha_n=\alpha_0+\delta n$, $n\in \mathbb{Z}$. Also, recall the function $\nu:\mathbb{Z}[i]\rightarrow \mathbb{Z}$ defined in (\ref{morm}) as it is used here and throughout the rest of the paper In the special case where $L$ is the real line, we have $\alpha_0=0$, $\delta=1$, and $\alpha_n=n$, $n\in\mathbb{Z}$. In this case, divisibility of integers on the line $L$ by a rational integer $r$ is periodic with period $r$. Our first theorem shows that this periodicity generalizes to arbitrary primitive Gaussian lines, specifically that divisibility by a Gaussian integer $\beta$ is periodic with period $\nu(\beta)$. Note that the periodicity of divisibility on the real line is a special case of the following theorem. \begin{theorem} \label{periodicity} Suppose $\beta\in\mathbb{Z}[i]$ divides some Gaussian integer $\alpha_t$ on $L$. Then $\beta$ divides $\alpha_{n}$ if and only if $n \equiv t \pmod {\nu(\beta)}$. \end{theorem} \begin{proof} Suppose $\beta $ divides $\alpha_{t}$ for some $t$. Then $\beta $ and $\delta$ are relatively prime, since any common divisor would also divide $\alpha_0= \alpha_t - \delta t$, but $\delta$ and $\alpha_0$ are relatively prime. Thus, $\beta$ divides $\alpha_n$ if and only if $\beta$ divides $\alpha_n-\alpha_t$, which in turn holds if and only if $\beta$ divides $n-t$ since $\alpha_n- \alpha_t = \delta(n-t)$. But $n-t \in \mathbb{Z}$, so $\beta$ divides $\alpha_n$ if and only if $\nu(\beta)$ divides $n-t$. \end{proof} Theorem \ref{periodicity} implies that consecutive Gaussian integers $\alpha_n$ and $\alpha_{n+1}$ on $L$ are always relatively prime over $\mathbb{Z}[i]$, just as consecutive rational integers on the real line are always relatively prime over $\mathbb{Z}$. Also, because Theorem \ref{periodicity} is about Gaussian integers that divide some element on $L$, a natural follow-up problem is to characterize those Gaussian integers that occur as divisors of elements on $L$. In this section we specialize to rational integer and Gaussian prime divisors, and in Section \ref{divisor} we give the complete characterization of the set of Gaussian integer divisors. We define the {\em divisor set of $L$}, denoted ${\mathcal{D}}(L)$, to be the set of Gaussian integers that divide some Gaussian integer on $L$. Our main result in Section \ref{divisor} (Theorem \ref{bigtheorem}) is a complete characterization of this set. Here we begin by characterizing two of its subsets, the {\em rational set} and the {\em Gaussian-prime set}, which we need for our work in Section~\ref{CRT}. We define the {\em rational set of $L$}, denoted $\mathcal{R}(L)$, to be the set of rational integers that divide some Gaussian integer on $L$, and the {\em Gaussian-prime set of $L$}, denoted ${\mathcal{G}\mathcal{P}}(L)$, to be the set of non-rational Gaussian primes that divide some Gaussian integer on $L$. For easy reference, below are the set theoretical definitions of these three sets for a given Gaussian line $L$: \begin{align*} \mathcal{R}(L) &=\{r \in \mathbb{Z} : r|\alpha_n \text{ for some } n \in \mathbb{Z} \}; \\ {\mathcal{G}\mathcal{P}}(L) &= \{ \pi\in\mathbb{Z}[i] : \pi\ \text{is a Gaussian prime, }\pi \not\in \mathbb{Z}, \text{ and }\pi|\alpha_n \text{ for some } n \in \mathbb{Z}\};\\ {\mathcal{D}}(L) &= \{ \beta\in\mathbb{Z}[i]:\beta|\alpha_n\ \text{for some } n \in \mathbb{Z} \}. \end{align*} Note that an element in any of these three sets does not necessarily lie on the line $L$, but simply divides some Gaussian integer that lies on $L$. In general, the divisor set ${\mathcal{D}}(L)$ of $L$ is not closed under multiplication. For example, suppose $1+2i$ divides $\alpha_0$ and $1-2i$ divides $\alpha_1$, so $1+2i, 1-2i\in{\mathcal{D}}(L)$. Since $\nu(1+2i)=\nu(1-2i)=5$, it follows from Theorem \ref{periodicity} that $1+2i$ and $1-2i$ both divide every fifth Gaussian integer on $L$, starting with $\alpha_0$ and $\alpha_1$ respectively. Thus, no integer on $L$ is divisible by their product $i.e.$, $(1+2i)( 1-2i)=5\notin{\mathcal{D}}(L)$, and ${\mathcal{D}}(L)$ is not closed under multiplication. Our first lemma shows that this type of restriction from Theorem~\ref{periodicity} is really the only property preventing the divisor set from being closed under multiplication. \begin{lemma} \label{ncrt} Let $\beta$ and $\gamma$ be in the divisor set ${\mathcal{D}}(L)$ of $L$. If $\nu (\beta)$ and $\nu (\gamma)$ are relatively prime, then $\beta \gamma$ is in ${\mathcal{D}}(L)$. \end{lemma} \begin{proof} Suppose $\beta,\gamma\in{\mathcal{D}}(L)$. Then, by Theorem \ref{periodicity}, there exist integers $s$ and $t$ such that $\beta|\alpha_n$ if and only if $n \equiv s \pmod{\nu (\beta)}$ and $\gamma|\alpha_n$ if and only if $n \equiv t \pmod{\nu (\gamma)}$. By the Chinese Remainder Theorem, there is an $n$ that satisfies both congruences. Therefore, $\beta \gamma \in {\mathcal{D}}(L)$. \end{proof} We use Lemma \ref{ncrt} to prove our next theorem and characterize the rational set of $L$. Recall from (\ref{D}) that $\Delta=ad-bc$ is a rational integer associated to $L$. \begin{theorem} \label{ZL} Let $r\in\mathbb{Z}$. Then $r$ is in the rational set $\mathcal{R}(L)$ of $L$ if and only if r divides~$\Delta$.\end{theorem} \begin{proof} Note that if $r,s\in\mathbb{Z}$ satisfy $rs\in\mathcal{R}(L)$, then $r\in\mathcal{R}(L)$ and $s\in\mathcal{R}(L)$ by the definition of the rational set. It follows from this and Lemma \ref{ncrt} that it is sufficient to prove Theorem \ref{ZL} for prime powers. Let $r=p^t$, where $p$ is a rational prime. Then $r\in\mathcal{R}(L)$ if and only if $p^t$ divides $\alpha_n$ for some $n\in\mathbb{Z}$. We have that $\alpha_n=\alpha_0+n\delta$, so ${\text{Re}}(\alpha_n)=a+nc$ and ${\text{Im}}(\alpha_n)=b+nd$. Thus, $p^t$ divides $\alpha_n$ if and only if $p^t$ divides both $a+nc$ and $b+nd$. Recall that $c$ and $d$ are relatively prime, so at least one of them is not divisible by $p$. With out loss of generality, we assume that $p$ does not divide $c$. Then $c$ has a multiplicative inverse modulo $p^t$. Thus we have: \begin{eqnarray*} p^t|\alpha_n &\Longleftrightarrow &a+nc\equiv 0\pmod{p^t} \quad {\rm and}\quad b+nd\equiv 0\pmod{p^t}\\ &\Longleftrightarrow &b\equiv -nd\pmod {p^t}, \text{ where } n\equiv -ac^{-1}\pmod {p^t}\\ &\Longleftrightarrow &b\equiv ac^{-1}d\pmod {p^t}\\ &\Longleftrightarrow &ad\equiv bc\pmod{p^t}\\ &\Longleftrightarrow &p^t|\,\Delta, \end{eqnarray*} as needed. \end{proof} Thus, the rational integers that divide some Gaussian integer $\alpha_n$ on $L$ are exactly the divisors of $\Delta$. Consequently, the rational set $\mathcal{R}(L)$ of $L$ is finite unless $\Delta=0$; that is, unless $L$ is the real or imaginary line. Our next theorem characterizes the Gaussian prime set of $L$ and shows, by contrast, that this set is always infinite. \begin{theorem} \label{PL} Let $\pi$ be a Gaussian prime with $\pi\not\in\mathbb{Z}$. Then $\pi\in{\mathcal{G}\mathcal{P}}(L)$ if and only if $\pi$ does not divide $\delta$. \end{theorem} \begin{proof} First suppose $\pi$ divides $\delta$. Then $\pi$ does not divide $\alpha_n=\alpha_0+\delta n$ for all $n$ since $\alpha_0$ and $\delta$ are relatively prime. Thus, $\pi \not \in {\mathcal{G}\mathcal{P}}(L)$ in this case. Conversely, suppose $\pi$ does not divide $\delta$. Let $\pi$ lie over the rational prime $p$. If $p$ divides $\Delta$, then $p$ divides some Gaussian integer $\alpha_n$ on $L$ by Theorem~\ref{ZL}. Thus $\pi$ also divides $\alpha_n$, and $\pi\in{\mathcal{G}\mathcal{P}}(L)$ as needed. Thus, from now on we assume $p$ does not divide $\Delta$, and show that $\pi\in{\mathcal{G}\mathcal{P}}(L)$ in this case as well. As in (\ref{N}), the norm of an arbitrary Gaussian integer $\alpha_n$ on $L$ can be viewed as a quadratic polynomial $$f(n)=N(\alpha_0+\delta n)=N(\delta)n^2+Tr(\alpha_0\overline\delta\,)n+N(\alpha_0),$$ with discriminant Disc$(f)=-4\Delta^2$. If $p\neq 2$, then $p\equiv 1\pmod 4$ since $\pi\notin\mathbb{Z}$. In this case, $-1$ is a square modulo $p$ and so Disc$(f)$ is a non-zero square modulo $p$. Therefore, $f(n)$ has two distinct roots modulo $p$, so there are $r,s\in\mathbb{Z}$, $r\not\equiv s\pmod p$, such that $N(\alpha_r)\equiv N(\alpha_s)\equiv 0\pmod p$. It follows from Theorem \ref{periodicity} that $\pi$ and $\overline\pi$ both divide exactly one of $\alpha_r$ and $\alpha_s$. Thus $\pi\in{\mathcal{G}\mathcal{P}}(L)$ in this case. If $p=2$, then Disc$(f)\equiv 0\pmod p$ and $f$ has a double root modulo $p$. It follows that $\pi$ divides either $\alpha_0$ or $\alpha_1$. Thus, $\pi\in{\mathcal{G}\mathcal{P}}(L)$ in this case as well. \end{proof} Since $\delta\neq 0$, it follows from Theorem \ref{PL} that the divisor set of a Gaussian line always contains infinitely many Gaussian primes. In particular, we have the following corollary to Theorem \ref{PL}. \begin{corollary} \label{minprimes} The divisor set ${\mathcal{D}}(L)$ of $L$ contains at least one Gaussian prime that lies over $p$ for every rational prime $p\equiv 1\pmod 4$. \end{corollary} \begin{proof} Let $\pi$ be a Gaussian prime that lies over the rational prime $p\equiv 1\pmod 4$. Suppose that neither $\pi$ nor $\overline{\pi}$ are in ${\mathcal{D}}(L)$. Then neither is in ${\mathcal{G}\mathcal{P}}(L)$ and so both divide $\delta$ by Theorem \ref{PL}. Thus $p$ divides $\delta$ and $p$ is a common divisor of $c$ and $d$, which contradicts $L$ being primitive.\end{proof} Taken together, Theorems \ref{ZL} and \ref{PL} imply that if a Gaussian prime $\pi$ divides $\delta$, and $\pi$ lies over $p$, then $p$ does not divide $\Delta$ (or, equivalently, $\pi$ does not divide $\Delta$). This can also be seen directly: If $\pi$ is a common divisor of both $\delta$ and $\Delta$, then $\pi$ divides $d$ since $d\alpha_0=\Delta+b\delta$ and $\alpha_0$ and $\delta$ are relatively prime. Now, $\delta=c+di$, so $\pi$ also divides $c$. But $c, d\in \mathbb{Z}$, so it follows that $p$ is a common divisor of $c$ and $d$, which contradicts $L$ being primitive. Theorems \ref{ZL} and \ref{PL} characterize the rational and Gaussian prime set of a Gaussian line. In Section \ref{divisor} we use these theorems to give a complete characterization of the divisor set as well. First we turn to some consequences of the theorems in this section. \section{The Chinese Remainder Theorem for Gaussian Lines.}\label{CRT} In this section we prove a theorem about Gaussian lines that is analogous to the Chinese Remainder Theorem for $\mathbb{Z}$. We also use the Chinese Remainder Theorem for $\mathbb{Z}[i]$ to prove that there are always infinitely many Gaussian lines that satisfy any given CRT-type divisibility properties. \medskip The Chinese Remainder Theorem (CRT) for $\mathbb{Z}$ implies that there will always be a solution to a system of linear congruences over $\mathbb{Z}$ when the moduli are pairwise relatively prime. It is well known that this theorem generalizes with the same proof to the Gaussian integers (or to any Euclidean domain). We state this more general version here since we will need it in our later work. \begin{theorem}[CRT for $\Z[i]$] \label{chinese} Let $\mu_1,\mu_2,\ldots, \mu_k$ be pairwise relatively prime Gaussian integers and $\beta_1, \beta_2,\ldots, \beta_k$ be arbitrary Gaussian integers. Then the system of $k$ congruences $$x \equiv \beta_j \pmod{\mu_j}, \ 1\leq j\leq k,$$ has a unique solution $\tau\in\mathbb{Z}[i]$ modulo $\mu_1\mu_2\ldots \mu_k$. \end{theorem} Note that CRT for $\mathbb{Z}$ is just Theorem \ref{chinese} with $\beta_j, \mu_j\in\mathbb{Z}$, $1\leq j\leq k$. In the spirit of this paper, we extend CRT for $\mathbb{Z}$ to CRT for Gaussian lines. First we restate CRT for $\mathbb{Z}$ in terms of divisibility since the analogous statement for Gaussian lines is given in terms of divisibility. \begin{theorem}[CRT for $\mathbb{Z}$] \label{crtZ} Let $m_1, m_2,\ldots, m_k$ be pairwise relatively prime rational integers and $b_1, b_2,\ldots, b_k$ be arbitrary rational integers. Then there is a unique rational integer $t$ modulo $m_1 m_2\cdots m_k$ such that $$m_1|{(t+b_1)}, \ m_2|{(t+b_2)}, \ \ldots, \ m_k|{(t+b_k)}.$$ \end{theorem} We use the function $\nu:\mathbb{Z}[i]\rightarrow\mathbb{Z}$ defined (\ref{morm}) to extend Theorem \ref{crtZ} to any Gaussian line. Since $\nu(n)=n$ for all $n\in\mathbb{Z}$, the following theorem is exactly CRT for $\mathbb{Z}$ when $L$ is the real line. \begin{theorem}[CRT for Gaussian lines] \label{crtGL} Let $L$ be a primitive Gaussian line, and suppose $\mu_1, \mu_2,\ldots, \mu_k$ are Gaussian integers in the divisor set ${\mathcal{D}}(L)$ of $L$ such that $\nu(\mu_1),\nu(\mu_2), \ldots, \nu(\mu_k)$ are pairwise relatively prime. Let $b_1, b_2,\ldots, b_k$ be arbitrary rational integers. Then there is a unique rational integer $t$ modulo $\nu(\mu_1)\nu(\mu_2) \cdots \nu(\mu_k)$ such that $$ \mu_1|\alpha_{t+b_1}, \ \mu_2|\alpha_{t+b_2}, \ \ldots, \ \mu_k|\alpha_{t+b_k}.$$ \end{theorem} \begin{proof} Since $\mu_j\in{\mathcal{D}}(L)$, $1\leq j\leq k$, it follows from Theorem \ref{periodicity} that for each $j$ there exists $m_j\in\mathbb{Z}$ such that $\mu_j$ divides the Gaussian integer $\alpha_{n}$ on $L$ if and only if $n \equiv m_j \pmod {\nu(\mu_j)}$. By Theorem \ref{crtZ}, the system of $k$ congruences $$x \equiv m_j-b_j \pmod {\nu(\mu_{j})}, \ 1\leq j\leq k,$$ has a unique solution $x\equiv t\pmod{\nu(\mu_1)\nu(\mu_2) \cdots \nu(\mu_k)}$. Thus, for $1\leq j\leq k$, we have $t+b_j\equiv m_j \pmod{\nu(\mu_j)}$ and $\alpha_{t+b_j}$ is divisible by $\mu_j$ as needed. \end{proof} Now, suppose you want to find a primitive Gaussian line that satisfies certain CRT-type divisibility properties. For instance, suppose you want a line where $2+i$ divides $\alpha_1$, $2+3i$ divides $\alpha_2$, and $4080 + 1397i$ divides $\alpha_3$. It follows from our next theorem that infinitely many such lines exist (one example in this case is the line defined by $\alpha_0=1$ and $\delta=6297+8234i$), and the proof shows how to construct them. \begin{theorem} \label{newchinese} Let $b_1, b_2, \ldots, b_k $ be rational integers and $\mu_1, \mu_2, \ldots, \mu_k$ be pairwise relatively prime Gaussian integers. Then there are infinitely many primitive Gaussian lines $L$ such that $\mu_j $ divides the Gaussian integer $ \alpha_{b_j}$ \hskip-.03in on $L$ for all $1 \leq j\leq k$. \end{theorem} \begin{proof} To show there are infinitely many primitive Gaussian lines $L$ that satisfy the desired divisibility conditions, we show that there are infinitely many Gaussian integers $\alpha_0=a+bi$ and $\delta=c+di$ that satisfy all of the following properties: \begin{quote} \begin{enumerate} \item[(a)] $N(\alpha_0+n\delta)>N(\alpha_0)$ for all $n\neq 0$, $n\in\mathbb{Z}$; \item[(b)] gcd$(c,d)=1$ and $c\geq 0$; \item[(c)] $\alpha_0$ and $\delta$ are relatively prime over $\Z[i]$; \item[(d)] $\mu_j $ divides $ \alpha_{b_j}\hskip-.035in=\alpha_0+b_j\delta$ for all $1 \leq j\leq k$. \end{enumerate} \end{quote} We first choose $\alpha_0$. For $1 \leq j\leq k$, let $\gamma_j\in\mathbb{Z}[i]$ be a common divisor of $\mu_j$ and $b_j$ with maximal norm (each $\gamma_j$ is uniquely defined up to multiplication by a unit in $\Z[i]$). Let $\lambda$ be any Gaussian integer that is relatively prime to both $\mu_1\mu_2\cdots\mu_k$ and $b_1b_2\cdots b_k$. Define $\alpha_0$ by $$\alpha_0=\lambda\prod_{j=1}^k \gamma_j\in\Z[i].$$ There are infinitely many possibilities for $\lambda$, so there are infinitely many possibilities for $\alpha_0$. For each $\alpha_0$, we show there are infinitely many $\delta=c+di\in\Z[i]$ such that the above properties (a)--(d) are satisfied. Property (d) is equivalent to $\delta$ being a solution to the system of $k$ congruences $$\alpha_0+b_j x\equiv 0\pmod{\mu_j}, \ \ 1 \leq j\leq k.$$ Dividing by $\gamma_j$ for each $j$, this is equivalent to $\delta$ being a solution to the system $$x\equiv -\left({\alpha_0\over\gamma_j}\right)\kappa_j^{-1} \pmod{\omega_j}, \ \ 1 \leq j\leq k,$$ where each $\kappa_j=b_j/\gamma_j\in\Z[i]$ is relatively prime to $\omega_j=\mu_j/\gamma_j\in\Z[i].$ Note that each $\alpha_0/\gamma_j$ is also relatively prime to $\omega_j$ since $\omega_1,\omega_2,\ldots,\omega_k$ are pairwise relatively prime. Thus, any solution to this latter system of congruences is relatively prime to the product $\omega_1\omega_2\cdots\omega_k$. Since $\delta$ will be a solution, we include an additional congruence to insure that any solution is also relatively prime to $\alpha_0$ and so property~(c) will automatically be satisfied. Let $\beta$ be the product of all the Gaussian primes that divide $\alpha_0$ but do not divide $\omega_1\omega_2\cdots\omega_k$, and let $\beta=1$ if no such Gaussian primes exist. Then $\delta$ is relatively prime to $\alpha_0$ if it is relatively prime to both $\beta$ and $\omega_1\omega_2\cdots\omega_k$. Thus, to insure that properties (c) and (d) are both satisfied, it is sufficient that $\delta$ be a solution to the following system of $k+1$ congruences: \begin{align*} x & \equiv 1\pmod\beta,\ \ {\text{and}} \\ x &\equiv -\left({\alpha_0\over\gamma_j}\right)\kappa_j^{-1} \pmod{\omega_j}, \ \ 1 \leq j\leq k. \end{align*} This system has a unique solution $\tau=r+si\in\Z[i]$ modulo $\beta\omega_1\omega_2\cdots\omega_k$ by CRT for Gaussian integers (Theorem \ref{chinese}) since $\beta,\omega_1,\omega_2,\ldots,\omega_k$ are pairwise relatively prime. Thus, it remains to construct $\delta=c+di$ that satisfies properties (a) and (b), and such that $\delta\equiv\tau\pmod{\beta\omega_1\omega_2\cdots\omega_k}$, so that properties (c) and (d) hold as well. To satisfy property (a), we construct $\delta=c+di$ such that $$N(\alpha_n)=N(\alpha_0+n\delta)=(c^2+d^2)n^2+2(ac+bd)n+a^2+b^2, \ \ n\in\mathbb{Z},$$ obtains its minimum value only when $n=0$. For any $c,d\in\mathbb{Z}$, the quadratic function, $$f(x)=(c^2+d^2)x^2+2(ac+bd)x+a^2+b^2, \ \ x\in\mathbb{R},$$ obtains its absolute minimum when $f^\prime(x)=0$, $i.e.$, when $x={-(ac+bd)/({c^2+d^2})}.$ Thus, since $f$ is symmetric, for property (a) to be satisfied and $f(0)$ to be the minimum integer value of $f$, it is sufficient that $c$ and $d$ satisfy \begin{equation}\label{1prime}-{1\over 2}<{{ac+bd}\over{c^2+d^2}}<{1\over 2}.\end{equation} For a fixed $d$, $$\lim_{c\rightarrow\infty}\left({{ac+bd}\over{c^2+d^2}}\right)=0,$$ so (\ref{1prime}) holds for all $c$ larger than some bound that depends on $d$. We use this fact to complete the proof. It is sufficient to choose $\delta=c+di$ such that (\ref{1prime}) holds, gcd$(c,d)=1$, $c\geq 0$, and $\delta\equiv\tau\equiv r+si\pmod{\beta\omega_1\omega_2\cdots\omega_k}$. Let $M=N(\beta\omega_1\omega_2\cdots\omega_k)\in\mathbb{Z}$. We first consider $s=0$. In this case, $\tau=r$ is relatively prime to $M$ since is a non-zero rational integer that is relatively prime to $\beta\omega_1\omega_2\cdots\omega_k$. It follows from Dirichlet's Theorem on Primes in Arithmetic Progressions, that there are infinitely many rational primes congruent to $r$ modulo $M$. Thus, we can choose a rational prime $p$ such that $p\equiv r\pmod{M}$, $p>M$, and $p$ is large enough so that (\ref{1prime}) holds for $c=p$ and $d=M$. Define $\delta$ by $\delta = p+Mi$. Then, (\ref{1prime}) holds and gcd$(c,d)=1$, since $p$ is prime and larger than $M$. Also, $\delta\equiv\tau\pmod{M}$, so $\delta\equiv\tau\pmod{\beta\omega_1\omega_2\cdots\omega_k}$, since $\beta\omega_1\omega_2\cdots\omega_k$ divides $M$. Thus, $\alpha_0$ and $\delta$ define a primitive Gaussian line that satisfies the divisibility conditions stated in Theorem~\ref{newchinese}. Moreover, according to Dirichlet's Theorem, there are infinitely many choices of the prime $p$. Thus, there are infinitely many choices for $\delta$, and so infinitely many primitive Gaussian lines with the same $\alpha_0$ that satisfy the conditions. Similarly, if $r=0$, then $\tau=si$, where $s$ is a non-zero rational integer that is relatively prime to $M$. Proceed as above to get a rational prime $p\equiv s\pmod{M}$, $p>M$, and $p$ is large enough so that (\ref{1prime}) holds for $c=M$ and $d=p$. Define $\delta$ by $\delta=M+pi$. Then $\alpha_0$ and $\delta$ define a primitive Gaussian line that satisfies the divisibility conditions stated in Theorem \ref{newchinese}. Again, by Dirichlet's Theorem, there are infinitely many choices of the prime $p$ and thus infinitely many primitive Gaussian lines with this $\alpha_0$ that satisfy these conditions. Finally, suppose $r$ and $s$ are both non-zero rational integers. Let $h$ be the smallest positive rational divisor of $r$ such that gcd$(r/h,M)=1$. Again by Dirichlet's Theorem, we can find a rational prime $p>s$ such that $p\equiv r/h\pmod M$ and $p$ is large enough so that (\ref{1prime}) holds for $c=ph$ and $d=s$. Define $\delta$ by $$\delta=ph+si.$$ Then $\delta\equiv\tau\pmod{\beta\omega_1\omega_2\cdots\omega_k}$. To see that gcd$(ph,s)=1$, first observe that gcd$(p,s)=1$ since $p>s$ is prime. Also, gcd$(h,s)=1$, since any common rational prime divisor $q$ of $h$ and $s$ is also a common divisor of $\tau$ and $M$. Hence, there is a Gaussian prime that lies over $q$ that divides both $\tau$ and $\beta\omega_1\omega_2\cdots\omega_k$, which is a contradiction since they are relatively prime. Thus, as above, $\alpha_0$ and $\delta$ define a primitive Gaussian line that satisfies the required divisibility conditions, and again there are infinitely many choices of $\delta$ by Dirichlet's Theorem. \end{proof} \section{The Divisor Set of a Gaussian Line.}\label{divisor} We now return to questions about divisibility on Gaussian lines related to those discussed in Section \ref{div}. For a given Gaussian line $L$, we first characterize those Gaussian-prime powers that exactly divide some Gaussian integer on $L$. Using this, our main theorem in this section gives a complete characterization of the divisor set ${\mathcal{D}}(L) $ of $L$. \medskip Theorem \ref{PL} in Section \ref{div} resolves the question of which Gaussian primes occur in the divisor set ${\mathcal{D}}(L)$ of $L$, but does not address division by prime powers. For example, Theorem~\ref{PL} does not answer the following question: If $\pi\in{\mathcal{D}}(L)$, then is $\pi^{100}$ guaranteed to be in ${\mathcal{D}}(L)$? Nor does it say anything about which prime powers $\pi^k$ {\em exactly divide} some Gaussian integer $\alpha_n$ on $L$ ($i.e.$, $\pi^k$ divides $\alpha_n$, but $\pi^{k+1}$ does not). For example, if $\pi^{50}\in{\mathcal{D}}(L)$, then certainly $\pi, \pi^2,\ldots,\pi^{49}\in{\mathcal{D}}(L)$, but is $\pi$ guaranteed to exactly divide some Gaussian integer on $L$? What about $\pi^2$ or $\pi^3$ or any other power of $\pi$? Our next theorem shows that the answer to all of these questions is {\em YES} whenever $\pi$ lies over a rational prime $p\equiv 1\pmod 4$, but is conditional for other values of $p$. We restrict to lines with $\Delta\neq 0$ since this simplifies the proof and exact division by all prime powers holds on the real and imaginary lines. \begin{theorem} \label{exact} Let $L$ be a primitive Gaussian line with $\Delta\neq 0$. Suppose $\pi$ is a Gaussian prime that lies over the rational prime $p$. \begin{enumerate} \item If $p\equiv 1\pmod 4$, then the following are equivalent: \begin{enumerate} \item $\pi$ does not divide $\delta$. \item $\pi^k\in{\mathcal{D}}(L)$ for some positive integer $k$. \item For every positive integers $r$, $\pi^r$ exactly divides some Gaussian integer on $L$. In particular, $\pi^r\in{\mathcal{D}}(L)$ for all positive integers $r$. \end{enumerate} \item If $p=2$, then the following are equivalent: \begin{enumerate} \item $1+i$ does not divide $\delta$. \item $(1+i)^k\in{\mathcal{D}}(L)$ for some positive integer $k$. \item Let $2^s$ be the exact power of $2$ that divides $\Delta$, and $\beta\in\Z[i]$ have 2-power norm. Then $\beta$ exactly divides some Gaussian integer $\alpha_n$ on $L$ if and only if $\beta$ is an associate of $2, 2^2, \ldots, 2^s$, or $2^{s}(1+i)$. That is, $(1+i)^t\in{\mathcal{D}}(L)$ if and only if $0\leq t\leq 2s+1$, but $(1+i)^t$ exactly divides a Gaussian integer on $L$ if and only if in addition $t$ is even or $t=2s+1$. \end{enumerate} \item If $p\equiv 3\pmod 4$ (so $\pi$ is an associate of $p$), then $p^k$ exactly divides some Gaussian integer $\alpha_n$ on $L$ if and only if $p^k$ divides $\Delta$. \end{enumerate} \end{theorem} \begin{proof} We consider the three cases separately. \medskip \noindent {\underline{\em{Case 1}}:} Suppose $p\equiv 1\pmod 4$. Statements 1(a) and 1(b) are equivalent by Theorem \ref{PL}. Since 1(c) trivially implies 1(b), we only need to show that 1(b) implies 1(c). For this, suppose that $\pi^k\in{\mathcal{D}}(L)$, say $\pi^k$ divides $\alpha_m$. Then $\pi^h$ exactly divides $\alpha_m$ for some $h\geq k$. Let $r$ be a positive integer. If $r< h$, then 1(c) holds since $\pi^r$ exactly divides $\alpha_n$ for $n=m+p^rq$, where $q$ is any integer not divisible by $p$. To see this, write $\alpha_n=\alpha_0+(m+p^rq)\delta=\alpha_m+p^rq\delta$, and use that $\pi^h$ exactly divides $\alpha_m$ while $\pi^r$ exactly divides $p^rq\delta$. Note that by considering the special case where $r=1$, this shows in general that if a Gaussian prime $\pi$ does not divide $\delta$ then $\pi$ {\em exactly} divides some Gaussian integer $\alpha_n$ on $L$. We use induction and the general fact for $r=1$ given above to show that 1(c) holds for $r\geq h$ as well. If $r=h$, then $\pi^r$ exactly divides $\alpha_m$ by hypothesis, so 1(c) holds in this case. Suppose it holds for some $t\geq h$, say $\pi^t$ exactly divides $\alpha_s$. Let $\omega=\alpha_s/\pi^t\in\Z[i]$. For $q\in\mathbb{Z}$, consider $$\alpha_{s+p^{t}q}=\alpha_s+p^{t}q\delta=\pi^t(\omega+\overline\pi^{t}\delta q),$$ where $p=\pi\overline\pi$ and $\overline\pi$ is not an associate of $\pi$ since $p\equiv 1\pmod 4$. Now, $\overline{\pi}^{t}\delta$ has no rational integer divisors since $\pi\nmid\delta$. Also, $\omega$ and $\overline{\pi}^{t}\delta$ are relatively prime since $\alpha_s$ and $p^t\delta$ are relatively prime. Thus, the numbers $\omega+\overline\pi^{t}\delta q$, $q\in\mathbb{Z}$, are the Gaussian integers on a different primitive Gaussian line $L^\prime$ with $\delta^\prime=\overline\pi^{t}\delta$. Since $\pi \nmid \delta^\prime$, it follows from the general result for $r=1$, that there is a $q_0\in\mathbb{Z}$ such that $\pi$ exactly divides the Gaussian integer $\omega+\overline\pi^{t}\delta q_0$ on $L^\prime$. Thus $\pi^{t+1}$ exactly divides the Gaussian integer $\alpha_n$ on $L$ for $n=s+p^{t}q_0$, and 1(c) holds for $r=t+1$. By induction it holds for all $r$. \medskip \noindent {\underline{\em{Case 2}}:} Suppose $p=2$. As above, it is sufficient to prove statement 2(b) implies 2(c). Suppose $(1+i)^k\in{\mathcal{D}}(L)$ for some positive integer $k$. Then $(1+i)\in{\mathcal{D}}(L)$ and $1+i$ does not divide $\delta$. Let $2^s$, $s\geq 0$, be the exact power of $2$ that divides $\Delta$. Then $2^s\in{\mathcal{D}}(L)$, but $2^{s+1}\not\in{\mathcal{D}}(L)$ by Theorem \ref{ZL}. Since $2$ ramifies in $\Z[i]$, this is equivalent to $(1+i)^{2s}\in{\mathcal{D}}(L)$, but $(1+i)^{2s+2}\not\in{\mathcal{D}}(L)$. We first claim that $(1+i)^{2s+1}\in{\mathcal{D}}(L)$. For this, note that since $(1+i)^{2s}\in{\mathcal{D}}(L)$, there is a Gaussian integer $\alpha_m$ on $L$ such that $(1+i)^{2s}$ divides $\alpha_m$. If $(1+i)^{2s+1}$ divides $\alpha_m$ then $(1+i)^{2s+1}\in{\mathcal{D}}(L)$ as claimed. So suppose $(1+i)^{2s+1}$ does not divide $\alpha_m$. By Theorem \ref{periodicity}, $(1+i)^{2s}$ divides $\alpha_{m+2^s}$ since $\nu((1+i)^{2s})=2^s$. Now, $$\alpha_{m+2^s}=\alpha_m+2^s\delta=2^s(\omega+\delta),$$ where $\omega=\alpha_m/2^s\in\Z[i]$ is not divisible by $1+i$. Since neither $\omega$ nor $\delta$ is divisible by $1+i$, their sum must be divisible by $1+i$. Thus, $(1+i)^{2s+1}$ divides $\alpha_{m+2^s}$, and $(1+i)^{2s+1}\in{\mathcal{D}}(L)$ in this case as well. Thus we have $(1+i)^t\in{\mathcal{D}}(L)$ if and only if $0\leq t\leq 2s+1$, and so it remains to consider exact division by $(1+i)^t$. We consider $t$ even and $t$ odd separately. First suppose that $t=2h$ is even. We claim that $(1+i)^t$ exactly divides some Gaussian integer $\alpha_n$ on $L$, or equivalently, that $2^h$ divides $\alpha_n$ but $2^h(1+i)$ does not. This is true when $t=s$ since $(1+i)^{2s}$ exactly divides either $\alpha_m$ or $\alpha_{m+2^s}$ by the preceding paragraph. So suppose that for some $h$, $0<h\leq s$, we have $(1+i)^{2h}$ exactly divides $\alpha_n$ for some $n$. Consider, $$\alpha_{n+2^{h-1}}=\alpha_n+2^{h-1}\delta=2^{h-1}(\omega+\delta),$$ where $\omega=\alpha_n/2^{h-1}\in\Z[i]$ is divisible by $(1+i)^2$ and $\delta$ is not divisible by $1+i$. Thus, $\omega+\delta$ is not divisible by $1+i$ , and so $(1+i)^{2h-2}$ exactly divides $\alpha_{n+2^{h-1}}$. The claim for odd $t$ follows by induction. Now suppose $t$ is odd and $(1+i)^t$ exactly divides some Gaussian integer $\alpha_r$ on $L$. For instance, this holds for $t=2s+1$ since $(1+i)^{2s+1}$ exactly divides either $\alpha_m$ or $\alpha_{m+2^s}$. Write $t=2j+1$, so $\nu((1+i)^t)=2^{j+1}$. Thus, by Theorem \ref{periodicity}, $(1+i)^t$ divides $\alpha_n$ for $n=r+2^{j+1}q$, $q\in\mathbb{Z}$. Now, $$\alpha_n=\alpha_{r+2^{j+1}q}=\alpha_r+2^{j+1}\delta q=(1+i)^t\left( \omega+\mu(1+i)\delta q\right),$$ where $\omega=\alpha_r/(1+i)^t\in\Z[i]$ is not divisible by $1+i$ and $\mu\in\Z[i]$ is a unit. Now, the real and imaginary parts of $\mu(1+i)\delta$ must be relatively prime since $1+i$ does not divide $\delta$ and the real and imaginary part of $\delta$ are relatively prime. Also, $\omega$ and $\mu(1+i)\delta$ are relatively prime over $\Z[i]$ since $1+i$ does not divide $\omega$ and $\alpha_r$ and $\delta$ are relatively prime. Thus, the numbers $\omega+(1+i)\delta q$, $q\in\mathbb{Z}$, are the Gaussian integers on a different primitive Gaussian line $L^\prime$ with $\delta^\prime=(1+i)\delta$. Since $1+i$ divides $\delta^\prime$, it follows from Theorem \ref{ZL} that none of the Gaussian integers $\omega+(1+i)\delta q$ are divisible by $1+i$, that is, $(1+i)\not\in{{\mathcal{D}}(L^\prime)}$. Thus, $(1+i)^{t+1}\not\in{\mathcal{D}}(L)$, or equivalently, $2^{j+1}\not\in{\mathcal{D}}(L)$. This is a contradiction unless $j=s$. Therefore, if $t$ is odd then $(1+i)^t$ exactly divides some Gaussian integer on $L$ if and only if $t=2s+1$. \medskip \noindent {\underline{\em{Case 3}}:} Suppose $p\equiv 3\pmod 4$. Then $p$ remains prime in $\Z[i]$ and $\pi$ is an associate of $p$. By Theorem \ref{ZL}, we know that then $p^k$ divides some Gaussian integer $\alpha_n$ on $L$ if and only if $p^k$ divides $\Delta$. For {\em exact} divisibility, let $s$ be such that $p^s$ exactly divides $\Delta$. Then $p^s$ exactly divides some $\alpha_m$ on $L$ since $p^{{s+1}}\not\in{\mathcal{D}}(L)$. Then, as in the case $p=2$, we have that $p^{s-1}$ exactly divides $$\alpha_{m+p^{s-1}}=\alpha_m+p^{s-1}\delta=p^{s-1}\left( \omega+\delta \right),$$ since $\omega=\alpha_m/p^{s-1}$ is divisible by $p$ but $\delta$ is not. Continue in the same way to get that $p^k$ exactly divides some Gaussian integer on $L$ for all $k$ with $0\leq k\leq s$. \end {proof} Putting Theorem \ref{exact} together with the results in Section \ref{div} yields a characterization of the divisor set ${\mathcal{D}}(L)$ of $L$ as follows. \begin{theorem}\label{bigtheorem} Let $L$ be a primitive Gaussian line with $\Delta\neq 0$. A Gaussian integer $\beta$ is in the divisor set ${\mathcal{D}}(L)$ of $L$ if and only if $\beta$ can be written as $$\beta= \mu r(1+i)^t\pi_1^{k_1}\pi_2^{k_2} \cdots \pi_m^{k_m},$$ where the variables in this expression are defined as follows: \begin{enumerate} \item[(a)] $\mu\in\{\pm1, \pm i\}$ is a unit in $\Z[i]$; \item[(b)] $r$ is a rational integer that divides $\Delta$; \item[(c)] $t=0$ if $1+i$ divides $\delta$, and $t\in \{0, 1\}$ otherwise; \item[(d)] For $1\leq j\leq m$, $\pi_j$ is a Gaussian prime such that $\pi_j$ does not divide $\delta$, $N(\pi_j)\neq 2$, and $N(\pi_j)\neq N(\pi_n)$ for $j\neq n$; \item[(e)] For $1\leq j\leq m$, $k_j \geq 0$ is a rational integer. \end{enumerate} \end{theorem} \begin{proof} By Lemma \ref{ncrt}, it is sufficient to characterize those $\beta\in{\mathcal{D}}(L)$ where $\nu(\beta)$ is a prime power. Thus, let $p$ be a rational prime and $\beta\in\Z[i]$ satisfy $\nu(\beta)=p^n$ for some positive integer $n$. First suppose $p\equiv 1\pmod 4$, and let $\pi$ be a Gaussian primes that lies over $p$. We may assume $\pi\in{\mathcal{G}\mathcal{P}}(L)$ by Corollary \ref{minprimes} of Theorem \ref{PL}. If $\overline\pi\not\in{\mathcal{G}\mathcal{P}}(L)$, then by Theorems \ref{ZL} and~\ref{exact}, $\beta\in{\mathcal{D}}(L)$ if and only if $\beta = \mu p^t\pi^k$, where $\mu$ is a unit in $\Z[i]$ and $t$ and $k$ are non-negative integers, and $p^t$ divides $\Delta$. If, in addition, $\overline\pi\in{\mathcal{G}\mathcal{P}}(L)$, then $\beta$ can also be of the form $\mu p^t\overline\pi^k$. If $p=2$ then, up to associates, $1+i$ is the only Gaussian prime that lies over $p$. Let $2^s$ be the power of $2$ that exactly divides $\Delta$. It follows from Theorem \ref{exact} that $\beta\in{\mathcal{D}}(L)$ if and only if $\beta = \mu 2^r(1+i)^t$, where $\mu$ is a unit in $\Z[i]$, $0\leq r\leq s$, and $t=0$ if $1+i$ divides $\delta$ and $t\in \{0, 1\}$ otherwise. Finally, if $p\equiv 3\pmod 4$, then $p$ remains prime in $\Z[i]$. In this case, it follows from Theorem \ref{ZL} that $\beta\in{\mathcal{D}}(L)$ if and only if $\beta=\mu p^r$, where $\mu$ is a unit in $\Z[i]$ and $p^r$ divides $\Delta$. \end{proof}

https://arxiv.org/abs/0909.3822

A derivation of Benford's Law ... and a vindication of Newcomb

We show how Benford's Law (BL) for first, second, ..., digits, emerges from the distribution of digits of numbers of the type $a^{R}$, with $a$ any real positive number and $R$ a set of real numbers uniformly distributed in an interval $[ P\log_a 10, (P +1) \log_a 10) $ for any integer $P$. The result is shown to be number base and scale invariant. A rule based on the mantissas of the logarithms allows for a determination of whether a set of numbers obeys BL or not. We show that BL applies to numbers obtained from the {\it multiplication} or {\it division} of numbers drawn from any distribution. We also argue that (most of) the real-life sets that obey BL are because they are obtained from such basic arithmetic operations. We exhibit that all these arguments were discussed in the original paper by Simon Newcomb in 1881, where he presented Benford's Law.

\section{Introduction.} Benford's Law (BL) asserts that in certain sets of numbers, most of them of real-life origin, the first digit is distributed non-uniformly in the form \begin{equation} P_B^{(1)}(d) = \log_{10} \left( 1 + \frac{1}{d} \right) , \label{BL} \end{equation} where $d$ is the first digit of the number and $\log_{10}$ is the logarithm base 10. In other words, $P_B^{(1)}(d)$ is the fraction of the numbers with first digit $d$ in the given set. There are also forms of Benford's Law for second, third, etc., digits, namely $P_B^{(n)}(d)$. Table 1 shows the values of $P_B^{(1)}(d)$ for $d = 1, 2, \dots , 9$. \begin{table}[htdp] \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|} \hline \hline $d$ & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 \\ \hline $P_B^{(1)}$ & $\> 0.3010 \>$& $\> 0.1761 \>$& $\> 0.1249 \>$& $\> 0.0969 \>$ & $\> 0.0792 \>$& $\> 0.0669 \>$ & $\> 0.0580 \>$ &$\> 0.0512 \>$& $\> 0.0458\>$ \\ \hline \hline \end{tabular} \caption{First digit Benford law.} \end{table} BL has been found to be obeyed quite well in a variety of situations, many of them checked by Franck Benford himself\cite{Benford}. These sets include population census, stock markets indeces, utilities bills, tax returns, areas of rivers, physical and mathematical constants, and molecular weights, among others\cite{Benford,Fewster}. At first sight, the Law is certainly baffling and counterintuitive\cite{Lines} since one's naive intuition is that digits of numbers should be uniformly or randomly distributed. Although Franck Benford has been credited with the law for his work of 1938\cite{Benford}, the law was originally discovered by the astronomer Simon Newcomb in 1881\cite{Newcomb} as a follow up of the observation that the pages of tables of logarithms in his university library were worn out following BL, as given by equation (\ref{BL}). What is rarely told is that Newcomb {\it derived} Benford's Law. His demonstration for us may now look obscure, and probably just sketchy, because he used arguments that were not so difficult to those familiar with concepts of log tables ... and certainly we are not. We shall advance a plausible explanation of Newcomb's observation of the worn pages of the log tables and argue why many sets of real-life origin also obey BL; alas, this argument was also used by Newcomb. We shall first prove a general result that appears to be known already\cite{Fewster,Hill,Pietronero,Raimi,others}, although to the best of our knowledge it has not been shown explicitely in the form here presented; we shall see that yields exactly all digit's Benford's distributions, allowing also for concluding that it is scale and number base invariant\cite{Hill}. We demonstrate that if $R$ is a set of real numbers uniformly distributed, then, the distributions of digits of $a^R$ obey BL for any real positive number $a$. Then, we discuss the main result of Newcomb, namely, the fact that a given set of numbers obeys BL if the mantissas of their logarithms are uniformly distributed. We then analyze two main type of sequences of numbers that obey BL, those that are obtained from multiplication of numbers drawn from any distribution and those that are part of a geometric progression of numbers uniformly distributed in an arbitrary interval. \section{A general result concerning Benford's Law.} Let $\{R_1, R_2, \dots , R_N\}$ be a sequence of real numbers drawn from a uniform distribution in the interval $R_i \in \left[ \left. P \log_a 10, (P +1) \log_a 10 \right) \right.$, with $P$ any integer. Then, the first, second, ..., digit distributions of the sequence $\{a^{R_1}, a^{R_2}, \dots , a^{R_N}\}$, with $a$ any real positive number, approaches Benford's Law, Eq.(\ref{BL}) and its generalizations, as $N \to \infty$ . Let us look first at the first digit distribution. In Fig. \ref{GR-2} we plot $a^R$ vs $R$ in a semi-log (base $a$) scale. In this graph, $a^R$ vs $R$ appears as a straight line. Now take in the $R$-axis the sequence $\{R_1, R_2, \dots , R_N\}$ within the interval $\left[ \left. P \log_a 10, (P +1) \log_a 10 \right) \right.$. Take any number of the sequence, say $R_i$. Then, in the logarithmic scale, $\log_a a^{R_i}$ must lie within any of the following ``bins": $b_1$, the interval $[\log_{a}\left(1\times 10^P\right), \log_{a} \left(2\times 10^P\right) )$; or $b_2$, the interval $[ \log_{a}\left(2 \times 10^P\right), \log_{a} \left( 3 \times 10^P\right) )$; $\dots$; or, $b_9$, the interval $[ \log_{a} \left(9 \times 10^P\right), \log_{a} \left( 10 \times 10^P\right) )$. The main point is this: if $ \log_{a} a^{R_i}$ lies within the bin $b_d$, then the first digit of $a^{R_i}$ is $d$. \begin{figure}[ htbp!] \centering \includegraphics[width=0.5\textwidth,keepaspectratio]{fig1.pdf} \caption{$a^R$ vs $R$ in semi-log scale. The dotted line shows an example of an arbitrary point $R_i$ in the chosen interval, such that the first digit of $a^{R_i}$ is 2 because it falls in bin $b_2$.} \label{GR-2} \end{figure} Since $R_i$ was drawn from a uniform distribution, it has the same chance to take any value within the interval $\left[ \left. P\log_a 10, (P +1) \log_a 10 \right) \right.$ and, therefore, the probability of $ \log_a a^{R_i}$ to fall within the bin $b_d$ is the length of the bin $b_d$ divided by the length of the full interval, namely \begin{eqnarray} P\left[ \log_a a^{R_i} \in b_d \right] &=& \frac{\log_{a} \left((d + 1) \times 10^P \right) - \log_{a} \left(d \times 10^P \right)}{\log_{a} \left(10^{P+1} \right) - \log_{a} \left(10^P \right)} \nonumber \\ &=& \frac{ \log_{a} \left(\frac{1 + d}{d} \right)}{\log_{a} 10} \nonumber \\ & = & \log_{10} \left( 1 + \frac{1}{d} \right) .\label{BL1} \end{eqnarray} This has the form of Benford's Law for the first digit $P_B^{(1)}(d)$, Eq. (\ref{BL}). Thus, as $N \to \infty$, the first digit distribution of the sequence $\{a^{R_1}, a^{R_2}, \dots , a^{R_N}\}$ will approach $P_B^{(1)}(d)$. We shall call this the {\it General Result} (GR). Note that GR is independent of the integer value of $P$ of the interval as long as the sequence is uniformly distributed. Clearly, the result holds if we change the interval to $\left[ \left. P \log_a 10, (P + M) \log_a 10 \right) \right.$ with $M$ any integer. Note that we never used the fact that neither $a$, nor $R$, nor $d$, are numbers base 10; the graph in Fig. \ref{GR-2} is plotted for numbers base 10 for illustration purposes, but the result would have been the same for any number base. Thus, we conclude that BL is base invariant, i.e. valid for any number base $K$, with $d = 0, 1, 2, \dots, K -1$. The second, third, ..., digit distributions follow right away with the same argument. For instance, the probability that the second digit of $a^{R_i}$ is $d$, equals the sum of the lenghts of the ``sub-bins", \begin{eqnarray} b_{1d} & = & \left[\log_{a}\left((1 + \frac{d}{10}) \times 10^P\right), \log_{a} \left((1 + \frac{ d+1}{10}) \times 10^P\right) \right) \nonumber \\ b_{2d} & = & \left[\log_{a}\left((2 + \frac{d}{10}) \times 10^P\right), \log_{a} \left((2 + \frac{ d+1}{10}) \times 10^P\right) \right)\nonumber \\ & \vdots & \nonumber \\ b_{9d} & = & \left[\log_{a}\left((9 + \frac{d}{10}) \times 10^P\right), \log_{a} \left((9 + \frac{ d+1}{10}) \times 10^P\right) \right)\nonumber \end{eqnarray} where $d$ can now take all values $0, 1, \dots, 9$. Thus, the second digit distribution is, \begin{eqnarray} P_B^{(2)}(d) &=& \frac{1}{\log_{a} \left(10^{P+1} \right) - \log_{a} \left(10^P \right)} \sum_{m=1}^9 \left[ \log_{a} \left((m + \frac{ d+1}{10}) \times 10^P \right) - \log_{a} \left((m + \frac{ d}{10}) \times 10^P \right) \right] \nonumber \\ &=& \frac{1}{\log_{a} \left(10^{P+1} \right) - \log_{a} \left(10^P \right)} \sum_{m=1}^9 \log_{a}\frac{ 10m + d + 1}{10 m + d } \nonumber \\ & = & \sum_{m=1}^9 \log_{10} \left( 1 + \frac{ 1}{10 m + d } \right). \label{BL2} \end{eqnarray} The argument is easily generalized to the $n$-th digit and the result is, \begin{equation} P_B^{(n)}(d) = \sum_{m=10^{n-2}}^{10^{n-1}-1} \log_{10} \left( 1 + \frac{ 1}{10 m + d } \right). \label{BLn} \end{equation} The present derivation is extremely simple. Although Newcomb never wrote the formulas for $P_B^{(n)}(d)$, given his mastery of log tables and numerical analysis\cite{Newcomb-PT}, it is clear that he knew them since he writes the values of $P_B^{(1)}(d)$ and $P_B^{(2)}(d)$ explicitely and mentions how $P_B^{(3)}(d)$ and $P_B^{(4)}(d)$ behave (the latter are almost uniform). Due to Newcomb's most important result, as we discuss in the next section, it seems to the author that he knew a derivation very similar to this one. Benford's Law and its generalizations have been rigorously shown by Hill\cite{Hill} to follow as a consequence of base invariance of the underlying law. We have no pretense of such a mathematical rigour here, but rather to show its simplicity to a wider audience. \subsection{Scale invariance.} A very important property of BL that follows from GR above is the fact that BL is scale invariant\cite{Pietronero}. Add to the values $R_i$ any constant value $c$. This is equivalent to consider a uniform sequence of numbers $R_i$ in the interval $\left[ \left. c + P \log_a 10, c + (P +1) \log_a 10 \right) \right.$. Referring to Fig. \ref{GR-2} one can see that in the semi-log graph this also amounts to shift the interval in the ordinate by a constant amount, $a^c$; one also sees, however, that the sizes of the bins $b_n$ remain unchanged. Thus, the sequence $\{a^c a^{R_1}, a^c a^{R_2}, \dots, a^c a^{R_N}\}$ also obeys BL. But this new sequence is the same as the original one $\{a^{R_1}, a^{R_2}, \dots , a^{R_N}\}$ multiplied by a constant, arbitrary, factor $a^c$. \subsection{The mantissa rule.} The sequences or sets of numbers that usually follow BL are not of the form $\{ a^R \}$. So, one can enquiry for a rule that tells us if some sequences do follow BL or not. The answer was also given by Newcomb in his two-page paper\cite{Newcomb}. We shall demonstrate now that for a given sequence of numbers $\{A_1, A_2, \dots , A_N\}$, if the mantissas of the logarithms of $A_i$, namely of $\{ \log_{10} A_1, \log_{10} A_2, \dots , \log_{10} A_N \}$, are uniformly distributed, then the sequence $\{A_1, A_2, \dots , A_N\}$ obeys BL. Before we give the demonstration, we note that GR can be restated much simpler for the case $a = 10$, namely, if a sequence of numbers $R_i$ are uniformly distributed in the interval $[0, 1)$, the sequence $10^{R_i}$ follows BL. We use this form below. The demonstration can be done writing the log of $A_i$ as, \begin{equation} \log_{10} A_i = C(A_i) + m(A_i) , \end{equation} where $C(A_i)$ is an integer, the so-called {\it characteristic} of the log, and $m(A_i)$ the fractional part of the logarithm or {\it mantissa}. Note that by definition the mantissas of logarithms base 10 are within the interval $[0, 1)$. It is clear, then, that when taking the ``antilogarithm" $10^{\log_{10} A_i} = 10^{C(A_i)} 10^{m(A_i)}$ the digits of $A_i$ will be determined only by $10^{m(A_i)}$ since the factor $10^{C(A_i)}$ just determines the position of the decimal point. Thus, the distribution of digits is determined by considering the sequence of the mantissas only, namely of the sequence $\{ m( A_1),m(A_2), \dots , m(A_N )\}$. Hence, if the latter are uniformly distributed, by GR the sequence $\{10^{m(A_1)}, 10^{m(A_2)}, \dots, 10^{m(A_N)} \}$ obeys BL and, therefore, so does the sequence $\{A_1, A_2, \dots , A_N\}$. This result is very useful since allows us to check if a sequence of numbers obeys BL by looking at the distribution of the mantissas of their logarithms. This is a simple operational rule, instead of a logical one by checking at the digits themselves. \section{Some sequences that obey BL.} The next question is which type of sequences or sets of numbers follow BL. Answering this in an exhaustive fashion appears as a difficult task. Here, we discuss two general type of sequences that can be shown quite clearly that obey BL. With these two, we shall conjecture about the general case. We discuss these cases below. \subsection{Products of variables with arbitrary distributions.} Consider the set of numbers $\{Q_1, Q_2, \dots , Q_N\}$, with $Q_i$ given by the product of $M$ numbers, \begin{equation} Q_i = R_1^{(i)} R_2^{(i)} \cdots R_M^{(i)} ,\label{prod} \end{equation} where $R$ are the absolute values of numbers drawn from an {\it arbitrary} distribution (up to a requirement to be given below). We now show that in the limit $N \to \infty$ and $M \to \infty$, the sequence $\{Q_1, Q_2, \dots , Q_N\}$ obeys BL. The idea is to use the mantissa rule. For this, we consider the log base 10 of the numbers $Q_i$, \begin{equation} \log_{10} Q_i = \log_{10} R_1^{(i)} + \log_{10} R_2^{(i)} + \cdots + \log_{10} R_M^{(i)} . \end{equation} We now introduce the requirement that the {\it distribution of the logarithms} of the numbers $R$ have finite first and second moments. Then, in the limit $M \to \infty$, by the Central Limit Theorem (CLT)\cite{Feller}, the distribution of $\log_{10} Q$ is the normal distribution. That is, the values of $\log_{10} Q$ are distributed as, \begin{equation} \rho(\log_{10} Q) = \frac{1}{\sqrt{2 \pi} \sigma} \> e^{- (\log_{10} Q - \log_{10} Q_0)^2/2 \sigma^2 }. \label{gauss} \end{equation} Note that this is not the log-normal distribution, but simply the normal distribution for the variable $\log_{10} Q$. The centroid $\log_{10} Q_0 \approx M c_0$ and $\sigma \approx \sqrt{M} \sigma_0$, where $c_0$ and $\sigma_0$ are the first and second moments of the distribution of the {\it logarithms} of the numbers $R$. This point will be further discussed below. We proceed to show that the mantissas of the sequence $\{\log_{10} Q_1, \log_{10} Q_2, \dots , \log_{10} Q_N\}$ are uniformly distributed in the interval $[0, 1)$, in the limits mentioned. Before we give the general condition, we can see the how this limit is achieved. Assume that the gaussian function given by Eq.(\ref{gauss}) is already wide enough such that it covers several orders of magnitude, or ``decades'', of the values of $Q$; see Fig. \ref{Gaus} where the decades are denoted by $P - 2$, $P - 1$, $ \dots$, $P + 3$ . The mantissas of the $\log_{10} Q$ are the decimal values within the intervals $P - L$ and $P - (L + 1)$. Thus, we can ``shift'' all intervals within all decades to a single interval, thus placing the mantissas within the same interval. Adding all the values of the mantissas yields, almost, a uniform distribution. This procedure is the same as considering the sum of an infinite number of gaussians each centered at $(\log_{10} Q_0 - P)$ with $P$ taking all the integer values; in the limit $M \to \infty$, equivalent to $\sigma \to \infty$, one gets the exact result, \begin{equation} \lim_{\sigma \to \infty} \> \frac{1}{\sqrt{2 \pi} \sigma} \sum_{P= -\infty}^{\infty} \> e^{- (\log_{10} Q - \log_{10} Q_0 + P)^2/2 \sigma^2 } = 1 . \end{equation} This proves that the mantissas are uniformly distributed in the limit, for logarithms normally distributed. Although the previous result is strictly valid only in the limit $\sigma \to \infty$, the convergence is extremely fast. For instance, for $\sigma \approx 1$, the sum differs from 1 in the eighth significant figure. One finds strong deviations from the uniform distribution as $\sigma$ becomes much smaller than 1, that is, when the gaussian covers less than one decade. \begin{figure} \begin{center} \begin{minipage}[h]{.50\textwidth} \includegraphics[width=.99\textwidth]{fig2a.pdf} \end{minipage}% \hfill \begin{minipage}[h]{.50\textwidth} \centering \includegraphics[width=.99\linewidth]{fig2b.pdf} \end{minipage} \end{center} \caption{First panel, normal distribution of $\log_{10}(A)$, covering 5 decades approximately. Second panel, the mantissas of the normal distribution within one decade, i.e. in the interval $[0, 1)$; the dotted line is the sum of only 5 decades, adding to 1 within 3 significant figures.} \label{Gaus} \end{figure} On the other hand, since $\sigma$ depends not only on $M$ but also on the second moment $\sigma_0$ of the distribution of logarithms of $R$, i.e. $\sigma \approx \sqrt{M} \sigma_0$, the convergence might be very slow if the width, or support, of the distribution of $R$ itself is very narrow. As particular examples, considering $R$ taken from a uniform distribution in the interval $[1, 10)$, requires $M$ to be less than 10 (about 4 or 5) to converge to BL. Conversely, for $R$ in the interval $[5, 6)$ takes $M \approx 400$ to yield BL. The result of this section, namely of the product of numbers obeying BL, is very robust and general\cite{Pietronero} in the sense that even if the distribution of the numbers $R$ lack second moment, the logarithm of $R$ may not\cite{Lorentz}. This is because the logarithm function is very ``slow'' and tends to smooth the original distribution. Moreover, even if the numbers $R$ are correlated, the action of the logarithm and the limit of very large products (i.e. large values of $M$) may again yield a normal distribution of the logarithms\cite{GUE,Flores}. \subsection{Generalized geometric sequences of variables uniformly distributed.} Here we consider a geometric sequence of products of the form, \begin{equation} \{Z^{(1)}, Z^{(2)}, \dots, Z^{(N)}, \dots \} = \{ R_1^{(1)}, R_1^{(2)} R_2^{(2)}, R_1^{(3)}R_2^{(3)}R_3^{(3)}, \dots, R_1^{(N)}R_2^{(N)}\cdots R_N^{(N)}, \dots \} ,\label{geo} \end{equation} where $R_i^{(J)}$ are numbers uniformly distributed in an arbitrary interval $[a, b]$, with $a$ and $b$ real positive numbers. This sequence obeys BL. Although this result may be generalized to arbitrary distributions, we restrict the results here to uniform distributions. We note that if $a = b$, the above sequence is a true geometric progression with ratio $a$. Thus, geometric progressions also obey BL (except if $a = 10^L$ with $L$ any integer). Again, we first consider the sequence of the logarithms of the products, $\log_{10} Z^{(J)}$. Since we do not have an analytic demonstration, we resort to a numerical one. In Fig. \ref{geofig} a particular example shows that the distribution of $\log_{10} Z^{(J)}$ becomes uniformly distributed as $J \to \infty$. As this distribution covers many decades of $Z^{(J)}$, obviously the mantissas of $\log_{10} Z^{(J)}$ also become uniform in $[0, 1)$. A numerical comparison with BL is also included. We have extensively verified that these results hold for any sequence of this type, including true geometric progressions\cite{Raimi}. \begin{figure} \begin{center} \begin{minipage}[h]{.33\textwidth} \includegraphics[width=.99\textwidth]{fig3a.pdf} \end{minipage}% \hfill \begin{minipage}[h]{.33\textwidth} \centering \includegraphics[width=.99\linewidth]{fig3b.pdf} \end{minipage} \begin{minipage}[h]{.33\textwidth} \includegraphics[width=.99\textwidth]{fig3c.pdf} \end{minipage}% \hfill \end{center} \caption{Numerical analysis of the first 10,000 terms of one realization of a generalized geometric sequence, as given by Eq.(\ref{geo}), for numbers uniformly distributed in the interval $[1.0, 9.9]$. First panel shows the uniform distribution of $\log_{10}(A)$ covering more than 40 decades. Second panel shows that the distribution of mantissas is uniform in $[0, 1)$. The third panel is a comparison of the exact Benford Law (circles) with the distribution of the first digit of the 10,000 terms considered (triangles).} \label{geofig} \end{figure} \subsection{A conjecture on the general case.} From the above two cases, it appears that a generalization is as follows: As long as the distribution of logarithms is wide enough, namely, covering many decades of the set considered, the mantissa distribution will tend to become uniformly distributed. An analogous argument was recently used by Fewster\cite{Fewster} to illustrate when Benford's law should be obeyed. \section{Why the pages of log tables wear out following BL?} Simon Newcomb initiates his article by pointing out that the log tables were worn out more at the beginning than at the end, i.e. following BL. That is, since the tables are for logarithms of numbers going in order from $1.000\dots$ to $9.999\dots$, he found that the pages for numbers starting with 1 were more used than those for numbers starting with 2, etc. Newcomb gave an explanation of these observation by assuming that ``natural" numbers, i.e. those appearing in Nature, were obtained by ratios of other numbers. Then, he argued that no matter the underlying law of the primitive numbers, their ratios (in the limit of many ratios) had the mantissas of their logarithms uniformly distributed. He then simply stated that this implied Benford Law. As we have seen, the mantissa rule is equivalent to GR. It is fairly evident that Newcomb certainly knew this result, and thus, that he must be credited with the derivation of BL. We mention, once more, that the arguments given in this article are essentially contained in Newcomb's original paper. An interesting aspect is why Newcomb considered that ``natural" numbers were the result of ratios, or products for that matter, of other numbers. In the light of the previous sections and a bit of second-guessing, we can advance an explanation for this assertion by Newcomb. Moreover, this may also well be the explanation for the agreement of actual real-life data with BL. To begin, we should recall why log tables were used in the first place. We are well into the era of electronic calculators, be it a pocket-size one or a huge supercomputer: numerical calculations are now their task not ours. But as recently as the early 1970's, not to mention in the XIX century, numerical calculations were done by hand and/or sliding rules. And the log tables were essential to realize those tedious and lengthy tasks. As a matter of fact, logarithms were invented (or discovered?) by John Napier in 1614 to perform lengthy calculations! In the words of Napier himself\cite{e}, {\it Seeing there is nothing that is so troublesome to mathematical practice, nor that doth more molest and hinder calculations, than the multiplications, divisions, square and cubic extractions of great numbers. ... I began therefore to consider in my mind by what certain and ready art I might remove those hindrances.} - John Napier, {\it Mirifici logarithmorum canonis descriptio} (1614). That is, the trouble appears when one must make calculations by hand, specially multiplications, of numbers with {\it many} digits. It is lengthy, tedious and prone to produce mistakes. Thus, one goes to the tables to find out the logarithms of the numbers involved, performs sums and subtractions which are much easier, and then taking antilogarithms the result is found. The point is, where did those {\it long} numbers come from? Those were definitely not made up, neither read out from somewhere else, nor measured. The long numbers came themselves from {\it multiplication, divisions or powers} of smaller numbers. The latter may be random, or measured or taken arbitrarily from somewhere else, indeed. But, we insist, the long numbers did arise from operations performed on smaller numbers. As we have seen in the previous section, multiplication of numbers tipically tend to BL, even if only few factors are involved, as long as they arise from wide distributions. In other words, the numbers that people looked their logarithms for, typically, obeyed already BL. Since in the XIX century numbers were not churned out from a computer but arised from arithmetic operations performed by real people, it seems to the author that for Newcomb these were ``naturally" produced. This may also explain why many sets of real-life data obey BL, that is, unless one asks a computer for a random number, numbers that quantify a property, be it the area of a lake or the weight of a molecule, usually arise from arithmetic operations performed on measured quantities with arbitrary constants and units. {\bf Acknowledgments}. I thank R. Esquivel and A. Robledo for several important references.

https://arxiv.org/abs/1107.4392

Lower bounds for sumsets of multisets in Z_p^2

The classical Cauchy-Davenport theorem implies the lower bound n+1 for the number of distinct subsums that can be formed from a sequence of n elements of the cyclic group Z_p (when p is prime and n<p). We generalize this theorem to a conjecture for the minimum number of distinct subsums that can be formed from elements of a multiset in (Z_p)^m; the conjecture is expected to be valid for multisets that are not "wasteful" by having too many elements in nontrivial subgroups. We prove this conjecture in (Z_p)^2 for multisets of size p+k, when k is not too large in terms of p.

\section{Introduction} Determining the number of elements in a particular abelian group that can be written as sums of given sets of elements is a topic that goes back at least two centuries. The most famous result of this type, involving the cyclic group ${\mathbb Z}_p$ of prime order $p$, was established by Cauchy in 1813~\cite{Cauchy} and rediscovered by Davenport in 1935~\cite{Dav1,Dav2}: \begin{lemma}[Cauchy--Davenport Theorem] Let $A$ and $B$ be subsets of ${\mathbb Z}_p$, and define $A+B$ to be the set of all elements of the form $a+b$ with $a\in A$ and $b\in B$. Then $\#(A+B) \ge \min\{ p, \#A + \#B - 1\}$. \end{lemma} \noindent The lower bound is easily seen to be best possible by taking $A$ and $B$ to be intervals, for example. It is also easy to see that the lower bound of $\#A + \#B - 1$ does not hold for general abelian groups ${\mathbf G}$ (take $A$ and $B$ to be the same nontrivial subgroup of ${\mathbf G}$). There is, however, a well-known generalization obtained by Kneser in 1953~\cite{Kne}, which we state in a slightly simplified form that will be quite useful for our purposes (see \cite[Theorem 4.1]{Nat} for an elementary proof): \begin{lemma}[Kneser's Theorem] Let $A$ and $B$ be subsets of a finite abelian group ${\mathbf G}$, and let $m$ be the largest cardinality of a proper subgroup of ${\mathbf G}$. Then $\#(A+B) \ge \min\{ \#{\mathbf G}, \#A + \#B - m\}$. \end{lemma} Given a sequence $A = (a_1,\dots,a_k)$ of (not necessarily distinct) elements of an abelian group ${\mathbf G}$, a related result involves its {\em sumset} $\Sigma A$, which is the set of all sums of any number of elements chosen from $A$ (not to be confused with $A+A$, which it contains but usually properly): \[ \Sigma A = \bigg\{ \sum_{j\in J} a_j \colon J\subseteq \{1,\dots,k\} \bigg\}. \] (Note that we allow $J$ to be empty, so that the group's identity element is always an element of $\Sigma A$.) When ${\mathbf G}={\mathbb Z}_p$, one can prove the following result by writing $\Sigma A = \{0, a_1\} + \cdots + \{ 0,a_k \}$ and applying the Cauchy--Davenport theorem inductively: \begin{lemma} \label{CD pre lemma} Let $A=(a_1,\dots,a_k)$ be a sequence of nonzero elements of ${\mathbb Z}_p$. Then $\#\Sigma A \ge \min\{p,k + 1\}$. \end{lemma} \noindent This result can also be proved directly by induction on $k$, and in fact such a proof will discover why the order $p$ of the cyclic group must be prime (intuitively, the sequence $A$ could lie completely within a nontrivial subgroup). For a formal proof, see \cite[Lemma 2]{DEF}. Again the lower bound is easily seen to be best possible, by taking $a_1=\cdots=a_k$. It is a bit misleading to phrase such results in terms of sequences, since the actual order of the elements in the sequence is irrelevant (given that we are considering only abelian groups). We prefer to use {\em multisets}, which are simply sets that are allowed to contain their elements with multiplicity. If we let $m_x$ denote the multiplicity with which the element $x$ occurs in the multiset $A$, then the definition of $\Sigma A$ can be written in the form \[ \Sigma A = \bigg\{ \sum_{x\in{\mathbf G}} \delta_x x \colon 0\le \delta_x \le m_x \bigg\}, \] where $\delta_x x$ denotes the group element $x+\cdots+x$ obtained by adding $\delta_x$ summands all equal to~$x$. When using multisets, we should choose our notation with care: the hypotheses of such results tend to involve the total number of elements of the multiset $A$ counting multiplicity, while the conclusions involve the number of distinct elements of $\Sigma A$. Consequently, throughout this paper, we use the following notational conventions: \begin{itemize} \item $|S|$ denotes the total number of elements of the multiset $S$, counted with multiplicity; \item $\#S$ denotes the number of distinct elements of the multiset $S$, or equivalently the number of elements of $S$ considered as a (mere) set. \end{itemize} In this notation, Lemma~\ref{CD pre lemma} can be restated as: \begin{lemma} \label{CD lemma} Let $A$ be a multiset contained in ${\mathbb Z}_p$ such that $0\notin A$. Then $\#\Sigma A \ge \min\{p,|A| + 1\}$. \end{lemma} \noindent The purpose of this paper is to improve, as far as possible, this lower bound for multisets contained in the larger abelian group $\F_p^2$. We cannot make any progress without some restriction upon our multisets: if a multiset is contained within a nontrivial subgroup of $\F_p^2$ (of cardinality $p$), then so is its sumset, in which case the lower bound $\min\{p,|A| + 1\}$ from Lemma~\ref {CD lemma} is the best we can do. Therefore we restrict to the following class of multisets. We use the symbol $\text{\bf 0}=(0,0)$ to denote the identity element of~$\F_p^2$. \begin{definition} \label{valid definition} A multiset $A$ contained in $\F_p^2$ is called {\em valid} if: \begin{itemize} \item $\text{\bf 0}\notin A$; and \item every nontrivial subgroup contains fewer than $p$ points of $A$, counting multiplicity. \end{itemize} \end{definition} \noindent The exact number $p$ in the second condition has been carefully chosen: any nontrivial subgroup of $\F_p^2$ is isomorphic to ${\mathbb Z}_p$, and so Lemma~\ref{CD lemma} applies to these nontrivial subgroups. In particular, any multiset $A$ containing $p-1$ nonzero elements of a nontrivial subgroup will automatically have that entire subgroup contained in its sumset $\Sigma A$, so allowing $p$ nonzero elements in a nontrivial subgroup would always be ``wasteful''. We believe that the following lower bound should hold for sumsets of valid multisets: \begin{conjecture} \label{2d conjecture} Let $A$ be a valid multiset contained in $\F_p^2$ such that $p \le |A| \le 2p-3$. Then $\#\Sigma A \ge (|A|+2-p)p$. In other words, if $|A| = p+k$ with $0\le k\le p-3$, then $\#\Sigma A \ge (k+2)p$. \end{conjecture} It is easy to see that this conjectured lower bound would be best possible: if $A$ is the multiset that contains the point $(1,0)$ with multiplicity $p-1$ and the point $(0,1)$ with multiplicity $k+1$, then the set $\Sigma A$ is precisely $\big\{ (s,t)\colon s\in{\mathbb Z}_p,\, 0\le t\le k+1 \big\}$, which has $(k+2)p$ distinct elements. Conjecture~\ref {2d conjecture} is actually part of a larger assertion (see Conjecture~\ref{any d conjecture}) concerning lower bounds for sumsets in ${\mathbb Z}_p^m$. One of our results completely resolves the first two cases of this conjecture: \begin{theorem} \label{conjecture true for k tiny} Let $p$ be a prime. \begin{enumerate} \item If $A$ is any valid multiset contained in $\F_p^2$ with $|A| = p$, then $\#\Sigma A \ge 2p$. \item Suppose that $p\ge5$. If $A$ is any valid multiset contained in $\F_p^2$ with $|A| = p+1$, then $\#\Sigma A \ge 3p$. \end{enumerate} \end{theorem} It turns out that proving part (b) of the theorem requires a certain amount of computation for a finite number of primes (see the remarks following the proof of the theorem in Section~\ref{thm proofs section}). Extending the conjecture to larger values of $k$ would require, by our methods, more and more computation to take care of small primes $p$ as $k$ grows. However, we are able to establish the conjecture when $p$ is large enough with respect to $k$, or equivalently when $k$ is small enough with respect to $p$: \begin{theorem} \label{conjecture true for k small in terms of p} Let $p$ be a prime, and let $2\le k\le \sqrt{p/(2\log p+1)}-1$ be an integer. If $A$ is any valid multiset contained in $\F_p^2$ with $|A| = p+k$, then $\#\Sigma A \ge (k+2)p$. \end{theorem} A contrapositive version of Theorem~\ref{conjecture true for k small in terms of p} is also enlightening: \begin{corollary} \label{main corollary} Let $p$ be a prime, and let $2\le k\le \sqrt{p/(2\log p+1)}-1$ be an integer. Let $A$ be a multiset contained in $\F_p^2 \setminus \{\text{\bf 0}\}$ with $|A| = p+k$. If $\#\Sigma A < (k+2)p$, then there exists a nontrivial subgroup of $\F_p^2$ that contains at least $p$ points of $A$, counting multiplicity. \end{corollary} Our methods of proof stem from two main ideas. First, we will obviously exploit the structure of $\F_p^2$ as a direct sum of cyclic groups of prime order, within which we can apply the known Lemma~\ref{CD lemma} after using projections. Section~\ref{direct projects section} contains several elementary lemmas in this vein (see in particular Lemma~\ref{sweep to the left lemma}). It is important for us to utilize the flexibility coming from the fact that $\F_p^2$ can be decomposed as the direct sum of two subgroups in many different ways. Second, our methods work best when there exists a single subgroup that contains many elements of the given multiset; however, by selectively replacing pairs of elements with their sums, we can increase the number of elements in a subgroup in a way that improves our lower bounds upon the sumset (see Lemma~\ref {any j lemma}). These methods, which appear in Section~\ref {thm proofs section}, combine to provide the proofs of Theorems \ref {conjecture true for k tiny} and~\ref{conjecture true for k small in terms of p}. Finally, Section~\ref{conjecture section} contains a generalization of Conjecture~\ref{2d conjecture} to higher-dimensional direct sums of ${\mathbb Z}_p$, together with examples demonstrating that the conjecture would be best possible. \section{Sumsets in abelian groups and direct products} \label{direct projects section} All of the results in this section are valid for general finite abelian groups and have correspondingly elementary proofs, although the last two lemmas seem rather less standard than the first few. In this section, ${\mathbf G}$, ${\mathbf H}$, and ${\mathbf K}$ denote finite abelian groups, and $e$ denotes a group's identity element. \begin{lemma} \label{a la carte lemma} Let $B_0,B_1,B_2,\dots,B_j$ be multisets in ${\mathbf G}$, and set $A = B_0 \cup B_1 \cup \dots \cup B_j$. For each $1\le i\le j$, specify an element $x_i \in \Sigma B_i$, and set $C = B_0 \cup \{ x_1, \dots ,x_j\}$. Then $\Sigma C \subseteq \Sigma A$. \end{lemma} \begin{proof} For each $1\le i\le j$, choose a submultiset $D_i \subseteq B_i$ such that the sum of the elements of $D_i$ equals $x_i$. By definition, every element $y$ of $\Sigma C$ equals the sum of the elements of some subset $E$ of $B_0$, plus $\sum_{i\in I} x_i$ for some $I\subseteq \{1,\dots,j\}$. But then $y$ equals the sum of the elements of $E \cup \bigcup_{i\in I} D_i$, which is an element of $\Sigma A$ since $E \cup \bigcup_{i\in I} D_i \subseteq B_0 \cup \bigcup_{1\le i\le j} B_i = A$. \end{proof} \begin{lemma} \label{kneser bound} Let $A_1,A_2,\dots,A_j$ be multisets in ${\mathbf G}$, and set $A = A_1 \cup \dots \cup A_j$. If $m$ is the largest cardinality of a proper subgroup of ${\mathbf G}$, then either $\Sigma A = {\mathbf G}$ or $\#\Sigma A \ge (\sum_{i=1}^j \# \Sigma A_i) - (j-1)m$. \end{lemma} \begin{proof} Since $\Sigma A = \Sigma A_1 + \Sigma A_2 + \cdots + \Sigma A_j$ (viewed as ordinary sets), this follows immediately by inductive application of Kneser's theorem. \end{proof} For the remainder of this section, we will be dealing with groups that can be decomposed into a direct sum. \begin{definition} A subgroup ${\mathbf H}$ of ${\mathbf G}$ is called an {\em internal direct summand} if there exists a subgroup ${\mathbf K}$ of ${\mathbf G}$ such that ${\mathbf G}$ is the internal direct sum of ${\mathbf H}$ and ${\mathbf K}$, or in other words, such that ${\mathbf H} \cap {\mathbf K} = \{e\}$ and ${\mathbf H} + {\mathbf K} = {\mathbf G}$. Equivalently, ${\mathbf H}$ is an internal direct summand of ${\mathbf G}$ if there exists a {\em projection homomorphism} $\pi_{\mathbf H}\colon {\mathbf G} \to {\mathbf H}$ that is the identity on ${\mathbf H}$. Note that this projection homorphism does depend on the choice of ${\mathbf K}$ but is uniquely determined by $\pi_{\mathbf H}^{-1}(e) = {\mathbf K}$. \end{definition} \begin{lemma} \label{pi commutes with Sigma lemma} For any homomorphism $f\colon {\mathbf G}\to {\mathbf H}$, and any subset $X$ of ${\mathbf G}$, we have $f(\Sigma X) = \Sigma (f(X))$. In particular, if ${\mathbf H}$ is an internal direct summand of ${\mathbf G}$, then $\pi_{\mathbf H}(\Sigma X) = \Sigma(\pi_{\mathbf H}(X))$ for any subset $X$ of ${\mathbf G}$. \end{lemma} \begin{proof} Given $y\in f(\Sigma X)$, there exists $x\in \Sigma X$ such that $f(x)=y$. Hence we can find $x_1,\dots,x_j\in X$ such that $x_1+\cdots+x_j = x$, and so $f(x_1+\cdots+x_j) = y$. But $f$ is a homomorphism, and so $f(x_1)+\cdots+f(x_j) = y$, so that $y \in \Sigma(f(X))$. This shows that $f(\Sigma X) \subseteq \Sigma(f(X))$; the proof of the reverse inclusion is similar. \end{proof} \begin{lemma} \label{pi split lemma} Let ${\mathbf G} = {\mathbf H} \oplus {\mathbf K}$, and let $D$ and $E$ be multisets contained in ${\mathbf H}$ and ${\mathbf K}$, respectively. For any $z\in {\mathbf G}$, \[ z \in \Sigma(D\cup E) \quad\text{if and only if}\quad \pi_{\mathbf H}(z) \in\Sigma D \text{ and } \pi_{\mathbf K}(z) \in\Sigma E. \] \end{lemma} \begin{proof} Since $z = \pi_{\mathbf H}(z) + \pi_{\mathbf K}(z)$, the ``if'' direction is obvious. For the converse, note that \[ \pi_{\mathbf H}(z) \in \pi_{\mathbf H}\big( \Sigma (D\cup E) \big) = \Sigma\big( \pi_{\mathbf H}(D \cup E) \big) \] by Lemma~\ref {pi commutes with Sigma lemma}. On the other hand, $\pi_{\mathbf H}(D) = D$ and $\pi_{\mathbf H}(E) = \{e\}$, and so \[ \pi_{\mathbf H}(z) \in \Sigma\big( \pi_{\mathbf H}(D) \cup \pi_{\mathbf H}(E) \big) = \Sigma \big( D \cup \{e\} \big) = \Sigma D \] (since the sumset is not affected by whether $e$ is an allowed summand). A similar argument shows that $\pi_{\mathbf K}(z) \in \Sigma E$, which completes the proof of the lemma. \end{proof} \begin{lemma} \label{direct product lemma} Let ${\mathbf H}$ and ${\mathbf K}$ be subgroups of ${\mathbf G}$ satisfying ${\mathbf H} \cap {\mathbf K} = \{e\}$. Let $D$ and $E$ be multisets contained in ${\mathbf H}$ and ${\mathbf K}$, respectively. Then $\#\Sigma(D\cup E) = \#\Sigma D \cdot \#\Sigma E$. \end{lemma} \begin{proof} Notice that every element of $\Sigma(D\cup E)$ is contained in ${\mathbf H}+{\mathbf K}$; therefore we may assume without loss of generality that ${\mathbf G} = {\mathbf H} \oplus {\mathbf K}$. In particular, we may assume that ${\mathbf H}$ and ${\mathbf K}$ are internal direct summands of ${\mathbf G}$, so that the projection maps $\pi_{\mathbf H}$ and $\pi_{\mathbf K}$ exist and every element $z\in {\mathbf G}$ has a unique representation $z=x+y$ where $x\in {\mathbf H}$ and $y\in {\mathbf K}$; note that $x=\pi_{\mathbf H}(z)$ and $y=\pi_{\mathbf K}(z)$ in this representation. To establish the lemma, it therefore suffices to show that $z = \pi_{\mathbf H}(z) + \pi_{\mathbf K}(z) \in \Sigma(D\cup E)$ if and only if $\pi_{\mathbf H}(z) \in\Sigma D \text{ and } \pi_{\mathbf K}(z) \in\Sigma E$; but this is exactly the statement of Lemma~\ref{pi split lemma}. \end{proof} The next lemma is a bit less standard yet still straightforward: in a direct product of two abelian groups, it characterizes the elements of a sumset that lie in a given coset of one of the direct summands. \begin{lemma} \label{orthogonal structure lemma} Let ${\mathbf H}$ and ${\mathbf K}$ be subgroups of ${\mathbf G}$ satisfying ${\mathbf H} \cap {\mathbf K} = \{e\}$. Let $D$ and $E$ be multisets contained in ${\mathbf H}$ and ${\mathbf K}$, respectively. For any $y\in {\mathbf K}$: \begin{enumerate} \item if $y\in \Sigma E$, then $({\mathbf H}+\{y\}) \cap \Sigma(D\cup E) = \Sigma D + \{y\}$; \item if $y\notin \Sigma E$, then $({\mathbf H}+\{y\}) \cap \Sigma(D\cup E) = \emptyset$. \end{enumerate} \end{lemma} \begin{proof} As in the proof of Lemma~\ref {direct product lemma}, we may assume without loss of generality that ${\mathbf G} = {\mathbf H} \oplus {\mathbf K}$. Suppose that $z$ is an element of $({\mathbf H}+\{y\}) \cap \Sigma(D\cup E)$. Since $z\in {\mathbf H}+\{y\}$, we may write $z=x+y$ for some $x\in{\mathbf H}$, whence $\pi_{\mathbf K}(z) = \pi_{\mathbf K}(x) + \pi_{\mathbf K}(y) = e+y = y$. On the other hand, since $z\in \Sigma(D\cup E)$, we see that $y \in \Sigma E$ by Lemma~\ref{pi split lemma}. In other words, the presence of any element $z\in ({\mathbf H}+\{y\}) \cap \Sigma(D\cup E)$ forces $y\in \Sigma E$, which establishes part (b) of the lemma. We continue under the assumption $y\in \Sigma E$ to prove part (a). The inclusions $\Sigma D + \{y\} \subseteq {\mathbf H}+\{y\}$ and $\Sigma D + \{y\} \subseteq \Sigma(D\cup E)$ are both obvious, and so $\Sigma D + \{y\} \subseteq ({\mathbf H}+\{y\}) \cap \Sigma(D\cup E)$. As~for the reverse inclusion, let $z \in ({\mathbf H}+\{y\}) \cap \Sigma(D\cup E)$ as above; then $\pi_{\mathbf H}(z) \in \Sigma D$ by Lemma~\ref{pi split lemma}, whence $z = \pi_{\mathbf H}(z) + \pi_{\mathbf K}(z) = \pi_{\mathbf H}(z) + y \in \Sigma D + \{y\}$ as required. \end{proof} Finally we can establish the lemma that we will make the most use of when we return to the setting ${\mathbf G}=\F_p^2$ in the next section. \begin{lemma} \label{sweep to the left lemma} Let ${\mathbf G} = {\mathbf H} \oplus {\mathbf K}$, and let $C$ be a multiset contained in ${\mathbf G}$. Let $D=C\cap {\mathbf H}$, let $F = C \setminus D$, and let $E = \pi_{\mathbf K}(F)$. Then $\#\Sigma C \ge \#\Sigma D \cdot \#\Sigma E$. \end{lemma} \begin{proof} Lemma~\ref {direct product lemma} tells us that $\#\Sigma (D\cup E) = \#\Sigma D \cdot \#\Sigma E$, and so it suffices to show that $\#\Sigma C \ge \#\Sigma (D\cup E)$. We accomplish this by showing that \begin{equation} \# \big( ({\mathbf H} + \{y\}) \cap \Sigma C \big) \ge \# \big( ({\mathbf H} + \{y\}) \cap \Sigma (D\cup E) \big) \label{one line at a time} \end{equation} for all $y\in{\mathbf K}$. For any $y\in{\mathbf K} \setminus \Sigma E$, Lemma~\ref{orthogonal structure lemma} tells us that $({\mathbf H} + \{y\}) \cap \Sigma (D\cup E) = \emptyset$, in which case the inequality~\eqref{one line at a time} holds trivially. For any $y \in \Sigma E$, Lemma~\ref{orthogonal structure lemma} tells us that $({\mathbf H} + \{y\}) \cap \Sigma (D\cup E) = \Sigma D + \{y\}$, and so the right-hand side of the inequality~\eqref{one line at a time} equals $\#\Sigma D$. On the other hand, since $\Sigma E = \Sigma (\pi_{\mathbf K}(F)) = \pi_{\mathbf K}(\Sigma F)$ by Lemma~\ref{pi commutes with Sigma lemma}, there exists at least one element $z\in \Sigma F$ satisfying $\pi_{\mathbf K}(z)=y$; as ${\mathbf G} = {\mathbf H} \oplus {\mathbf K}$, this is equivalent to saying that $z \in {\mathbf H} + \{y\}$. Since $\Sigma D \subseteq {\mathbf H}$, we have $\Sigma D + \{z\} \subseteq {\mathbf H} + \{y\}$ as well. But the inclusion $\Sigma D + \{z\} \subseteq \Sigma D + \Sigma F = \Sigma C$ is trivial, and therefore $\Sigma D + \{z\} \subseteq ({\mathbf H} + \{y\}) \cap \Sigma C$; in particular, the left-hand side of the inequality~\eqref{one line at a time} is at least $\#\Sigma D$. Combined with the observation that the right-hand side equals $\#\Sigma D$, this lower bound establishes the inequality~\eqref{one line at a time} and hence the lemma. \end{proof} These lemmas might be valuable for studying sumsets in more general abelian groups. They will prove to be particularly useful for studying sumsets in $\F_p^2$, however, essentially because there are many ways of writing $\F_p^2$ as an internal direct sum of two subgroups (which are simply lines through $\text{\bf 0}$). \section{Lower bounds for sumsets} \label{thm proofs section} In this section we establish Theorems~\ref {conjecture true for k tiny} and~\ref {conjecture true for k small in terms of p}; the proofs employ two combinatorial propositions which we defer to the next section. It would be possible to prove these two theorems at the same time, at the expense of a bit of clarity; however, we find it illuminating to give complete proofs of Theorem~\ref {conjecture true for k tiny} (the cases $|A|=p$ and $|A|=p+1$) first, as the proofs will illustrate the methods used to prove the more general Theorem~\ref {conjecture true for k small in terms of p}. Seeing the limitations of the proof of Theorem~\ref {conjecture true for k tiny} will also motivate the formulation of our main technical tool, Lemma~\ref {any j lemma}. Throughout this section, $A$ will denote a valid multiset contained in $\F_p^2$. For any $x\in\F_p^2$, we let $\lin x$ denotes the subgroup of $\F_p^2$ generated by $x$ (that is, the line passing through both the origin $\text{\bf 0}$ and $x$), and we let $m_x$ denote the multiplicity with which $x$ appears in $A$, so that $|A| = \sum_{x\in\F_p^2} m_x$. The fact that $A$ is valid means that $m_\text{\bf 0}=0$ and $\sum_{t\in\lin x} m_t < p$ for every $x\in\F_p^2\setminus\{\orig\}$. Our first lemma quantifies the notion that we can establish sufficiently good lower bounds for the cardinality of $\Sigma A$ if we know that there are enough elements of $A$ lying in one subgroup of~$\F_p^2$. Naturally, the method of proof is to partition $A$ into the elements lying in that subgroup and all remaining elements, project the remaining elements onto a complementary subgroup, and then use Lemma~\ref{CD lemma} in each subgroup separately. \begin{lemma} \label{conjecture true if enough on line lemma} Let $A$ be any valid multiset contained in $\F_p^2$. Suppose that for some $x\in\F_p^2\setminus\{\orig\}$, \begin{equation} \label{symmetric bound} \sum_{y\in\lin x} m_y \ge |A| - (p-1). \end{equation} Then $\#\Sigma A \ge (|A|+2-p)p$. \end{lemma} \begin{remark} The conclusion is worse than trivial if $|A| < p-1$; also, the fact that $A$ is valid means that the left-hand side of equation~\eqref{symmetric bound} is at most $p-1$, and so the lemma is vacuous if $|A| > 2p-2$. Therefore in practice the lemma will be applied only to multisets $A$ satisfying $p-1 \le |A| \le 2p-2$. \end{remark} \begin{proof} Let $D = A \cap \lin x$; note that $|D| \le p-1$ since $A$ is a valid multiset, and note also that $|D| = \sum_{y\in\lin x} m_y \ge |A| - (p-1)$ by assumption. Set $F = A \setminus D$. Choose any nontrivial subgroup ${\mathbf K}$ of $\F_p^2$ other than $\lin x$, and set $E = \pi_{\mathbf K}(F)$. Then by Lemma~\ref{sweep to the left lemma}, we know that $\#\Sigma A \ge \#\Sigma D \cdot \#\Sigma E$. By Lemma~\ref {CD lemma} and the fact that $\text{\bf 0}\notin D\cup E$, we obtain \begin{align} \#\Sigma A &\ge \min \big\{ p, 1 + |D| \big\} \cdot \min \big\{ p, 1 + |E| \big\} \notag \\ &= \min \big\{ p, 1 + |D| \big\} \cdot \min \big\{ p, 1 + |A| - |D| \big\}, \label{actually special case} \end{align} since $|E| = |F| = |A|-|D|$. The inequalities $|D| \le p-1$ and $|A| - |D| \le p-1$ ensure that $p$ is the larger element in both minima, and so we have simply \[ \#\Sigma A \ge (1+|D|)(1+|A|-|D|) = \tfrac14|A|^2 + |A| + 1 - \big( |D| - \tfrac12|A| \big)^2. \] The pair of inequalities $|D| \le p-1$ and $|A| - |D| \le p-1$ is equivalent to the inequality $\big| |D| - \tfrac12|A| \big| \le p-1- \tfrac12|A|$; therefore \[ |\Sigma A| \ge \tfrac14|A|^2 + |A| + 1 - \big( p-1- \tfrac12|A| \big)^2 = (|A|+2-p)p, \] as claimed. \end{proof} This lemma alone is sufficient to establish Theorem~\ref{conjecture true for k tiny}. \begin{proof}[Proof of Theorem~\ref{conjecture true for k tiny}(a)] When $|A|=p$, the right-hand side of the inequality~\eqref {symmetric bound} equals 1, and so the inequality holds for any $x\in A$. Therefore Lemma~\ref {conjecture true if enough on line lemma} automatically applies, yielding $\#\Sigma A \ge (|A|+2-p)p = 2p$ as desired. (In fact essentially the same proof gives the more general statement: if $A$ is a multiset contained in $\F_p^2\setminus\{\text{\bf 0}\}$ but not contained in any proper subgroup, and $|A|\ge p$, then $\#\Sigma A \ge 2|A|$.) \end{proof} \begin{proof}[Proof of Theorem~\ref{conjecture true for k tiny}(b)] We are assuming that $|A| = p+1$. Suppose first that there exists a nontrivial subgroup of $\F_p^2$ that contains at least two points of $A$ (including possibly two copies of the same point). Choosing any nonzero element $x$ in that subgroup, we see that the inequality~\eqref {symmetric bound} is satisfied, and so Lemma~\ref {conjecture true if enough on line lemma} yields $\#\Sigma A \ge (|A|+2-p)p = 3p$ as desired. From now on we may assume that there does not exist a nontrivial subgroup of $\F_p^2$ that contains at least two points of $A$. Since there are only $p+1$ nontrivial subgroups of $\F_p^2$, it must be the case that $A$ consists of exactly one point from each of these $p+1$ subgroups; in particular, the elements of $A$ are distinct. We can verify the assertion for $p\le11$ by exhaustive computation (see the remarks after the end of this proof), so from now on we may assume that $p\ge13$. Suppose first that all sums of pairs of distinct elements from $A$ are distinct. All these sums are elements of $\Sigma A$, and thus $\#\Sigma A \ge \binom{p+1}2 >3p$ since $p\ge13$. The only remaining case is when two pairs of distinct elements from $A$ sum to the same point of $\F_p^2$. Specifically, suppose that there exist $x_1,y_1,x_2,y_2\in A$ such that $x_1+y_1=x_2+y_2$. Partition $A = B_0 \cup B_1 \cup B_2$ where $B_1=\{x_1,y_1\}$ and $B_2=\{x_2,y_2\}$ and hence $B_0 = A \setminus \{x_1,y_1,x_2,y_2\}$; note that this really is a partition of $A$, as the fact that $x_1+y_1=x_2+y_2$ forces all four elements to be distinct. Moreover, if we define $z=x_1+y_1=x_2+y_2$, then we know that $z\ne\text{\bf 0}$ since $x_1$ and $y_1$ are in different subgroups. Define $C$ to be the multiset $B_0 \cup \{ z, z \}$; by Lemma~\ref {a la carte lemma}, we know that $\#\Sigma A \ge \#\Sigma C$. Define $D = C \cap \lin z$; we claim that $|D| = 3$. To see this, note that $A$ has exactly one point in every nontrivial subgroup, and in particular $A$ has exactly one point in $\lin z$. Furthermore, that point cannot be $x_1$ for example, since then $y_1 = z-x_1$ would also be in that subgroup; similarly that point cannot be $x_2$, $y_1$, or $y_2$. We conclude that $B_0$ has exactly one point in $\lin z$, whence $C$ has exactly three points in $\lin z$. Now define $F = C \setminus D$, so that $|F| = |C| - |D| = (|B_0|+2)-3 = (|A|-4+2)-3 = p-4$. Let ${\mathbf K}$ be any nontrivial subgroup other than $\lin z$, and set $E = \pi_{\mathbf K}(F)$. The lower bounds $\#\Sigma D \ge 4$ and $\#\Sigma E \ge p-3$ then follow from Lemma~\ref {CD lemma}. By Lemma~\ref{sweep to the left lemma}, we conclude that $\#\Sigma C \ge \#\Sigma D \cdot \#\Sigma E = 4(p-3) > 3p$ since $p\ge13$. \end{proof} \begin{remark} The computation that verifies Theorem~\ref{conjecture true for k tiny}(b) for $p\le11$ should be done a little bit intelligently, since there are $10^{12}$ subsets $A$ of ${\mathbb Z}_{11}^2$ (for example) consisting of exactly one nonzero element from each nontrivial subgroup. We describe the computation in the hardest case $p=11$. Let us write the elements of ${\mathbb Z}_{11}^2$ as ordered pairs $(s,t)$ with $s$ and $t$ considered modulo~$11$. By separately dilating the two coordinates of ${\mathbb Z}_{11}^2$ (which does not alter the cardinality of $\Sigma A$), we may assume without loss of generality that $A$ contains both $(1,0)$ and $(0,1)$. We also know every such $A$ contains a subset of the form $\{ (i,i), (j,2j), (k,3k), (\ell,4\ell) \}$ for some integers $1\le i,j,k,\ell \le 10$. Therefore the cardinality of every such $\Sigma A$ is at least as large as the cardinality of one of the subsumsets $\Sigma \big( \{ (1,0), (0,1), (i,i), (j,2j), (k,3k), (\ell,4\ell) \} \big)$. There are $10^4$ such subsumsets, and direct computation shows that all of them have more than $33$ distinct elements except for the sixteen cases $\Sigma \big( \{ (1,0), (0,1), \pm(1,1), \pm(1,2), \pm(1,3), \pm(1,4) \} \big)$, which each contain $32$ distinct elements. It is then easily checked that any subsumset of the form $\Sigma \big( \{ (1,0), (0,1), \pm(1,1), \pm(1,2), \pm(1,3), \pm(1,4), (m,5m) \} \big)$ with $1\le m\le10$ contains more than 33 distinct elements. This concludes the verification of Theorem~\ref{conjecture true for k tiny}(b) for $p=11$, and the cases $p\le7$ are verified even more quickly. \end{remark} We now foreshadow the proof of Theorem~\ref{conjecture true for k small in terms of p} by reviewing the structure of the proof of Theorem~\ref{conjecture true for k tiny}(b). In that proof, we quickly showed that the desired lower bound held if there were enough elements of $A$ in the same subgroup. Also, the desired lower bound certainly held if there were enough distinct sums of pairs of elements of~$A$. If however no subgroup contained enough elements of $A$ and there were only a few distinct sums of pairs of elements of $A$, then we showed that we could find multiple pairs of elements summing to the same point in $\F_p^2$. Replacing those elements in $A$ with multiple copies of their joint sum, we found that the corresponding subgroup now contained enough elements to carry the argument through. The following lemma quantifies the final part of this strategy, where we replace $j$ pairs of elements of $A$ with their joint sum and then use our earlier ideas to bound the cardinality of the sumset from below. \begin{lemma} \label{any j lemma} Let $A$ be any valid multiset contained in $\F_p^2$, and let $z\in\F_p^2\setminus\{\orig\}$. For any integer $j$ satisfying \begin{equation} 0\le j \le \tfrac12 \sum_{t\in\F_p^2\setminus\lin z} \min\{m_t,m_{z-t}\}, \label{allows choosing pairs} \end{equation} we have \[ \#\Sigma A \ge \min\bigg\{ p, 1 + j + \sum_{y\in\lin z} m_y \bigg\} \min\bigg\{ p, 1 + |A| - 2j - \sum_{y\in\lin z} m_y \bigg\}. \] \end{lemma} \begin{remark} This can be seen as a generalization of Lemma~\ref {conjecture true if enough on line lemma}, as equation~\eqref {actually special case} is the special case $j=0$ of this lemma. \end{remark} \begin{proof} Partition $A = B_0 \cup B_1 \cup \cdots \cup B_j$, where for each $1\le i\le j$, the multiset $B_i$ has exactly two elements, neither contained in $\lin z$, that sum to $z$ (the complimentary submultiset $B_0$ is unrestricted). The upper bound~\eqref{allows choosing pairs} for $j$ is exactly what is required for such a partition to be possible; the factor of $\frac12$ arises because the sum on the right-hand side of~\eqref{allows choosing pairs} double-counts the pairs $(t,z-t)$ and $(z-t,t)$. Then set $C$ equal to $B_0$ with $j$ additional copies of $z$ inserted. By Lemma~\ref{a la carte lemma}, we know that $\#\Sigma A \ge \#\Sigma C$. Now let $D$ be the intersection of $C$ with the subgroup $\lin z$, and let $F = C \setminus D$. Let ${\mathbf K}$ be any nontrivial subgroup other than $\lin z$, and set $E = \pi_{\mathbf K}(F)$. By Lemma~\ref{sweep to the left lemma}, we know that $\#\Sigma C \ge \#\Sigma D \cdot \#\Sigma E$. However, the number of elements of $D$ (counting multiplicity) is $j$ more than the number of elements of $B_0 \cap \lin z$; this is the same as $j$ more than the number of elements of $A \cap \lin z$ (since no elements of $B_1,\dots,B_j$ lie on $\lin z$), or in other words $j + \sum_{y\in\lin z} m_y$. Similarly, the number of elements of $E$ (equivalently, of $F$) is equal to the number of elements of $B_0 \setminus \lin z$; this is the same as $2j$ less than the number of elements of $A \setminus \lin z$, or in other words $|A| - 2j - \sum_{y\in\lin z} m_y$. The lower bounds $\#\Sigma D \ge \min\big\{ p, 1 + j + \sum_{y\in\lin z} m_y \big\}$ and $\#\Sigma E \ge \min\big\{ p, 1 + |A| - 2j - \sum_{y\in\lin z} m_y \big\}$ then follow from Lemma~\ref {CD lemma}; the chain of inequalities $\#\Sigma A \ge \#\Sigma C \ge \#\Sigma D \cdot \#\Sigma E$ establishes the lemma. \end{proof} We are now ready to use Lemma~\ref{any j lemma} to establish Conjecture~\ref{2d conjecture} when $|A|=p+k$, for all but finitely many primes $p$ depending on~$k$. Let $H_k = 1 + \tfrac12 + \cdots + \tfrac1k$ denote the $k$th harmonic number. \begin{theorem} \label{conjecture true for p large in terms of k} Let $k\ge2$ be any integer, and let $A$ be any valid multiset contained in $\F_p^2$ such that $|A| = p+k$. If $p\ge 4(k+1)^2 H_k - 2k$, then $\#\Sigma A \ge (k+2)p$. \end{theorem} \begin{remark} Using the elementary bound $H_k \le \gamma + \log(k+1)$, where $\gamma$ denotes the Euler--Mascheroni constant, we see that Theorem~\ref{conjecture true for p large in terms of k} holds as long as $p \ge 4(k+1)^2 (\gamma + \log(k+1))$. Theorem~\ref{conjecture true for k small in terms of p} can thus be readily deduced from Theorem~\ref {conjecture true for p large in terms of k} as follows: If $k+1 \le \sqrt{p/(2\log p+1)}$ then $p \ge 4(k+1)^2 (\tfrac14 + \tfrac12 \log p)$. In this case we certainly have $p \ge (1 + 2\log 2) (k+1)^2$, whence $\log p \ge \frac45 + 2 \log(k+1)$ and $\tfrac14 + \tfrac12 \log p \ge \gamma + \log(k+1)$. \end{remark} \begin{proof} If there are $k+1$ elements of $A$ in some nontrivial subgroup, then we are done by Lemma~\ref {conjecture true if enough on line lemma}. Therefore we may assume that there are at most $k$ points in each subgroup; in particular, $m_x\le k$ for all $x\in\F_p^2$. We now argue that if $\Sigma A$ is small, then there must be lots of pairs of elements of $A$ that add to the same element of $\F_p^2$, at which point we will be able to invoke Lemma~\ref {any j lemma}. We may assume that $\Sigma A \ne \F_p^2$, for otherwise we are done immediately. For each $1 \le i \le k$, we define the level set $A_i = \{x \in \F_p^2 : m_x \ge i\}$. Notice that $A$ can be written precisely as the multiset union $A_1 \cup A_2 \cup \cdots \cup A_k$, and so $\sum_{i=1}^k \#A_i = |A| = p+k$. Let $B_i$ be the multiset formed by the sums of pairs of elements of $A_i$ not in the same subgroup: \[ B_i = \big\{ x + y \colon x,y\in A_i,\, \lin x\ne\lin y \big\}. \] Note that $\text{\bf 0} \notin B_i$ (the restriction $\lin x\ne\lin y$ ensures that $x \ne -y$) and that every element of $B_i$ occurs with even multiplicity (the restriction $\lin x\ne\lin y$ ensures that $x \ne y$). It is not hard to estimate the relative sizes of $\#A_i$ and $|B_i|$: for each $x \in A_i$ there are at most $k$ elements of $A$ lying in the subgroup $\lin x$. Since each such $x$ occurs with multiplicity at least $i$ in $A$, there are at most $k/i$ distinct values of $y$ excluded by the condition $\lin x \ne \lin y$. Hence $|B_i| \ge \#A_i (\#A_i - k/i)$, which implies that \begin{equation}\label{bound on a_i} \#A_i \le \frac{k}{i} + \sqrt{|B_i|}. \end{equation} Since $\sum_{i=1}^k \#A_i$ is fixed, this shows that $|B_i|$ cannot be very small on average. At the same time, $\#B_i$ cannot get very large: if $\sum_{i=1}^k \#B_i \ge (2k+1)p$, then (under our assumption that $\Sigma A \ne \F_p^2$) Lemma~\ref{kneser bound} already yields \[ \#\Sigma A \ge \sum_{i=1}^k \#\Sigma A_i - (k-1)p > \sum_{i=1}^k \#B_i - (k-1)p \ge (k+2)p. \] where the middle inequality holds because $B_i \subseteq \Sigma A_i$. We may therefore assume henceforth that \begin{equation} \label{bound on b_i} \sum_{i=1}^k \#B_i < (2k+1)p. \end{equation} Let us now introduce the weighted height parameter \begin{equation} \label{eta def} \eta = \max_{1\le i \le k} \left\{ \frac{i|B_i|}{2\#B_i} : \#B_i > 0 \right\}. \end{equation} We shall show shortly that $\eta > k+1$. Assuming so, then for some $1 \le i \le k$, we have \[ \frac{|B_i|}{2\#B_i} > \frac{k+1}{i}, \] so by the pigeonhole principle, there exists some $z \in B_i$ (in particular $z\ne\text{\bf 0}$) occurring with multiplicity greater than $2(k+1)/i$; since this multiplicity is an even integer, it must be at least $2(k+2)/i.$ For each solution $x+y = z$ corresponding to an occurrence of $z$ in $B_i$, we have by construction that $x,y \notin \lin z$ and $m_x, m_y \ge i$, so for this particular choice of~$z$, \[ \tfrac12 \sum_{t \in \F_p^2 \setminus\lin z } \min\{m_t, m_{z-t}\} \ge k+2. \] Furthermore, $\sum_{y\in\lin z} m_y \le k$ by assumption. Therefore we are free to apply Lemma~\ref{any j lemma} with $j = (k+2) - \sum_{y\in\lin z} m_y,$ which gives the lower bound \[ \#\Sigma A \ge \min\{p,k+3\} \min\bigg\{p,p - k - 3 + \sum_{y\in\lin z} m_y\bigg\} \ge (k+3)(p-k-3) \ge (k+2)p \] (the last step used the inequality $p \ge (k+3)^2$, which certainly holds under the hypotheses of the theorem). It remains only to verify that $\eta > k+1$. Summing the inequality~\eqref{bound on a_i} over all $1\le i\le k$ yields \[ p+k = \sum_{i=1}^k \#A_i \le k H_k + \sum_{i=1}^k \sqrt{|B_i|} \le kH_k + \sqrt{2\eta} \sum_{i=1}^k \sqrt{\frac{\#B_i}{i}}, \] using the definition~\eqref{eta def} of~$\eta$. We estimate the rightmost sum using Cauchy--Schwarz together with the inequality~\eqref{bound on b_i}: \[ \sum_{i=1}^k \sqrt{\frac{\#B_i}{i}} \le \bigg(\sum_{i=1}^k \#B_i\bigg)^{1/2}\bigg(\sum_{i=1}^k \frac1i\bigg)^{1/2} < \sqrt{(2k+1) p H_k}. \] Combining the previous two inequalities gives $p+k - kH_k < \sqrt{ \eta(4k+2) p H_k}$, so that \[ \eta > \frac{(p+k-kH_k)^2}{(4k+2) pH_k} > \frac{p(p+2(k-kH_k))}{(4k+2) pH_k} = \frac{(p+2k)-2kH_k}{(4k+2)H_k} \ge \frac{4(k+1)^2H_k-2kH_k}{(4k+2)H_k} \] by the hypothesis on the size of~$p$. In other words, \[ \eta > \frac{2(k+1)^2-k}{2k+1} = k+1+\frac1{2k+1}, \] which completes the proof of the theorem. \end{proof} \section{A wider conjecture} \label{conjecture section} As mentioned earlier, Conjecture~\ref{2d conjecture} is just one part of a more far-reaching conjecture concerning sumsets of multisets in ${\mathbb Z}_p^m$. Before formulating that wider conjecture, we must expand the definition of a valid multiset to ${\mathbb Z}_p^m$. \begin{definition} \label{more valid definition} Let $p$ be an odd prime, and let $m$ be a positive integer. A multiset $A$ contained in ${\mathbb Z}_p^m$ is {\em valid} if: \begin{itemize} \item $\text{\bf 0}\notin A$; and \item for each $1\le d\le m$, every subgroup of ${\mathbb Z}_p^m$ that is isomorphic to ${\mathbb Z}_p^d$ contains fewer than $dp$ points of $A$, counting multiplicity. \end{itemize} \end{definition} \noindent When $m=1$, a multiset contained in ${\mathbb Z}_p$ is valid precisely when it does not contain $0$; when $m=2$ and $|A| < 2p$, this definition of valid agrees with Definition~\ref{valid definition} for multisets contained in $\F_p^2$. Note that in particular, Definition~\ref{more valid definition}(b) implies that every valid multiset contained in ${\mathbb Z}_p^m$ has at most $mp-1$ elements, counting multiplicity. We now give an example showing that this upper bound $mp-1$ can in fact be achieved. Throughout this section, let $\{x_1,\dots,x_m\}$ denote a generating set for ${\mathbb Z}_p^m$, and let ${\mathbf K}_d = \lin{x_1,\dots,x_d}$ denote the subgroup of ${\mathbb Z}_p^m$ generated by $\{x_1,\dots,x_d\}$, so that ${\mathbf K}_d \cong {\mathbb Z}_p^d$. \begin{example} \label{valid example} Let $A_1$ be the multiset consisting of $p-1$ copies of $x_1$; for $2\le j\le m$ let $A_j = \{ x_j + ax_1 \colon 0\le a\le p-1 \}$; and define $B_m = \bigcup_{j=1}^m A_j$. Then $|B_m| = (p-1) + (m-1)p = mp-1$ and $\text{\bf 0}\notin B_m$. To verify that $B_m$ is a valid subset of ${\mathbb Z}_p^m$, let ${\mathbf H}$ be any subgroup of ${\mathbb Z}_p^m$ that is isomorphic to ${\mathbb Z}_p^d$; we need to show that $B_m$ contains fewer than $dp$ points of ${\mathbf H}$. First suppose that $x_1\notin{\mathbf H}$, which implies that $bx_1\notin{\mathbf H}$ for every nonzero multiple $bx_1$ of~$x_1$. Then for each $2\le j\le m$, at most one of the elements of $A_j$ can be in ${\mathbf H}$, since the difference of any two such elements is a nonzero multiple of $x_1$. Therefore $|B_m\cap{\mathbf H}| = \ell$ for some $1\le\ell\le m-1$, and in fact all $\ell$ of these elements are of the form $x_j+ax_1$ for $\ell$ distinct values of~$j$. Since no such element is in the subgroup spanned by the others, we conclude that $d\ge\ell$, and so the necessary inequality $|B_m\cap{\mathbf H}|=\ell\le d<dp$ is amply satisfied. Now suppose that $x_1\in{\mathbf H}$. Then for each $2\le j\le m$, either all or none of the elements of $A_j$ are in ${\mathbf H}$. By reindexing the $x_i$, we may choose an integer $1\le\ell\le m$ such that ${\mathbf H}$ contains $A_1 \cup \cdots \cup A_\ell$ and is disjoint from $A_{\ell+1} \cup \cdots \cup A_m$. In particular, $|B_m\cap{\mathbf H}| = (p-1)+(\ell-1)p = \ell p-1$. But ${\mathbf H}$ contains $\{x_1,\dots,x_\ell\}$ and hence $d\ge\ell$, so that $\ell p-1\le dp-1$ as required. \end{example} We may now state our wider conjecture; Conjecture~\ref{2d conjecture} is the special case $q=1$ of part (a) of this conjecture. \begin{conjecture} \label {any d conjecture} Let $p$ be an odd prime. Let $m$ be a positive integer, and let $A$ be a valid multiset of ${\mathbb Z}_p^m$ with $|A|\ge p$. Write $|A| = qp+k$ with $0\le k\le p-1$. \begin{enumerate} \item If $0\le k\le p-3$, then $\#\Sigma A \ge (k+2)p^q$. \item If $k=p-2$, then $\#\Sigma A \ge p^{q+1}-1$. \item If $k=p-1$, then $\#\Sigma A \ge p^{q+1}$. \end{enumerate} In particular, if $|A| = mp-1$ then $\Sigma A = {\mathbb Z}_p^m$. \end{conjecture} We remark that the quantity $dp$ in Definition~\ref{more valid definition}, bounding the number of elements in a valid multiset that can lie in a rank-$d$ subgroup, has been carefully chosen in light of this conjecture: by Conjecture~\ref {any d conjecture}(c), any valid multiset $A$ with at least $dp-1$ elements counting multiplicity must satisfy $\#\Sigma A \ge p^d$. In particular, if $A$ is a valid multiset contained in a subgroup ${\mathbf H} < {\mathbb Z}_p^m$ that is isomorphic to ${\mathbb Z}_p^d$, then $|A|\ge dp-1$ implies that $\Sigma A = {\mathbf H}$. Therefore allowing $dp$ elements in such a subgroup would always be ``wasteful''. Of course, the validity of Definition~\ref{more valid definition} for rank-$d$ subgroups depends crucially upon the truth of Conjecture~\ref {any d conjecture}(c) for $q=d-1$. The conjecture is restricted to multisets $A$ with $|A|\ge p$ because we already know the truth for smaller multisets, for which the definition of ``valid'' is simply the condition that $\text{\bf 0} \notin A$: when $|A|\le p-1$, the best possible lower bound is $\#\Sigma A \ge |A|+1$ as in Lemma~\ref {CD lemma}. We remark that Peng~\cite[Theorem 2]{P1} has proved Conjecture~\ref {any d conjecture}(c) in the case $m=2$ and $q=1$, under even a slightly weaker hypothesis; in other words, he has shown that if $A$ is a valid multiset contained in $\F_p^2$ with $|A| = 2p-1$, then $\Sigma A = \F_p^2$. (We remark that Mann and Wou~\cite{MW} have proved in the case that $A$ is actually a set---that is, a multiset with distinct elements---that $\#A = 2p-2$ suffices to force $\Sigma A = \F_p^2$.) Peng considers the higher-rank groups ${\mathbb Z}_p^m$ as well, but the multisets he allows (see \cite[Theorem 1]{P2}) form a much wider class than our valid multisets, and so his conclusions are much weaker than Conjecture~\ref{any d conjecture} for $q\ge2$. Finally, we mention that we have completely verified Conjecture~\ref{any d conjecture} by exhaustive computation for the groups ${\mathbb Z}_p^2$ with $p\le 7$ and also for the group ${\mathbb Z}_3^3$. It is easy to see that all of the lower bounds in Conjecture~\ref{any d conjecture}(a), if true, would be best possible. Given $q\ge1$ and $0\le k\le p-3$, let $A'$ be any valid multiset contained in ${\mathbf K}_q$ with $|A'| = qp-1$ (such as the one given in Example~\ref{valid example} with $m=q$), and let $A$ be the union of $A'$ with $k+1$ copies of $x_{q+1}$. Then $\Sigma A = \{ y + ax_{q+1}\colon y\in \Sigma A',\, 0\le a\le k+1\}$ and thus $\#\Sigma A' = (k+2)\#\Sigma A \le (k+2)p^q$ since $\Sigma A$ is contained in~${\mathbf K}_q$. Similarly, the fact that there exists a valid multiset contained in ${\mathbf K}_{q+1}$ with $qp+(q-1)=(q+1)p-1$ elements (such as the one given in Example~\ref{valid example} with $m=q+1$) shows that the lower bound in Conjecture~\ref{any d conjecture}(c) would be best possible, since the sumset of this multiset would still be contained in ${\mathbf K}_{q+1}$ and thus would have at most $p^{q+1}$ distinct elements. The lower bound in Conjecture~\ref{any d conjecture}(b) might seem counterintuitive, especially in comparison with the pattern established in Conjecture~\ref{any d conjecture}(a). However, we can give an explicit example showing that the lower bound $p^{q+1}-1$ for $\#\Sigma A$ cannot be increased: \begin{example} \label{border example} When $p$ is an odd prime, define $B'_m$ to be the set $B_m$ from Example~\ref{valid example} with one copy of $x_1$ removed, so that $B'_m$ contains $x_1$ with multiplicity only $p-2$. Since $B_m$ is a valid multiset contained in ${\mathbb Z}_p^m$, so is $B'_m$. We have $|B'_m| = |B_m|-1 = (mp-1)-1 = (m-1)p + (m-2)$, and we claim that $-x_1\notin \Sigma B'_m$; this will imply that $\#\Sigma B'_m \le p^m-1$, and so the lower bound for $\#\Sigma A$ in Conjecture~\ref{any d conjecture}(b) cannot be increased. (In fact it is not hard to show that every other element of ${\mathbb Z}_p^m$ is in $\Sigma B'_m$, and so $\#\Sigma B'_m$ is exactly equal to $p^m-1$.) Suppose for the sake of contradition that $-x_1\in \Sigma B'_m$, and let $C$ be a submultiset of $B'_m$ such that $-x_1 = \sum_{y\in C}y$. For each $2\le j\le m$, define $\ell_j = |C\cap A_j|= \#\big( C \cap \{ x_j + ax_1\colon 0\le a\le p-1 \} \big)$. Then \[ -x_1 = \sum_{y\in C} y = t x_1 + \ell_2 x_2 + \ell_3 x_3 + \cdots + \ell_m x_m \] for some integer~$t$. It follows from this equation that each $\ell_j$ must equal either $0$ or $p$. However, if $\ell_j = p$ then \[ \sum_{y \in C\cap A_j} y = \sum_{0\le a\le p-1} (x_j + ax_1) = px_j + \frac{p(p-1)}2 x_1 = \text{\bf 0} \] (since $p$ is odd). So in either case, if $s = |C\cap A_1|$ is the multiplicity with which $x_1$ appears in $C$, then \[ -x_1 = \sum_{y\in C} y = s x_1 + \sum_{j=2}^m \sum_{y\in C\cap A_j} y = s x_1 + \text{\bf 0} + \cdots + \text{\bf 0}. \] This is a contradiction, however, since $s$ must lie between $0$ and $p-2$. Therefore $-x_1$ is indeed not an element of $\Sigma B'_m$, as claimed. \end{example} The line of questioning in this section turns out to be uninteresting when $p=2$: when the multiset $A$ does not contain $\text{\bf 0}$, the condition that no rank-$1$ subgroup of ${\mathbb Z}_2^m$ contain $2$ points of $A$ is simply equivalent to $A$ not containing any element with multiplicity greater than~$1$. It is easy to check that if $A$ consists of any $q$ points in ${\mathbb Z}_2^m$ that do not lie in any subgroup isomorphic to ${\mathbb Z}_2^{q-1}$, then $\Sigma A$ fills out the entire rank-$q$ subgroup generated by~$A$. In other words, the analogous definition of ``valid'' for multisets in ${\mathbb Z}_2^m$ would simply be a set of $q$ points that generate a rank-$q$ subgroup of ${\mathbb Z}_2^m$, and we would always have $\#\Sigma A = 2^{|A|} = 2^{\#A}$ for valid (multi)sets in~${\mathbb Z}_2^m$. \section*{Acknowledgments} The collaboration leading to this paper was made possible thanks to Jean--Jacques Risler, Richard Kenyon, and especially Ivar Ekeland; the authors also thank the University of British Columbia and the Institut d'\'Etudes Politiques de Paris for their undergraduate exchange program. The first author thanks Andrew Granville for conversations that explored this topic and eventually led to the formulation of the conjectures herein.

https://arxiv.org/abs/1102.5347

On the maximum number of isosceles right triangles in a finite point set

Let $Q$ be a finite set of points in the plane. For any set $P$ of points in the plane, $S_{Q}(P)$ denotes the number of similar copies of $Q$ contained in $P$. For a fixed $n$, Erdős and Purdy asked to determine the maximum possible value of $S_{Q}(P)$, denoted by $S_{Q}(n)$, over all sets $P$ of $n$ points in the plane. We consider this problem when $Q=\triangle$ is the set of vertices of an isosceles right triangle. We give exact solutions when $n\leq9$, and provide new upper and lower bounds for $S_{\triangle}(n)$.

\section{Introduction} In the 1970s Paul Erd\H{o}s and George Purdy \cite{D,E,F} posed the question, \textquotedblleft Given a finite set of points $Q$, what is the maximum number of similar copies $S_{Q}(n)$ that can be determined by $n$ points in the plane?\textquotedblright. This problem remains open in general. However, there has been some progress regarding the order of magnitude of this maximum as a function of $n$. Elekes and Erd\H{o}s \cite{B} noted that $S_{Q}\left( n\right) \leq n\left( n-1\right) $ for any pattern $Q$ and they also gave a quadratic lower bound for $S_{Q}(n)$ when $\left\vert Q\right\vert =3$ or when all the coordinates of the points in $Q$ are algebraic numbers. They also proved a slightly subquadratic lower bound for all other patterns $Q$. Later, Laczkovich and Ruzsa \cite{LR97} characterized precisely those patterns $Q$ for which $S_{Q}\left( n\right) =\Theta(n^{2})$. In spite of this, the coefficient of the quadratic term is not known for any non-trivial pattern; it is not even known if $\lim_{n\rightarrow \infty}S_Q(n)/n^2$ exists! Apart from being a natural question in Discrete Geometry, this problem also arose in connection to optimization of algorithms designed to look for patterns among data obtained from scanners, digital cameras, telescopes, etc. (See \cite{Bra, BrMoPa, BP05} for further references.) Our paper considers the case when $Q$ is the set of vertices of an isosceles right triangle. The case when $Q$ is the set of vertices of an equilateral triangle has been considered in \cite{A}. To avoid redundancy, we refer to an isosceles right triangle as an ${\mathsf{IRT}}$ for the remainder of the paper. We begin with some definitions. Let $P$ denote a finite set of points in the plane. We define $S_{\triangle}(P)$ to be the number of triplets in $P$ that are the vertices of an ${\mathsf{IRT}}$. Furthermore, let \[ S_{\triangle}(n)=\max_{|P|=n}S_{\triangle}(P). \] As it was mentioned before, Elekes and Erd\H{o}s established that $S_{\triangle}(n)=\Theta(n^2)$ and it is implicit from their work that $1/18 \leq \liminf_{n\rightarrow\infty S_{\triangle}(n)/n^{2} \leq 1$. The main goal of this paper is to derive improved constants that bound the function $S_{\triangle}(n)/n^{2}$. Specifically, in Sections \ref{sec: lower} and \ref{sec: upper}, we prove the following result: \begin{theorem} \[ 0.433064 < \liminf_{n\rightarrow\infty \frac{S_{\triangle}(n)}{n^{2}}\leq\frac{2}{3} < 0.66667. \] \end{theorem} We then proceed to determine, in Section \ref{sec: small cases}, the exact values of $S_{\triangle }(n)$ when $3\leq n\leq9$. Several ideas for the proofs of these bounds come from the equivalent bounds for equilateral triangles in \cite{A}. \section{Lower Bound\label{sec: lower}} We use the following definition. For $z\in P$, let $R_{\pi/2}(z,P)$ be the $\pi/2$ counterclockwise rotation of $P$ with center $z$. Furthermore, let $\deg _{\pi/2}(z)$ be the number of isosceles right triangles in $P$ such that $z$ is the right-angle vertex of the triangle. If $z\in P$, then $\deg_{\pi/2}(z)$ can be computed by simply rotating our point set $P$ by $\pi/2$ about $z$ and counting the number of points in the intersection other than $z$. Therefore, \begin{equation} \deg_{\pi/2}(z)=|P\cap R_{\pi/2}(z,P)|-1. \label{degpi/2 \end{equation} Due to the fact that an ${\mathsf{IRT}}$ has only one right angle, then \[ S_{\triangle}(P) = \sum_{z\in P} \deg_{\pi/2}(z). \] That is, the sum computes the number of ${\mathsf{IRT}}$s in $P$. From this identity an initial $5/12$ lower bound can be derived for $\liminf_{n\rightarrow \infty}S_\triangle(n)/n^2$ using the set \[ P=\left\{ (x,y)\in\mathbb{Z}^{2}:0\leq x\leq \sqrt{n},0\leq y\leq \sqrt{n}\right\}. \] We now improve this bound. The following theorem generalizes our method for finding a lower bound. We denote by $\Lambda$ the lattice generated by the points $(1,0)$ and $(0,1)$; furthermore, we refer to points in $\Lambda$ as \emph{lattice points}. The next result provides a formula for the leading term of $S_{\triangle}(P)$ when our points in $P$ are lattice points enclosed by a given shape. This theorem, its proof, and notation, are similar to Theorem 2 in \cite{A}, where the authors obtained a similar result for equilateral triangles in place of ${\mathsf{IRT}}$s. \begin{theorem} \label{integral} Let K be a compact set with finite perimeter and area 1. Define $f_{K}:\mathbb{C}\rightarrow\mathbb{R}^{+}$ as $f_{K}(z)=Area(K\cap R_{\pi/2}(z,K))$ where $z\in K$. If $K_{n}$ is a similar copy of $K$ intersecting $\Lambda$ in exactly $n$ points, then \[ S_{\triangle}(K_{n}\cap\Lambda)=\left( \int_{K}f_{K}(z)\,dz\right) n^{2}+O(n^{3/2}). \] \end{theorem} \begin{proof} Given a compact set $L$ with finite area and perimeter, we have that \[ \left\vert rL\cap\Lambda\right\vert ={\mathrm{Area}(rL)} +O(r)=r^{2}\mathrm{Area}(L)+O(r), \] where $rL$ is the scaling of $L$ by a factor $r$. Therefore, \begin{align*} S_{\triangle}(K_{n}\cap\Lambda) & =\sum_{z\in K_{n}\cap\Lambda}|(\Lambda \cap K_{n})\cap R_{\pi/2}(z,(K_{n}\cap \Lambda))|-1\\ & =\sum_{z\in K_{n}\cap\Lambda}\mathrm{Area}(K_{n}\cap R_{\pi/2 (z,K_{n}))+O(\sqrt{n}). \end{align*} We see that each error term in the sum is bounded by the perimeter of $K_{n}$, which is finite by hypothesis. Thus, \begin{align*} S_{\triangle}(K_{n}\cap\Lambda) & =n^{2}\sum_{z\in K_{n}\cap\Lambda}\frac {1}{n^{2}}\mathrm{Area}(K_{n}\cap R_{\pi/2}(z,K_{n}))+O(n^{3/2})\\ & =n^{2}\sum_{z\in K_{n}\cap\Lambda}\frac{1}{n}\mathrm{Area}(\frac{1 {\sqrt{n}}\left( K_{n}\cap R_{\pi/2}(z,K_{n})\right) )+O(n^{3/2})\\ & =n^{2}\sum_{z\in K_{n}\cap\Lambda}\frac{1}{n}\mathrm{Area}\left( \frac {1}{\sqrt{n}}K_{n}\cap R_{\pi/2}\left( \frac{z}{\sqrt{n}},\frac{1}{\sqrt{n }K_{n}\right) \right) +O(n^{3/2})\text{. \end{align*} The last sum is a Riemann approximation for the function $f_{(1/\sqrt{n )K_{n}}$ over the region $(1/\sqrt{n})K_{n}$, thus \[ S_{\triangle}(K_{n}\cap\Lambda)=n^{2}\left( \int_{\frac{1}{\sqrt{n}}K_{n }f_{\frac{1}{\sqrt{n}}K_{n}}(z)\,dz+O\left( \frac{1}{\sqrt{n}}\right) \right) +O(n^{3/2}). \] Since \[ \mathrm{Area}\left( \frac{1}{\sqrt{n}}K_{n}\right) =\frac{1}{n \mathrm{Area}(K_{n})=\frac{1}{n}(n+O(\sqrt{n}))=1+O\left( \frac{1}{\sqrt{n }\right) =\mathrm{Area}(K)+O\left( \frac{1}{\sqrt{n}}\right) , \] it follows that, \[ \int_{\frac{1}{\sqrt{n}}K_{n}}f_{\frac{1}{\sqrt{n}}K_{n}}(z)\,dz=\int_{K f_{K}(z)\,dz+O\left( \frac{1}{\sqrt{n}}\right) \text{. \] As a result, \begin{align*} S_{\triangle}(K_{n}\cap\Lambda) & =n^{2}\int_{\frac{1}{\sqrt{n}}K_{n }f_{\frac{1}{\sqrt{n}}K_{n}}(z)\,dz+O(n^{3/2})\\ & =n^{2}\int_{K}f_{K}(z)\,dz+O(n^{3/2})\text{.}\qedhere \end{align*} \end{proof} The importance of this theorem can be seen immediately. Although our $5/12$ lower bound for $\liminf_{n\rightarrow\infty} S_{\triangle}(n)/n^{2}$ was derived by summing the degrees of each point in a square lattice, the same result can be obtained by letting $K$ be the square $\{ (x,y):|x| \leq \frac{1}{2}, |y| \leq \frac{1}{2} \}$. It follows that $f_K(x,y)=(1-|x|-|y|)(1-||x|-|y||)$ and \[ S_{\triangle}(K_{n}\cap\Lambda)=\left( \int_{K}f_{K}(z)\,dz\right) n^{2}+O(n^{3/2}) = \frac{5}{12}n^{2}+O(n^{3/2})\text{. \] An improved lower bound will follow provided that we find a set $K$ such that the value for the integral in Theorem \ref{integral} is larger than $5/12$. We get a larger value for the integral by letting $K$ be the circle $\{ z \in \mathbb{C} : |z|\leq 1/\sqrt{\pi} \}$. In this case \begin{equation} \label{circle function} f_K(z)=\frac{2}{\pi} \arccos (\frac{\sqrt{2 \pi}}{2} |z|)-|z|\sqrt{\frac{2}{\pi}-|z|^2} \end{equation} and \[ S_{\triangle}(K_{n}\cap\Lambda)=\left( \int_{K}f_{K}(z)\,dz\right) n^{2}+O(n^{3/2}) = \left( \frac{3}{4}-\frac{1}{\pi}\right)n^{2}+O(n^{3/2})\text{. \] It was conjectured in \cite{A} that not only does $\lim_{n\rightarrow \infty} E(n)/n^2$ exist, but it is attained by the uniform lattice in the shape of a circle. ($E(n)$ denotes the maximum number of equilateral triangles determined by $n$ points in the plane.) The corresponding conjecture in the case of the isosceles right triangle turns out to be false. That is, if $\lim_{n\rightarrow \infty}S_{\triangle}(n)/n^2$ exists, then it must be strictly greater than $3/4-1/\pi$. Define $\overline{\Lambda}$ to be the translation of $\Lambda$ by the vector $(1/2,1/2)$. The following lemma will help us to improve our lower bound. \begin{lemma} \label{Lem: lambda} If $(j,k)\in \mathbb{R}^2$ and $\Lambda^\prime=\Lambda$ or $\Lambda^\prime=\overline{\Lambda}$, then \[ R_{\pi/2}((j,k),\Lambda^\prime) \cap \Lambda^\prime=\left\{ \begin{array}{l} \Lambda^\prime \text{ if } (j,k)\in \Lambda \cup \overline{\Lambda}, \\ \varnothing \text{ else.} \end{array} \right. \] \end{lemma} \begin{proof} Observe that \[R_{\pi/2} ((j,k),(s,t)) = \begin{pmatrix} 0 & -1\\ 1& 0\\ \end{pmatrix} \begin{pmatrix} s-j\\ t-k\\ \end{pmatrix} +\begin{pmatrix} j\\ k\\ \end{pmatrix} = \begin{pmatrix} k-t+j\\ s-j+k\\ \end{pmatrix}. \] First suppose $(s,t)\in \Lambda$. Since $s,t\in \mathbb{Z}$, then $(k-t+j,s-j+k)\in \Lambda$ if and only if $k-j\in \mathbb{Z}$ and $k+j\in \mathbb{Z}$. This can only happen when either both $j$ and $k$ are half-integers (i.e., $(j,k)\in \overline{\Lambda}$), or both $j$ and $k$ are integers (i.e., $(j,k)\in \Lambda$). Now suppose $(s,t)\in \overline{\Lambda}$. In this case, because both $s$ and $t$ are half-integers, we conclude that $(k-t+j, s-j+k)\in \overline{\Lambda}$ if and only if both $k-j\in \mathbb{Z}$ and $k+j\in \mathbb{Z}$. Once again this occurs if and only if $(j,k)\in \Lambda \cup \overline{\Lambda}$. \end{proof} Recall that if $K$ denotes the circle of area 1, then $(3/4-1/\pi)n^2$ is the leading term of $S_{\triangle}(K_n\cap \Lambda)$. The previous lemma implies that, if we were to adjoin a point $z\in \mathbb{R}^2$ to $K_n\cap \Lambda$ such that $z$ has half-integer coordinates and is located near the center of the circle formed by the points of $K_n\cap \Lambda$, then $\deg_{\pi/2}(z)$ will approximately equal $|K_n\cap \Lambda|$. We obtain the next theorem by further exploiting this idea. \begin{theorem} \[.43169\approx \frac{3}{4}-\frac{1}{\pi} < .433064 < \liminf_{n\rightarrow \infty}\frac{S_{\triangle}(n)}{n^2}\] \end{theorem} \begin{proof} Let $K$ be the circle of area 1, $A= K_{m_1}\cap \Lambda$, and $B = K_{m_2}\cap\overline{\Lambda}$. Moreover, position $B$ so that its points are centered on the circle formed by the points in $A$ (See Figure \ref{Fig: Centered}). We let $n=m_1+m_2=|A \cup B|$ and $m_2=x\cdot m_1$, where $0<x<1$ is a constant to be determined. \begin{figure}[h] \centering \includegraphics{AUBandPlot.eps} \caption{(a) Set $B$ (gray points) centered on set $A$ (black points), (b) Plot of the $n^2$ coefficient of $S_\triangle(A \cup B)$ as $x$ ranges from 0 to 1.} \label{Fig: Centered} \end{figure} We proceed to maximize the leading coefficient of $S_{\triangle}(A\cup B)$ as $x$ varies from 0 to $1$. By Lemma \ref{Lem: lambda}, there cannot exist an \irt$\:$whose right-angle vertex lies in $A$ while one $\pi/4$ vertex lies in $A$ and the other lies in $B$. Similarly, there cannot exist an \irt whose right angle-vertex lies in $B$ while one $\pi/4$ vertex lies in $A$ and the other lies in $B$. Therefore, each \irt with vertices in $A\cup B$ must fall under one of the following four cases: \bigskip \\ \noindent \textit{Case 1: All three vertices in $A$}. Using Theorem \ref{integral}, it follows that there are $(3/4 - 1/\pi)m_1^2 + O(m_1^{3/2})$ {\irt}s in this case. Since $m_1=n/(1+x)$, the number of {\irt}s in terms of $n$ equals \begin{equation}\label{Case 1} \left(\frac{3}{4} - \frac{1}{\pi}\right)\frac{n^2}{(1+x)^2} + O(n^{3/2}). \end{equation} \\ \noindent \textit{Case 2: All three vertices in $B$}. By Theorem \ref{integral}, there are $(3/4 - 1/\pi)m_2^2 + O(m_2^{3/2})$ {\irt}s in this case. This time $m_2=nx/(1+x)$ and the number of {\irt}s in terms of $n$ equals \begin{equation}\label{Case 2} \left(\frac{3}{4} - \frac{1}{\pi}\right)\frac{n^2x^2}{(1+x)^2} + O(n^{3/2}). \end{equation} \\ \noindent \textit{Case 3: Right-angle vertex in $B$, $\pi/4$ vertices in $A$}. The relationship given by Lemma \ref{Lem: lambda} allows us to slightly adapt the proof of Theorem \ref{integral} in order to compute the number of {\irt}s in this case. The integral approximation to the number of {\irt}s in this case is given by \[\sum_{z\in K_{m_2}\cap \overline{\Lambda}} |(K_{m_1}\cap\Lambda)\cap R_{\pi/2}(z,(K_{m_1}\cap\Lambda))| = m_1^2\left(\int_{\frac{1}{\sqrt{m_1}}K_{m_2}} f_{\frac{1}{\sqrt{m_1}}K_{m_1}}(z)\,dz \right) + O(m_1^{3/2}).\] But \[\mathrm{Area}\left(\frac{1}{\sqrt{m_1}}K_{m_2}\right) = \mathrm{Area}\left(\sqrt{\frac{m_2}{m_1}}K\right) + O(\sqrt{m_1}),\] so \[m_1^2\left(\int_{\frac{1}{\sqrt{m_1}}K_{m_2}} f_{\frac{1}{\sqrt{m_1}}K_{m_1}}(z)\,dz \right) + O(m_1^{3/2}) = m_1^2\left(\int_{\sqrt{\frac{m_2}{m_1}}K} f_{K}(z)\,dz \right) + O(m_1^{3/2}).\] Expressing this value in terms of $n$ gives \begin{equation}\label{Case 3} \left(\int_{\sqrt{x}K} f_{K}(z)\,dz \right)\frac{n^2}{(1+x)^2} + O(n^{3/2}). \end{equation} \\ \noindent \textit{Case 4: Right-angle vertex in $A$, $\pi/4$ vertices in $B$}. As in Case 3, the number of {\irt}s is given by \begin{equation}\label{Eqn: integral} \sum_{z\in K_{m_1}\cap \Lambda} |(K_{m_2}\cap\overline{\Lambda})\cap R_{\pi/2}(z,(K_{m_2}\cap \overline{\Lambda}))| = m_2^2\left(\int_{\frac{1}{\sqrt{m_2}}K_{m_1}} f_{\frac{1}{\sqrt{m_2}}K_{m_2}}(z)\,dz \right) + O(m_2^{3/2}). \end{equation} Now recall that $f_{(1/\sqrt{m_2})K_{m_2}}(z) = \mathrm{Area}\left((1/\sqrt{m_2})K_{m_2}\cap R_{\pi/2}(z,(1/\sqrt{m_2})K_{m_2})\right)$. It follows that $f_{(1/\sqrt{m_2})K_{m_2}}(z_0) = 0$ if and only if $z_0$ is farther than $\sqrt{2/\pi}$ from the center of $(1/\sqrt{m_2})K_{m_2}$. Thus for small enough values of $m_2$, the region of integration in Equation (\ref{Eqn: integral}) is actually $(\sqrt{2/m_2})K_{m_2}$, so it does not depend on $m_1$. We consider two subcases. First, if $x \leq 1/2$ (i.e., $m_2 \leq m_1/2$), then \[\sqrt{\frac{2}{\pi}} = \frac{1}{\sqrt{m_2}}\frac{\sqrt{2m_2}}{\sqrt{\pi}} \leq \frac{1}{\sqrt{m_2}}\frac{\sqrt{2}}{\sqrt{\pi}}\frac{\sqrt{m_1}}{\sqrt{2}} = \frac{1}{\sqrt{m_2}}\sqrt{\frac{m_1}{\pi}}.\] The left side of the above inequality is the radius of $(\sqrt{2/m_2})K_{m_2}$, meanwhile the right side is the radius of $(1/\sqrt{m_2})K_{m_1}$, thus the region of integration where $f_{\frac{1}{\sqrt{m_2}}K_{m_2}}$ is nonzero equals $(\sqrt{2/m_2})K_{m_2}$. Hence, the number of {\irt}s equals \begin{align}\nonumber m_2^2\left(\int_{\sqrt{\frac{2}{m_2}}K_{m_2}} f_{\frac{1}{\sqrt{m_2}}K_{m_2}}(z)\,dz \right) + O(m_2^{3/2}) &= m_2^2\left(\int_{\sqrt{2}K} f_{K}(z)\,dz \right) + O(m_2^{3/2})\\ \label{Case 4A} &= \left(\int_{\sqrt{2}K} f_{K}(z)\,dz \right)x^2 n^2 + O(n^{3/2}). \end{align} Now we consider the case $x>1/2$ (i.e., $m_2> m_1/2$). In this case, $f_{\frac{1}{\sqrt{m_2}}K_{m_2}}$ is nonzero for all points in $\frac{1}{\sqrt{m_2}}K_{m_1}$. Thus the number of {\irt}s in this case equals \begin{align}\nonumber m_2^2\left(\int_{\frac{1}{\sqrt{m_2}}K_{m_1}} f_{\frac{1}{\sqrt{m_2}}K_{m_2}}(z)\,dz \right) + O(m_2^{3/2}) &= m_2^2\left(\int_{\sqrt{\frac{m_1}{m_2}}K} f_{K}(z)\,dz \right) + O(m_2^{3/2})\\ \label{Case 4B} &=\left(\int_{\sqrt{\frac{1}{x}}K} f_{K}(z)\,dz \right)\frac{n^2x^2}{(1+x)^2} + O(n^{3/2}) \end{align} By Equation (\ref{circle function}), we have that for $t>0$, \begin{align*} \int_{tK}f_{K}(z)\,dz =&2\pi \int_{0}^{t/\sqrt{\pi }}\left( \frac{2}{\pi } \arccos ( \frac{\sqrt{2\pi }}{2}r) -r\sqrt{\frac{2}{\pi }-r^{2}} \right) r\,dr \\ =&\frac{1}{2\pi }\left( 4t^2\arccos ( \frac{t}{\sqrt{2}}) +2\arcsin ( \frac{t}{\sqrt{2}}) -t(t^{2}+1)\sqrt{2-t^{2}}\right). \end{align*} Therefore, putting all four cases together (i.e., expressions (\ref{Case 1}), (\ref{Case 2}), (\ref{Case 3}), and either (\ref{Case 4A}) or (\ref{Case 4B})), we obtain that the $n^2$ coefficient of $S_\triangle (A\cup B)$ equals \begin{equation*} \frac{1}{4\pi (x+1)^{2}}\left(8x\arccos \sqrt{\frac{x}{2}}+4\arcsin \sqrt{\frac{ }{2}}+(5\pi -4)x^{2}+ (3\pi -4)-2(x+1)\sqrt{2x-x^{2}}\right) \end{equation*} if $0<x\leq1/2$, or \begin{multline*} \frac{1}{4\pi (x+1)^{2}}\left(8x\left( \arccos \sqrt{\frac{x}{2}}+\arccos \sqrt{ \frac{1}{2x}}\right) +4\arcsin \sqrt{\frac{x}{2}}+4x^{2}\arcsin \sqrt{\frac{1 }{2x}}+ \right.\\ \left. (3\pi -4)(x^{2}+1)-2(x+1)\left( \sqrt{2x-x^{2}}+\sqrt{2x-1}\right) \right) \end{multline*} if $1/2<x<1$. Letting $x$ vary from 0 to 1, it turns out that this coefficient is maximized (see Figure \ref{Fig: Centered}) when $x\approx .0356067$ (this corresponds to when the radius of $B$ is approximately 18.87\% of the radius of $A$). Letting $x$ equal this value gives $0.433064$ as a decimal approximation to the maximum value attained by the $n^2$ coefficient. \end{proof} At this point, one might be tempted to further increase the quadratic coefficient by placing a third set of lattice points arranged in a circle and centered on the circle formed by $B$. It turns out that forming such a configuration does not improve the results in the previous theorem. This is due to Lemma \ref{Lem: lambda}. More specifically, given our construction from the previous theorem, there is no place to adjoin a point $z$ to the center of $A\cup B$ such that $z\in \Lambda$ or $z\in \overline{\Lambda}$. Hence, if we were to add the point $z$ to the center of $A\cup B$, then any new {\irt}s would have their right-angle vertex located at $z$ with one $\pi/4$ vertex in $A$ and the other $\pi/4$ vertex in $B$. Doing so can produce at most $2m_2 = 2xm_1\approx .0712m_1$ new {\irt}s (recall that $x\approx .0356066$ in our construction). On the other hand, adding $z$ to the perimeter of $A$, gives us $m_1f_{K}(1/\sqrt{\pi}) \approx .1817m_1$ new {\irt}s. \section{Upper Bound\label{sec: upper}} We now turn our attention to finding an upper bound for $S_{\triangle }(n)/n^{2}$. It is easy to see that $S_{\triangle}(n)\leq n^{2}-n$, since any pair of points can be the vertices of at most 6 ${\mathsf{IRT}}$s. Our next theorem improves this bound. The idea is to prove that there exists a point in $P$ that does not belong to many ${\mathsf{IRT}}$s. First, we need the following definition. For every $z\in P$, let $R_{\pi/4}^{+}(z,P)$ and $R_{\pi/4}^{-}(z,P)$ be the dilations of $P$, centered at $z$, by a factor of $\sqrt{2}$ and $1/\sqrt{2}$, respectively; followed by a $\pi/4$ counterclockwise rotation with center $z$. Furthermore, let $\deg_{\pi/4 ^{+}(z)$ and $\deg_{\pi/4}^{-}(z)$ be the number of isosceles right triangles $zxy$ with $x,y\in P$ such that $zxy$ is ordered in counterclockwise order, and $zy$, respectively $zx$, is the hypotenuse of the triangle $zxy$. Much like the case of $\deg_{\pi/2}$, $\deg_{\pi/4}^{+}$ and $\deg_{\pi/4 ^{-}$ can be computed with the following identities, \[ \deg_{\pi/4}^{+}\left( z\right) =\left\vert P\cap R_{\pi/4}^{+}(z,P)\right\vert -1\text{ and }\deg_{\pi/4}^{-}\left( z\right) =\left\vert P\cap R_{\pi/4 ^{-}(z,P)\right\vert -1\text{. \] \begin{theorem} \label{theorem3}For $n\geq3$, \[ S_{\triangle}(n)\leq\left\lfloor \frac{2}{3}(n-1)^{2}-\frac{5}{3}\right\rfloor . \] \end{theorem} \begin{proof} By induction on $n$. If $n=3$, then $S_{\triangle}(3)\leq1 = \left\lfloor \left( 2\cdot4-5\right) /3\right\rfloor$. Now suppose the theorem holds for $n=k$. We must show this implies the theorem holds for $n=k+1$. Suppose that there is a point $z\in P$ such that $\deg_{\pi/2}(z)+\deg_{\pi/4 ^{+}(z)+\deg_{\pi/4}^{-}(z)\leq\lfloor(4n-5)/3\rfloor$. Then by induction, \begin{align*} S_{\triangle}(k+1) & \leq\deg_{\pi/2}(z)+\deg_{\pi/4}^{+}(z)+\deg_{\pi /4}^{-}(z)+S_{\triangle}(k)\\ & \leq\left\lfloor \frac{4k-1}{3}\right\rfloor +\left\lfloor \frac{2 {3}(k-1)^{2}-\frac{5}{3}\right\rfloor =\left\lfloor \frac{2}{3}k^{2}-\frac {5}{3}\right\rfloor . \end{align*} The last equality can be verified by considering the three possible residues of $k$ when divided by 3. Hence, our theorem is proved if we can find a point $z\in P$ with the desired property. Let $x,y\in P$ be points such that $x$ and $y$ form the diameter of $P$. In other words, if $w\in P$, then the distance from $w$ to any other point in $P$ is less than or equal to the distance from $x$ to $y$. We now prove that either $x$ or $y$ is a point with the desired property mentioned above. We begin by analyzing $\deg_{\pi/4}^{-}$. We use the same notation from Theorem 1 in \cite{A}. Define $N_{x}=P \cap R_{\pi/4}^{-}(x,P)\backslash\{x\}$ and $N_{y}=P \cap R_{\pi/4 ^{-}(y,P)\backslash\{y\}$. It follows from our identities that, $\deg_{\pi /4}^{-}(x)=\vert N_{x}\vert $ and $\deg_{\pi/4}^{-}(y)=\vert N_{y}\vert $. Furthermore, by the Inclusion-Exclusion Principle for finite sets, we have $\vert N_{x}\vert +\vert N_{y}\vert =\vert N_{x}\cup N_{y}\vert +\vert N_{x}\cap N_{y}\vert .$ We shall prove by contradiction that $|N_{x}\cap N_{y}|\leq1$. Suppose that there are two points $u,v\in N_{x}\cap N_{y}$. This means that there are points $u_{x},v_{x},u_{y},v_{y}\in P$ such that the triangles $xu_{x}u,xv_{x v,yu_{y}u,yv_{y}v$ are ${\mathsf{IRT}}$s oriented counterclockwise with right angle at either $u$ or $v$. \begin{figure}[h] \begin{center} \includegraphics[scale=.91]{thelem4.eps} \end{center} \caption{Proof of Theorem 4. \label{theoremlemma4 \end{figure} But notice that the line segments $u_{x}u_{y}$ and $v_{x}v_{y}$ are simply the $(\pi/2$)-counterclockwise rotations of $xy$ about centers $u$ and $v$ respectively. Hence, $u_{x}u_{y}v_{x}v_{y}$ is a parallelogram with two sides having length $xy$ as shown in Figure \ref{theoremlemma4}(a). This is a contradiction since one of the diagonals of the parallelogram is longer than any of it sides. Thus, $|N_{x}\cap N_{y}|\leq1$. Furthermore, $x\notin N_{y}$ and $y\notin N_{x}$, so $|N_{x}\cup N_{y}|\leq n-2$ and thus \[ \deg_{\pi/4}^{-}(x)+\deg_{\pi/4}^{-}(y)=\left\vert N_{x} \cup N_{y} \right\vert +\left\vert N_x \cap N_{y}\right\vert \leq n-2+1=n-1\text{. \] This also implies that \[ \deg_{\pi/4}^{+}(x)+\deg_{\pi/4}^{+}(y)\leq n-1, \] since we can follow the exact same argument applied to the reflection of $P$ about the line $xy$. We now look at $\deg_{\pi/2}(x)$ and $\deg_{\pi/2}(y)$. First we need the following lemma. \begin{lemma} For every $p\in P$, at most one of $R_{\pi/2}(x,p)$ or $R_{\pi/2}(y,p)$ belongs to $P$. \end{lemma} \begin{proof} Let $p_{x}=R_{\pi/2}(x,p)$ and $p_{y}=R_{\pi/2}(y,p)$ (see Figure \ref{theoremlemma4}(b)). Note that the distance $p_{x}p_{y}$ is exactly the distance $xy$ but scaled by $\sqrt{2}$. This contradicts the fact that $xy$ is the diameter of $P$. \end{proof} Let us define a graph $G$ with vertex set $V(G)=P\backslash\{x,y\}$ and where $uv$ is an edge of $G$, (i.e., $uv\in E(G)$) if and only if $v=R_{\pi/2}(x,u)$ or $v=R_{\pi/2}(y,u)$. \begin{lemma} \label{inequalitylemma \[ 0\leq\deg_{\pi/2}(x)+\deg_{\pi/2}(y)-|E(G)|\leq1\text{. \] \end{lemma} \begin{proof} The left inequality follows from the fact each edge counts an ${\mathsf{IRT}}$ in either $\deg_{\pi/2}(x)$ or $\deg_{\pi/2}(y)$ and possibly in both. However, if $uv$ is an edge of $G$ so that $v=R_{\pi/2}(x,u)$ and $u=R_{\pi /2}(y,v)$, then $xuyv$ is a square, so this can only happen for at most one edge. \end{proof} Now, let $\deg_{G}(u)$ be the number of edges in $E(G)$ incident to $u$. We prove the following lemma. \begin{lemma} \label{deglemma} For every $u\in V(G)$, $\deg_{G}(u)\leq2$. \end{lemma} \begin{proof} Suppose $uv_{1}\in E(G)$, see Figure \ref{Fig: lemma56}(a). Without loss of generality we can assume that $u=R_{\pi/2}(y,v_{1})$. If $v_{3}=R_{\pi /2}(y,u)\in P$, then we conclude that $xv_{3}>xy$ or $xv_1 >xy$, because $\angle xyv_{3 \geq\pi/2$ or $\angle xyv_1 \geq \pi/2$. This contradicts the fact that $xy$ is the diameter of $P$. Similarly, if $v_{2}$ and $v_{4}$ are defined as $u=R_{\pi/2}(x,v_{4})$ and $v_{2}=R_{\pi/2}(x,u)$, then at most one of $v_{2}$ or $v_{4}$ can be in $P$. \end{proof} \begin{figure}[h] \begin{center} \includegraphics[scale=.91]{lemma56.eps} \end{center} \caption{Proof of Lemmas \ref{deglemma} and \ref{pathlemma}. \label{Fig: lemma56 \end{figure} We still need one more lemma for our proof. \begin{lemma} \label{pathlemma} All paths in $G$ have length at most 2. \end{lemma} \begin{proof} We prove this lemma by contradiction. Suppose we can have a path of length 3 or more. To assist us, let us place our points on a cartesian coordinate system with our diameter $xy$ relabeled as the points $(0,0)$ and $(r,0)$, furthermore, assume $p,q\geq0$ and that the four vertices of the path of length $3$ are $(p,-q)$, $(q,p)$, $(r-p,q-r)$, and $(r-q,r-p)$. Our aim is to show that the distance between $(r-q,r-p)$ and $(r-p,q-r)$ contradicts that $r$ is the diameter of $P$. Now, if paths of length 3 were possible, then the distance between every pair of points in Figure \ref{Fig: lemma56}(b) must be less than or equal to $r$. Since $d((p,-q),(q,p))\leq r$ then $p^{2}+q^{2}\leq r^{2}/2$. Now let us analyze the square of the distance from $(r-q,r-p)$ to $(r-p,q-r)$. Because $2(p^2+q^2)\geq (p+q)^2$, it follows that \begin{align*} d^{2}((r-q,r-p),(r-p,q-r)) & =(-q+p)^{2}+(2r-p-q)^{2}\\ & =4r^{2}-4r(p+q)+2(p^{2}+q^{2})\\ & \geq 4r^{2}-4\sqrt{2}r\sqrt{p^2+q^2}+2(p^{2}+q^{2})=\left(2r-\sqrt{2(p^2+q^2)}\right)^2. \end{align*} But $\sqrt{2(p^2+q^2)} \leq r$, so $(2r-\sqrt{2(p^2+q^2)})\geq r$ and thus \[ d^{2}((r-q,r-p),(r-p,q-r))\geq r^{2}. \] Equality occur if and only if $p=r/2$ and $q=r/2$; otherwise, $d((r-q,r-p),(r-p,q-r))$ is strictly greater than $r$, contradicting the fact that the diameter of $P$ is $r$. Therefore if $p\neq r/2$ or $q\neq r/2$ then there is no path of length 3. In the case that $p=r/2$ and $q=r/2$ the points $(q,p)$ and $(r-q,r-p)$ become the same and so do the points $(p,-q)$ and $(r-p,q-r)$. Thus we are left with a path of length 1. \end{proof} It follows from Lemmas \ref{deglemma} and \ref{pathlemma} that all paths of length 2 are disjoint. In other words, $G$ is the union of disjoint paths of length less than or equal to 2. Let $a$ denote the number of paths of length 2 and $b$ denote the number of paths of length 1, then \[ \left\vert E(G)\right\vert =2a+b\text{ \textup{and} }3a+2b\leq n-2. \] Recall from Lemma \ref{inequalitylemma} that either $\deg_{\pi/2}(x)+\deg_{\pi/2}(y)=\left\vert E(G)\right\vert$ or $\deg_{\pi/2}(x)+\deg_{\pi/2}(y)=\left\vert E(G)\right\vert + 1$. If $\deg_{\pi/2}(x)+\deg_{\pi/2}(y)=\left\vert E(G)\right\vert$, then \[ 2\left\vert E(G)\right\vert =4a+2b\leq n-2+a\leq n-2+\frac{n-2}{3},\] so $\deg_{\pi/2}(x)+\deg_{\pi/2}(y)=\left\vert E(G)\right\vert \leq\frac{2}{3}\left( n-2\right).$ Moreover, if $\deg_{\pi/2}(x)+\deg_{\pi/2}(y)=\left\vert E(G)\right\vert +1$, then $b\geq1$ and we get a minor improvement, \[ 2\left\vert E(G)\right\vert =4a+2b\leq n-2+a\leq n-4+\frac{n-2}{3}, \] so $\deg_{\pi/2}(x)+\deg_{\pi/2}(y)=\left\vert E(G)\right\vert +1\leq \left( 2n-7\right)/3 < \frac {2}{3}\left( n-2\right) $. We are now ready to put everything together. Between the two points $x$ and $y$, we derived the following bounds: \begin{align*} \deg_{\pi/2}(x)+\deg_{\pi/2}(y) & \leq\frac{2}{3}(n-2),\\ \deg_{\pi/4}^{+}(x)+\deg_{\pi/4}^{+}(y) & \leq(n-1)\text{, and}\\ \deg_{\pi/4}^{-}(x)+\deg_{\pi/4}^{-}(y) & \leq(n-1)\text{. \end{align*} Because the degree of a point must take on an integer value, it must be the case that either $x$ or $y$ satisfies $\deg_{\pi/2}+\deg_{\pi/4}^{+}+\deg _{\pi/4}^{-}\leq\left\lfloor (4n-5)/3\right\rfloor $. \end{proof} \section{Small Cases\label{sec: small cases}} In this section we determine the exact values of $S_{\triangle}(n)$ when $3\leq n\leq9$. \begin{theorem} \label{smallcases} For $3\leq n\leq9$, $S_{\triangle}(3)=1$, $S_{\triangle }(4)=4$, $S_{\triangle}(5)=8$, $S_{\triangle}(6)=11$, $S_{\triangle}(7)=15$, $S_{\triangle}(8)=20$, and $S_{\triangle}(9)=28$. \end{theorem} \begin{figure}[h] \begin{center} \includegraphics[ height=2.75in ] {optimalsets.eps} \caption{Optimal sets achieving equality for $S_{\triangle}(n)$. \label{Fig: optimal} \end{center} \end{figure} \begin{proof} We begin with $n=3$. Since $3$ points uniquely determine a triangle, and there is an \irt$\:$with 3 points (Figure \ref{Fig: smallcases}(a)), this situation becomes trivial and we therefore conclude that $S_{\triangle}(3)=1.$ Now let $n=4$. In Figure \ref{Fig: smallcases}(b) we show a point-set $P$ such that $S_{\triangle}(P)=4$. This implies that $S_{\triangle}(4)\geq4$. However, $S_{\triangle}(4)$ is also bounded above by $\tbinom{4}{3}=4$. Hence, $S_{\triangle}(4)=4$. To continue with the proof for the remaining values of $n$, we need the following two lemmas. \begin{lemma} \label{p=4} Suppose $|P|=4$ and $S_{\triangle}(P)\geq2$. The sets in Figure \ref{Fig: smallcases}(b)--(e), not counting symmetric repetitions, are the only possibilities for such a set $P$. \end{lemma} \begin{proof} Having $S_{\triangle}(P)\geq2$ implies that we must always have more than one ${\mathsf{IRT}}$ in $P$. Hence, we can begin with a single ${\mathsf{IRT}}$ and examine the possible ways of adding a point and producing more ${\mathsf{IRT}}$s. We accomplish this task in Figure \ref{Fig: smallcases}(a). The 10 numbers in the figure indicate the location of a point, and the total number of ${\mathsf{IRT}}$s after its addition to the set of black dots. All other locations not labeled with a number do not increase the number of ${\mathsf{IRT}}$s. Therefore, except for symmetries, all the possibilities for $P$ are shown in Figures \ref{Fig: smallcases}(b)--(e). \end{proof} \begin{figure}[ph] \centering \includegraphics[height = 7.5in]{smallcases1.eps} \caption{{}Proof of Theorem 5. Each circle with a number indicates the location of a point and the total number of ${\mathsf{IRT}}$s resulting from its addition to the base set of black dots. \label{Fig: smallcases \end{figure} \begin{lemma} \label{sumlemma} Let $P$ be a finite set with $|P|=n$. Suppose that $S_{\triangle}(A)\leq b$ for all $A\subseteq P$ with $|A|=k$. Then \[ S_{\triangle}(P)\leq\left\lfloor \frac{n\left( n-1\right) \left( n-2\right) b}{k\left( k-1\right) \left( k-2\right) }\right\rfloor . \] \end{lemma} \begin{proof} Suppose that within $P$, every $k$-point configuration contains at most $b$ ${\mathsf{IRT}}$s. The number of ${\mathsf{IRT}}$s in $P$ can then be counted by adding all the ${\mathsf{IRT}}$s in every $k$-point subset of $P$. However, in doing so, we end up counting a fixed ${\mathsf{IRT}}$ exactly $\tbinom {n-3}{k-3}$ times. Because $S_{\triangle}(A)\leq b$ we get, \[ \binom{n-3}{k-3}S_{\triangle}(P) = \sum_{\substack{A\subseteq P\\\left\vert A\right\vert =k}}S_{\triangle}(A)\leq\binom{n}{k}b. \] Notice that $S_{\triangle}(P)$ can only take on integer values so, \[ S_{\triangle}(P)\leq\left\lfloor \frac{\binom{n}{k}b}{\binom{n-3}{k-3 }\right\rfloor =\left\lfloor \frac{n\left( n-1\right) \left( n-2\right) b}{k\left( k-1\right) \left( k-2\right) }\right\rfloor .\qedhere \] \end{proof} Now suppose $|P|=5$. If $S_{\triangle}(A)\leq1$ for all $A\subseteq P$ with $|A|=4$, then by Lemma \ref{sumlemma}, $S_{\triangle}(P)\leq2$. Otherwise, by Lemma \ref{p=4}, $P$ must contain one of the 4 sets shown in Figures \ref{Fig: smallcases}(b)--\ref{Fig: smallcases}(e). The result now follows by examining the possibilities for producing more ${\mathsf{IRT}}$s by placing a fifth point in the 4 distinct sets. In Figures \ref{Fig: smallcases}(b), \ref{Fig: smallcases}(c), \ref{Fig: smallcases}(d), and \ref{Fig: smallcases (e) we accomplish this task. In the same way as we did in Lemma \ref{p=4}, every number in a figure indicates the location of a point, and the total number of ${\mathsf{IRT}}$s after its addition to the set of black dots. It follows that the maximum value achieved by placing a fifth point is $8$ and so $S_{\triangle}(5)=8$. The point-set that uniquely achieves equality is shown in Figure \ref{Fig: smallcases}(f). Moreover, there is exactly one set $P$ with $S_{\triangle}(P)=6$ (shown in Figure \ref{Fig: smallcases}(g)), and two sets $P$ with $S_{\triangle}(P)=5$ (Figures \ref{Fig: smallcases}(h) and \ref{Fig: smallcases}(i)). Now suppose $|P|=6$. If $S_{\triangle}(A)\leq4$ for all $A\subseteq P$ with $|A|=5$, then by Lemma \ref{sumlemma}, $S_{\triangle}(P)\leq8$. Otherwise, $P$ must contain one of the sets in Figures \ref{Fig: smallcases}(f)--\ref{Fig: smallcases}(i). We now check all possibilities for adding more ${\mathsf{IRT}}$s by joining a sixth point to our 4 distinct sets. This is shown in Figures \ref{Fig: smallcases}(f)--\ref{Fig: smallcases}(i). It follows that the maximum value achieved is $11$ and so $S_{\triangle}(6)=11$. The point-set that uniquely achieves equality is shown in Figure \ref{Fig: smallcases}(j). Also, except for symmetries, there are exactly 3 sets $P$ with $S_{\triangle}(P)=10$ (Figures \ref{Fig: smallcases}(k)--\ref{Fig: smallcases}(m)) and only one set $P$ with $S_{\triangle}(P)=9$ (Figure \ref{Fig: smallcases}(n)). Now suppose $|P|=7$. If $S_{\triangle}(A)\leq8$ for all $A\subseteq P$ with $|A|=6$, then by the Lemma \ref{sumlemma}, $S_{\triangle}(P)\leq14$. Otherwise, $P$ must contain one of the sets in Figures \ref{Fig: smallcases (j)--\ref{Fig: smallcases}(n). We now check all possibilities for adding more ${\mathsf{IRT}}$s by joining a seventh point to our 5 distinct configurations. We complete this task in Figures \ref{Fig: smallcases (j)--\ref{Fig: smallcases}(n). Because the maximum value achieved is $15$, we deduce that $S_{\triangle}(7)=15$. In this case, there are exactly two point-sets that achieve 15 ${\mathsf{IRT}}$s. The proof for the values $n=8$ and $n=9$ follows along the same lines, but there are many more intermediate sets to be considered. We omit the details. All optimal sets are shown in Figure \ref{Fig: optimal}. \end{proof} Inspired by our method used to prove exact values of $S_\triangle(n)$, a computer algorithm was devised to construct the best 1-point extension of a given base set. This algorithm, together with appropriate heuristic choices for some initial sets, lead to the construction of point sets with many ${\mathsf{IRT}}$s giving us our best lower bounds for $S_\triangle(n)$ when $10\leq n\leq25$. These lower bounds are shown in Table 1 and the point-sets achieving them in Figure \ref{Fig: bestconst}. \begin{table}[ph] \centering \begin{tabular}{||c||c|c|c|c|c|c|c|c||} \multicolumn{9}{c}{}\\ \hline $n$ & 10 & 11 & 12 & 13 & 14 & 15 & 16 & 17\\ \hline $S_{\triangle}(n)\geq$ & 35 & 43 & 52 & 64 & 74 & 85 & 97 & 112\\ \hline \multicolumn{9}{c}{}\\ \hline $n$ & 18 & 19 & 20 & 21 & 22 & 23 & 24 & 25\\ \hline $S_{\triangle}(n)\geq$ & 124 & 139 & 156 & 176 & 192 & 210 & 229 & 252\\ \hline \end{tabular} \caption{Best lower bounds for $S_\triangle (n).$\label{tab: small values} \end{table} \begin{figure} [pht] \begin{center} \includegraphics[ height=1.69in {bestconstr.eps \caption{Best constructions $A_{n}$ for $n\leq25$. Each set $A_{n}$ is obtained as the union of the starting set (in white) and the points with label $\leq n$. The value $S_{\triangle}(A_n)$ is given by Table \ref{tab: small values}. \label{Fig: bestconst \end{center} \end{figure} \bigskip \noindent \textbf{Acknowledgements.} We thank Virgilio Cerna who, as part of the CURM mini-grant that supported this project, helped to implement the program that found the best lower bounds for smaller values of $n$. We also thank an anonymous referee for some useful suggestions and improvements to the presentation.

https://arxiv.org/abs/1912.13291

A doubly stochastic block Gauss-Seidel algorithm for solving linear equations

We propose a simple doubly stochastic block Gauss--Seidel algorithm for solving linear systems of equations. By varying the row partition parameter and the column partition parameter of the coefficient matrix, we recover the Landweber algorithm, the randomized Kaczmarz algorithm, the randomized Gauss--Seidel algorithm, and the doubly stochastic Gauss--Seidel algorithm. For general (consistent or inconsistent) linear systems, we show the exponential convergence of the {\it norms of the expected iterates} via exact formulas. For consistent linear systems, we prove the exponential convergence of the {\it expected norms of the error and the residual}. Numerical experiments are given to illustrate the efficiency of the proposed algorithm.

\section{Introduction} \label{intro} Randomized iterative algorithms for solving a linear system of equations \begin{equation}\label{lineq}{\bf Ax=b},\quad \mbf A\in\mbbr^{m\times n}, \quad \mbf b\in\mbbr^m\end{equation} have attracted much attention recently; see, for example, \cite{strohmer2009rando,leventhal2010rando,zouzias2013rando,needell2014paved,needell2015rando,ma2015conve,gower2015rando,ma2018itera,bai2018conve,bai2018greed,bai2018relax,bai2019parti,bai2019greed,du2019tight,razaviyayn2019linea,necoara2019faste,zhang2019new,liu2019varia,wu2020proje,chen2020error,niu2020greed,richtarik2017stoch}. At each step, to generate the next iterate from the current iterate, the randomized Kaczmarz algorithm \cite{strohmer2009rando} uses a randomly picked row, the randomized Gauss--Seidel (i.e., randomized coordinate descent) algorithm \cite{leventhal2010rando} uses a randomly picked column, and the doubly stochastic Gauss--Seidel algorithm \cite{razaviyayn2019linea} uses a randomly picked entry of the coefficient matrix $\mbf A$. It is natural to ask whether one can design a randomized algorithm which uses a randomly picked submatrix of $\mbf A$. In this paper, we propose a doubly stochastic block Gauss--Seidel (DSBGS) algorithm which uses a submatrix of $\mbf A$ at each step (see Algorithm 1 in \S 2). We can view DSBGS as a stochastic gradient descent for solving the following optimization problem \begin{equation}\label{opt}\min_{\mbf x\in\mbbr^n}\l\{f(\mbf x):=\frac{1}{2\|\mbf A\|_\rmf^2}\|\mbf b-\mbf A\mbf x\|_2^2\r\}.\end{equation} The Landweber iterative algorithm \cite{landweber1951itera}, the randomized Kaczmarz (RK) algorithm, the randomized Gauss--Seidel (RGS) algorithm, and the doubly stochastic Gauss--Seidel (DSGS) algorithm are special cases of our algorithm. Our algorithm does not need to use projections and Moore-Penrose pseudoinverses of submatrices, so it is different from the block algorithms in \cite{needell2014paved,needell2015rando,gower2015rando}. Numerical experiments for both synthetic data and real-world data are given to illustrate the efficiency of DSBGS. {\it Main contributions}. We propose a simple doubly stochastic block Gauss--Seidel algorithm for solving linear equations and prove its convergence theory. More specifically, we show the exponential convergence of the norms of the expected iterates via exact formulas (see Theorems \ref{main1} and \ref{main2}) for general (consistent or inconsistent) linear systems, and prove the exponential convergence of the expected norms of the error and the residual (see Theorems \ref{main3} and \ref{main4}) for consistent linear systems. {\it Organization of this paper}. In the rest of this section, we give some notation. In Section 2 we describe the doubly stochastic block Gauss--Seidel algorithm and prove its convergence theory. In Section 3 we report the numerical results. Finally, we present brief concluding remarks in Section 4. {\it Notation}. For any random variables $\bm\xi$ and $\bm\zeta$, we use $\mbbe[\bm\xi]$ and $\mbbe[\bm\xi\ |\bm\zeta]$ to denote the expectation of $\bm\xi$ and the conditional expectation of $\bm\xi$ given $\bm\zeta$, respectively. For an integer $m\geq 1$, let $[m]:=\{1,2,3,\ldots,m\}$. Lowercase (upper-case) boldface letters are reserved for column vectors (matrices). For any vector $\mbf u\in\mbbr^m$, we use $\mbf u_i$, $\bf u^\rmt$ and $\|\mbf u\|_2$ to denote, the $i$th entry, the transpose and the Euclidean norm of $\mbf u$, respectively. We use $\mbf I$ to denote the identity matrix whose order is clear from the context. For any matrix $\mbf A\in\mbbr^{m\times n}$, we use $\mbf A_{i,j}$, $\mbf A_{i,:}$, $\mbf A_{:,j}$ $\mbf A^\rmt$, $\mbf A^\dag$, $\|\mbf A\|_\rmf$, $\ran(\mbf A)$, $\rank(\mbf A)$, $\sigma_{1}(\mbf A)\geq\sigma_{2}(\mbf A)\geq\cdots\geq\sigma_{r}(\mbf A)>0$ to denote the $(i,j)$ entry, the $i$th row, the $j$th column, the transpose, the Moore-Penrose pseudoinverse, the Frobenius norm, the column space, the rank, and all the nonzero singular values of $\mbf A$, respectively. Obviously, $\rank(\mbf A)=r$. We call a matrix $\mbf A\in\mbbr^{m\times n}$ full column rank if $\rank(\mbf A)=n$ and rank deficient if $\rank(\mbf A)<n$. For index sets $\mcali\subseteq[m]$ and $\mcalj\subseteq[n]$, let $\mbf A_{\mcali,:}$, $\mbf A_{:,\mcalj}$, and $\mbf A_{\mcali,\mcalj}$ denote the row submatrix indexed by $\mcali$, the column submatrix indexed by $\mcalj$, and the submatrix that lies in the rows indexed by $\mcali$ and the columns indexed by $\mcalj$, respectively. The linear system (\ref{lineq}) is called consistent if $\mbf b\in\ran(\mbf A)$, i.e., a solution exists; otherwise, it is called inconsistent. \section{A doubly stochastic block Gauss--Seidel algorithm} Let $\{\mcali_1,\mcali_2,\ldots,\mcali_s\}$ denote a partition of $[m]$ such that, for $i,j=1,2,\ldots,s$ and $i\neq j,$ $$ \mcali_i\neq\emptyset,\quad\mcali_i\cap\mcali_j=\emptyset,\quad \bigcup_{i=1}^s\mcali_i=[m].$$ Let $\{\mcalj_1,\mcalj_2,\ldots,\mcalj_t\}$ denote a partition of $[n]$ such that, for $i,j=1,2,\ldots,t$ and $i\neq j,$ $$ \mcalj_i\neq\emptyset,\quad\mcalj_i\cap\mcalj_j=\emptyset,\quad \bigcup_{i=1}^t\mcalj_i=[n].$$ Let $$\mcalp=\{\mcali_1,\mcali_2,\ldots,\mcali_s\}\times \{\mcalj_1,\mcalj_2,\ldots,\mcalj_t\}.$$ We propose the following doubly stochastic block Gauss--Seidel algorithm (Algorithm 1) for solving the linear system $\bf Ax=b$. \begin{center} \begin{tabular*}{132mm}{l} \toprule {\bf Algorithm 1:} A doubly stochastic block Gauss--Seidel algorithm\\ \hline \qquad Let $\alpha>0$. Initialize $\mbf x^0\in\mbbr^n$\\ \qquad {\bf for} $k=1,2,\ldots,$ {\bf do}\\ \qquad \qquad Pick $(\mcali,\mcalj)\in\mcalp$ with probability $\dsp\frac{\|\mbf A_{\mcali,\mcalj}\|^2_\rmf}{\|\mbf A\|_\rmf^2}$\\ \qquad \qquad Set $\dsp\mbf x^k=\mbf x^{k-1}-\alpha \frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2} (\mbf A\mbf x^{k-1}-\mbf b)$\\ \bottomrule \end{tabular*} \end{center} Here we consider constant step size for simplicity. By varying the row partition parameter $s$ and the column partition parameter $t$, we recover the following well-known algorithms as special cases: \bit \item[$\bullet$] Landweber \cite{landweber1951itera} ($s=1$ and $t=1$), $$\mbf x^k=\mbf x^{k-1}-\alpha\frac{{\bf A}^\rmt}{\|\mbf A\|_\rmf^2}(\mbf A\mbf x^{k-1}-\mbf b).$$ \item[$\bullet$] Randomized Kaczmarz \cite{strohmer2009rando} ($s=m$ and $t=1$), $$\mbf x^k=\mbf x^{k-1}-\alpha\frac{\mbf A_{i,:}\mbf x^{k-1}-\mbf b_i}{\|\mbf A_{i,:}\|_2^2}(\mbf A_{i,:})^\rmt.$$ \item[$\bullet$] Randomized Gauss--Seidel \cite{leventhal2010rando} ($s=1$ and $t=n$), $$\dsp\mbf x^k=\mbf x^{k-1}-\alpha\frac{(\mbf A_{:,j})^\rmt(\mbf A\mbf x^{k-1}-\mbf b)}{\|\mbf A_{:,j}\|_2^2} \mbf I_{:,j}.$$ \item[$\bullet$] Doubly Stochastic Gauss--Seidel \cite{razaviyayn2019linea} ($s=m$ and $t=n$), $$\dsp\mbf x^k=\mbf x^{k-1}-\alpha \frac{\mbf A_{i,j}(\mbf A_{i,:}\mbf x^{k-1}-\mbf b_i)}{|\mbf A_{i,j}|^2}\mbf I_{:,j}.$$ \eit The conditional expectation of $\mbf x^k$ given $\mbf x^{k-1}$ is \beqas\mbbe[\mbf x^k\ |\mbf x^{k-1}]&=&\mbf x^{k-1}-\alpha\mbbe\bem\dsp\frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\eem(\mbf A\mbf x^{k-1}-\mbf b)\\ &=&\mbf x^{k-1}-\alpha\l(\sum_{(\mcali,\mcalj)\in\mcalp}\frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\frac{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}{\|\mbf A\|_\rmf^2}\r)(\mbf A\mbf x^{k-1}-\mbf b)\\ &=&\mbf x^{k-1}-\alpha\frac{{\bf A}^\rmt}{\|\mbf A\|_\rmf^2}(\mbf A\mbf x^{k-1}-\mbf b). \eeqas Note that the gradient of the objective function of the optimization problem (\ref{opt}) is $$\nabla f(\mbf x)=\frac{\mbf A^\rmt}{\|\mbf A\|_\rmf^2}(\mbf A\mbf x-\mbf b).$$ It follows $$\mbbe[\mbf x^k\ |\mbf x^{k-1}]=\mbf x^{k-1}-\alpha\nabla f(\mbf x^{k-1}).$$ Therefore, DSBGS can be viewed as a stochastic gradient descent method for solving the optimization problem (\ref{opt}). \subsection{The exponential convergence of the norms of the expected iterates} In this subsection we show the exponential convergence of the norms of the expected iterates for general (consistent or inconsistent) linear systems. The following lemma will be used to prove Theorems \ref{main1} and \ref{main2}. Its proof (via singular value decomposition) is straightforward and we omit the details. \begin{lemma}\label{leq} Let $\alpha>0$ and $\mbf A$ be any nonzero real matrix. For every $\mbf u\in\ran(\mbf A)$, it holds $$\l\|\l(\mbf I-\frac{\alpha\bf AA^\rmt}{\|\mbf A\|_\rmf^2}\r)^k\mbf u\r\|_2\leq\l(\max_{1\leq i\leq r}\l|1-\frac{\alpha\sigma_i^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r|\r)^k\|\mbf u\|_2.$$ \end{lemma} In Theorem \ref{main1}, we show the exponential convergence of the norm of the expected error for {\it consistent} linear systems. \begin{theorem}\label{main1} Let $\mbf x^k$ denote the $k$th iterate of {\rm DSBGS} applied to the consistent linear system $\bf Ax=b$ with arbitrary $\mbf x^0\in\mbbr^n$. It holds $$\|\mbbe[{\bf x}^k-\mbf x^0_\star]\|_2\leq\l(\max_{1\leq i\leq r}\l|1-\frac{\alpha\sigma_i^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r|\r)^k\|{\bf x}^0-\mbf x^0_\star\|_2,$$ where $$\mbf x^0_\star:=(\mbf I-\mbf A^\dag\mbf A)\mbf x^0+\mbf A^\dag\mbf b$$ is the orthogonal projection of $\mbf x^0$ onto the solution set $\{\mbf x\in\mbbr^n\ |\ \bf A x= b\}$. \end{theorem} \proof The conditional expectation of $\mbf x^k-\mbf x_\star^0$ given $\mbf x^{k-1}$ is \beqas \mbbe[{\bf x}^k-\mbf x_\star^0\ |\mbf x^{k-1}]&=&\mbbe[\mbf x^k \ |\mbf x^{k-1}] -\mbf x_\star^0\\ &=&\mbf x^{k-1}-\alpha\frac{{\bf A}^\rmt}{\|\mbf A\|_\rmf^2}(\mbf A\mbf x^{k-1}-\mbf b)-\mbf x_\star^0\\&=& \mbf x^{k-1}-\alpha\frac{{\bf A}^\rmt}{\|\mbf A\|_\rmf^2}(\mbf A\mbf x^{k-1}-\mbf A\mbf x_\star^0)-\mbf x_\star^0\\ &=&\l(\mbf I-\frac{\alpha\mbf A^\rmt\mbf A}{\|\mbf A\|_\rmf^2}\r) ({\bf x}^{k-1}-\mbf x_\star^0).\eeqas Taking expectation gives \beqas\mbbe[{\bf x}^k-\mbf x_\star^0]&=&\mbbe[\mbbe[{\bf x}^k-\mbf x_\star^0\ |\mbf x^{k-1}]]\\ &=& \l(\mbf I-\frac{\alpha\mbf A^\rmt\mbf A}{\|\mbf A\|_\rmf^2}\r) \mbbe[{\bf x}^{k-1}-\mbf x_\star^0]\\ &=& \l(\mbf I-\frac{\alpha\mbf A^\rmt\mbf A}{\|\mbf A\|_\rmf^2}\r)^k ({\bf x}^0-\mbf x_\star^0). \eeqas Applying the norms to both sides we obtain $$\|\mbbe[{\bf x}^k-\mbf x_\star^0]\|_2\leq\l(\max_{1\leq i\leq r}\l|1-\frac{\alpha\sigma_i^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r|\r)^k\|{\bf x}^0-\mbf x_\star^0\|_2.$$ Here the inequality follows from the fact that $${\bf x}^0-\mbf x_\star^0={\bf A^\dag Ax}^0-\mbf A^\dag\mbf b\in\ran(\mbf A^\rmt)$$ and Lemma \ref{leq}. \qed \begin{remark} If $\mbf x^0\in\ran(\mbf A^\rmt)$, then $\mbf x^0_\star=\mbf A^\dag\mbf b$. \end{remark} \begin{remark} To ensure convergence of the expected iterate, it suffices to have $$\dsp\max_{1\leq i\leq r}\l|1-\frac{\alpha\sigma_i^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r|<1\quad i.e.,\quad 0<\alpha<\dsp\frac{2\|\mbf A\|_\rmf^2}{\sigma_1^2(\mbf A)}.$$ \end{remark} In Theorem \ref{main2}, we show the exponential convergence of $\|\mbbe[{\bf Ax}^k-{\bf Ax}_\star]\|_2$ for the {\it consistent or inconsistent} linear system $\bf Ax=b$, where $\mbf x_\star$ is any solution of the normal equations $$\mbf A^\rmt\mbf A\mbf x=\mbf A^\rmt\mbf b.$$ \begin{theorem}\label{main2} Let $\mbf x^k$ denote the $k$th iterate of {\rm DSBGS} applied to the consistent or inconsistent linear system $\bf Ax=b$ with arbitrary $\mbf x^0\in\mbbr^n$. It holds $$\|\mbbe[{\bf Ax}^k-{\bf Ax}_\star]\|_2\leq\l(\max_{1\leq i\leq r}\l|1-\frac{\alpha\sigma_i^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r|\r)^k\|{\bf Ax}^0-{\bf Ax}_\star\|_2,$$ where $\mbf x_\star$ is any solution of $\mbf A^\rmt\mbf A\mbf x=\mbf A^\rmt\mbf b$. \end{theorem} \proof The conditional expectation of ${\bf Ax}^k-{\bf Ax}_\star$ given $\mbf x^{k-1}$ is \beqas \mbbe[{\bf Ax}^k-{\bf Ax}_\star\ |\mbf x^{k-1}]&=&\mbf A(\mbbe[\mbf x^k \ |\mbf x^{k-1}] - {\bf x_\star})\\ &=& \mbf A\l(\mbf x^{k-1}-\alpha\frac{{\bf A}^\rmt}{\|\mbf A\|_\rmf^2}(\mbf A\mbf x^{k-1}-\mbf b)- {\bf x_\star}\r) \\ &=& \mbf A\l(\mbf x^{k-1}-\frac{\alpha\mbf A^\rmt}{\|\mbf A\|_\rmf^2}(\mbf A\mbf x^{k-1}-\mbf A\mbf x_\star)- {\bf x_\star}\r) \qquad{\rm (by\ \bf A^\rmt b=\bf A^\rmt Ax_\star)} \\ &=& \mbf A\mbf x^{k-1} - {\bf Ax_\star}-\frac{\alpha\mbf A\mbf A^\rmt}{\|\mbf A\|_\rmf^2}(\mbf A\mbf x^{k-1}-\mbf A\mbf x_\star)\\ &=&\l(\mbf I-\frac{\alpha\mbf A\mbf A^\rmt}{\|\mbf A\|_\rmf^2}\r) ({\bf Ax}^{k-1}-{\bf Ax_\star}).\eeqas Taking expectation gives \beqas\mbbe[{\bf Ax}^k-{\bf Ax}_\star]&=&\mbbe[\mbbe[{\bf Ax}^k-{\bf Ax}_\star\ |\mbf x^{k-1}]]\\ &=& \l(\mbf I-\frac{\alpha\mbf A\mbf A^\rmt}{\|\mbf A\|_\rmf^2}\r) \mbbe[{\bf Ax}^{k-1}-{\bf Ax}_\star]\\ &=& \l(\mbf I-\frac{\alpha\mbf A\mbf A^\rmt}{\|\mbf A\|_\rmf^2}\r)^k ({\bf Ax}^0-{\bf Ax}_\star). \eeqas Applying the norms to both sides we obtain $$\|\mbbe[{\bf Ax}^k-{\bf Ax}_\star]\|_2\leq\l(\max_{1\leq i\leq r}\l|1-\frac{\alpha\sigma_i^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r|\r)^k\|{\bf Ax}^0-{\bf Ax}_\star\|_2.$$ Here the inequality follows from the fact that ${\bf Ax}^0-{\bf Ax}_\star\in\ran(\mbf A)$ and Lemma \ref{leq}. \qed \subsection{The exponential convergence of the expected norms of the error and the residual} In this subsection we prove the exponential convergence of the expected norms of the error or the residual for consistent linear systems. {The convergence depends on the positive number $\rho$ defined as $$\beta=\max_{(\mcali,\mcalj)\in\mcalp}\frac{\|\mbf A_{\mcali,\mcalj}\|_2^2}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}.$$} The following two lemmas will be used. Their proofs are straightforward and we omit the details. \begin{lemma}\label{lemma2} For any vector $\mbf u\in\mbbr^m$ and any matrix $\mbf A\in\mbbr^{m\times n}$, it holds $${\mbf u^\rmt\mbf A\mbf A^\rmt\mbf u\leq\|\mbf A\|_2^2\mbf u^\rmt\mbf u.}$$ \end{lemma} \begin{lemma}\label{lemma3} For any matrix $\mbf A\in\mbbr^{m\times n}$ with rank $r$ and any vector $\mbf u\in\ran(\mbf A)$, it holds $$\mbf u^\rmt\mbf A\mbf A^\rmt\mbf u\geq\sigma_r^2({\mbf A})\|\mbf u\|_2^2.$$ \end{lemma} For {\it full column rank consistent} linear systems, we prove the exponential convergence of the expected norm of the error in the following theorem. We recall that in this case $\bf A^\dag b$ is the unique solution of $\bf Ax=b$. \begin{theorem}\label{main3} Let $\mbf x^k$ denote the $k$th iterate of {\rm DSBGS} applied to the full column rank consistent linear system $\bf Ax=b$ with arbitrary $\mbf x^0\in\mbbr^n$. Assume $0<\alpha<2/{(t\beta)}$. It holds $$\mbbe[\|{\bf x}^k-\mbf A^\dag\mbf b\|_2^2]\leq\l(1-\frac{(2\alpha-t{\beta}\alpha^2)\sigma_n^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)^k \|\mbf x^0-\mbf A^\dag\mbf b\|_2^2.$$ \end{theorem} \proof Note that \beqas \|\mbf x^k-\mbf A^\dag\mbf b\|_2^2&=&\l\|\mbf x^{k-1}-\alpha\l(\frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)(\mbf A\mbf x^{k-1}-\mbf b)-\mbf A^\dag\mbf b\r\|_2^2\\ &=& \l\|\mbf x^{k-1}-\alpha\l(\frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)\mbf A(\mbf x^{k-1}-\mbf A^\dag\mbf b)-\mbf A^\dag\mbf b\r\|_2^2\\ &=& \l\|\mbf x^{k-1}-\mbf A^\dag\mbf b-\alpha\l(\frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt\mbf A}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)(\mbf x^{k-1}-\mbf A^\dag\mbf b)\r\|_2^2 \\ &=& \|\mbf x^{k-1}-\mbf A^\dag\mbf b\|_2^2-2\alpha(\mbf x^{k-1}-\mbf A^\dag\mbf b)^\rmt\l(\frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt\mbf A}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)(\mbf x^{k-1}-\mbf A^\dag\mbf b)\\ && + \alpha^2(\mbf x^{k-1}-\mbf A^\dag\mbf b)^\rmt\l(\frac{\mbf A^\rmt\mbf I_{:,\mcali}\mbf A_{\mcali,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt\mbf A}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^4}\r)(\mbf x^{k-1}-\mbf A^\dag\mbf b)\\ &\leq& \|\mbf x^{k-1}-\mbf A^\dag\mbf b\|_2^2-2\alpha(\mbf x^{k-1}-\mbf A^\dag\mbf b)^\rmt\l(\frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt\mbf A}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)(\mbf x^{k-1}-\mbf A^\dag\mbf b)\\ && + {\alpha^2(\mbf x^{k-1}-\mbf A^\dag\mbf b)^\rmt\l(\frac{\|\mbf A_{\mcali,\mcalj}\|_2^2}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\cdot\frac{\mbf A^\rmt\mbf I_{:,\mcali}(\mbf I_{:,\mcali})^\rmt\mbf A}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)(\mbf x^{k-1}-\mbf A^\dag\mbf b)\quad \mbox{(by Lemma \ref{lemma2})} }\\ & \leq & { \|\mbf x^{k-1}-\mbf A^\dag\mbf b\|_2^2-2\alpha(\mbf x^{k-1}-\mbf A^\dag\mbf b)^\rmt\l(\frac{\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt\mbf A}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)(\mbf x^{k-1}-\mbf A^\dag\mbf b)}\\ && {+ \beta\alpha^2(\mbf x^{k-1}-\mbf A^\dag\mbf b)^\rmt\l(\frac{\mbf A^\rmt\mbf I_{:,\mcali}(\mbf I_{:,\mcali})^\rmt\mbf A}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)(\mbf x^{k-1}-\mbf A^\dag\mbf b)}. \eeqas Taking expectation gives \beqas \mbbe[\|{\bf x}^k-\mbf A^\dag\mbf b\|_2^2\ |\mbf x^{k-1}]&\leq&\|\mbf x^{k-1}-\mbf A^\dag\mbf b\|_2^2-(2\alpha-t{\beta}\alpha^2)(\mbf x^{k-1}-\mbf A^\dag\mbf b)^\rmt\l(\frac{\mbf A ^\rmt\mbf A}{\|\mbf A\|_\rmf^2}\r)(\mbf x^{k-1}-\mbf A^\dag\mbf b)\\ &\leq&\l(1-\frac{(2\alpha-t{\beta}\alpha^2)\sigma_n^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)\|\mbf x^{k-1}-\mbf A^\dag\mbf b\|_2^2.\quad ({\rm by\ Lemma\ \ref{lemma3}}) \eeqas Taking expectation again gives \beqas \mbbe[\|{\bf x}^k-\mbf A^\dag\mbf b\|_2^2]&=&\mbbe[\mbbe[\|{\bf x}^k-\mbf A^\dag\mbf b\|_2^2\ |\mbf x^{k-1}]] \\ &\leq&\l(1-\frac{(2\alpha-t{\beta}\alpha^2)\sigma_n^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)\mbbe[\|\mbf x^{k-1}-\mbf A^\dag\mbf b\|_2^2]\\ &\leq&\l(1-\frac{(2\alpha-t{\beta}\alpha^2)\sigma_n^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)^k \|\mbf x^0-\mbf A^\dag\mbf b\|_2^2.\qquad \qed\eeqas \begin{remark} For the special case $s=m$, $t=1$, $\alpha=1$ (i.e., the randomized Kaczmarz algorithm{, we have $\beta=1$ in this case}) and the special case $s=m$, $t=n$, $\alpha=1/n$ (i.e., the doubly stochastic Gauss--Seidel algorithm{, we have $\beta=1$ in this case}), the results of Theorem \ref{main3} are given in \cite[Theorem 2]{strohmer2009rando} and \cite[Theorem 1]{razaviyayn2019linea}, respectively. \end{remark} \begin{remark} For rank deficient consistent linear systems, if $t=1$ and $\mbf x^0\in\mbbr^n$, we can show $\mbf x^k-\mbf x_\star^0\in\ran(\mbf A^\rmt)$ by induction, where $\mbf x^0_\star=(\mbf I-\mbf A^\dag\mbf A)\mbf x^0+\mbf A^\dag\mbf b$ is the orthogonal projection of $\mbf x^0$ onto the solution set $\{\mbf x\in\mbbr^n\ |\ \bf A x= b\}$. Then by the same approach as used in the proof of Theorem \ref{main3}, for any $s\in[m]$ and $t=1$, we can prove the convergence bound $$\mbbe[\|{\bf x}^k-\mbf x_\star^0\|_2^2]\leq\l(1-\frac{(2\alpha-{\beta}\alpha^2)\sigma_r^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)^k \|\mbf x^0-\mbf x_\star^0\|_2^2.$$ This result for the special case $s=m$ and $t=1$ (i.e., the randomized Kaczmarz algorithm{, we have $\beta=1$ in this case}) with $\mbf x^0\in\ran(\mbf A^\rmt) $ is given in \cite[Theorem 3.4]{zouzias2013rando}. \end{remark} Next, we prove the exponential convergence of the expected norm of the residual for {\it full column rank or rank-deficient consistent} linear systems. \begin{theorem}\label{main4} Let $\mbf x^k$ denote the $k$th iterate of {\rm DSBGS} applied to the consistent linear system (full column rank or rank deficient) $\bf Ax=b$ with arbitrary $\mbf x^0\in\mbbr^n$. If $t=n$ and $0<\alpha<2\sigma_r^2(\mbf A)/{(\beta\|\mbf A\|_\rmf^2)}$, then $$\mbbe[\|\mbf A\mbf x^k-\mbf b\|_2^2]\leq\l(1+{\beta}\alpha^2-\frac{2\alpha\sigma_r^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)^k\|\mbf A\mbf x^0-\mbf b\|_2^2.$$ If $t<n$ and $0<\alpha<2\sigma_r^2(\mbf A)/(t\rho{\beta})$, then $$\mbbe[\|\mbf A\mbf x^k-\mbf b\|_2^2]\leq\l(1-\frac{2\alpha\sigma_r^2(\mbf A)-t\rho{\beta}\alpha^2}{\|\mbf A\|_\rmf^2}\r)^k\|\mbf A\mbf x^0-\mbf b\|_2^2,$$ where $$\rho=\max_{1\leq j\leq t}\sigma_1^2(\mbf A_{:,\mcalj_j}).$$ \end{theorem} \proof Note that \beqas \|{\bf Ax}^k-\mbf b\|_2^2&=&\l\|{\bf Ax}^{k-1}-\alpha\l(\frac{\mbf A\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)(\mbf A\mbf x^{k-1}-\mbf b)-\mbf b\r\|_2^2\\ &=& \|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)\\ && + \alpha^2({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf I_{:,\mcali}\mbf A_{\mcali,\mcalj}(\mbf I_{:,\mcalj})^\rmt\mbf A^\rmt\mbf A \mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^4}\r)({\bf Ax}^{k-1}-\mbf b). \eeqas If $t=n$, then it follows from $(\mbf I_{:,\mcalj})^\rmt\mbf A^\rmt\mbf A \mbf I_{:,\mcalj}=\|\mbf A_{:,\mcalj}\|_\rmf^2$ (since $\mbf A \mbf I_{:,\mcalj}=\mbf A_{:,\mcalj}$ is a column vector) that \beqas \|{\bf Ax}^k-\mbf b\|_2^2 &=& \|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)\\ && + \alpha^2({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\|\mbf A_{:,\mcalj}\|_\rmf^2\mbf I_{:,\mcali}\mbf A_{\mcali,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^4}\r)({\bf Ax}^{k-1}-\mbf b)\\ &\leq & \|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)\\ && + \alpha^2({\bf Ax}^{k-1}-\mbf b)^\rmt\l({\frac{\|\mbf A_{\mcali,\mcalj}\|_2^2}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}}\frac{\|\mbf A_{:,\mcalj}\|_\rmf^2\mbf I_{:,\mcali}(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b) \quad ({\rm by\ Lemma\ \ref{lemma2}})\\ &\leq & {\|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)}\\ && {+ \beta\alpha^2({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\|\mbf A_{:,\mcalj}\|_\rmf^2\mbf I_{:,\mcali}(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)} \eeqas Taking expectation gives \beqas\mbbe[\|{\bf Ax}^k-\mbf b\|_2^2\ |\mbf x^{k-1}] &\leq&(1+{\beta}\alpha^2)\|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf A^\rmt}{\|\mbf A \|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)\\ &\leq& \l(1+{\beta}\alpha^2-\frac{2\alpha\sigma_r^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)\|\mbf A\mbf x^{k-1}-\mbf b\|_2^2.\eeqas The last inequality follows from $\mbf A\mbf x^{k-1}-\mbf b\in\ran(\mbf A)$ and Lemma \ref{lemma3}. Taking expectation again gives \beqas \mbbe[\|{\bf Ax}^k-\mbf b\|_2^2]&=&\mbbe[ \mbbe[\|{\bf Ax}^k-\mbf b\|_2^2\ |\mbf x^{k-1}]]\\ &\leq& \l(1+{\beta}\alpha^2-\frac{2\alpha\sigma_r^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)\mbbe[\|\mbf A\mbf x^{k-1}-\mbf b\|_2^2]\\ &\leq& \l(1+{\beta}\alpha^2-\frac{2\alpha\sigma_r^2(\mbf A)}{\|\mbf A\|_\rmf^2}\r)^k\|\mbf A\mbf x^0-\mbf b\|_2^2.\eeqas If $t<n$, then it follows from $(\mbf I_{:,\mcalj})^\rmt\mbf A^\rmt\mbf A \mbf I_{:,\mcalj}=\mbf A_{:,\mcalj}^\rmt\mbf A_{:,\mcalj}\preceq\rho\mbf I$ (since $\rho=\max_{1\leq j\leq t}\sigma_1^2(\mbf A_{:,\mcalj_j})$) that \beqas \|{\bf Ax}^k-\mbf b\|_2^2 &\leq& \|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)\\ && + \alpha^2({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\rho\mbf I_{:,\mcali}\mbf A_{\mcali,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^4}\r)({\bf Ax}^{k-1}-\mbf b)\\ &\leq& \|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)\\ && + \alpha^2({\bf Ax}^{k-1}-\mbf b)^\rmt\l({\frac{\|\mbf A_{\mcali,\mcalj}\|_2^2}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}}\frac{\rho\mbf I_{:,\mcali}(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b) \quad ({\rm by\ Lemma\ \ref{lemma2}})\\ &\leq& \|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf I_{:,\mcalj}(\mbf A_{\mcali,\mcalj})^\rmt(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)\\ && + \alpha^2({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\rho{\beta}\mbf I_{:,\mcali}(\mbf I_{:,\mcali})^\rmt}{\|\mbf A_{\mcali,\mcalj}\|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b). \eeqas Taking expectation gives \beqas\mbbe[\|{\bf Ax}^k-\mbf b\|_2^2\ |\mbf x^{k-1}] &\leq&\l(1+\frac{t\rho{\beta}\alpha^2}{\|\mbf A\|_\rmf^2}\r)\|{\bf Ax}^{k-1}-\mbf b\|_2^2-2\alpha({\bf Ax}^{k-1}-\mbf b)^\rmt\l(\frac{\mbf A\mbf A^\rmt}{\|\mbf A \|_\rmf^2}\r)({\bf Ax}^{k-1}-\mbf b)\\&\leq& \l(1-\frac{2\alpha\sigma_r^2(\mbf A)-t\rho{\beta}\alpha^2}{\|\mbf A\|_\rmf^2}\r)\|\mbf A\mbf x^{k-1}-\mbf b\|_2^2.\eeqas The last inequality follows from $\mbf A\mbf x^{k-1}-\mbf b\in\ran(\mbf A)$ and Lemma \ref{lemma3}. Taking expectation again gives \beqas \mbbe[\|{\bf Ax}^k-\mbf b\|_2^2]&=&\mbbe[ \mbbe[\|{\bf Ax}^k-\mbf b\|_2^2\ |\mbf x^{k-1}]]\\ &\leq& \l(1-\frac{2\alpha\sigma_r^2(\mbf A)-t\rho{\beta}\alpha^2}{\|\mbf A\|_\rmf^2}\r)\mbbe[\|\mbf A\mbf x^{k-1}-\mbf b\|_2^2]\\ &\leq& \l(1-\frac{2\alpha\sigma_r^2(\mbf A)-t\rho{\beta}\alpha^2}{\|\mbf A\|_\rmf^2}\r)^k\|\mbf A\mbf x^0-\mbf b\|_2^2.\qquad \qed\eeqas \begin{remark} For the special case $s=1$, $t=n$, $\alpha=\sigma_r^2(\mbf A)/\|\mbf A\|_\rmf^2$ (i.e., the randomized Gauss--Seidel algorithm{, we have $\beta=1$ in this case}) and the special case $s=m$, $t=n$, $\alpha=\sigma_r^2(\mbf A)/\|\mbf A\|_\rmf^2$ (i.e., the doubly stochastic Gauss--Seidel algorithm{, we have $\beta=1$ in this case}), the results of Theorem \ref{main4} are given in \cite[Theorem 3.2]{leventhal2010rando} and \cite[Theorem 2]{razaviyayn2019linea}, respectively. \end{remark} \begin{remark} Let $\mbf x_\star$ be any solution of $\bf A^\rmt A x=A^\rmt b$. We have $(\mbf A_{:,\mcalj})^\rmt\mbf b=(\mbf A_{:,\mcalj})^\rmt\mbf A\mbf x_\star$. Then for $s=1$ and $t\in[n]$, by the same approach as used in the proof of Theorem \ref{main4}, we can prove that the convergence bounds (replacing $\mbf b$ by $\mbf A\mbf x_\star$) in Theorem \ref{main4} still hold for {\it inconsistent} linear systems. The result for the special case $s=1$ and $t=n$ (i.e., the randomized Gauss--Seidel algorithm) was already given in the literature, for example, \cite[Theorem 3.2]{leventhal2010rando}, \cite[Lemma 4.2]{ma2015conve} and \cite[Theorem 3]{du2019tight}. \end{remark} \section{Numerical results} In this section, we compare the performance of the doubly stochastic block Gauss--Seidel (DSBGS) algorithm proposed in this paper against the randomized Kaczmarz (RK) algorithm for solving consistent linear systems. We only run on small or medium-scale problems. The purpose is to demonstrate that even in these simple examples, DSBGS offers significant advantages over RK. All experiments are performed using MATLAB (R2019a) on a laptop with 2.7-GHz Intel Core i7 processor, 16 GB memory, and macOS Sierra (version 10.12.6). We use DSBGS($\alpha,\ell,\tau$) to denote the doubly stochastic block Gauss--Seidel algorithm employing the step size $\alpha$, the row partition $\{\mcali_i\}_{i=1}^s$ with $s=\lceil\frac{m}{\ell}\rceil$: \beqas\mcali_i&=&\{(i-1)\ell+1,(i-1)\ell+2,\ldots,i\ell\},\quad i=1,2,\ldots,s-1,\\ \mcali_s&=&\{(s-1)\ell+1,(s-1)\ell+2,\ldots,m\},\eeqas and the column partition $\{\mcalj_j\}_{j=1}^t$ with $t=\lceil\frac{n}{\tau}\rceil$: \beqas\mcalj_j&=&\{(j-1)\tau+1,(j-1)\tau+2,\ldots,j\tau\},\quad j=1,2,\ldots,t-1,\\ \mcalj_t&=&\{(t-1)\tau+1,(t-1)\tau+2,\ldots,n\}.\eeqas The randomized Kaczmarz algorithm with step size $\alpha=1$ is the special case DSBGS($1,1,n$). To construct a consistent linear system, for a given coefficient matrix $\mbf A$, we set $\bf b = Ax$ where $\mbf x$ is a vector with entries generated from a standard normal distribution. All algorithms are started from the initial guess $\mbf x^0 = \mbf 0$, terminated if $\|\mbf x^k-\mbf A^\dag\mbf b\|_2\leq10^{-5}$. We report the average number of iterations (denoted as ITER) of RK and DSBGS. We also report the average computing time in seconds (denoted as CPU) and the speed-up of DSBGS against RK, which is defined as $$\mbox{speed-up}=\frac{\mbox{CPU of RK}}{\mbox{CPU of DSBGS}}.$$ In each experiment, ITER and CPU are averaged over 20 trials. \subsection{Synthetic data} Two types of coefficient matrices are generated as follows. \bit \item[$\bullet$] Type I: For given $m$, $n$, $r = rank(\mbf A)$, and $\kappa>1$, we construct a matrix $\mbf A$ by $$\bf A = UDV^\rmt,$$ where $\mbf U\in\mbbr^{m\times r} $ and $\mbf V\in \mbbr^{n\times r}$. Entries of $\mbf U$ and $\mbf V$ are generated from a standard normal distribution, and then, columns are orthonormalized: $${\tt [U,\sim] = qr(randn(m,r),0);\qquad [V,\sim] = qr(randn(n,r),0);}$$ The matrix $\mbf D$ is an $r\times r$ diagonal matrix whose diagonal entries are uniformly distributed numbers in $(1,\kappa)$: $${\tt D = diag(1+(\kappa-1).*rand(r,1));}$$ So the condition number of $\mbf A$, which is defined as $\sigma_1(\mbf A)/\sigma_r(\mbf A)$, is upper bounded by $\kappa$. \item[$\bullet$] Type II: For given $m$, $n$, entries of $\mbf A$ are generated from a standard normal distribution: $$\tt A=randn(m,n);$$ So $\mbf A$ is a full (column or row) rank matrix with probability one. \eit In Figures \ref{fig1} and \ref{fig2} we plot the error $\|\mbf x^k-\mbf A^\dag\mbf b\|_2$ of DSBGS with different step size $\alpha$ and different block size for full column rank consistent linear systems. From these figures, we observe that appropriate step size and block size improve the convergence remarkably. In Tables \ref{tab1} and \ref{tab2}, we report the average numbers of iterations and the average computing times for RK and DSBGS. From these results we see DSBGS vastly outperforms RK in terms of computing times with significant speed-ups for general (overdetermined or underdetermined, full column rank or rank deficient) consistent linear systems. It should be noted that the step size $\alpha\in(0,2/{(t\beta)})$ given in Theorem \ref{main3} is a sufficient condition for DSBGS's convergence. Numerical experiments show that DSBGS with some large step size (for example $\alpha=15$ in the experiment for Figure 2) converges much faster than with step size $\alpha\in(0,2/{(t\beta)})$. \begin{figure} \centerline{\epsfig{figure=fig1.pdf,height=3in}} \caption{The average (20 trials of each case) convergence history of DSBGS with different step size ($\alpha=5,10,15,17$) and fixed block size ($\ell=50,\tau=50$) for a full column rank consistent linear system with random coefficient matrix $\mbf A$ of Type II ({\tt A=randn(500,250)}).} \label{fig1} \end{figure} \begin{figure} \centerline{\epsfig{figure=fig2.pdf,height=3in}} \caption{The average (20 trials of each case) convergence history of DSBGS with different step size and different block size for a full column rank consistent linear system with random coefficient matrix $\mbf A$ of Type II ({\tt A=randn(500,250)}).} \label{fig2} \end{figure} \begin{table} \caption{The average (20 trials of each experiment) {\rm ITER} and {\rm CPU} of {\rm RK} and {\rm DSBGS}($\alpha,\ell,\tau$) for consistent linear systems with random coefficient matrices $\mbf A$ of Type I: ${\bf A=UDV^\rmt}$.} \label{tab1} \begin{center} \begin{tabular}{ccr|rr|rrr|c} \toprule $m\times n$ & rank & $\kappa$ & \multicolumn{2}{|c|}{RK: ITER, CPU} & \multicolumn{3}{|c|}{DSBGS: ITER, CPU, $(\alpha,\ell,\tau)$} & speed-up\\ \noalign{\smallskip} \hline \noalign{\smallskip} $125\times 250$ & 100 & 2 & 3162.55 & 0.0805 & 628.85 & 0.0258 & $(5,5,n)$ & 3.12 \\ $125\times 250$ & 100 & 10 & 26413.55 & 0.6813 & 5255.70 & 0.1751 & $(5,5,n)$ & 3.89 \\ $125\times 250$ & 100 & 20 & 158506.00 & 3.8662 &31751.35 & 1.0482 & $(5,5,n)$ & 3.69 \\ $250\times 500$ & 200 & 2 & 6791.10 & 0.2270 & 638.40 & 0.0482 & $(10,10,n)$ & 4.71 \\ $250\times 500$ & 200 & 10 & 69568.35 & 2.3030 & 6912.85 & 0.5141 & $(10,10,n)$ & 4.48 \\ $250\times 500$ & 200 & 20 & 252768.05 & 8.3663 & 25328.15 & 1.8639 & $(10,10,n)$ & 4.49 \\ $250\times 125$ & 125 & 2 & 4215.20 & 0.1138 & 975.25 & 0.0291 & $(5,25,25)$ & 3.90 \\ $250\times 125$ & 125 & 10 & 38675.35 & 1.0213 & 8241.10 & 0.2396 & $(5,25,25)$ & 4.26 \\ $250\times 125$ & 125 & 20 & 101769.10 & 2.6832 & 23687.75 & 0.6682 & $(5,25,25)$ & 4.02 \\ $500\times 250$ & 250 & 2 & 8637.00 & 0.2758 & 993.80 & 0.0602 & $(10,50,50)$ & 4.58 \\ $500\times 250$ & 250 & 10 & 85328.80 & 2.7336 & 9732.80 & 0.5871 & $(10,50,50)$ & 4.66 \\ $500\times 250$ & 250 & 20 & 448211.30 & 14.1407 & 45301.45 & 2.7288 & $(10,50,50)$ & 5.18 \\ \bottomrule \end{tabular} \end{center} \end{table} \begin{table} \caption{The average (20 trials of each experiment) {\rm ITER} and {\rm CPU} of {\rm RK} and {\rm DSBGS}($\alpha,\ell,\tau$) for consistent linear systems with random coefficient matrices $\mbf A$ of Type II: {\tt A=randn(m,n)}. Here $\kappa(\mbf A)=\sigma_1(\mbf A)/\sigma_r(\mbf A)$.} \label{tab2} \begin{center} \begin{tabular}{rc|rr|rrr|c} \toprule $m\times n$ & $\kappa(\mbf A)$ & \multicolumn{2}{|c|}{RK: ITER, CPU} & \multicolumn{3}{|c|}{DSBGS: ITER, CPU, $(\alpha,\ell,\tau)$} & speed-up \\ \noalign{\smallskip} \hline \noalign{\smallskip} $125\times 250$ & 5.66 & 16118.20 & 0.4160 & 3210.75 & 0.1063 & $(5,5,n)$ & 3.91 \\ $125\times 500$ & 2.87 & 5876.40 & 0.1719 & 1134.60 & 0.0756 & $(5,5,n)$ & 2.28 \\ $125\times 1000$ & 2.09 & 3867.00 & 0.1571 & 727.65 & 0.0691 & $(5,5,n)$ & 2.27 \\ $250\times 500$ & 5.56 & 30557.05 & 1.0139 & 3091.35 & 0.2273 & $(10,10,n)$ & 4.46 \\ $250\times 1000$ & 2.96 & 12015.30 & 0.5085 & 1174.55 & 0.1310 & $(10,10,n)$ & 3.88 \\ $250\times 2000$ & 2.06 & 7927.15 & 0.4963 & 751.75 & 0.1959 & $(10,10,n)$ & 2.53 \\ $500\times 750$ & 9.66 & 173700.40 & 7.2715 & 17381.35 & 1.7435 & $(10,10,n)$ & 4.17 \\ $500\times 1500$ & 3.66 & 33019.55 & 2.1238 & 3325.55 & 0.5445 & $(10,10,n)$ & 3.90 \\ $500\times 3000$ & 2.34 & 18520.60 & 2.4642 & 1813.70 & 0.7393 & $(10,10,n)$ & 3.33 \\ $250\times 125$ & 5.33 & 13981.95 & 0.3709 & 3053.15 & 0.0864 & $(5,25,25)$ & 4.29 \\ $500\times 125$ & 2.96 & 6095.05 & 0.1811 & 1338.45 & 0.0393 & $(5,25,25)$ & 4.60 \\ $1000\times 125$ & 2.07 & 4112.30 & 0.1488 & 1003.05 & 0.0344 & $(5,25,25)$ & 4.32 \\ $500\times 250$ & 5.77 & 32563.80 & 1.0330 & 6958.60 & 0.4097 & $(5,50,25)$ & 2.52 \\ $1000\times 250$ & 2.87 & 11965.45 & 0.4537 & 2724.85 & 0.1763 & $(5,50,25)$ & 2.57 \\ $2000\times 250$ & 2.12 & 8489.30 & 0.4305 & 2086.25 & 0.1430 & $(5,50,25)$ & 3.01 \\ $750\times 500$ & 9.50 & 148619.65 & 5.9663 & 32345.90 & 2.8518 & $(5,50,50)$ & 2.09 \\ $1500\times 500$ & 3.68 & 33655.05 & 1.6979 & 7212.25 & 0.6977 & $(5,50,50)$ & 2.43 \\ $3000\times 500$ & 2.36 & 18990.70 & 1.3557 & 4378.90 & 0.5279 & $(5,50,50)$ & 2.57 \\ \bottomrule \end{tabular} \end{center} \end{table} \subsection{Real-world data} We test RK and DSBGS on eight real-world problems from the University of Florida sparse matrix collection \cite{davis2011unive}: {\tt abtaha1}, {\tt WorldCities}, {\tt cari}, {\tt df2177}, {\tt flower\_5\_1}, {\tt football}, {\tt relat6}, {\tt Sandi\_authors}. The first two matrices are of full column rank and the last six matrices are rank-deficient. In Table \ref{tab3}, we report the average numbers of iterations and the average computing times of RK and DSBGS. We observe that DSBGS based on good choices of step size and block size significantly outperforms RK. Moreover, good step size and block size are problem dependent. \begin{table} \caption{The average (20 trials of each experiment) ITER and {\rm CPU} of {\rm RK} and {\rm DSBGS}($\alpha,\ell,\tau$) for consistent linear systems with coefficient matrices from the University of Florida sparse matrix collection. Here $\kappa(\mbf A)=\sigma_1(\mbf A)/\sigma_r(\mbf A)$. The first two matrices are of full column rank and the last six matrices are rank deficient.} \label{tab3} \begin{center} {\footnotesize\begin{tabular}{lrr|rr|rrr|c} \toprule Matrix & $m \times n$ & $\kappa(\mbf A)$ & \multicolumn{2}{|c|}{RK: ITER, CPU} & \multicolumn{3}{|c|}{DSBGS: ITER, CPU, $(\alpha,\ell,\tau)$} & speed-up \\ \noalign{\smallskip} \hline \noalign{\smallskip} {\tt abtaha1} & $14596\times209$ & 12.23 & 1.83e05 & 42.2491 & 3.91e04 & 2.8988 & $(5,10,n)$ & 14.57 \\ {\tt WorldCities} & $315\times 100$ & 66.00& 7.38e04 & 1.9794 & 3.09e04 & 0.8624 & $(2.5,10,n)$ & 2.30 \\ {\tt cari} & $400\times1200$ & 3.13 & 9.79e03 & 0.5035 & 2.76e03 & 0.3186 & $(2.5,5,n)$ & 1.58 \\ {\tt df2177} & $630\times10358$ & 2.01 & 1.63e04 & 8.5768 & 2.66e03 & 5.0829 & $(5,10,n)$ & 1.69 \\ {\tt flower\_5\_1} & $211\times201 $ & 13.70 & 9.55e04 & 2.6053 & 3.83e04 & 1.2001 & $(2.5,5,n)$ & 2.17 \\ {\tt football} & $35\times35 $ & 166.47 & 7.88e05 & 15.1897 & 3.94e05 & 8.4485 & $(2,4,n)$ & 1.80 \\ {\tt relat6} & $2340\times 157$ & 7.74 & 2.66e04 & 1.4331 & 1.03e04 & 0.3969 & $(2.5,10,n)$ & 3.61 \\ {\tt Sandi\_authors} & $86\times86 $ & 189.58 & 2.16e06 & 45.6432 & 8.66e05 & 21.5159 & $(2.5,5,n)$ & 2.12 \\ \bottomrule \end{tabular}} \end{center} \end{table} \section{Concluding remarks} We have proposed a doubly stochastic block Gauss--Seidel algorithm for solving linear systems and prove its convergence theory. The randomized Kaczmarz algorithm, the randomized Gauss--Seidel algorithm, and the doubly stochastic Gauss--Seidel algorithm are special cases of the doubly stochastic block Gauss--Seidel algorithm. Numerical experiments show that appropriate step size and block size significantly improve the performance. Finding appropriate variable step size, proposing more effective sampling strategies for submatrices, and designing other block variants via the ideas in \cite{necoara2019faste} should be valuable topics in the future study. \section*{Acknowledgments} The research of the first author was supported by the National Natural Science Foundation of China (No.11771364) and the Fundamental Research Funds for the Central Universities (No.20720180008).

https://arxiv.org/abs/1711.00819

The Steklov Problem on Rectangles and Cuboids

This paper is a brief account of the Steklov eigenvalue problem on a 2-dimensional rectangular domain, and then on a 3-dimensional rectangular box. It is divided into four sections. Section 1 relies heavily on real analytic methods to show the existence of an eigenfunction class which always produces the first non-trivial Steklov eigenvalue on a rectangle. Section 2 lists all possible Steklov eigenfunctions on a cuboid. The very brief section 3 gives the 3-dimensional analogue of the analytic results in Section 1. The analogue is given as conjecture, but is expected to derivate from standard (albeit tedious) real analysis methods, should one wish to expound on these calculations. Section 4 deals with the special cases of the square and the cube.

\section{Analysis of the 1st Steklov eigenspace on the rectangle} Consider the rectangle defined on $[-1,1]\times[-a,a]$, where $a \leq 1$. One can classify the Steklov eigenfunctions on this domain into 4 parity classes (Auchmuty $\&$ Cho, 2014). Generally, each parity class has two eigenfunctions, but for the case of the square $(a=1)$, there is an additional eigenfunction $s(x,y)=xy$ in Class II, whose eigenvalue is simple and equal to $1$. \vspace{1mm} \\ One can consider the smallest nontrivial eigenvalue $\sigma_1$, and ask: \textit{Among all rectangular domains with constant perimeter, which domains maximize the first Steklov eigenvalue, $\sigma_1$?} \vspace{1mm} \\ \noindent One may normalize the eigenvalue $\sigma_1$ by multiplying with the perimeter of the domain, $L=4(1+a)$. The product, $\sigma_1L$, is an invariant with respect to a scaling of the domain. \vspace{1mm} \\ \noindent That is, one can define an equivalence relation between similar rectangles, with rectangles in the same equivalence class having the same first Steklov invariant. We identify any rectangle with a rectangle defined $[-1,1] \times [-a,a]$, where $a \leq 1$, so that we may restrict our analysis to such rectangles. Since similar rectangles have the same Steklov invariants, the question then becomes: \textit{Among all values of $a \in (0,1]$, which maximizes $4\sigma_1(1+a)$?} \vspace{1mm} \\ \noindent We show that $a=1$ does. \\ As mentioned above, we consider the eigenfunctions by parity class. In the table below, we only need to consider positive $\nu$. \FloatBarrier \begin{table}[!htbp] \centering \begin{tabular}{l|c|c|r} Class & Eigenfunction & Determining Equation & Eigenvalue $\sigma$\\\hline\hline I(i) & cosh$(\nu x)$cos$(\nu y)$, & tan$(\nu a)+$tanh$(\nu)=0$ & $\nu$tanh$(\nu)$ \\ I(ii) & cos$(\nu x)$cosh$(\nu y)$ & tan$(\nu)+$tanh$(\nu a)=0$ & $\nu$tanh$(\nu a)$ \\\hline II(i) & sinh$(\nu x)$sin$(\nu y)$, & tan$(\nu a)=$tanh$(\nu)$ & $\nu$coth$(\nu)$ \\ II(ii) & sin$(\nu x)$sinh$(\nu y)$ & tan$(\nu)=$tanh$(\nu a)$ & $\nu$coth$(\nu a)$ \\\hline III(i) & cosh$(\nu x)$sin$(\nu y)$, & tan$(\nu a)=$coth$(\nu)$ & $\nu$tanh$(\nu)$ \\ III(ii) & cos$(\nu x)$sinh$(\nu y)$ & tan$(\nu)+$coth$(\nu a)=0$ & $\nu$coth$(\nu a)$ \\\hline IV(i) & sinh$(\nu x)$cos$(\nu y)$, & tan$(\nu a)+$coth$(\nu)=0$ & $\nu$coth$(\nu)$ \\ IV(ii) & sin$(\nu x)$cosh$(\nu y)$ & tan$(\nu)=$coth$(\nu a)$ & $\nu$tanh$(\nu a)$ \end{tabular} \caption{\label{eightclasses}The eigenfunctions fall into eight classes.} \end{table} \FloatBarrier \vspace{1mm} \noindent (We have omitted $s(x,y)=xy$, which only arises when $a=1$.) \vspace{2mm} \noindent Then, for instance, the 1st Steklov eigenfunction from class IV(ii) has value $\nu$ given by the 1st positive solution of tan$\nu = $coth$(\nu a)$. Of course, for given $a$, minimizing $\nu$ is equivalent to minimizing $\nu$tanh$(\nu a)$. In fact, we have the following simple lemma. \begin{lemma}\label{increasing} For given $a \in (0,1]$, the maps $\nu \mapsto \nu$tanh$(\nu a)$ and $\nu \mapsto \nu$coth$(\nu a)$ are increasing on $\nu \in (0,\infty)$ (Auchmuty $\&$ Cho, 2014). \end{lemma} \vspace{1mm} \noindent We claim that the 1st nontrivial Steklov eigenvalue for a rectangle always comes from class IV(ii). \vspace{1mm} \noindent By considering intersections of graphs (see Figure 1 of Girouard $\&$ Polterovich, 2014), it is trivial that III(i) is a better minimizer of $\nu$ -- and therefore $\sigma$ -- than I(i). Likewise IV(ii) is a better minimizer of $\nu$ than I(ii), II(ii), and III(ii). Since tanh$(\nu a) <$ coth$(\nu a)$, we have that IV(ii) gives the smallest value of $\sigma$ compared with I(ii), II(ii), and III(ii). \vspace{1mm} \noindent It remains to compare II(i), III(i), IV(ii). \begin{claim} Fix $a \in (0,1]$. For a rectangle on $[-1,1] \times [-a,a]$, let the candidate for the 1st Steklov eigenvalue from class II(i) be $\sigma_1^{II(i)}$, and that from class III(i) be $\sigma_1^{III(i)}$. Then $\sigma_1^{III(i)} < \sigma_1^{II(i)}$. \end{claim} \begin{proof} We want to show that III(i) is a better minimizer of $\sigma_1$ than II(i). For II(i), an eigenvalue $\sigma$ is given by an intersection of the graphs $\sigma = \nu$coth$(\nu)$ and $\sigma = \nu$cot$(\nu a)$. For III(i), an eigenvalue $\sigma$ is given by an intersection of the graphs $\sigma = \nu$tanh$(\nu)$ and $\sigma = \nu$cot$(\nu a)$. \vspace{1mm} \noindent Naturally, in either case, the candidate for $\sigma_1$ is the intersection which has the lowest height. \vspace{1mm} \noindent We show that the candidate for $\sigma_1$ is smaller in the case of III(i). \FloatBarrier \begin{figure}[!htbp] \centering \includegraphics[width=1\textwidth]{IIIi.JPG} \caption{\label{fig:IIIi}The graphs for $a = 0.5$.} \end{figure} \FloatBarrier \vspace{1mm} \noindent In each case, the lowest intersection must also be the intersection nearest to to the $\sigma$-axis, by Lemma \ref{increasing}. Hence we need only consider the graphs on $0<\nu< \nu_0$ where $\nu_0=\frac{\pi}{2a}$. Figure \ref{fig:IIIi} compares the lowest intersections for case III(i) and II(i). By the intermediate value theorem, there exists a candidate for $\sigma_1$ in $0<\nu< \nu_0$ for both cases III(i) and II(i). By the strict monotonicity of the graphs on this interval, these candidates are unique. Finally, the candidate from class III(i) has a lower height than the candidate from class II(i), because the graph of $\sigma = \nu$cot$(\nu a)$ is strictly decreasing on $0<\nu< \nu_0$. Otherwise we would derive a contradiction by the mean value theorem. \end{proof} \begin{claim}\label{iffsquare} Fix $a \in (0,1]$. For a rectangle on $[-1,1] \times [-a,a]$, let the candidate for the 1st Steklov eigenvalue from class III(i) be $\sigma_1^{III(i)}$, and that from class IV(ii) be $\sigma_1^{IV(ii)}$. Then $\sigma_1^{IV(ii)} \leq \sigma_1^{III(i)}$, with equality iff $a = 1$. \end{claim} \begin{proof} Similar to before, we want to show that IV(ii) is a better minimizer of $\sigma_1$ than III(i). For III(i), an eigenvalue $\sigma$ is given by an intersection of the graphs $\sigma = \nu$tanh$(\nu)$ and $\sigma =\nu$cot$(\nu a)$. For IV(ii), an eigenvalue $\sigma$ is given by an intersection of the graphs $\sigma =\nu$coth$(\nu a)$ and $\sigma =\nu$cot$(\nu)$. \vspace{1mm} \noindent We show that the candidate for $\sigma_1$ is smaller in the case of IV(ii). \FloatBarrier \begin{figure}[!htbp] \centering \includegraphics[width=1\textwidth]{III_i_.JPG} \caption{\label{fig:III(i)}The candidate for $\sigma_1$ of Class III(i), in the case $a = 0.5$.} \end{figure} \FloatBarrier \FloatBarrier \begin{figure}[!htbp] \centering \includegraphics[width=1\textwidth]{IV_ii_.JPG} \caption{\label{fig:IV(ii)}The candidate for $\sigma_1$ of Class IV(ii), in the case $a = 0.5$.} \end{figure} \FloatBarrier \vspace{1mm} \noindent In each case, the lowest intersection must also be the intersection nearest to to the $\sigma$-axis, by $(*)$. Hence we need only consider the graphs on $0<\nu< \nu_0$ where $\nu_0=\frac{\pi}{2a}$. Figures \ref{fig:III(i)} and \ref{fig:IV(ii)} give the lowest intersection for class III(i) and class IV(ii) respectively. These intersections can be shown to exist in $0<\nu< \nu_0$ by the intermediate value theorem. (A sample calculation for this is done explicitly in the proof of Claim \ref{diff}.) Since it is clear that both graphs in Figure \ref{fig:IV(ii)} are lower than the respective graphs in Figure \ref{fig:III(i)}, we can conclude that $\sigma_1^{IV(ii)} < \sigma_1^{III(i)}$. \end{proof} \vspace{1mm} \noindent We have just shown that the 1st Steklov invariant on a given rectangle is \textit{always} given by eigenfunction IV(ii)! Moreover, for $a \neq 1$, this is the only eigenfunction in the 1st eigenspace, because $\sigma_1^{IV(ii)}$ is strictly smaller than any other eigenvalue. Let us state this as a theorem. \begin{theorem} On any rectangular domain that is not a square, the 1st Steklov eigenspace has basis $\{$sin$(\nu x)$cosh$(\nu y)$$\}$. \end{theorem} \vspace{3mm} \noindent Below is a graph of this invariant against the length $a \in (0,1]$. \FloatBarrier \begin{figure}[!htbp] \centering \includegraphics[width=1\textwidth]{SmallestSteklovInvariantOnRectangle.JPG} \caption{\label{fig:smallest}The 1st Steklov invariant for a rectangle on $[-1,1]\times[-a,a]$.} \end{figure} \FloatBarrier \vspace{3mm} \noindent For completeness, we also plot the invariant given by each of the 8 eigenfunctions against $a$ in Figure \ref{fig:all}. \FloatBarrier \begin{figure}[!htbp] \centering \includegraphics[width=1\textwidth]{SmallestSteklovInvariantForEachOf8EigenfunctionsOnRectangle.JPG} \caption{\label{fig:all}The 1st Steklov invariant for each of 8 determining equations.} \end{figure} \FloatBarrier \vspace{1mm} \noindent The lowest graph in Figure \ref{fig:all} is, of course, the same graph as in Figure \ref{fig:smallest}. Figure \ref{fig:smallest} suggests that $a = 1$ maximizes the 1st Steklov invariant. In fact, we now show that the graph is differentiable, increasing with $a$, and tends to the origin (in the first quadrant). \begin{claim}\label{diff} The graph in Figure \ref{fig:smallest} is differentiable on $a \in (0,1)$. \end{claim} \begin{proof} Fix $a \in (0,1)$. Define $k_a:(0,\frac{\pi}{2}) \to \mathbb{R}$, $k_a(\nu) = \nu$cot$(\nu)-\nu$tanh$(\nu a)$. Since $\lim_{\nu \downarrow 0} k_a(\nu)=1>0$ and $\lim_{\nu \uparrow \frac{\pi}{2}} k_a(\nu)=-\frac{\pi}{2}\text{tanh}(\frac{\pi}{2} a) <0$, the intermediate value theorem guarantees some root $\nu_1^{IV(ii)} \in (0,\frac{\pi}{2})$. That is, cot$(\nu_1^{IV(ii)})-$tanh$(\nu_1^{IV(ii)} a)=0$. In fact, $\nu_1^{IV(ii)}$ is unique since $k_a$ is strictly decreasing. (Clearly $\sigma_1^{IV(ii)}=\nu_1^{IV(ii)}$tanh$(\nu_1^{IV(ii)}a)$.) The above allows me to define $h:\mathbb{R}\times (0,\frac{\pi}{2}) \to \mathbb{R}$, $h(a,\nu)=k_a(\nu)$. Next, we show that $\nu$ is a differentiable function of $a \in (0,1)$ by applying the implicit function theorem to $h$. At each root $(a,\nu_1^{IV(ii)})$ of $h$, we check that $h_{\nu} \neq 0$: \begin{align*} h_{\nu}( \ (a,\nu_1^{IV(ii)}) \ ) &= -\text{tanh}(\nu_1^{IV(ii)} a) - \nu_1^{IV(ii)} a \text{sech}^2(\nu_1^{IV(ii)} a) + \text{cot}(\nu_1^{IV(ii)}) - \nu_1^{IV(ii)} \text{cosec}^2(\nu_1^{IV(ii)}) \\ &= - \nu_1^{IV(ii)} a \text{sech}^2(\nu_1^{IV(ii)} a) - \nu_1^{IV(ii)} \text{cosec}^2(\nu_1^{IV(ii)}) \\ &< 0. \end{align*} \vspace{1mm} \noindent Therefore the implicit function theorem ensures that $\nu$ is locally a differentiable function of $a$ for each $a \in (0,1)$, and hence globally differentiable on $(0,1)$. It is also continuous on $[0,1]$. The composition $\sigma = \nu$tanh$(\nu a)$ is therefore a differentiable function of $a$. \end{proof} \begin{claim} The graph in Figure \ref{fig:smallest} is strictly increasing on $(0,1] \ni a$. Moreover, $\sigma \downarrow 0$ as $a \downarrow 0$. \end{claim} \begin{proof} Consider again Figure \ref{fig:IV(ii)}, which displays (for the special case $a=0.5$) the intersection of $\sigma=\nu $tanh$(\nu a)$ with $\sigma=\nu $cot$(\nu)$. Clearly as $a \downarrow 0$, the intersection height decreases (continuously!) to $0$. Moreover, for $a_1 < a_2$, we have $\sigma_1 < \sigma_2$. \end{proof} \vspace{1mm} \noindent (Alternatively, by examining the \textit{Rayleigh quotient}, one can see easily that $\sigma \downarrow 0$ as $a \downarrow 0$.) \section{Finding Steklov eigenfunctions on a cuboid} Now we move to the 3D case. The notation here is such that $(a,b,c)$ give the dimensions of a cube, but similarly to the 2D case, we may let $c=1$ be the longest side without loss of generality. \vspace{1mm} \noindent Let us consider a cuboid on $[-a,a]\times[-b,b]\times[-c,c]$. we use separation of variables to solve Laplace's equation with a Steklov condition on the faces of the cuboid. This treatment will assume the completeness of the set of eigenfunctions obtained by separation of variables. For the 2D case, completeness can be shown using 'sloshing' methods (see Girouard and Polterovich, 2014). \\ By symmetry, we only need to consider eigenfunctions which are odd or even in $x$, $y$, and $z$. This gives 8 parity classes of eigenfunctions. In the tables below, we need only consider $\lambda_i>0$. \\ \\ 1. CLASS 000: Even in $x$, $y$, $z$. \begin{align*} & Eigenfunction & & Eigenvalue \\\Xhline{5\arrayrulewidth} s(x,y,z) & = 1 & \sigma&=0 \\\hline s(x,y,z) & = \text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{cosh}(\lambda_1y)\text{cosh}(\lambda_2z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1b) = \lambda_2 \text{tanh} (\lambda_2c) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{cosh}(\lambda_1x)\text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{cosh}(\lambda_2z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1a) = \lambda_2 \text{tanh} (\lambda_2c) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{cosh}(\lambda_1x)\text{cosh}(\lambda_2y)\text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1a) = \lambda_2 \text{tanh} (\lambda_2b) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{cos}(\lambda_1y)\text{cos}(\lambda_2z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1b) = -\lambda_2 \text{tan} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{cos}(\lambda_1x)\text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{cos}(\lambda_2z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1a) = -\lambda_2 \text{tan} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{cos}(\lambda_1x)\text{cos}(\lambda_2y)\text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1b) = -\lambda_2 \text{tan} (\lambda_2a) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline \end{align*} 2. CLASS 001: Even in $x$, $y$, Odd in $z$. \begin{align*} & Eigenfunction & & Eigenvalue \\\Xhline{5\arrayrulewidth} s(x,y,z) & = \text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{cosh}(\lambda_1y)\text{sinh}(\lambda_2z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1b) = \lambda_2 \text{coth} (\lambda_2c) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{cosh}(\lambda_1x)\text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{sinh}(\lambda_2z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1a) = \lambda_2 \text{coth} (\lambda_2c) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{cosh}(\lambda_1x)\text{cosh}(\lambda_2y)\text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1a) = \lambda_2 \text{tanh} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{cos}(\lambda_1y)\text{sin}(\lambda_2z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1b) =\lambda_2 \text{cot} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{cos}(\lambda_1x)\text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{sin}(\lambda_2z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1a) = \lambda_2 \text{cot} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{cos}(\lambda_1x)\text{cos}(\lambda_2y)\text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1a) = -\lambda_2 \text{tan} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = z \text{cosh} (\lambda x) \text{cos} (\lambda y) & \sigma&= 1/c, \text{coth}(\lambda a) = \lambda c = -\text{cot}(\lambda b) \\\hline s(x,y,z) & = z \text{cos} (\lambda x) \text{cosh} (\lambda y) & \sigma&= 1/c, \text{coth}(\lambda b) = \lambda c = -\text{cot}(\lambda a) \\\hline \end{align*} 3. CLASS 010: Even in $x$, $z$, Odd in $y$. \begin{align*} & Eigenfunction & & Eigenvalue \\\Xhline{5\arrayrulewidth} s(x,y,z) & = \text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{sinh}(\lambda_2y)\text{cosh}(\lambda_1z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1c) = \lambda_2 \text{coth} (\lambda_2b) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{cosh}(\lambda_1x)\text{sinh}(\lambda_2y) \text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1 a) = \lambda_2 \text{coth} (\lambda_2b) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{cosh}(\lambda_1x)\text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}y) \text{cosh}(\lambda_2z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1 a) = \lambda_2 \text{tanh} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{sin}(\lambda_2y)\text{cos}(\lambda_1z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1c) =\lambda_2 \text{cot} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{cos}(\lambda_1x)\text{sin}(\lambda_2y) \text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1 a) = \lambda_2 \text{cot} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{cos}(\lambda_1x) \text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}y) \text{cos}(\lambda_2z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1a) = -\lambda_2 \text{tan} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = y \text{cosh} (\lambda x) \text{cos} (\lambda z) & \sigma&= 1/b, \text{coth}(\lambda a) = \lambda b = -\text{cot}(\lambda c) \\\hline s(x,y,z) & = y \text{cos} (\lambda x) \text{cosh} (\lambda z) & \sigma&= 1/b, \text{coth}(\lambda c) = \lambda b = -\text{cot}(\lambda a) \\\hline \end{align*} 4. CLASS 011: Odd in $y$, $z$, Even in $x$. \begin{align*} & Eigenfunction & & Eigenvalue \\\Xhline{5\arrayrulewidth} s(x,y,z) & = \text{cosh}(\lambda_2x)\text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}y) \text{sinh}(\lambda_1z) & \sigma&= \lambda_1 \text{coth} (\lambda_1c) = \lambda_2 \text{tanh} (\lambda_2a) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{cosh}(\lambda_2x)\text{sinh}(\lambda_1y) \text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{coth} (\lambda_1 b) = \lambda_2 \text{tanh} (\lambda_2 a) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{sinh}(\lambda_1y) \text{sinh}(\lambda_2z) & \sigma&= \lambda_1 \text{coth} (\lambda_1 b) = \lambda_2 \text{coth} (\lambda_2c) = - \sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{cos}(\lambda_2x)\text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}y) \text{sin}(\lambda_1z) & \sigma&= \lambda_1 \text{cot} (\lambda_1c) = - \lambda_2 \text{tan} (\lambda_2 a) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2} b) \\\hline s(x,y,z) & = \text{cos}(\lambda_2x)\text{sin}(\lambda_1y)\text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{cot} (\lambda_1 b) = - \lambda_2 \text{tan} (\lambda_2 a) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}x) \text{sin}(\lambda_1y) \text{sin}(\lambda_2z) & \sigma&= \lambda_1 \text{cot} (\lambda_1 b) = \lambda_2 \text{cot} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = y \text{cos} (\lambda x) \text{sinh} (\lambda z) & \sigma&= 1/b, \text{tanh}(\lambda c) = \lambda b = -\text{cot}(\lambda a) \\\hline s(x,y,z) & = y \text{cosh} (\lambda x) \text{sin} (\lambda z) & \sigma&= 1/b, \text{tan}(\lambda c) = \lambda b = \text{coth}(\lambda a) \\\hline s(x,y,z) & = z \text{cos} (\lambda x) \text{sinh} (\lambda y) & \sigma&= 1/c, \text{tanh}(\lambda b) = \lambda c = -\text{cot}(\lambda a) \\\hline s(x,y,z) & = z \text{cosh} (\lambda x) \text{sin} (\lambda y) & \sigma&= 1/c, \text{tan}(\lambda b) = \lambda c = \text{coth}(\lambda a) \\\hline \end{align*} 5. CLASS 100: Even in $y$, $z$, Odd in $x$. \begin{align*} & Eigenfunction & & Eigenvalue \\\Xhline{5\arrayrulewidth} s(x,y,z) & = \text{sinh}(\lambda_2x) \text{cosh}(\lambda_1y) \text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1b) = \lambda_2 \text{coth} (\lambda_2 a) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2} c) \\\hline s(x,y,z) & = \text{sinh}(\lambda_2x) \text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}y) \text{cosh}(\lambda_1z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1 c) = \lambda_2 \text{coth} (\lambda_2a) = -\sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{cosh}(\lambda_2y)\text{cosh}(\lambda_1z) & \sigma&= \lambda_1 \text{tanh} (\lambda_1 c) = \lambda_2 \text{tanh} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{sin}(\lambda_2x)\text{cos}(\lambda_1y)\text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1b) =\lambda_2 \text{cot} (\lambda_2 a) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{sin}(\lambda_2x)\text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{cos}(\lambda_1z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1 c) = \lambda_2 \text{cot} (\lambda_2 a) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{cos}(\lambda_2y)\text{cos}(\lambda_1z) & \sigma&= -\lambda_1 \text{tan} (\lambda_1 c) = -\lambda_2 \text{tan} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = x \text{cos} (\lambda y) \text{cosh} (\lambda z) & \sigma&= 1/a, \text{coth}(\lambda c) = \lambda a = -\text{cot}(\lambda b) \\\hline s(x,y,z) & = x \text{cosh} (\lambda y) \text{cos} (\lambda z) & \sigma&= 1/a, \text{coth}(\lambda b) = \lambda a = -\text{cot}(\lambda c) \\\hline \end{align*} 6. CLASS 101: Odd in $x$, $z$, Even in $y$. \begin{align*} & Eigenfunction & & Eigenvalue \\\Xhline{5\arrayrulewidth} s(x,y,z) & = \text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{cosh}(\lambda_2y) \text{sinh}(\lambda_1z) & \sigma&= \lambda_1 \text{coth} (\lambda_1c) = \lambda_2 \text{tanh} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2} a) \\\hline s(x,y,z) & = \text{sinh}(\lambda_1x)\text{cosh}(\lambda_2y) \text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{coth} (\lambda_1 a) = \lambda_2 \text{tanh} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{sinh}(\lambda_1x)\text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}y) \text{sinh}(\lambda_2z) & \sigma&= \lambda_1 \text{coth} (\lambda_1 a) = \lambda_2 \text{coth} (\lambda_2c) = - \sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{cos}(\lambda_2y) \text{sin}(\lambda_1z) & \sigma&= \lambda_1 \text{cot} (\lambda_1c) = - \lambda_2 \text{tan} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{sin}(\lambda_1x)\text{cos}(\lambda_2y)\text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{cot} (\lambda_1 a) = - \lambda_2 \text{tan} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{sin}(\lambda_1x)\text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}y) \text{sin}(\lambda_2z) & \sigma&= \lambda_1 \text{cot} (\lambda_1a) = \lambda_2 \text{cot} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = x \text{cos} (\lambda y) \text{sinh} (\lambda z) & \sigma&= 1/a, \text{tanh}(\lambda c) = \lambda a = -\text{cot}(\lambda b) \\\hline s(x,y,z) & = x \text{cosh} (\lambda y) \text{sin} (\lambda z) & \sigma&= 1/a, \text{tan}(\lambda c) = \lambda a = \text{coth}(\lambda b) \\\hline s(x,y,z) & = z \text{sinh} (\lambda x) \text{cos} (\lambda y) & \sigma&= 1/c, \text{tanh}(\lambda a) = \lambda c = -\text{cot}(\lambda b) \\\hline s(x,y,z) & = z \text{sin} (\lambda x) \text{cosh} (\lambda y) & \sigma&= 1/c, \text{tan}(\lambda a) = \lambda c = \text{coth}(\lambda b) \\\hline \end{align*} 7. CLASS 110: Odd in $x$, $y$, Even in $z$. \begin{align*} & Eigenfunction & & Eigenvalue \\\Xhline{5\arrayrulewidth} s(x,y,z) & = \text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{sinh}(\lambda_1y)\text{cosh}(\lambda_2z) & \sigma&= \lambda_1 \text{coth} (\lambda_1b) = \lambda_2 \text{tanh} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{sinh}(\lambda_1x)\text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{cosh}(\lambda_2z) & \sigma&= \lambda_1 \text{coth} (\lambda_1 a) = \lambda_2 \text{tanh} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{sinh}(\lambda_1x)\text{sinh}(\lambda_2y)\text{cos}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{coth} (\lambda_1 a) = \lambda_2 \text{coth} (\lambda_2b) = - \sqrt{\lambda_1^2 + \lambda_2^2}\text{tan}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{sin}(\lambda_1y)\text{cos}(\lambda_2z) & \sigma&= \lambda_1 \text{cot} (\lambda_1b) = - \lambda_2 \text{tan} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{sin}(\lambda_1x)\text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{cos}(\lambda_2z) & \sigma&= \lambda_1 \text{cot} (\lambda_1 a) = - \lambda_2 \text{tan} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{sin}(\lambda_1x)\text{sin}(\lambda_2y)\text{cosh}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{cot} (\lambda_1a) = \lambda_2 \text{cot} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{tanh}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = x \text{sinh} (\lambda y) \text{cos} (\lambda z) & \sigma&= 1/a, \text{tanh}(\lambda b) = \lambda a = -\text{cot}(\lambda c) \\\hline s(x,y,z) & = x \text{sin} (\lambda y) \text{cosh} (\lambda z) & \sigma&= 1/a, \text{tan}(\lambda b) = \lambda a = \text{coth}(\lambda c) \\\hline s(x,y,z) & = y \text{sinh} (\lambda x) \text{cos} (\lambda z) & \sigma&= 1/b, \text{tanh}(\lambda a) = \lambda b = -\text{cot}(\lambda c) \\\hline s(x,y,z) & = y \text{sin} (\lambda x) \text{cosh} (\lambda z) & \sigma&= 1/b, \text{tan}(\lambda a) = \lambda b = \text{coth}(\lambda c) \\\hline \end{align*} 8. CLASS 111: Odd in $x$, $y$, $z$. \begin{align*} & Eigenfunction & & Eigenvalue \\\Xhline{5\arrayrulewidth} s(x,y,z) & = xyz & \sigma &=1/a =1/b =1/c \\\hline s(x,y,z) & = \text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{sinh}(\lambda_1y)\text{sinh}(\lambda_2z) & \sigma&= \lambda_1 \text{coth} (\lambda_1b) = \lambda_2 \text{coth} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{sinh}(\lambda_1x)\text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{sinh}(\lambda_2z) & \sigma&= \lambda_1 \text{coth} (\lambda_1a) = \lambda_2 \text{coth} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{sinh}(\lambda_1x)\text{sinh}(\lambda_2y)\text{sin}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{coth} (\lambda_1a) = \lambda_2 \text{coth} (\lambda_2b) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{cot}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = \text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}x)\text{sin}(\lambda_1y)\text{sin}(\lambda_2z) & \sigma&= \lambda_1 \text{cot} (\lambda_1b) = \lambda_2 \text{cot} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}a) \\\hline s(x,y,z) & = \text{sin}(\lambda_1x)\text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}y)\text{sin}(\lambda_2z) & \sigma&= \lambda_1 \text{cot} (\lambda_1 a) = \lambda_2 \text{cot} (\lambda_2c) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}b) \\\hline s(x,y,z) & = \text{sin}(\lambda_1x)\text{sin}(\lambda_2y)\text{sinh}(\sqrt{\lambda_1^2+\lambda_2^2}z) & \sigma&= \lambda_1 \text{cot} (\lambda_1b) = \lambda_2 \text{cot} (\lambda_2a) = \sqrt{\lambda_1^2 + \lambda_2^2}\text{coth}(\sqrt{\lambda_1^2 + \lambda_2^2}c) \\\hline s(x,y,z) & = x \text{sinh} (\lambda y) \text{sin} (\lambda z) & \sigma&= 1/a, \text{tanh}(\lambda b) = \lambda a = \text{tan}(\lambda c) \\\hline s(x,y,z) & = x \text{sin} (\lambda y) \text{sinh} (\lambda z) & \sigma&= 1/a, \text{tan}(\lambda b) = \lambda a = \text{tanh}(\lambda c) \\\hline s(x,y,z) & = y \text{sinh} (\lambda x) \text{sin} (\lambda z) & \sigma&= 1/b, \text{tanh}(\lambda a) = \lambda b = \text{tan}(\lambda c) \\\hline s(x,y,z) & = y \text{sin} (\lambda x) \text{sinh} (\lambda z) & \sigma&= 1/b, \text{tan}(\lambda a) = \lambda b = \text{tanh}(\lambda c) \\\hline s(x,y,z) & = z \text{sinh} (\lambda x) \text{sin} (\lambda y) & \sigma&= 1/c, \text{tanh}(\lambda a) = \lambda c = \text{tan}(\lambda b) \\\hline s(x,y,z) & = z \text{sin} (\lambda x) \text{sinh} (\lambda y) & \sigma&= 1/c, \text{tan}(\lambda a) = \lambda c = \text{tanh}(\lambda b) \\\hline \end{align*} \vspace{1mm} \noindent We may plot an analogue of Figure \ref{fig:smallest} by considering the cuboid with $c=1$ and $a,b \leq 1$ without loss of generality. \FloatBarrier \begin{figure}[!htbp] \centering \includegraphics[width=1\textwidth]{SmallestSteklovInvariantOnCuboid.JPG} \caption{\label{fig:smallestoncuboid}The 1st Steklov invariant for a cuboid on $[-a,a]\times[-b,b]\times[-1,1]$.} \end{figure} \FloatBarrier \vspace{1mm} \noindent Drawing similarities with the 2D scenario, it would seem likely that the following holds: \begin{conjecture} The surface in Figure \ref{fig:smallestoncuboid} is differentiable and tends to the origin (in the 1st octant). \end{conjecture} \section{Analysis of the 1st Steklov eigenspace on the cuboid} We may assume, without loss of generality, that $0 \leq a \leq b \leq c=1$. Drawing similarities with the 2D scenario, it is likely that the following holds: \begin{conjecture} $\sigma_1$ is always due to the same one eigenfunction, namely cosh$(\lambda_1 x)$cosh$(\lambda_2 y)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}z)$. Further, if $a<b<1$, then this is the unique eigenfunction which gives $\sigma_1$. Otherwise, there may be other eigenfunctions that give this eigenvalue and $\sigma_1$ becomes a multiple eigenvalue. \end{conjecture} \section{The square and the cube} Now we will perform some explicit numerical calculations for the square and the cube. \vspace{1mm} \noindent It is important to note that for the square on $[-1,1]\times[-a,a]$ and the cube on $[-a,a]\times[-b,b]\times[-1,1]$, there are redundancies in the determining equations (sometimes across parity classes); however the eigenfunctions are still distinct, which gives rise to an eigenvalue being multiple. For instance, for the square, parity classes I(i) and I(ii) give the same eigenvalue, as do classes III(i) and IV(ii). The latter is an example of multiplicity across classes. \vspace{1mm} \noindent Let us collect, for the square, the eigenfunctions according to determining equation, along with their corresponding candidate for $\sigma_1$. We include the eigenfunction $s(x,y)=xy$. \FloatBarrier \begin{table}[!htbp] \centering \begin{tabular}{l|c|c|r} Eigenfunctions & Determining Equation & Eigenvalue $\sigma$\\\hline\hline cosh$(\nu x)$cos$(\nu y)$, & tan$(\nu)+$tanh$(\nu)=0$ & $\nu$tanh$(\nu)$ \\ cos$(\nu x)$cosh$(\nu y)$ & Candidate for $\nu_1 = 2.3650203...$ & Candidate for $\sigma_1 = 2.3236377...$ \\\hline sinh$(\nu x)$sin$(\nu y)$, & tan$(\nu)=$tanh$(\nu)$ & $\nu$coth$(\nu)$ \\ sin$(\nu x)$sinh$(\nu y)$ & Candidate for $\nu_1 = 3.9266023...$ & Candidate for $\sigma_1 = 3.9296545...$ \\\hline cosh$(\nu x)$sin$(\nu y)$, & tan$(\nu)=$coth$(\nu)$ & $\nu$tanh$(\nu)$ \\ sin$(\nu x)$cosh$(\nu y)$ & Candidate for $\nu_1 = 0.9375520...$ & Candidate for $\sigma_1 = 0.6882527...$ \\\hline sinh$(\nu x)$cos$(\nu y)$, & tan$(\nu)+$coth$(\nu)=0$ & $\nu$coth$(\nu)$ \\ cos$(\nu x)$sinh$(\nu y)$ & Candidate for $\nu_1 = 2.3470455...$ & Candidate for $\sigma_1 = 2.3903892...$ \\\hline $xy$ & - & $1$ \\ & & Candidate for $\sigma_1 = 1$ \end{tabular} \caption{\label{tablesquare}For the square, eigenfunctions collect into pairs.} \end{table} \FloatBarrier \vspace{1mm} \noindent Therefore, comparing all the candidates in Table \ref{tablesquare}, we see that the correct value for $\sigma_1$ is $0.6882527...$, whose two eigenfunctions are cosh$(\nu x)$sin$(\nu y)$ and sin$(\nu x)$cosh$(\nu y)$. The perimeter of our square is 8 units, so the first Steklov invariant for any square is $\sigma_1L=8\sigma_1=5.506...$, which, as one can double-check, is the maximum in Figure \ref{fig:smallest}. \vspace{1mm} \noindent (Of course, we already know from the previous section that class IV(ii) must give $\sigma_1$. However, since $a=1$, class III(i) also gives $\sigma_1$. See Claim \ref{iffsquare}.) \vspace{1mm} \noindent We perform a similar calculation for the cube, collecting the Steklov eigenfunctions according to the candidate for $\sigma_1$ which they produce. For brevity, we omit those eigenfunctions such as $s(x,y,z)=xyz$ which give $\sigma_1 = 1$; the following table lists only the cases with non-degenerate determining equations. \FloatBarrier \begin{table}[!htbp] \centering \begin{tabular}{l|c|c|r} Eigenfunctions & Determining Equation & Eigenvalue $\sigma$\\\hline\hline cosh$(\lambda_1 x)$cosh$(\lambda_2 y)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$tanh$\lambda_1$ = $\lambda_2$tanh$\lambda_2$ = $-\sqrt{\lambda_1^2+\lambda_2^2}$tan$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\lambda_1$tanh$\lambda_1$ \\ cosh$(\lambda_1 x)$cosh$(\lambda_2 z)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(1.8041319,1.8041319)$ & Candidate for $\sigma_1$ = 1.7089319 \\ cosh$(\lambda_1 y)$cosh$(\lambda_2 z)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline cos$(\lambda_1 x)$cos$(\lambda_2 y)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $-\lambda_1$tan$\lambda_1$ = $-\lambda_2$tan$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$tanh$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\sqrt{\lambda_1^2+\lambda_2^2}$tanh$(\sqrt{\lambda_1^2+\lambda_2^2})$ \\ cos$(\lambda_1 x)$cos$(\lambda_2 z)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(2.1882115, 2.1882115)$ & Candidate for $\sigma_1$ = 3.0819274 \\ cos$(\lambda_1 y)$cos$(\lambda_2 z)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline cosh$(\lambda_1 x)$sinh$(\lambda_2 y)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$tanh$\lambda_1$ = $\lambda_2$coth$\lambda_2$ = $-\sqrt{\lambda_1^2+\lambda_2^2}$tan$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\lambda_1$tanh$\lambda_1$ \\ cosh$(\lambda_1 x)$sinh$(\lambda_2 z)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(1.883677,1.677149)$ & Candidate for $\sigma_1$ = 1.7985693 \\ cosh$(\lambda_1 y)$sinh$(\lambda_2 z)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\ sinh$(\lambda_1 x)$cosh$(\lambda_2 y)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & & \\ sinh$(\lambda_1 x)$cosh$(\lambda_2 z)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & & \\ sinh$(\lambda_1 y)$cosh$(\lambda_2 z)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline cosh$(\lambda_1 x)$cosh$(\lambda_2 y)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$tanh$\lambda_1$ = $\lambda_2$tanh$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$cot$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\lambda_1$tanh$\lambda_1$ \\ cosh$(\lambda_1 x)$cosh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(,)$ & Candidate for $\sigma_1$ = 0.5315091 \\ cosh$(\lambda_1 y)$cosh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline cos$(\lambda_1 x)$sin$(\lambda_2 y)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $-\lambda_1$tan$\lambda_1$ = $\lambda_2$cot$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$tanh$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\sqrt{\lambda_1^2+\lambda_2^2}$tanh$(\sqrt{\lambda_1^2+\lambda_2^2})$ \\ cos$(\lambda_1 x)$sin$(\lambda_2 z)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(2.0017440,3.8679675)$ & Candidate for $\sigma_1$ = 4.3538085 \\ cos$(\lambda_1 y)$sin$(\lambda_2 z)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\ sin$(\lambda_1 x)$cos$(\lambda_2 y)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & & \\ sin$(\lambda_1 x)$cos$(\lambda_2 z)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & & \\ sin$(\lambda_1 y)$cos$(\lambda_2 z)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline cos$(\lambda_1 x)$cos$(\lambda_2 y)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $-\lambda_1$tan$\lambda_1$ = $-\lambda_2$tan$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$coth$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\sqrt{\lambda_1^2+\lambda_2^2}$coth$(\sqrt{\lambda_1^2+\lambda_2^2})$ \\ cos$(\lambda_1 x)$cos$(\lambda_2 z)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(2.1843218,2.1843218)$ & Candidate for $\sigma_1$ = 3.1019388 \\ cos$(\lambda_1 y)$cos$(\lambda_2 z)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline sinh$(\lambda_1 x)$cosh$(\lambda_2 y)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$coth$\lambda_1$ = $\lambda_2$tanh$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$cot$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\lambda_1$coth$\lambda_1$ \\ sinh$(\lambda_1 x)$cosh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(2.9145019,2.8791763)$ & Candidate for $\sigma_1$ = 2.8974090 \\ sinh$(\lambda_1 y)$cosh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\ cosh$(\lambda_1 x)$sinh$(\lambda_2 y)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & & \\ cosh$(\lambda_1 x)$sinh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & & \\ cosh$(\lambda_1 y)$sinh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline sinh$(\lambda_1 x)$sinh$(\lambda_2 y)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$coth$\lambda_1$ = $\lambda_2$coth$\lambda_2$ = $-\sqrt{\lambda_1^2+\lambda_2^2}$tan$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\lambda_1$coth$\lambda_1$ \\ sinh$(\lambda_1 x)$sinh$(\lambda_2 z)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(1.7665698,1.7665698)$ & Candidate for $\sigma_1$ = 1.8728895 \\ sinh$(\lambda_1 y)$sinh$(\lambda_2 z)$cos$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline sin$(\lambda_1 x)$cos$(\lambda_2 y)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$cot$\lambda_1$ = $-\lambda_2$tan$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$coth$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\sqrt{\lambda_1^2+\lambda_2^2}$coth$(\sqrt{\lambda_1^2+\lambda_2^2})$ \\ sin$(\lambda_1 x)$cos$(\lambda_2 z)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(2.0014790,3.8676451)$ & Candidate for $\sigma_1$ = 4.3562732 \\ sin$(\lambda_1 y)$cos$(\lambda_2 z)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\ cos$(\lambda_1 x)$sin$(\lambda_2 y)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & & \\ cos$(\lambda_1 x)$sin$(\lambda_2 z)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & & \\ cos$(\lambda_1 y)$sin$(\lambda_2 z)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline sin$(\lambda_1 x)$sin$(\lambda_2 y)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$cot$\lambda_1$ = $\lambda_2$cot$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$tanh$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\sqrt{\lambda_1^2+\lambda_2^2}$tanh$(\sqrt{\lambda_1^2+\lambda_2^2})$ \\ sin$(\lambda_1 x)$sin$(\lambda_2 z)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(0.7371448,0.7371448)$ & Candidate for $\sigma_1$ = 0.8119520 \\ sin$(\lambda_1 y)$sin$(\lambda_2 z)$cosh$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline sinh$(\lambda_1 x)$sinh$(\lambda_2 y)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$coth$\lambda_1$ = $\lambda_2$coth$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$cot$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\lambda_1$coth$\lambda_1$ \\ sinh$(\lambda_1 x)$sinh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(2.8949112,2.8949112)$ & Candidate for $\sigma_1$ = 2.9126739 \\ sinh$(\lambda_1 y)$sinh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline sin$(\lambda_1 x)$sin$(\lambda_2 y)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}z)$ & $\lambda_1$cot$\lambda_1$ = $\lambda_2$cot$\lambda_2$ = $\sqrt{\lambda_1^2+\lambda_2^2}$coth$(\sqrt{\lambda_1^2+\lambda_2^2})$ & $\sqrt{\lambda_1^2+\lambda_2^2}$coth$(\sqrt{\lambda_1^2+\lambda_2^2})$ \\ sin$(\lambda_1 x)$sin$(\lambda_2 z)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ & Candidate for $((\lambda_1)_1,(\lambda_2)_1)$ = $(3.7570495,3.7570495)$ & Candidate for $\sigma_1$ = 5.3135282 \\ sin$(\lambda_1 y)$sin$(\lambda_2 z)$sinh$(\sqrt{\lambda_1^2+\lambda_2^2}x)$ & & \\\hline \end{tabular} \caption{\label{tablecube}For the cube, eigenfunctions collect to give twelve categories with non-degenerate determining equations. (The other cases produce candidates of $\sigma_1 = 1$). (Calculations are rounded up to 7 decimal places.)} \end{table} \FloatBarrier \vspace{1mm} \noindent Comparing all the candidates in Table \ref{tablecube}, we see that the correct value for $\sigma_1$ is $0.5315091$, whose three eigenfunctions are cosh$(\lambda_1 x)$cosh$(\lambda_2 y)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}z)$, cosh$(\lambda_1 x)$cosh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}y)$ and cosh$(\lambda_1 y)$cosh$(\lambda_2 z)$sin$(\sqrt{\lambda_1^2+\lambda_2^2}x)$. \\ The surface area of our cube is 24 square units, so the first Steklov invariant for any cube is $\sigma_1 \sqrt{24}=2.603...$, which, as one can double-check, is the maximum in Figure \ref{fig:smallestoncuboid}. \input{Bibliography.tex} \end{document}

https://arxiv.org/abs/1907.07172

Ordinal pattern probabilities for symmetric random walks

An ordinal pattern for a finite sequence of real numbers is a permutation that records the relative positions in the sequence. For random walks with steps drawn uniformly from $[-1,1]$, we show an ordinal pattern occurs with probability $\frac{|[1,w]|}{2^n n!}$, where $[1,w]$ is a weak order interval in the affine Weyl group $\widetilde{A}_n$. For random walks with steps drawn from a symmetric Laplace distribution, the probability is $\frac{1}{2^n \prod_{j=1}^n \mathrm{lev}(\pi)_j}$, where $\mathrm{lev}(\pi)_j$ measures how often $j$ occurs between consecutive values in $\pi$. Permutations whose consecutive values are at most two positions apart in $\pi$ are shown to occur with the same probability for any choice of symmetric continuous step distribution. For random walks with steps from a mean zero normal distribution, ordinal pattern probabilities are determined by a matrix whose $ij$-th entry measures how often $i$ and $j$ are between consecutive values.

\section{Introduction} \label{s:introduction} Let $(a_1,\ldots,a_n) \in \R^n$ be an arbitrary finite sequence of real numbers. A permutation $\pi$ such that $\pi(i) = j$ if $a_i$ is the $j$-th largest position is called the \emph{ordinal pattern for $(a_1,\ldots,a_n)$}. For a given sequence $Z_1,Z_2,\ldots$ of continuous random variables, it is natural to ask what the probability is that a given permutation $\pi \in S_n$ occurs as an ordinal pattern in a length $n$ subsequence of outcomes. It is known \cite{bandt-shiha} that for exchangeable random variables, such as those that are independent and identically distributed, this probability is $1/n!$ for all $\pi \in S_n$. By contrast, the distribution on $S_n$ is never uniform for ordinal patterns in positions of a random walk when $n \geq 3$. Exact probabilities have been calculated for ordinal pattern occurrence in a random walk for a few cases. For $n = 3,4$, Bandt and Shiha \cite{bandt-shiha}, DeFord and Moore \cite{deford-moore}, and Zare \cite{zare} gave values for the case of normally distributed steps of mean zero. For $n = 3,4$, DeFord and Moore \cite{deford-moore} gave piece-wise polynomials for the case of uniform distributions on $[b-1,b]$ for $b \in \left[\frac{1}{2},1\right)$. In \cite{elizalde-martinez} and \cite{martinez}, Elizalde and Martinez showed that certain pairs of permutations have the same probability of occurring as an ordinal pattern in a random walk regardless of choice of continuous step distribution. Furthermore, Martinez \cite{martinez} gave a detailed description of regions of steps that generate a given ordinal pattern $\pi$ in terms of a hyperplane arrangement equivalent to the braid arrangement. In this paper, we use hyperplane arrangements and the tools developed in \cite{elizalde-martinez} and \cite{martinez} to find probabilities for ordinal pattern occurrence in certain random walks. \subsection{Main results} \label{ss:results} Let $X_1, X_2,\ldots$ be independent and identically distributed continuous random variables called \emph{steps} with probability density function $f:\R \ra \R$. Let $Z_i := \sum_{j = 1}^{i - 1} X_j$ be the \emph{positions} of the random walk. For $\pi \in S_{n+1}$, let $\PRB(f,\pi)$ denote the probability that $\pi$ occurs as an ordinal pattern in a length $n + 1$ consecutive subsequence of positions generated by $n$ steps drawn from $f$. All of the main results of the paper are statements about $\PRB(f,\pi)$ for various choices of density function $f$. Let $\mathrm{lev}(\pi)_j$ denote the number of positions $i$ such that $\pi(i) \leq j < \pi(i+1)$ or such that $\pi(i+1) \leq j < \pi(i)$. In Section \ref{s:laplace}, we use a direct calculation to show that when $f$ is a Laplace distribution, the value of $\PRB(f,\pi)$ is computed from the values of $\mathrm{lev}(\pi)_1,\ldots,\mathrm{lev}(\pi)_n$. \\ \\ \noindent \textbf{Theorem \ref{t:laplace}.} Let $\pi \in S_{n+1}$ and let $f$ be the density function for a Laplace distribution with mean zero. Then \[ \PRB(f,\pi) = \frac{1}{2^n \prod_{j=1}^n \mathrm{lev}(\pi)_j}. \] We say a permutation is \emph{almost consecutive} if its consecutive values are at most two positions apart in its $1$-line notation. Recall that a function is symmetric if $f(-x) = f(x)$ for all $x \in \R$. In Section \ref{s:universal}, we show that $\PRB(f,\pi)$ does not depend upon the choice of symmetric density function $f$ if $\pi$ is almost consecutive. \\ \\ \noindent \textbf{Theorem \ref{t:acformula}.} Let $\pi \in S_{n+1}$ be an almost consecutive permutation. Let $f:\R \ra \R$ be a symmetric density function for a continuous probability distribution. Then \[ \PRB(f,\pi) = \frac{1}{2^n \PROD_{i=1}^n \mathrm{lev}(\pi)_i}. \] In Section \ref{s:affineA}, we show that when $f$ is the density function for the uniform distribution on $[-1,1]$, we calculate $\PRB(f,\pi)$ by counting regions of the {\typeA} inside a rational polytope constructed from $\pi$. The counted regions correspond to elements in a weak order interval of the {\affWeyl}. \\ \\ \noindent \textbf{Theorem \ref{t:weakordertheorem}.} Let $f = \frac{1}{2}\Chi_{[-1,1]}$ be the uniform density function on $[-1,1]$. Let $\pi \in S_{n+1}$. Then \[ \PRB(f,\pi) = \frac{|[1,w]|}{2^n n!}, \] where $[1,w]$ is a weak order interval of the {\affWeyl}. \\ A corollary is that $1 (n + 1)$ or $(n + 1) 1$ appears in the $1$-line notation for $\pi \in S_{n+1}$ if and only if $\PRB(f,\pi) = \frac{1}{2^n n!}$. By contrast, for the Laplace and normal distributions, the other values in the $1$-line notation for $\pi$ typically influence the value of $\PRB(f,\pi)$. In Section \ref{s:normaldistribution}, we show that when $f$ is the density function for a mean zero normal distribution, we can sometimes compare $\PRB(f,\pi)$ to $\PRB(f,\tau)$. Let $\mathrm{lev}(\pi)_{i,j}$ be the number of positions $k \in [n]$ satisfying $\pi(k) \leq i,j < \pi(k+1)$ or $\pi(k+1) \leq i,j < \pi(k)$. \\ \\ \noindent \textbf{Theorem \ref{t:comparegaussian}.} Let $f:\R \ra \R$ be the density function for the normal distribution with mean zero and any variance. Let $\pi, \tau \in S_{n+1}$. Suppose $\mathrm{lev}(\pi)_{i,j} \leq \mathrm{lev}(\tau)_{i,j}$ for all $(i,j) \in \Phi^+$. Then $\PRB(f,\pi) \geq \PRB(f,\tau)$. \subsection{The larger context for these results} \label{ss:greatercontext} Using ordinal patterns to analyze a time series is sometimes called \emph{ordinal analysis}. Bandt and Pompe \cite{bandt-pompe} introduced permutation entropy, which involves computing ordinal pattern frequencies in a time series. Subsequently, many papers suggested applying ordinal analysis to a variety of applied contexts. The survey \cite{amigo} shows that ordinal pattern frequency often captures qualitative features of a time series. For example, iterated map dynamical systems with deterministic chaotic behavior tend to have forbidden ordinal patterns, whereas white noise does not. We can interpret the results of this paper as providing a kind of fingerprint for certain random processes. DeFord and Moore \cite{deford-moore} define KL divergence for the distribution of patterns of length $n$ in $Z$ from those in $X$ by \[ \mathcal{D}_{\text{KL}_n}(X || Z) = \sum_{\pi \in S_n} P_X(\pi) \log\left(\frac{P_X(\pi)}{P_Z(\pi)}\right), \] where $X$ and $Z$ are random variables. Thus, comparisons to walks with steps from symmetric uniform and Laplace densities can be made via this version of KL divergence. By Theorem \ref{t:acformula}, there exists a value $p_n$ for the probability that an almost consecutive permutation arises from a length $n$ sequence generated by $n - 1$ steps from a symmetric density function. Thus, for fixed $n$, we have a Bernoulli trial whose ``success'' probability is the same for any random walk whose steps have a symmetric density function. In \cite{denoncourt}, the following values are determined for $p_n$: \begin{center} \begin{tabular} { c | c } $n$ & $p_n$\\ \hline $2$ & $1$\\ $3$ & $1$\\ $4$ & $2/3$\\ $5$ & $5/12$\\ $6$ & $251/960$\\ $7$ & $463/2880$\\ $8$ & $15281 / 161280$ \end{tabular} \end{center} Although the symmetric density function hypothesis is limited in scope, the above values hold for the mean zero Gaussian distribution. Thus, after transforming a sequence generated by a random walk to have mean zero steps, it is reasonable to expect to see values close to those given above on a long enough time scale. \section{Ordinal pattern preliminaries} \label{s:ordinal} Throughout the paper, random walks have $n$ steps and $n + 1$ positions. The $n$ steps are outcomes of $n$ independent and identically distributed continuous random variables $X_1,\ldots,X_n$. Every tuple $(x_1,\ldots,x_n) \in \R^n$ of steps generates a tuple $(z_1,\ldots,z_{n+1}) \in \R^{n+1}$ of walk positions, where $z_1 = 0$, and $z_i = x_1 + \cdots + x_{i-1}$ for $i > 1$. We say a walk $(z_1,\ldots,z_{n+1})$ has \emph{ordinal pattern $\pi \in S_{n+1}$} if $\pi(i) = j$ whenever $z_i$ is the $j$-th largest position of $(z_1,\ldots,z_{n+1})$. We refer to $(x_1,\ldots,x_n)$ as \emph{step coordinates} for the random walk and the generated tuple $(z_1,\ldots,z_{n+1})$ as \emph{walk coordinates} for the random walk. Ordinal pattern probabilities are calculated as integrals over regions of $\R^n$, but the ordinal patterns themselves are permutations in $S_{n+1}$ derived from walk coordinates in $\R^{n+1}$. Following \cite{elizalde-martinez}, we define a map $p:\R^n \setminus Z \ra S_{n+1}$ by \begin{equation*} p(x_1,\ldots,x_n) = \pi, \end{equation*} where \begin{equation*} \pi(i) = \left|\{j \in \{1,\ldots,n+1\}: z_i \geq z_j\} \right| \end{equation*} and $Z$ is the measure zero set of steps $(x_1,\ldots,x_n)$ such that $z_i = z_j$ for some $i \neq j$. \begin{definition} \label{d:generate} Let $\pi \in S_{n+1}$. Let $\bm{x} \in \R^n$. We say $\bm{x}$ \emph{generates $\pi$} if $p(\bm{x}) = \pi$. We denote the region of tuples that generate $\pi$ by $D_\pi$. Thus, \[ D_\pi = \{\bm{x} \in \R^n : p(\bm{x}) = \pi \}. \] \end{definition} We interpret $p$ as mapping a tuple of steps to the ordinal pattern of the generated walk positions. The region $D_\pi$ contains all such tuples for a fixed $\pi \in S_{n+1}$. In \cite[Section 2]{elizalde-martinez}, Elizalde and Martinez define an \emph{edge diagram} as a collection of oriented vertical line segments connecting $(i,\pi(i))$ and $(i,\pi(i+1))$. The orientation is downward if $\pi(i) > \pi(i+1)$ and upward otherwise. A \emph{level} is a vertical interval, denoted $\ubar{j}$, whose $y$-coordinates are in $[j,j+1]$. In the edge diagram, the \emph{edge} $e_i$ is a formal sum $e_i = \sum_{j=\pi(i)}^{\pi(i+1) - 1} \ubar{j}$ if $\pi(i) < \pi(i+1)$ and $e_i = -\sum_{j = \pi(i+1)}^{\pi(i) - 1} \ubar{j}$ if $\pi(i) > \pi(i+1)$. In Section \ref{s:laplace}, we introduce a tuple $\mathrm{lev}(\pi)$ that records the number of edges that contain $\ubar{j}$ or its negation. (See figure \ref{f:edgediagram} for an example of an edge diagram and $\mathrm{lev}(\pi)$.) An edge $e_i$ contains $\ubar{j}$ or its negation if $\pi(i) \leq j < \pi(i+1)$ or $\pi(i+1) \leq j < \pi(i)$. We primarily use the edge diagram as a visual description of $\mathrm{lev}(\pi)$. The edge diagram can be read off from the matrix given in the next definition, as can properties of the region $D_\pi$. \begin{figure}[htb] \centering \begin{tikzpicture}[scale=.5] \foreach \x in {1,2,...,6} \draw[dotted] (0,\x)--(15,\x); \draw(20.5,3.5) node{$\implies \mathrm{lev}(\pi) = (2,4,3,2,2)$}; \draw[thick,*->-*](4.5,3) node[left]{3}--(4.5,1); \draw[thick,*->-*](5.5,1) node[left]{1}--(5.5,5); \draw[thick,*->-*](6.5,5) node[left]{5}--(6.5,6); \draw[thick,*->-*](7.5,6) node[left]{6}--(7.5,2); \draw[thick,*->-*](8.5,2) node[left]{2}--(8.5,4) node[right]{4}; \draw(0.5,7) node{Levels}; \draw(6.0,8) node{Edge diagram}; \draw(6.0,7) node{for $\pi = 315624$}; \draw(11.5,7) node{Level count}; {\foreach \x in {1,2,...,5} \draw (0.5,0.5+\x) node{$\ubar{\x}$}; } \draw (11.5,1.5) node{$2$}; \draw (11.5,2.5) node{$4$}; \draw (11.5,3.5) node{$3$}; \draw (11.5,4.5) node{$2$}; \draw (11.5,5.5) node{$2$}; \end{tikzpicture} \caption{The edge diagram for a permutation and its level count. Figure adapted from \cite{elizalde-martinez}.} \label{f:edgediagram} \end{figure} \begin{definition} \label{d:represent} (Martinez \cite[Section 2.1]{martinez}) Let $L:S_{n+1} \ra GL_n(\C)$ be defined by $L(\pi) = L_\pi$, where $L_\pi$ is a matrix whose entries are given by \[ \left(L_\pi\right)_{ij} := \begin{cases} 1, & \;\;\; \text{if } \pi(i) \leq j < \pi(i+1),\\ -1, & \;\;\; \text{if } \pi(i+1) \leq j < \pi(i),\\ 0 & \;\;\; \text{otherwise.} \end{cases} \] Thus, the $i$-th coordinate of $L_\pi(x_1,\ldots,x_n)$ is given by \[ L_\pi(x_1,\ldots,x_n)_i = \begin{cases} x_{\pi(i)} + \cdots + x_{\pi(i+1) - 1}, & \;\; \text{if } \pi(i) < \pi(i+1),\\ -(x_{\pi(i+1)} + \cdots + x_{\pi(i) - 1}), & \;\; \text{if } \pi(i+1) < \pi(i). \end{cases} \] \end{definition} Thus, the edges in the edge diagram may be expressed as $e_i = \sum_{j=1}^n (L_\pi)_{i,j} \ubar{j}$. That is, the orientation of edges and which levels appear in edges of an edge diagram can be read from the rows of $L_\pi$. \begin{example} \label{ex:ellpi} Let $\pi = 315624$. Then \[ L_\pi = \begin{bmatrix} -1 & -1 & 0 & 0 & 0\\ 1 & 1 & 1 & 1 & 0\\ 0 & 0 & 0 & 0 & 1\\ 0 & -1 & -1 & -1 & -1\\ 0 & 1 & 1 & 0 & 0 \end{bmatrix} . \] \end{example} The next three lemmas capture the basic properties of the representation $L$. \begin{lemma} \label{l:represent} (Martinez \cite[Lemma 2.1.3]{martinez}) The function $L:S_{n+1} \ra GL_n(\C)$ given in Definition \ref{d:represent} is a homomorphism. \end{lemma} \begin{lemma} \label{l:image} (Martinez \cite[Lemma 2.1.4]{martinez}) Let $\pi \in S_{n+1}$. Then \[ L_\pi\left(\R_{> 0}^n\right) = D_\pi. \] \end{lemma} \begin{lemma} \label{l:det} (Martinez \cite[Lemma 2.1.3]{martinez}) Let $\pi \in S_{n+1}$. Then $\mathrm{det}(L_\pi) = \pm 1$. \end{lemma} Since steps are given by independent and identically distributed continuous random variables, the probability that an ordinal pattern $\pi$ occurs depends only on $\pi$ and the associated probability density function $f:\R \ra \R$. The associated joint density function for a walk of $n$ steps is always the function $g:\R^n \ra \R$ defined by $g(x_1,\ldots,x_n) = \prod_{i=1}^n f(x_i)$. Suppose $f:\R \ra \R$ is a density function. Denote the probability that an ordinal pattern $\pi \in S_{n+1}$ occurs in a random walk of $n$ steps by $\PRB(f,\pi)$, where $f$ is the density function for the steps. Thus, \begin{equation} \label{e:density} \PRB(f,\pi) = \int_{D_\pi} f(x_1) f(x_2) \cdots f(x_n) \; \D x_1 \D x_2 \cdots \D x_n. \end{equation} Since $L_\pi$ is invertible and linear, the change-of-variables theorem applies. Since the determinant of $L_\pi$ is $\pm 1$, the absolute value of the Jacobian is always $1$. \begin{lemma} \label{l:general} Let $\pi \in S_{n+1}$. Let $f:\R \ra \R$ be a density function for a continuous distribution. Let $g:\R^n \ra \R$ be the joint density function for the independent steps produced by $f$ defined by $g(x_1,\ldots,x_n) = \prod_{i=1}^n f(x_i)$. Then \[ \PRB(f,\pi) = \INT_{D_\pi} g (\bm{x}) \; \D \bm{x} = \INT_{\R_{> 0}^n} g(L_\pi(\bm{x})) \; \D \bm{x} = \INT_{\R_{> 0}^n} \prod_{i=1}^n f(L_\pi(\bm{x})_i) \; \D \bm{x}, \] where $L_\pi(\bm{x})_i$ is the $i$-th coordinate of $L_\pi(\bm{x})$. \end{lemma} \begin{proof} The second equality follows from Lemma \ref{l:image}, Lemma \ref{l:det}, and the change-of-variables theorem. \end{proof} \begin{example} \label{ex:generalintegral} Suppose $f$ is the density function of a continuous distribution. Lemma \ref{l:general} implies \[ \PRB(f,2413) = \INT_0^\infty \INT_0^\infty \INT_0^\infty f(y + z) \cdot f(-x - y - z) \cdot f(x + y) \; \D x \; \D y \; \D z. \] \end{example} In principle, Lemma \ref{l:general} allows for the calculation of $\PRB(f,\pi)$ for any suitable density function $f$. However, evaluation of the integral is probably computationally infeasible in general. The main reason to use the second integral of Lemma \ref{l:general} instead of the first is that the integration always occurs over $\R_{>0}^n$. Since all probability distributions in this paper are assumed to be continuous, the existence of a (not necessarily continuous) density function is guaranteed. Also, walk positions overlap with probability zero, which implies that ordinal pattern probabilities add to $1$. This is also true of many discrete distributions, but we do not address the discrete distribution case in this paper. Ordinal pattern probabilities are invariant under changes in scale. If $\bm{x} \in D_\pi$, then $c \bm{x} \in D_\pi$ for any $c > 0$. Thus, the region of integration in \eqref{e:density} does not change under the substitution of $c\bm{x}$ for $\bm{x}$, which proves the next lemma. This property is called \emph{scale invariance}. \begin{lemma} \label{l:scale-invariance} Let $f:\R \ra \R$ be a density function of a continuous probability distribution. Let $c \in \R_{>0}$ and let $h:\R \ra \R$ be defined by $h(x) = c f(cx)$. Then $h$ is also a density function and $\PRB(f,\pi) = \PRB(h,\pi)$ for all $\pi \in S_{n+1}$. \end{lemma} \section{Pattern probabilities when steps are from Laplace densities} \label{s:laplace} The Laplace distribution, also called the double exponential distribution, has density function $f:\R \ra \R$ given by \[ f(x) = \frac{1}{2b}\EXP{-\frac{|x - \mu|}{b}}, \] where $\mu$ is the mean and $b$ is a scale parameter. In this section, we restrict our attention to the mean zero case. A mean zero Laplace distribution arises as the distribution for a random variable expressed as the difference of two identically distributed exponential random variables. By Lemma \ref{l:scale-invariance}, ordinal pattern probabilities are scale invariant. Thus, we lose no generality by restricting our attention to the choice $b = 1$. For the remainder of the section, the density function $f$ is defined by \[ f(x) = \frac{1}{2}\EXP{-|x|}. \] \begin{definition} \label{d:levelcount} Let $\pi \in S_{n+1}$. Denote the number of $i \in [n]$ such that $\pi(i) \leq j < \pi(i+1)$ or $\pi(i+1) \leq j < \pi(i)$ by $\mathrm{lev}(\pi)_j$. We call the tuple $(\mathrm{lev}(\pi)_1,\ldots,\mathrm{lev}(\pi)_n)$ the \emph{level count of $\pi$}. \end{definition} Note that $\mathrm{lev}(\pi)_j$ counts the number of times level $\ubar{j}$ is contained in an edge of the edge diagram of $\pi$. (See figure \ref{f:edgediagram}.) Alternativly, it is the sum of the absolute values of the entries in column $j$ of $L_\pi$. \begin{example} Let $\pi = 315624$. Then $\mathrm{lev}(\pi) = (2,4,3,2,2)$ as shown in Figure \ref{f:edgediagram}. \end{example} Recall from Lemma \ref{l:general} that $\PRB(f,\pi) = \int_{\R_{>0}^n} \prod_{i=1}^n f(L_\pi(\bm{x})_i) \D \bm{x}$. \begin{theorem} \label{t:laplace} Let $\pi \in S_{n+1}$ and let $f$ be the density function for a mean zero Laplace distribution. Then \[ \PRB(f,\pi) = \frac{1}{2^n \prod_{j=1}^n \mathrm{lev}(\pi)_j}. \] \end{theorem} \begin{proof} Every factor of the last integrand in Lemma \ref{l:general} has the form $f(\pm(x_a + \cdots + x_b))$, where each $x_k > 0$. Thus, \[ f(\pm(x_a + \cdots + x_b)) = \EXP{-\left|\pm(x_a + \cdots + x_b)\right|} = \EXP{-(x_a + \cdots + x_b)}. \] The term $-x_j$ appears in the above sum whenever $\pi(i) \leq j < \pi(i+1)$ or $\pi(i+1) \leq j < \pi(i)$. Thus, by Definition \ref{d:levelcount}, there are $\mathrm{lev}(\pi)_j$ factors of $\prod_{i=1}^n f(L_\pi(\bm{x})_i)$ contributing $-x_j$ inside the exponential for the overall product. By Lemma \ref{l:general}, \begin{align*} \PRB(f,\pi) &= \INT_{\R_{> 0}^n} \prod_{i=1}^n f(L_\pi(\bm{x})_i) \; \D \bm{x}\\ &= \INT_{\R_{> 0}^n} \frac{1}{2^n}\PROD_{j=1}^n \EXP{-\mathrm{lev}(\pi)_j \; x_j} \; \textrm{d} x_1 \cdots \textrm{d} x_n\\ &= \frac{1}{2^n} \PROD_{j = 1}^n \INT_0^{\infty} \EXP{-\mathrm{lev}(\pi)_j \; x_j} \; \textrm{d} x_j\\ &= \frac{1}{2^n \PROD_{j = 1}^n \mathrm{lev}(\pi)_j}. \end{align*} \end{proof} \section{Universal pattern probabilities for symmetric step densities} \label{s:universal} Martinez \cite[Section 5]{martinez} introduced a hyperplane arrangement $\mathcal{D}_n$ in $\R^n$ such that for any $\pi \in S_{n+1}$, the set $D_\pi$ is a region of $\mathcal{D}_n$. Furthermore, in \cite[Lemma 5.1.2]{martinez} it was shown that the walls of $D_\pi$ are defined by the row vectors of $L_{\pi^{-1}}$. This allows us to show that for certain $\pi$, we may express $D_\pi$ as a union of cells of the type $B$ Coxeter arrangement, which establishes the main result of this section. Namely, permutations whose consecutive values are at most two positions apart have the same ordinal pattern probabilities as the Laplace distribution. Thus, when $\pi$ is such a permutation, we have $\PRB(f,\pi) = \frac{1}{2^n \PROD_{j=1}^n \mathrm{lev}(\pi)_j}$, \emph{regardless of choice of symmetric density function} for the steps in the random walk. \subsection{Hyperplane arrangement preliminaries, notation, and terminology} \label{ss:hyperplaneterminology} The hyperplane arrangement notation and terminology we use in this section is similar to that found in \cite[Section 1.4]{abramenko-brown} or \cite[Chapter 2]{bjorner-etal}. In particular, a \emph{hyperplane arrangement} is a set $\HYP = \{H_i\}_{i \in I}$ of finitely many hyperplanes. In this section, the arrangements under consideration are \emph{central}, which means they pass through the origin. Thus, associated to each $H_i$ is a linear function $f_i:\R^n \ra \R$ such that $H_i = \{\bm{x} \in \R^n :\;\; f_i(\bm{x}) = 0\}$. For each $H_i \in \HYP$, let \begin{align*} H_i^+ &:= \{\bm{x} \in \R^n : \;\;f_i(\bm{x}) > 0\};\\ H_i^0 &:= \{\bm{x} \in \R^n : \;\;f_i(\bm{x}) = 0\}; \text{ and}\\ H_i^- &:= \{\bm{x} \in \R^n : \;\;f_i(\bm{x}) < 0\}. \end{align*} A \emph{cell with respect to $\HYP$} is a nonempty set $C$ obtained by choosing for each $i \in I$ a sign $\sigma_i \in \{+,0,-\}$ such that $\bm{x} \in H_i^{\sigma_i}$ for all $\bm{x} \in C$. The sequence $(\sigma_1,\ldots,\sigma_n)$ is called the \emph{sign sequence for $C$}. The cell $C$ is represented by \begin{equation} \label{e:halfspaces} C = \bigcap_{i \in I} H_i^{\sigma_i}. \end{equation} The intersection may be redundant. Cells such that $\sigma_i \neq 0$ for all $i \in I$ are called \emph{regions}. The regions of $\HYP$, denoted $\REG{\HYP}$, are the nonempty convex open subsets that partition $\R^n \setminus \cup_{i \in I} H_i$. Note that the collection of all cells partition $\R^n$. However, the cells that are not regions have measure zero and thus contribute nothing to the probability calculations of this section. \subsection{A hyperplane arrangement for steps of a random walk} \label{ss:braidarrangement} Let $H_{i,j}$ be the hyperplane defined by \[ H_{i,j} = \left\{(x_1,\ldots,x_n) \in \R^n : \;\; x_i + x_{i+1} + \cdots + x_{j-1} = 0\right\}. \] Let \[ \CS_n := \left\{ H_{i,j} : \;\; i,j \in [n] \text{ and } i < j \right\} \] be the hyperplane arrangement defined in \cite[Section 5.1]{martinez}. As noted in \cite[Section 5.1]{martinez}, the arrangement $\CS_n$ is obtained from the standard braid arrangement via a linear substitution. The arrangements have the same face poset and the same number of regions. However, since the geometry is different and the calculation of $\PRB(f,\pi)$ is not always uniform across regions, we distinguish between the two arrangements in this paper. The next lemma motivates the choice of the arrangement $\CS_n$. \begin{lemma} \label{l:regions} (Martinez \cite[Lemma 5.1.1]{martinez}) The set of regions of $\CS_n$ is $\{D_\pi : \;\;\pi \in S_{n+1}\}$. \end{lemma} We say a cell is a \emph{face} of the region $R$ if the cell's sign sequence matches $R$'s sign sequence except for one hyperplane $H$ whose sign is $0$. In this case, we say that $H$ is a \emph{wall} of $R$. For convenience, let $H_{j,i} = H_{i,j}$ when $j > i$. \begin{lemma} \label{l:walls} (Martinez \cite[Lemma 5.1.2]{martinez}) Let $\pi \in S_{n+1}$. The set of walls of $D_\pi$ is \[ \left\{ H_{\pi^{-1}(i), \pi^{-1}(i + 1)} \in \CS_n: \;\;i \in [n]\right\}. \] \end{lemma} Suppose $\HYP$ and $\HYP'$ be hyperplane arrangements such that $\HYP \subseteq \HYP'$. Then, the cells of $\HYP$ may be written as a union of cells of $\HYP'$. The next lemma follows from the fact that the collection of walls $\HYP_R$ of a region $R$ forms a hyperplane arrangement in its own right. We use this in Section \ref{ss:typeb} to express $D_\pi$ as a union of cells from the type $B$ Coxeter arrangement. \begin{lemma} \label{l:chamberunion} Let $\HYP_R$ be the collection of walls for a region $R$ of a hyperplane arrangement $\HYP$. If $\HYP'$ is any hyperplane arrangement such that $\HYP_R \subseteq \HYP'$, then \[ R = \bigcup_{j = 1}^k C_j, \] for some collection of cells $C_1, \ldots, C_k$ of $\HYP'$. \end{lemma} \subsection{The type \texorpdfstring{$B$}{B} Coxeter arrangement} \label{ss:typeb} We represent a signed permutation on $[n]$ as a pair $(\omega,\epsilon)$, where $\omega \in S_n$ is a permutation on $[n]$ and $\epsilon \in \{-1,+1\}^n$ is a choice of sign for each position. A signed permutation $(\omega,\epsilon)$ acts on $\R^n$ by mapping $(x_1,\ldots,x_n)$ to $\left(\epsilon_1 x_{\omega(1)},\ldots,\epsilon_n x_{\omega(n)}\right)$. The \emph{type $B$ Coxeter arrangement} is defined by the following hyperplanes: \begin{align} \label{e:ijplus} X_{i,j,+} = \{ (x_1,\ldots,x_n) \in \R^n : x_i + x_j &= 0\};\\ \label{e:ijminus} X_{i,j,-} = \{ (x_1,\ldots,x_n) \in \R^n : x_i - x_j &= 0\}; \text{ and}\\ \label{e:coords} X_i = \{ (x_1,\ldots,x_n) \in \R^n : x_i &= 0\}, \end{align} where $i,j \in [n]$ and $i < j$. It is known that the group $B_n$ of all signed permutations acts simply transitively on the regions of the type $B$ hyperplane arrangement, which implies that there are $2^n n!$ regions. See \cite[Section 7]{bjorner-wachs} or \cite[Section 1.15]{humphreys}, for example. Furthermore, the group $B_n$ is generated by reflections so that every group element can be represented as a matrix with determinant $\pm 1$. For a symmetric density function $f:\R \ra \R$, we have $f(-x) = f(x)$ for all $x \in \R$. Let $g:\R^n \ra \R$ be the joint density function defined by $g(x_1,\ldots,x_n) = \prod_{i=1}^n f(x_i)$. Since $f$ is symmetric and products are invariant under permutations, we have $g(w \cdot \bm{x}) = g(\bm{x})$ for any signed permutation $w \in B_n$. \begin{lemma} \label{l:probregions} Let $f:\R \ra \R$ be a symmetric density function of a continuous probability distribution. Let $g:\R^n \ra \R$ be the joint density for the random walk of $n$ steps given by $g(x_1,\ldots,x_n) = \prod_{i=1}^n f(x_i)$. Then, for any signed permutation $w \in B_n$, and any region $R$ of the type $B$ hyperplane arrangement, we have \[ \int_{R} g(\bm{x}) \D \bm{x} = \frac{1}{2^n n!}. \] \end{lemma} \begin{proof} Let $R_i$ and $R_j$ be arbitrary regions. Since $B_n$ acts simply transitively on regions, there exists $w \in B_n$ such that $w(R_i) = R_j$. Since the absolute value of the Jacobian for $w$ is $1$, the fact that $g(w \cdot \bm{x}) = g(\bm{x})$ and the change-of-variables theorem imply \[ \int_{R_i} g(\bm{x}) \D \bm{x} = \int_{R_i} g(w \cdot \bm{x}) \D \bm{x} = \int_{R_j} g(\bm{x}) \D \bm{x}. \] Since there are $2^n n!$ regions, the result follows. \end{proof} Recall Lemma \ref{l:chamberunion}: If the walls of a region $R$ lie in an arrangement $\HYP'$ distinct from the one that defined $R$, then we can write $R$ as a union of cells from $\HYP'$. Thus, if the walls of a region $D_\pi \in \CS_n$ are type $B$ hyperplanes, the value $\PRB(f,\pi)$ can be calculated by counting type B regions contained in $D_\pi$. \begin{lemma} \label{l:countregions} Suppose $f:\R \ra \R$ is a symmetric density function. Suppose the walls of $D_\pi$ are hyperplanes in the type $B$ Coxeter arrangement. Then \[ \PRB(f,\pi) = \frac{1}{2^n \PROD_{i=1}^n \mathrm{lev}(\pi)_i}. \] \end{lemma} \begin{proof} The hypothesis and Lemma \ref{l:chamberunion} imply that \[ D_\pi = \bigcup_{i = 1}^k R_i \; \cup \; \bigcup_{j=1}^{\ell} C_j, \] where each $R_i$ is a region of the type $B$ Coxeter arrangement and each cell $C_j$ is a measure $0$ cell of the arrangement. Thus, \[ \PRB(f,\pi) = \int_{D_\pi} f(x_1)\cdots f(x_n) \D x_1 \cdots \D x_n = \sum_{i=1}^k \int_{R_i} f(x_1)\cdots f(x_n) \D x_1 \cdots \D x_n. \] Since $f$ is symmetric, the joint density function $g(x_1,\ldots,x_n) = \prod_{i=1}^n f(x_i)$ is invariant under the action of $B_n$. By Lemma \ref{l:probregions}, the hypotheses imply $\PRB(f,\pi) = \frac{k}{2^n n!}$, where $k$ is the number of type $B$ regions contained in $D_\pi$. Since $k$ depends only on $\pi$, not on the choice of symmetric density function, we may choose $f$ to be the Laplace distribution. The result then follows from Theorem \ref{t:laplace}. \end{proof} \subsection{Almost consecutive permutations} \label{ss:almostconsecutive} It remains to identify the permutations $\pi \in S_{n+1}$ such that the walls of $D_\pi$ are hyperplanes of the type $B$ Coxeter arrangement. Recall from Lemma \ref{l:walls} that the set of walls for $D_\pi$ is the set of all $H_{\pi^{-1}(i),\pi^{-1}(i+1)}$ such that $i \in [n]$. In Lemma \ref{l:acwalls}, we show the walls of $D_\pi$ are type $B$ hyperplanes if $\pi$ is a permutation whose consecutive values occur no more than two positions apart in its $1$-line notation. These permutations (or their inverses) are called \emph{key permutations} in \cite{page} and \emph{3-determined permutations} in \cite{avgustinovich-kitaev}. In both papers it is shown that the counting sequence for these permutations, which is sequence \href{http://oeis.org/A003274/}{A003274} of the OEIS, grows asymptotically like $(1.4655\ldots)^n$. \begin{definition} \label{d:almostconsecutive} We say $\pi \in S_{n+1}$ is \emph{almost consecutive} if $|\pi^{-1}(i+1) - \pi^{-1}(i)| \leq 2$ for all $i \in [n]$. \end{definition} \begin{example} Let $\pi = 1423$. Then $\pi$ is almost consecutive since all instances of consecutive values are at most two positions apart in the $1$-line notation. By contrast, the permutation $2413$ is not almost consecutive, since the values $2$ and $3$ are three positions apart. \end{example} \begin{lemma} \label{l:acwalls} Let $\pi \in S_{n+1}$ be an almost consecutive permutation. Then the walls of $D_\pi$ are hyperplanes of the type $B$ Coxeter arrangement. \end{lemma} \begin{proof} By Lemma \ref{l:walls}, the walls of $D_\pi$ are $H_{\pi^{-1}(i), \pi^{-1}(i+1)}$, where $i \in [n]$. Definition \ref{d:almostconsecutive} then implies that every wall of $D_\pi$ has the form $H_{k,k+1}$ or $H_{k,k+2}$ for some $k$. Thus, a given wall of $D_\pi$ is defined by an equation of the form $x_k = 0$ or $x_k + x_{k+1} = 0$, which is a hyperplane of the form~\eqref{e:coords} or~\eqref{e:ijplus}. In either case, a wall of $D_\pi$ is a hyperplane in the type $B$ Coxeter arrangement. \end{proof} \begin{theorem} \label{t:acformula} Let $\pi \in S_{n+1}$ be an almost consecutive permutation. Let $f:\R \ra \R$ be a symmetric density function. Then \[ \PRB(f,\pi) = \frac{1}{2^n \PROD_{i=1}^n \mathrm{lev}(\pi)_i}. \] \end{theorem} \begin{proof} The result follows from Lemma \ref{l:acwalls} and Lemma \ref{l:countregions}. \end{proof} \section{Uniform random walk patterns and Affine \texorpdfstring{$A$}{A}} \label{s:affineA} Throughout this section, the only density function $f:\R \ra \R$ under consideration is the uniform density function on $[-1,1]$. It is defined by $f(x) = \frac{1}{2}$ for $x \in [-1,1]$ and $f(x) = 0$ otherwise. In Section \ref{ss:polytope}, we show that $\PRB(f,\pi)$ is related to the volume of a rational polytope derived from $\pi$. This rational polytope turns out to be an alcoved polytope, which is a union of the regions (called alcoves) of the {\typeA}. A lot is known about the {\typeA} and the {\affWeyl} that acts upon its regions. Thus, the early sections of the chapter are devoted to translating everything into the language of type $A$ root systems. The main result, Theorem \ref{t:weakordertheorem}, states that $\PRB(f,\pi)$ can be computed by counting the number of elements of a weak order interval of $\A_n$. \subsection{The polytope of steps that generate \texorpdfstring{$\pi$}{p}} \label{ss:polytope} We now define a rational polytope $P_\pi$ that is used to reduce the problem of calculating $\PRB(f,\pi)$ to the problem of calculating the volume of $P_\pi$. \begin{definition} \label{d:rationalpolytope} Let $\pi \in S_{n+1}$. Let $m_i = \mathrm{min}\{\pi(i),\pi(i+1)\}$ and $M_i = \mathrm{max}\{\pi(i),\pi(i+1)\}$. We call the rational polytope $P_\pi$ satisfying \begin{align} \label{e:rationalcoords} &x_i \geq 0,\\ \label{e:linearsystem} 0 \leq &x_{m_i} + \cdots + x_{M_i - 1} \leq 1, \end{align} for all $i \in [n]$ the \emph{polytope of steps for $\pi$}. \end{definition} \begin{example} \label{ex:linearsystem} Let $\pi = 2413$. The system of inequalities defining $P_\pi$ is given by \begin{align*} 0 \leq x_2 + x_3 &\leq 1,\\ 0 \leq x_1 + x_2 + x_3 &\leq 1, \\ 0 \leq x_1 + x_2 &\leq 1, \text{ and}\\ x_1,x_2,x_3 &\geq 0. \end{align*} \end{example} Recall that Lemma \ref{l:general} expresses $\PRB(f,\pi)$ as $\int_{\R_{>0}^n} \prod_{i=1}^n f(L_\pi(\bm{x})_i) \D \bm{x}$. Also recall Definition \ref{d:represent}, which expresses the $i$-th coordinate of $L_\pi(x_1,\ldots,x_n)$ as \[ L_\pi(x_1,\ldots,x_n)_i = \begin{cases} x_{\pi(i)} + \cdots + x_{\pi(i+1) - 1}, & \;\; \text{if } \pi(i) < \pi(i+1),\\ -(x_{\pi(i+1)} + \cdots + x_{\pi(i) - 1}), & \;\; \text{if } \pi(i+1) < \pi(i). \end{cases} \] \begin{lemma} \label{l:volumepolytope} Let $f = \frac{1}{2} \Chi_{[-1,1]}$ be the uniform density function on $[-1,1]$. Let $\pi \in S_{n+1}$. Let $P_\pi$ be the polytope of steps for $\pi$. Then \[ \PRB(f,\pi) = \frac{1}{2^n} \cdot \mathrm{volume}(P_\pi). \] \end{lemma} \begin{proof} Let $m_i = \mathrm{min}\{\pi(i),\pi(i+1)\}$ and $M_i = \mathrm{max}\{\pi(i),\pi(i+1)\}$. By Lemma \ref{l:general}, we have \begin{equation} \label{e:integralpolytope} \PRB(f,\pi) = \INT_{\R_{>0}^n} \prod_{i=1}^n \frac{1}{2} \Chi_{[-1,1]}(L_\pi(\bm{x})_i) \; \D \bm{x} = \frac{1}{2^n} \INT_{\R_{>0}^n} \prod_{i=1}^n \Chi_{[-1,1]}(L_\pi(\bm{x})_i) \;\D \bm{x}. \end{equation} Let $\bm{x} = (x_1,\ldots,x_n) \in \R_{\geq 0}^n$. Then $x_j \geq 0$ for all $j \in [n]$. Thus $L_\pi(x_1,\ldots,x_n)_i \in [-1,1]$ if and only if \[ 0 \leq x_{m_i} + \cdots + x_{M_i - 1} \leq 1. \] The last integrand of \eqref{e:integralpolytope} is $1$ if the system of inequalities defining $P_\pi$ in Definition \ref{d:rationalpolytope} is satisfied, and $0$ otherwise. Thus, the last integral of \eqref{e:integralpolytope} calculates the volume of $P_\pi$. \end{proof} \begin{remark} A consequence of the coordinate inequalities $x_i \geq 0$ and those that have the form $0 \leq x_{m_i} + \cdots + x_{M_i - 1} \leq 1$ is that $x_a + \cdots + x_b \leq 1$ for any $a,b \in [m_i, M_i)$. In particular, it is a consequence of Lemma \ref{l:existcoordinate} that $x_i \leq 1$ for all $i \in [n]$, which implies $P_\pi \subseteq [0,1]^n$. \end{remark} \subsection{Type \texorpdfstring{$A$}{A} root system preliminaries} Let $\epsilon_1,\ldots,\epsilon_{n+1}$ be the standard basis of $\R^{n+1}$. Let $(\cdot,\cdot)$ be the standard inner product on $\R^{n+1}$. Let \[ V = \left\{\lambda \in \R^{n+1} : \; \;(\lambda, \epsilon_1 + \cdots + \epsilon_{n+1}) = 0 \right\}. \] The set \[ \Phi = \left\{\epsilon_i - \epsilon_j \in V : \; i,j \in [n+1] \text{ and } i < j\right\} \] is called the \emph{root system of type $A_n$}. The sets \begin{align*} \Phi^+ &= \{\epsilon_i - \epsilon_j \in V : \; i,j \in [n+1] \text{ and } i < j\} \text{ and}\\ \Phi^- &= \{-\lambda \in V : \; \lambda \in \Phi^+\}, \end{align*} respectively, are called the set of \emph{positive roots} and the set of \emph{negative roots}, respectively. \begin{notation*} We often abbreviate $\epsilon_i - \epsilon_j \in \Phi^+$ by $(i,j) \in \Phi^+$. \end{notation*} Let $\alpha_i = \epsilon_i - \epsilon_{i+1}$. Then \[ \Delta = \{ \epsilon_i - \epsilon_{i+1} \in V: \;\; i \in [n]\} = \{ \alpha_i \in V: \;\; i \in [n]\} \] is a basis for $V$. The vectors $\alpha_1,\ldots,\alpha_n$ contained in $\Delta$ are called \emph{simple roots}. There is a dual basis to $\Delta$ consisting of vectors $\omega_1,\ldots,\omega_n$ satisfying $(\omega_i, \alpha_j) = \delta_{ij}$. The dual basis is called the basis of \emph{fundamental coweights}. The \emph{\Weyl} is the group generated by reflections about the hyperplanes orthogonal to the simple roots. Explicitly, the reflection $s_i$ about the hyperplane orthogonal to $\alpha_i$ is given by \begin{equation} \label{e:reflection} s_i (\lambda) = \lambda - (\lambda, \alpha_i) \alpha_i. \end{equation} The map that sends the adjacent transposition $(i \;\; i+1) \in S_{n+1}$ to the reflection $s_i \in A_n$ is called the \emph{geometric representation}. It is a faithful representation of the symmetric group as a Coxeter group. See \cite[Section 4.2]{bjorner-brenti2}, for example. The representation $L$ given in Definition \ref{d:represent} is closely related to the geometric representation of $S_{n+1}$ as the {\Weyl}. \begin{lemma} \label{l:matrixgeometric} The matrix representation of $s_i$ in the basis of simple roots is $I + M$, where $I$ is the identity matrix, and $M$ is the matrix whose only nonzero entries $A$ are given by $M_{i,i-1} = 1$, $M_{i,i} = -2$, and $M_{i,i+1} = 1$. \end{lemma} \begin{proof} This follows directly from \eqref{e:reflection} and appears in the proof of \cite[Proposition 4.2.1]{bjorner-brenti2}. \end{proof} \begin{lemma} \label{l:coxeterrep} Let $\pi \in S_{n+1}$. The matrix representation of $\pi$ in the basis of simple roots is $L_\pi^T$. Consequently, the matrix $L_\pi$ is the matrix representation of $\pi$ in the basis of fundamental coweights. \end{lemma} \begin{proof} Recall from Lemma \ref{l:represent} that the function $L:S_{n+1} \ra GL_n(\R)$ that maps $\pi$ to $L_\pi$ is a representation. Thus, it suffices to check the result for the adjacent transpositions. Let $\pi$ be the adjacent transposition $(i \;\; i+1)$. We may exhaustively check that $L_\pi^T$ is the geometric representation given in Lemma \ref{l:matrixgeometric}. Note that $\pi(j) = j$ and $\pi(j+1) = j + 1$ except for $j \in \{i-1,i,i+1\}$. Thus all rows of $L_\pi$ match the identity matrix except rows $i-1$, $i$, and $i+1$. Since $\pi(i-1) = i-1$, and $\pi(i) = i + 1$, the $(i-1)$-st row of $L_\pi$, if it exists, has a $1$ in columns $i-1$ and $i$ and $0$'s in all other positions. Similarly, if row $i+1$ exists, there is a $1$ in columns $i$ and $i+1$ and $0$'s in all other positions. Since $\pi(i) = i+1$ and $\pi(i) = i$, the only nonzero entry of row $i$ is a $-1$ in column $i$. In summary, we may wite $L_\pi$ as $I + N$, where the only nonzero entries of $N$ are given by $N_{i-1,i} = 1$, $N_{i,i} = -2$, and $N_{i+1,i} = 1$. This is the transpose of the matrix for the geometric representation given in Lemma \ref{l:matrixgeometric}. \end{proof} \subsection{The affine arrangement of type \texorpdfstring{$A_n$}{An} in step coordinates} The definition of the {\typeA} and its connected components involve inner products of the form $(\lambda, \epsilon_i - \epsilon_j)$. Note that $\lambda$, expressed in the basis of fundamental coweights as $x_1 \omega_1 + \cdots + x_n \omega_n$, satisfies \begin{align*} (\lambda, \epsilon_i - \epsilon_j) &= (x_1 \omega_1 + \cdots + x_n \omega_n\; , \; \alpha_i + \cdots + \alpha_{j-1})\\ &= x_i + \cdots + x_{j-1}. \end{align*} The linear isomorphism mapping $x_1 \omega_1 + \cdots + x_n \omega_n$ to $(x_1,\ldots,x_n)$ translates results about the {\typeA} to results about $P_\pi$. We refer to the image $(x_1,\ldots,x_n) \in \R^n$ of this isomorphism as \emph{step coordinates} in reference to the steps of the random walk. Whenever it makes sense, we expand the standard results and definitions about the {\typeA} into the basis $\omega_1,\ldots,\omega_n$ in anticipation of what is needed to calculate the volume of $P_\pi$. \begin{definition} \label{d:consecutiveaffine} Let $(i,j) \in \Phi^+$ and $a \in \Z$. Let \begin{align*} H_{ij}^a &= \{\lambda \in V : (\lambda, \epsilon_i - \epsilon_j) = a\}\\ &= \{x_1 \omega_1 + \cdots + x_n \omega_n \in V : x_i + \cdots + x_{j-1} = a\}. \end{align*} The collection of all hyperplanes of the form $H_{ij}^a$ is called the \emph{\typeA}. The connected components of $V \setminus \cup H_{ij}^a$ are called \emph{alcoves}. The group generated by the set of reflections about hyperplanes of the form $H_{ij}^a$ is the \emph{\affWeyl}. \end{definition} Let $\alc$ be an alcove of the affine walk arrangement. For any $\lambda \in \alc$, and any pair $(i,j) \in \Phi^+$, Definition \ref{d:consecutiveaffine} implies the existence of an integer $k_\alc (i,j)$ such that $(\lambda, \epsilon_i - \epsilon_j)$ is strictly between $k_\alc (i,j)$ and $k_\alc (i,j) + 1$. \begin{definition} \label{d:address} Let $\alc$ be an alcove of the affine walk arrangement. Let $\Phi^+$ be the set of positive roots. The function $k_\alc : \Phi^+ \ra \Z$ such that \begin{align*} \alc &= \left\{ \lambda \in V : \; \; k_\alc (i,j) < (\lambda, \epsilon_i - \epsilon_j) < k_\alc (i,j) + 1 \text{ for all } (i,j) \in \Phi^+ \right\}\\ &= \left\{ x_1 \omega_1 + \cdots + x_n \omega_n \in V: \;\; k_\alc (i,j) < x_i + \cdots + x_{j-1} < k_\alc (i,j) + 1 \text{ for all } (i,j) \in \Phi^+\right\} \end{align*} is called the \emph{address of $\alc$}. The alcove \begin{align*} \alc_{\circ} &= \left\{\lambda \in V : \;\; 0 < (\lambda,\epsilon_i - \epsilon_j) < 1 \textrm{ for all } (i,j) \in \Phi^+ \right\}\\ &= \left\{x_1 \omega_1 + \cdots + x_n\omega_n \in V : \;\; 0 < x_i + \cdots + x_{j-1} < 1 \textrm{ for all } (i,j) \in \Phi^+ \right\} \end{align*} is called the \emph{fundamental alcove}. \end{definition} Thus, the fundamental alcove is the unique alcove whose address is the constant zero function from $\Phi^+$ to $\Z$. \begin{example} \label{ex:alcove} Every point in the unit hypercube that is not in the measure zero union of hyperplanes of the affine walk arrangement lies in some alcove. For example, the point $(0.2, 0.05, 0.8, 0.4, 0.7)$ in step coordinates is in the alcove $\alc$ whose address is shown in Figure \ref{f:alcove}. \end{example} \begin{figure}[htb] \centering \begin{tikzpicture}[scale=.25] \foreach \i in {1,2,3,4,5} { \pgfmathsetmacro{\a}{int(\i + 1)} \foreach \j in {\a,...,6} { \draw (2*\i,2*\j) node{$\i \j$}; } } \draw (14.5,9.5) node{$k_\alc$}; \draw[very thick,->] (13,7) .. controls (14,9.5) and (15,6) .. (17,7); \draw (20, 4) node{$0$}; \draw (20, 6) node{$0$}; \draw (20, 8) node{$1$}; \draw (20, 10) node{$1$}; \draw (20, 12) node{$2$}; \draw (22, 6) node{$0$}; \draw (22, 8) node{$0$}; \draw (22, 10) node{$1$}; \draw (22, 12) node{$1$}; \draw (24, 8) node{$0$}; \draw (24, 10) node{$1$}; \draw (24, 12) node{$1$}; \draw (26, 10) node{$0$}; \draw (26, 12) node{$1$}; \draw (28, 12) node{$0$}; \end{tikzpicture} \caption{The address of the alcove $\alc$ containing the point in Example \ref{ex:alcove}.} \label{f:alcove} \end{figure} The group $\A_n$ has generating set $s_1,\ldots,s_{n+1}$, where $s_1,\ldots,s_n$ are the same generators from $A_n$ that reflect about the hyperplanes $H_{i \; i+1}^0$. The generator $s_{n+1}$ reflects about the hyperplane $H_{1 n}^1$. Thus, the action of $s_i$ is to swap the $i$-th and $(i+1)$-st coordinates of elements of $V$. The action of $s_{n+1}$ is to swap the first and last coordinates, add one to the first coordinate and subtract one from the last coordinate. See \cite[page 86]{shi1} or \cite[Section 4.3]{humphreys}, for example. The first part of the next lemma provides a correspondence between the the group $\A_n$ and the alcoves of the {\typeA}. The last part provides the link to calculating the volume of $P_\pi$. Recall that Lemma \ref{l:coxeterrep} identifies $L_\pi^T$ as the matrix representing $\pi$ in the geometric representation. Also recall from Lemma \ref{l:det} that the determinant of $L_\pi$ is $\pm 1$. \begin{lemma} The following are true about the {\affWeyl}. \label{l:affineprelims} \begin{enumerate}[(i)] \item The {\affWeyl} acts simply transitively on the alcoves of the {\typeA}. \item Every element of $\A_n$ is a product of an element of $A_n$ and a translation. \item Elements of $\A_n$ acting on step coordinates are volume-preserving on $\R^n$ relative to the standard inner product on $\R^n$ and Lebesgue measure. \end{enumerate} \end{lemma} \begin{proof} Part (i) is \cite[Theorem 4.5]{humphreys}. Part (ii) is \cite[Proposition 4.2]{humphreys}. Lemma \ref{l:coxeterrep} shows that elements of $A_n$ expressed as matrices relative to the basis of fundamental coweights have the form $L_\pi$ for some $\pi \in S_{n+1}$. Since translation preserves volume in any basis under any inner product, part (ii) and Lemma \ref{l:det} prove part (iii). \end{proof} \begin{remark} When we convert to coordinates in $\R^n$ via the basis of fundamental coweights, we are calculating volumes and integrals with a standard Lebesgue measure on $\R^n$ equipped with the standard inner product. This is not the same inner product as the one on $V$. To see this difference in inner product visually, compare \cite[Figure 5.1]{martinez} to a standard centrally-symmetric representation of the braid arrangement in the plane. \end{remark} Part (i) of Lemma \ref{l:affineprelims} ensures that $w(\alc_\circ)$ in the next definition is an alcove. \begin{definition} \label{d:alcovelabel} Let $w \in \A_n$. The \emph{alcove of $w$}, denoted $\alc_w$, is the alcove $w(\alc_\circ)$. \end{definition} \subsection{Computing the volume of \texorpdfstring{$P_\pi$}{Pp} by counting alcoves} \begin{lemma} \label{l:polytopeunion} Let $\pi \in S_{n+1}$ and let $P_\pi$ be the polytope of steps for $\pi$. Let $\alc$ be an alcove of the {\typeA} expressed in step coordinates. Then $\alc \subseteq P_\pi$ or $\alc \cap P_\pi = \emptyset$. \end{lemma} \begin{proof} The address $k_\alc: \Phi^+ \ra \Z$ for $\alc$ determines a system of inequalities where each inequality has the form $k_\alc(i,j) < x_i + \cdots + x_{j-1} < k_\alc(i,j) + 1$, for each $(i,j) \in \Phi^+$. This includes the pairs $(m_i, M_i - 1) \in \Phi^+$ in Definition \ref{d:rationalpolytope}. If $k_\alc(m_i,M_i - 1) = 0$ for all $i \in [n]$, then every $\bm{x} \in \alc$ satisfies all the inequalities that define $P_\pi$, which implies $\alc \subseteq P_\pi$. Otherwise, the sum of coordinates $x_{m_i} + \cdots + x_{M_i - 1}$ is incompatible with $P_\pi$ for some $i \in [n]$, which implies $\alc \cap P_\pi = \emptyset$. \end{proof} \begin{lemma} \label{l:affinealcovecount} Let \begin{align*} P &= \left\{ \lambda \in V : \;\; 0 < (\lambda, \alpha_i) < 1 \text{ for all } i \in [n]\right\}\\ &= \left\{ x_1 \omega_1 + \cdots + x_n \omega_n : \;\; 0 < x_i < 1 \text{ for all } i \in [n]\right\}. \end{align*} In step coordinates, the parallelepiped $P$ is the unit cube $[0,1]^n$. There are $n!$ alcoves of the {\typeA} contained in $P$. \end{lemma} \begin{proof} See the proof of \cite[Theorem 4.9]{humphreys} or \cite[Section 3]{lam-postnikov2}. \end{proof} \begin{corollary} \label{c:polytopevolume} In step coordinates, each alcove of the {\typeA} has volume $1/n!$. Thus, \[ \PRB(f,\pi) = \frac{K_\pi}{2^n n!}, \] where $K_\pi$ is the number of alcoves contained in $P_\pi$. \end{corollary} \begin{proof} The set of points not in any alcove has measure zero. Thus part (iii) of Lemma \ref{l:affineprelims} and Lemma \ref{l:affinealcovecount} show that alcoves have volume $1/n!$ in step coordinates. The result then follows from Lemma \ref{l:polytopeunion} and Lemma \ref{l:volumepolytope}. \end{proof} Not every function from $\Phi^+$ to $\Z$ is the address of an alcove. A characterization of such functions is given by Shi's Theorem. See \cite[Lemma 6.1.3]{shi1} or \cite[Theorem 5.2]{shi2}. \begin{theorem} \label{t:shi} (Shi's Theorem) A function $k:\Phi^+ \ra \Z$ is the address of an alcove if and only if \[ k(i,t) + k(t,j) \leq k(i,j) \leq k(i,t) + k(t,j) + 1 \] for all $i,t,j$ satisfying $1 \leq i < t < j \leq n + 1$. \end{theorem} Shi's Theorem and Corollary \ref{c:polytopevolume} provide a straightforward, though inefficient, method for computing $\PRB(f,\pi)$. This method, and an alternative one based on \cite{stanley}, is given in \cite{denoncourt}. \begin{proposition} \label{p:computeuniform} Let $\pi \in S_{n+1}$. Let $f$ be the uniform density function on $[-1,1]$. Let $K_\pi$ denote the number of functions $k:\Phi^+ \ra \N$ satisfying the inequalities \[ k(i,t) + k(t,j) \leq k(i,j) \leq k(i,t) + k(t,j) + 1, \] where $i < t < j$, and also satisfying the equalities $k(i,j) = 0$ whenever there exists $c$ such that $\pi(c) \leq i < j \leq \pi(c+1)$ or $\pi(c+1) \leq i < j \leq \pi(c)$. Then \[ \PRB(f,\pi) = \frac{K_\pi}{2^n n!}. \] \end{proposition} \subsection{A characterization of the weak order in terms of alcove addresses} \label{ss:weakordersection} Recall that $\A_n$ is generated by reflections $s_1,\ldots,s_{n+1}$. The \emph{length of $w$}, denoted $\ell(w)$, is the smallest number of generators in an expression of $w$ as a product of generators. Define a relation $\ra$ by the condition $w \ra ws$ if $s$ is a generator and $\ell(ws) > \ell(w)$. The \emph{weak order} on $\A_n$ is defined as the transitive closure of the relation $\ra$. The main result of this section, Lemma \ref{l:weakcharacter1}, characterizes the weak order on $\A_n$ in terms of alcove addresses. It might be folklore or known. There is an indirect way to prove the lemma by combining \cite[Theorem 4.1]{shi3} with \cite[Theorem 5.3]{bjorner-brenti2}. The approach given below uses a geometric characterization of the weak order on $\A_n$ given in \cite{humphreys}. For a given hyperplane $H_{ij}^a$ of the {\typeA}, two sides of the hyperplane are determined by the conditions $(\lambda, \epsilon_i - \epsilon_j) > a$ and $(\lambda, \epsilon_i - \epsilon_j) < a$. We say a hyperplane $H$ \emph{separates} $\alc$ from $\alc_\circ$ if $\alc$ and $\alc_\circ$ lie on two sides of $H$. Based on the conditions for determining sides, we determine whether $H_{ij}^a$ separates $\alc$ and $\alc_\circ$ from the the address of $\alc$. \begin{lemma} \label{l:separatingaddress} Let $H_{ij}^a$ be a hyperplane in the {\typeA}, let $\alc_\circ$ denote the fundamental alcove, and let $\alc$ be an arbitrary alcove. If $a > 0$, then $H_{ij}^a$ separates $\alc$ from $\alc_\circ$ if and only if $k_\alc(i,j) \geq a$. If $a \leq 0$, then $H_{ij}^a$ separates $\alc$ from $\alc_\circ$ if and only if $k_\alc(i,j) \leq a - 1$. \end{lemma} \begin{proof} Suppose $a > 0$. Since $k_{\alc_\circ}(i,j) = 0$, we have $\alc_\circ$ on the side of $H_{ij}^a$ where $(\lambda,\epsilon_i - \epsilon_j) < a$. Note that $\alc$ is on the side where $(\lambda, \epsilon_i - \epsilon_j) > a$ if and only $k_\alc(i,j) \geq a$. Thus $H_{ij}^a$ separates $\alc_\circ$ and $\alc$ if and only if $k_\alc(i,j) \geq a$. The argument for $a < 0$ is similar. \end{proof} \begin{lemma} \label{l:separatingweak} Let $\mathcal{L}(w)$ be the set of hyperplanes separating $\alc_w$ from $\alc_\circ$. Then $u \leq w$ in the weak order if and only if $\mathcal{L}(u) \subseteq \mathcal{L}(w)$. \end{lemma} \begin{proof} This is \cite[Theorem 4.5]{humphreys}. \end{proof} Let $k:\Phi^+ \ra \Z$ and $k':\Phi^+ \ra \Z$ be addresses. We write $k' \leq_A k$ if $k'(i,j) \leq k(i,j)$ whenever both are nonnegative or $k'(i,j) \geq k(i,j)$ whenever both are nonpositive. We write $k' \leq k$ if $k'(i,j) \leq k(i,j)$ for all $(i,j) \in \Phi^+$, which is the standard notation for function comparison. \begin{lemma} \label{l:weakcharacter1} Let $u,w \in \A_n$. Then $u \leq w$ if and only if $k_u \leq_A k_w$. \end{lemma} \begin{proof} The result follows from Lemma \ref{l:separatingaddress} and Lemma \ref{l:separatingweak}. \end{proof} The addresses of alcoves in $P_\pi$ are all greater than equal to $0$. Thus, we simplify the previous lemma to characterize weak order as a comparison of addresses as functions. \begin{corollary} \label{c:weakcharacter2} Let $u,w \in \A_n$. Suppose $k_u(i,j) \geq 0$ and $k_w(i,j) \geq 0$ for all $(i,j) \in \Phi^+$. Then $u \leq w$ if and only if $k_u \leq k_w$. \end{corollary} \subsection{Ideals in the root poset determine the alcoves in \texorpdfstring{$P_\pi$}{Pp}} \label{ss:rootposet} If we set $k(i,j) = 0$ whenever required by Proposition \ref{p:computeuniform}, and greedily set $k(i,j)$ to the maximum amount allowed by Shi's theorem, then we obtain a maximal address satisfying the system of linear inequalities defining the polytope $P_\pi$. By Corollary \ref{c:weakcharacter2}, if this turns out to be a unique maximum address satisfying the system, then the alcoves in $P_\pi$ correspond to a weak order interval of $\A_n$. We use a construction due to Sommers \cite{sommers} to show that this is the case. There is a standard order $\leq$ on $\Phi^+$, called \emph{the root poset}, such that $(i',j') \leq (i,j)$ if and only if $i \leq i' < j' \leq j$, which is equivalent to $[i',j'] \subseteq [i,j]$. Recall that an ideal is a down-closed subset of a poset. \begin{definition} \label{d:ideal} Let $\pi \in S_{n+1}$. For $i \in [n]$, we say $(\pi(i),\pi(i+1))$ is a \emph{consecutive root for $\pi$} if $\pi(i) < \pi(i+1)$. Similarly, if $\pi(i+1) < \pi(i)$, we say $(\pi(i+1),\pi(i))$ is a \emph{consecutive root for $\pi$}. Denote the collection of consecutive roots for $\pi$ by $C_\pi$. Define the \emph{root ideal of $\pi$}, denoted $\IDEAL_\pi$, by \begin{equation*} \IDEAL_\pi := \{(i',j') : (i',j') \leq (i,j) \text{ for some } (i,j) \in C_\pi\} \end{equation*} \end{definition} The motivation for defining $\IDEAL_\pi$ comes from the next lemma, which states that the address of any alcove in $P_\pi$ is $0$ on the ideal $I_\pi$. \begin{lemma} \label{l:ideal} Let $k_\alc:\Phi^+ \ra \Z$ be the address of an alcove $\alc$ in the polytope $P_\pi$ of steps for $\pi$. For any $(i',j') \in \IDEAL_\pi$ and any $(i,j) \in C_\pi$ such that $k(i,j) = 0$, we have $k(i',j') = 0$. \end{lemma} \begin{proof} Given that $(i,j) \in C_\pi$, we know $x_i + \cdots + x_{j-1} \leq 1$. Thus, if $i' > i$ and $j' < j$, we know $x_{i'} + \cdots + x_{j' - 1} \leq 1$. It follows that $k(i',j') = 0$. \end{proof} In the next definition, it is more convenient to regard elements of $\Phi^+$ as vectors, rather than using our abbreviation as pairs $(i,j)$ of integers. \begin{definition} \label{d:rootideal} For a fixed root $\alpha \in \Phi^+$ and a fixed ideal $\IDEAL$ of $\Phi^+$, let $\alpha_{\IDEAL}$ be defined by \[ \alpha_{\IDEAL} = \text{min} \; \left\{k : \sum_{i = 1}^{k + 1} \gamma_i = \alpha \text{ with } \gamma_i \in \IDEAL \right\}. \] \end{definition} In other words, the smallest number of joins needed to express $\alpha$ as a join in the root poset using only elements of $\IDEAL$ is $\alpha_{\IDEAL} + 1$. The value of $\alpha_{\IDEAL}$ is zero for any element of $\IDEAL$. As in Section \ref{ss:weakordersection}, we write $k' \leq k$ if $k'(i,j) \leq k(i,j)$ for all $(i,j) \in \Phi^+$ for addresses that are always nonnegative. The next lemma is a dual version of \cite[Theorem 2]{armstrong}. \begin{lemma} \label{l:sommers} (Sommers \cite[Section 5]{sommers}) For any ideal $\IDEAL$ of $\Phi^+$ that contains all the simple roots, there exists a unique maximum address $k_\IDEAL:\Phi^+ \ra \Z$ such that $k_\IDEAL(i,j) = 0$ for all $(i,j) \in \IDEAL$. It is defined by \[ k_\IDEAL(i,j) = (i,j)_{\IDEAL}. \] \end{lemma} \begin{proof} In the proof of \cite[Lemma 5.1 part (2)]{sommers}, it is shown that $k(i,j) \leq (i,j)_\IDEAL$ for any address satisfying $k(i,j) = 0$ for all $(a,b) \in \IDEAL$. In \cite[Lemma 5.2 part (2)]{sommers}, it is shown that there exists an address $k$ such that $k(i,j) = (i,j)_\IDEAL$ for all $(i,j) \in \Phi^+$. Since any address $k'$ satisfying $k'(i,j) = 0$ for all $(i,j) \in \IDEAL$ must also satisfy $k'(i,j) \leq (i,j)_\IDEAL = k(i,j)$, it follows that $k$ is the unique maximum address such that $k(i,j)$ is zero on $\IDEAL$. \end{proof} \begin{example} \label{ex:ideal} The alcove address of Figure \ref{f:alcove} has $k(i,j) = 0$ for any $(i,j)$ where $j = i + 1$ as well as $(1,3)$ and $(2,4)$. The maximum alcove guaranteed by Lemma \ref{l:sommers} is obtained by filling the entries with the maximum possible value that the conditions of Shi's theorem allows. The address of this alcove is given in Figure \ref{f:maximumalcove}. Its values are, as expected, larger than those of Figure \ref{f:alcove}. \end{example} \begin{figure}[htb] \centering \begin{tikzpicture}[scale=.25] \foreach \i in {1,2,3,4,5} { \pgfmathsetmacro{\a}{int(\i + 1)} \foreach \j in {\a,...,6} { \draw (2*\i,2*\j) node{$\i \j$}; } } \draw (14.5,9.5) node{$k_\IDEAL$}; \draw[very thick,->] (13,7) .. controls (14,9.5) and (15,6) .. (17,7); \draw (20, 4) node{$0$}; \draw (20, 6) node{$0$}; \draw (20, 8) node{$1$}; \draw (20, 10) node{$1$}; \draw (20, 12) node{$2$}; \draw (22, 6) node{$0$}; \draw (22, 8) node{$0$}; \draw (22, 10) node{$1$}; \draw (22, 12) node{$2$}; \draw (24, 8) node{$0$}; \draw (24, 10) node{$1$}; \draw (24, 12) node{$2$}; \draw (26, 10) node{$0$}; \draw (26, 12) node{$1$}; \draw (28, 12) node{$0$}; \end{tikzpicture} \caption{The maximum address that is zero on the ideal $\IDEAL$ of Example \ref{ex:ideal}.} \label{f:maximumalcove} \end{figure} To apply Lemma \ref{l:sommers} to $\IDEAL_\pi$ requires that $\IDEAL_\pi$ contain all the simple roots. \begin{lemma} \label{l:existcoordinate} Let $\pi \in S_{n+1}$. Let $m_i = \mathrm{min}\{\pi(i),\pi(i+1)\}$ and $M_i = \mathrm{max}\{\pi(i),\pi(i+1)\}$. For any $j \in [n]$, there exists $i \in [n]$ such that $m_i \leq j < M_i$. Thus every $(j,j+1) \in \Delta$ is in $\IDEAL_\pi$. \end{lemma} \begin{proof} Suppose otherwise. Let $k$ be such that $\pi(k) = j$. Then both $\pi(k-1)$ and $\pi(k+1)$, if defined, must be greater than $j$. If there exists an index $b$ such that $\pi(b) > j$ and $\pi(b+1) < j$, or vice versa, then $i = b$ is such that $m_i \leq j < M_i$. Thus, to the left of $k - 1$ and to the right of $k + 1$, the values must stay above $j$. Since $\pi$ is a permutation, this implies $j = 1$. However, one of $\pi(k-1)$ or $\pi(k+1)$ is defined, and $\pi(k)$ is the minimum of the two values, which implies $m_i \leq j < M_i$ for either $i = k - 1$ or $i = k$. \end{proof} \begin{corollary} Let $f = \frac{1}{2}\Chi_{[-1,1]}$ be the uniform density function on $[-1,1]$. Let $\tau,\pi \in S_{n+1}$. Suppose $\IDEAL_\pi \subseteq \IDEAL_\tau$. Then \[ \PRB(f,\pi) \geq \PRB(f,\tau). \] \end{corollary} \begin{proof} Lemma \ref{l:existcoordinate} implies that $\IDEAL_\pi$ and $\IDEAL_\tau$ contain all the simple roots. The hypothesis $\IDEAL_\pi \subseteq \IDEAL_\tau$ implies that $k_{\IDEAL_\tau}(i,j) = 0$ whenever $k_{\IDEAL_\pi}(i,j) = 0$. Lemma \ref{l:sommers} implies that $k_{\IDEAL_\tau}(i,j) \leq k_{\IDEAL_\pi}(i,j)$ for all $(i,j) \in \Phi^+$. The result then follows from Corollary \ref{c:weakcharacter2}. \end{proof} \begin{theorem} \label{t:weakordertheorem} Let $f = \frac{1}{2}\Chi_{[-1,1]}$ be the uniform density function on $[-1,1]$. Let $\pi \in S_{n+1}$ and let $\IDEAL_\pi$ be the root ideal of $\pi$. Let the address of $w \in \A_n$ be given by $k_w(i,j) = (i,j)_{\IDEAL_\pi}$ for all $(i,j) \in \Phi^+$. Then, \[ \PRB(f,\pi) = \frac{|[1,w]|}{2^n n!}, \] where $[1,w]$ consists of all $v \in \A_n$ such that $v \leq w$ in the weak order on $\A_n$. \end{theorem} \begin{proof} Lemma \ref{l:existcoordinate} implies that we may apply Lemma \ref{l:sommers} to $\IDEAL_\pi$. The result then follows from Lemma \ref{l:sommers} and Corollary \ref{c:weakcharacter2}. \end{proof} Weak order intervals of $\A_n$ satisfy condition $(1)$ of \cite[Proposition 3.5]{lam-postnikov2}, which implies $P_\pi$ is an \emph{alcoved polytope} in the sense of \cite{lam-postnikov1} and \cite{lam-postnikov2}. Thus \cite[Theorem 3.2]{lam-postnikov1} provides yet another computational approach to calculating the volume of $P_\pi$, although we do not pursue that approach in this paper. The next proposition is somewhat surprising, in the sense that two consecutive entries of $\pi$ can completely determine $\PRB(f,\pi)$. This is not the case for the Laplace or normal density functions, and it is reasonable to suspect that a typical density function does not exhibit this property. \begin{proposition} \label{p:minimumprob} Let $f = \frac{1}{2}\Chi_{[-1,1]}$ be the uniform density function on $[-1,1]$. Let $\pi \in S_{n+1}$. Then $1 (n + 1)$ or $(n + 1) 1$ occur in consecutive positions in the $1$-line notation for $\pi$ if and only if \[ \PRB(f,\pi) = \frac{1}{2^n n!} \] \end{proposition} \begin{proof} If $1$ and $n + 1$ are consecutive in the $1$-line notation, then $\IDEAL_\pi$ is all of $\Phi^+$. Thus $w = 1$ in Theorem \ref{t:weakordertheorem}, which implies there is only one element of $\A_n$ in the interval $[1,w]$. Conversely, if there are no consecutive occurrences of $1$ and $(n+1)$, then the ideal $\IDEAL_\pi$ does not contain $(1,n)$. Thus $(1,n)$ is the join of at least $2$ elements of $\IDEAL_\pi$, which implies $k_{\IDEAL_\pi}(1,n) > 0$. This implies $[1,w]$ contains more than one element. \end{proof} \section{Pattern probability comparisons for the normal distribution} \label{s:normaldistribution} \begin{figure}[htb] \label{f:edgediagram2} \centering \begin{tikzpicture}[scale=.5] \foreach \x in {1,2,...,6} \draw[dotted] (0,\x)--(9,\x); \draw(15.5,3.5) node{$\implies \mathrm{lev}(\pi) = \begin{bmatrix} 2 & 2 & 1 & 1 & 0\\ 0 & 4 & 3 & 2 & 1\\ 0 & 0 & 3 & 2 & 1\\ 0 & 0 & 0 & 2 & 1\\0 & 0 & 0 & 0 & 2 \end{bmatrix}$}; \draw[thick,*->-*](4.5,3) node[left]{3}--(4.5,1); \draw[thick,*->-*](5.5,1) node[left]{1}--(5.5,5); \draw[thick,*->-*](6.5,5) node[left]{5}--(6.5,6); \draw[thick,*->-*](7.5,6) node[left]{6}--(7.5,2); \draw[thick,*->-*](8.5,2) node[left]{2}--(8.5,4) node[right]{4}; \draw(0.5,7) node{Levels}; \draw(6.0,8) node{Edge diagram}; \draw(6.0,7) node{for $\pi = 315624$}; {\foreach \x in {1,2,...,5} \draw (0.5,0.5+\x) node{$\ubar{\x}$}; } \end{tikzpicture} \caption{The edge diagram for a permutation and its level count matrix. Figure adapted from \cite{elizalde-martinez}.} \end{figure} As Zare \cite{zare} suggested, when $f$ is a normal distribution, we calculate $\PRB(f,\pi)$ by finding the volume of a spherical simplex. General equations exist to compute such volumes. See \cite{aomoto} or \cite{ribando}, for example. However, they appear to be computationally intensive, as is Lemma \ref{l:general} when it is applied to the normal distribution. Nonetheless, there are a few direct comparisons we can make involving alcoves and levels of the edge diagram. Recall that the alcoves of Section \ref{s:affineA} are simplices of volume $1/n!$, by Corollary \ref{c:polytopevolume}. For a given origin-centered ball $B$ in $\R^n$, we obtain an underestimate for $\PRB(f,\pi)$ by counting all alcoves in $D_\pi$ that are fully contained in $B$. Similarly, we obtain an overestimate for $\PRB(f,\pi)$ by counting all alcoves in $D_\pi$ that intersect $B$ or are fully contained in $B$. The address of an alcove and the radius of the ball suffice to determine whether an alcove is fully contained in $B$ or intersects $B$ or is disjoint from $B$. \begin{proposition} \label{p:bounds} Let $B$ be an origin-centered ball in $\R^n$. Let $m_\pi$ be the number of alcoves fully contained in $D_\pi$ and $B$. Let $M_\pi$ be the number of alcoves fully contained in $D_\pi$ that have nonempty intersection with $B$. Then \[ \frac{m_\pi}{n! \text{ volume}(B)} \leq \PRB(f,\pi) \leq \frac{M_\pi}{n! \text{ volume}(B)}. \] \end{proposition} Note that hypercubes with integer-valued vertices could be used instead of alcoves, but one would need to determine whether the hypercube is fully contained in $D_\pi$ or intersects $D_\pi$ or is disjoint from $D_\pi$. For alcoves, this is directly determined from the alcove's address. For $\pi \in S_{n+1}$, we defined $\mathrm{lev}(\pi)$ on $[n]$ to measure how often a value lies between two consecutive values of $\pi$. We extend the definition of $\mathrm{lev}(\pi)$ to arbitrary pairs of $[n]$. \begin{definition} \label{d:levelmatrix} Let $\pi \in S_{n+1}$. Denote the number of positions $k$ such that $\pi(k) \leq i,j < \pi(k+1)$ or $\pi(k+1) \leq i,j < \pi(k)$ by $\mathrm{lev}(\pi)_{i,j}$. Note that $\mathrm{lev}(\pi)_{i,i}$ is the same as $\mathrm{lev}(\pi)_i$ defined in Definition \ref{d:levelcount}. \end{definition} The measure of the spherical simplex that determines $\PRB(f,\pi)$ for the normal distribution is completely determined by the values of $\mathrm{lev}(\pi)$, as will be seen in the proof of Theorem \ref{t:comparegaussian}. Although such measures may be difficult to calculate, we can sometimes use $\mathrm{lev}(\pi)$ and $\mathrm{lev}(\tau)$ to compare $\PRB(f,\pi)$ and $\PRB(f,\tau)$. Recall that Lemma \ref{l:general} expresses $\PRB(f,\pi)$ as $\int_{D_\pi} g(L_\pi(\bm{x})) \D \bm{x}$, where $g$ is the joint density function defined by $g(x_1,\ldots,x_n) = f(x_1)f(x_2)\cdots f(x_n)$. \begin{theorem} \label{t:comparegaussian} Let $f:\R \ra \R$ be the density function for a normal distribution with mean zero and any variance. Let $\pi, \tau \in S_{n+1}$ and suppose $\mathrm{lev}(\pi)_{i,j} \leq \mathrm{lev}(\tau)_{i,j}$ for all $(i,j) \in \Phi^+$. Then $\PRB(f,\pi) \geq \PRB(f,\tau)$. \end{theorem} \begin{proof} By scale invariance (Lemma \ref{l:scale-invariance}), we may assume $f$ is given by $f(x) = Ke^{-x^2}$ for some $K \in \R$. Every factor in Lemma \ref{l:general} has the form $f(\pm(x_a + \cdots + x_b))$. We have \begin{align*} f(\pm(x_a + \cdots + x_b)) &= \EXP{-(x_a + \cdots + x_b)^2}\\ &= \EXP{-\SUM_{k=a}^b x_k^2} \; \cdot \; \EXP{-\sum 2 x_k x_{k'}}, \end{align*} where the sum in the second exponential is over all pairs between $a$ and $b$ (inclusive). By Definition \ref{d:levelmatrix}, there are $\mathrm{lev}(\pi)_{j,j}$ factors contributing one term of the form $-x_j^2$ and $\mathrm{lev}(\pi)_{i,j}$ factors contributing one term of the form $-2x_i x_j$ to the overall product of exponentials. Thus, if $\mathrm{lev}(\pi)_{i,j} \leq \mathrm{lev}(\tau)_{i,j}$ for all $(i,j)$, the integrand in Lemma \ref{l:general} for $\pi$ is always at least as large as the integrand for $\tau$. \end{proof} In \cite[Lemma 2.3]{elizalde-martinez}, Elnitsky and Martinez showed that if $L_\pi$ can be obtained from $L_\tau$ by a permutation of rows and columns, then $\PRB(f,\pi) = \PRB(f,\tau)$ for any choice of density function $f$, symmetric or otherwise. By including their guaranteed equalities, we obtain more comparable pairs of permutations than what is guaranteed by Theorem \ref{t:comparegaussian}. Write $\pi \equiv \tau$ if $L_\pi$ can be obtained from $L_\tau$ by a permutation of rows and columns. Write $\pi \leq_{\mathrm{lev}} \tau$ if $\mathrm{lev}(\pi)_{i,j} \leq \mathrm{lev}(\tau)_{i,j}$ for all $i,j \in [n]$. Define $\leqslant$ as the transitive closure of of the relation $\equiv$ and the partial order $\leq_{\mathrm{lev}}$. We then have a broader collection of comparable pairs of permutations for the normal distribution. \begin{corollary} \label{c:comparegaussian} Let $f$ be a normal distribution with mean zero. Let $\pi, \tau \in S_{n+1}$. If $\pi \leqslant \tau$, then $\PRB(f,\pi) \geq \PRB(f,\tau)$. \end{corollary} \section{Concluding remarks and problems} \label{s:postscript} In \cite{zare}, Zare asks which permutations occur most frequently in random walks with a normal or uniform distribution of mean zero for its steps. Our results provide an imprecise heuristic: permutations with large consecutive changes in its $1$-line notation are less likely to occur than permutations with small consecutive changes. In other words, for permutations where $\mathrm{lev}(\pi)_{i,j}$ is large, we expect $\PRB(f,\pi)$ to be small, and vice versa, for a large class of symmetric density functions of a continuous probability distribution. As a general problem, we would like to know what general hypotheses are needed to prove $\PRB(f,\pi) \geq \PRB(f,\tau)$ whenever $\mathrm{lev}(\pi)_{i,j} \leq \mathrm{lev}(\tau)_{i,j}$ for all $(i,j) \in \Phi^+$. However, this question is probably too open-ended. We have evidence for the following more precise conjecture. \\ \\ \textbf{Conjecture:} Let $f:\R \ra \R$ be a density function that is log-concave on $\R_{>0}$ and symmetric on $\R$. Let $\pi,\tau \in S_{n+1}$ and suppose $\mathrm{lev}(\pi)_{i,j} \leq \mathrm{lev}(\tau)_{i,j}$ for all $(i,j) \in \Phi^+$. Then $\PRB(f,\pi) \geq \PRB(f,\tau)$. \\ \\ From the perspective of computation, Proposition \ref{p:computeuniform} provides a direct approach to computing ordinal pattern probabilities when the steps are uniform. Theorem \ref{t:weakordertheorem} reduces the problem to finding the size of a weak order interval in $\A_n$. A lot is known about these intervals, which is enough to make the computation easier in some cases. For example, in \cite{lapointe-morse}, Lapointe and Morse show that the weak order on the quotient $\PB{k+1}$ is order-isomorphic to the $k$-Young lattice. Furthermore, some intervals of the $k$-Young lattice are intervals of the Young lattice. The size of intervals of the Young lattice is given by a classical determinant formula due to Kreweras, thus providing an alternative calculation to Proposition \ref{p:computeuniform} for some permutations. (See \cite[Section 2.3.7]{kreweras}.) However, the affine symmetric group contains many weak order intervals isomorphic to weak order intervals of the symmetric group. By \cite[Theorem 1.4]{dittmer-pak}, computing the size of weak order intervals in $S_n$ is $\#P$-complete. Unless there is something special about the weak order intervals in Theorem \ref{t:weakordertheorem}, computing $\PRB(f,\pi)$ is hard when $f$ is uniform. \\ \\ \textbf{Conjecture}: Computing $\PRB(f,\pi)$ for the uniform density function $f$ on $[-1,1]$ and arbitrary $\pi \in S_n$ is $\#P$-complete. \begin{center} ACKNOWLEDGEMENTS \end{center} The author would like to thank Jim and Kate Daly and Emily Pavey for their support. The author also thanks Dana Ernst, Michael Falk, and Jim Swift of NAU's Algebra, Combinatorics, Geometry, and Topology Seminar for their comments on a talk based on an earlier version of this paper.

https://arxiv.org/abs/2012.02824

High order steady-state diffusion approximations

We derive and analyze new diffusion approximations of stationary distributions of Markov chains that are based on second- and higher-order terms in the expansion of the Markov chain generator. Our approximations achieve a higher degree of accuracy compared to diffusion approximations widely used for the past fifty years, while retaining a similar computational complexity. To support our approximations, we present a combination of theoretical and numerical results across three different models. Our approximations are derived recursively through Stein/Poisson equations, and the theoretical results are proved using Stein's method.

\section{Introduction.}\label{intro} \section{Introduction} \label{fse1} We propose a new class of approximations for stationary distributions of Markov chains. The new approximations will be numerically demonstrated to be accurate in three models: the $M/M/n$ queue known as the Erlang-C model, the hospital model proposed in \cite{DaiShi2017}, and the autoregressive (AR(1)) model studied in \cite{BlanGlyn2018}. In addition to numerical results, for the Erlang-C model we provide theoretical guarantees that our approximation achieves higher-order accuracy. Consider a one-dimensional, positive-recurrent, discrete-time Markov chain (DTMC) $X=\{X(n), n\geq 0\}$ taking values on a subset of $\mathbb{R}$. We introduce our approach in the DTMC setting, but continuous-time Markov chains (CTMC) can be treated analogously; see Section~\ref{fse3} where we treat the Erlang-C model. Call $\mathbb{E} \big( X(1) - X(0) | X(0) = x \big)$ the drift of the DTMC. We center and scale our DTMC by defining $\tilde X=\{\tilde X(n), n\geq 0\}$, where $\tilde X(n)=\delta(X(n)-R)$ for some constants $\delta>0$ and $R\in \mathbb{R}$. We typically take $R$ to be the point where the drift of $X$ equals zero, which also happens to be the equilibrium of the corresponding fluid model; c.f., \cite{Stol2015} or \cite{Ying2016}. The scaling parameter $\delta$ is related to stochastic fluctuations around $R$. Let $\tilde X(0)$ have the stationary distribution of $\tilde X$, let $W=\tilde X(0)$, and let $W'=\tilde X(1)$. Stationarity implies that \ben{\label{f0} \mathbb{E} f(W')-\mathbb{E} f(W)=0 } for all $f:\mathbb{R}\to \mathbb{R}$ such that the expectations exist. \blue{Setting $\Delta = W' - W$, for sufficiently smooth $f(x)$ we can expand} the left-hand side to get \begin{align} 0 =&\ \mathbb{E} f(W') - \mathbb{E} f(W) = \mathbb{E} \bigg[\sum_{i=1}^{n} \frac{1}{i!} \Delta^i f^{(i)}(W) + \frac{1}{(n+1)!} \Delta^{n+1} f^{(n+1)}(\xi)\bigg], \quad n \geq 0, \label{eq:taylorgeneric} \end{align} where $\xi = \xi^{(n)}$ lies between $W$ and $W'$. Note that $\Delta$ equals our scaling term $\delta$ multiplied by the displacement of the unscaled DTMC. Informally, the DTMCs we consider are those where the moments of the (unscaled) displacement are bounded by a constant independent of $\delta$, while $\delta$ itself is close to zero. In this setting, the right-hand side of \eqref{eq:taylorgeneric} is governed by its lower-order terms when $\delta$ is small. This motivates our approximations of $W$. Letting $\mathcal{W}$ be the state space of $\tilde{X}$, for each $x \in \mathcal{W}$ let $ b(x)= \mathbb{E}(\Delta |W=x) $ be the drift of the DTMC at state $x$. Let $(\underline{w}, \overline{w})$ be the smallest interval containing $\mathcal{W}$, and assume $b(x)$ is extended to be defined on all of $(\underline{w}, \overline{w})$. The precise form of the extension is unimportant for the time being and we will see in our examples that this extension often has a natural form. We approximate $W$ by a continuous random variable $Y \in (\underline{w}, \overline{w})$ with density \ben{\label{f10} \frac{\kappa}{v(x)} \exp\Big(\int_0^x \frac{b(y)}{v(y)} dy \Big),\quad x\in (\underline{w}, \overline{w}), } where $\kappa$ is the normalizing constant and $v(x):(\underline{w}, \overline{w})\to\mathbb{R}_+$ is some function to be specified. We note that the distribution of $Y$ is determined for a given \emph{fixed} set of system parameters of the Markov chain. In particular, $Y$ is well defined even when no limit is studied, so the stationary distribution of the unscaled DTMC $X$ would then be approximated by $Y/\delta + R$. To discuss how to choose $v(x)$, suppose $(\underline{w}, \overline{w})=\mathbb{R}$ and consider the diffusion process $\{Y(t), t\ge 0\}$ given by \begin{align} Y(t) = Y(0)+ \int_0^t b(Y(s))ds + \int_0^t \sqrt{2 v(Y(s))} dB(s), \label{eq:sde} \end{align} where $\{B(t), t\ge 0\}$ is the standard Brownian motion. Under mild regularity conditions on $b(x)$ and $v(x)$, the above diffusion process is well defined and has a unique stationary distribution whose density is given by (\ref{f10}); \blue{for a proof, see Chapter 15.5 of \cite{KarlTayl1981}}. Furthermore, the stationary density in (\ref{f10}) is characterized by \begin{align}\label{eq:bar} \mathbb{E} b(Y)f'(Y)+ \mathbb{E} v(Y)f''(Y)=0 \quad \text{ for all suitable } f:\mathbb{R}\to \mathbb{R}. \end{align} When one or both of the endpoints of $(\underline{w}, \overline{w})$ are finite, we would account for this by adding suitable boundary reflection terms. In this paper we think of $Y$ as being a \emph{diffusion approximation} of $W$. The characterization equation (\ref{eq:bar}) is well known for Markov processes; c.f., \cite{EthiKurt1986}. A related version is called the \emph{basic adjoint relationship} in the context of multidimensional reflecting Brownian motions by \cite{HarrWill1987}. Equation (\ref{eq:bar}) is known in the Stein research community as the \emph{Stein equation}; see, for example, \cite{ChenGoldShao2011}. Ensuring that $Y$ is a good approximation of $W$ requires a careful choice of $v(x)$. If we consider \eqref{eq:taylorgeneric} with $n = 2$ and ignore the third-order error term, then a natural choice is to use $v(x)=v_1(x)$, where $v_1(x)$ is an extension of $\frac{1}{2}\mathbb{E}(\Delta^2|W=x)$ to all of $(\underline{w}, \overline{w})$. Choosing a diffusion approximation in such a way was done in \cite{MandMassReim1998} and \cite{ WardGlyn2003}, as well as more recently in \cite{DaiShi2017}. Despite this natural choice, most of the literature in the last fifty years did not use $v_1 $ to develop diffusion approximations. Instead, the typical choice is $v(x) = v_0$, where \begin{align} \label{eq:v0} v_0= v_1(0) = \frac{1}{2}\mathbb{E}(\Delta^2|W=0); \end{align} i.e., $v_0$ is $v_1(x)$ evaluated at the fluid equilibrium $x = 0$. For examples, see \cite{HalfWhit1981,HarrNguy1993,Gurv2014, Ward2012}. It is usually the case that $v_0$ and $v_1(W)$ are asymptotically close, so using $v_0$ is enough to prove a limit theorem, which is the focus of most of the diffusion approximation literature. We however, show that using $v_0$ instead of $v_1$ can lead to significant excess error. \blue{One such case is the Erlang-C model. It was shown in \cite{BravDaiFeng2016} that for a large class of performance measures, the $v_0$ approximation error is at most $C/\sqrt{R}$, where $R$ is a parameter known as the offered load and $C>0$ is a constant. In Section~\ref{fse3} we prove this upper bound is tight. On the other hand, the $v_1$ error vanishes at a faster rate of $1/R$. Moreover, the $v_1$ error is much smaller than the $v_0$ error, even in cases when $R$ is small. } Given the performance of the $v_1$ approximation in the Erlang-C model, it is natural to wonder whether the $v_1$ error vanishes at a faster rate (compared to the $v_0$ error) for other models as well. The answer is mixed; e.g., it is not true for the model in Section~\ref{fse4}. In this paper we provide other options for $v(x)$ beyond $v_0$ and $v_1(x)$. For $n \geq 1$, we define a $v_n$ approximation to be one that uses information from the first $n+1$ terms of the Taylor expansion in \eqref{eq:taylorgeneric}; $v_n$ approximations are not unique. We adopt the convention that $v_n$ can refer to either the function $v_n(x)$, or the $v_n$ approximation itself. As a preview, we can use third-order information from the Taylor expansion is by setting \begin{align} \label{eq:v2} v(x) = v_2(x) = v_2^{(\eta)}(x)=\max \bigg\{\frac{a(x)}{2} -\frac{b(x)c(x)}{3a(x)} -\frac{a(x)}{6}\Big(\frac{c(x)}{a(x)}\Big)', \eta \bigg\} \quad \text{ for } x\in (\underline{w}, \overline{w}), \end{align} where $a(x)$ and $c(x)$ are extensions of $ \mathbb{E}(\Delta^2|W=x)$ and $ \mathbb{E}(\Delta^3|W=x)$ to $(\underline{w}, \overline{w})$, respectively, and $\eta>0$ is a tuneable parameter selected to keep $v_2(x)$ positive. We formally motivate and derive \eqref{eq:v2} in Section~\ref{sec:v2def}, where we also elaborate on the need for $\eta$ and how to choose it. Going beyond $v_2$, we derive $v_3$ for the hospital model of Section~\ref{fse4} and the AR(1) model of Section~\ref{fse5}. In both cases, numerical work suggests that finding an approximation that achieves either a faster convergence rate of the error to zero, or a significantly lower approximation error than $v_0$, requires us to go as far as $v_3$. For a discussion on how to determine which $v_n$ to use, see Section~\ref{sec:hospcompare}. This paper is limited to the setting where the Markov chain is one-dimensional because the derivation of $v_n$ for $n \geq 2$ exploits the one-dimensional nature of the Poisson equation; for an example, see Section~\ref{sec:v2def}. At present we do not know how to generalize this to the multidimensional setting. The theoretical framework underpinning our work is Stein's method, which was pioneered by \cite{Stei1972}. Specifically, we use the generator comparison framework of Stein's method, which dates back to \cite{Barb1988} and was popularized recently in queueing theory by \cite{Gurv2014}. We remark that deriving the $v_n$ approximations requires only algebra, which is handy from a practical standpoint as one can derive and implement the approximations numerically without worrying about justifying them theoretically. In addition to deriving the $v_n$ approximations, we also use Stein's method to provide theoretical guarantees. For the Erlang-C model, Theorem~\ref{thm:md-high} establishes Cram\'er-type moderate-deviations error bounds. If $Y$ is an approximation of $W$, then moderate-deviations bounds refer to bounds on the relative error \begin{align*} \bigg| \frac{\mathbb{P}(Y \geq z)}{\mathbb{P}(W \geq z)} - 1 \bigg| \quad \text{ and } \quad \bigg| \frac{\mathbb{P}(Y \leq z)}{\mathbb{P}(W\leq z)} - 1 \bigg| . \end{align*} Compared to the Kolmogorov distance $\sup_{z \in \mathbb{R}} \big| \mathbb{P}(W \geq z) - \mathbb{P}(Y \geq z) \big|$, the relative error is a much more informative measure when the value being approximated is small, as is the case in the approximation of small tail probabilities. For many stochastic systems modeling service operations such as customer call centers and hospital operations, these small probabilities represent important performance metrics; e.g.,\ at most $1\%$ of customers waiting more than 10 minutes before getting into service. To summarize, our main contribution is to present a new family of $v_n$ approximations for Markov chains. Using a combination of theoretical and numerical results, we show that the $v_n$ approximations perform significantly better than the traditional $v_0$ approximation across three separate models. Our results suggest that $v_1, v_2, v_3, \ldots$ can, and should, be used whenever possible to achieve much greater approximation accuracy. Before moving on to the main body of the paper, we first provide a brief review of related literature. \subsection{Literature Review} \label{sec:literature} \paragraph{Steady-state diffusion approximations.} In the last fifty years, diffusion approximations have been a major research theme in the applied probability community for approximate steady-state analysis of many stochastic systems; c.f., \cite{King1961a,HalfWhit1981,HarrNguy1993,MandZelt2009}. Some of these approximations were initially motivated by \emph{process-level limit theorems} that establish functional central limits in certain asymptotic parameter regions; e.g., \cite{Reim1984,Bram1998,Will1998a}. The pioneering paper of \cite{GamaZeev2006} initiated a wave of research providing \emph{steady-state limit theorems}, justifying steady-state approximations on top of process-level convergence. For some examples of these, see \cite{Tezc2008, ZhanZwar2008, BudhLee2009, Kats2010, GamaStol2012, DaiDiekGao2014, Gurv2014a, YeYao2016}. Steady-state limit theorems do \emph{not} provide a rate of convergence or an error bound. Recently, building on earlier work by \cite{GurvHuanMand2014}, \cite{Gurv2014} developed a general approach to proving the rate of convergence for steady-state performance measures of many stochastic systems. In the setting of the $M/Ph/n+M$ queue with phase-type service time distributions, \cite{BravDai2017} refined the approach in \cite{Gurv2014}, casting it into the Stein framework that has been extensively studied in the last fifty years. The Stein framework allows one to obtain an error bound, not just a limit theorem, for approximate steady-state analysis of a stochastic system with a \emph{fixed} set of system parameters. Readers are referred to \cite{BravDaiFeng2016} for a tutorial introduction to using Stein's method for steady-state diffusion approximations of Erlang-A and Erlang-C models, where error bounds were established under a variety of metrics, including the Wasserstein distance, Kolmogorov distance, and moment difference. \paragraph{Stein's method and moderate deviations.} Stein's method was first introduced by \cite{Stei1972}. We refer the reader to the book by \cite{ChenGoldShao2011} for an introduction to Stein's method. Moderate deviations date back to \cite{Cram1938}, who obtained expansions for tail probabilities for sums of independent random variables about the normal distribution. Stein's method for moderate deviations for general dependent random variables was first studied in \cite{ChFaSh13a}. See \cite{ChFaSh13b}, \cite{ShZhZh18}, \cite{Zh19}, \cite{FaLuSh19} for further developments. \paragraph{Refined mean-field approximations.} First-order approximations, such as mean-field, or fluid model approximations capture the deterministic flow of the Markov chain while ignoring the stochastic effects. A recent series of papers, \cite{GastHoud2017}, \cite{GastLateMass2018}, \cite{GastBortTrib2019}, explored refined mean-field approximations for computing moments of the Markov chain stationary distribution. In those papers, the authors were able to explicitly compute correction terms to the mean-field approximation, which significantly improves the accuracy of the approximation and speeds up the rate at which the approximation error converges to zero. However, the computation of these correction terms rests on assuming that the mean-field model is globally exponentially stable and that the drift of the Markov chain is differentiable. These assumptions fail to hold even for some basic queueing models; e.g., the Erlang-C model. \subsection{Notation and Organization of the Paper} \label{sec:notation} For $a, b \in \mathbb{R}$, we use $a^+, a^-, a \wedge b$, and $a \vee b$ to denote $\max(a,0)$, $\max(-a,0)$, $\min(a,b)$, and $\max(a, b)$, respectively. We adopt the convention that $\sum_{l=k_1}^{k_2}=0$ if $k_2<k_1$. In Section~\ref{sec:v2def}, we derive several versions of $v_2$ and discuss how to analyze the approximation error using Stein's method. In Sections~\ref{fse3}--\ref{fse5}, we study the performance of various $v_n$ approximations for three different Markov chains. To keep the main paper a reasonable length, some details of the proofs are left to the Appendix. \section{Deriving the Diffusion Approximations}\label{sec:v2def} \blue{In the previous section, we said that for $n \geq 1$, a $v_n$ approximation is one that uses information from the first $n+1$ terms of the Taylor expansion in \eqref{eq:taylorgeneric}. In this section, we justify $v_2(x)$ proposed in \eqref{eq:v2} by tapping into the third-order terms in \eqref{eq:taylorgeneric}. For examples of accessing fourth-order terms, we refer the reader to the derivations of $v_3$ for the models in Sections~\ref{fse4} and \ref{fse5}.} What follows can be repeated for continuous-time Markov chains (CTMC), with the identity $\mathbb{E} G f(W)=0$ replacing $ \mathbb{E} f(W') - \mathbb{E} f(W) = 0$, where $G$ is the generator of the CTMC. \blue{As our starting point, we recall from \eqref{eq:taylorgeneric} that \begin{align*} 0 =&\ \mathbb{E} f(W') - \mathbb{E} f(W) = \mathbb{E} \bigg[\sum_{i=1}^{n} \frac{1}{i!} \Delta^i f^{(i)}(W) + \frac{1}{(n+1)!} \Delta^{n+1} f^{(n+1)}(\xi)\bigg], \end{align*} where $\Delta = W' - W$, and that $b(x), a(x)$, and $c(x)$ are extensions of $\mathbb{E}(\Delta|W=x)$, $\mathbb{E}(\Delta^2|W=x)$, and $ \mathbb{E}(\Delta^3|W=x)$ to $(\underline{w}, \overline{w})$, respectively. Let $d(x)$ be an extension of $ \mathbb{E}(\Delta^4|W=x)$ to $(\underline{w}, \overline{w})$. Setting $n = 3$ in the expansion above yields } \begin{align} \mathbb{E} b(W) f'(W)+\frac{1}{2}\mathbb{E} a(W)f''(W) + \frac{1}{6} \mathbb{E} c(W) f'''(W) = -\frac{1}{24} \mathbb{E} d(W) f^{(4)}(\xi_1) \label{eq:taylorthird} \end{align} where $\xi_1$ lies between $W$ and $W'$. We implicitly assume $f(x)$ is sufficiently differentiable and the expectations above exist. Since $\Delta$ is small, we treat the right-hand side as error and use the left-hand side to derive a diffusion approximation. The challenge to overcome is that the stationary density of the diffusion is characterized by \eqref{eq:bar}, which considers only the first two derivatives of a function $f(x)$, whereas the left-hand side of \eqref{eq:taylorthird} contains three derivatives. \blue{We therefore convert $f'''(W)$ into an expression involving $f''(W)$ plus some error.} Consider \eqref{eq:taylorgeneric} again, but with $ n = 2$: \begin{align} \mathbb{E} b(W) f'(W)+\frac{1}{2}\mathbb{E} a(W)f''(W) = -\frac{1}{6} \mathbb{E} c(W) f'''(\xi_2), \label{eq:taylorsecond} \end{align} for some $\xi_2$ between $W$ and $W'$. Fix $f(x)$ and let $g(x) = \int_{0}^{x} \frac{c(y)}{a(y)} f''(y) dy$. Note that \begin{align*} g''(x)=&\ \Big( \frac{c(x)}{a(x)} f''(x)\Big)' =\Big(\frac{c(x)}{a(x)}\Big)'f''(x)+\frac{c(x)}{a(x)}f'''(x). \end{align*} Evaluating \eqref{eq:taylorsecond} with $g(x)$ in place of $f(x)$ there yields \begin{align*} \mathbb{E} \frac{b(W)c(W)}{a(W)}f''(W)+\mathbb{E} \frac{a(W)}{2}\Big(\frac{c(W)}{a(W)}\Big)'f''(W)+\frac{1}{2}\mathbb{E} c(W)f'''(W) = -\frac{1}{6} \mathbb{E} c(W) g'''(\xi_2). \end{align*} Rearranging terms, we have \ben{\label{f3} \frac{1}{6}\mathbb{E} c(W)f'''(W) = -\mathbb{E}\Big( \frac{b(W)c(W)}{3a(W)} +\frac{a(W)}{6}\Big(\frac{c(W)}{a(W)}\Big)' \Big)f''(W) -\frac{1}{18} \mathbb{E} c(W) g'''(\xi_2). } Substituting \eqref{f3} into \eqref{eq:taylorthird}, we obtain \begin{align} & \mathbb{E} b(W)f'(W)+\mathbb{E}\Big(\frac{a(W)}{2}-\frac{b(W)c(W)}{3a(W)}-\frac{a(W)}{6}\Big(\frac{c(W)}{a(W)}\Big)'\Big)f''(W) \notag \\ =&\ \frac{1}{18} \mathbb{E} c(W) g'''(\xi_2) -\frac{1}{24} \mathbb{E} d(W) f^{(4)}(\xi_1). \label{f4} \end{align} The left-hand side resembles the generator of a diffusion process. Define \begin{align} \underline{v}_2(x) = \frac{a(x)}{2}-\frac{b(x)c(x)}{3a(x)}-\frac{a(x)}{6}\Big(\frac{c(x)}{a(x)}\Big)' , \quad x \in (\underline{w}, \overline{w}), \label{eq:underlinev2} \end{align} and let $v_2(x) = (\underline v_2(x) \vee \eta)$ for some $\eta > 0$ to recover the $v_2(x)$ in \eqref{eq:v2}. The value of $\eta$ should be chosen close to zero, and if $\inf_{x \in (\underline{w}, \overline{w})} \underline v_2(x) > 0$, then we can pick $v_2(x) = \underline v_2(x)$. We enforce $v(x) > 0$ because there may be issues with the integrability of the density in \eqref{f10} if $v(x)$ is allowed to be negative. For instance, in all three examples considered in this paper, $b(x) > 0$ when $x$ is to the left of the fluid equilibrium of $W$, and $b(x) < 0$ when $x$ is to the right of the fluid equilibrium; i.e.,\ the DTMC drifts back toward its equilibrium. This drift toward the equilibrium is intimately tied to the positive recurrence of the DTMC and can therefore be thought of as a reasonable assumption even if we go beyond this paper's three examples. Now, if $v(x)$ is allowed to be negative, it may be that $\kappa = \infty$ in \eqref{f10}; e.g.,\ if $v(x) < 0$ for $x > K$ for some threshold $K$. Conversely, $\inf_{x \in (\underline{w}, \overline{w})} v(x) > 0$ is sufficient to ensure that $\kappa < \infty$ in all three of our examples. Another, more intuitive, reason that $v(x) > 0$ is that a diffusion coefficient cannot be negative. \subsection{The $v_2$ Approximation Error}\label{sec:theoryv2} Let us discuss the error of our $v_2$ approximation. For simplicity, let us assume that $(\underline{w}, \overline{w}) = \mathbb{R}$ and that $\inf_{x \in \mathbb{R}} \underline v_2(x) >0$, i.e., $v_2(x)$ equals the untruncated version $\underline v_2(x)$. We discuss in Section~\ref{sec:hybrid} what happens when the latter assumption does not hold. Suppose $Y$ is a random variable with density as in \eqref{f10} and with $v(x)$ there equal to $v_2(x)$, \blue{i.e., \begin{align*} \frac{\kappa}{v_2(x)} \exp\Big(\int_0^x \frac{b(y)}{v_2(y)} dy \Big),\quad x\in (\underline{w}, \overline{w}), \end{align*}} and assume for simplicity that $(\underline{w}, \overline{w}) = \mathbb{R}$. Fix a test function $h: \mathbb{R} \to \mathbb{R}$ with $\mathbb{E} \abs{h(Y)} < \infty$, and let $f_h(x)$ be the solution to the Poisson equation \begin{align} b(x) f_h'(x) + v_2(x) f_h''(x) = \mathbb{E} h(Y) - h(x), \quad x \in \mathbb{R}. \label{eq:poissonde} \end{align} Assume that $\mathbb{E} \abs{f_h (W)} < \infty$, which is typically true in practice, and take expected values with respect to $W$ to get \begin{align*} \mathbb{E} h(Y) - \mathbb{E} h(W) =&\ \mathbb{E} b(W) f_h'(W) + \mathbb{E} v_2(W) f_h''(W) = \frac{1}{18} \mathbb{E} c(W) g_{h}'''(\xi_2) -\frac{1}{24} \mathbb{E} d(W) f_{h}^{(4)}(\xi_1). \end{align*} The last equality follows from \eqref{f4}, and $g_h(x) = \int_{0}^{x} \frac{c(y)}{a(y)} f_h''(y) dy$. We have again made an implicit assumption that $f_h(x)$ is sufficiently regular. The regularity of $f_h(x)$ is entirely determined by the regularity of $b(x)$, $v(x)$, and $h(x)$. The right-hand side equals \begin{align} &\frac{1}{18} \mathbb{E} \Big[c(W) \Big(\frac{c(x)}{a(x)}\Big)''\Big|_{x = \xi_2}f_h''(\xi_2)\Big] + \frac{2}{18} \mathbb{E} \Big[ c(W) \Big(\frac{c(x)}{a(x)}\Big)'\Big|_{x = \xi_2} f_h'''(\xi_2)\Big] \notag \\ &+ \frac{1}{18} \mathbb{E} \Big[c(W) \frac{c(\xi_2)}{a(\xi_2)} f_h^{(4)}(\xi_2)\Big] -\frac{1}{24} \mathbb{E} d(W) f_h^{(4)}(\xi_1) \label{eq:v2taylorerror} \end{align} because \begin{align*} g_h'''(x) =&\ \Big( \frac{c(x)}{a(x)} f_h''(x)\Big)'' = \Big(\frac{c(x)}{a(x)}\Big)''f_h''(x) + 2\Big(\frac{c(x)}{a(x)}\Big)'f_h'''(x)+\frac{c(x)}{a(x)}f_h^{(4)}(x). \end{align*} Note that \eqref{eq:v2taylorerror} contains a term involving $f_h''(x)$ that is not captured by $\underline v_2(x)$. To capture that term, we can consider \begin{align*} \overline v_2(x) = \frac{a(x)}{2}-\frac{b(x)c(x)}{3a(x)}-\frac{a(x)}{6}\Big(\frac{c(x)}{a(x)}\Big)' - \frac{1}{18} c(x) \Big(\frac{c(x)}{a(x)}\Big)'' , \quad x \in \mathbb{R}. \end{align*} Truncating $\overline v_2(x)$ produces yet another $v_2$ approximation with error \begin{align} &\frac{1}{18} \mathbb{E} \Big[c(W) \Big( \Big(\frac{c(x)}{a(x)}\Big)''\Big|_{x = \xi_2}f_h''(\xi_2) - \Big(\frac{c(W)}{a(W)}\Big)'' f_h''(W)\Big)\Big] \notag \\ &+ \frac{2}{18} \mathbb{E} \Big[ c(W) \Big(\frac{c(x)}{a(x)}\Big)'\Big|_{x = \xi_2} f_h'''(\xi_2)\Big] + \frac{1}{18} \mathbb{E} \Big[c(W) \frac{c(\xi_2)}{a(\xi_2)} f_h^{(4)}(\xi_2)\Big] -\frac{1}{24} \mathbb{E} d(W) f_h^{(4)}(\xi_1) \label{eq:v2taylorerroralt} \end{align} in place of \eqref{eq:v2taylorerror}. In order to decide between $\underline v_2(x)$ and $\overline v_2(x)$, let us compare the two error terms in \eqref{eq:v2taylorerror} and \eqref{eq:v2taylorerroralt}. We stress that the following is an informal discussion meant to develop intuition. Theoretical guarantees for $v_2$ must be established on a case-by-case basis and fall outside the scope of this paper. Consider first the error term \eqref{eq:v2taylorerror}. Recall that $a(x),c(x)$, and $d(x)$ equal $\mathbb{E}(\Delta^k|W=x)$ for $k = 2,3,4$, respectively. Now $\Delta = W' - W$ equals $\delta$ times the one-step displacement of the Markov chain. Let us assume that the displacement is bounded by a constant independent of $\delta$, in which case $\mathbb{E}(\Delta^k|W=x)$ shrinks at the rate of at least $\delta^k$ as $\delta \to 0$. In particular, $d(x)$ shrinks at least as fast as $\delta^4$. Since $a(x)$ is the extension of the strictly positive function $\mathbb{E} (\Delta^{2} | W=x)$, we assume that this extension is also strictly positive. Furthermore, we assume that $a(x)$ is of order $\delta^2$, as opposed to merely shrinking at a rate of at least $\delta^2$. Formally, we assume that $\inf\{\delta^{-2} \abs{a(x)} : \delta \in (0,1),\ x \in (\underline{w}, \overline{w})\} > 0$, which implies that, provided the derivatives exist, $c(x) \frac{c(x)}{a(x)}$, $c(x) \Big(\frac{c(x)}{a(x)}\Big)'$, and $c(x) \Big(\frac{c(x)}{a(x)}\Big)''$ all shrink at a rate of at least $\delta^4$ as $\delta \to 0$, making them comparable to $d(x)$. Now consider the error term \eqref{eq:v2taylorerroralt}, focusing on the first line there. Provided $a(x),c(x),$ and $f_h(x)$ are sufficiently differentiable, the mean value theorem implies \begin{align*} &\mathbb{E} \Big[c(W) \Big( \Big(\frac{c(x)}{a(x)}\Big)''\Big|_{x = \xi_2}f_h''(\xi_2) - \Big(\frac{c(W)}{a(W)}\Big)'' f_h''(W)\Big)\Big] \\ =&\ \mathbb{E} \Big[c(W) (\xi_2 - W) \Big( \Big(\frac{c(x)}{a(x)}\Big)'' f_h''(x)\Big)'\Big|_{x = \xi_3}\Big]. \end{align*} Under the two assumptions from before, the terms in front of the derivatives of $f_h(x)$ above shrink at a rate of at least $\delta^5$. If the rest of the terms in \eqref{eq:v2taylorerroralt} shrink at the rate of $\delta^4$, then using $\overline v_2(x)$ instead of $\underline v_2(x)$ as the $v_2$ approximation would not make the error converge to zero faster. For this reason and also because $\underline v_2(x)$ is simpler than $\overline v_2(x)$, we work with $\underline v_2(x)$ in the models we consider. \section{Erlang-C Model}\label{fse3} \blue{In this section we consider the Erlang-C model. We prove that the $v_1$ error converges to zero at a faster rate than the $v_0$ error. We also conduct numerical experiments where we observe that the $v_1$ error is much smaller, often by a factor of 10, than the $v_0$ error. After defining the model, we introduce the approximations in Section~\ref{sec:ecv0v1} and then present theoretical and numerical results in Sections~\ref{sec:erlangctheory} and \ref{sec:erlangcnumerical}, respectively. } \blue{The Erlang-C, or $M/M/n$, system has a single buffer served by $n$ homogeneous servers working in} a first-come-first-served manner. Customers arrive according to a Poisson process with rate $\lambda$, and service times are i.i.d., exponentially distributed with mean $1/\mu$. We let $R = \lambda/\mu$ and $\rho = \frac{\lambda}{n\mu} = R/n$ be the offered load and utilization, respectively. Let $X(t)$ be the number of customers in the system at time $t$. We assume that $\rho < 1$, implying that $X = \{X(t), t \geq 0\}$ a positive recurrent CTMC. Set $\delta = 1/\sqrt{R}$, $\tilde X = \{ \tilde X(t) = \delta(X(t) - R),\ t \geq 0\}$, and let $W$ be the random variable having the stationary distribution of $\tilde X$. The support of $W$ is $\mathcal{W} =\{\delta(k-R): k\in \mathbb{Z}_+\}$, so we let \begin{align} (\underline{w}, \overline{w}) = (-\delta R, \infty) = (-\sqrt{R}, \infty). \label{eq:smallestinterval} \end{align} The generator of $\tilde X$ satisfies (cf. Eq. (3.6) of \cite{BravDaiFeng2016}) \ben{ \label{eq:gx} G_{\tilde X}f(x)=\lambda (f(x+\delta)-f(x))+\mu\big[ (x/\delta+R) \wedge n \big] (f(x-\delta)-f(x)), } where $x = \delta (k-R)$ for some integer $k \geq 0$. Proposition 1.1 in \cite{Hend1997} states that \ben{ \label{eq:gxbar} \mathbb{E} G_{\tilde X} f(W) = \mathbb{E} \Big[ \lambda (f(W+\delta)-f(W))+\mu\big[ (W/\delta+ R) \wedge n \big] (f(W-\delta)-f(W)) \Big]=0 } for all $f(x)$ such that $\mathbb{E} \abs{f(W)} < \infty$. \subsection{The $v_0$ and $v_1$ Approximations} \label{sec:ecv0v1} Let us perform Taylor expansion on the left-hand side of \eqref{eq:gxbar}: \begin{align} \mathbb{E} b(W) f'(W)+\mathbb{E} \frac{a(W)}{2}f''(W) =&\ -\frac{1}{6}\big( \delta^3 \lambda f^{'''}(\xi_1) - \delta^3 \mu\big[ (W/\delta+ R) \wedge n \big]f^{'''}(\xi_2)\big), \label{f6} \end{align} where $\xi_{1} \in (W,W+\delta)$, $\xi_2 \in (W-\delta, W)$, \begin{align} & b(x) = \delta \big(\lambda -\mu \big[(x/\delta+R)\wedge n\big]\big), \quad \text{ and } \label{eq:tb}\\ & a(x) = \delta^2 \big(\lambda+ \mu \blue{[(x/\delta+R)\wedge n]} \big) = 2\mu - \delta b(x), \quad x\in \mathcal{W}. \label{eq:ta} \end{align} The second equality in \eqref{eq:ta} holds because $\delta^2 = 1/R = \mu/\lambda$. Let \begin{align} \beta= \delta(n-R)>0, \quad \text{ or } \quad n = R +\beta \sqrt{R}. \label{eq:staffing} \end{align} When $\beta$ is fixed and $R,n \to \infty$, the asymptotic regime is known as the Halfin-Whitt regime; see \cite{HalfWhit1981}. It is also known as the \emph{quality and efficiency--driven regime} because in this parameter region, the system simultaneously achieves short average waiting time (quality) and high server utilization (efficiency); \cite{GansKoolMand2003}. Some of our results assume that $\beta$ is fixed, while others do not. By considering the cases when $x \leq \beta$ and $x > \beta$ in \eqref{eq:tb}, we see that $b(x) = -(\mu x \wedge \mu \beta)$ for $x \in \mathcal{W}$, and we extend $b(x)$ to the entire real line via \begin{align} b(x) = -(\mu x \wedge \mu \beta), \quad x \in \mathbb{R}. \label{eq:bdef} \end{align} We also want a strictly positive extension of $a(x)$ to $\mathbb{R}$. Since $\mathcal{W} \subset [ -\sqrt{R}, \infty)$, we define \begin{align} a(x) = 2\mu - \delta b(-\sqrt{R} \vee x), \quad x \in \mathbb{R}, \label{eq:adef} \end{align} and since $b(x)$ is nonincreasing and $b(-\sqrt{R}) = \mu \sqrt{R} = \mu/\delta$, we have $a(x) \geq a(-\sqrt{R}) = \mu$. Recall from \eqref{f10} that our diffusion approximations all have density of the form \begin{align*} \frac{\kappa}{v(x)} \exp\Big(\int_0^x \frac{b(y)}{v(y)} dy \Big),\quad x\in \mathbb{R}, \end{align*} for some normalizing constant $\kappa > 0$. The $v_0$ and $v_1$ approximations are obtained by setting \begin{align*} v(x) = v_0 = \frac{1}{2} a(0) = \mu \quad \text{ and } \quad v(x) = v_1(x) = \frac{1}{2} a(x), \quad x \in \mathbb{R}. \end{align*} Let $Y_0$ and $Y_1$ be the random variables corresponding to $v_0$ and $v_1$, respectively. \begin{remark} To better approximate $W$, we can use a diffusion process defined on $[-\sqrt{R},\infty)$ with a reflecting condition at the left boundary of $x = -\sqrt{R}$. However, our theorems in Section~\ref{sec:erlangctheory} are intended for the asymptotic regimes when $R \to \infty$. Since the probability of an empty system shrinks rapidly as $R$ grows, the choice between a reflected diffusion on $[-\sqrt{R}, \infty)$ and a diffusion defined on $\mathbb{R}$ is inconsequential. \end{remark} \subsection{Theoretical Guarantees for the Approximations}\label{sec:erlangctheory} We now present several theoretical results showing that the $v_1$ error vanishes faster than the $v_0$ error. Define the class of all Lipschitz-$1$ functions by \begin{align*} \text{\rm Lip(1)} = \big\{h: \mathbb{R} \to \mathbb{R} \ \big|\ \abs{h(x)-h(y)} \leq \abs{x-y} \text{ for all $x,y\in \mathbb{R}$}\big\}. \end{align*} It was shown in \cite{BravDaiFeng2016} that \begin{align} \sup_{h \in \text{\rm Lip(1)}} \big| \mathbb{E} h(W) - \mathbb{E} h(Y_0) \big| \leq \frac{205}{\sqrt{R}}, \quad \text{ if } R < n. \label{eq:oldmain} \end{align} The quantity on the left-hand side above is known as the Wasserstein distance and, as was shown in \cite{GibbSu2002}, convergence in the Wasserstein distance implies convergence in distribution. To add to the result of \cite{BravDaiFeng2016}, we prove the following lower bound in Section~\ref{sec:lowerbound} of the electronic companion. \begin{proposition} \label{prop:lowerbound} Assume $n = R + \beta \sqrt{R}$ for some fixed $\beta > 0$. There exists a constant $C(\beta) > 0$ depending only on $\beta$ such that \begin{align*} \big| \mathbb{E} W - \mathbb{E} Y_0 \big| \geq \frac{C(\beta)}{\sqrt{R}}. \end{align*} \end{proposition} An immediate implication of Proposition~\ref{prop:lowerbound} is that the Wasserstein distance between $W$ and $Y_0$ is at least $C(\beta)/\sqrt{R}$. The assumption that $\beta$ is fixed can likely be removed (with additional effort), but that is not the focus of our paper. We turn to the $v_1$ approximation. Define $W_2 = \{h: \mathbb{R} \to \mathbb{R}\ |\ h(x), h'(x) \in \text{\rm Lip(1)} \}$ and for two random variables $U,V$, define the $W_2$ distance as \begin{align*} d_{W_2}(U,V) = \sup_{h \in W_2} \big| \mathbb{E} h(U) - \mathbb{E} h(V) \big|. \end{align*} Although $W_2 \subset \text{\rm Lip(1)}$, it still rich enough to imply convergence in distribution. In particular, Lemma 3.5 of \cite{Brav2017} shows that by approximating the indicator function of a half line by Lipschitz functions with bounded second derivatives, convergence in the $d_{W_2}$ distance implies convergence in distribution. The following result first appeared as Theorem 3.1 in \cite{Brav2017}. \begin{theorem} \label{thm:w2} There exists a constant $C > 0$ (independent of $\lambda, n$, and $\mu$) such that for all $n \geq 1, \lambda > 0$, and $\mu > 0$ satisfying $1 \leq R < n $, \begin{align*} \sup_{h \in W_2} \big| \mathbb{E} h(W) - \mathbb{E} h(Y_1) \big| \le \frac{C}{R}. \end{align*} \end{theorem} Note that $h(x) = x$ belongs to $W_2$, so Theorem~\ref{thm:w2} and Proposition~\ref{prop:lowerbound} tell us that the the $v_1$ approximation error of $\mathbb{E} (W)$ is guaranteed to vanish faster than the $v_0$ error as $R \to \infty$. Error bounds of the flavor of Theorem~\ref{thm:w2} were established in \cite{GurvHuanMand2014, Gurv2014, BravDai2017, BravDaiFeng2016}, all of which studied convergence rates for steady-state diffusion approximations of various models. The rate of $1/R$ is an order of magnitude better than the rates in any of the previously mentioned papers, where the authors obtained rates that would be equivalent to $1/\sqrt{R}$ in our model. Going beyond error bounds for smooth test functions, we now present moderate-deviations bounds for our two approximations. Namely, we are interested in the relative error of approximating the cumulative distribution function (CDF) and complementary CDF (CCDF). We define the relative error of the right tail to be \begin{align*} \abs{\frac{\mathbb{P}(Y_i \geq z)}{\mathbb{P}(W \geq z)} - 1}, \quad i = 0,1. \end{align*} The relative error for the left tail is defined similarly. The first result is for $v_0$. \begin{theorem} \label{thm:md-std} Assume that $n = R + \beta \sqrt{R}$ for some fixed $\beta > 0$. \blue{There exist positive constants $c_0$ and $C$ depending only on $\beta$ such that \begin{align} & \left|\frac{\mathbb{P}(Y_0\geq z)}{\mathbb{P}(W\geq z)}-1\right|\leq \frac{C}{\sqrt{R}}\left(1+z\right) \quad \text{ for } 0<z\leq c_0 R^{1/2}\ \text{and} \label{f12} \\ &\left|\frac{\mathbb{P}(Y_0\leq -z)}{\mathbb{P}(W\leq -z)}-1\right|\leq \frac{C}{\sqrt{R}}\left(1+z^3\right),\ \text{ for } 0<z\leq \min\{c_0 R^{1/6}, R^{1/2}\}. \label{f15} \end{align} } \end{theorem} The second result presents analogous bounds for the $v_1$ approximation. \begin{theorem} \label{thm:md-high} Assume $n = R + \beta \sqrt{R}$ for some fixed $\beta > 0$. \blue{There exist positive constants $c_1$ and $C$ depending only on $\beta$ such that \begin{align} & \left|\frac{\mathbb{P}(Y_1\geq z)}{\mathbb{P}(W\geq z)}-1\right|\leq \frac{C}{\sqrt{R}}\left(1+\frac{z}{\sqrt{R}}\right) \quad \text{ for } 0< z\leq c_1 R\ \text{and} \label{f13}\\ &\left|\frac{\mathbb{P}(Y_1\leq -z)}{\mathbb{P}(W\leq -z)}-1\right|\leq \frac{C}{\sqrt{R}}\left(1+z+\frac{z^4}{\sqrt{R}}\right),\ \text{ for } 0<z\leq \min\{c_1 R^{1/4}, R^{1/2}\} \label{f16}. \end{align} } \end{theorem} Inequality \eqref{f13} follows from Theorem~4.1 of \cite{Brav2017}. We prove \eqref{f16} in Section \ref{fse8}; Theorem \ref{thm:md-std} follows from a similar and simpler proof in Section~\ref{fap3}. These are called moderate deviations bounds because they cover the case when $z$ is ``moderately'' far from the origin, with ``moderately'' being quantified by intervals of the form $z \in [0,c_0R^{1/2}]$, $z \in [0,c_1R]$, etc. In contrast, large-deviations results focus on understanding the behavior of $\mathbb{P}(W \geq z)$ as $z \to \infty$. To compare the two theorems, suppose $z = c_0 \sqrt{R}$ and consider the upper bounds in \eqref{f12} and \eqref{f13}. The $v_0$ error is guaranteed to be bounded as $R$ grows, while the $v_1$ error shrinks at a rate of at least $1/\sqrt{R}$. \subsection{Numerical Results}\label{sec:erlangcnumerical} Although the Erlang-C system depends on three parameters $\lambda$, $\mu$, and $n$, the stationary distribution depends on only $\rho$ and $n$; see Appendix C in \cite{Alle1990}. Figure~\ref{fig:ecmoment} displays the relative errors of the $v_0$ and $v_1$ approximations of $\mathbb{E} (W)$ when $5 \leq n \leq 100$ and $0.5 \leq \rho \leq 0.99$. Note that the $v_1$ error is about ten times smaller. We also compare how well $v_0$ and $v_1$ approximate the CCDF of $W$. Figure~\ref{fig:ecmd1} plots the relative error of approximating $\mathbb{P}(W/\delta +R \geq z)$ for various values of $z$ when $n = 10$ and $\rho \in [0.5,0.99]$, and shows that the $v_1$ error is again much smaller. In results not reported in the paper, we observed that the $v_1$ error remains much smaller even as we vary $n$. \begin{figure}[h!] \centering \includegraphics[scale=0.55]{scaledmean.pdf} \caption{The errors increase towards the bottom-right corner of each plot. This is due to the fact that $\mathbb{E} (W)$ is very close to zero in that region and not because approximations perform poorly. \label{fig:ecmoment}} \end{figure} \begin{figure} \centering \includegraphics[scale=0.24]{moddev3dn10part1.pdf} \caption{The $v_1$ approximation is much more accurate. } \label{fig:ecmd1} \end{figure} \section{Hospital Model}\label{fse4} In this section we consider the discrete-time model for hospital inpatient flow proposed by \cite{DaiShi2017}. Numerical experiments presented later suggest that both the $v_1$ and $v_2$ errors vanish at the same rate as the $v_0$ approximation error. To observe a faster convergence rate, we have to resort to the $v_3$ approximation. Consider a discrete-time queueing model with $N$ identical servers. Let $X(n)$ be the number of customers in the system at the end of time unit $n$. Given $X(0)$, we define \begin{align*} X(n) = X(n-1) + A(n) - D(n), \quad n \geq 1, \end{align*} where $A(n) \sim$ Poisson$(\Lambda)$ represents the number of new arrivals in the time period $[n-1,n)$. At the end of each time period, every customer in service flips a coin and, with probability $\mu \in (0,1)$, departs the system at the start of the next time period. Thus, conditioned on $X(n-1) = k$, we have $D(n) \sim $ Binomial$(k\wedge N,\mu)$. Assuming $ \Lambda < N \mu$, \cite{DaiShi2017} showed that $X=\{X(n): n=1, 2, \ldots\}$ is a positive-recurrent DTMC. \blue{This DTMC is similar to the Erlang-C model, but unlike the Erlang-C model where the customer count only changes by one at a time, the jump size $X(1) - X(0)$ is not bounded because $A(n)$ is unbounded. As a result, computing the stationary distribution takes a long time when $\Lambda$ is large and the utilization $\rho = \Lambda/(N\mu)$ is near one because the state-space truncation has to be large to account for potential arrivals. } We are interested in the scaled DTMC $\tilde X = \{ \tilde X(n) = \delta (X(n) - N)\}$. To stay consistent with \cite{DaiShi2017}, we center $X(n)$ around $N$. We consider the parameter ranges studied in \cite{DaiShi2017}, which, given some constant $\beta>0$, are \begin{align}\label{eq:hosqed} \Lambda=\sqrt{N}-\beta, \quad \mu=\delta=1/\sqrt{N} \end{align} \subsection{Motivating the Need for a $v_3$ Approximation} \label{sec:hospcompare} This section contains an informal discussion aimed at explaining why the $v_0$, $v_1$, and $v_2$ errors vanish at the same rate of $\delta = 1/\sqrt{N}$, and why we need the $v_3$ approximation to observe a convergence rate of $\delta^2 = 1/N$. Initialize $\tilde X(0)$ according to the stationary distribution of $\tilde X$, let $W = \tilde X(0)$, $W' = \tilde X(1)$, and set $\Delta = W' - W$. The support of $W$ is $\mathcal{W}=\{ \delta(k-N): k \in \mathbb{N} \} \subset [-\delta N, \infty)$. As we are accustomed to doing by this point, we let $b(x) = \mathbb{E}(\Delta | W = x)$. We know from (37) and (38) of \cite{DaiShi2017} that for $x \in \mathcal{W}$, \begin{align} b(x) = \mathbb{E}(\Delta| W= x) =&\ \delta( x^{-} - \beta ), \quad \text{ and } \label{eq:1}\\ \mathbb{E}(\Delta^2|W=x) =&\ 2\delta + \big(b^{2}(x)-\delta b(x) -\delta^2 - 2\delta^2\beta\big) + \delta^3 x^{-}. \label{eq:2} \end{align} For the higher moments of $\Delta$, let us use $\epsilon(x)$ to represent a generic function that may change from line to line but always satisfies the property \begin{align} \abs{\epsilon(x)} \leq C (1 + \abs{x})^{5} \label{eq:epsdef} \end{align} for some constant $C>0$ that depends only on $\beta$. We show in Section~\ref{sec:hosmecproof} that \begin{align} \mathbb{E}(\Delta^3|W=x) = \delta^2 \epsilon(x), \quad \mathbb{E}(\Delta^4|W=x) = \delta^2 \epsilon(x), \quad \mathbb{E}(\Delta^5 |W=x) = \delta^3 \epsilon(x), \quad x \in \mathcal{W}. \label{eq:345} \end{align} All of our $v_n$ approximations share the same drift $b(x)$, but the diffusion coefficients vary. As always, $v_1(x) = \frac{1}{2} \mathbb{E}(\Delta^2|x )$. Since $b(x) = 0$ at $x = -\beta$, \eqref{eq:2} implies that \begin{align} v_0 =&\ v_1(-\beta) = \frac{1}{2}\big( 2\delta - \delta^2 (1+2\beta) + \delta^3 \beta\big). \label{eq:hospv0} \end{align} The following informal discussion assumes that all functions are sufficiently differentiable and that all expectations exist. Our starting point, as always, is the Poisson equation \begin{align} b(x) f_h'(x) + v(x) f_h''(x) = \mathbb{E} h(Y) - h(x), \quad x \in \mathbb{R}, \label{eq:hosppoisson} \end{align} where $v(x)$ is a temporary placeholder and $Y$ has density given by \eqref{f10}. The Taylor expansion in \eqref{eq:taylorgeneric} tells us that \begin{align*} \mathbb{E} b(W) f_h'(W) + \mathbb{E} \bigg[\sum_{i=2}^{n} \frac{1}{i!} \Delta^i f_h^{(i)}(W) + \frac{1}{(n+1)!} \Delta^{n+1} f_h^{(n+1)}(\xi)\bigg] = 0, \end{align*} where $\xi = \xi^{(n)}$ lies between $W$ and $W'$. Subtracting this equation from \eqref{eq:hosppoisson} and taking expected values there with respect to $W$ we see that for $n \geq 1$, \begin{align} \mathbb{E} h(W) - \mathbb{E} h(Y) =&\ -\mathbb{E} v(W) f_h''(W) + \mathbb{E} \bigg[\sum_{i=2}^{n} \frac{1}{i!} \Delta^i f_h^{(i)}(W) + \frac{1}{(n+1)!} \Delta^{n+1} f_h^{(n+1)}(\xi)\bigg]. \label{eq:hosperr} \end{align} Consider the $v_0$ error. When $v(x) = v_0$, equation \eqref{eq:hosperr} with $n = 2$ there becomes \begin{align} \mathbb{E} h(W) - \mathbb{E} h(Y) =&\ \mathbb{E} \Big(\frac{1}{2}\Delta^2 - v_0 \Big) f_h''(W) + \frac{1}{6} \mathbb{E} \Delta^3 f_h'''(\xi). \label{eq:v0errorhosp} \end{align} Note that $f_h(x)$ depends on $v(x)$ because it solves \eqref{eq:hosppoisson}. In Lemma~3 of \cite{DaiShi2017}, the authors proved that $\abs{f_h''(x)} \leq C/\delta$ and $\abs{f_h'''(x)} \leq C/\delta$ for some constant $C > 0$ dependent only on $\beta$. Assuming that $f_h''(x)$ and $f_h'''(x)$ indeed grow at the rate of $C/\delta$ as $\delta \to 0$, the forms of $\mathbb{E} (\Delta^2 | W = x)$ and $v_0$ in \eqref{eq:2} and \eqref{eq:hospv0} yield \begin{align*} \frac{1}{2}\mathbb{E} (\Delta^2 | W = x) - v_0 =&\ \frac{1}{2} \Big(b^{2}(x)-\delta b(x) + \delta^3 (x^{-}-\beta)\Big) = \frac{1}{2} \Big(b^{2}(x)-\delta b(x) + \delta^2 b(x)\Big). \end{align*} Since $b(x) = \delta(x^{-}-\beta)$, this quantity is of order $\delta^2$. Now \eqref{eq:345} says that $\mathbb{E} (\Delta^3| W =x)$ is also of order $\delta^2$. Therefore, we expect both $ \mathbb{E} \big(\Delta^2/2 - v_0 \big) f_h''(W)$ and $\mathbb{E} \Delta^3 f_h'''(\xi)$ to be of order $\delta$, so even if $ \mathbb{E} \big(\Delta^2/2 - v_0 \big) f_h''(W)$ were not present in \eqref{eq:v0errorhosp}, the approximation error would still be of order $\delta$ due to $\mathbb{E} \Delta^3 f_h'''(\xi)$. We believe this is why the $v_0$ and $v_1$ errors appear to vanish at the same rate despite $v_1(x)$ capturing the entire second order term of \begin{align} \mathbb{E} \bigg[\sum_{i=2}^{n} \frac{1}{i!} \Delta^i f_h^{(i)}(W) + \frac{1}{(n+1)!} \Delta^{n+1} f_h^{(n+1)}(\xi)\bigg]. \label{eq:hosptaylorerror} \end{align} Going beyond $v_0$ and $v_1$, we see that for the error to be of order $\delta^2$, the diffusion approximation must capture all the terms in \eqref{eq:hosptaylorerror} that are of order $\delta$. If we assume for the moment that the derivatives of $f_h(x)$ are all of order $1/\delta$, we see that our approximation has to capture all terms of order $\delta^2$ or larger in the functions $\{ \mathbb{E}(\Delta^i|x)\}_{i=1}^{\infty}$. From \eqref{eq:1}, \eqref{eq:2}, and \eqref{eq:345} we see that $\mathbb{E} (\Delta^{1} | x)$ through $\mathbb{E} (\Delta^{4} | x)$ are all of order $\delta$ or $\delta^2$, while $\mathbb{E}(\Delta^{5} | x)$ is of order $\delta^3$. Thus, if we want the error to be of order $\delta^2$, our approximation must capture the terms of order $\delta$ and $\delta^2$ in $\mathbb{E} (\Delta^{1} | x)$ through $\mathbb{E} (\Delta^{4} | x)$ and can ignore terms of order $\delta^3$ like $\mathbb{E}(\Delta^{5} | x)$. We remark that $v_2(x)$ in \eqref{eq:v2} depends only on $\mathbb{E} (\Delta^{1} | x)$ through $\mathbb{E} (\Delta^{3} | x)$ and not on $\mathbb{E} (\Delta^{4} | x)$. We suspect this is why we observe the $v_2$ error to be of order $\delta$. In Section~\ref{app:hospital_proofs} we derive a $v_3$ approximation of the form \begin{align} \label{eq:hosv3} v_3(x)=\max\Big\{ \delta +\frac{1}{2} \Big(\delta^{2} 1(x<0)- \delta b(x)-\delta^2-2\delta^2\beta\Big), \delta/2 \Big\} \end{align} and in the following section we present numerical results that suggest that the error of this approximation converges to zero at a rate of $\delta^2$. \subsection{Numerical Results} \label{sec:hospital_numeric} In Figure~\ref{fig:hospmoments} we compare the $v_n$ approximations for $\mathbb{E} (W)$ when $\beta = 1$ and $N \in \{4,16,64,256\}$. The values of $\mathbb{E} (W)$ are estimated using a simulation; the width of the $95\%$ confidence intervals (CIs) is on the order of $10^{-4}$. Though we do not report them, the $v_0, v_1$, and $v_2$ approximation errors appear to decay at the rate of $1/\sqrt{N}$, but the $v_3$ error in the table appears to decay linearly in $N$. When it comes to approximating the CCDF, $v_3$ also outperforms the other approximations; see Figure~\ref{fig:hosp64} for an example when $N = 64$ and $\beta = 1$. Our findings were consistent for other values of $\beta$ and $N$. \begin{figure}[H] \centering \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width=.8\linewidth]{moment_comparison_beta1.pdf} \end{subfigure}% \begin{subfigure}[t]{0.45\textwidth} \centering \setlength\tabcolsep{3pt} \renewcommand{\arraystretch}{0.5} \vspace{-3.5cm} \begin{tabular}{rc | c | c } $N$ & $\beta$ & $\mathbb{E} W$ &$95\% \text{CI for } \big| \mathbb{E} (W) - \mathbb{E} (Y_3) \big|$ \\ \hline 4 & 1 & -0.933 & $[0.0007, 0.0008] $ \\ 16 & 1 & -0.865 & $[0.0016, 0.0017]$ \\ 64 & 1 & -0.823 &$[0.0004, 0.0005]$ \\ 256 & 1 & -0.801 & $[0,0.0002]$ \\~\\ \end{tabular} \end{subfigure} \caption{$\beta = 1$. $Y_n$ corresponds to the $v_n$ approximation. } \label{fig:hospmoments} \end{figure} \vspace{-1cm} \begin{figure} \centering \includegraphics[width=3.5in]{N64beta1.pdf} \caption{$N = 64$ and $\beta = 1$. $\mathbb{P}(W \geq x) \approx 10^{-6}$ for $x$ at the right endpoint of the $x$-axis. } \label{fig:hosp64} \end{figure} \section{AR(1) Model}\label{fse5} In this section we consider the first-order autoregressive model with random coefficient and general error distribution, which we refer to as the AR$(1)$ model. \cite{BlanGlyn2018} studied this model and used an Edgeworth expansion to approximate its stationary distribution. Following the notation of \cite{BlanGlyn2018}, we compare the performance their expansion to our diffusion approximations. We encounter the issue that for large values of $x$, the untruncated $v_3(x)$ becomes sufficiently negative to make our usual solution of truncation from below perform poorly when approximating the tail of the distribution. We resolve this via a hybrid approximation that uses both $v_3(x)$ and $v_2(x)$ to construct a diffusion coefficient $\hat v_3(x)$ that combines the extra accuracy of $v_3(x)$ with the positivity of $v_2(x)$. To introduce the model, let $\{X_n, n\geq 1\}$, $\{Z_n, n\geq 1\}$ be two independent sequences of i.i.d.\ random variables. We assume that both $X_1$ and $Z_1$ are exponentially distributed with unit mean so that we can compare our approximations to those of \cite{BlanGlyn2018}. Given $D_{0} \in \mathbb{R}$ and $\alpha>0$, consider the DTMC $D=\{D_n, n\geq 0\}$ defined as \begin{align} D_{n+1} = e^{-\alpha Z_{n+1}} D_n + X_{n+1} \label{eq:defar1} \end{align} and let $D_\infty > 0$ denote the random variable having the stationary distribution of $D$. Using \eqref{eq:defar1} we can see that $D_{\infty}$ is equal in distribution to $\sum_{k=0}^\infty X_k e^{-\alpha\sum_{j=0}^{k-1}Z_j}$, which is the random variable studied in Section 4 of \cite{BlanGlyn2018}. We consider $\tilde D = \{ \tilde D_n = \delta (D_n - R),\ n \geq 0\}$, where $\delta = \sqrt{\alpha}$ and, to be consistent with \cite{BlanGlyn2018}, we choose $R = 1/\alpha$. The asymptotic regime we consider is $\alpha \to 0$, so going forward we assume that $\alpha \in (0,1)$. It follows from \eqref{eq:defar1} that \[ \tilde D_{n+1} = e^{-\alpha Z_{n+1}} \tilde D_n + \delta\big(X_{n+1} + R(e^{-\alpha Z_{n+1}} - 1)\big). \] Let $W = \delta(D_{\infty}-R)$ and $W' = e^{-\alpha Z}W + \delta\big(X + R(e^{-\alpha Z} - 1)\big)$, where $(X, Z)$ is an independent copy of $(X_1, Z_1)$, which is also independent of $W$. Since $D_{\infty} > 0$, the support of $W$ is $\mathcal{W} = (-1/\sqrt{\alpha}, \infty)$, which grows as $\alpha \to 0$. Stationarity implies that $\mathbb{E} f(W')-\mathbb{E} f(W)=0$ provided $\mathbb{E} \abs{f(W)} < \infty$. Note that the one-step jump size \begin{align*} \Delta = W' - W = W\big(e^{-\alpha Z}-1\big) + \delta\big(X + R(e^{-\alpha Z} - 1)\big) = \delta \big(D_{\infty}(e^{-\alpha Z} - 1) + X\big) \end{align*} does not depend on the choice of $R$. To present our diffusion approximations, we need expressions for $\mathbb{E}(\Delta^{k} | W = x)$. The following lemma is proved in Section~\ref{app:ar1proof}. \begin{lemma}\label{lem:ar1} Recall that $\delta = \sqrt{\alpha}$. For any $k \geq 1$, \begin{align*} \mathbb{E}(\Delta^{k} | D_{\infty} = d) = \delta^{k} k! \bigg(1 + \sum_{i=1}^{k} (-1)^{i}d^{i} \prod_{j=1}^{i} \frac{ \alpha }{1 + j \alpha} \bigg), \quad d > 0. \end{align*} \end{lemma} The relationship between $D_{\infty}$ and $W$ implies that \begin{align*} \mathbb{E}(\Delta^{k} | W = x) = \mathbb{E}(\Delta^{k} | D_{\infty} = x/\delta + R) =&\ \delta^{k} k! \bigg(1 + \sum_{i=1}^{k} (-1)^{i}\Big( x\sqrt{\alpha} + 1 \Big)^{i} \prod_{j=1}^{i} \frac{1}{1 + j \alpha} \bigg), \end{align*} for $x \in \mathcal{W}$, where we used the facts that $\delta = \sqrt{\alpha}$ and $R = 1/\alpha$ in the second equality. Extending $\mathbb{E}(\Delta^{k} | W = x)$ to all $x\in \mathbb{R}$ in the obvious way, we now state the $v_0$, $v_1$, and $v_2$ approximations, whose forms are all standard. Namely, $v_{2}(x)$ follows from \eqref{eq:v2}, $v_1(x) = \mathbb{E}(\Delta^{2} | W = x)/2$, and $v_0 = \mathbb{E}(\Delta^{2} | W = x^{\ast})$, where $x^{\ast} = 1+1/\alpha$ solves $\mathbb{E}(\Delta | W = x) = 0$. To present $v_3(x)$, let us note that $\mathbb{E}(\Delta^{k} | W = x)$ takes the form $\delta^k p_{k}(x)$ for some degree-$k$ polynomial $p_{k}(x)$; we omit the dependence on $\alpha$ to ease notation. Given a truncation level $\eta > 0$, we let $v_3(x) = (\underline{v}_3(x) \vee \eta)$, where \begin{align*} \underline{v}_3(x) =&\ \delta^2 \Big(\frac{p_{2}(x)}{2}-\frac{p_{1}(x)\bar p_3(x)}{ \underline{p}_2(x)}-\delta \underline{p}_2(x)\Big(\frac{\bar p_3(x)}{\underline{p}_2(x)}\Big)'\Big), \\ \bar p_3(x) =&\ \frac{1}{6} \Big( p_3(x) - \frac{p_1(x)p_4(x)}{ 2 p_2(x)} - \frac{1}{4} \delta p_2(x)\Big(\frac{p_4(x)}{p_2(x)}\Big)' \Big),\\ \underline p_2(x) =&\ \Big(\frac{p_2(x)}{2}-\frac{p_1(x)p_3(x)}{3p_2(x)}-\frac{p_2(x)}{6}\Big(\frac{p_3(x)}{p_2(x)}\Big)'\Big). \end{align*} We derive $v_3(x)$ by successively applying the ``trick'' we used in Section~\ref{sec:v2def} to derive the $v_2(x)$ approximation in order to gain access to higher order terms in the Taylor expansion. The details are left to Section~\ref{app:ar1proof} of the electronic companion. Before presenting our numerical results, we discuss how to modify the $v_3$ approximation to overcome the issue that $\underline{v}_3(x)$ becomes negative for large values of $x$. \subsection{Hybrid Approximation} \label{sec:hybrid} Figure~\ref{fig:v3plots} displays $\underline{v}_3(x)$ for several values of $\alpha$. When $\alpha = 0.001$ or $0.01$, the plots of $\underline{v}_3(x)$ are very close to zero but remain nonnegative. However, $\underline{v}_3(x)$ is negative when $\alpha = 0.1, 0.5$, or $0.9$. The behavior of $\underline{v}_3(x)$ in the left tail is not as important because the left boundary of the support of $W = \sqrt{\alpha} (D - 1/\alpha)$ is $-1/\sqrt{\alpha}$; we therefore ignore the negativity in the left part of $\underline v_3(x)$ for $x < 0$. The farther $\underline{v}_3(x)$ drops below zero, the worse we expect the truncated $\underline{v}_3(x)$ to perform. For example, the plot with $\alpha = 0.9$ in Figure~\ref{fig:arbig} of Section~\ref{sec:ar1numerical} shows that while $v_3$ performs well in regions where $\underline{v}_3(x)>0$, it does not perform as well when estimating $\mathbb{P}(W > x)$ for large $x$. To improve upon the $v_3$ approximation, we propose the hybrid approximation $\hat v_3(x) = \underline{v}_3(x) 1(x \leq K) + \underline{v}_2(x) 1(x > K)$. The threshold $K$ is numerically chosen to equal the right-most point of intersection of $\underline{v}_2(x)$ and $\underline{v}_3(x)$. The idea is for $\hat v_3(x)$ to enjoy the increased accuracy of $\underline{v}_3(x)$ in the center with the performance of $\underline{v}_2(x)$ far in the tail. We expect $\hat v_3(x)$ to outperform a truncated $\underline{v}_3(x)$ when $\underline{v}_3(x)$ drops far below zero; e.g., when $\alpha = 0.9$. If $\underline{v}_3(x)$ is nonnegative, we expect little benefit from $\hat v_3(x)$; e.g., when $\alpha = 0.001$. Our expectations are consistent with our numerical findings in Section~\ref{sec:ar1numerical}. \begin{figure} \includegraphics[width=2.5in]{v3plots.pdf} \centering \caption{ When $\alpha = 0.001$ and $0.01$, $\underline{v}_3(x)$ is nonnegative at all points plotted. \label{fig:v3plots}} \end{figure} Lastly, we remark on what can be done in the case when both $\underline{v}_2(x)$ and $\underline{v}_3(x)$ are negative in the same region: instead of falling back on $\underline{v}_2(x)$, we can combine $\underline{v}_3(x)$ with $v_1(x)$, which is always positive because $v_1(x) = \frac{1}{2} \mathbb{E} (\Delta^{2} | W=x) > 0 $ for all $x \in \mathcal{W}$. \subsection{Numerical Results} \label{sec:ar1numerical} It is well known that Edgeworth expansions, obtained for the probability distribution at a particular point, can suffer from two issues: (1) they may not be a proper probability distribution function, and (2) they may not be sufficiently accurate in the tails. Our diffusions approximate the entire distribution. We compare the quality of the $v_n$ and $\hat v_3$ approximations to the Edgeworth expansion of \cite{BlanGlyn2018}. Figure~\ref{fig:arbig} displays the relative error of approximating the CCDF of $W$ for different values of $\alpha$ and contains two plots: one where $\alpha$ is close to zero and one where $\alpha$ is far from zero. In the latter plot, the hybrid approximation is the best performer because $\alpha = 0.9$ and $\underline{v}_3(x)$ is negative in Figure~\ref{fig:v3plots}, whereas in the former plot $v_3$ is the best performer because $\alpha = 0.001$ and $\underline{v}_3(x)$ is nonnegative in Figure~\ref{fig:v3plots}. In addition to estimating the CCDF of $W$, Table~\ref{tab:ar1log} compares the performance of our approximations when estimating the expectation of a smooth test function like $\mathbb{E} \log(W + \delta R) = \mathbb{E} \log(\alpha D_{\infty})$. \begin{table} \begin{center} \renewcommand{\arraystretch}{0.5} \begin{tabular}{|c|c|c|c|c|c| } \hline & $\alpha = 0.64$ & $\alpha = 0.32$ &$\alpha = 0.16$ & $\alpha = 0.08$& $\alpha = 0.04$\\ \hline $\abs{\mathbb{E} f(Y_0) - \mathbb{E} f(W)}$ & 0.095 & 0.039 & 0.011 & 0.002 & $3.6 \times 10^{-4}$ \\ \hline $\abs{\mathbb{E} f(Y_1) - \mathbb{E} f(W)}$ & 0.104 & 0.037 & 0.008 & $9.4 \times 10^{-4}$ & $1.0\times 10^{-4}$ \\ \hline $\abs{\mathbb{E} f(Y_2) - \mathbb{E} f(W)}$ & 0.034 & 0.011 & 0.003 & $4.7 \times 10^{-4}$ & $7.4 \times 10^{-5}$ \\ \hline $\abs{\mathbb{E} f(Y_3) - \mathbb{E} f(W)}$ & 0.0201 & 0.005 & $7.9 \times 10^{-4}$ & $8.9 \times 10^{-5}$ &$7.8\times 10^{-6}$ \\ \hline $\abs{\mathbb{E} f(\hat{Y}_3) - \mathbb{E} f(W)}$ & 0.0194 & 0.005 & $7.8 \times 10^{-4}$ & $8.9 \times 10^{-5}$ & $7.8 \times 10^{-6}$ \\ \hline $\abs{\mathbb{E} f( {Y}_e) - \mathbb{E} f(W)}$ & 0.148 & 0.053 & 0.009 & $7.3 \times 10^{-4}$ & $6.3 \times 10^{-4}$ \\ \hline $ \mathbb{E} f(W) $ & 0.510 & 0.721 & 0.994 & 1.302 & 1.629 \\ \hline \end{tabular} \end{center} \caption{ $f(W) = \log(W + \delta R)$. The random variable $Y_n$ corresponds to the $v_n$-approximation, $\hat Y_3$ corresponds to the $\hat v_3$-approximation, and $Y_e$ corresponds to the Edgeworth expansion estimate. \label{tab:ar1log}} \end{table} \begin{figure} \centering \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width=.8\linewidth]{armoddevalpha0.001.pdf} \vspace{-.3cm} \caption{\footnotesize{$\alpha = 0.001$}} \end{subfigure}% \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width=.8\linewidth]{armoddevalpha0.9.pdf} \vspace{-.3cm} \caption{\footnotesize{$\alpha = 0.9$}} \end{subfigure}% \caption{The plots exclude $v_0$ and $v_1$, the worst performing approximations.} \label{fig:arbig} \end{figure} \section{Conclusion} We have outlined a general procedure to derive $v_n$ approximations for one-dimensional Markov chains. Although the expressions for $v_n(x)$ get more complicated as $n$ increases, the diffusion approximations remain computationally tractable. A natural question is how to extend this work to the multi-dimensional setting. Another direction worth exploring relates to establishing theoretical guarantees for the approximations. The only results we have are for the $v_1$ error in Section~\ref{fse3}, a key ingredient of which are bounds on the derivatives to the solution of the Poisson equation, also known as Stein factor bounds; c.f., Lemma~\ref{lem:higherders} of the electronic companion. Since the Poisson equation depends on the diffusion coefficient, we have to reestablish Stein factor bounds for each new $v_n(x)$; the difficulty of this grows with the complexity of the expression for $v_n(x)$. The prelimit generator approach, recently proposed by \cite{Brav2022}, may offer a simpler avenue for theoretical guarantees because it uses Stein factor for the Markov chain instead of the diffusion. \ECSwitch \ECHead{Accompanying Proofs} This e-companion contains the proofs of certain theoretical results in the paper. It is divided into three main sections. The first section is about the Erlang-C model, and contains the proofs of Proposition~\ref{prop:lowerbound}, Theorem~\ref{thm:md-high}, and Theorem~\ref{thm:md-std}. The second and third sections derive the $v_3$ approximation for the hospital model and AR(1) model, respectively. \section{Companion for the Erlang-C Model} \blue{To prepare for the arguments to come, let us recall the notation related to the Erlang-C model. The Erlang-C model is defined by the customer arrival rate $\lambda >0$, the service rate $\mu > 0$, and the number of servers $n > 0$. Additional important quantities include \begin{align*} R=\frac{\lambda}{\mu} <n,\ \beta=\frac{n-R}{\sqrt{R}}>0, \text{ and } \delta=\frac{1}{\sqrt{R}}. \end{align*} We study $W$, which has the stationary distribution of the CTMC $\{\tilde X(t) = \delta (X(t) - R)\}$, where $X(t)$ is the number of customers in the system at time $t \geq 0$. Equation \eqref{eq:gxbar} states that \ben{ \label{eq:gxbarf} \mathbb{E} G_{\tilde X} f(W) = \mathbb{E} \Big[ \lambda (f(W+\delta)-f(W))+\mu\big[ (W/\delta+ R) \wedge n \big] (f(W-\delta)-f(W)) \Big]=0 } for all $f(x)$ satisfying $\mathbb{E} \abs{f(W)} < \infty$. We also note that the support of $W$ is \begin{align} \mathcal{W} = \{-\sqrt{R}, -\sqrt{R}+\delta, -\sqrt{R}+2\delta,\dots\}. \label{f29} \end{align} To define $v_0(x)$ and $v_1(x)$, we recall from \eqref{eq:tb} and \eqref{eq:ta} that for $x \in \mathcal{W}$, \begin{align} b(x) =&\ \delta \big(\lambda -\mu \big[(x/\delta+R)\wedge n\big]\big) \quad \text{ and } \quad a(x)= \delta^2 \big(\lambda+ \mu \blue{[(x/\delta+R)\wedge n]} \big), \label{eq:tabeff} \end{align} and from \eqref{eq:bdef} and \eqref{eq:adef} that the extensions of these to $\mathbb{R}$ are \begin{align} b(x) =&\ -(\mu x \wedge \mu \beta) \quad \text{ and } \quad a(x)= 2\mu - \delta b( -\sqrt{R}\vee x), \quad x \in \mathbb{R}. \label{eq:adeff} \end{align} We define $v_1(x) = \frac{1}{2} a(x)$ and $v_0(x) = v_0 = v_1(0) = \mu$, and for $n \in \{0,1\}$ we define the $v_n$ approximation to be the random variable $Y_n$ with density \begin{equation} \label{eq:stddenf} \frac{\kappa}{v_n(x)} \exp\Big({\int_0^x \frac{b(y)}{v_n(y)}dy}\Big), \quad x \in \mathbb{R}, \end{equation} where $\kappa > 0$ is a normalization constant that depends on $n$. Lastly, assuming that $-z$ belongs to $ \mathcal{W}$ and setting $f(x)=1(x\geq -z)$ in \eqref{eq:gxbarf}, we get \ben{\label{f26} \lambda \P(W=-z-\delta)=\mu[(-z/\delta+R)\wedge n] \P(W=-z), } which are the flow-balance equations for the CTMC. } \subsection{Proving Proposition~\ref{prop:lowerbound}} \label{sec:lowerbound} We repeat the statement of Proposition~\ref{prop:lowerbound} for convenience. \begin{proposition} \label{prop:lowerboundec} Assume that $n = R + \beta \sqrt{R}$ for some fixed $\beta > 0$. There exists a constant $C(\beta) > 0$ depending only on $\beta$ such that \begin{align*} \big| \mathbb{E} W - \mathbb{E} Y_0 \big| \geq \frac{C(\beta)}{\sqrt{R}}. \end{align*} \end{proposition} We prove the proposition with the help of four auxiliary lemmas. The lemmas are proved at the end of this section after we prove Proposition~\ref{prop:lowerboundec}. We use $C = C(\beta)>0$ to denote a constant that may change from line to line, but does not depend on anything other than $\beta$. \begin{lemma} \label{lem:momentequiv} For any $\beta > 0$, \begin{align} \beta \mathbb{P}(W \geq \beta) = -\mathbb{E} \big( W 1(W < \beta)\big) \quad \text{ and } \quad \beta \mathbb{P}(Y_0 \geq \beta) = -\mathbb{E} \big( Y_0 1(Y_0 < \beta)\big). \label{eq:bizero} \end{align} Consequently, \begin{align*} \mathbb{E} W = \mathbb{E} (W-\beta)^{+} \quad \text{ and } \quad \mathbb{E} Y_0 = \mathbb{E} (Y_0-\beta)^{+}. \end{align*} \end{lemma} Lemma~\ref{lem:momentequiv} implies that \begin{align*} \abs{ \mathbb{E} W - \mathbb{E} Y_0} = \abs{ \mathbb{E} (W-\beta)^{+} - \mathbb{E} (Y_0-\beta)^{+}}. \end{align*} The next lemma rewrites the right-hand side above using the Poisson equation so that we can bound it from below. \begin{lemma} \label{lem:expliciterror} Fix $h \in \text{\rm Lip(1)}$ and let $f_h(x)$ be the solution the the Poisson equation \begin{align} b(x) f_h'(x) + v_0 f_h''(x) = \mathbb{E} h(Y_0) - h(x), \quad x \in \mathbb{R}. \label{eq:poissonv0} \end{align} Then $f_h'''(x-) = \lim_{y \uparrow x} f_h'''(y)$ is defined for all $x \in \mathbb{R}$ and \begin{align*} \mathbb{E} h(W) - \mathbb{E} h(Y_0) = - \frac{1}{2} \delta \mathbb{E} b(W) f_h''(W)+ \mathbb{E} \varepsilon(W), \end{align*} where \begin{align*} \varepsilon(W) =&\ \frac{1}{6}\delta^2 b(W) f_h'''(W-)+ \lambda(\varepsilon_{+}(W) + \varepsilon_{-}(W)) - \frac{1}{\delta} b(W) \varepsilon_{-}(W), \\ \varepsilon_{+}(W) =&\ \frac{1}{2}\int_{W}^{W+\delta} (W+\delta - y)^{2} (f_h'''(y) - f_h'''(W-)) dy, \\ \varepsilon_{-}(W) =&\ - \frac{1}{2}\int_{W-\delta}^{W} (y- (W-\delta))^{2} (f_h'''(y) - f_h'''(W-)) dy. \end{align*} \end{lemma} Our plan is to show that for any $h \in \text{\rm Lip(1)}$, the term $\mathbb{E} \varepsilon(W)$ vanishes at a rate of at least $\delta^2$. We then fix $h(x) = (x-\beta)^{+}$ and show that $\abs{\mathbb{E} b(W) f_h''(W)}$ can be bounded away from zero by a constant independent of $\delta$, which implies Proposition~\ref{prop:lowerboundec}. The following two lemmas are needed for this. The first one is for the upper bound on $\mathbb{E} \varepsilon(W)$, and the second is for the lower bound on $\abs{\mathbb{E} b(W) f_h''(W)}$. \begin{lemma} \label{lem:higherders} Assume that $n = R + \beta \sqrt{R}$ for some fixed $\beta > 0$. There exists $C = C(\beta) > 0$ depending only on $\beta$ such that for any $h \in \text{\rm Lip(1)}$, \begin{align*} \abs{ f_h^{'''}(x-)} \leq \frac{C}{\mu } \quad \text{ and } \quad \mathbb{E} \abs{b(W)} \leq \mu \mathbb{E} \abs{W} \leq&\ \mu C , \quad x \in \mathbb{R}. \end{align*} Additionally, if $h(x) = (x-\beta)^{+}$, then for all $x \neq \beta$, $f_h^{(4)}(x)$ exists and $\abs{f_h^{(4)}(x)} \leq \frac{C}{\mu }(1 + \abs{x})$. \end{lemma} \begin{lemma} \label{lem:secondder} If $h(x) = (x-\beta)^{+}$, then $f_h''(x) = \frac{1}{\mu \beta}$ for $x \geq \beta$ and \begin{align*} f_h''(x) =\frac{1}{\mu \beta} \frac{1 + x e^{x^2/2} \int_{-\infty}^{x} e^{-y^2/2} dy }{1 + \beta e^{\beta^2/2} \int_{-\infty}^{\beta} e^{-y^2/2} dy } , \quad x \leq \beta. \end{align*} \end{lemma} Before proving Proposition~\ref{prop:lowerboundec}, let us remark that we restrict ourselves to $h(x) = (x-\beta)^{+}$ to keep the proof simple. Our arguments can likely be extended to work for other $h(x)$ at the expense of added complexity. \proof{Proof of Proposition~\ref{prop:lowerboundec}} Lemma~\ref{lem:expliciterror} implies that \begin{align*} \mathbb{E} h(W) - \mathbb{E} h(Y_0) = - \frac{1}{2} \delta \mathbb{E} b(W) f_h''(W)+ \mathbb{E} \varepsilon(W). \end{align*} In the first part of the proof, we show that $\mathbb{E} \abs{\varepsilon(W)} \leq C \delta^2$. For convenience, we recall that \begin{align*} \varepsilon(W) =&\ \frac{1}{6}\delta^2 b(W) f_h'''(W-)+ \lambda(\varepsilon_{+}(W) + \varepsilon_{-}(W)) - \frac{1}{\delta} b(W) \varepsilon_{-}(W), \\ \varepsilon_{+}(W) =&\ \frac{1}{2}\int_{W}^{W+\delta} (W+\delta - y)^{2} (f_h'''(y) - f_h'''(W-)) dy, \\ \varepsilon_{-}(W) =&\ - \frac{1}{2}\int_{W-\delta}^{W} (y- (W-\delta))^{2} (f_h'''(y) - f_h'''(W-)) dy. \end{align*} Lemma~\ref{lem:higherders} immediately implies that $\mathbb{E} \abs{b(W) f_h'''(W-)} \leq C$. Next, we show that $\mathbb{E} \abs{\varepsilon_{+}(W)} \leq \frac{C}{\mu} \delta^4 $. By considering the cases when $W = \beta$ and $W \neq \beta$ and using Lemma~\ref{lem:higherders}, we see that \begin{align*} \mathbb{E} \abs{\varepsilon_{+}(W)} \leq&\ \frac{1}{2}\mathbb{P}(W = \beta)\bigg|\int_{\beta}^{\beta+\delta} (\beta+\delta - y)^{2} (\abs{f_h'''(y)} + \abs{f_h'''(\beta-)}) dy\bigg| \\ &+ \frac{1}{2}\mathbb{E} \Bigg[1(W \neq \beta)\bigg|\int_{W}^{W+\delta} (W+\delta - y)^{2}\int_{W}^{y} \abs{f_h^{(4)}(u)} du dy \bigg|\Bigg]\\ \leq&\ \frac{C}{\mu}\mathbb{P}(W = \beta)\bigg|\int_{\beta}^{\beta+\delta} (\beta+\delta - y)^{2} dy\bigg|\\ &+\frac{C}{\mu}\mathbb{E} \Bigg[1(W \neq \beta) \bigg|\int_{W}^{W+\delta} \delta (1+\abs{W}+\delta) (W+\delta - y)^{2} dy \bigg|\Bigg]\\ \leq&\ \frac{C}{\mu} \delta^3 \mathbb{P}(W = \beta) + \frac{C}{\mu} \delta^4. \end{align*} This argument can be repeated to show $\mathbb{E} \abs{\varepsilon_{-}(W)} \leq \frac{C}{\mu} \delta^4$. It was shown in (3.29) of \cite{Brav2017} that $\mathbb{P}(W = \beta)\leq C \delta$, but for completeness we repeat the argument at the end of the proof. Combining these results, we arrive at \begin{align*} \mathbb{E}\abs{\varepsilon(W)} \leq C \delta^2 + \lambda \frac{C}{\mu} \delta^4 + \mathbb{E} \abs{b(W)} \frac{C}{\mu} \delta^3 \leq C \delta^2, \end{align*} where in the last inequality we use the bound on $\mathbb{E} \abs{b(W)}$ from Lemma~\ref{lem:higherders} and the fact that $\delta^2 = 1/R = \mu/\lambda$. We now show that $\abs{\mathbb{E} b(W) f_h''(W)} \geq C$. Combining the fact that $b(x) = -(\mu x \wedge \mu \beta)$ with the form of $f_h''(x)$ from Lemma~\ref{lem:secondder}, we have \begin{align*} \mathbb{E} b(W) f_h''(W) =&\ \frac{-\mu \beta}{\mu \beta} \mathbb{P}(W \geq \beta) - \frac{\mu}{\mu \beta} \frac{\mathbb{E}\bigg[ W\Big(1+ W e^{W^2/2}\int_{-\infty}^{W} e^{-y^2/2} dy \Big) 1(W < \beta) \bigg]}{1 + \beta e^{\beta^2/2} \int_{-\infty}^{\beta} e^{-y^2/2} dy } \\ \leq&\ -\mathbb{P}(W \geq \beta) - \frac{1}{\beta} \frac{\mathbb{E}\big( W 1(W < \beta) \big)}{1 + \beta e^{\beta^2/2} \int_{-\infty}^{\beta} e^{-y^2/2} dy }. \end{align*} From Lemma~\ref{lem:momentequiv} we know that $\mathbb{E} \big( W 1(W < \beta)\big) = -\beta\mathbb{P}(W \geq \beta)$, so \begin{align*} \mathbb{E} b(W) f_h''(W) =&\ -\mathbb{P}(W \geq \beta) + \frac{1}{ \beta} \frac{ \beta \mathbb{P}(W \geq \beta) - \mathbb{E}\bigg[ \Big( W^2 e^{W^2/2}\int_{-\infty}^{W} e^{-y^2/2} dy \Big) 1(W < \beta) \bigg]}{1 + \beta e^{\beta^2/2} \int_{-\infty}^{\beta} e^{-y^2/2} dy } \\ \leq&\ \mathbb{P}(W \geq \beta)\Big( -1 + \frac{1}{1 + \beta e^{\beta^2/2} \int_{-\infty}^{\beta} e^{-y^2/2} dy} \Big)\\ \leq&\ -C \mathbb{P}(W \geq \beta). \end{align*} Proposition 1 of \cite{HalfWhit1981} tells us that $\mathbb{P}(W \geq \beta)$ converges to a positive constant (depending on $\beta$) as $R \to \infty$. This implies the lower bound on $\abs{\mathbb{E} b(W) f_h''(W)}$. Lastly, we prove that $\mathbb{P}(W = \beta)\leq C \delta$. Let $\phi_0(x)$ be the density of $Y_0$. Since $W$ is grid valued, for any $z \in (0,\delta)$ we have \begin{align*} \mathbb{P}(W = \beta) =&\ \mathbb{P}( \beta - z \leq W \leq \beta + z) \\ \leq&\ \mathbb{P}( \beta - z \leq Y_{0} \leq \beta + z) + \abs{\mathbb{P}( \beta - z \leq W \leq \beta + z) - \mathbb{P}( \beta - z \leq Y_0 \leq \beta + z)}\\ \leq&\ 2z \sup_{x \in \mathbb{R}} \phi_0(x) + 2\sup_{x \in \mathbb{R}} \abs{\mathbb{P}(W \leq x) - \mathbb{P}(Y_0 \leq x)}. \end{align*} To reach the desired conclusion, we use Lemma 7 and Theorem 3 of \cite{BravDaiFeng2016}. The former says that $\phi_0(x) \leq \sqrt{2/\pi}$, while the latter result says that $\sup_{x \in \mathbb{R}} \abs{\mathbb{P}(W \leq x) - \mathbb{P}(Y_0 \leq x)} \leq C \delta$. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{lem:momentequiv}.} \proof{Proof of Lemma~\ref{lem:momentequiv} } If $f(x) = x$, then $\mathbb{E} G_{\tilde X} f(W) = 0$ because $\mathbb{E} |W| < \infty$; see Lemma~2 of \cite{BravDaiFeng2016} for a proof of the latter fact. It follows from \eqref{eq:gxbarf} that \begin{align*} \lambda - \mathbb{E} \Big( \mu\big[ (W/\delta+ R) \wedge n \big] \Big) = \mathbb{E} b(W) = - \mu \mathbb{E}(W \wedge \beta) = 0, \end{align*} implying that \begin{align} -\beta \mathbb{P}(W \geq \beta) = \mathbb{E} \big( W 1(W < \beta)\big). \label{eq:leminterm} \end{align} Adding $ \mathbb{E} \big( W 1(W \geq \beta)\big)$ to both sides proves that $\mathbb{E} W = \mathbb{E} (W-\beta)^{+}$. The claim about $Y_0$ follows similarly. The density of $Y_0$ is given by \eqref{eq:stddenf}, so \begin{align*} \mathbb{E} b(Y_0) = \kappa \int_{-\infty}^{\infty} \frac{b(x)}{\mu} \exp\Big(\int_0^x \frac{b(y)}{\mu} dy \Big) dx = 0. \end{align*} The last equality follows from integration by parts and the fact that $\lim_{x\to \pm \infty} \exp\big(\int_0^x \frac{b(y)}{\mu} dy \big) = 0$. Therefore \begin{align*} \frac{1}{\mu} \mathbb{E} b(Y_0) = -\beta \mathbb{P}(Y_0 \geq \beta) - \mathbb{E} \big(Y_0 1(Y_0 < \beta)\big) = 0 \end{align*} and $\mathbb{E} Y_0 = \mathbb{E} (Y_0 -\beta)^{+}$. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{lem:expliciterror}.} \proof{Proof of Lemma~\ref{lem:expliciterror}} First, note that $f_h''(x) = -\frac{b(x)}{v_0} f_h'(x) + \frac{1}{v_0}( \mathbb{E} h(Y_0) - h(x))$ is differentiable almost everywhere because $f_h'(x)$ is continuously differentiable and both $b(x)$ and $h(x)$ are Lipschitz functions. The former statement follows, for example, from (B.1) of \cite{BravDaiFeng2016}. Therefore, $f_h'''(x)$ exists almost everywhere. Now \eqref{eq:gxbarf} implies that \begin{align*} \mathbb{E} G_{\tilde X} f_h(W) = \mathbb{E} \Big[ \lambda (f_h(W+\delta)-f_h(W))+\mu\big[ (W/\delta+ R) \wedge n \big] (f_h(W-\delta)-f_h(W)) \Big]=0, \end{align*} provided $\mathbb{E} \abs{f_h(W)} < \infty$. The integrability of $f_h(W)$ has already been established in \cite{BravDaiFeng2016}; see Lemma 1 and Remark 2 there. Since $f_h'''(x)$ does not exist everywhere (for instance at $x = \beta$), to perform Taylor expansion we need to use the integral form of the remainder term. We claim that \begin{align} f_h(x+\delta) - f_h(x) =&\ \delta f_h'(x) + \frac{1}{2}\delta^2 f_h''(x) + \frac{1}{6}\delta^{3} f_h'''(x-) + \varepsilon_+(x), \notag \\ f_h(x-\delta) - f_h(x) =&\ -\delta f_h'(x) + \frac{1}{2}\delta^2 f_h''(x) - \frac{1}{6}\delta^{3} f_h'''(x-) + \varepsilon_{-}(x). \label{eq:vareps} \end{align} To verify the claim, note that \begin{align*} f_h(x+\delta) - f_h(x) = \int_{x}^{x+\delta} f_h'(y) dy =&\ \delta f_h'(x) + \int_{x}^{x+\delta} (f_h'(y)-f_h'(x)) dy \\ =&\ \delta f_h'(x) + \int_{x}^{x+\delta} \int_{x}^{y} f_h''(u) du dy\\ =&\ \delta f_h'(x) + \int_{x}^{x+\delta} (x+\delta - u) f_h''(u) du. \end{align*} A similar treatment of $f_h(x+\delta) - f_h(x)$ yields \eqref{eq:vareps}. Letting $s(W) = \mu\big[ (W/\delta+ R) \wedge n \big]$, we therefore have \begin{align*} \mathbb{E} G_{\tilde X} f_h(W) =&\ \mathbb{E} \Big[ \delta (\lambda - s(W)) f_h'(W) + \frac{1}{2}\delta^2 (\lambda + s(W)) f_h''(W) \\ &+ \frac{1}{6}\delta^3 (\lambda - s(W)) f_h'''(W-) + \lambda \varepsilon_{+}(W) + s(W) \varepsilon_{-}(W)\Big]. \end{align*} We know from \eqref{eq:tabeff} that $\delta(\lambda - s(W)) = b(W)$ and $\delta^2(\lambda + s(W)) = a(W)$, and consequently $s(W) = \lambda - b(W) / \delta$. Therefore, \begin{align*} & \mathbb{E} G_{\tilde X} f_h(W) \\ =&\ \mathbb{E} \Big[b(W)f_h'(W) + \frac{1}{2} a(W) f_h''(W) + \frac{1}{6}\delta^2 b(W) f_h'''(W-) + \lambda (\varepsilon_{+}(W) + \varepsilon_{-}(W)) - \frac{1}{\delta} b(W) \varepsilon_{-}(W)\Big] = 0. \end{align*} Taking expected values in the Poisson equation \eqref{eq:poissonv0}, we get \begin{align*} \mathbb{E} h(Y_0) - \mathbb{E} h(W) =&\ \mathbb{E} \Big[ b(W) f_h'(W) + v_0 f_h''(W)\Big] - \mathbb{E} G_{\tilde X} f_h(W) \\ =&\ - \frac{1}{2}\mathbb{E} \big(a(W) - a(0)\big) f_h''(W) - \mathbb{E} \varepsilon(W). \end{align*} We conclude by noting that $a(W) - a(0) = -\delta b(W)$. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{lem:higherders}.} \proof{Proof of Lemma~\ref{lem:higherders}} To bound $\abs{f^{(4)}(x)}$, first note that the Poisson equation \eqref{eq:poissonv0} implies that \begin{align*} f_h'''(x) = -\frac{b(x)}{v_0} f_h''(x) -\frac{b'(x)}{v_0} f_h'(x) -\frac{1}{v_0} h'(x). \end{align*} Since $b(x)$ and $h(x)$ are piece-wise linear with a kink at $x = \beta$, the derivative above exists for all $x \neq \beta$. Differentiating again, we get \begin{align*} f_h^{(4)}(x) = -\frac{b(x)}{v_0} f_h'''(x) -\frac{b'(x)}{v_0} f_h''(x) -\frac{b'(x)}{v_0} f_h''(x) , \quad x \neq \beta. \end{align*} We conclude that $|f_h^{(4)}(x)| \leq (C/\mu)(1+\abs{x})$ for $x \neq \beta$ because $v_0 = \mu$, $\abs{b(x)} \leq \mu \abs{x}$, $\abs{b'(x)} \leq \mu$, $\abs{f_h'''(x)} \leq C/\mu$, and $\abs{f_h''(x)} \leq C/\mu $, where the last two inequalities follow from Lemma 3 of \cite{BravDaiFeng2016}. To conclude the proof, we note that $\mathbb{E} \abs{b(W)} \leq \mu\mathbb{E} |W| \leq \mu C$, where the last inequality follows from Lemma 2 of \cite{BravDaiFeng2016}, which tell us that $\mathbb{E} \abs{W} \leq C $. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{lem:secondder}.} \proof{Proof of Lemma~\ref{lem:secondder}} First assume that $x \geq \beta$. In (B.8) of \cite{BravDaiFeng2016} it is shown that \begin{align*} f_h''(x) =&\ e^{-\int_{0}^{x} \frac{b(u)}{v_0} du}\int_{x}^{\infty} \frac{1}{\mu}\big( h'(y) + f_h'(y) b'(y)\big) e^{\int_{0}^{y} \frac{b(u)}{v_0} du} dy = e^{-\int_{\beta}^{x} \frac{b(u)}{v_0} du}\int_{x}^{\infty} \frac{1}{\mu} e^{\int_{\beta}^{y} \frac{b(u)}{v_0} du} dy, \end{align*} where in the second equality we use $b'(x) = 0$ and $h'(x) = 1$ for $x \geq \beta$. Since $b(x)/v_0 = -\beta$ for $x \geq \beta$, \begin{align*} f_h''(x) = e^{\beta(x-\beta)} \int_{x}^{\infty} \frac{1}{\mu } e^{-\beta(y-\beta)} dy = \frac{1}{\mu \beta}. \end{align*} Now suppose that $x \leq \beta$. The Poisson equation \eqref{eq:poissonv0} and the fact that $h(x) = 0$ imply that \begin{align*} f_h''(x) =&\ -\frac{b(x)}{v_0} f_h'(x) + \frac{1}{v_0} \mathbb{E} h(Y_0) = x f_h'(x) + \frac{\mathbb{E} h(Y_0)}{\mu }. \end{align*} One can verify by differentiating that \begin{align*} f_h'(x) = e^{-\int_{0}^{x} \frac{b(u)}{v_0} du} \int_{-\infty}^{x} \frac{1}{v_0} \big(\mathbb{E} h(Y_0) - h(y) \big) e^{\int_{0}^{y} \frac{b(u)}{v_0} du} dy = \frac{\mathbb{E} h(Y_0)}{\mu }e^{ \frac{1}{2}x^2 } \int_{-\infty}^{x} e^{- \frac{1}{2} y^2} dy. \end{align*} The first equality appears as equation (B.1) in \cite{BravDaiFeng2016}. The second equality follows from the form of $b(x)$ in \eqref{eq:adeff} and the fact that $h(x) = 0$ for $x \leq \beta$. Lastly, since the density of $Y_0$ is given by \eqref{eq:stddenf}, we have \begin{align*} \mathbb{E} h(Y_0) = \frac{\int_{-\infty}^{\infty} h(y) e^{\int_{0}^{y} \frac{b(u)}{v_0} du} dy}{\int_{-\infty}^{\infty} e^{\int_{0}^{y} \frac{b(u)}{v_0} du}dy} =&\ \frac{\int_{\beta}^{\infty} (y-\beta)^{+} e^{\int_{\beta}^{y} \frac{b(u)}{v_0} du}dy}{\int_{-\infty}^{\beta} e^{\int_{\beta}^{y} \frac{b(u)}{v_0} du}dy + \int_{\beta}^{\infty} e^{\int_{\beta}^{y} \frac{b(u)}{v_0} du}dy}\\ =&\ \frac{\int_{\beta}^{\infty} (y-\beta) e^{-\beta(y-\beta)}dy}{\int_{-\infty}^{\beta} e^{-\frac{1}{2}(y^2-\beta^2) }dy + \int_{\beta}^{\infty} e^{-\beta(y-\beta)}dy} = \frac{1/\beta^2 }{ e^{\frac{1}{2}\beta^2 } \int_{-\infty}^{\beta} e^{-\frac{1}{2} y^2 }dy + 1/\beta}. \end{align*} This verifies the form of $f_h''(x)$ when $x \leq \beta$. \hfill $\square$\endproof \subsection{Proving Theorem~\ref{thm:md-high}}\label{fse8} We first recall Theorem~\ref{thm:md-high}. \begin{theorem} \label{thm:md-highec} Assume $n = R + \beta \sqrt{R}$ for some fixed $\beta > 0$. \blue{There exist positive constants $c_1$ and $C$ depending only on $\beta$ such that \begin{align} & \left|\frac{\mathbb{P}(Y_1\geq z)}{\mathbb{P}(W\geq z)}-1\right|\leq \frac{C}{\sqrt{R}}\left(1+\frac{z}{\sqrt{R}}\right) \quad \text{ for } 0< z\leq c_1 R\ \text{and} \label{f13ec}\\ &\left|\frac{\mathbb{P}(Y_1\leq -z)}{\mathbb{P}(W\leq -z)}-1\right|\leq \frac{C}{\sqrt{R}}\left(1+z+\frac{z^4}{\sqrt{R}}\right),\ \text{ for } 0<z\leq \min\{c_1 R^{1/4}, R^{1/2}\} \label{f16ec}. \end{align} } \end{theorem} We use \blue{$c_1,\ C,\ C_1, K$} to denote positive constants which may differ from line to line, but will only depend on $\beta$. We first prove \eqref{f16ec}, and then \eqref{f13ec}. \blue{ Note that if $R \leq K$ for some $K > 0$, then $n$ is also bounded because $\beta$ is fixed and $n = R + \beta \sqrt{R}$. We argue that \eqref{f16ec} holds trivially in such a case. Observe that for any $K > 0$, \magenta{because $\{W = -\sqrt{R}\}$ corresponds to an empty system,} \begin{align*} \inf_{\substack{0 < R \leq K \\ 0 < z \leq \sqrt{R}}} \mathbb{P}(W \leq -z) \geq \inf_{\substack{0 < R \leq K }} \mathbb{P}(W = -\sqrt{R}) \geq L(K,\beta) > 0, \end{align*} \magenta{where $L(K,\beta)$ is a positive constant depending only on $K$ and $\beta$. The second-last inequality is true because $\inf_{\substack{0 < R \leq K }} \mathbb{P}(W = -\sqrt{R})$ is at least as large as $\mathbb{P}(W = -\sqrt{R})$ when $R = K$, because for each $n$, the probability of an empty system decreases in $R$.} We can choose a sufficiently large $C$ (that depends on $K$) to ensure \eqref{f16ec} trivially holds. Therefore, in the following, we assume $R\geq C_1$ for a sufficiently large $C_1$. Since \eqref{f16ec} requires $ 0 < z \magenta{\leq} c_1 R^{1/4}$, we can also assume that \ben{\label{fEC4} \delta = \frac{1}{\sqrt{R}} <\min\{1/2,\beta\}, \quad 0<z<\sqrt{R}-2,\quad \delta(z+1)<1/2. } If not, we simply increase the value of $C_1$ and decrease the value of $c_1$ until \eqref{fEC4} holds. Without loss of generality let us therefore assume \eqref{fEC4} going forward. } Given $z \in \mathbb{R}$, we let $f_z(x)$ be the solution \magenta{(cf. \eqref{f18})} to the Poisson equation \ben{\label{f17} \frac{1}{2}a(x)f_z''(x)+b(x)f_z'(x)=\P(Y_1 \leq -z) - 1(x\leq -z), \quad x \in \mathbb{R}. } The following object will be of use. Define\magenta{, for $W$ in its support \eqref{f29},} \begin{align} K_W(y) = \begin{cases} (\lambda - b(W)/\delta) (y+\delta) \geq 0, \quad y \in [-\delta, 0], \\ \lambda(\delta-y) \geq 0, \quad y \in [0,\delta]. \end{cases} \label{eq:kdef} \end{align} It can be checked that \begin{align} \int_{-\delta}^{0} K_W(y) dy =&\ \frac{1}{2} \delta^2 \lambda -\frac{1}{2}\delta b(W), \quad \int_{0}^{\delta} K_W(y) dy = \frac{1}{2}\delta^2 \lambda, \notag \\ \int_{-\delta}^{\delta} K_W(y) dy =&\ \frac{1}{2}a(W) = \mu - \frac{\delta}{2} b(W), \quad \text{ and } \quad \int_{-\delta}^\delta y K_W(y) dy=\frac{\delta^2 b(W)}{6}. \label{MD:k3} \end{align} Our first result is an expression for $\mathbb{P}(Y_1 \leq -z) - \mathbb{P}(W \leq -z)$, and is proved in Section~\ref{sec:prooftaylormdhigh}. \begin{lemma} \label{lem:taylormdhigh} For any $z \in \mathbb{R}$, \begin{align} &\mathbb{P}(Y_1 \leq -z) - \mathbb{P}(W \leq -z) \notag \\ =&\ \mathbb{E} \bigg[\int_{-\delta}^{\delta} \Big(\frac{2b(W+y)}{a(W+y)}f_z'(W+y) - \frac{2b(W)}{a(W)}f_z'(W) \Big) K_W(y) dy \bigg] \notag \\ &\magenta{-}\mathbb{E} \bigg[\int_{-\delta}^{\delta} \Big( \frac{2}{a(W)} 1(W\leq -z) - \frac{2}{a(W+y)}1(W+y\leq -z) \Big) K_W(y) dy \bigg] \notag \\ &\magenta{-}\mathbb{P}(Y_1 \leq -z)\mathbb{E} \bigg[\int_{-\delta}^{\delta} \Big( \frac{2}{a(W+y)} -\frac{2}{a(W)} \Big) K_W(y) dy \bigg]. \label{f20} \end{align} \end{lemma} The bulk of the effort to prove \eqref{f16ec} comes from the first term. The following lemma is proved in Section~\ref{fap4}. \begin{lemma}\label{l4} For $x \in \mathbb{R}$, define $r(x) = 2b(x)/a(x)$. There exists constants $c_1,C,C_1 > 0$ depending only on $\beta$ such that for any $R\geq C_1$ and $0<z\leq c_1 R^{1/4}$ satisfying \eqref{fEC4}, \besn{\label{f21} \left| \mathbb{E} \bigg[ \int_{-\delta}^\delta \big(r(W+y)f_z'(W+y)-r(W)f_z'(W)\big)K_W(y)dy \bigg] \right|\leq&\ C \delta^2 (z\vee 1)^4 \P(Y_1\leq -z). } \end{lemma} \proof{Proof of \eqref{f16ec}} We first bound the third term in \eqref{f20}. Using the form of $a(x)$ in \eqref{eq:adeff} and the assumption that $\delta < 1/2$ in \eqref{fEC4}, it is not hard to check that \ben{\label{fEC6} \mu \leq a(x) \leq 2\mu+\delta\mu \beta \leq C \mu, \quad \text{ and } \quad \abs{a'(x)} \leq \delta \mu, } from which it follows that \begin{align} \frac{1}{a(x)} \leq 1/\mu \quad \text{ and } \quad \abs{\frac{1}{a(x)} - \frac{1}{a(y)}} = \frac{\abs{a(y) - a(x)}}{a(y)a(x)} \leq \frac{\delta \abs{y-x}}{\mu}. \label{eq:alipsch} \end{align} Therefore, the third term in \eqref{f20} satisfies \begin{align} \P(Y_1\leq -z)\left| \mathbb{E} \bigg[\int_{-\delta}^{\delta} \Big( \frac{2}{a(W+y)} -\frac{2}{a(W)} \Big) K_W(y) dy \bigg] \right| \leq&\ \P(Y_1\leq -z)\left| \mathbb{E} \bigg[\frac{2\delta^2}{\mu } \int_{-\delta}^{\delta} K_W(y) dy \bigg] \right| \notag \\ =&\ \P(Y_1\leq -z)\left| \mathbb{E} \bigg[\frac{2\delta^2}{\mu } \frac{1}{2}a(W) \bigg] \right| \notag \\ \leq&\ C\delta^2 \P(Y_1\leq -z). \label{f24} \end{align} The equality is due to \eqref{MD:k3}. The second term in \eqref{f20} is bounded similarly. Namely, \begin{align} & \left| \mathbb{E} \bigg[\int_{-\delta}^{\delta} \Big( \frac{2}{a(W)} 1(W\leq -z) - \frac{2}{a(W+y)}1(W+y\leq -z) \Big) K_W(y) dy \bigg] \right| \notag \\ \leq &\ \left| \mathbb{E} \bigg[\int_{-\delta}^{\delta} \Big( \frac{2}{a(W)} - \frac{2}{a(W+y)} \Big)1(W\leq -z) K_W(y) dy \bigg] \right| \notag \\ &+\left| \mathbb{E} \bigg[\int_{-\delta}^{\delta} \frac{2}{a(W+y)}\Big( 1(W\leq -z) -1(W+y\leq -z) \Big) K_W(y) dy \bigg] \right| \notag \\ \leq &\ C \delta^2 \mathbb{P}(W \leq -z) + \frac{2}{\mu } \mathbb{E} \bigg[\int_{-\delta}^{\delta} \Big| 1(W\leq -z) -1(W+y\leq -z) \Big| K_W(y) dy \bigg]. \label{f22} \end{align} Letting $-\tilde{z}$ denote the smallest value in the support of $\{W: W > -z \}$ \magenta{(cf. \eqref{f29})}, we have \begin{align} \frac{2}{\mu } \mathbb{E} \bigg[\int_{-\delta}^{\delta} \Big| 1(W\leq -z) -1(W+y\leq -z) \Big| K_W(y) dy \bigg] \leq C \P(W=-\tilde{z})+ C \P(W=-\tilde{z}-\delta). \label{eq:interm30} \end{align} At the end of this proof we argue that \ben{\label{f23} \P(W=-\tilde{z})+ \P(W=-\tilde{z}-\delta)\leq C\delta (z\vee 1) \P(W\leq -z). } Combining the bounds in \magenta{\eqref{f24},} \eqref{f22}, \eqref{eq:interm30} and \eqref{f23} with Lemma~\ref{l4} yields \begin{align*} |\P(Y_1\leq -z)-\P(W\leq -z)|\leq C\delta^2 (z\vee 1)^4 \P(Y_1\leq -z)+C\delta (z\vee 1) \P(W\leq -z). \end{align*} Dividing both sides by $P(W\leq -z)$, which is allowed because our assumption that $z < \sqrt{R} -2$ \magenta{in \eqref{fEC4}} implies that $P(W\leq -z)\geq \P(W=-\sqrt{R})>0$, we arrive at \ben{\label{fEC5} \left|\frac{\P(Y_1\leq -z)}{\P(W\leq -z)}-1 \right|\leq C \bigg( \delta^2 (z\vee 1)^4 \frac{\P(Y_1\leq -z)}{\P(W\leq -z)}+\delta(z\vee 1) \bigg). } Since we assumed $z \leq c_1 R^{1/4}$ and $R \geq C_1$, it follows that $\delta^2(z\vee 1)^4 \leq c_1^4 \vee (1/\magenta{C_1})$, which can be made arbitrarily small (without affecting $C$ above) by decreasing $c_1$ and increasing $C_1$. Therefore, without loss of generality we can assume $C\delta^2(z\vee 1)^4<1/2$, so \begin{align*} \frac{1}{2}\frac{\P(Y_1\leq -z)}{\P(W\leq -z)} \leq (1-C\delta^2 (z\vee 1)^4)\frac{\P(Y_1\leq -z)}{\P(W\leq -z)}\leq 1+C\delta (z\vee 1) \leq C, \end{align*} where the second inequality is due to \eqref{fEC5} and the last inequality follows from \eqref{fEC4}. Combining the upper bound above with \eqref{fEC5} proves \eqref{f16ec}. It remains to verify \eqref{f23}. Equation \eqref{f26} implies that for $-y$ in the support of $W$, \begin{align*} \lambda \P(W=-y-\delta)=\mu[(-y/\delta+R)\wedge n] \P(W=-y). \end{align*} Since $\beta = \delta(n-R)$, the set $\{-y/\delta+R \leq n\}$ equals $\{-y \leq \beta\}$, so dividing both sides above by $\lambda$ and using the fact that $\mu/\lambda = 1/R$, it follows that \ben{\label{f27} \P(W=- y-\delta)=(1-\delta y) \P(W=- y) \quad \text{ for $-y\leq \beta$}. } Recall that $-\tilde{z}$ denotes the smallest value in the support of $\{W: W > -z \}$, so $-\tilde z - \delta \leq -z < -\tilde z$ and the set $\{W \leq -z\}$ equals $\{W \leq -\tilde z - \delta\}$. Also recall that we assumed in \eqref{fEC4} that $0 < z < \sqrt{R} -2$ and $\delta < \beta$. Therefore, \begin{align*} \P(W\leq -z) =&\ \P(W=-\tilde z-\delta)+\P(W=-\tilde z-2\delta)+\cdots +\P(W=-\sqrt{R})\\ =&\ \P(W=-\tilde z) \Big( (1-\delta \tilde z) +\big((1-\delta \tilde z)(1-\delta (\tilde z+\delta))\big)\\ &+\big((1-\delta \tilde z)(1-\delta (\tilde z+\delta))(1-\delta (\tilde z+2\delta))\big)+\cdots \Big)\\ \geq&\ \P(W=-\tilde z) \Big( (1-\delta \tilde z) +\big((1-\delta \tilde z)(1-\delta (\tilde z+\delta))\big)+\cdots\\ & +\big((1-\delta \tilde z)\cdots(1-\delta(\tilde z+ \lfloor \frac{1}{\delta}\rfloor \delta))\big) \Big). \end{align*} The second equality follows from \eqref{f27} with $\tilde z$ in place of $y$. This requires that $-\tilde z \leq \beta$, which follows from the fact that $-\tilde z - \delta \leq -z < 0$ and therefore $-\tilde z < \beta$ by our assumption that $\delta < \beta$ in \eqref{fEC4}. In the last inequality, $\lfloor \cdot \rfloor$ denotes the integer part and the inequality itself follows from the fact that $-(\tilde z+ \lfloor \frac{1}{\delta}\rfloor \delta)\geq -(\tilde z+1)>-(z+1)>-\sqrt{R}$. Now for any $0 \leq k \leq \lfloor \frac{1}{\delta}\rfloor$ we have $1 - \delta (\tilde z + \delta k) \geq 1 - \delta(\tilde z + 1) > 1 - \delta(z + 1) > 1/2$, where the last inequality follows from our assumption in \eqref{fEC4}. Therefore, the right-hand side above is bounded from below by \begin{align*} & \P(W=-\tilde z) \Big( \big(1-\delta (\tilde z+1)\big) +\big(1-\delta (\tilde z+1)\big)^2+\cdots +\big(1-\delta (\tilde z+1)\big)^{\lfloor \frac{1}{\delta}\rfloor +1}\Big)\\ =&\ \P(W=-\tilde z)\big(1-\delta (\tilde z+1)\big) \frac{1-\big(1-\delta (\tilde z+1)\big)^{\lfloor \frac{1}{\delta}\rfloor+1}}{\delta (\tilde z+1)} \\ \geq & \P(W=-\tilde z)\big(1-\delta (\tilde z+1)\big) \frac{1-(e^{-\delta (\tilde z+1)})^{\lfloor \frac{1}{\delta}\rfloor+1}}{\delta (\tilde z+1)} \\ \geq& \P(W=-\tilde z)[1-\delta ( z+1)] \frac{1-e^{- 1/2}}{\delta ( z+1)} \\ \geq& \P(W=-\tilde z) \frac{1-e^{- 1/2}}{2\delta ( z+1)}. \end{align*} The first inequality is true because $1-x \leq e^{-x}$ and $\delta(\tilde z+1)\geq \delta(z-\delta+1)>0$. The second inequality is true because $0<\delta(\tilde z+1)<\delta (z+1)$ and $\delta (\tilde z+1)(\lfloor 1/\delta\rfloor+1)>(\delta/2) (\lfloor 1/\delta\rfloor+1) > 1/2$ (because $\lfloor 1/\delta\rfloor+1 > 1/\delta$, $\delta < 1/2$ by \eqref{fEC4}, and $-\tilde z - \delta \leq -z < 0$ from it follows that $\tilde z > -\delta > -1/2$). The last inequality follows from $\delta(z+1)<1/2$ in \eqref{fEC4}. This implies that $\P(W=-\tilde z)\leq C\delta(z\vee 1) \P(W\leq \magenta{-z})$. Using $-\tilde z < \delta < 1/2$ and \eqref{f27} again we get \be{ \P(W=-\tilde z-\delta)=(1-\delta \tilde z) \P(W=-\tilde z)\leq (1+\delta^2) \P(W=-\tilde z)\leq \frac{5}{4} \P(W=-\tilde z), } This proves \eqref{f23}. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{lem:taylormdhigh}. } \label{sec:prooftaylormdhigh} \proof{Proof of Lemma~\ref{lem:taylormdhigh}} Lemma~\ref{fEC3}, which we state in Section~\ref{fap4}, implies that $f_z'(x)$ is bounded and absolutely continuous with bounded $f_z''(x)$. Lemma~\ref{lem:higherders} implies $\mathbb{E} \abs{W} < \infty$, which when combined with the fact that $f_z'(x)$ is bounded implies $\mathbb{E} \abs{f_z(W)} < \infty$, and in turn $\mathbb{E} G_{\tilde X} f_z(W) = 0$ due to \eqref{eq:gxbarf}. Letting \begin{align*} G_Y f(x)=\frac{1}{2}a(x)f''(x)+b(x) f'(x), \end{align*} taking expected values with respect to $W$ in the Poisson equation \eqref{f17} and subtracting $\mathbb{E} G_{\tilde X} f_z(W)=0$ from it, we get \ben{ \label{eq:f171} \mathbb{E} G_Y f_z(W) - \mathbb{E} G_{\tilde X} f_z(W) =\P(Y_1\leq -z) - \P(W\leq -z). } To prove the lemma we work on the left-hand side. Similar to \eqref{eq:vareps}, for any function $f: \mathbb{R} \to \mathbb{R}$ with an absolutely continuous derivative, and any $x,\delta \in \mathbb{R}$, \begin{align*} f(x+\delta) -f(x) = \delta f'(x) + \int_x^{x+\delta} (x+\delta -y)f''(y)dy= \delta f'(x) + \int_{0}^{\delta} (\delta -y)f''(x+y)dy. \end{align*} Applying this expansion to $G_{\tilde X}$ in \eqref{eq:gxbarf}, we get \begin{align*} G_{\tilde X} f_z(W) =&\ \delta \big( \lambda - \mu \big[(W/\delta+R)\wedge n\big]\big) f_z'(W) \\ &+ \lambda \int_{0}^{\delta} (\delta -y) f_z''(W+y) dy + \mu \big[(W/\delta+R)\wedge n\big] \int_{-\delta}^{0} (y+\delta)f_z''(W+y) dy. \end{align*} Recall from \eqref{eq:tabeff} that $b(W) =\delta \big( \lambda - \mu \big[(W/\delta+R)\wedge n\big]\big) $, and consequently $\mu \big[(W/\delta+R)\wedge n\big] = \lambda - b(W)/\delta$. We recall $K_W(y)$ from \eqref{eq:kdef}, so \begin{align} G_{\tilde X} f_z(W) =&\ b(W)f_z'(W) +\int_{-\delta}^{\delta} f_z''(W+y) K_{W}(y) dy \notag \\ =&\ b(W)f_z'(W) + \frac{1}{2} a(W) f_z''(W) + \int_{-\delta}^{\delta} \big(f_z''(W+y) - f_z''(W)\big) K_{W}(y) dy . \label{eq:steinx} \end{align} The last equality follows from $\int_{-\delta}^{\delta} K_W(y) dy = \frac{1}{2}a(W)$ in \eqref{MD:k3}. Thus, \begin{align} \mathbb{P}(Y_1 \leq -z) - \mathbb{P}(W \leq -z) =&\ \mathbb{E} G_{Y} f_z(W) - \mathbb{E} G_{\tilde X} f_z(W) \notag \\ =&\ \mathbb{E} \bigg[ \int_{-\delta}^{\delta} \big(f_z''(W) - f_z''(W+y)\big) K_{W}(y) dy\bigg]. \label{MD:exp1} \end{align} In the last equality above we used \eqref{eq:steinx}. The lemma follows from \eqref{f17}; i.e., $f_z''(x) = -\frac{2b(x)}{a(x)}f_z'(x) + \frac{2}{a(x)}\big(\mathbb{P}(Y_1 \leq -z ) - 1(x \leq -z) \big)$. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{l4}.} \label{fap4} We first present a series of intermediary lemmas that represent the main steps in the proof, and then use them to prove Lemma~\ref{l4}. We remind the reader that \eqref{eq:adeff} implies that \begin{align} r(x) = \begin{cases} -2x, \quad x \leq -1/\delta,\\ \frac{-2x}{2 +\delta x}, \quad x \in [-1/\delta, \beta], \\ \frac{-2\beta}{2 + \delta\beta}, \quad x \geq \beta, \end{cases} \quad r'(x) = \begin{cases} -2, \quad x \leq -1/\delta, \\ \frac{-4}{(2+\delta x)^2}, \quad x \in (-1/\delta, \beta], \\ 0, \quad x > \beta, \end{cases}\label{eq:rform} \end{align} where $r'(x)$ is interpreted as the derivative from the left at the points $x = -1/\delta, \beta$. In particular, note that $\abs{r'(x)} \leq 4$. The first lemma decomposes \begin{align*} \mathbb{E} \bigg[ \int_{-\delta}^\delta \big(r(W+y)f_z'(W+y)-r(W)f_z'(W)\big)K_W(y)dy \bigg] \end{align*} into a more convenient form. It is proved in Section~\ref{sec:errorexpproof}. \begin{lemma} \label{lem:error_expansion} Let $f_z(x)$ solve \eqref{f17}, then \besn{\label{26} & \int_{-\delta}^\delta \big(r(W+y)f_z'(W+y)-r(W)f_z'(W)\big)K_W(y)dy \\ =&\int_{-\delta}^\delta K_W(y) r(W+y) y f_z''(W) dy\\ &- \int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \int_0^s r(W+u)f_z''(W+u) du ds dy \\ &- \int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \int_0^s r'(W+u)f_z'(W+u) du ds dy \\ &-\int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \bigg[\frac{ 1(W+s\leq -z)}{a(W+s)/2}-\frac{ 1(W\leq -z)}{a(W)/2} \bigg]ds dy \\ &\magenta{+} \P(Y_1\leq -z) \int_{-\delta}^\delta K_W(y) r(W+y) \int_0^y \bigg[\frac{2}{a(W+s)}-\frac{2}{a(W)}\bigg] ds dy \\ &+1(W=-1/\delta) f_z'(W) \int_0^\delta K_W(y) \int_0^y r'(W+s) ds dy\\ &+1(W=\beta) f_z'(W) \int_{-\delta}^0 K_W(y) \int_0^y r'(W+s) ds dy\\ &+1(W\in [-1/\delta+\delta, \beta-\delta]) f_z'(W) \int_{-\delta}^\delta K_W(y) \int_0^y \int_0^s r''(W+u) duds dy\\ &+1(W\in [-1/\delta+\delta, \beta-\delta]) f_z'(W) r'(W) \frac{\delta^2 b(W)}{6}. } \end{lemma} Next we need a bound on $f_z'(x)$ and $f_z''(x)$ as provided by the following lemma. \begin{lemma}\label{fEC3} Let $f_z(x)$ solve \eqref{f17}. There exists a constant $C$ that depends only on $\beta$ such that for any $z > 0$, \begin{align} \abs{f_z'(x)} \leq&\ \frac{C}{\mu } \bigg(1(x \leq -z) \Big( \frac{\mu}{\abs{b(x)}} \wedge 1 \Big) \notag \\ & \qquad + \P(Y_1\leq -z) \Big( 1(-z < x < 0) e^{-\int_{0}^{x} r(u)du} + 1(x \geq 0)\Big( \frac{\mu}{\abs{b(x)}} \wedge 1 \Big) \Big) \bigg), \label{eq:mdleft1} \\ \abs{f_z''(x)} \leq&\ \frac{C}{\mu } \bigg(1(x \leq -z) + \P(Y_1\leq -z) \Big( 1(-z < x < 0)(1 + \abs{x}) e^{-\int_{0}^{x} r(u)du} + 1(x \geq 0) \Big) \bigg)\magenta{.} \label{eq:mdleft2} \end{align} \end{lemma} \proof{ Proof of Lemma \ref{fEC3} } We recall from \eqref{eq:stddenf} that the density of $Y_1$ is \begin{align*} \kappa\frac{2}{a(x)} \exp\Big({\int_0^x \frac{2b(y)}{a(y)}dy}\Big), \quad x \in \mathbb{R}, \end{align*} where $\kappa$ is the normalizing constant. One may verify that the solution to the Poisson equation $\frac{1}{2}a(x)f_z''(x)+b(x)f_z'(x)=\P(Y_1 \leq -z) - 1(x\leq -z)$ satisfies \ben{\label{f18} f_z'(x)= \begin{cases} -\P(Y_1\geq -z) e^{-\int_0^x r(u)du} \frac{1}{\kappa} \P(Y_1\leq x), & x\leq -z, \\ -\P(Y_1\leq -z) e^{-\int_0^x r(u)du} \frac{1}{\kappa} \P(Y_1\geq x), & x\geq -z\magenta{.} \end{cases} } Rearranging the Poisson equation and using \eqref{f18} yields \besn{\label{f19} f_z''(x)=&-r(x) f_z'(x) +\frac{2(\P(Y_1\leq -z) - 1(x\leq -z))}{a(x)}\\ =& \begin{cases} -\P(Y_1\geq -z) \left[ \frac{2}{a(x)} -\frac{r(x)}{\kappa} e^{-\int_0^x r(u) du} \P(Y_1\leq x) \right], & x\leq -z, \\ -\P(Y_1\leq -z) \left[ -\frac{2}{a(x)} -\frac{r(x)}{\kappa} e^{-\int_0^x r(u) du} \P(Y_1\geq x)\right], & x>-z. \end{cases} } We will start by proving the bound on $f_z''(x)$. The form of the density of $Y_1$ implies \begin{align*} \frac{1}{\kappa} \P(Y_1\leq x) = \int_{-\infty}^{x} \frac{2}{ a(y)} e^{\int_{0}^{y}r(u)du} dy. \end{align*} Since $b(x)$ is nonincreasing and $b(x) > 0$ when $x < 0$, then for $x < 0$, \begin{align} \frac{1}{\kappa} \P(Y_1\leq x) = \int_{-\infty}^{x} \frac{2}{ a(y)} e^{\int_{0}^{y}r(u)du} dy \leq \frac{1}{b(x)} \int_{-\infty}^{x} \frac{2 b(y)}{ a(y)} e^{\int_{0}^{y}r(u)du} dy = \frac{1}{b(x)} e^{\int_{0}^{x}r(u)du}, \label{eq:main1} \end{align} and since $b(x) < 0$ for $x > 0$, then for $x \magenta{>} 0$, \begin{align} \frac{1}{\kappa} \P(Y_1\geq x) \leq \frac{1}{b(x)} \int_{x}^{\infty} \frac{2 b(y)}{ a(y)} e^{\int_{0}^{y}r(u)du} dy = \frac{-1}{b(x)} e^{\int_{0}^{x}r(u)du}= \frac{ 1}{\abs{b(x)}} e^{\int_{0}^{x}r(u)du}. \label{eq:main2} \end{align} Applying \eqref{eq:main1} and \eqref{eq:main2} to the form of $f_z''(x)$ above, we get the desired upper bound when $x\leq -z$ or $x \geq 0$. When $-z < x < 0$, we use the fact that $\abs{r(x)} \leq C \abs{x}$ to get \begin{align*} \abs{f_z''(x)} \leq&\ \magenta{\frac{2}{a(x)}\P(Y_1\leq -z)+C} \P(Y_1\leq -z)\magenta{\abs{x}} e^{-\int_0^x r(u) du} \int_{x}^{\infty} \frac{2}{ a(y)} e^{\int_{0}^{y} r(u)du} dy \\ \leq&\ \frac{C}{\mu } \P(Y_1\leq -z)(1 + \abs{x}) e^{-\int_0^x r(u) du}. \end{align*} The last inequality follows from the facts that $a(x) \geq \mu$, and that $\int_{x}^{\infty} e^{\int_{0}^{y} r(u)du} dy$ can be bounded by a constant that depends only on $\beta$, which is evident from the form of $r(x)$. This establishes the bound on $f_z''(x)$. Repeating the same procedure with $f_z'(x)$ gives us the bound \begin{align*} \abs{f_z'(x)} \leq&\ C \bigg(1(x \leq -z) \frac{1}{\abs{b(x)}} + \P(Y_1\leq -z) \Big( 1(-z < x < 0) \magenta{\frac{1}{\mu}} e^{-\int_{0}^{x} r(u)du} + 1(x \geq 0)\frac{1}{\abs{b(x)}} \Big) \bigg). \end{align*} To conclude the proof, we require the following two inequalities from Lemma B.8 of \cite{Brav2017}: \begin{align} & e^{-\int_{0}^{x} r(u)du}\int_{-\infty}^{x} \frac{2}{ a(y)} e^{\int_{0}^{y}r(u)du} dy\leq \begin{cases} \frac{3}{ \mu }, \quad x \leq 0,\\ \frac{1}{\mu}e^{\beta^2} (3 + \beta), \quad x \in [0,\beta], \end{cases} \label{DS:fbound1}\\ & e^{-\int_{0}^{x}r(u)du}\int_{x}^{\infty} \frac{2}{ a(y)} e^{\int_{0}^{y} r(u)du} dy \leq \begin{cases} \frac{1}{\mu } \Big( 2 + \frac{1}{\beta} \Big), \quad x \in [0,\beta], \\ \frac{1}{\mu \beta}, \quad x \geq \beta. \end{cases} \label{DS:fbound2} \end{align} By repeating the bounding procedure discussed above, but using \eqref{DS:fbound1} and \eqref{DS:fbound2} in place of \eqref{eq:main1} and \eqref{eq:main2}, we arrive at \begin{align*} \abs{f_z'(x)} \leq&\ \frac{C}{\mu } \bigg(1(x \leq -z) + \P(Y_1\leq -z) \Big( 1(-z < x < 0) e^{-\int_{0}^{x} r(u)du} + 1(x \geq 0) \Big) \bigg), \end{align*} which establishes the desired bound on $f_z'(x)$. \hfill $\square$\endproof Lastly, we will need the following lemmas, which are proved in Section~\ref{fap5}. \begin{lemma}\label{fl2} There exist constant $c_1,C_1 > 0$ such that for any integer $k \geq 0$, some $C(k) > 0$, and any $R\geq C_1$ and $0<z\leq c_1 R^{1/4}$, \ben{\label{11} \mathbb{E}|1(W\leq -z) W^k|\leq C(k) (z\vee 1)^{k+1} \P(Y_1\leq -z), } The constants $c_1,C_1,C(k)$ depend on $\beta$. \end{lemma} \begin{lemma}\label{fl3} There exist constant $c_1,C_1 > 0$ such that for any integer $k \geq 0$, some $C(k) > 0$, and any $R\geq C_1$ and $0<z\leq c_1 R^{1/4}$, \ben{\label{14} \mathbb{E}|1(-z\leq W\leq 0)W^k e^{-\int_0^W r(u)du}|\leq C(k) (z\vee 1)^{k+1}, } The constants $c_1,C_1,C(k)$ depend on $\beta$. \end{lemma} We are now ready to prove Lemma~\ref{l4}. \proof{Proof of Lemma~\ref{l4}} Again, in the following, we C to denote a constant whose value may change from line to line but only depends on $\beta$. We take expected values on both sides of \eqref{26} and bound the terms on the right-hand side one line at a time. For the first line, we need to bound \bes{ \left| \mathbb{E} \bigg[ \int_{-\delta}^\delta K_W(y) r(W+y) y f_z''(W) dy \bigg] \right| \leq &\left| \mathbb{E} \bigg[ r(W) f_z''(W) \int_{-\delta}^\delta y K_W(y) dy \bigg]\right|\\ &+\left| \mathbb{E} \bigg[ f_z''(W) \int_{-\delta}^\delta y K_W(y) (r(W+y)-r(W)) dy \bigg] \right|. } Using \eqref{MD:k3}, it follows that \begin{align*} \left| \mathbb{E} \bigg[ r(W) f_z''(W) \int_{-\delta}^\delta y K_W(y) dy \bigg]\right| =&\ \left| \mathbb{E} \bigg[ r(W) f_z''(W) \frac{ \delta^2 b(W)}{6}\bigg]\right| \leq C \mu \delta^2 \mathbb{E} \abs{W^2 f_{z}''(W)}, \end{align*} where we used $\abs{r(x)} \leq C \abs{x}$ and $\abs{b(x)}\leq \mu \abs{x}$ in the inequality. Furthermore, since $\abs{a(x)} \leq C \mu$ and recalling from \eqref{eq:rform} that $\abs{r'(x)} \leq 4$, we have \begin{align*} \left| \mathbb{E} \bigg[ f_z''(W) \int_{-\delta}^\delta y K_W(y) (r(W+y)-r(W)) dy \bigg] \right| \leq&\ \magenta{ \mathbb{E} \bigg[ C |f_z''(W)|\delta^2 \int_{-\delta}^\delta K_W(y) dy \bigg] } \\ =&\ \magenta{ \mathbb{E} \bigg[ C |f_z''(W)|\delta^2 \frac{a(W)}{2} \bigg] } \\ \leq &\ C \mu \delta^2 \mathbb{E} \abs{f_z''(W)}. \end{align*} Therefore, \begin{align*} \left| \mathbb{E} \bigg[ \int_{-\delta}^\delta K_W(y) r(W+y) y f_z''(W) dy \bigg] \right| \leq C \mu \delta^2 \mathbb{E} \abs{(1 + W^2) f_z''(W)}. \end{align*} Applying the bound on $\abs{f_z''(x)}$ from Lemma~\ref{fEC3}, we see that the right-hand side above is bounded by \begin{align*} & \magenta{C\delta^2} \mathbb{E} \bigg((1 + W^2)1(W \leq -z) \\ &\qquad + \P(Y_1\leq -z)(1 + W^2) \Big( 1(-z < W < 0)(1 + \abs{W}) e^{-\int_{0}^{W} r(u)du} + 1(W \geq 0) \Big) \bigg)\\ \leq&\ C \P(Y_1\leq -z) \magenta{\delta^2} (z\vee 1)^4, \end{align*} where the inequality is due to Lemmas~\ref{fl2} and \ref{fl3} and the fact that $\mathbb{E} W^2 \leq C$, which was proved in Lemma A.1 of \cite{Brav2017}. Following the same argument, the second line \begin{align*} &\abs{\int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \int_0^s r(W+u)f_z''(W+u) du ds dy}\\ &\magenta{\leq C\mu\delta^2\mathbb{E} (1+W^2) \sup_{-\delta\leq u\leq \delta}|f_z''(W+u)|} \\ &\magenta{\leq C\delta^2 \mathbb{E} \bigg( (1+W^2)1(W\leq -z+\delta)} \\ &\magenta{\quad + \P(Y_1\leq -z) (1+W^2) \Big( 1(-z-\delta<W<\delta)(1+|W|)e^{\sup_{-\delta\leq u\leq \delta}(-\int_0^{W+u} r(v)dv)}+1(W\geq -\delta) \Big) \bigg),} \end{align*} \magenta{which} is also bounded by $C \P(Y_1\leq -z) \delta^2 (z\vee 1)^4$ from simple modifications of Lemmas~\ref{fl2} and \ref{fl3}. For the third line, we use the representation $f_z'(W+u) - f_z'(W) = \int_{0}^{u} f_z''(W+v) dv$ and the triangle inequality to get \begin{align*} &\left| \mathbb{E} \bigg[\int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \int_0^s r'(W+u)f_z'(W+u) du ds dy \bigg]\right| \\ \leq&\ \left| \mathbb{E} \bigg[\int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \int_0^s r'(W+u)\int_{0}^{u} f_z''(W+v) dvdu ds dy \bigg]\right| \\ &+ \left| \mathbb{E} \bigg[\int_{-\delta}^\delta K_W(y) f_z'(W) (r(W+y)-r(W))\int_0^y \int_0^s r'(W+u) du ds dy \bigg]\right| \\ &+ \left| \mathbb{E} \bigg[\int_{-\delta}^\delta K_W(y) f_z'(W) r(W) \int_0^y \int_0^s r'(W+u) du ds dy \bigg]\right| \\ \leq&\magenta{\ C \mathbb{E} \bigg[ (\abs{W} + \delta)\int_{-\delta}^\delta K_W(y) \int_0^y \int_0^s \int_{0}^{u} |f_z''(W+v)| dvdu ds dy \bigg]} \\ &\magenta{+ C \delta^3 \mathbb{E} \bigg[\abs{f_z'(W)} \int_{-\delta}^\delta K_W(y)dy \bigg] + C\delta^2 \mathbb{E} \bigg[\abs{r(W) f_z'(W)}\int_{-\delta}^\delta K_W(y) dy \bigg]}\\ \leq&\magenta{\ C \mathbb{E} \bigg[ (\abs{W} + \delta)\int_{-\delta}^\delta K_W(y) \int_0^y \int_0^s \int_{0}^{u} |f_z''(W+v)| dvdu ds dy \bigg]} \\ &\magenta{+ C \mu \delta^3 \mathbb{E} \abs{f_z'(W)} + C\delta^2 \mathbb{E} \abs{b(W) f_z'(W)} .} \end{align*} In the second inequality we used $\abs{r'(x)}\leq 4$ and the last inequality is due to \eqref{MD:k3}, that $r(x) = 2b(x)/a(x)$ and the fact that $\abs{a(x)} \leq C \mu$. The right-hand side can be bounded by $C\P(Y_1\leq -z) \delta^2 (z\vee 1)^4$ as follows. The first term on the right-hand side above can be bounded by repeating the procedure used to bound lines one and two of \eqref{26}. The second and third terms can be bounded by combining the bound on $f_z'(x)$ from Lemma~\ref{fEC3} with Lemmas~\ref{fl2} and \ref{fl3}. For the fourth line, repeating the arguments from \eqref{f22}, \eqref{eq:interm30} and \eqref{f23}, we have \bes{ &\left| \mathbb{E} \bigg[\int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \bigg[\frac{ 1(W+s\leq -z)}{a(W+s)/2}-\frac{ 1(W\leq -z)}{a(W)/2} \bigg]ds dy \bigg] \right|\\ \leq & \left| \mathbb{E} \bigg[ \int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \bigg[\frac{2}{a(W+s)}-\frac{2}{a(W)} \bigg]1(W\leq -z)ds dy \bigg] \right|\\ &+\left| \mathbb{E} \bigg[ \int_{-\delta}^\delta K_W(y) r(W+y)\int_0^y \bigg[\frac{2(1(W+s\leq -z)-1(W\leq -z))}{a(W+s)} \bigg]ds dy \bigg] \right|\\ \leq & C\delta^3 \magenta{\mathbb{E}(1+|W|) 1(W\leq -z)} + C\delta \magenta{(z\vee 1)} [\P(W=-\tilde{z})+\P(W=-\tilde{z}-\delta)]\\ \leq & \magenta{C\P(Y_1\leq -z) \delta^3 (z\vee 1)^2+ C\delta^2 (z\vee 1)^2 \P(W\leq -z)}\\ \leq & C\P(Y_1\leq -z) \delta^2 (z\vee 1)^3, } where in the last \magenta{two inequalities} we used Lemma~\ref{fl2}. The fifth line is bounded as follows. Using \eqref{MD:k3}, \eqref{eq:alipsch} and the fact that $\abs{r(x)} \leq C\abs{x}$, we get \bes{ &\left| \mathbb{E} \bigg[ \int_{-\delta}^\delta K_W(y) r(W+y) \P(Y_1\leq -z) \int_0^y \bigg[\frac{2}{a(W+s)}-\frac{2}{a(W)}\bigg] ds dy \bigg] \right|\\ \leq & C\P(Y_1\leq -z) \delta^3 \magenta{\mathbb{E}(1+|W|)}\leq C\P(Y_1\leq -z) \delta^3. } The sixth line is bounded as follows. Using $\abs{r'(x)} \leq 4$, \eqref{MD:k3}, and $\abs{1(W=-1/\delta) f_z'(W)} \leq 1(W=-1/\delta) C/\mu$, which follows from Lemma~\ref{fEC3} and the fact that $z < \sqrt{R}-2 < \sqrt{R} = 1/\delta$ from \eqref{fEC4}, we have \bes{ \left|\mathbb{E} \bigg[ 1(W=-1/\delta) f_z'(W) \int_0^\delta K_W(y) \int_0^y r'(W+s) ds dy\bigg]\right| \leq & C \delta \P(W=-1/\delta), } and now we argue that \begin{align} C \delta \P(W=-1/\delta) \leq C \delta^2 (z \vee 1)\P(Y_1 \leq -z). \label{eq:toargue} \end{align} For any integer $0 < k < R$, \magenta{from \eqref{f26}} \begin{align*} \P(W=-1/\delta) = \P(W = -\sqrt{R}) \leq \frac{1}{k} \sum_{i=0}^{k} \P(W = \delta(i-R)) = \frac{1}{k} \P(W \leq \delta (k-R)). \end{align*} Choose $k$ such that $\delta (k-R)$ is the largest element in the support of $W$ that is less than or equal to $-z$, and use the fact that $z \leq c_1 R^{1/4}$ to conclude that \begin{align*} \P(W=-1/\delta) \leq C \delta \P(W \leq -z). \end{align*} Using Lemma~\ref{fl2} implies \eqref{eq:toargue}. The seventh line is bounded as follows. Using $\abs{r'(x)} \leq 4$, \eqref{MD:k3}, and $\abs{1(W=\beta) f_z'(W)} \leq 1(W=\beta) C/\mu$, which follows from Lemma~\ref{fEC3}, we have \bes{ & \left| \mathbb{E} \bigg[1(W=\beta) f_z'(W) \int_{-\delta}^0 K_W(y) \int_0^y r'(W+s) ds dy\bigg]\right| \\ \leq & C \delta \P(W=\beta)\P(Y_1\leq -z)\leq C\delta^2 \P(Y_1\leq -z), } where we used $\P(W=\beta)\leq C\delta$, which was argued at the end of the proof of Proposition~\ref{prop:lowerboundec}. The eighth line is bounded as follows. Using $\abs{r''(x)} \leq C \delta$ and Lemma~\ref{fEC3}, \bes{ & \left| \mathbb{E} \bigg[1(W\in [-1/\delta+\delta, \beta-\delta]) f_z'(W) \int_{-\delta}^\delta K_W(y) \int_0^y \int_0^s r''(W+u) duds dy \bigg] \right| \\ \leq & C \delta^3 \P(W\leq -z) +C \delta^3 \P(Y_1\leq -z) \mathbb{E}|1(-z\leq W\leq 0) e^{-\int_0^W r(u)du}| +C \delta^3 \P(Y_1\leq -z)\\ \leq & C \delta^3 (z\vee 1) \P(Y_1\leq -z), } where in the last inequality we used \magenta{Lemmas~\ref{fl2} and \ref{fl3}}. The ninth line is bounded similarly. Using $\abs{r'(x)} \leq 4$ and the bound on $\abs{f_z'(x) b(x)} $ from Lemma~\ref{fEC3}, we have \bes{ & \left| E\{1(W\in [-1/\delta+\delta, \beta-\delta]) f_z'(W) r'(W) \frac{\delta^2 b(W)}{6}\right| \\ \leq&\ C \delta^2 \mathbb{E} \abs{W 1(W\leq -z)} +C \delta^2 \P(Y_1\leq -z) \mathbb{E}|1(-z\leq W\leq 0) W e^{-\int_0^W r(u)du}| +C \delta^2 \P(Y_1\leq -z) \\ \leq&\ C \delta^2 (z\vee 1)^2 \P(W\leq -z) \magenta{+C\delta^2 \P(Y_1\leq -z)} \leq C \delta^2 (z\vee 1)^3 \P(Y_1\leq -z). } The second inequality is due to \magenta{Lemmas~\ref{fl2} and \ref{fl3}. } The last inequality is due to Lemma~\ref{fl2} with $k = 0$ there. Combining the bounds proves Lemma~\ref{l4}. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{lem:error_expansion}.} \label{sec:errorexpproof} \proof{Proof of Lemma~\ref{lem:error_expansion}} Assume for now that for all $y \in (-\delta, \delta)$, \begin{align} &r(W+y)f_z'(W+y) - r(W)f_z'(W) \notag \\ =&\ y r(W+y)f_z''(W) - r(W+y)\int_{0}^{y}\int_{0}^{s} \Big(r(W+u) f_z''(W+u) + r'(W+u)f_z'(W+u) \Big)duds \notag \\ &- r(W+y)\int_{0}^{y}\Big(\frac{2}{a(W+s)} 1(W+s \leq -z) - \frac{2}{a(W)} 1(W \leq -z) \Big)ds \notag \\ &+ \mathbb{P}(Y_1 \leq -z)r(W+y)\int_{0}^{y}\Big(\frac{2}{a(W+s)} - \frac{2}{a(W)} \Big)ds + f_z'(W) \int_{0}^{y}r'(W+s)ds. \label{eq:interm} \end{align} We postpone verifying \eqref{eq:interm} to the end of this proof, but \eqref{eq:interm} implies \begin{align*} &\int_{-\delta}^{\delta} \Big(r(W+y)f_z'(W+y) - r(W)f_z'(W) \Big) K_W(y) dy \notag \\ =&\ \int_{-\delta}^{\delta} K_W(y) y r(W+y)f_z''(W) dy \\ &- \int_{-\delta}^{\delta} K_W(y) r(W+y)\int_{0}^{y}\int_{0}^{s} r(W+u) f_z''(W+u) duds dy \\ &- \int_{-\delta}^{\delta} K_W(y) r(W+y)\int_{0}^{y}\int_{0}^{s} r'(W+u)f_z'(W+u) duds dy\\ &- \int_{-\delta}^{\delta} K_W(y) r(W+y) \int_{0}^{y}\Big(\frac{2}{a(W+s)} 1(W+s \leq -z) - \frac{2}{a(W)} 1(W \leq -z) \Big)ds dy\\ &+ \mathbb{P}(Y_1\magenta{\leq -z}) \int_{-\delta}^{\delta}K_W(y)r(W+y)\int_{0}^{y}\Big(\frac{2}{a(W+s)} - \frac{2}{a(W)} \Big)ds dy \\ &+ f_z'(W)\int_{-\delta}^{\delta}K_W(y) \int_{0}^{y}r'(W+s)dsdy. \end{align*} We are almost done, but the last term on the right-hand side above requires some additional manipulations. Since $r'(x) = 0$ for $x > \beta$ and $K_W(y)=0$ for $W = -1/\delta$ and $y \in [-\delta, 0]$, \begin{align*} &\int_{-\delta}^{\delta} K_W(y)\int_{0}^{y}r'(W+s)ds dy \\ =&\ 1(W = -1/\delta)\int_{0}^{\delta} K_W(y)\int_{0}^{y}r'(W+s)ds dy\\ &+ 1(W = \beta) \int_{-\delta}^{0} K_W(y)\int_{0}^{y}r'(W+s)ds dy \\ &+ 1(W \in [-1/\delta + \delta, \beta - \delta]) \int_{-\delta}^{\delta} K_W(y)\int_{0}^{y}r'(W+s)ds dy, \end{align*} and for $W \in [-1/\delta + \delta, \beta - \delta]$, \begin{align*} &\int_{-\delta}^{\delta} K_W(y)\int_{0}^{y}r'(W+s)ds dy \\ =&\ \int_{-\delta}^{\delta} K_W(y)\int_{0}^{y} \big(r'(W+s) - r'(W)\big)ds dy +r'(W) \int_{-\delta}^{\delta} yK_W(y)dy \\ =&\ \int_{-\delta}^{\delta} K_W(y)\int_{0}^{y} \int_{0}^{s}r''(W+u)duds dy + r'(W) \frac{1}{6}\delta^2b(W). \end{align*} To conclude the proof, we verify \eqref{eq:interm}: \begin{align*} &r(W+y)f_z'(W+y) - r(W)f_z'(W) \notag \\ =&\ r(W+y)f_z'(W) + r(W+y) \int_{0}^{y} f_z''(W+s) ds - r(W)f_z'(W) \notag \\ =&\ r(W+y) \int_{0}^{y} f_z''(W+s) ds \magenta{+} f_z'(W) \int_{0}^{y} r'(W+s) ds. \end{align*} Now \begin{align*} \int_{0}^{y} f_z''(W+s)ds =&\ yf_z''(W) + \int_{0}^{y} \big(f_z''(W+s) - f_z''(W) \big)ds \\ =&\ y f_z''(W) - \int_{0}^{y}\Big(r(W+s)f_z'(W+s) - r(W)f_z'(W) \Big)ds \\ &- \int_{0}^{y}\Big(\frac{2}{a(W+s)} 1(W+s \leq -z) - \frac{2}{a(W)} 1(W \leq -z) \Big)ds\\ &+ \mathbb{P}(Y_1 \leq -z)\int_{0}^{y}\Big(\frac{2}{a(W+s)} - \frac{2}{a(W)} \Big)ds, \end{align*} and the fundamental theorem of calculus tells us that \begin{align*} &r(W+s)f_z'(W+s) - r(W)f_z'(W)\\ =&\ \int_{0}^{s} \Big(r(W+u) f_z''(W+u) + r'(W+u)f_z'(W+u) \Big)du, \end{align*} which proves \eqref{eq:interm}. \hfill $\square$\endproof \subsubsection{Proving Lemmas~\ref{fl2} and \ref{fl3}.}\label{fap5} We first prove an upper bound on $E e^{-tW}$ for $0\leq t\leq c_1 R^{1/4}$. Note that $\mathbb{E} e^{-t W} < \infty$ for $ t \geq 0$ because $W\geq -\sqrt{R}$ (cf. \eqref{f29}). \begin{lemma}\label{l1} There exist $c_1,C_1>0$ such that if $R\geq C_1$ and $0\leq t\leq c_1 R^{1/4}$, then \ben{\label{04} \mathbb{E} e^{-tW}\leq C e^{\frac{t^2}{2}-\frac{\delta t^3}{6}}. } \end{lemma} \proof{Proof of Lemma \ref{l1}} For $0\leq s\leq t$, define $h(s)=\mathbb{E} e^{-sW}$, so $h'(s)= -\mathbb{E} \big(W e^{-sW}\big)$. We will shortly prove that \ben{\label{05} \Big(1+\frac{\delta s}{2}+ \frac{\delta^2 s^2}{6}\Big) h'(s)\leq (s+C\delta^2 s^3) h(s). } from which we have \begin{align*} h'(s)\leq \Big(\frac{s}{1+\frac{\delta s}{2}+ \frac{\delta^2 s^2}{6}}+\frac{C \delta^2 s^3}{1+\frac{\delta s}{2}+ \frac{\delta^2 s^2}{6}}\Big) h(s) \leq&\ \Big(\frac{s}{1+\frac{\delta s}{2}}+C \delta^2 t^3\Big) h(s) \\ \leq&\ \Big(s\big(1-\delta s/2 + (\delta s/2)^2 \big) +C\delta^2 t^3\Big)h(s) \\ \leq&\ \Big(s-\delta s^2/2 +C\delta^2 t^3\Big)h(s). \end{align*} Above we have used $0 \leq s \leq t$. The \magenta{third} inequality follows from the inequality $1/(1+x) \leq (1-x+x^2)$ for $x \geq 0$. Since $h(0) = 1$, \be{ \log (h(t)) = \int_{0}^{t} \frac{h'(s)}{h(s)} ds \leq \int_0^t (s-\frac{\delta s^2}{2}+C\delta^2 t^3) ds\leq \frac{t^2}{2}-\frac{\delta t^3}{6}+C\delta^2 t^4\leq \frac{t^2}{2}-\frac{\delta t^3}{6}+\magenta{C}c_1^4, } which implies \eqref{04}. We are left to prove \eqref{05}. Recall from \eqref{eq:gxbarf} that $\mathbb{E} G_{\magenta{\tilde X}} f(W) = 0$ provided $\mathbb{E} \abs{f(W)} < \infty$. Choose $f(x) = -e^{-sx}/s$, so that $f'(x) = e^{-sx}$ and $f''(x) = -s e^{-sx}$. The form of $G_{\magenta{\tilde X}} f(x)$ in \eqref{eq:steinx} implies \begin{align} \mathbb{E} \bigg[ \frac{a(W)}{2}f''(W)+b(W)f'(W)+\int_{-\delta}^\delta (f''(W+y)-f''(W)) K_W(y) dy \bigg] =0 \label{eq:identstein} \end{align} Using the form of $b(x)$ from \eqref{eq:adeff}, we have \besn{\label{06} \mathbb{E}[b(W)f'(W)] =&\ \mathbb{E}[1(W\leq \beta) (-\mu W) e^{-sW}+1(W>\beta)(-\mu \beta) e^{-sW}]\\ =&\mu h'(s)+\mathbb{E}[1(W>\beta)\mu(W-\beta)e^{-sW}]\\ \geq & \mu h'(s). } Similarly, using the form of $a(x)$ from \eqref{eq:adeff} and the fact that $W \geq -1/\delta$, we have \besn{\label{07} -\mathbb{E}\Big[\frac{a(W)}{2}f''(W)\Big] =&\ \mu \mathbb{E}\Big[1(-\frac{1}{\delta}\leq W\leq \beta) (1+\frac{\delta W}{2})s e^{-sW}+1(W>\beta)(1+\frac{\delta \beta}{2})s e^{-sW}\Big]\\ =&\ \mu \mathbb{E}\Big[(1+\frac{\delta W}{2})s e^{-sW}\Big]\magenta{-}\mu \mathbb{E}\Big[1(W>\beta) \frac{\delta (W-\beta)}{2}s e^{-sW}\Big]\\ \leq&\ \mu \mathbb{E}\Big[(1+\frac{\delta W}{2})s e^{-sW}\Big]=\mu sh(s)-\mu \frac{\delta s}{2}h'(s). } Lastly, \begin{align*} -\mathbb{E} \bigg[ \int_{-\delta}^\delta (f''(W+y)-f''(W)) K_W(y) dy\bigg] =&\ \mathbb{E} \bigg[\int_{-\delta}^\delta s (e^{-s(W+y)}-e^{-sW}) K_W(y)dy\bigg] \nonumber\\ \leq&\ \mathbb{E} \bigg[se^{-sW} \int_{-\delta}^\delta (-sy+s^2 y^2e^{\abs{sy}} )K_W(y)dy\bigg] \\ \leq&\ \mathbb{E} \bigg[se^{-sW} \int_{-\delta}^\delta (-sy+ C s^2 y^2 )K_W(y)dy\bigg]. \end{align*} The first inequality is due to $e^{-x}-1 \leq -x + x^2 e^{\abs{x}}$, and the second inequality is due to $|sy|\leq c_1 R^{1/4}\frac{1}{\sqrt{R}}\leq C$. Recall from \eqref{MD:k3} that $\int_{-\delta}^\delta K_W(y)dy = \frac{1}{2} a(W) $ and that $\int_{-\delta}^\delta y K_W(y)dy = \frac{\delta^2 b(W)}{6}$, and from \eqref{fEC6} that $\abs{a(x)} \leq C \mu$. Also note from \eqref{eq:adeff} that $b(x) = -\mu x - 1(x > \beta)(\mu \beta - \mu x)$. Therefore, the right-hand side above is bounded by \begin{align} & \mathbb{E} \bigg[s^2e^{-sW} \frac{-\delta^2 b(W)}{6}\bigg]+ C\mu \delta^2 s^3 Ee^{-sW} \nonumber\\ =&\ \frac{\delta^2 s^2}{6} \mathbb{E} \big[\mu W e^{-sW}\big] +\frac{\delta^2 s^2}{6} \mathbb{E} \Big[1(W>\beta) (\mu \beta - \mu W) e^{-sW}\Big]+ C\mu \delta^2 s^3 \mathbb{E} e^{-sW}\nonumber\\ =&\ -\mu \frac{\delta^2 s^2}{6} h'(s) +\frac{\delta^2 s^2}{6} \mathbb{E} \Big[1(W>\beta) (\mu \beta - \mu W) e^{-sW}\Big]+C\mu \delta^2 s^3 h(s)\nonumber\\ \leq&\ -\mu \frac{\delta^2 s^2}{6} h'(s)+C\mu \delta^2 s^3 h(s).\label{08} \end{align} Combining \eqref{06}--\eqref{08} with \eqref{eq:identstein} concludes the proof. \hfill $\square$\endproof Now we are ready to prove Lemmas~\ref{fl2} and \ref{fl3}. \proof{Proof of Lemma~\ref{fl2}} Suppose $k \geq 1$. We prove the lemma by showing that \begin{align} \mathbb{E}|1(W\leq -z) W^k| \leq&\ C(k) (z \vee 1)^{k} e^{-\frac{z^2}{2}-\frac{\delta z^3}{6}}, \notag \\ \text{ and } \quad C \frac{1}{z} e^{-\frac{z^2}{2}-\frac{\delta z^3}{6}} \leq&\ \mathbb{P}(Y_1 \leq -z). \label{eq:l1part1} \end{align} Let us start with the first inequality. \magenta{For $0<z<1$, the first inequality holds because of Lemma \ref{l1}. Therefore, we only need to consider the case $z\geq 1$.} Integration by parts yields \be{ \mathbb{E}|1(W\leq -z) W^k|=z^k \P(W\leq -z)+\int_{-\infty}^{-z} k (-y)^{k-1} \P(W\leq y) dy. } For the first term, note that \ben{\label{12} z^k\P(W\leq -z) \leq z^{k} \mathbb{E}\Big( e^{z(-z -W)} 1(W \leq -z) \Big) \leq z^k e^{-z^2} \mathbb{E} e^{-zW}\leq Cz^k e^{-\frac{z^2}{2}-\frac{\delta z^3}{6}}, } where the last inequality is due to Lemma~\ref{l1}. For the second term, we have \bes{ \int_{-\infty}^{-z} k (-y)^{k-1} \P(W\leq y) dy \magenta{\leq}&\ \int_{-\infty}^{-z} k (-y)^{k-1} \mathbb{E}\Big( e^{z(y-W)} 1(W \leq y) \Big) dy\\ \leq&\ \int_{-\infty}^{-z} k (-y)^{k-1} e^{zy} \mathbb{E} e^{-zW} dy\\ \leq&\ C e^{\frac{z^2}{2}-\frac{\delta z^3}{6}} \int_{-\infty}^{-z} k (-y)^{k-1}e^{zy} dy\\ \leq&\ C(k) z^{k-2} e^{-\frac{z^2}{2}-\frac{\delta z^3}{6}}. } The second-last inequality is due to Lemma~\ref{l1} and in the last inequality we used $\int_{-\infty}^{-z} (-y)^{k-1}e^{zy}dy\leq C(k)z^{k-2}e^{-z^2}$ for $k\geq 1$ \magenta{and $z\geq 1$}. This proves the first inequality in \eqref{eq:l1part1}, and we now argue the second one. Recall that $r(x) = 2b(x)/a(x)$ and that the density of $Y_1$ in \eqref{eq:stddenf} implies \bes{ \P(Y_1\leq -z)= \int_{-\infty}^{-z} \frac{2\kappa}{a(y)} e^{\int_{0}^{y} r(u) du} dy \geq \int_{-z-1}^{-z} \frac{2\kappa}{a(y)} e^{\int_{0}^{y} r(u) du} dy = &\ \int_{-z-1}^{-z} \frac{2\kappa}{a(y)} e^{\int_{0}^{y} \frac{-2u}{2 + \delta u} du} dy. } The last equality follows from the assumption $z \leq \sqrt{R} - 2 $ in \eqref{fEC4} so that $-z-1 \geq -\sqrt{R} $, meaning $a(x) = 2\mu - \delta b(x)$ for $x \in [-z-1, \magenta{0}]$. We now argue that $2\kappa/a(x) \geq C$ for some constant $C> 0 $ that depends only on $\beta$. The density of $Y_1$ is given by \eqref{eq:stddenf}, so we know that the normalizing constant $\kappa$ satisfies \begin{align*} \frac{1}{\kappa} =&\ \int_{-\infty}^{\infty} \frac{2}{a(x)} e^{\int_{0}^{x} 2b(y)/a(y) dy} dx \\ =&\ \int_{-\infty}^{0} \frac{2}{a(x)} e^{\int_{x}^{0} -2\abs{b(y)}/a(y) dy} dx + \int_{0}^{\infty} \frac{2}{a(x)} e^{\int_{0}^{x} -2\abs{b(y)}/a(y) dy} dx \\ \leq &\ \frac{C}{\mu} \int_{-\infty}^{0} e^{\int_{x}^{0} -\frac{2\abs{b(y)}} {\mu (2+\delta \beta)} dy} dx + \frac{C}{\mu} \int_{0}^{\infty} e^{\int_{0}^{x} -\frac{2\abs{b(y)}} {\mu (2+\delta \beta)} dy} dx. \end{align*} The second equality is true because $b(x)$ in \eqref{eq:adeff} is nonincreasing and $b(0) = 0$, meaning $b(x) = -\abs{b(x)} 1(x \geq 0) + \abs{b(x)} 1(x < 0)$. The inequality is true because $a(x) \geq \mu$ and is \magenta{nondecreasing} with $ a(x) \leq a(\beta) = \mu(2 + \delta \beta)$. Since $b(x)/\mu$ is a function that depends only on $\beta$, the right-hand side is a quantity that increases in $\delta$. In \eqref{fEC4} we assumed $\delta < 1/2$, so $ 1/\kappa \leq \sup_{ \delta \in (0,1/2)} 1/\kappa \leq C/\mu$. Combining this with the fact that $a(x) \leq C\mu$ from \eqref{fEC6} we get $2\kappa/a(x) \geq C$. Therefore, \bes{ \P(Y_1\leq -z) \geq &\ C \int_{-z-1}^{-z} e^{\int_{0}^{y} \frac{-2u}{2 + \delta u} du} dy = C\int_{-z-1}^{-z} e^{ \frac{4\log(\delta y+2) -\magenta{2} \delta y}{\delta^2} -\frac{4\log(2)}{\delta^2}} dy . } Taylor expansion tells us that \begin{align} \frac{4\log(\delta y+2) - \magenta{2}\delta y}{\delta^2} -\frac{4\log(2)}{\delta^2} = \frac{-y^2}{2} + \frac{\delta y^3}{6} - \frac{\delta^2 \xi^4}{16} \label{eq:logtaylor} \end{align} for some $\xi$ between $0$ and $y$. Recall that $\delta = 1/\sqrt{R}$ and $z\leq c_1 R^{1/4}$, meaning $\delta^2 \abs{\xi}^4 \leq C$ when $y \in [-z-1,-z]$, so \bes{ \P(Y_1\leq -z) \geq C\int_{-z-1}^{-z} e^{-\frac{y^2}{2}+\frac{\delta y^3}{6}}dy =&\ C\int_{-1}^0 e^{-\frac{(-z+y)^2}{2}+\frac{\delta(-z+y)^3}{6}}dy. } Now for $y \in [-1,0]$, and $z\leq c_1 R^{1/4}$ we use the fact that $\delta z^2 \leq C$ to get \begin{align*} -\frac{(-z+y)^2}{2}+\frac{\delta(-z+y)^3}{6} =&\ -\frac{z^2}{2}+zy - \frac{y^2}{2}+ \delta\frac{z^2y}{2} - \delta \frac{z y^2}{2} + \delta \frac{y^3}{6} -\frac{\delta z^3}{6}\\ \geq&\ -\frac{z^2}{2}+zy - \frac{y}{2}(y - \delta z^2) -\frac{\delta z^3}{6} - C\\ \geq&\ -\frac{z^2}{2}+zy -\frac{\delta z^3}{6} - C, \end{align*} so \besn{\label{f30} \P(Y_1\leq -z) \geq C\int_{-1}^0 e^{-\frac{z^2}{2}+zy-\frac{\delta z^3}{6}}dy \geq C \frac{1}{z}e^{-\frac{z^2}{2}-\frac{\delta z^3}{6}}. } This proves \eqref{eq:l1part1} for $k\geq 1$. For $k=0$, the lemma follows from \eqref{12} and \eqref{f30}. \hfill $\square$\endproof \proof{Proof of Lemma~\ref{fl3}} \magenta{The bound \eqref{14} trivially holds if $0<z<1$. Therefore, we assume $z\geq 1$ in the following.} Using integration by parts, we have for $k\geq 1$, \besn{\label{15} &\mathbb{E}|1(-z\leq W\leq 0)W^k e^{-\int_0^W r(u)du}|\\ =&\ -z^k e^{-\int_0^{\magenta{-z}} r(u)du} \P(W\leq -z)+\int_{-z}^0 \big(k(-y)^{k-1}+(-y)^kr(y)\big)e^{-\int_0^y r(u)du}\P(W\leq y) dy\\ \leq&\ C(k)\int_{-z}^0 \big((-y)^{k-1}+(-y)^{k+1}\big)e^{-\int_0^y r(u)du}\P(W\leq y) dy. } The last inequality is due to $r(y) = 2b(y)/a(y) = 2\mu(-y)/a(y) \leq C (-y)$ when $y \leq 0$, because $a(y) \geq \mu$. The same argument tells us that for $k=0$, \ben{\label{16} \mathbb{E}[1(-z\leq W\leq 0) e^{-\int_0^W r(u)du}]\leq \magenta{1+}C\int_{-z}^0 (-y)e^{-\int_0^y r(u)du}\P(W\leq y) dy. } In what follows we prove that for $k\geq 0$, \ben{\label{13} \int_{-z}^0 (-y)^k e^{-\int_0^y r(u)du}\P(W\leq y) dy\leq Cz^{k\vee 1}. } Lemma~\ref{fl3} follows from \eqref{15}, \eqref{16} and \eqref{13}. Without loss of generality assume that $z$ is an integer. If not, increase it to the nearest integer. The taylor expansion in \eqref{eq:logtaylor} implies that for $0\leq -y\leq z \leq c_1 R^{1/4}$, \be{ -\int_0^y r(u)du = -\int_0^y \frac{2b(u)}{a(u)} du = -\frac{4\log(\delta y+2) - \magenta{2}\delta y}{\delta^2} + \frac{4\log(2)}{\delta^2} \leq \frac{y^2}{2}-\frac{\delta y^3}{6}+C. } Thus, \bes{ &\int_{-z}^0 (-y)^k e^{-\int_0^y r(u)du}\P(W\leq y) dy\\ \leq &C\sum_{j=-z}^{-1} |j|^k \int_j^{j+1} e^{\frac{y^2}{2}-\frac{\delta y^3}{6}} e^{|j|y}e^{-|j|y} \P(W\leq y) dy\\ \leq &C \sum_{j=-z}^{-1} |j|^k \sup_{j\leq y\leq j+1} [e^{\frac{y^2}{2}+|j|y}] \sup_{j\leq y\leq j+1} [e^{-\frac{\delta y^3}{6}}] \int_{j}^{j+1} e^{-|j|y} \P(W\leq y) dy\\ \leq & C \sum_{j=-z}^{-1} |j|^k e^{-\frac{j^2}{2}} \sup_{j\leq y\leq j+1} [e^{-\frac{\delta y^3}{6}}] \int_{-\infty}^{\infty} e^{-|j|y} \P(W\leq y) dy\\ =&C \sum_{j=-z}^{-1} |j|^k e^{-\frac{j^2}{2}} \sup_{j\leq y\leq j+1} [e^{-\frac{\delta y^3}{6}}] \frac{1}{|j|} E e^{-|j| W}. } We used $\sup_{j\leq y\leq j+1} [e^{\frac{y^2}{2}+|j|y}]=e^{-\frac{j^2}{2}+\frac{1}{2}}$ in the last inequality and integration by parts in the last equality. Invoking Lemma~\ref{l1}, we have \bes{ \int_{-z}^0 (-y)^k e^{-\int_0^y r(u)du}\P(W\leq y) dy \leq&\ C \sum_{j=-z}^{-1} |j|^{k-1} e^{-\frac{j^2}{2}} \sup_{j\leq y\leq j+1} [e^{-\frac{\delta y^3}{6}}] e^{\frac{j^2}{2}} e^{-\frac{\delta |j|^3}{6}}\\ \leq&\ C \sum_{j=-z}^{-1} |j|^{k-1} \leq C z^{k\vee 1}, } where we used \be{ \sup_{j\leq y\leq j+1} [-\frac{\delta y^3}{6} -\frac{\delta |j|^3}{6}]\leq C \delta j^2\leq C, \ \text{for}\ 1\leq -j\leq z \leq c_1 R^{1/4}. } This proves \eqref{13}. \hfill $\square$\endproof \subsubsection{Proof of \eqref{f13ec}.} \proof{Proof of \eqref{f13ec}.} Inequality \eqref{f13ec} contains an upper bound on $\left|\frac{\mathbb{P}(Y_1\geq z)}{\mathbb{P}(W\geq z)}-1\right|$. \magenta{A} similar upper bound on $\left|\frac{\mathbb{P}(W\geq z)}{\mathbb{P}(Y_1\geq z)}-1\right|$ is a consequence of Theorem~4.1 of \cite{Brav2017}. The following simple modification of the argument in \cite{Brav2017} implies \eqref{f13ec}. It follows from (4.8), (4.9), (4.11), and (4.12) of \cite{Brav2017} that there exist some $c_1,C_1 > 0$ such that for $R\geq C_1$ and $0<z\leq c_1 R$, \bes{ &|P(Y_1\geq z)-P(W\geq z)|\\ \leq& C \delta^2 \P(Y_1\geq z)+C\delta^2 \P(W\geq z)+C\delta^2 \min\{(z\vee 1), \frac{1}{\delta^2}\}\P(Y_1\geq z)+C\delta \P(W\geq z). } Dividing both sides by $\P(W\geq z)$ we have \be{ \left|\frac{\P(Y_1\geq z)}{\P(W\geq z)}-1 \right|\leq C\delta^2 (z\vee 1) \frac{\P(Y_1\geq z)}{\P(W\geq z)}+C\delta. } Note that $C\delta^2 (z\vee 1) \leq C(\delta^2 z + \delta^2) = C(z/R + 1/R) \leq C(c_1 + 1/C_1) $, and choose $c_1$ small enough and $C_1$ large enough so that $C(c_1 + 1/C_1) < 1/2$. Then, repeating the argument used below \eqref{fEC5} implies \magenta{\eqref{f13ec} for $R\geq C_1$, and the argument above \eqref{fEC4} implies \eqref{f13ec} for $R<C_1$.} \hfill $\square$\endproof \subsection{Proof of Theorem~\ref{thm:md-std}}\label{fap3} We first recall Theorem~\ref{thm:md-std}. \begin{theorem} \label{thm:md-stdec} Assume $n = R + \beta \sqrt{R}$ for some fixed $\beta > 0$. \blue{There exist positive constants $c_0$ and $C$ depending only on $\beta$ such that \begin{align} & \left|\frac{\mathbb{P}(Y_0\geq z)}{\mathbb{P}(W\geq z)}-1\right|\leq \frac{C}{\sqrt{R}}\left(1+z\right) \quad \text{ for } 0<z\leq c_0 R^{1/2}\ \text{and} \label{f12ec} \\ &\left|\frac{\mathbb{P}(Y_0\leq -z)}{\mathbb{P}(W\leq -z)}-1\right|\leq \frac{C}{\sqrt{R}}\left(1+z^3\right),\ \text{ for } 0<z\leq \min\{c_0 R^{1/6}, R^{1/2}\}. \label{f15ec} \end{align} } \end{theorem} Theorem~\ref{thm:md-stdec} follows from a similar and simpler proof than that of Theorem~\ref{thm:md-highec}. In particular, the bound \eqref{f15ec} can be proved by a simple adaption of the arguments in \cite{ChFaSh13a} and therefore its proof is omitted. The proof of \eqref{f12ec} below may be useful for other exponential approximation problems. We \magenta{will use $c_0,\ C,\ C_0$ to denote positive constants} that may differ from line to line, but will only depend on $\beta$. We require the following two lemmas. The first one is proved in Section~\ref{fap7}, and the second in Section~\ref{fap6}. \begin{lemma}\label{l2} There exist $C,C_0 > 0$ that depend only on $\beta$, such that for any $R \geq C_0$ and any $z > \beta$, \ben{\label{5} \P(W\geq \magenta{z+\delta})\leq C e^{-(\beta- 3\beta^2 \delta)z}. } \end{lemma} To state the second lemma we let $f_z(x)$ solve the Poisson equation \ben{\label{8} b(x) f_z'(x) + \mu f_z''(x)= 1(x\geq z) - \P(Y_0\geq z). } \begin{lemma}\label{lem:eq9} There exist $C_0,C > 0$ depending on $\beta$ such that for all $R \geq C_0$ \magenta{and any $z>\beta$}, \begin{align} &\mathbb{E} \Big( \int_{-\delta}^{\delta} \big( b(W+y)f_z'(W+y) - b(W) f_z'(W)\big) \frac{K_W(y)}{\mu } dy \Big) \notag \\ \leq&\ C\delta \P(Y_0\geq z) \mathbb{E} \Big( \sup_{|s|\leq \delta}\big(1+e^{\beta (W+s)}1(\beta \leq W+s \leq z)\magenta{\big)\Big)}, \label{eq:lem91} \end{align} where $K_W(y)$ is defined in \eqref{eq:kdef}. Furthermore, \begin{align} & - \frac{\delta}{2\mu } \mathbb{E} \big( b^2(W) f_z'(W)\big) \leq C\delta \Big(\P(W\geq z)+\P(Y_0\geq z)\big(1+\mathbb{E} e^{\beta W}1(\beta\leq W\leq z) + |W|\big) \Big), \label{eq:lem92}\\ &- \frac{\delta}{2\mu } \mathbb{E} \big( b(W)\big( \mathbb{P}(Y_0 \geq z) -1(W \geq \delta + z) \big)\big) \leq C\delta \P(W\geq z+\delta)\magenta{.} \label{eq:lem93} \end{align} \end{lemma} \proof{Proof of \eqref{f12ec}} Note that if $R$ is bounded, then, for fixed $\beta$, $n$ is also bounded and $P(W\geq z)$ is bounded away from 0 for $z$ in the bounded range $0<z\leq c R^{1/2}$. By choosing a sufficiently large $C$, \eqref{f12ec} trivially holds. Therefore, in the following, we assume $R\geq C_0$ for a sufficiently large $C_0$. Additionally, the result for finite range $z \in (0, \beta+\delta]$ follows from the Berry-Esseen bound in Theorem 3 of \cite{BravDaiFeng2016}, so we assume $z > \beta + \delta$. Recall that the density of $Y_0$ is given in \eqref{eq:stddenf}, and that $b(x) =-(\mu x \wedge \mu \beta)$. Just like in \eqref{f18}, \magenta{one may verify that the solution to the Poisson equation \eqref{8} satisfies} \ben{\label{f33} f_z'(x)= \begin{cases} -\P(Y_0\geq z)e^{-\int_{0}^{x} \frac{b(u)}{\mu} du} \int_{-\infty}^{x} \frac{1}{\mu } e^{\int_{0}^{y} \frac{b(u)}{\mu} du} dy , & x<z, \\ -\P(Y_0\leq z)e^{-\int_{0}^{x} \frac{b(u)}{\mu} du} \int_{x}^{\infty} \frac{1}{\mu } e^{\int_{0}^{y} \frac{b(u)}{\mu} du} dy, & x\geq z, \end{cases} } \magenta{so} \eqref{eq:gxbarf} implies $\mathbb{E} G_{\magenta{\tilde X}} f_z(W) = 0$. Taking expected values in \eqref{8} then gives us \begin{align*} &\P(W\geq z) - \P(Y_0\geq z) \\ =&\ \mathbb{E}\Big( b(W) f_z'(W) + \mu f_z''(W)\Big) - \mathbb{E} G_{\magenta{\tilde X}} f_z(W) \\ =&\ \mathbb{E}\bigg(\mu f_z''(W)- \int_{-\delta}^{\delta}f_{\magenta{z}}''(W+y) K_W(y) dy \bigg) \\ =&\ \mathbb{E}\big( 1(W\geq z) - \P(Y_0\geq z) -b(W)f_z'(W)\big)\\ &-\mathbb{E}\bigg(\int_{-\delta}^{\delta} \big( 1(W+y\geq z) - \P(Y_0\geq z) - b(W+y) f_z'(W+y)\big) \frac{1}{\mu}K_W(y) dy\bigg), \end{align*} where $K_W(y)$ is defined in \eqref{eq:kdef}, the second equality is due to \eqref{eq:steinx}, and the final equality follows from $\mu f_z''(x) = - b(x) f_z'(x) + 1(x \geq z) - \mathbb{P}(Y_0 \geq z)$. Using $K_W(y) \geq 0$ and recalling from \eqref{MD:k3} that $\int_{-\delta}^{\delta} K_W(y) dy = \mu - \frac{\delta}{2}b(W)$, we have \begin{align*} &b(W)f_z'(W) = \int_{-\delta}^{\delta}b(W) f_z'(W) \frac{K_W(y)}{\mu } dy + \frac{\delta}{2\mu } b^2(W) f_z'(W), \\ &-\mathbb{E}\bigg(\int_{-\delta}^{\delta}1(W+y\geq z) \frac{1}{\mu}K_W(y) dy\bigg) \leq -\mathbb{E}\Big(1(W \geq z+\delta) \big(1 - \frac{\delta}{2 \mu} b(W)\big) \Big),\\ &\mathbb{E}\bigg(\int_{-\delta}^{\delta} \P(Y_0\geq z) \frac{1}{\mu}K_W(y) dy\bigg) = \P(Y_0\geq z) \mathbb{E}\Big( 1 - \frac{\delta}{2 \mu} b(W) \Big), \end{align*} and therefore, \begin{align} &\P(W\geq z) - \P(Y_0\geq z) \notag \\ \leq&\ \P(z \leq W < z+\delta) + \mathbb{E} \Big( \int_{-\delta}^{\delta} \big( b(W+y)f_z'(W+y) - b(W) f_z'(W)\big) \frac{K_W(y)}{\mu } dy \Big) \notag \\ &- \frac{\delta}{2\mu } \mathbb{E} \big( b^2(W) f_z'(W)\big) - \frac{\delta}{2\mu } \mathbb{E} \big( b(W)\big( \mathbb{P}(Y_0 \geq z) -1(W \geq \delta + z) \big)\big)\magenta{.} \label{10} \end{align} Subtracting $\P(z \leq W < z+\delta)$ from both sides and using Lemma~\ref{lem:eq9} to bound the remaining three terms, we get \begin{align*} &\P(W\geq z+\delta) - \P(Y_0\geq z) \notag \\ \leq&\ C\delta\sup_{|s|\leq \delta}\Big( \P(\magenta{W}\geq z)+ \P(Y_0\geq z)\big(1+\mathbb{E}|\magenta{W}|+\mathbb{E} e^{\beta(W+s)}1(\beta\leq W+s\leq z)\big) \Big). \end{align*} We now argue that the right-hand side above can be bounded by $ C\delta \P(Y_0\geq z) (z\vee 1)$. Note that $\mathbb{E} \abs{W} \leq C$ due to Lemma~\ref{lem:higherders}. Next, since $b(x)/\mu = -(x \wedge \beta)$ only depends on $\beta$ and $z \geq \beta$, then $\mathbb{P}(Y_0 \geq z) = \frac{\int_{z}^{\infty} e^{-\beta y} dy}{\int_{-\infty}^{\infty} e^{\int_{0}^{y} b(u)/\mu du }} = C e^{-\beta z}$ for some $C>0$ depending only on $\beta$. Thus, Lemma~\ref{l2} tells us that for \magenta{$R\geq C_0$ and $\beta+\delta<z\leq c_0 R^{1/2}$}, \be{ \P(\magenta{W}\geq z) = \frac{ \P(\magenta{W}\geq z)}{\P(Y_0\geq z)}\P(Y_0\geq z)\leq C\P(Y_0\geq z). } Moreover, \magenta{for $|s|\leq \delta$,} \bes{ &\mathbb{E} \big(e^{\beta(W+s)}1(\beta\leq W+s\leq z)\big)\\ &\magenta{= e^{\beta^2}\P(\beta\leq W+s\leq z) +\int_\beta^z \beta e^{\beta y} \P(y<W+s\leq z) dy } \\ &\leq C+\beta\int_{\beta}^z e^{\beta u} \P(W+s\geq u)du\\ &\leq C+ \int_{\beta}^z C du\\ &\leq C(z\vee 1). } The \magenta{first} equality follows from integration by parts, and the second-last inequality is due to Lemma~\ref{l2}. Thus we have shown that \be{ \P(W\geq z+\delta)-\P(Y_0\geq z)\leq C\delta \P(Y_0\geq z) (z\vee 1). } Note that \magenta{$|\P(Y_0\geq z)/\P(Y_0\geq z+\delta)-1|=|e^{\beta \delta}-1|\leq \delta C$}. Therefore, \be{ \P(W\geq z+\delta)-\P(Y_0\geq z+\delta)(1+\delta C)\leq C\delta \P(Y_0\geq z+\delta)(1+\delta C)(z\vee 1). } Dividing both sides of the above inequality by $\P(W\geq z+\delta)$, we obtain \be{ 1-\frac{\P(Y_0\geq z+\delta)}{\P(W\geq z+\delta)}(1+\delta C)\leq C\delta \frac{\P(Y_0\geq z+\delta)}{\P(W\geq z+\delta)}(1+\delta C)(z\vee 1). } Choose $C_0$ large enough and $c_0$ small enough and \magenta{since} $R\geq C_0$ and $0<z\leq c_0 R^{1/2}$, the constant in front of $\frac{\P(Y_0\geq z+\delta)}{\P(W\geq z+\delta)}$ on the right-hand side can be made less than \magenta{1/2, implying } \be{ 1-\frac{\P(Y_0\geq z+\delta)}{\P(W\geq z+\delta)} \leq C\delta (z\vee 1). } A similar argument can be repeated to show $\frac{\P(Y_0\geq z+\delta)}{\P(W\geq z+\delta)} -1 \leq C\delta (z\vee 1)$, implying \eqref{f12ec}. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{l2}.}\label{fap7} We require the following auxiliary result, whose proof is provided after the proof of Lemma~\ref{l2}. \begin{lemma}\label{l3} There exist constants $C,C_0 > 0$ depending only on $\beta$ such that for $R \geq C_0$, \ben{\label{6} \mathbb{E} e^{(\beta-3\beta^2 \delta)W}\leq C/\delta. } \end{lemma} \proof{Proof of Lemma~\ref{l2}} Since $\delta = 1/\sqrt{R}$, we can choose $C_0$ large enough so that $(\beta-3\beta^2 \delta)>\beta/2$. For notational convenience, we set $\nu = 3\beta^2 \delta$. Define \be{ f''(x)= \begin{cases} (\beta-\nu)e^{(\beta-\nu)x}, & x\leq z\\ \text{linear interpolation}, & z<x\leq z+\Delta\\ 0, & x>z+\Delta, \end{cases} } with $0<\Delta\leq \delta$ to be chosen, $f'(x) = 1 + \int_{0}^{x} f''(y) dy$ and $f(x) = \int_{0}^{x} f'(y) dy$. Note that $f(x)$ grows linearly in $x$ when $x \geq z + \Delta$ because $f'(x) = f'(z+\Delta)$ for $x \geq z + \Delta$, so $\mathbb{E} \abs{f(W)} < \infty$ because $W$ is bounded from below and $\mathbb{E} \abs{W} < \infty$. Therefore, $\mathbb{E} G_{\magenta{\tilde X}} f(W) = 0$ due to \eqref{eq:gxbarf}, implying $ \mathbb{E}\big(-b(W)f'(W)\big) = \mathbb{E} \Big(\int_{-\delta}^{\delta} f''(W+y) K_{W}(y) dy\Big)$ if we use the form of $G_{\magenta{\tilde X}} f(x)$ in \eqref{eq:steinx}. Note also that $f'(0) = 1$, $f'(x) \geq 0$, and $f'(x) \geq e^{(\beta-\nu)z}$ for $x \geq z + \Delta$. Using $b(x) = -(\mu x \wedge \mu \beta)$ \magenta{and the assumption that $z>\beta$} we therefore have \bes{ \mathbb{E}\big(-b(W)f'(W)\big) \geq&\ \mathbb{E}\big(\mu We^{(\beta-\nu)W}1(W<\beta)\big) \\ &\quad + \mathbb{E}\big(\mu \beta e^{(\beta-\nu)W}1(\beta\leq W\leq z)\big) +\mathbb{E}\big(\mu \beta e^{(\beta-\nu)z}1(W> z+\Delta)\big)\\ \geq&\ -\mu C+ \mathbb{E}\big(\mu \beta e^{(\beta-\nu)W}1(\beta\leq W\leq z)\big) +\mathbb{E}\big(\mu \beta e^{(\beta-\nu)z}1(W> z+\Delta)\big), } where the second inequality is due to $-\abs{x} e^{-(\beta-\nu)\abs{x}} 1(x < \beta) \geq -C$. Recalling from \eqref{MD:k3} that $\int_{-\delta}^{\delta} K_{W}(y)dy = \mu-\delta b(W)/2 $, we have \bes{ \mathbb{E} \Big(\int_{-\delta}^{\delta} f''(W+y) K_{W}(y) dy\Big) &\leq\mathbb{E}\Big( \sup_{|s|\leq \delta} f''(W+s) \int_{-\delta}^{\delta} K_{W}(y)dy \Big)\\ &=\mathbb{E} \Big( \mu \sup_{|s|\leq \delta} f''(W+s) \big(1-\frac{\delta}{2\mu}b(W)\big)1(W\leq z+\Delta+\delta)\Big)\\ &\leq \mathbb{E}\Big( \mu (\beta-\nu) e^{(\beta-\nu)(W+\delta)}\big(1-\frac{\delta}{2\mu}b(W)\big)1(W\leq z+\Delta+\delta)\Big)\\ &=\mathbb{E}\Big( \mu (\beta-\nu) e^{(\beta-\nu)(W+\delta)}\big(1+\frac{\delta}{2}\beta\big)1(\beta\leq W\leq z+\Delta+\delta)\Big)\\ &\quad +\mathbb{E}\Big( \mu (\beta-\nu) e^{(\beta-\nu)(W+\delta)}\big(1+\frac{\delta}{2}W\big)1(W<\beta )\Big)\\ &\leq \mathbb{E}\Big( \mu (\beta+C\delta) e^{(\beta-\nu)W}1(\beta\leq W\leq z)\Big)\\ &\quad +\mathbb{E} \Big( \mu (\beta+C\delta)e^{(\beta-\nu)z}1(z\leq W\leq z+\Delta+\delta)\Big)+ \mu C. } Combining the inequalities above, we have \bes{ &\mathbb{E}\big(\beta e^{(\beta-\nu)z}1(W> z+\Delta+\delta)\big)\\ &\leq C+C\delta \mathbb{E} \big( e^{(\beta-\nu)W}1(\beta\leq W\leq z+\Delta+\delta)\big)+\beta e^{(\beta-\nu)z} \P(z\leq W\leq z+\Delta). } Without loss of generality we assume $z$ does not belong to the support of $W$ and let $\Delta\to 0$, and observe that $\P(z\leq W\leq z+\Delta) \to 0$. Therefore, we have \be{ \magenta{\beta e^{(\beta-\nu)z}}\P(W>z+\magenta{\delta})\leq C+ C\delta \mathbb{E} e^{(\beta-\nu)W}\leq C } where we \magenta{have used Lemma~\ref{l3}.} \hfill $\square$\endproof \proof{Proof of Lemma~\ref{l3}} Since $\delta = 1/\sqrt{R}$, we can choose $C_0$ large enough so that $(\beta-3\beta^2 \delta)>\beta/2$. For notational convenience, we set $\nu = 3\beta^2 \delta$. Fix $M > \beta $ and let $f(x) = \int_{0}^{x}e^{(\beta-\nu)(y \wedge M)} dy$. We recall from \eqref{eq:steinx} that $G_{\magenta{\tilde X}} f(x) = b(W)f'(W) +\int_{-\delta}^{\delta} f''(W+y) K_{W}(y) dy$ where $K_W(y)$ is defined in \eqref{eq:kdef}. Since $\beta - \nu > \beta/2$ by assumption, the function $f(x)$ grows linearly for $x \geq M$, so $\mathbb{E} \abs{ f(W)} < \infty$ because $\mathbb{E} \abs{W} < \infty$. Therefore, $\mathbb{E} G_{\magenta{\tilde X}} f(W) = 0$, or $ \mathbb{E}\big(-b(W)f'(W)\big) = \mathbb{E} \Big(\int_{-\delta}^{\delta} f''(W+y) K_{W}(y) dy\Big)$, due to \eqref{eq:gxbarf}. Now \besn{\label{f40} \mathbb{E}\big(-b(W)f'(W)\big)=&\ \mathbb{E}\big(\mu \beta e^{(\beta-\nu)(W \wedge M)} 1(W\geq \beta) \big)+\mathbb{E}\big(\mu W e^{(\beta-\nu)W} 1(W< \beta) \big)\\ =&\ \mathbb{E}\big(\mu \beta e^{(\beta-\nu)(W \wedge M)} \big)+\mathbb{E}\big(\mu (W-\beta) e^{(\beta-\nu)W} 1(W< \beta)\big)\\ \geq&\ \mathbb{E}\big(\mu \beta e^{(\beta-\nu)(W \wedge M)} \big)-\mu C } where in the last inequality we used the fact that $\abs{(x-\beta) e^{(\beta-\nu)x}} 1(x<\beta) \leq C$ if $(\beta-\nu)>\beta/2$. Furthermore, since $f''(x) = (\beta-\nu) e^{(\beta-\nu)(x \wedge M)} 1(x < M)$ and $\int_{-\delta}^{\delta} K_{W}(y)dy = \mu -\delta b(W)/2 $, we have \besn{\label{f41} \mathbb{E} \bigg(\int_{-\delta}^{\delta} f''(W+y) K_{W}(y) dy\bigg) \leq&\ \mathbb{E} \bigg(\int_{-\delta}^{\delta}(\beta-\nu)e^{(\beta-\nu) ((W+y)\wedge M ) }K_W(y) dy\bigg)\\ \leq&\ (\beta-\nu) e^{(\beta-\nu)\delta}\mathbb{E}\Big(e^{(\beta-\nu)(W \wedge M)}(\mu-\frac{\delta}{2}b(W))\Big)\\ =&\ \mu (\beta-\nu) e^{(\beta-\nu)\delta}\mathbb{E}\Big(e^{(\beta-\nu)(W \wedge M)}(1+\frac{\delta}{2}W)1(W \leq \beta)\Big) \\ &+ \mu (\beta-\nu) e^{(\beta-\nu)\delta}\mathbb{E}\Big(e^{(\beta-\nu)(W \wedge M)}(1+\frac{\delta}{2}\beta)1(W \geq \beta)\Big)\\ \leq&\ \mu C+ \mu (\beta-\nu)e^{(\beta-\nu)\delta}(1+\frac{\delta}{2}\beta)\mathbb{E}\big(e^{(\beta-\nu)(W \wedge M)}\big).\\ } Divide both sides of \eqref{f40} and \eqref{f41} by \magenta{$\mu \delta$} and combine these two inequalities, and also substitute $3 \beta^2 \delta$ for $\nu$, to get \be{ \frac{\Big(\beta-(\beta-3\beta^2\delta)e^{(\beta-3\beta^2\delta)\delta} (1+\frac{\delta}{2}\beta) \Big)}{\delta} \mathbb{E} \big( e^{(\beta-3 \beta^2 \delta)(W\wedge M)}\big)\leq C/\delta. } Since the coefficient in front of the expected value on the left-hand side converges to a positive constant as $\delta \to 0$, for sufficiently small $\delta$ (or sufficiently large $C_0$), we have \be{ \mathbb{E} \big(e^{(\beta-3\beta^2\delta)(W\wedge M)}\big)\leq C/\delta. } We conclude by letting $M \to \infty$. \hfill $\square$\endproof \subsubsection{Proof of Lemma~\ref{lem:eq9}.}\label{fap6} We begin by proving \eqref{eq:lem91}. \begin{align*} &\mathbb{E} \Big( \int_{-\delta}^{\delta} \big( b(W+y)f_z'(W+y) - b(W) f_z'(W)\big) \frac{K_W(y)}{\mu } dy \Big) \\ =&\ \mathbb{E} \Big( \int_{-\delta}^{\delta}\frac{K_W(y)}{\mu } \int_{0}^{y}\big( b(x)f_z'(x)\big)'\big|_{x=W+s} ds dy \Big) \\ \leq&\ \delta \mathbb{E} \Big( \sup_{|s|\leq \delta} \abs{\big( b(x)f_z'(x)\big)'\big|_{x=W+s}} \big( 1 - \frac{\delta}{2\mu}b(W) \big) \Big)\\ \leq&\ C \delta \mathbb{E} \Big( \sup_{|s|\leq \delta} \abs{\big( b(x)f_z'(x)\big)'\big|_{x=W+s}} \Big). \end{align*} The first inequality is due to $K_W(y) \geq 0$ and $\int_{-\delta}^{\delta} K_{W}(y)/\mu dy = 1-\delta b(W)/(2\mu)$ from \eqref{MD:k3}, and the last inequality is true because $\big| \frac{\delta}{2\mu}b(W)\big| = \big|\frac{\delta}{2}(W \wedge \beta) \big| \leq C$ since $W\geq -1/\delta$. To bound the right-hand side we note that \magenta{(cf. \eqref{f33})} \be{ -b(x)f_z'(x)= \begin{cases} \P(Y_0\geq z)b(x) e^{-\int_{0}^{x} \frac{b(u)}{\mu} du} \int_{-\infty}^{x} \frac{1}{\mu } e^{\int_{0}^{y} \frac{b(u)}{\mu} du} dy , & x<z, \\ \P(Y_0\leq z)b(x) e^{-\int_{0}^{x} \frac{b(u)}{\mu} du} \int_{x}^{\infty} \frac{1}{\mu } e^{\int_{0}^{y} \frac{b(u)}{\mu} du} dy, & x\geq z, \end{cases} } so for \magenta{$x> z>\beta$}, \bes{ (-b(x)f_z'(x))' =&\ -\beta \P(Y_0\leq z)\Big(-1+ \beta e^{-\int_{0}^{x} \frac{b(u)}{\mu} du}\int_{x}^{\infty} e^{ \int_{0}^{y} \frac{b(u)}{\mu} du} dy \Big)\\ =&\ -\beta \P(Y_0\leq z)\Big(-1+ \beta e^{- \frac{\beta^2}{2} + \beta x}\int_{x}^{\infty} e^{ \frac{\beta^2}{2} - \beta y} dy \Big)\\ =&\ 0. } For $\beta<x<z$, \bes{ \abs{(b(x)f_z'(x))'} =&\ \beta \P(Y_0\geq z) \Big(1+ \beta e^{-\frac{\beta^2}{2}+\beta x} \int_{-\infty}^{x} e^{\int_{0}^{y} \frac{b(u)}{\mu} du} dy \Big) \leq \beta \P(Y_0\geq z) (1 + C e^{\beta x}). } In the inequality above we used the fact that $\int_{-\infty}^{x} e^{\int_{0}^{y} \frac{b(u)}{\mu} du} dy \leq C$ because $b(x)/\mu = -(x \wedge \beta)$ depends only on $\beta$. Lastly, for $x<\beta$, \bes{ \abs{(b(x)f_z'(x))'} =&\ \P(Y_0\geq z)\Big|x+(e^{\frac{x^2}{2}}+x^2 e^{\frac{x^2}{2}})\int_{-\infty}^{x} e^{- \frac{y^2}{2}} dy \Big|. } When $ -1 \leq x\magenta{ < \beta}$, the right-hand side is bounded by $C \P(Y_0\geq z)$ and when $x < -1$, we use the bound $\frac{1}{-x-1} e^{-\frac{x^2}{2}}\leq \int_{-\infty}^{x} e^{- \frac{y^2}{2}} dy \leq \frac{1}{-x} e^{-\frac{x^2}{2}}$ to conclude that \begin{align*} \P(Y_0\geq z)\Big|x+(e^{\frac{x^2}{2}}+x^2 e^{\frac{x^2}{2}})\int_{-\infty}^{x} e^{- \frac{y^2}{2}} dy \Big| \leq \P(Y_0\geq z)\Big|x+C-x \Big| \leq C \P(Y_0\geq z). \end{align*} Combining the three cases yields \eqref{eq:lem91}. We now prove the bound on $- \frac{\delta}{2\mu } \mathbb{E} \big( b^2(W) f_z'(W)\big)$ in \eqref{eq:lem92}. From the form of $b(x) f_z'(x)$ above, we have for $x>z\magenta{>\beta}$, \bes{ \magenta{-}\frac{1}{\mu}b^2(x)f_z'(x)&= \magenta{\beta} \P(Y_0\leq z), } for $\beta\leq x\leq z$, \bes{ \magenta{-}\frac{1}{\mu}b^2(x)f_z'(x)&= \P(Y_0\geq z) \beta e^{-\frac{\beta^2}{2}+\beta x} \int_{-\infty}^{x} e^{\magenta{\int_{0}^{y}} \frac{b(u)}{\mu} du} dy \leq C e^{ \beta x}\P(Y_0\geq z), } and for $x<\beta$, \be{ \magenta{-}\frac{1}{\mu}b^2(x)f(x) = \P(Y_0\geq z) x^2 e^{\frac{x^2}{2}}\int_{-\infty}^{x} e^{- \frac{y^2}{2}} dy \leq \P(Y_0\geq z) \magenta{(1+|x|)}. } Combining the three cases implies \eqref{eq:lem92}. Lastly we prove \eqref{eq:lem93}. In the proof of Lemma~\ref{lem:momentequiv} we showed that $\mathbb{E} b(W) = 0$. Furthermore, $-\mu\beta\leq b(x)\leq 0$ for $x\geq 0$, so \begin{align*} -\frac{\delta}{2\mu} \mathbb{E} \big( b(W) \big(\P(Y_0\geq z)-1(W\geq z+\delta)\big) \big)= \frac{\delta}{2\mu} \mathbb{E} \big( b(W)1(W\geq z+\delta) \big) \leq C\delta \P(W\geq z+\delta). \end{align*} \hfill $\square$\endproof \section{Companion for the Hospital Model} \label{app:hospital_proofs} In this portion of the electronic companion, we motivate the $v_3$ approximation for the hospital model presented in Section~\ref{sec:hospcompare}, where we suggested using \begin{align} \label{eq:hosv3ec} v_3(x)=\max\Big\{ \delta +\frac{1}{2} \Big(\delta^{2} 1(x<0)- \delta^2 ( x^{-} - \beta )-\delta^2-2\delta^2\beta\Big), \delta/2 \Big\}. \end{align} We recall that $\tilde X = \{ \tilde X(n) = \delta (X(n) - N)\}$, where $X(n)$ is the customer count at the end of time unit $n$, that $W$ and $W'$ have the distributions of $\tilde X(0)$ and $\tilde X(1)$ when $\tilde X(0)$ is initialized according to the stationary distribution of $\tilde X$, and that $\Delta = W' - W$. We use $\epsilon(x)$ and $\epsilon_i(x)$ to denote generic functions satisfying \begin{align} \abs{\epsilon(x)} \leq C (1 + \abs{x})^{5}. \label{eq:epsdefec} \end{align} The following lemma gives us the conditional moments of $\Delta$. It is proved in Section~\ref{sec:hosmecproof}. \begin{lemma}\label{lem:hosmec} For the hospital model with $N$ servers, $\Lambda=\sqrt{N}-\beta$ and $\mu=\delta=1/\sqrt{N}$, \begin{align} b(x) = \mathbb{E}(\Delta| W= x) =&\ \delta( x^{-} - \beta ), \label{eq:1ec}\\ \mathbb{E}(\Delta^2|W=x) =&\ 2\delta + \Big(b^{2}(x)-\delta b(x) -\delta^2 - 2\delta^2\beta\Big) + \delta^3 x^{-} \label{eq:2ec} \\ \mathbb{E}(\Delta^3|W=x) =&\ 6\delta b(x) + \delta^{3} \epsilon(x) , \label{eq:3ec}\\ \mathbb{E}(\Delta^4|W=x) =&\ 12\delta^2 + \delta^{3} \epsilon(x), \label{eq:4ec}\\ \mathbb{E}(\Delta^5|W=x)=&\ \delta^{3} \epsilon(x). \label{eq:5ec} \end{align} \end{lemma} To derive $v_3(x)$, we begin with the Taylor expansion in \eqref{eq:taylorgeneric} with $n = 4$, which says that for sufficiently smooth $f(x)$, \begin{align} - \mathbb{E} \Delta f'(W) = \mathbb{E} \bigg[\sum_{i=2}^{4} \frac{1}{i!} \Delta^i f^{(i)}(W) + \frac{1}{5!} \Delta^{5} f^{(5)}(\xi_1)\bigg], \label{eq:txec} \end{align} where $\xi_i$ denote numbers lying between $W$ and $W'$. By combining \eqref{eq:txec} with Lemma~\ref{lem:hosmec}, we will show that \begin{align} &- \mathbb{E} b(W) f'(W) - \frac{1}{2} \mathbb{E}\Big(2 \delta + \delta^{2} 1(W<0)- \delta^2( W^{-} - \beta ) -\delta^2-2\delta^2\beta\Big) f''(W)\notag \\ =&\ \delta^3 \Big(\frac{1}{2}\mathbb{E} \epsilon_0(W) f''(W)+ \frac{1}{6} \mathbb{E} \epsilon_3(W) f'''(W) + \frac{1}{24} \mathbb{E} \epsilon_4(W) f^{(4)}(W) + \frac{1}{120} \mathbb{E} \epsilon_5(W) f^{(5)}(\xi_1)\Big) \notag \\ &+ \frac{1}{2}\delta^3 \mathbb{E} \Big(\epsilon_1(W) f^{(4)}(W) + \epsilon_2(W) f^{(5)}(\xi)\Big) \notag \\ &+\frac{1}{2} \delta^3 \mathbb{E} \Big(\epsilon_6(W)f''(W)+\epsilon_7(W)f'''(W) + \epsilon_2(W) \Big(\frac{d^2}{d x^2} \big( (x^{-}-\beta) f''(x) \big)\big|_{x = \xi_3}\Big)\Big). \label{eq:toprovehospec} \end{align} Truncating the term in front of $f''(W)$ on the left-hand side from below by $\delta/2$ gives us $v_3(x)$ in \eqref{eq:hosv3ec}. The truncation level $\delta/ 2$ is chosen because the support of $W$ is in $[-\delta N, \infty)$ and the term in front of $f''(W)$ on the left-hand side of \eqref{eq:toprovehospec} equals $\delta \big(\frac{1}{2} - \frac{1}{2}\delta \beta \big) \approx\frac{\delta}{2}$ when evaluated at the point $W = -\delta N$. We could have chosen $\delta \big(\frac{1}{2} - \frac{1}{2}\delta \beta \big)$ instead (when this quantity is positive), but in practice this does not make a big difference. Let us now prove \eqref{eq:toprovehospec}. Combining \eqref{eq:txec} with Lemma~\ref{lem:hosmec} yields \begin{align*} - \mathbb{E} b(W) f'(W) =&\ \frac{1}{2} \mathbb{E}\Big(2\delta + b^{2}(W)-\delta b(W) -\delta^2 - 2\delta^2\beta + \delta^3 W^{-}\Big) f''(W) \notag \\ &+ \frac{1}{6} \mathbb{E} \big(6\delta b(W) + \delta^{3} \epsilon_3(W)\big)f'''(W) + \frac{1}{24} \mathbb{E}\big(12\delta^2 + \delta^3 \epsilon_4(W) \big) f^{(4)}(W) \notag \\ &+ \frac{1}{120} \delta^3 \mathbb{E} \epsilon_5(W) f^{(5)}(\xi_1). \end{align*} Let us write $W^{-}f''(W)$ as $\epsilon_{0}(W) f''(W)$ and rearrange the right-hand side above into the more convenient form: \begin{align} &- \mathbb{E} b(W) f'(W) - \frac{1}{2} \mathbb{E}\Big(2\delta + b^{2}(W)-\delta b(W) -\delta^2 - 2\delta^2\beta \Big) f''(W) \notag \\ =&\ \frac{1}{2} \delta \mathbb{E} b(W) f'''(W) + \frac{1}{2} \delta \big(\mathbb{E} b(W) f'''(W) + \delta \mathbb{E} f^{(4)}(W)\big) \notag \\ &+\delta^3 \Big( \frac{1}{2} \mathbb{E} \epsilon_0(W) f''(W)+ \frac{1}{6} \mathbb{E} \epsilon_3(W) f'''(W) + \frac{1}{24} \mathbb{E} \epsilon_4(W) f^{(4)}(W) + \frac{1}{120} \mathbb{E} \epsilon_5(W) f^{(5)}(\xi_1)\Big). \label{1} \end{align} The last row is considered as error because of the $\delta^3$ there. We wish to transform the first row on the right-hand side into an expression involving $f''(x)$ plus error. To this end we require the following lemma, which is proved at the end of this section. \begin{lemma} \label{lem:hospaux} Suppose that $g \in C^{3}(\mathbb{R})$ is such that $\mathbb{E} g(W') - \mathbb{E} g(W) = 0$. Then \begin{align*} \mathbb{E} b(W) g'(W) + \delta \mathbb{E} g''(W) =&\ \delta^2 \mathbb{E} \Big(\epsilon_1(W) g''(W) + \epsilon_2(W) g'''(\xi)\Big), \end{align*} where $\epsilon_i(x)$ are generic functions satisfying \eqref{eq:epsdefec} and $\xi$ lies between $W$ and $W'$. \end{lemma} We apply Lemma~\ref{lem:hospaux} with $g(x) = \frac{1}{2} \delta f''(x)$ to get \begin{align} & \frac{1}{2}\delta \Big(\mathbb{E} b(W) f'''(W) + \delta \mathbb{E} f^{(4)}(W)\Big) = \frac{1}{2}\delta^3 \mathbb{E} \Big(\epsilon_1(W) f^{(4)}(W) + \epsilon_2(W) f^{(5)}(\xi)\Big). \label{eq:gfdelta} \end{align} The left-hand side above coincides with one of the terms in the second row of \eqref{1}. Next we choose $g(x) = \int_{0}^{x} b(y) f''(y) dy$ and note that $g''(x) = b'(x)f''(x) + b(x)f'''(x)$. Applying Lemma~\ref{lem:hospaux} with our new choice of $g(x)$, we get \begin{align*} & \mathbb{E} b^2(W) f''(W) + \delta \mathbb{E} \big(b'(W)f''(W) + b(W)f'''(W)\big) \\ =&\ \delta^2 \mathbb{E} \Big(\epsilon_1(W) \big(b'(W)f''(W) + b(W)f'''(W)\big) + \epsilon_2(W) \Big(\frac{d^2}{d x^2} \big( b(x) f''(x) \big)\big|_{x = \xi_3}\Big)\Big)\\ =&\ \delta^3 \mathbb{E} \Big(\epsilon_6(W)f''(W)+\epsilon_7(W)f'''(W) + \epsilon_2(W) \Big(\frac{d^2}{d x^2} \big( (x^{-}-\beta) f''(x) \big)\big|_{x = \xi_3}\Big)\Big). \end{align*} The last equality follows from the fact that $b(x) = \delta(x^{-}-\beta)$. Multiplying both sides by $1/2$ and rearranging terms, we get \begin{align} & \frac{1}{2}\delta \mathbb{E} b(W) f'''(W) \notag \\ =&\ - \frac{1}{2}\mathbb{E}\big( b^2(W) + \delta b'(W)\big) f''(W) \notag \\ &+\frac{1}{2} \delta^3 \mathbb{E} \Big(\epsilon_6(W)f''(W)+\epsilon_7(W)f'''(W) + \epsilon_2(W) \Big(\frac{d^2}{d x^2} \big( (x^{-}-\beta) f''(x) \big)\big|_{x = \xi_3}\Big)\Big). \label{f25} \end{align} Plugging \eqref{eq:gfdelta} and \eqref{f25} into \eqref{1}, we conclude that \begin{align*} &- \mathbb{E} b(W) f'(W) - \frac{1}{2} \mathbb{E}\Big(2\delta + b^{2}(W)-\delta b(W) -\delta^2 - 2\delta^2\beta \Big) f''(W)\notag \\ =&\ - \frac{1}{2}\mathbb{E}\big( b^2(W) + \delta b'(W)\big) f''(W) \\ &+ \delta^3 \Big(\frac{1}{2}\mathbb{E} \epsilon_0(W) f''(W)+ \frac{1}{6} \mathbb{E} \epsilon_3(W) f'''(W) + \frac{1}{24} \mathbb{E} \epsilon_4(W) f^{(4)}(W) + \frac{1}{120} \mathbb{E} \epsilon_5(W) f^{(5)}(\xi_1)\Big) \\ &+ \frac{1}{2}\delta^3 \mathbb{E} \Big(\epsilon_1(W) f^{(4)}(W) + \epsilon_2(W) f^{(5)}(\xi)\Big) \\ &+\frac{1}{2} \delta^3 \mathbb{E} \Big(\epsilon_6(W)f''(W)+\epsilon_7(W)f'''(W) + \epsilon_2(W) \Big(\frac{d^2}{d x^2} \big( (x^{-}-\beta) f''(x) \big)\big|_{x = \xi_3}\Big)\Big). \end{align*} To conclude \eqref{eq:toprovehospec}, we move $- \frac{1}{2}\mathbb{E}\big( b^2(W) + \delta b'(W)\big) f''(W)$ to the left-hand side and note that \begin{align*} & 2\delta + b^{2}(W)-\delta b(W) -\delta^2 - 2\delta^2\beta - b^2(W) - \delta b'(W) \\ =&\ 2\delta + \delta^2( W^{-} - \beta )^2 -\delta^2( W^{-} - \beta ) -\delta^2 - 2\delta^2\beta - \delta^2( W^{-} - \beta )^2 + \delta^2 1(W<0) \\ =&\ 2\delta - \delta^2(W^{-}-\beta)- \delta^2 - 2\delta^2\beta + \delta^2 1(W < 0), \end{align*} which coincides with the term in front of $f''(W)$ on the left-hand side of \eqref{eq:toprovehospec}. \proof{Proof of Lemma~\ref{lem:hospaux}} Since $\mathbb{E} g(W') - \mathbb{E} g(W) = 0$, performing a third-order Taylor expansion gives us \begin{align*} 0 =&\ \mathbb{E} \Delta g'(W) + \frac{1}{2} \mathbb{E}\Delta^2 g''(W) + \frac{1}{6} \mathbb{E}\Delta^{3} g'''(\xi) \\ =&\ \mathbb{E} b(W) g'(W) + \frac{1}{2} \mathbb{E}\Big(2\delta + b^{2}(W)-\delta b(W) -\delta^2 - 2\delta^2\beta + \delta^3 W^{-}\Big) g''(W) \notag \\ &+ \frac{1}{6} \mathbb{E} \big(6\delta b(W) + \delta^{3} \epsilon_0(W)\big)g'''(\xi). \end{align*} The second equality is due to Lemma~\ref{lem:hosmec}. We set \begin{align*} \epsilon_1(x) =&\ -\frac{b^{2}(x)-\delta b(x) -\delta^2 - 2\delta^2\beta + \delta^3 x^{-}}{2\delta^2} \quad \text{ and } \quad \epsilon_2(x) = - \frac{ 6\delta b(x) + \delta^{3} \epsilon_0(x) }{6\delta^2} \end{align*} and note that both $\epsilon_1(x)$ and $\epsilon_2(x)$ satisfy \eqref{eq:epsdefec} because $b(x) = \delta(x^{-}-\beta)$. \hfill $\square$\endproof \subsection{Proof of Lemma~\ref{lem:hosmec} } \label{sec:hosmecproof} By the definition of the hospital model the change in customers $\Delta = W'-W$ satisfies \begin{align*} \Delta = \delta(A - D), \end{align*} where $A$ is a mean $\Lambda$ Poisson random variable, and conditioned on $W=x=\delta(k-N)$, $D\sim \text{Binomial}(k\wedge N, \mu)$; see \cite{DaiShi2017}. To prove the lemma, we utilize the following Stein identities for a mean $\Lambda$ Poisson random variable $X$ and a Binomial($k, \mu$) random varabile $Y$: \begin{align} & \mathbb{E} X f(X) =\Lambda \mathbb{E} f(X+1) \quad \text{ for each } f:\mathbb{Z}_+\to \mathbb{R} \text{ with } \mathbb{E} \abs{ X f(X)} <\infty, \label{eq:steinpoisson} \\ & \mathbb{E} Yf(Y) = \mu k \mathbb{E} f(Y+1) - \mu \mathbb{E}\Big[Y\Big(f(Y+1)-f(Y)\Big)\Big] \text{ for each } f:\mathbb{Z}_+\to \mathbb{R}. \label{eq:steinbinom} \end{align} See, for example, Lectures VII and VIII of \cite{Stei1986}. \proof{Proof of Lemma~\ref{lem:hosmec}} We prove \eqref{eq:1ec}--\eqref{eq:5ec} in sequence. Using the facts that $\Lambda=\sqrt{N}-\beta$ and $\mu=\delta=1/\sqrt{N}$, for $x = \delta(k - N)$ we have \begin{align} \mathbb{E}(A-D|W=x)=\Lambda - (k\wedge N) \mu = \mu \big(\Lambda/\mu - N - (k- N \wedge 0)\big) =&\ -\beta - (x \wedge 0) = x^{-} - \beta. \label{eq:6} \end{align} For the remainder of the proof we adopt the convention that all expectations are conditioned on $W=x$. We now prove \eqref{eq:2ec}: \begin{align} \mathbb{E} \Delta^2 &=\delta^2\Big(\mathbb{E}\big[A(A-D)\big]-\mathbb{E}\big[D(A-D)\big]\Big)\nonumber\\ &=\delta^2\Big(\Lambda\mathbb{E} (A-D+1)-\mathbb{E}\big[D(A-D)\big] \Big) \nonumber\\ &=\delta^2\Big(\Lambda\mathbb{E} (A-D+1)-\mu(n\wedge N) \mathbb{E} (A-D-1) - \mu \mathbb{E} D \Big) \nonumber\\ & =\delta^2\Big(2\Lambda + ( \Lambda - \mu(n \wedge N) )\mathbb{E} (A-D-1) + \mu(x^{-} - \sqrt{N})\Big) \nonumber\\ &= \delta^2\Big(2\Lambda + (x^{-} - \beta)^2 - (x^{-} - \beta) + \mu(x^{-} - \sqrt{N})\Big).\label{eq:8} \end{align} We used \eqref{eq:steinpoisson} in the second equality and \eqref{eq:steinbinom} in the third equality. The fourth and fifth equalities are due to \begin{align*} \mathbb{E} D =&\ \mathbb{E}(D-A)+\mathbb{E} A =-(x^{-} - \beta)+\Lambda = -x^- + \sqrt{N} \\ &\ \text{ and } \Lambda-\mu(n\wedge N) = \mathbb{E}(A-D) = x^- - \beta, \text{ respectively}. \end{align*} Lastly, bringing the $\delta^2$ term in \eqref{eq:8} inside the parenthesis and recalling that $\delta^2 \Lambda = \delta - \delta^2 \beta$, $\delta (x^- - \beta) = b(x)$, and $\mu = \delta = 1/\sqrt{N}$ yields (\ref{eq:2ec}). We now prove \eqref{eq:3ec}. Note that $\mathbb{E}\Delta^3 = \delta^3 \mathbb{E} (A-D)^3$ equals \begin{align*} & \delta^3\Big(\mathbb{E}\big[A(A-D)^2\big]-\mathbb{E}\big[D(A-D)^2\big]\Big)\\ =&\delta^3\Big(\Lambda \mathbb{E}(A-D+1)^2-\mu (n\wedge N)\mathbb{E}(A-D-1)^2 + \mu \mathbb{E} D\big[(A-D-1)^2-(A-D)^2\big] \Big) \\ =&\delta^3\Big(4 \Lambda \mathbb{E} (A-D) + \big(\Lambda-(n\wedge N)\mu\big) \mathbb{E} (A-D-1)^2 + \mu \mathbb{E} D\big[-2(A-D)+1\big] \Big). \end{align*} The first equality is due to \eqref{eq:steinpoisson} and \eqref{eq:steinbinom} and to get the second equality we use $\mathbb{E} (A-D+1)^2 = \mathbb{E} (A-D- 1 + 2)^2 = \mathbb{E}\big[(A-D- 1)^2 + 4(A-D)\big]$. Let us analyze the terms above one by one. First, we have \begin{align*} \delta^3 4 \Lambda \mathbb{E} (A-D) =&\ 4 (x^- - \beta) \delta^3 \Lambda = 4 (x^- - \beta) (\delta^2 - \delta^3 \beta) = 4\delta^2(x^- - \beta) + \delta^3\epsilon(x). \end{align*} Second, since $\delta(A-D) = \Delta$, we have \begin{align*} \delta^3 \big(\Lambda-(n\wedge N)\mu\big) \mathbb{E} (A-D-1)^2 =&\ \delta (x^- - \beta) \mathbb{E} \big(\Delta^2 - 2\delta \mathbb{E} \Delta + \delta^2 \big)\\ =&\ \delta (x^- - \beta) \mathbb{E} \big(\Delta^2 - 2\delta^2(x^- - \beta) + \delta^2 \big) \\ =&\ \delta (x^- - \beta) \mathbb{E} \Delta^2 + \delta^3 \epsilon(x)\\ =&\ 2\delta^2(x^- - \beta) + \delta^3 \epsilon(x). \end{align*} The last equality is due to \eqref{eq:2ec}, which says that $\mathbb{E} \Delta^2 = 2\delta + \delta^2 \epsilon(x)$. Lastly, \begin{align*} \delta^3 \mu \mathbb{E} D\big[-2(A-D)+1\big] =&\ \delta^4 \big( 2 \mathbb{E} (A-D)^2 - 2\mathbb{E} A(A-D) + \mathbb{E} D\big) \\ =&\ \delta^2 \big( 2 \mathbb{E} \Delta^2 - 2\delta^2\Lambda\mathbb{E} (A-D+1) + \delta^2(- x^- + \sqrt{N})\big) \\ =&\ \delta^3 \epsilon(x). \end{align*} The second equality is due to \eqref{eq:steinpoisson} and the last equality follows from $\mathbb{E} \Delta^2 = \delta \epsilon(x)$ and $\delta^2 \Lambda\mathbb{E}(A-D +1) = \delta (1-\delta \beta)(x^- - \beta+1) = \delta \epsilon(x)$. Putting the pieces together yields $\mathbb{E} \Delta^3 = 6\delta^2 (x^- - \beta) + \delta^3 \epsilon(x)$, which proves \eqref{eq:3ec}. We now prove (\ref{eq:4ec}): \begin{align*} \mathbb{E} \Delta^4 =&\delta^4\Big(\mathbb{E}\big[A(A-D)^3\big]-\mathbb{E}\big[D(A-D)^3\big]\Big)\\ &= \delta^4\Big(\Lambda \mathbb{E}\big[(A-D+1)^3\big]-\mu(n\wedge N) \mathbb{E}(A-D-1)^3 \Big)\\ & \quad + \mu \mathbb{E}\big[D(A-D-1)^3-D(A-D)^3]\Big)\\ =&\delta^4\Big(\Lambda \mathbb{E}\big[(A-D+1)^3-(A-D-1)^3\big] + \big(\Lambda-(n\wedge N)\mu\big) \mathbb{E}(A-D-1)^3 \\ &+ \mu \mathbb{E} D\big[(A-D-1)^3-(A-D)^3\big] \Big) \\ =&\delta^4\Big(\Lambda \mathbb{E}\big[6(A-D)^2+2\big] + (x^{-} - \beta) \mathbb{E}(A-D-1)^3 \\ &+ \mu \mathbb{E} D\big[-3(A-D)^2+3(A-D)-1\big] \Big). \end{align*} Let us analyze the terms above one by one. First, \begin{align*} \delta^4 \Lambda \mathbb{E}\big[6(A-D)^2+2\big] = \delta^2 \Lambda \mathbb{E}\big[6\Delta^2+2\delta^2\big] = 12\delta^2 + \delta^3 \epsilon(x). \end{align*} Second, \begin{align*} \delta^4 (x^{-} - \beta) \mathbb{E} (A-D-1)^3 = \delta (x^{-} - \beta) \big( \mathbb{E} \Delta^3 - 3 \delta \mathbb{E} \Delta^2 + 3 \delta^2 \mathbb{E} \Delta - 1\big) = \delta^3 \epsilon(x), \end{align*} and third, \begin{align*} & \delta^4 \mu \mathbb{E} D\big[-3(A-D)^2+3(A-D)-1\big] \\ =&\ \delta^5 \mathbb{E} \big[ 3 (A-D)^3 - 3(A-D)^2 + D \big] + \delta^5 \mathbb{E} A\big[-3(A-D)^2+3(A-D)\big]\\ =&\ \delta^2 \mathbb{E} \big[ 3 \Delta^3 - 3\delta \Delta^2 + \delta^3 D \big] + \delta^5 \Lambda \mathbb{E} \big[-3(A-D+1)^2+3(A-D+1)\big]\\ =&\ \delta^3 \epsilon(x). \end{align*} Putting the pieces together yields $\mathbb{E} \Delta^4 = 12\delta^2 + \delta^3 \epsilon(x)$. The proof of (\ref{eq:5ec}) is analogous to the proof of (\ref{eq:4ec}) and is omitted. \hfill $\square$\endproof \section{Companion for the AR(1) Model} \label{app:ar1proof} In this section we derive the $v_3$ approximation for the AR(1) model and then prove Lemma~\ref{lem:ar1} in Section~\ref{sec:ar1pf}. We recall that $W = \delta(D_{\infty}-R)$, $W' = e^{-\alpha Z}W + \delta\big(X + R(e^{-\alpha Z} - 1)\big)$, and $\Delta = W' - W$, where $\delta = \sqrt{\alpha}$, $R = 1/\alpha$, $X$ and $Z$ are independent unit-mean exponentially distributed random variables that are also independent of $W$, and $D_\infty > 0$ has the limiting distribution of the AR(1) model defined by \eqref{eq:defar1}. The asymptotic regime we consider is $\alpha \to 0$, so we assume that $\alpha \in (0,1)$. We recall Lemma~\ref{lem:ar1}: \begin{lemma}\label{lem:ar1ec} Recall that $\delta = \sqrt{\alpha}$. For any $k \geq 1$, \begin{align*} \mathbb{E}(\Delta^{k} | D_{\infty} = d) = \delta^{k} k! \bigg(1 + \sum_{i=1}^{k} (-1)^{i}d^{i} \prod_{j=1}^{i} \frac{ \alpha }{1 + j \alpha} \bigg), \quad d > 0. \end{align*} \end{lemma} We also recall that for $x \geq -1/\sqrt{\alpha}$, \begin{align*} \mathbb{E}(\Delta^{k} | W = x) = \mathbb{E}(\Delta^{k} | D_{\infty} = x/\delta + R) =&\ \delta^{k} k! \bigg(1 + \sum_{i=1}^{k} (-1)^{i}\Big( x\sqrt{\alpha} + 1 \Big)^{i} \prod_{j=1}^{i} \frac{1}{1 + j \alpha} \bigg), \end{align*} and observe that $\mathbb{E}(\Delta^{k} | W = x) = \delta^k p_{k}(x)$ for some degree-$k$ polynomial $p_{k}(x)$. To derive $v_3(x)$, we start with the Taylor expansion in \eqref{eq:taylorgeneric} with $n = 4$; i.e., for any function $f: \mathbb{R} \to \mathbb{R}$ satisfying $\mathbb{E} f(W') - \mathbb{E} f(W) = 0$, \begin{align} & \delta \mathbb{E} p_{1}(W) f'(W)+\frac{1}{2}\delta^2 \mathbb{E} p_{2}(W) f''(W) + \frac{1}{6} \delta^3 \mathbb{E} p_{3}(W) f'''(W)+\frac{1}{24}\delta^4 \mathbb{E} p_{4}(W) f^{(4)}(W) \notag \\ =&\ -\frac{1}{120} \delta^5 \mathbb{E} p_{5}(W) f^{(5)}(\xi). \label{eq:taylorfourth} \end{align} Since $\sup_{\alpha \in (0,1)} \abs{p_{k}(x)} < \infty$ for each $x \in \mathbb{R}$, the right-hand side is of order $\delta^5$. When deriving $v_3$, we want it to account for all terms of order $\delta, \ldots, \delta^4$ and treat terms of order $\delta^5$ as error. The following lemma is the basis for the $v_3$ approximation. It converts the third and fourth derivative terms in \eqref{eq:taylorfourth} into expressions involving $f''(x)$ plus error. Its proof is similar to the $v_2$ derivation in Section~\ref{sec:v2def}, so we postpone it until the end of this section. \begin{lemma} \label{lem:ar1taylor} Define \begin{align*} \bar p_3(x) =&\ \frac{1}{6} \Big( p_3(x) - \frac{p_1(x)p_4(x)}{ 2 p_2(x)} - \frac{1}{4} \delta p_2(x)\Big(\frac{p_4(x)}{p_2(x)}\Big)' \Big),\\ \underline p_2(x) =&\ \Big(\frac{p_2(x)}{2}-\frac{p_1(x)p_3(x)}{3p_2(x)}-\frac{p_2(x)}{6}\Big(\frac{p_3(x)}{p_2(x)}\Big)'\Big). \end{align*} Let $W$ and $W'$ be as in Section~\ref{fse5}. If $f \in C^{5}(\mathbb{R})$ is such that $\mathbb{E} f(W') - \mathbb{E} f(W) = 0$, then \begin{align} &\delta \mathbb{E} p_{1}(W) f'(W)+\delta^2 \mathbb{E} \Big(\frac{p_{2}(W)}{2}-\frac{p_{1}(W)\bar p_3(W)}{ \underline{p}_2(W)}-\delta \underline{p}_2(W)\Big(\frac{\bar p_3(W)}{\underline{p}_2(W)}\Big)'\Big)f''(W) \notag \\ =&\ \ -\frac{1}{120} \delta^5 \mathbb{E} p_5(W) f^{(5)}(\xi_1)+ \frac{1}{72} \delta^5 \mathbb{E} p_3(W) \Big( \frac{ p_4(x) }{p_2(x) }f'''(x) \Big)''\Big|_{x = \xi_2 } \notag \\ & \hspace{1.2cm} + \frac{1}{24}\delta^5 \mathbb{E} p_4(W) \Big( \frac{\bar p_3(x) }{\underline{p}_2(x) }f''(x) \Big)'''\Big|_{x = \xi_3 } - \frac{1}{18} \delta^5 \mathbb{E} p_3(W) \Big(\frac{p_3(x)}{p_2(x)} \frac{\bar p_3(x) }{\underline{p}_2(x) }f''(x) \Big)'' \Big|_{x = \xi_4 }. \label{eq:arlem} \end{align} \end{lemma} We choose \begin{align*} \underline{v}_3(x) = \delta^2 \Big(\frac{p_{2}(x)}{2}-\frac{p_{1}(x)\bar p_3(x)}{ \underline{p}_2(x)}-\delta \underline{p}_2(x)\Big(\frac{\bar p_3(x)}{\underline{p}_2(x)}\Big)'\Big) \end{align*} based on the term in front of $f''(W)$ on the left-hand side of \eqref{eq:arlem}, which is the basis for the $v_3$ approximation in Section~\ref{fse5}. Our choice is based on the presumption that all the terms on the right-hand side of \eqref{eq:arlem} are of order $\delta^5$ when $\delta$ is close to zero. We do not prove this claim rigorously in this paper. Nevertheless, our $v_3$ approximation performs quite well numerically. \proof{Proof of Lemma~\ref{lem:ar1taylor}} In Section~\ref{sec:v2def}, we used $a(x), \ldots, d(x)$ to represent $\mathbb{E}(\Delta^{k} | W = x)$; i.e., \begin{align} b(x) =&\ \mathbb{E}(\Delta | W = x), \quad a(x) = \mathbb{E}(\Delta^{2} | W = x), \quad c(x) = \mathbb{E}(\Delta^{3} | W = x), \notag \\ d(x) =&\ \mathbb{E}(\Delta^{4} | W = x), \quad e(x) = \mathbb{E}(\Delta^{5} | W = x). \label{eq:abcd} \end{align} Since this proof relies heavily on Section~\ref{sec:v2def}, we use this notation and then convert to use $\delta^k p_k(x) = \mathbb{E}(\Delta^{k} | W = x)$ at the end. Our starting point is equation \eqref{eq:taylorfourth}, which we recall for convenience: \begin{align} & \mathbb{E} b(W) f'(W)+\frac{1}{2} \mathbb{E} a(W) f''(W) + \frac{1}{6} \mathbb{E} c(W) f'''(W)+\frac{1}{24} \mathbb{E} d(W) f^{(4)}(W) \notag \\ =&\ -\frac{1}{120} \mathbb{E} e(W) f^{(5)}(\xi). \label{eq:taylorfourthinpf} \end{align} Our proof relies on several key equations from Section~\ref{sec:v2def}, which we recall as we go. Let $g_{1}(x) = \int_{0}^{x}\frac{d(y)}{a(y)} f'''(y) dy $ and use \eqref{eq:taylorsecond}, or \begin{align} \mathbb{E} b(W) f'(W)+\frac{1}{2}\mathbb{E} a(W)f''(W) = -\frac{1}{6} \mathbb{E} c(W) f'''(\xi_2), \label{eq:taylorsecondec} \end{align} with $g_{1}(x)$ in place of $f(x)$ there to get \begin{align*} \mathbb{E} \frac{b(W)d(W)}{a(W)}f'''(W)+\mathbb{E} \frac{a(W)}{2}\Big(\frac{d(W)}{a(W)}\Big)'f'''(W)+\frac{1}{2}\mathbb{E} d(W)f^{(4)}(W) = -\frac{1}{6} \mathbb{E} c(W) g_{1}'''(\xi_2). \end{align*} We multiply both sides by $1/12$ and subtract the result from \eqref{eq:taylorfourthinpf} to get \begin{align} &\mathbb{E} b(W) f'(W)+\frac{1}{2}\mathbb{E} a(W)f''(W) + \mathbb{E} \bar c(W) f'''(W) \notag \\ =&\ \frac{1}{72} \mathbb{E} c(W) g_{1}'''(\xi_2) -\frac{1}{120} \mathbb{E} e(W) f^{(5)}(\xi_1), \label{eq:arintermv3} \end{align} where $\bar c(x) = \frac{1}{6} c(x) - \frac{b(x)d(x)}{ 12 a(x)} - \frac{a(x)}{24}\Big(\frac{d(x)}{a(x)}\Big)' $. Note that $\bar c(x) = \delta^3 \bar p_3(x)$. Next, let $g_2(x) = \int_{0}^{x}\frac{\bar c(y) }{\underline{v}_2(y) }f''(y) dy $, where \begin{align*} \underline{v}_2(x) = \frac{a(x)}{2}-\frac{b(x)c(x)}{3a(x)}-\frac{a(x)}{6}\Big(\frac{c(x)}{a(x)}\Big)', \quad x \in \mathbb{R} \end{align*} is identical to $\underline{v}_2(x)$ defined in \eqref{eq:underlinev2}, and note that $\underline{v}_2(x) = \delta^2 \underline{p}_2(x)$. We use \eqref{f4}, or \begin{align} & \mathbb{E} b(W)f'(W)+\mathbb{E}\Big(\frac{a(W)}{2}-\frac{b(W)c(W)}{3a(W)}-\frac{a(W)}{6}\Big(\frac{c(W)}{a(W)}\Big)'\Big)f''(W) \notag \\ =&\ \frac{1}{18} \mathbb{E} c(W) g'''(\xi_2) -\frac{1}{24} \mathbb{E} d(W) f^{(4)}(\xi_1), \label{f4ec} \end{align} with $g_2(x)$ in place of $f(x)$ there to get \begin{align*} & \mathbb{E} b(W)\frac{\bar c(W) }{\underline{v}_2(W) }f''(W)+\mathbb{E} \underline{v}_2(W) \Big(\frac{\bar c(W) }{\underline{v}_2(W) }\Big)' f''(W) + \mathbb{E} \underline{v}_2(W) \frac{\bar c(W) }{\underline{v}_2(W) }f'''(W) \notag \\ =&\ \frac{1}{18} \mathbb{E} c(W) \Big(\frac{c(x)}{a(x)} g_{2}''(x) \Big)'' \Big|_{x = \xi_4 } -\frac{1}{24} \mathbb{E} d(W) g_2^{(4)}(\xi_3). \end{align*} Subtracting the equation above from \eqref{eq:arintermv3}, we conclude that \begin{align*} &\mathbb{E} b(W) f'(W)+\mathbb{E} \Big(\frac{a(W)}{2}-\frac{b(W)\bar c(W)}{ \underline{v}_2(W)}-\underline{v}_2(W)\Big(\frac{\bar c(W)}{\underline{v}_2(W)}\Big)'\Big)f''(W) \notag \\ =&\ \ - \frac{1}{18} \mathbb{E} c(W) \Big(\frac{c(x)}{a(x)} g_{2}''(x) \Big)'' \Big|_{x = \xi_4 } + \frac{1}{24} \mathbb{E} d(W) g_2^{(4)}(\xi_3) \\ &+ \frac{1}{72} \mathbb{E} c(W) g_{1}'''(\xi_2) -\frac{1}{120} \mathbb{E} e(W) f^{(5)}(\xi_1). \end{align*} To conclude, we note that $g_2^{(4)}(\xi_3) =\Big( \frac{\bar c(x) }{\underline{v}_2(x) }f''(x) \Big)'''\Big|_{x = \xi_3 }$ and $g_{1}'''(\xi_2) = \Big( \frac{ d(x) }{a(x) }f'''(x) \Big)''\Big|_{x = \xi_2 }$ and then substitute $\delta p_1(x)$ for $b(x)$, $\delta^2 p_2(x)$ for $a(x)$, etc., where they appear above. \hfill $\square$\endproof \subsection{Proof of Lemma~\ref{lem:ar1}} \label{sec:ar1pf} \proof{} Recall that $\Delta = W' - W = \delta \big(D_{\infty}(e^{-\alpha Z} - 1) + X\big)$, so \begin{align*} \mathbb{E}(\Delta^{k} | D_{\infty} = d) =&\ \delta^k \mathbb{E} \big(d(e^{-\alpha Z} - 1) + X \big)^{k} = \delta^{k} \sum_{i=0}^{k} {k \choose i} \mathbb{E} \Big(X^{i}\big( e^{-\alpha Z} - 1\big)^{k-i}\Big) d^{k-i}. \end{align*} Since $X$ and $Z$ are independent and exponentially distributed with mean $1$, we have $\mathbb{E} X^i = i!$ and \begin{align*} {k \choose i} \mathbb{E} \Big(X^{i}\big( e^{-\alpha Z} - 1\big)^{k-i}\Big) =&\ \frac{k!}{(k-i)!} \mathbb{E} \big( e^{-\alpha Z} - 1\big)^{k-i} = \frac{k!}{(k-i)!} \int_{0}^{\infty} \big(e^{-\alpha z} - 1\big)^{k-i} e^{-z} dz. \end{align*} Using integration by parts, \begin{align*} &\int_{0}^{\infty} \big(e^{-\alpha z} - 1\big)^{k-i} e^{-z} dz \\ =&\ (k-i) (-\alpha) \int_{0}^{\infty} \big(e^{-\alpha z} - 1\big)^{k-i-1} e^{-(1+\alpha)z} dz\\ =&\ (k-i)(k-i-1) (-\alpha)^2 \frac{1}{1 + \alpha} \int_{0}^{\infty} \big(e^{-\alpha z} - 1\big)^{k-i-2} e^{-(1+2\alpha)z} dz\\ & \ldots \\ =&\ (k-i)! (-\alpha)^{k-i} \frac{1}{1 + \alpha}\frac{1}{1 + 2\alpha} \ldots \frac{1}{1+(k-i-1)\alpha} \int_{0}^{\infty} e^{-(1+(k-i)\alpha)z} dz\\ =&\ (k-i)! (-\alpha)^{k-i} \frac{1}{1 + \alpha}\frac{1}{1 + 2\alpha} \ldots \frac{1}{1+(k-i)\alpha}. \end{align*} \hfill $\square$\endproof \ACKNOWLEDGMENT{We thank Zhuosong Zhang for proving Lemma~\ref{lem:hosmec}. We thank Yige Hong and Zhuoyang Liu for producing some figures of this paper. Xiao Fang is partially supported by Hong Kong RGC grants 24301617, 14302418 and 14304917, a CUHK direct grant and a CUHK start-up grant. J. G. Dai is partially supported by NSF grant CMMI-1537795.} \bibliographystyle{informs2014}

https://arxiv.org/abs/1906.03125

A new Federer-type characterization of sets of finite perimeter in metric spaces

Federer's characterization states that a set $E\subset \mathbb{R}^n$ is of finite perimeter if and only if $\mathcal H^{n-1}(\partial^*E)<\infty$. Here the measure-theoretic boundary $\partial^*E$ consists of those points where both $E$ and its complement have positive upper density. We show that the characterization remains true if $\partial^*E$ is replaced by a smaller boundary consisting of those points where the \emph{lower} densities of both $E$ and its complement are at least a given number. This result is new even in Euclidean spaces but we prove it in a more general complete metric space that is equipped with a doubling measure and supports a Poincaré inequality.

\section{Introduction} Federer's \cite{Fed} characterization of sets of finite perimeter states that a set $E\subset {\mathbb R}^n$ is of finite perimeter if and only if $\mathcal H^{n-1}(\partial^*E)<\infty$, where $\mathcal H^{n-1}$ is the $n-1$-dimensional Hausdorff measure and $\partial^*E$ is the measure-theoretic boundary; see Section \ref{sec:preliminaries} for definitions. A similar characterization holds also in the abstract setting of complete metric spaces $(X,d,\mu)$ that are equipped with a doubling measure $\mu$ and support a Poincar\'e inequality; in such spaces one replaces the $n-1$-dimensional Hausdorff measure with the \emph{codimension one} Hausdorff measure $\mathcal H$. The ``only if'' direction of the characterization was shown in metric spaces by Ambrosio \cite{A1}, and the ``if'' direction was recently shown by the author \cite{L-Fedchar}. Federer also showed that if a set $E\subset {\mathbb R}^n$ is of finite perimeter, then $\mathcal H^{n-1}(\partial^*E\setminus \Sigma_{1/2}E)=0$, where the boundary $\Sigma_{1/2}E$ consists of those points where both $E$ and its complement have density exactly $1/2$. In metric spaces we similarly have $\mathcal H(\partial^*E\setminus \Sigma_{\gamma}E)=0$, where $0<\gamma\le 1/2$ is a suitable constant depending on the space and the \emph{strong boundary} $\Sigma_{\gamma}E$ is defined by \[ \Sigma_{\gamma} E:=\left\{x\in X:\, \liminf_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))}\ge \gamma\ \ \textrm{and}\ \ \liminf_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))}\ge \gamma\right\}. \] This raises the natural question of whether the condition $\mathcal H(\Sigma_{\beta} E)<\infty$ for some $\beta>0$, which appears much weaker than $\mathcal H(\partial^* E)<\infty$, is already enough to imply that $E$ is of finite perimeter. Recently Chleb\'ik \cite{Chl} posed this question in Euclidean spaces and noted that the (positive) answer is known only when $n=1$. In the current paper we show that this characterization does indeed hold in every Euclidean space and even in the much more general metric spaces that we consider. \begin{theorem}\label{thm:main theorem} Let $(X,d,\mu)$ be a complete metric space with $\mu$ doubling and supporting a $(1,1)$-Poincar\'e inequality. Let $\Omega\subset X$ be an open set and let $E\subset X$ be a $\mu$-measurable set with $\mathcal H(\Sigma_{\beta} E\cap \Omega)<\infty$, where $0<\beta\le 1/2$ only depends on the doubling constant of the measure and the constants in the Poincar\'e inequality. Then $P(E,\Omega)<\infty$. \end{theorem} Explicitly, in the Euclidean space ${\mathbb R}^n$ with $n\ge 2$, we can take (see \eqref{eq:choice of beta in Euclidean space}) \[ \beta= \frac{n^{13n/2}}{2^{26n^2+64n+15}\omega_n^{13}}, \] where $\omega_n$ is the volume of the Euclidean unit ball. Our strategy is to show that if $\mathcal H(\Sigma_{\beta}E\cap \Omega)<\infty$, then $\mathcal H((\partial^*E\setminus \Sigma_{\beta}E)\cap \Omega)=0$ and so the result follows from the previously known Federer's characterization. Our proof consists essentially of two steps. First in Section \ref{sec:strong boundary points}, we show that for every point in the measure-theoretic boundary $\partial^*E$, arbitrarily close there is a point in the strong boundary $\Sigma_{\beta} E$. Then, after some preliminary results concerning connected components of sets of finite perimeter as well as functions of least gradient in Sections \ref{sec:components} and \ref{sec:least gradient}, in Section \ref{sec:constructing a quasiconvex space} we show that there exists an open set $V$ containing a suitable part of $\Sigma_{\beta}E$ such that $X\setminus V$ is itself a metric space with rather good properties. Thus we can apply the first step in this space. In Section \ref{sec:proof of the main result} we combine the two steps to prove Theorem \ref{thm:main theorem}. \paragraph{Acknowledgments.} The author wishes to thank Nageswari Shanmugalingam for many helpful comments as well as for discussions on constructing spaces where the Mazurkiewicz metric agrees with the ordinary one; Anders Bj\"orn also for discussions on constructing such spaces; and Olli Saari for discussions on finding strong boundary points. \section{Notation and definitions}\label{sec:preliminaries} In this section we introduce the notation, definitions, and assumptions that are employed in the paper. Throughout this paper, $(X,d,\mu)$ is a complete metric space that is equip\-ped with a metric $d$ and a Borel regular outer measure $\mu$ satisfying a doubling property, meaning that there exists a constant $C_d\ge 1$ such that \[ 0<\mu(B(x,2r))\le C_d\mu(B(x,r))<\infty \] for every ball $B(x,r):=\{y\in X:\,d(y,x)<r\}$, with $x\in X$ and $r>0$. Closed balls are denoted by $\overline{B}(x,r):=\{y\in X:\,d(y,x)\le r\}$. By iterating the doubling condition, we obtain that for every $x\in X$ and $y\in B(x,R)$ with $0<r\le R<\infty$, we have \begin{equation}\label{eq:homogenous dimension} \frac{\mu(B(y,r))}{\mu(B(x,R))}\ge \frac{1}{C_d^2}\left(\frac{r}{R}\right)^{s}, \end{equation} where $s>1$ only depends on the doubling constant $C_d$. Given a ball $B=B(x,r)$ and $\beta>0$, we sometimes abbreviate $\beta B:=B(x,\beta r)$; note that in a metric space, a ball (as a set) does not necessarily have a unique center point and radius, but these will be prescribed for all the balls that we consider. We assume that $X$ consists of at least $2$ points. When we want to state that a constant $C$ depends on the parameters $a,b, \ldots$, we write $C=C(a,b,\ldots)$. When a property holds outside a set of $\mu$-measure zero, we say that it holds almost everywhere, abbreviated a.e. All functions defined on $X$ or its subsets will take values in $[-\infty,\infty]$. As a complete metric space equipped with a doubling measure, $X$ is proper, that is, closed and bounded sets are compact. Since $X$ is proper, for any open set $\Omega\subset X$ we define $L_{\mathrm{loc}}^1(\Omega)$ to be the space of functions that are in $L^1(\Omega')$ for every open $\Omega'\Subset\Omega$. Here $\Omega'\Subset\Omega$ means that $\overline{\Omega'}$ is a compact subset of $\Omega$. Other local spaces of functions are defined analogously. For any set $A\subset X$ and $0<R<\infty$, the restricted Hausdorff content of codimension one is defined by \[ \mathcal{H}_{R}(A):=\inf\left\{ \sum_{j\in I} \frac{\mu(B(x_{j},r_{j}))}{r_{j}}:\,A\subset \bigcup_{j\in I}B(x_{j},r_{j}),\,r_{j}\le R,\,I\subset{\mathbb N}\right\}. \] The codimension one Hausdorff measure of $A\subset X$ is then defined by \[ \mathcal{H}(A):=\lim_{R\rightarrow 0}\mathcal{H}_{R}(A). \] In the Euclidean space ${\mathbb R}^n$ (equipped with the Euclidean metric and the $n$-dimensional Lebesgue measure) this is comparable to the $n-1$-dimensional Hausdorff measure. By a curve we mean a rectifiable continuous mapping from a compact interval of the real line into $X$. The length of a curve $\gamma$ is denoted by $\ell_{\gamma}$. We will assume every curve to be parametrized by arc-length, which can always be done (see e.g. \cite[Theorem~3.2]{Hj}). A nonnegative Borel function $g$ on $X$ is an upper gradient of a function $u$ on $X$ if for all nonconstant curves $\gamma$, we have \begin{equation}\label{eq:definition of upper gradient} |u(x)-u(y)|\le \int_{\gamma} g\,ds:=\int_0^{\ell_{\gamma}} g(\gamma(s))\,ds, \end{equation} where $x$ and $y$ are the end points of $\gamma$. We interpret $|u(x)-u(y)|=\infty$ whenever at least one of $|u(x)|$, $|u(y)|$ is infinite. Upper gradients were originally introduced in \cite{HK}. The $1$-modulus of a family of curves $\Gamma$ is defined by \[ \Mod_{1}(\Gamma):=\inf\int_{X}\rho\, d\mu \] where the infimum is taken over all nonnegative Borel functions $\rho$ such that $\int_{\gamma}\rho\,ds\ge 1$ for every curve $\gamma\in\Gamma$. A property is said to hold for $1$-a.e. curve if it fails only for a curve family with zero $1$-modulus. If $g$ is a nonnegative $\mu$-measurable function on $X$ and (\ref{eq:definition of upper gradient}) holds for $1$-a.e. curve, we say that $g$ is a $1$-weak upper gradient of $u$. By only considering curves $\gamma$ in a set $A\subset X$, we can talk about a function $g$ being a ($1$-weak) upper gradient of $u$ in $A$.\label{curve discussion} Given an open set $\Omega\subset X$, we let \[ \Vert u\Vert_{N^{1,1}(\Omega)}:=\Vert u\Vert_{L^1(\Omega)}+\inf \Vert g\Vert_{L^1(\Omega)}, \] where the infimum is taken over all upper gradients $g$ of $u$ in $\Omega$. Then we define the Newton-Sobolev space \[ N^{1,1}(\Omega):=\{u:\|u\|_{N^{1,1}(\Omega)}<\infty\}. \] In ${\mathbb R}^n$ this coincides, up to a choice of pointwise representatives, with the usual Sobolev space $W^{1,1}(\Omega)$; this is shown in Theorem 4.5 of \cite{S}, where the Newton-Sobolev space was originally introduced. We understand Newton-Sobolev functions to be defined at every point $x\in \Omega$ (even though $\Vert \cdot\Vert_{N^{1,1}(\Omega)}$ is then only a seminorm). It is known that for every $u\in N_{\mathrm{loc}}^{1,1}(\Omega)$ there exists a minimal $1$-weak upper gradient of $u$ in $\Omega$, always denoted by $g_{u}$, satisfying $g_{u}\le g$ a.e. in $\Omega$ for any other $1$-weak upper gradient $g\in L_{\mathrm{loc}}^{1}(\Omega)$ of $u$ in $\Omega$, see \cite[Theorem 2.25]{BB}. In ${\mathbb R}^n$, the minimal $1$-weak upper gradient coincides (a.e.) with $|\nabla u|$, see \cite[Corollary A.4]{BB}. We will assume throughout the paper that $X$ supports a $(1,1)$-Poincar\'e inequality, meaning that there exist constants $C_P\ge 1$ and $\lambda \ge 1$ such that for every ball $B(x,r)$, every $u\in L^1_{\mathrm{loc}}(X)$, and every upper gradient $g$ of $u$, we have \[ \vint{B(x,r)}|u-u_{B(x,r)}|\, d\mu \le C_P r\vint{B(x,\lambda r)}g\,d\mu, \] where \[ u_{B(x,r)}:=\vint{B(x,r)}u\,d\mu :=\frac 1{\mu(B(x,r))}\int_{B(x,r)}u\,d\mu. \] As \label{quasiconvex and geodesic}a complete metric space equipped with a doubling measure and supporting a Poincar\'e inequality, $X$ is \emph{quasiconvex}, meaning that for every pair of points $x,y\in X$ there is a curve $\gamma$ with $\gamma(0)=x$, $\gamma(\ell_{\gamma})=y$, and $\ell_{\gamma}\le Cd(x,y)$, where $C$ is a constant and only depends on $C_d$ and $C_P$, see e.g. \cite[Theorem 4.32]{BB}. Thus a biLipschitz change in the metric gives a geodesic space (see \cite[Section 4.7]{BB}). Since Theorem \ref{thm:main theorem} is easily seen to be invariant under such a biLipschitz change in the metric, we can assume that $X$ is geodesic. By \cite[Theorem 4.39]{BB}, in the Poincar\'e inequality we can now choose $\lambda=1$. The $1$-capacity of a set $A\subset X$ is defined by \[ \capa_1(A):=\inf \Vert u\Vert_{N^{1,1}(X)}, \] where the infimum is taken over all functions $u\in N^{1,1}(X)$ satisfying $u\ge 1$ in $A$. The variational $1$-capacity of a set $A\subset \Omega$ with respect to an open set $\Omega\subset X$ is defined by \[ \rcapa_1(A,\Omega):=\inf \int_X g_u \,d\mu, \] where the infimum is taken over functions $u\in N^{1,1}(X)$ satisfying $u=0$ in $X\setminus\Omega$ and $u\ge 1$ in $A$, and $g_u$ is the minimal $1$-weak upper gradient of $u$ (in $X$). By truncation, we see that we can assume $0\le u\le 1$ on $X$. The variational $1$-capacity is an outer capacity in the sense that if $A\Subset \Omega$, then \begin{equation}\label{eq:rcapa outer capacity} \rcapa_{1}(A,\Omega) =\inf_{\substack{V\textrm{ open} \\A\subset V\subset \Omega}}\rcapa_{1}(V,\Omega); \end{equation} see \cite[Theorem 6.19(vii)]{BB}. For basic properties satisfied by capacities, such as monotonicity and countable subadditivity, see e.g. \cite{BB}. We say that a set $U\subset X$ is $1$-quasiopen\label{quasiopen} if for every $\varepsilon>0$ there exists an open set $G\subset X$ such that $\capa_1(G)<\varepsilon$ and $U\cup G$ is open. Next we present the definition and basic properties of functions of bounded variation on metric spaces, following \cite{M}. See also e.g. \cite{AFP, EvGa, Fed, Giu84, Zie89} for the classical theory in the Euclidean setting. Given an open set $\Omega\subset X$ and a function $u\in L^1_{\mathrm{loc}}(\Omega)$, we define the total variation of $u$ in $\Omega$ by \[ \|Du\|(\Omega):=\inf\left\{\liminf_{i\to\infty}\int_\Omega g_{u_i}\,d\mu:\, u_i\in N^{1,1}_{\mathrm{loc}}(\Omega),\, u_i\to u\textrm{ in } L^1_{\mathrm{loc}}(\Omega)\right\}, \] where each $g_{u_i}$ is the minimal $1$-weak upper gradient of $u_i$ in $\Omega$. In ${\mathbb R}^n$ this agrees with the usual Euclidean definition involving distributional derivatives, see e.g. \cite[Proposition 3.6, Theorem 3.9]{AFP}. (In \cite{M}, local Lipschitz constants were used in place of upper gradients, but the theory can be developed similarly with either definition.) We say that a function $u\in L^1(\Omega)$ is of bounded variation, and denote $u\in\mathrm{BV}(\Omega)$, if $\|Du\|(\Omega)<\infty$. For an arbitrary set $A\subset X$, we define \[ \|Du\|(A):=\inf\{\|Du\|(W):\, A\subset W,\,W\subset X \text{ is open}\}. \] If $u\in L^1_{\mathrm{loc}}(\Omega)$ and $\Vert Du\Vert(\Omega)<\infty$, then $\|Du\|(\cdot)$ is a Borel regular outer measure on $\Omega$ by \cite[Theorem 3.4]{M}. A $\mu$-measurable set $E\subset X$ is said to be of finite perimeter if $\|D\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_E\|(X)<\infty$, where $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_E$ is the characteristic function of $E$. The perimeter of $E$ in $\Omega$ is also denoted by \[ P(E,\Omega):=\|D\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_E\|(\Omega). \] The measure-theoretic interior of a set $E\subset X$ is defined by \begin{equation}\label{eq:measure theoretic interior} I_E:= \left\{x\in X:\,\lim_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))}=0\right\}, \end{equation} and the measure-theoretic exterior by \[ O_E:= \left\{x\in X:\,\lim_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))}=0\right\}. \] The measure-theoretic boundary $\partial^{*}E$ is defined as the set of points $x\in X$ at which both $E$ and its complement have nonzero upper density, i.e. \[ \limsup_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))}>0\quad \textrm{and}\quad\limsup_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))}>0. \] Note that the space $X$ is always partitioned into the disjoint sets $I_E$, $O_E$, and $\partial^*E$. By Lebesgue's differentiation theorem (see e.g. \cite[Chapter 1]{Hei}), for a $\mu$-measurable set $E$ we have $\mu(E\Delta I_E)=0$, where $\Delta$ is the symmetric difference. Given a number $0<\gamma\le 1/2$, we also define the strong boundary \begin{equation}\label{eq:strong boundary} \Sigma_{\gamma} E:=\left\{x\in X:\, \liminf_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))}\ge \gamma\ \, \textrm{and}\ \, \liminf_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))}\ge \gamma\right\}. \end{equation} For an open set $\Omega\subset X$ and a $\mu$-measurable set $E\subset X$ with $P(E,\Omega)<\infty$, we have $\mathcal H((\partial^*E\setminus \Sigma_{\gamma}E)\cap\Omega)=0$ for $\gamma \in (0,1/2]$ that only depends on $C_d$ and $C_P$, see \cite[Theorem 5.4]{A1}. Moreover, for any Borel set $A\subset\Omega$ we have \begin{equation}\label{eq:def of theta} P(E,A)=\int_{\partial^{*}E\cap A}\theta_E\,d\mathcal H, \end{equation} where $\theta_E\colon \Omega\to [\alpha,C_d]$ with $\alpha=\alpha(C_d,C_P)>0$, see \cite[Theorem 5.3]{A1} and \cite[Theorem 4.6]{AMP}. The following coarea formula is given in \cite[Proposition 4.2]{M}: if $\Omega\subset X$ is an open set and $u\in L^1_{\mathrm{loc}}(\Omega)$, then \begin{equation}\label{eq:coarea} \|Du\|(\Omega)=\int_{-\infty}^{\infty}P(\{u>t\},\Omega)\,dt, \end{equation} where we abbreviate $\{u>t\}:=\{x\in \Omega:\,u(x)>t\}$. If $\Vert Du\Vert(\Omega)<\infty$, then \eqref{eq:coarea} holds with $\Omega$ replaced by any Borel set $A\subset \Omega$. We know that for an open set $\Omega\subset X$, an arbitrary set $A\subset \Omega$, and any $\mu$-measurable sets $E_1,E_2\subset X$, we have \begin{equation}\label{eq:lattice property of sets of finite perimeter} P(E_1\cap E_2,A)+P(E_1\cup E_2,A)\le P(E_1,A)+P(E_2,A); \end{equation} for a proof in the case $A=\Omega$ see \cite[Proposition 4.7]{M}, and then the general case follows by approximation. Using this fact as well as the lower semicontinuity of the total variation with respect to $L_{\mathrm{loc}}^1$-convergence in open sets, we have for any $E_1,E_2\ldots \subset X$ that \begin{equation}\label{eq:perimeter of countable union} P\Bigg(\bigcup_{j=1}^{\infty}E_j,\Omega\Bigg) \le \sum_{j=1}^{\infty}P(E_j,\Omega). \end{equation} Applying the Poincar\'e inequality to sequences of approximating $N^{1,1}_{\mathrm{loc}}$-functions in the definition of the total variation, we get the following $\mathrm{BV}$ version: for every ball $B(x,r)$ and every $u\in L^1_{\mathrm{loc}}(X)$, we have \[ \int_{B(x,r)}|u-u_{B(x,r)}|\,d\mu \le C_P r \Vert Du\Vert (B(x, r)). \] Recall here and from now on that we take the constant $\lambda$ to be $1$, and so it does not appear in the inequalities. For a $\mu$-measurable set $E\subset X$, by considering the two cases $(\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_E)_{B(x,r)}\le 1/2$ and $(\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_E)_{B(x,r)}\ge 1/2$, from the above we get the relative isoperimetric inequality \begin{equation}\label{eq:relative isoperimetric inequality} \min\{\mu(B(x,r)\cap E),\,\mu(B(x,r)\setminus E)\}\le 2 C_P rP(E,B(x,r)). \end{equation} From the $(1,1)$-Poincar\'e inequality, by \cite[Theorem 4.21, Theorem 5.51]{BB} we also get the following Sobolev inequality: if $x\in X$, $0<r<\frac{1}{4}\diam X$, and $u\in N^{1,1}(X)$ with $u=0$ in $X\setminus B(x,r)$, then \begin{equation}\label{eq:sobolev inequality} \int_{B(x,r)} |u|\,d\mu \le C_S r \int_{B(x,r)} g_u\,d\mu \end{equation} for a constant $C_S=C_S(C_d,C_P)\ge 1$. For any $\mu$-measurable set $E\subset B(x,r)$, applying the Sobolev inequality to a suitable sequence approximating $u$, we get the isoperimetric inequality \begin{equation}\label{eq:isop inequality with zero boundary values} \mu(E)\le C_S r P(E,X). \end{equation} The lower and upper approximate limits of a function $u$ on an open set $\Omega$ are defined respectively by \begin{equation}\label{eq:lower approximate limit} u^{\wedge}(x): =\sup\left\{t\in{\mathbb R}:\,\lim_{r\to 0}\frac{\mu(B(x,r)\cap\{u<t\})}{\mu(B(x,r))}=0\right\} \end{equation} and \begin{equation}\label{eq:upper approximate limit} u^{\vee}(x): =\inf\left\{t\in{\mathbb R}:\,\lim_{r\to 0}\frac{\mu(B(x,r)\cap\{u>t\})}{\mu(B(x,r))}=0\right\} \end{equation} for $x\in \Omega$. Unlike Newton-Sobolev functions, we understand $\mathrm{BV}$ functions to be equivalence classes of a.e. defined functions, but $u^{\wedge}$ and $u^{\vee}$ are pointwise defined. The $\mathrm{BV}$-capacity of a set $A\subset X$ is defined by \[ \capa_{\mathrm{BV}}(A):=\inf \left(\Vert u\Vert_{L^1(X)}+\Vert Du\Vert(X)\right), \] where the infimum is taken over all $u\in\mathrm{BV}(X)$ with $u\ge 1$ in a neighborhood of $A$. By \cite[Theorem 4.3]{HaKi} we know that for some constant $C_{\textrm{cap}}=C_{\textrm{cap}}(C_d,C_P)\ge 1$ and every $A\subset X$, we have \begin{equation}\label{eq:Newtonian and BV capacities are comparable} \capa_1(A)\le C_{\textrm{cap}}\capa_{\mathrm{BV}}(A). \end{equation} We also define a variational $\mathrm{BV}$-capacity for any $A\subset\Omega$, with $\Omega\subset X$ open, by \[ \rcapa^{\vee}_{\mathrm{BV}}(A,\Omega):=\inf \Vert Du\Vert(X), \] where the infimum is taken over functions $u\in \mathrm{BV}(X)$ such that $u^{\wedge}=u^{\vee}= 0$ $\mathcal H$-a.e. in $X\setminus \Omega$ and $u^{\vee}\ge 1$ $\mathcal H$-a.e. in $A$. By \cite[Theorem 5.7]{L-SS} we know that \begin{equation}\label{eq:variational one and BV capacity} \rcapa_{1}(A,\Omega)\le C_{\textrm{r}}\rcapa^{\vee}_{\mathrm{BV}}(A,\Omega) \end{equation} for a constant $C_{\textrm{r}}=C_{\textrm{r}}(C_d,C_P)\ge 1$. \textbf{Standing assumptions:} In Section \ref{sec:strong boundary points} we will consider a different metric space $Z$ (which will later be taken to be a subset of $X$), but in Sections \ref{sec:components} to \ref{sec:proof of the main result} we will assume that $(X,d,\mu)$ is a complete, geodesic metric space that is equipped with the doubling Radon measure $\mu$ and supports a $(1,1)$-Poincar\'e inequality with $\lambda=1$. \section{Strong boundary points}\label{sec:strong boundary points} In this section we consider a complete metric space $(Z,\widehat{d},\mu)$ where $\mu$ is a Borel regular outer measure and doubling with constant $\widehat{C}_d\ge 1$. We define the Mazurkiewicz metric \begin{equation}\label{eq:widehat d c} \widehat{d}_M(x,y):=\inf\{\diam F:\,F\subset Z\textrm{ is a continuum containing }x,y\}, \quad x,y\in Z, \end{equation} and we assume the space to be ``geodesic'' in the sense that $\widehat{d}_M = \widehat{d}$. As usual, a continuum means a compact connected set. \begin{definition} We say that $(x_0,\ldots,x_m)$ is an $\varepsilon$-chain from $x_0$ to $x_m$ if $\widehat{d}(x_j,x_{j+1})<\varepsilon$ for all $j=0,\ldots,m-1$. \end{definition} The following proposition gives the existence of a strong boundary point. \begin{proposition}\label{prop:strong boundary point} Let $x_0\in Z$, $R>0$, and let $E\subset Z$ be a $\mu$-measurable set such that \begin{equation}\label{eq:half measure assumption} \frac{1}{2\widehat{C}_d^2}\le \frac{\mu(B(x_0,R)\cap E)}{\mu(B(x_0,R))}\le 1-\frac{1}{2\widehat{C}_d^2}. \end{equation} Then there exists a point $x\in B(x_0,6 R)$ such that \begin{equation}\label{eq:desired density point} \frac{1}{4 \widehat{C}_d^{12}}\le \liminf_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))} \le \limsup_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))} \le 1-\frac{1}{4 \widehat{C}_d^{12}}. \end{equation} \end{proposition} \begin{proof} The proof is by suitable iteration, where we consider two options. \textbf{Case 1.} Suppose that \begin{equation}\label{eq:E smaller than half everywhere} \frac{\mu(B(x,2^{-2}R)\cap E)}{\mu(B(x,2^{-2}R))}<\frac{1}{2} \end{equation} for all $x\in B(x_0,R)$; the case ``$>$'' is considered analogously. Define a ``bad'' set \[ P:=\left\{x\in B(x_0,R):\,\frac{\mu(B(x,2^{-2j}R)\cap E)}{\mu(B(x,2^{-2j}R))} \le \frac{1}{4\widehat{C}_d^6}\ \ \textrm{for some }j\in{\mathbb N}\right\}. \] For every $x\in P$ there is a radius $r_x\le R/20\le R$ such that \[ \frac{\mu(B(x,5r_x)\cap E)}{\mu(B(x,5r_x))}\le \frac{1}{4\widehat{C}_d^6}. \] Thus $\{B(x,r_x)\}_{x\in P}$ is a covering of $P$. By the $5$-covering theorem, pick a countable collection of pairwise disjoint balls $\{B(x_j,r_j)\}_{j=1}^{\infty}$ such that $P\subset \bigcup_{j=1}^{\infty}B(x_j,5r_j)$. Now \begin{align*} \mu(P\cap E)\le \sum_{j=1}^{\infty}\mu(B(x_j,5r_j)\cap E) &\le \frac{1}{4 \widehat{C}_d^6}\sum_{j=1}^{\infty}\mu(B(x_j,5r_j))\\ &\le \frac{1}{4 \widehat{C}_d^3}\sum_{j=1}^{\infty}\mu(B(x_j,r_j))\\ &\le \frac{1}{4 \widehat{C}_d^3}\mu(B(x_0,2R))\\ &\le \frac{1}{4 \widehat{C}_d^2}\mu(B(x_0,R)). \end{align*} Thus \begin{align*} \mu(P) &=\mu(P\cap E)+\mu(P\setminus E)\\ &\le \frac{1}{4\widehat{C}_d^2}\mu(B(x_0,R))+\mu(B(x_0,R)\setminus E)\\ &\le \frac{1}{4\widehat{C}_d^2}\mu(B(x_0,R))+\Bigg(1-\frac{1}{2\widehat{C}_d^2}\Bigg)\mu(B(x_0,R))\quad\textrm{by }\eqref{eq:half measure assumption}\\ &\le \Bigg(1-\frac{1}{4\widehat{C}_d^2}\Bigg)\mu(B(x_0,R)). \end{align*} In particular, there is a point $y\in B(x_0,R)\setminus P$. Now there are two options. \textbf{Case 1(a).} The first option is that for each $j\in{\mathbb N}$, we have \[ \frac{\mu(B(y,2^{-2j}R)\cap E)}{\mu(B(y,2^{-2j}R))}<\frac{1}{2} \] and then in fact \[ \frac{1}{4\widehat{C}_d^6}\le \frac{\mu(B(y,2^{-2j}R)\cap E)}{\mu(B(y,2^{-2j}R))}<\frac{1}{2}, \] for all $j\in{\mathbb N}$, since $y\in B(x_0,R)\setminus P$. From this we easily find that \eqref{eq:desired density point} holds (with $x=y$). \textbf{Case 1(b).} The second option is that there is a smallest index $l\ge 2$ such that \[ \frac{\mu(B(y,2^{-2l}R)\cap E)}{\mu(B(y,2^{-2l}R))}\ge\frac{1}{2}. \] Then \[ \frac{1}{2\widehat{C}_d^2}\le \frac{\mu(B(y,2^{-2l+2}R)\cap E)}{\mu(B(y,2^{-2l+2}R))} < \frac{1}{2}, \] and also \[ \frac{1}{4\widehat{C}_d^6}\le\frac{\mu(B(y,2^{-2j}R)\cap E)}{\mu(B(y,2^{-2j}R))} <\frac{1}{2}\quad\textrm{for all }j=1,\ldots,l-2. \] Note that regardless of the direction of the inequality in \eqref{eq:E smaller than half everywhere}, we get \[ \frac{1}{2\widehat{C}_d^2}\le \frac{\mu(B(y,2^{-2l+2}R)\cap E)}{\mu(B(y,2^{-2l+2}R))} < 1-\frac{1}{2\widehat{C}_d^2} \] and \begin{equation}\label{eq:doubling constant six estimate} \frac{1}{4\widehat{C}_d^6}\le\frac{\mu(B(y,2^{-2j}R)\cap E)}{\mu(B(y,2^{-2j}R))} \le1-\frac{1}{4\widehat{C}_d^6}\quad\textrm{for all }j=1,\ldots,l-2. \end{equation} \textbf{Case 2.} Alternatively, suppose that we find two points $x,y\in B(x_0,R)$ such that \[ \frac{\mu(B(x,2^{-2}R)\cap E)}{\mu(B(x,2^{-2}R))}\ge \frac{1}{2} \] and \[ \frac{\mu(B(y,2^{-2}R)\cap E)}{\mu(B(y,2^{-2}R))}\le \frac{1}{2}. \] Then, using the fact that $\widehat{d}_M=\widehat{d}$, we find a continuum $F$ that contains $x$ and $y$ and is contained in $B(x_0,3 R)$. Since $F$ is connected, for every $\varepsilon>0$ there is an $\varepsilon$-chain in $F$ from $x$ to $y$. In particular, we find an $R/4$-chain in $F$ from $x$ to $y$. Let $z$ be the last point in the chain for which we have \[ \frac{\mu(B(z,2^{-2}R)\cap E)}{\mu(B(z,2^{-2}R))}\ge \frac{1}{2}. \] If $z=y$, then we have \[ \frac{\mu(B(z,2^{-2}R)\cap E)}{\mu(B(z,2^{-2}R))}= \frac{1}{2}. \] Else there exists $w\in F$ with $\widehat{d}(z,w)<R/4$ and \[ \frac{\mu(B(w,2^{-2}R)\cap E)}{\mu(B(w,2^{-2}R))}< \frac{1}{2}\quad\textrm{and thus} \quad \frac{\mu(B(w,2^{-2}R)\setminus E)}{\mu(B(z,2^{-1}R))}\ge \frac{1}{2\widehat{C}_d^2}. \] Now \begin{align*} \frac{\mu(B(z,2^{-1}R)\cap E)}{\mu(B(z,2^{-1}R))} &= \frac{\mu(B(z,2^{-1}R))-\mu(B(z,2^{-1}R)\setminus E)}{\mu(B(z,2^{-1}R))}\\ &\le \frac{\mu(B(z,2^{-1}R))-\mu(B(w,2^{-2}R)\setminus E)}{\mu(B(z,2^{-1}R))}\\ &\le 1-\frac{1}{2\widehat{C}_d^2}. \end{align*} Conversely, \[ \frac{\mu(B(z,2^{-1}R)\cap E)}{\mu(B(z,2^{-1}R))} \ge\frac{\mu(B(z,2^{-2}R)\cap E)}{\widehat{C}_d\mu(B(z,2^{-2}R))} \ge \frac{1}{2\widehat{C}_d}. \] In conclusion, there is $z\in B(x_0,3R)$ with \[ \frac{1}{2\widehat{C}_d^2}\le \frac{\mu(B(z,2^{-1}R)\cap E)}{\mu(B(z,2^{-1}R))}\le 1-\frac{1}{2\widehat{C}_d^2}; \] note that this holds also in the case $z=y$. To summarize, in Case 1(a) we obtain infinitely many balls (and then we are done), in Case 1(b) we obtain the $l-1$ new balls $B(y,2^{-2}R),\ldots,B(y,2^{-2l+2}R)$, where $B(y,2^{-2l+2}R)$ satisfies \eqref{eq:half measure assumption}, and in Case (2) we obtain one new ball satisfying \eqref{eq:half measure assumption}. By iterating the procedure and concatenating the new balls obtained in each step to the previous list of balls, we find a sequence of balls with center points $x_k\in B(x_{k-1},3 r_{k-1})$ and radii $r_k$ such that $r_0=R$, $r_k\in [r_{k-1}/4,r_{k-1}/2]$, and (recall \eqref{eq:doubling constant six estimate}) \[ \frac{1}{4\widehat{C}_d^6}\le \frac{\mu(B(x_k,r_k)\cap E)}{\mu(B(x_k,r_k))} \le 1-\frac{1}{4\widehat{C}_d^6} \] for all $k\in{\mathbb N}$. (Note that several consecutive balls in this sequence will have the same center points if they are obtained from Case 1.) By completeness of the space we find $x\in Z$ such that $x_k\to x$. For each $l=0,1,\ldots$ we have \[ d(x,x_l)\le \sum_{k=l}^{\infty}d(x_k,x_{k+1}) \le 3\sum_{k=l}^{\infty}r_k\le 6 r_l. \] In particular, $d(x,x_0)\le 6 R$. Now $B(x_l,r_l)\subset B(x,7 r_l)\subset B(x_l,13 r_l)$ for all $l\in{\mathbb N}$, and so \[ \frac{\mu(B(x,7 r_l)\cap E)}{\mu(B(x,7 r_l))} \ge \frac{\mu(B(x_l,r_l)\cap E)}{\mu(B(x_l,13 r_l))} \ge \frac{1}{\widehat{C}_d^{4}}\frac{\mu(B(x_l,r_l)\cap E)}{\mu(B(x_l,r_l))} \ge \frac{1}{4 \widehat{C}_d^{10}} \] and similarly \[ \frac{\mu(B(x,7 r_l)\setminus E)}{\mu(B(x,7 r_l))} \ge \frac{\mu(B(x_l,r_l)\setminus E)}{\mu(B(x_l,13 r_l))} \ge \frac{1}{\widehat{C}_d^{4}} \frac{\mu(B(x_l,r_l)\setminus E)}{\mu(B(x_l,r_l))} \ge \frac{1}{4 \widehat{C}_d^{10}}. \] It follows that \[ \liminf_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))} \ge \frac{1}{4 \widehat{C}_d^{12}} \] and \[ \liminf_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))} \ge \frac{1}{4 \widehat{C}_d^{12}}, \] proving \eqref{eq:desired density point}. \end{proof} \begin{corollary}\label{cor:density points} Let $x_0\in Z$, $R>0$, and let $E\subset Z$ be a $\mu$-measurable set such that \[ 0< \mu(B(x_0,R)\cap E)<\mu(B(x_0,R)). \] Then there exists a point $x\in B(x_0,9 R)$ such that \begin{equation}\label{eq:strong boundary point} \frac{1}{4 \widehat{C}_d^{12}}\le \liminf_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))} \le \limsup_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))} \le 1-\frac{1}{4 \widehat{C}_d^{12}}. \end{equation} \end{corollary} \begin{proof} Again consider two cases. The first is that we find two points $y,z\in B(x_0,R)$ such that \[ \frac{\mu(B(y,2^{-1}R)\cap E)}{\mu(B(y,2^{-1}R))}\ge \frac{1}{2} \quad\textrm{and}\quad \frac{\mu(B(z,2^{-1}R)\cap E)}{\mu(B(z,2^{-1}R))}\le \frac{1}{2}. \] Then just as in the proof of Proposition \ref{prop:strong boundary point} Case 2, we find $w\in B(x_0,3R)$ with \[ \frac{1}{2\widehat{C}_d^2}\le \frac{\mu(B(w,R)\cap E)}{\mu(B(w,R))} \le 1-\frac{1}{2\widehat{C}_d^2}. \] Now Proposition \ref{prop:strong boundary point} gives a point $x\in B(w,6R)\subset B(x_0,9R)$ such that \eqref{eq:strong boundary point} holds. The second possible case is that for all $y\in B(x_0,R)$ we have \[ \frac{\mu(B(y,2^{-1}R)\cap E)}{\mu(B(y,2^{-1}R))}< \frac{1}{2} \] (the case ``$>$'' being analogous). By Lebesgue's differentiation theorem, we find a point $y\in I_E\cap B(x_0,R)$ (recall \eqref{eq:measure theoretic interior}) and then it is easy to find a radius $0<r\le R/2$ such that \[ \frac{1}{2\widehat{C}_d}\le \frac{\mu(B(y,r)\cap E)}{\mu(B(y,r))} <\frac{1}{2}. \] Now Proposition \ref{prop:strong boundary point} again gives a point $x\in B(y,6r)\subset B(x_0,4R)$ such that \eqref{eq:strong boundary point} holds. \end{proof} \section{Components of sets of finite perimeter}\label{sec:components} In Sections \ref{sec:components} to \ref{sec:proof of the main result} we assume that $(X,d,\mu)$ is a complete, geodesic metric space that is equipped with the doubling measure $\mu$ and supports a $(1,1)$-Poincar\'e inequality. In this section we consider connected components, or components for short, of sets of finite perimeter. The following is the main result of the section. \begin{proposition}\label{prop:connected components} Let $B(x,R)$ be a ball with $0<R<\frac{1}{4}\diam X$ and let $F\subset X$ be a closed set with $P(F,X)<\infty$. Denote the components of $F\cap \overline{B}(x,R)$ having nonzero $\mu$-measure by $F_1,F_2,\ldots$. Then $\mu\left(\overline{B}(x,R)\cap F\setminus \bigcup_{j=1}^{\infty}F_j\right)=0$, $P(F_j,B(x,R))<\infty$ for all $j\in{\mathbb N}$, and for any sets $A_j\subset F_j$ with $P(A_j,B(x,R))<\infty$ for all $j\in{\mathbb N}$ we have \[ P\Bigg(\bigcup_{j=1}^{\infty}A_j,B(x,R)\Bigg)=\sum_{j=1}^{\infty}P(A_j,B(x,R)). \] \end{proposition} Of course, there may be only finitely many $F_j$'s, and so we will always understand that some $F_j$'s can be empty. In fact, supposing that $\mu(F\cap B(x,R))>0$, we will know only after Lemma \ref{lem:H has measure zero} that any $F_j$'s are nonempty. Next we gather a number of preliminary results. Recall the definition of $1$-quasiopen sets from page \pageref{quasiopen}. \begin{proposition}[{\cite[Proposition 4.2]{L-Fed}}]\label{prop:set of finite perimeter is quasiopen} Let $\Omega\subset X$ be open and let $F\subset X$ be $\mu$-measurable with $P(F,\Omega)<\infty$. Then the sets $I_F\cap\Omega$ and $O_F\cap\Omega$ are $1$-quasiopen. \end{proposition} \begin{proposition}\label{prop:ae curve goes through boundary} Let $F\subset X$ with $P(F,X)<\infty$. Then for $1$-a.e. curve $\gamma$, $\gamma^{-1}(I_F)$ and $\gamma^{-1}(O_F)$ are relatively open subsets of $[0,\ell_{\gamma}]$. \end{proposition} \begin{proof} By Proposition \ref{prop:set of finite perimeter is quasiopen}, the sets $I_F$ and $O_F$ are $1$-quasiopen. Then by \cite[Remark 3.5]{S2}, they are also \emph{$1$-path open}, meaning that for $1$-a.e. curve $\gamma$ in $X$, the sets $\gamma^{-1}(I_F)$ and $\gamma^{-1}(O_F)$ are relatively open subsets of $[0,\ell_{\gamma}]$. \end{proof} For any set $A\subset X$, we define the \emph{measure-theoretic closure} as \begin{equation}\label{eq:measure theoretic closure} \overline{A}^m:=I_A\cup \partial^*A. \end{equation} \begin{lemma}\label{lem:inner capacity} Let $B(x,R)$ be a ball with $0<R<\frac{1}{4}\diam X$ and let $E_1\supset E_2\supset \ldots$ such that $P(E_j,B(x,R))<\infty$ for all $j\in{\mathbb N}$, and $\mu(E_j)\to 0$ and $P(E_j,B(x,R))\to 0$ as $j\to\infty$. Let $0<r<R$. Then \[ \capa_1(\overline{E_j}^m\cap B(x,r))\to 0. \] \end{lemma} \begin{proof} Take a cutoff function $\eta\in \Lip_c(B(x,R))$ with $0\le \eta\le 1$ on $X$, $\eta=1$ in $B(x,r)$, and $g_\eta\le 2/(R-r)$, where $g_{\eta}$ is the minimal $1$-weak upper gradient of $\eta$. Then for all $j\in{\mathbb N}$, by a Leibniz rule (see \cite[Proposition 4.2]{KKST3}) we have \[ \Vert D(\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{E_j}\eta)\Vert(X)=\Vert D(\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{E_j}\eta)\Vert(B(x,R)) \le \frac{2\mu(E_j)}{R-r}+P(E_j,B(x,R))\to 0 \] as $j\to\infty$. By \eqref{eq:variational one and BV capacity} and the fact that $(\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{E_j}\eta)^{\vee}=1$ in $\overline{E_j}^m\cap B(x,r)$, we get \begin{align*} \rcapa_1(\overline{E_j}^m\cap B(x,r),B(x,R)) &\le C_{\textrm{r}}\rcapa_{\mathrm{BV}}^{\vee}(\overline{E_j}^m\cap B(x,r),B(x,R))\\ &\le \Vert D(\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{E_j}\eta)\Vert(X)\to 0\quad\textrm{as }j\to\infty. \end{align*} Then by the Sobolev inequality \eqref{eq:sobolev inequality} we easily get \[ \capa_1(\overline{E_j}^m\cap B(x,r))\to 0. \] \end{proof} The variation measure is always absolutely continuous with respect to the $1$-capacity, in the following sense. \begin{lemma}[{\cite[Lemma 3.8]{L-SA}}]\label{lem:variation measure and capacity} Let $\Omega\subset X$ be an open set and let $u\in L^1_{\mathrm{loc}}(\Omega)$ with $\Vert Du\Vert(\Omega)<\infty$. Then for every $\varepsilon>0$ there exists $\delta>0$ such that if $A\subset \Omega$ with $\capa_1 (A)<\delta$, then $\Vert Du\Vert(A)<\varepsilon$. \end{lemma} \begin{lemma}\label{lem:coincidence of perimeter} Let $\Omega\subset X$ be open, let $F_1\subset F_2\subset X$ with $P(F_1,\Omega)<\infty$ and $P(F_2,\Omega)<\infty$, and let $A\subset \Omega$ such that for all $x\in A$, we have \[ \lim_{r\to 0}\frac{\mu(B(x,r)\cap (F_2\setminus F_1))}{\mu(B(x,r))}=0. \] Then $P(F_1,A)=P(F_2,A)$. \end{lemma} \begin{proof} First note that $P(F_2\setminus F_1,\Omega)<\infty$ by \eqref{eq:lattice property of sets of finite perimeter}, and then by \eqref{eq:def of theta} we have \[ P(F_2\setminus F_1,A)=0. \] Using \eqref{eq:lattice property of sets of finite perimeter} again, we have \[ P(F_2,A)\le P(F_1,A)+P(F_2\setminus F_1,A)=P(F_1,A) \] and \[ P(F_1,A)\le P(F_2,A)+P(F_2\setminus F_1,A)=P(F_2,A). \] \end{proof} The following lemma says that perimeter can always be controlled by the measure of a suitable ``curve boundary''. \begin{lemma}\label{lem:perimeter controlled by boundary} Let $\Omega\subset X$ be open, let $E\subset X$ be closed, and let $A\subset \Omega$ be such that $1$-a.e. curve $\gamma$ in $\Omega$ with $\gamma(0)\in I_E$ and $\gamma(\ell_{\gamma})\in X\setminus E$ intersects $A$. Then $P(E,\Omega)\le C_d\mathcal H(A)$. \end{lemma} \begin{proof} We can assume that $\mathcal H(A)<\infty$. Fix $\varepsilon>0$. We find a covering of $A$ by balls $\{B_j=B(x_j,r_j)\}_{j\in I}$, with $I\subset{\mathbb N}$, such that $r_j\le \varepsilon$ and \begin{equation}\label{eq:covering for A} \sum_{j\in I}\frac{\mu(B_j)}{r_j}\le \mathcal{H}(A)+\varepsilon. \end{equation} Denote the exceptional family of curves by $\Gamma$. Take a nonnegative Borel function $\rho$ such that $\Vert \rho\Vert_{L^1(\Omega)}<\varepsilon$ and $\int_{\gamma}\rho\,ds\ge 1$ for all $\gamma\in\Gamma$. Let \[ g:=\sum_{j\in I}\frac{\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{2B_j}}{r_j}+\rho. \] Then let \[ u(x):=\min\left\{1,\inf \int_{\gamma}g\,ds\right\}, \] where the infimum is taken over curves $\gamma$ (also constant curves) in $\Omega$ with $\gamma(0)= x$ and $\gamma(\ell_{\gamma})\in \Omega\setminus \left(E\cup\bigcup_{j\in I}2B_j\right)$. We know that $g$ is an upper gradient of $u$ in $\Omega$, see \cite[Lemma 5.25]{BB}. Moreover, $u$ is $\mu$-measurable by \cite[Theorem 1.11]{JJRRS}; strictly speaking this result is written for functions defined on the whole space, but the proof clearly works also for functions defined in an open set such as $\Omega$. If $x\in \Omega\setminus \left(E\cup\bigcup_{j\in I}2B_j\right)$, clearly $u(x)=0$. If $x\in I_E\setminus \bigcup_{j\in I}2B_j$, consider any curve $\gamma$ in $\Omega$ with $\gamma(0)= x$ and $\gamma(\ell_{\gamma})\in \Omega\setminus \left(E\cup\bigcup_{j\in I}2B_j\right)$. Then either $\int_{\gamma}\rho\,ds\ge 1$ or there is $t$ such that $\gamma(t)\in A$. In the latter case, for some $j\in I$ we have $\gamma(t)\in B_j$. Then \[ \int_{\gamma}g\,ds\ge \int_{\gamma}\frac{\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{2B_j}}{r_j}\,ds\ge 1. \] Thus $u(x)=1$, and so by Lebesgue's differentiation theorem we have $u=\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_E$ a.e. in $\Omega\setminus \bigcup_{j\in I}2B_j$. Thus \begin{align*} \int_{\Omega}|u-\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_E|\, d\mu &\le \int_{\Omega} \text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{\bigcup_{j\in I}2B_j}\, d\mu\le \sum_{j\in I}\mu(2B_j) \le \varepsilon\sum_{j\in I} \frac{\mu(2B_j)}{r_j} \le \varepsilon (C_d\mathcal{H}(A)+\varepsilon). \end{align*} Moreover, using \eqref{eq:covering for A} we get \[ \int_{\Omega}g\,d\mu\le \sum_{j\in I}\int_{\Omega}\frac{\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{2B_j}}{r_j}\,d\mu +\int_{\Omega}\rho\,d\mu\le C_d\mathcal H(A)+C_d \varepsilon +\varepsilon. \] Now for each $i\in{\mathbb N}$, use the above construction to obtain functions $u_i\in N^{1,1}_{\mathrm{loc}}(\Omega)$ and upper gradients $g_i\in L^1(\Omega)$ corresponding to $\varepsilon=1/i$. We have \[ \int_{\Omega}|u_i-\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_E|\, d\mu\le i^{-1} (C_d\mathcal{H}(A)+i^{-1})\to 0 \quad\textrm{as }i\to \infty \] and thus \[ P(E,\Omega)\le \liminf_{i\to\infty}\int_{\Omega}g_i\,d\mu \le \liminf_{i\to\infty}(C_d\mathcal H(A)+C_d i^{-1}+i^{-1})=C_d\mathcal H(A). \] \end{proof} \begin{proposition}\label{prop:sum of perimeters of components} Let $B(x,R)$ be a ball with $0<R<\frac{1}{4}\diam X$ and let $F\subset X$ be a closed set with $P(F,X)<\infty$. Denote the components of $F\cap \overline{B}(x,R)$ having nonzero $\mu$-measure by $F_1,F_2,\ldots$. Then \[ \sum_{j=1}^{\infty}P(F_j,B(x,R))<\infty, \] and for any sets $A_j\subset F_j$ with $P(A_j,B(x,R))<\infty$ for all $j\in{\mathbb N}$ we have \begin{equation}\label{eq:perimeter of union and sum of Ajs is the same} P\Bigg(\bigcup_{j=1}^{\infty}A_j,B(x,R)\Bigg) =\sum_{j=1}^{\infty}P(A_j,B(x,R)). \end{equation} \end{proposition} \begin{proof} Let $\Gamma_b$ be the exceptional family of curves of Proposition \ref{prop:ae curve goes through boundary}; then $\Mod_1(\Gamma_b)=0$. Consider a component $F_j$; it is a closed set. Consider a curve $\gamma\notin \Gamma_b$ in $B(x,R)$ with $\gamma(0)\in I_{F_j}$ and $\gamma(\ell_{\gamma})\in X\setminus F_j$. Then $\gamma(0)\in I_F$. Take \[ t:=\max\{s\in [0,\ell_{\gamma}]:\,\gamma([0,s])\subset F_j\}. \] Clearly $t<\ell_{\gamma}$. There cannot exist $\delta>0$ such that $\gamma(s)\in F$ for all $s\in (t,t+\delta)$ because this would connect $F_j$ with at least one other component of $F\cap \overline{B}(x,R)$. Thus there are points $s_j\searrow t$ with $\gamma(s_j)\in X\setminus F\subset O_F$. By Proposition \ref{prop:ae curve goes through boundary}, this implies that either $\gamma(t)\in\partial^*F$ or $\gamma(t)\in O_{F}$. In the latter case, there is a point $\widetilde{t}\in (0,t)$ with $\gamma(\widetilde{t})\in\partial^*F$. In both cases, we have found $t$ such that $\gamma(t)\in \partial^*F\cap F_j$. Thus by Lemma \ref{lem:perimeter controlled by boundary}, \[ P(F_j,B(x,R))\le C_d\mathcal H(\partial^*F\cap F_j) \] and so \begin{equation}\label{eq:perimeter sum is finite} \begin{split} \sum_{j=1}^{\infty}P(F_j,B(x,R)) &\le C_d \sum_{j=1}^{\infty} \mathcal H(\partial^*F\cap F_j)\\ &\le C_d \mathcal H(\partial^*F)\\ &\le C_d\alpha^{-1}P(F,X)\quad\textrm{by }\eqref{eq:def of theta}\\ &<\infty, \end{split} \end{equation} as desired. Next note that one inequality in \eqref{eq:perimeter of union and sum of Ajs is the same} follows from \eqref{eq:perimeter of countable union}. To prove the other one, note that the sets $F_j$ are closed and then in fact compact, and so for any $\mu$-measurable sets $A_j\subset F_j$ with $P(A_j,B(x,R))<\infty$ for all $j\in{\mathbb N}$, we have \begin{equation}\label{eq:distance between Aj and Ak} \dist(A_j,A_k)\ge \dist(F_j,F_k)>0 \end{equation} for all $j\neq k$. Take $N,M\in{\mathbb N}$ with $N\le M$. We have (recall \eqref{eq:measure theoretic closure}) \begin{equation}\label{eq:perimeter of union of Ajs} \begin{split} P\Bigg(\bigcup_{j=1}^{\infty}A_j,B(x,R)\Bigg) &\ge P\Bigg(\bigcup_{j=1}^{\infty}A_j,B(x,R)\setminus \overline{\bigcup_{j=M+1}^{\infty}A_j}^m\Bigg)\\ &=P\Bigg(\bigcup_{j=1}^{M}A_j,B(x,R)\setminus \overline{\bigcup_{j=M+1}^{\infty}A_j}^m\Bigg)\quad\textrm{by Lemma }\ref{lem:coincidence of perimeter}\\ &=\sum_{j=1}^{M} P\Bigg(A_j,B(x,R)\setminus\overline{\bigcup_{j=M+1}^{\infty}A_j}^m\Bigg)\quad\textrm{by }\eqref{eq:distance between Aj and Ak}\\ &\ge \sum_{j=1}^{N} P\Bigg(A_j,B(x,R)\setminus\overline{\bigcup_{j=M+1}^{\infty}A_j}^m\Bigg). \end{split} \end{equation} By \eqref{eq:perimeter of countable union} and \eqref{eq:perimeter sum is finite}, we have \[ P\Bigg(\bigcup_{j=M+1}^{\infty}F_j,B(x,R)\Bigg) \le \sum_{j=M+1}^{\infty}P(F_j,B(x,R)) \to 0\quad\textrm{as }M\to \infty. \] Then by Lemma \ref{lem:inner capacity} we have \[ \capa_1\Bigg(\overline{\bigcup_{j=M+1}^{\infty}A_j}^m\cap B(x,r)\Bigg) \le \capa_1\Bigg(\overline{\bigcup_{j=M+1}^{\infty}F_j}^m\cap B(x,r)\Bigg) \to 0\quad\textrm{as }M\to \infty \] for all $0<r<R$. From \eqref{eq:perimeter of union of Ajs} and Lemma \ref{lem:variation measure and capacity} we now get \[ P\Bigg(\bigcup_{j=1}^{\infty}A_j,B(x,R)\Bigg)\ge \sum_{j=1}^{N} P(A_j,B(x,r)). \] Letting $r\nearrow R$ and $N\to\infty$, we get the conclusion. \end{proof} For any nonnegative $g\in L^1_{\mathrm{loc}}(X)$, define the centered Hardy-Littlewood maximal function \[ \mathcal M g(x):=\sup_{r>0}\,\vint{B(x,r)}g\,d\mu,\quad x\in X. \] Recall the definition of the exponent $s>1$ from \eqref{eq:homogenous dimension}. The argument in the following lemma was inspired by the study of the so-called $\textrm{MEC}_p$-property in \cite{JJRRS}. \begin{lemma}\label{lem:finding a positive measure component} Let $B(x_0,r)$ be a ball and let $V\subset X$ be an open set with \[ \capa_1(V\cap B(x_0,r))< \frac{1}{20 \cdot 10^s C_P C_d^7}\frac{\mu(B(x_0,r))}{r}. \] Then there is a connected subset of $\overline{B}(x_0,r/2)\setminus V$ with measure at least $\mu(B(x_0,r))/(4\cdot 10^s C_d^2)$. \end{lemma} \begin{proof} Take $u\in N^{1,1}(X)$ with $u=1$ in $V\cap B(x_0,r)$ and \[ \Vert u\Vert_{N^{1,1}(X)}<\frac{1}{20 \cdot 10^s C_P C_d^7}\frac{\mu(B(x_0,r))}{r}. \] Thus there is an upper gradient $g$ of $u$ with \[ \Vert g\Vert_{L^1(X)}<\frac{1}{20 \cdot 10^s C_P C_d^7}\frac{\mu(B(x_0,r))}{r}. \] By the Vitali-Carath\'eodory theorem (see e.g. \cite[p. 108]{HKST15}) we can assume that $g$ is lower semicontinuous. We define \[ A:=\{\mathcal M g> (10C_P C_d^2 r)^{-1}\}\quad\textrm{and}\quad D:=\{u\ge 1/2\}. \] Then by the weak $L^1$-boundedness of the maximal function (see e.g. \cite[Lemma 3.12]{BB}) as well as \eqref{eq:homogenous dimension}, we estimate \[ \mu(A)\le 10 C_P C_d^5 r\Vert g\Vert_{L^1(X)}\le \frac{1}{2 \cdot 10^s C_d^2}\mu(B(x_0,r)) \le \frac{1}{2}\mu(B(x_0,r/10)). \] Similarly, \[ \mu(D)\le 2\Vert u\Vert_{L^1(X)}\le \frac{1}{4}\mu(B(x_0,r/10)), \] and then \begin{equation}\label{eq:measure of complement of A D} \mu(B(x_0,r/10)\setminus (A\cup D))\ge \frac{1}{4}\mu(B(x_0,r/10)) \ge \frac{\mu(B(x_0,r))}{4\cdot 10^s C_d^2}. \end{equation} In particular, we can fix $x\in B(x_0,r/10)\setminus (A\cup D)$. Let $\delta:=(100C_P C_d^2 r)^{-1}$. For every $k\in{\mathbb N}$, let $g_k:=\min\{g,k\}$ and \[ v_k(y):=\inf\int_{\gamma}(g_k+\delta)\,ds,\quad y\in B(x_0,r/2), \] where the infimum is taken over curves $\gamma$ (also constant curves) in $B(x_0,r/2)$ with $\gamma(0)= x$ and $\gamma(\ell_{\gamma})=y$. Then $g_k+\delta\le g+\delta$ is an upper gradient of $v_k$ in $B(x_0,r/2)$ (see \cite[Lemma 5.25]{BB}) and $v_k$ is $\mu$-measurable by \cite[Theorem 1.11]{JJRRS}. Since the space is geodesic, each $v_k$ is $(k+\delta)$-Lipschitz in $B(x_0,r/10)$ and thus all points in $B(x_0,r/10)$ are Lebesgue points of $v_k$. Define $B_j:=B(x,2^{-j+1}r/10)$, for $j=0,1\ldots$. By the Poincar\'e inequality, \begin{equation}\label{eq:telescope at x} \begin{split} |v_k(x)-(v_k)_{B_0}| \le \sum_{j=0}^{\infty}|(v_k)_{B_{j+1}}-(v_k)_{B_{j}}| &\le C_d\sum_{j=0}^{\infty}\, \vint{B_j}|v_k-(v_k)_{B_j}|\,d\mu\\ &\le C_d C_P \sum_{j=0}^{\infty}\frac{2^{-j+1}r}{10}\vint{B_j}(g+\delta)\,d\mu\\ &\le C_d C_P r(\mathcal M g(x)+ \delta)\\ &\le 1/8. \end{split} \end{equation} Similarly, for every $y\in B(x_0,r/10)\setminus (A\cup D)$ we have \begin{equation}\label{eq:telescope at y} |v_k(y)-(v_k)_{B(y,r/5)}|\le 1/8 \end{equation} and \begin{equation}\label{eq:middle term} \begin{split} |(v_k)_{B(x,r/5)}-(v_k)_{B(y,r/5)}| &\le 2C_d^2\vint{B(x,2r/5)}|v_k-(v_k)_{B(x,2r/5)}|\,d\mu\\ &\le 2C_d^2 C_P r\vint{B(x,2r/5)}(g+\delta)\,d\mu\\ &\le 2C_d^2 C_P r(\mathcal Mg(x)+\delta)\\ &\le 1/4. \end{split} \end{equation} Combining \eqref{eq:telescope at x}, \eqref{eq:telescope at y}, and \eqref{eq:middle term}, we get \[ v_k(y)= |v_k(x)-v_k(y)|\le 1/2. \] This means that there is a curve $\gamma_k$ in $B(x_0,r/2)$ with $\gamma_k(0)= x$, $\gamma_k(\ell_{\gamma_k})=y$, and $\int_{\gamma_k}(g_k+\delta)\,ds\le 1/2$, for every $k\in{\mathbb N}$. Note that \[ \ell_{\gamma_k}\le \frac{1}{\delta}\int_{\gamma_k}(g_k+\delta)\,ds\le \frac{1}{2\delta}. \] Consider the reparametrizations $\widetilde{\gamma}_k(t):=\gamma_k(t\ell_{\gamma_k})$, $t\in [0,1]$. By the Arzela-Ascoli theorem (see e.g. \cite[p. 169]{Roy}), passing to a subsequence (not relabeled) we find $\widetilde{\gamma}\colon [0,1]\to X$ such that $\widetilde{\gamma}_k\to \widetilde{\gamma}$ uniformly. It is straightforward to check that $\widetilde{\gamma}$ is continuous and rectifiable. Let $\gamma$ be the parametrization of $\widetilde{\gamma}$ by arc-length; then $\gamma(0)= x$ and $\gamma(\ell_{\gamma})=y$, and by \cite[Lemma 2.2]{JJRRS}, we have for every $k_0\in{\mathbb N}$ that \[ \int_{\gamma}g_{k_0}\,ds\le \liminf_{k\to\infty}\int_{\gamma_k}g_{k_0}\,ds \le \liminf_{k\to\infty}\int_{\gamma_k}g_{k}\,ds\le 1/2. \] Letting $k_0\to\infty$, we obtain \[ \int_{\gamma}g\,ds\le 1/2. \] Note that if $\gamma$ intersected a point $z\in V$, then we would have \[ \int_{\gamma}g\,ds \ge |u(x)-u(z)|> |1/2-1|=1/2, \] so this is not possible. Thus $\gamma$ is in $\overline{B}(x_0,r/2)\setminus V$; let us denote this curve, and also its image, by $\gamma_y$. Define the desired connected set as the union \[ \bigcup_{y\in B(x_0,r/10)\setminus (A\cup D)}\gamma_y. \] By \eqref{eq:measure of complement of A D} this has measure at least $\mu(B(x_0,r))/(4\cdot 10^s C_d^2)$. \end{proof} \begin{lemma}\label{lem:H has measure zero} Let $B(x,R)$ be a ball with $0<R<\frac{1}{4}\diam X$ and let $F\subset X$ be a closed set with $P(F,X)<\infty$. Denote the components of $F\cap \overline{B}(x,R)$ having nonzero $\mu$-measure by $F_1,F_2,\ldots$, and $H:=\overline{B}(x,R)\cap F\setminus \bigcup_{j=1}^{\infty} F_j$. Then $\mu(H)=0$. \end{lemma} \begin{proof} It follows from Proposition \ref{prop:sum of perimeters of components} that $P\left(\bigcup_{j=1}^{\infty} F_j,B(x,R)\right)<\infty$, and then by \eqref{eq:lattice property of sets of finite perimeter} also $P(H,B(x,R))<\infty$. By \eqref{eq:def of theta} and a standard covering argument (see e.g. the proof of \cite[Lemma 2.6]{KKST3}), we find that \[ \lim_{r\to 0}r\frac{P\left(\bigcup_{j=1}^{\infty} F_j,B(y,r)\right)}{\mu(B(y,r))}=0 \] for all $y\in B(x,R)\setminus \left(\partial^*\big(\bigcup_{j=1}^{\infty} F_j\big)\cup N\right)$, with $\mathcal H(N)=0$, in particular for all $y\in B(x,R)\cap I_H\setminus N$. Take $y\in B(x,R)\cap I_H\setminus N$ (if it exists). We find arbitrarily small $r>0$ such that $B(y,r) \subset B(x,R)$ and \begin{equation}\label{eq:complement of H small} \frac{\mu(B(y,r)\setminus H)}{\mu(B(y,r))}\le \frac{1}{80 \cdot 10^s C_P C_d^8 C_{\textrm{cap}}} \end{equation} and \[ r\frac{P\left(\bigcup_{j=1}^{\infty} F_j,B(y,r)\right)}{\mu(B(y,r))} \le \frac{1}{80 \cdot 10^s C_P C_d^8 C_{\textrm{Cap}}}. \] Now suppose that \[ P(H,B(y,r))\le \frac{1}{80 \cdot 10^s C_P C_d^8 C_{\textrm{cap}}}\frac{\mu(B(y,r))}{r}. \] Then since $H\cup\bigcup_{j=1}^{\infty}F_j=F\cap \overline{B}(x,R)$, by \eqref{eq:lattice property of sets of finite perimeter} we get \begin{align*} P(F,B(y,r)) &\le P(H,B(y,r))+P\Bigg(\bigcup_{j=1}^{\infty}F_j,B(y,r)\Bigg)\\ &\le \frac{1}{40 \cdot 10^s C_P C_d^8 C_{\textrm{cap}}}\frac{\mu(B(y,r))}{r}. \end{align*} Define the Lipschitz function \[ \eta:=\max\left\{0,1-\frac{\dist(\cdot,B(y,r/2))}{r/2}\right\}, \] so that $0\le \eta\le 1$ on $X$, $\eta=1$ in $B(y,r/2)$, $\eta=0$ in $X\setminus B(y,r)$, and $g_{\eta}\le (2/r)\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{B(y,r)}$ (see \cite[Corollary 2.21]{BB}). Then by a Leibniz rule (see \cite[Proposition 4.2]{KKST3}), we have \begin{align*} \capa_{\mathrm{BV}}(B(y,r/2)\setminus F) &\le \Vert D(\eta\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{X\setminus F})\Vert(X)\\ &\le P(F,B(y,r))+2\frac{\mu(B(y,r)\setminus F)}{r}\\ &\le P(F,B(y,r))+2\frac{\mu(B(y,r)\setminus H)}{r}\\ &\le \frac{1}{20 \cdot 10^s C_P C_d^8 C_{\textrm{cap}}}\frac{\mu(B(y,r))}{r}. \end{align*} Then by \eqref{eq:Newtonian and BV capacities are comparable}, \[ \capa_{1}(B(y,r/2)\setminus F)\le \frac{1}{20 \cdot 10^s C_P C_d^8}\frac{\mu(B(y,r))}{r} <\frac{1}{20 \cdot 10^s C_P C_d^7}\frac{\mu(B(y,r/2))}{r/2}. \] Then by Lemma \ref{lem:finding a positive measure component}, there is a connected subset of $F\cap \overline{B}(y,r/4)$ with measure at least \[ \frac{\mu(B(y,r/2))}{4\cdot 10^s C_d^2}\ge \frac{\mu(B(y,r))}{4\cdot 10^s C_d^3}. \] By \eqref{eq:complement of H small} this must be (partially) contained in $H$, a contradiction since $H$ contains no components of nonzero measure. Thus for all $y\in I_H\cap B(x,R)\setminus N$, we have \[ \limsup_{r\to 0}r\frac{P(H,B(y,r))}{\mu(B(y,r))} \ge \frac{1}{80 \cdot 10^s C_P C_d^8 C_{\textrm{cap}}}. \] By a simple covering argument, it follows that \[ \mu(I_H\cap B(x,R)\setminus N)\le \varepsilon\cdot 80 \cdot 10^s C_P C_d^{11} C_{\textrm{cap}} P(H,B(x,R)) \] for every $\varepsilon>0$. Thus $\mu(H\cap B(x,R)\setminus N)=0$ and so $\mu(H\cap B(x,R))=0$. Since the space $X$ is geodesic, by \cite[Corollary 2.2]{Buc} we know that $\mu(\{y\in X:\,d(y,x)=R\})=0$ and so in fact $\mu(H)=0$. \end{proof} \begin{proof}[Proof of Proposition \ref{prop:connected components}] This follows from Proposition \ref{prop:sum of perimeters of components} and Lemma \ref{lem:H has measure zero}. \end{proof} \section{Functions of least gradient}\label{sec:least gradient} In this section we consider functions of least gradient, or more precisely superminimizers and solutions of obstacle problems in the case $p=1$. We will follow the definitions and theory developed in \cite{L-WC}. Throughout this section the symbol $\Omega$ will always denote a nonempty open subset of $X$. We denote by $\mathrm{BV}_c(\Omega)$ the class of functions $\varphi\in\mathrm{BV}(\Omega)$ with compact support in $\Omega$, that is, $\supp \varphi\Subset \Omega$. \begin{definition} We say that $u\in\mathrm{BV}_{\mathrm{loc}}(\Omega)$ is a $1$-minimizer in $\Omega$ (often called function of least gradient) if for all $\varphi\in \mathrm{BV}_c(\Omega)$, we have \begin{equation}\label{eq:definition of 1minimizer} \Vert Du\Vert(\supp\varphi)\le \Vert D(u+\varphi)\Vert(\supp\varphi). \end{equation} We say that $u\in\mathrm{BV}_{\mathrm{loc}}(\Omega)$ is a $1$-superminimizer in $\Omega$ if \eqref{eq:definition of 1minimizer} holds for all nonnegative $\varphi\in \mathrm{BV}_c(\Omega)$. We say that $u\in\mathrm{BV}_{\mathrm{loc}}(\Omega)$ is a $1$-subminimizer in $\Omega$ if \eqref{eq:definition of 1minimizer} holds for all nonpositive $\varphi\in \mathrm{BV}_c(\Omega)$, or equivalently if $-u$ is a $1$-superminimizer in $\Omega$. \end{definition} Equivalently, we can replace $\supp\varphi$ by any set $A\Subset \Omega$ containing $\supp\varphi$ in the above definitions. If $\Omega$ is bounded, and $\psi\colon\Omega\to\overline{{\mathbb R}}$ and $f\in L^1_{\mathrm{loc}}(X)$ with $\Vert Df\Vert(X)<\infty$, we define the class of admissible functions \[ \mathcal K_{\psi,f}(\Omega):=\{u\in\mathrm{BV}_{\mathrm{loc}}(X):\,u\ge \psi\textrm{ in }\Omega\textrm{ and }u=f\textrm{ in }X\setminus\Omega\}. \] The (in)equalities above are understood in the a.e. sense. For brevity, we sometimes write $\mathcal K_{\psi,f}$ instead of $\mathcal K_{\psi,f}(\Omega)$. By using a cutoff function, it is easy to show that $\Vert Du\Vert(X)<\infty$ for every $u\in\mathcal K_{\psi,f}(\Omega)$. \begin{definition} We say that $u\in\mathcal K_{\psi,f}(\Omega)$ is a solution of the $\mathcal K_{\psi,f}$-obstacle problem if $\Vert Du\Vert(X)\le \Vert Dv\Vert(X)$ for all $v\in\mathcal K_{\psi,f}(\Omega)$. \end{definition} Whenever the characteristic function of a set $E$ is a solution of an obstacle problem, for simplicity we will call $E$ a solution as well. Similarly, if $\psi=\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_A$ for some $A\subset X$, we let $\mathcal K_{A, f}:=\mathcal K_{\psi, f}$. Now we list some properties of superminimizers and solutions of obstacle problems derived mostly in \cite{L-WC}. \begin{lemma}[{\cite[Lemma 3.6]{L-WC}}]\label{lem:solutions from capacity} If $x\in X$, $0<r<R<\frac 18 \diam X$, and $A\subset B(x,r)$, then there exists $E\subset X$ that is a solution of the $\mathcal K_{A,0}(B(x,R))$-obstacle problem with \[ P(E,X)\le \rcapa_1(A,B(x,R)). \] \end{lemma} \begin{proposition}[{\cite[Proposition 3.7]{L-WC}}]\label{prop:solutions are superminimizers} If $u\in\mathcal K_{\psi,f}(\Omega)$ is a solution of the $\mathcal K_{\psi,f}$-obstacle problem, then $u$ is a $1$-superminimizer in $\Omega$. \end{proposition} The following fact and its proof are similar to \cite[Lemma 3.2]{KKLS}. \begin{lemma}\label{lem:subminimizer char} Let $F\subset X$ with $P(F,\Omega)<\infty$ and suppose that for every $H\Subset \Omega$, we have \[ P(F,\Omega)\le P(F\setminus H,\Omega). \] Then $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F$ is a $1$-subminimizer in $\Omega$. \end{lemma} \begin{proof} Take a nonnegative $\varphi\in\mathrm{BV}_c(\Omega)$. Observe that for every $0<s<1$, we have $\supp\{\varphi\ge s\}\Subset \Omega$. Thus by the coarea formula \eqref{eq:coarea}, \begin{align*} \Vert D(\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F-\varphi)\Vert(\supp\varphi) &\ge\int_0^1 P(\{\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F-\varphi>t\},\supp\varphi)\,dt\\ &=\int_0^1 P(F\setminus \{\varphi\ge 1-t\},\supp\varphi)\,dt\\ &\ge\int_0^1 P(F,\supp\varphi)\,dt=\Vert D\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F\Vert(\supp\varphi). \end{align*} \end{proof} \begin{proposition}\label{prop:components are subminimizers} Let $B(x,R)$ be a ball and let $F\subset X$ be a closed set with $P(F,X)<\infty$ and such that $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F$ is a $1$-subminimizer in $B(x,R)$. Denote the components of $F\cap \overline{B}(x,R)$ with nonzero $\mu$-measure by $F_1,F_2,\ldots$. Then each $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{F_k}$ is a $1$-subminimizer in $B(x,R)$. \end{proposition} \begin{proof} Fix $k\in{\mathbb N}$ and take $H\Subset B(x,R)$. We can assume that $H\subset F_k$ and that $P(F_k\setminus H,B(x,R))<\infty$. Now \begin{align*} \sum_{\substack{j\in{\mathbb N}\\ j\neq k}}P(F_j,B(x,R))+P(F_k,B(x,R)) &=\sum_{j=1}^{\infty}P(F_j,B(x,R))\\ &= P(F,B(x,R))\quad\textrm{by Proposition }\ref{prop:connected components}\\ &\le P(F\setminus H,B(x,R))\\ &= \sum_{j=1}^{\infty}P(F_j\setminus H,B(x,R))\quad\textrm{by Proposition }\ref{prop:connected components}\\ &=\sum_{\substack{j\in{\mathbb N}\\ j\neq k}}P(F_j,B(x,R))+P(F_k\setminus H,B(x,R)). \end{align*} Note that since $\sum_{j=1}^{\infty}P(F_j,B(x,R))= P(F,B(x,R))<\infty$, we now get \[ P(F_k,B(x,R))\le P(F_k\setminus H,B(x,R)). \] By Lemma \ref{lem:subminimizer char}, $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{F_k}$ is a $1$-subminimizer in $B(x,R)$. \end{proof} We have the following weak Harnack inequality. We denote the positive part of a function by $u_+:=\max\{u,0\}$. \begin{theorem}[{\cite[Theorem 3.10]{L-WC}}]\label{thm:weak Harnack} Suppose $k\in{\mathbb R}$ and $0<R<\tfrac 14 \diam X$ with $B(x,R)\Subset \Omega$, and assume either that \begin{enumerate}[{(a)}] \item $u$ is a $1$-subminimizer in $\Omega$, or \item $\Omega$ is bounded, $u$ is a solution of the $\mathcal K_{\psi, f}(\Omega)$-obstacle problem, and $\psi\le k$ a.e. in $B(x,R)$. \end{enumerate} Then for any $0<r<R$ and some constant $C_1=C_1(C_d,C_P)$, \[ \esssup_{B(x,r)}u\le C_1\left(\frac{R}{R-r}\right)^{s}\vint{B(x,R)}(u-k)_+\,d\mu+k. \] \end{theorem} For later reference, let us note that a close look at the proof of the above theorem reveals that we can take \begin{equation}\label{eq:C1} C_1=2^{(s+1)^2}(6\widetilde{C}_S C_d)^s, \end{equation} where $\widetilde{C}$ is the constant from an $(s/(s-1),1)$-Sobolev inequality with zero boundary values. \begin{corollary}\label{cor:weak Harnack} Suppose $k\in{\mathbb R}$, $x\in X$, $0<R<\tfrac 14 \diam X$, and assume that $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F$ is a $1$-subminimizer in $B(x,R)$ with $\mu(F\cap B(x,R/2))>0$. Then \[ \frac{\mu(B(x,R)\cap F)}{\mu(B(x,R))}\ge (2^s C_1)^{-1}. \] \end{corollary} \begin{proof} Let $0<\varepsilon<R/2$. Applying Theorem \ref{thm:weak Harnack}(i) with $\Omega=B(x,R)$, $u=\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F$, $k=0$, and $R/2,\, R-\varepsilon$ in place of $r,\, R$, we get \[ 1\le C_1\left(\frac{R-\varepsilon}{R-\varepsilon-R/2}\right)^{s}\frac{\mu(B(x,R-\varepsilon)\cap F)}{\mu(B(x,R-\varepsilon))}. \] Letting $\varepsilon\to 0$, we get the result. \end{proof} Recall the definitions of the lower and upper approximate limits $u^{\wedge}$ and $u^{\vee}$ from \eqref{eq:lower approximate limit} and \eqref{eq:upper approximate limit}. \begin{theorem}[{\cite[Theorem 3.11]{L-WC}}]\label{thm:superminimizers are lsc} Let $u$ be a $1$-superminimizer in $\Omega$. Then $u^{\wedge}\colon\Omega\to (-\infty,\infty]$ is lower semicontinuous. \end{theorem} \begin{lemma}\label{lem:smallness in annuli} Let $B=B(x,R)$ be a ball with $0<R<\frac{1}{32} \diam X$, and suppose that $W\subset B$. Let $V\subset 4B$ be a solution of the $\mathcal K_{W,0}(4B)$-obstacle problem (as guaranteed by Lemma \ref{lem:solutions from capacity}). Then for all $y\in 3 B\setminus 2 B$, \[ \text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V^{\vee}(y)\le C_2 R \frac{\rcapa_1(W,4B)}{\mu(B)} \] for some constant $C_2=C_2(C_d,C_P)$. \end{lemma} \begin{proof} By Lemma \ref{lem:solutions from capacity} we know that \[ P(V,X)\le \rcapa_1(W, 4B), \] and thus by the isoperimetric inequality \eqref{eq:isop inequality with zero boundary values}, \begin{equation}\label{eq:E1 has small measure} \mu(V)\le 4C_S R P(V,X)\le 4C_S R \rcapa_1(W,4B). \end{equation} For any $z\in 3 B\setminus 2 B$ we have $B(z,R)\subset 4 B\setminus B$. Since now $W\cap B(z,R)=\emptyset$, we can apply Theorem \ref{thm:weak Harnack}(b) with $k=0$ to get \begin{align*} \sup_{B(z,R/2)} \text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V^{\vee } &\le \esssup_{B(z,R/2)}\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V\\ &\le C_1\left(\frac{R}{R-R/2}\right)^s\vint{B(z,R)}(\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V)_+\,d\mu\\ &= \frac{2^s C_1}{\mu(B(z,R))}\int_{B(z,R)} (\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V)_+\,d\mu\\ &\le \frac{2^s C_1 C_d^2}{\mu(B)}\mu(V)\\ &\le 2^{s+2} C_1 C_d^2 C_S R \frac{\rcapa_1(W,4B)}{\mu(B)}\quad\textrm{by }\eqref{eq:E1 has small measure}. \end{align*} Thus we can choose $C_2=2^{s+2} C_1 C_d^2 C_S$. \end{proof} \section{Constructing a ``geodesic'' space}\label{sec:constructing a quasiconvex space} In this section we construct a suitable space where the Mazurkiewicz metric agrees with the ordinary one; this space will be needed in the proof of the main result. Recall that in Section \ref{sec:strong boundary points}, in the space $(Z,\widehat{d},\mu)$ we defined the Mazurkiewicz metric $\widehat{d}_M$; given a set $V\subset X$ we now define \[ d_{M}^V(x,y):=\inf\{\diam K: K\subset X\setminus V\textrm{ is a continuum containing }x,y\}, \quad x,y\in X\setminus V. \] If $V=\emptyset$, we leave it out of the notation, consistent with \eqref{eq:widehat d c}. \begin{lemma}\label{lem:new metric lemma} Let $V\subset X$ be a bounded open set and let $B(x_0,R_0)$ be a ball such that $V\Subset B(x_0,R_0)$, and $\overline{B}(x_0,R_0)\setminus V$ is connected. Moreover, suppose there is $R>0$ such that for every $x\in X\setminus V$ and $0<r\le R$, the connected components of $\overline{B}(x,r)\setminus V$ intersecting $B(x,r/2)$ are finite in number. Then $d_M^V$ is a metric on $X\setminus V$ such that $d\le d_M^V$, $d_M^V$ induces the same topology on $X\setminus V$ as $d$, $(d_M^V)_M=d_M^V$, and $(X\setminus V,d_M^V)$ is complete. \end{lemma} Note that explicitly, for $x,y\in X\setminus V$, \[ (d_{M}^V)_M(x,y)= \inf\{\diam_{d_M^V} K:\, K\subset X\setminus V \textrm{ is a }d_M^V\textrm{-continuum containing }x,y\}. \] \begin{proof} Since $V\Subset B(x_0,R_0)$ and $\overline{B}(x_0,R_0)\setminus V$ is connected, also every $\overline{B}(x_0,r)\setminus V$ with $r\ge R_0$ is connected, by the fact that $X$ is geodesic. Thus we have for all $x,y\in X\setminus V$ \[ d_M^V(x,y)\le 2\max\{R_0,d(x,x_0),d(y,x_0)\}<\infty. \] Obviously $d\le d_M^V$ and $d_M^V(x,x)=0$ for all $x\in X\setminus V$. If $d_M^V(x,y)=0$ then $d(x,y)=0$ and so $x=y$. Obviously also $d_M^V(x,y)=d_M^V(y,x)$ for all $x,y\in X\setminus V$. Finally, take $x,y,z\in X\setminus V$. Take a continuum $K_1\subset X\setminus V$ containing $x,y$ and a continuum $K_2\subset X\setminus V$ containing $y,z$. Then $K_1\cup K_2\subset X\setminus V$ is a continuum containing $x,z$ and so \[ d_M^V(x,z)\le \diam(K_1\cup K_2)\le \diam(K_1)+\diam (K_2). \] Taking infimum over $K_1$ and $K_2$, we conclude that the triangle inequality holds. Hence $d_M^V$ is a metric on $X\setminus V$. To show that the topologies induced on $X\setminus V$ by $d$ and $d_M^V$ are the same, take a sequence $x_j\to x$ with respect to $d$ in $X\setminus V$. Fix $\varepsilon\in (0,R)$. Consider the components of $\overline{B}(x,\varepsilon/2)\setminus V$ intersecting $B(x,\varepsilon/4)$. By assumption there are only finitely many. Each of them not containing $x$ is at a nonzero distance from $x$ and so for large $j$, every $x_j$ belongs to the component containing $x$; denote it $F_1$. For such $j$, we have \[ d_M^V(x_j,x)\le \diam F_1\le \varepsilon. \] We conclude that $x_j\to x$ also with respect to $d_M^V$. Since we had $d\le d_M^V$, it follows that the topologies are the same. If $x,y\in X\setminus V$, and $\varepsilon>0$, we can take a continuum $K$ containing $x$ and $y$, with $\diam K<d_M^V(x,y)+\varepsilon$. The set $K$ is still a continuum in the metric space $(X\setminus V,d_M^V)$, and for every $z,w\in K$, \[ d_M^V(z,w)\le \diam K< d_M^V(x,y)+\varepsilon. \] It follows that $\diam_{d_M^V} K\le d_M^V(x,y)+\varepsilon$, and so $(d_M^V)_M(x,y)\le d_M^V(x,y)+\varepsilon$, showing that $(d_M^V)_M=d_M^V$. Finally let $(x_j)$ be a Cauchy sequence in $(X\setminus V, d_M^V)$. Since $d\le d_M^V$, it is also a Cauchy sequence in $(X,d)$, and so $x_j\to x\in X\setminus V$ with respect to $d$. But as we showed before, this implies that $x_j\to x$ with respect to $d_M^V$. \end{proof} Let $B$ be a ball and let $B_1,B_2\subset B$ be two other balls, and let $u\in L^1(B)$ such that $u=1$ in $B_1$ and $u=0$ in $B_2$. Then we have \begin{equation}\label{eq:KoLa result} \int_{B}|u-u_{B}|\,d\mu\ge \frac{1}{2}\min\{\mu(B_1),\mu(B_2)\}; \end{equation} this follows easily by considering the cases $u_{B}\le 1/2$ and $u_{B}\ge 1/2$. We have the following \emph{linear local connectedness}; versions of this property have been proved before e.g. in \cite{HK}, but they assume certain growth bounds on the measure, which we do not want to assume. \begin{lemma}\label{lem:lin loc connectedness} Let $B(x_0,R)$ be a ball and let $V\subset B(x_0,2R)$ with \begin{equation}\label{eq:V capacity in 4B} \rcapa_1(V,B(x_0,3R))<\frac{1}{12C_P C_d^3}\frac{\mu(B(x_0,R))}{R}. \end{equation} Then every pair of points $y,z\in B(x_0,5R)\setminus B(x_0,4R)$ can be joined by a curve in $B(x_0,6R)\setminus V$. \end{lemma} \begin{proof} If $d(y,z)\le 2R$, then the result is clear since the space is geodesic. Thus assume that $d(y,z)> 2R$. Consider the disjoint balls $B_1:=B(y,R)$ and $B_2:=B(z,R)$ which both belong to $B(x_0,6R)\setminus B(x_0,3R)$. Denote by $\Gamma$ the family of curves $\gamma$ in $B(x_0,6R)$ with $\gamma(0)\in B_1$ and $\gamma(\ell_{\gamma})\in B_2$. Note that $\Mod_1(\Gamma)<\infty$ since $\dist(B_1,B_2)>0$. Let $\varepsilon>0$. Let $g\in L^1(B(x_0,6R))$ such that $\int_{\gamma}g\,ds\ge 1$ for all $\gamma\in\Gamma$ and \[ \int_{B(x_0,6R)}g\,d\mu<\Mod_1(\Gamma)+\varepsilon. \] Let \[ u(x):=\min\left\{1,\inf \int_{\gamma}g\,ds\right\},\quad x\in B(x_0,6R), \] where the infimum is taken over curves $\gamma$ (also constant curves) in $B(x_0,6R)$ with $\gamma(0)=x$ and $\gamma(\ell_{\gamma})\in B_1$. Then $u=1$ in $B_2$. Moreover, $g$ is an upper gradient of $u$ in $B(x_0,6R)$, see \cite[Lemma 5.25]{BB}, and $u$ is $\mu$-measurable by \cite[Theorem 1.11]{JJRRS}. In total, $u\in N^{1,1}(B(x_0,6R))$ with $u=0$ in $B_1$ and $u=1$ in $B_2$. Thus using the Poincar\'e inequality, \begin{align*} \Mod_1(\Gamma) &>\int_{B(x_0,6R)}g\,d\mu-\varepsilon\\ &\ge \frac{1}{6C_P R}\int_{B(x_0,6R)}|u-u_{B(x_0,6R)}|\,d\mu-\varepsilon\\ &\ge \frac{1}{12C_P R}\min\{\mu(B_1),\mu(B_2)\}-\varepsilon\quad\textrm{by }\eqref{eq:KoLa result}\\ &\ge \frac{1}{12C_P C_d^3 R}\mu(B(x_0,R))-\varepsilon \end{align*} and so \[ \Mod_1(\Gamma)\ge \frac{1}{12C_P C_d^3 R}\mu(B(x_0,R)). \] On the other hand, by \eqref{eq:V capacity in 4B} we find a function $v\in N^{1,1}(X)$ such that $v=1$ in $V$, $v=0$ in $X\setminus B(x_0,3R)$, and $v$ has an upper gradient $\widetilde{g}$ satisfying \[ \int_X \widetilde{g}\,d\mu< \frac{1}{12C_P C_d^3}\frac{\mu(B(x_0,R))}{R}. \] Denote the family of all curves intersecting $V$ by $\Gamma_V$. Now $\int_{\gamma}\widetilde{g}\,ds\ge 1$ for all $\gamma\in \Gamma\cap \Gamma_V$, and so \[ \Mod_1(\Gamma\cap \Gamma_V)\le \int_X \widetilde{g}\,d\mu < \frac{1}{12C_P C_d^3}\frac{\mu(B(x_0,R))}{R}. \] Thus $\Gamma\setminus \Gamma_V$ is nonempty. Take a curve $\gamma\in \Gamma\setminus \Gamma_V$. Now we get the required curve by concatenating three curves: the first going from $y$ to $\gamma(0)$ inside $B(y,R)$ (using the fact that the space is geodesic), the second $\gamma$, and the third going from $\gamma(\ell_{\gamma})$ to $z$ inside $B(z,R)$. \end{proof} By using an argument involving Lipschitz cutoff functions, it is easy to see that for any ball $B(x,r)$ and any set $A\subset B(x,r)$, we have \begin{equation}\label{eq:capacity and Hausdorff measure} \rcapa_1(A,B(x,3r))\le C_d \mathcal H(A). \end{equation} In the following proposition we construct the space in which the metric and Mazurkiewicz metric agree. \begin{proposition}\label{prop:constructing the quasiconvex space} Let $B=B(x,R)$ be a ball with $0<R<\frac{1}{32} \diam X$, and let $A\subset B$ with \[ \mathcal H(A) \le \frac{1}{24 C_P C_S C_2 C_r C_d^4} \frac{\mu(B)}{R}. \] Let $\varepsilon>0$. Then we find an open set $V$ with $A\subset V\subset 2B$ and \[ P(V,X)\le C_d\mathcal H(A)+\varepsilon, \] and such that the following hold: the space $(Z,d_M^V,\mu)$ with $Z=X\setminus V$ is a complete metric space with $(d_M^V)_M=d_M^V$, $\mu$ in $Z$ is a Borel regular outer measure and doubling with constant $2^s C_1 C_d^2$, and for every $y\in X\setminus V$ and $r>0$ we have \[ \frac{\mu(B_Z(y,r))}{\mu(B(y,r))}\ge (2^s C_1 C_d)^{-1} \] where $B_Z(y,r)$ denotes an open ball in $Z$, defined with respect to the metric $d_M^V$. \end{proposition} \begin{proof} Using the fact that $\rcapa_1$ is an outer capacity in the sense of \eqref{eq:rcapa outer capacity}, as well as \eqref{eq:capacity and Hausdorff measure}, we find an open set $W$, with $A\subset W\subset B$, such that (note that the first inequality is obvious) \[ \rcapa_1(W,4B)\le \rcapa_1(W,3B)\le \rcapa_1(A,3B)+\varepsilon\le C_d\mathcal H(A)+\varepsilon. \] We can assume that \[ \varepsilon<\frac{1}{24 C_P C_S C_2 C_r C_d^3}\frac{\mu(B)}{R}. \] Take a solution $V$ of the $\mathcal K_{W,0}(4B)$-obstacle problem. By Lemma \ref{lem:solutions from capacity}, we have \[ P(V,X)\le \rcapa_1(W,4B)\le C_d\mathcal H(A)+\varepsilon. \] By Theorem \ref{thm:superminimizers are lsc}, the function $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V^{\wedge}$ is lower semicontinuous, and by redefining $V$ in a set of measure zero, we get $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V=\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V^{\wedge}$ and so $V$ is open. By Lemma \ref{lem:smallness in annuli} we know that for all $y\in 3 B\setminus 2B$, \[ \text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V^{\vee}(y)\le C_2 R \frac{\rcapa_1(W,4B)}{\mu(B)} \le C_2 R \frac{C_d \mathcal H(A)+\varepsilon}{\mu(B)}<1 \] and so $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V^{\vee}=0$ in $3 B\setminus 2B$. Then in fact $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V=\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V^{\vee}=0$ in $4B\setminus 2B$, that is, $V\subset 2B$, because else we could remove the parts of $V$ inside $4B\setminus 3B$ to decrease $P(V,X)$. By the isoperimetric inequality \eqref{eq:isop inequality with zero boundary values}, \begin{equation}\label{eq:measure of V} \mu(V)\le 2C_S R P(V,X)\le 2 C_S C_d R \mathcal H(A) +2C_S R \varepsilon \le \frac{\mu(B)}{2C_d^2}. \end{equation} Moreover, by \eqref{eq:variational one and BV capacity} we get \begin{align*} \rcapa_1(V,3B) &\le C_r \rcapa_{\mathrm{BV}}^{\vee}(V,3B)\\ &\le C_r P(V,X)\le C_r C_d \mathcal H(A)+C_r\varepsilon < \frac{1}{12C_P C_d^3}\frac{\mu(B)}{R}. \end{align*} By Lemma \ref{lem:lin loc connectedness}, $5B\setminus 4B$ belongs to one component of $6\overline{B}\setminus V$. Since the space is geodesic, in fact $6\overline{B}\setminus 4B$ belongs to one component of $6\overline{B}\setminus V$. Call this component $F_1$. Moreover, denote $F:=X\setminus V$; $F$ is a closed set with $P(F,X)=P(V,X)<\infty$. Consider all components of $F\cap 6\overline{B}$. Suppose there is another component $F_2$ with nonzero $\mu$-measure. Denote by $F_1,F_2,\ldots$ all the components with nonzero $\mu$-measure (as usual, some of these may be empty). By the relative isoperimetric inequality \eqref{eq:relative isoperimetric inequality}, we have \begin{equation}\label{eq:F2 has perimeter} P(F_2,6B)>0. \end{equation} Now the set $\widetilde{V}:=V\cup \bigcup_{j=2}^{\infty}F_j\subset 4B$ is admissible for the $\mathcal K_{W,0}(4B)$-obstacle problem, with \begin{align*} P(\widetilde{V},X) &=P(\widetilde{V},6B)\\ &=P\Bigg(X\setminus \Big(V\cup \bigcup_{j=2}^{\infty}F_j\Big),6B\Bigg)\\ &=P\Bigg(F\setminus \bigcup_{j=2}^{\infty}F_j,6B\Bigg)\\ &= P(F,6B)-\sum_{j=2}^{\infty}P(F_j,6B)\quad\textrm{by Proposition }\ref{prop:connected components}\\ &<P(F,6B)\quad\textrm{by }\eqref{eq:F2 has perimeter}\\ &=P(V,6B)=P(V,X). \end{align*} This is a contradiction with the fact that $V$ is a solution of the $\mathcal K_{W,0}(4B)$-obstacle problem. Thus by Proposition \ref{prop:connected components}, $F\cap 6\overline{B}$ is the union of $F_1$ and a set of measure zero $N$. Suppose \[ y\in 6\overline{B}\cap F\setminus F_1 =4B\cap F\setminus F_1. \] Now $y$ is at a nonzero distance from $F_1$. Thus for small $\delta>0$, \[ \mu(B(y,\delta)\cap F)\le \mu(N)=0. \] Note that since we had $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V=\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_V^{\wedge}$, it follows that $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F=\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F^{\vee}$. Thus in fact such $y$ cannot exist and $F\cap 6\overline{B}=F_1$ is connected. If $y\in F\setminus B(x,3R)$ and $0<r\le R$, then $\overline{B}(y,r)\cap F=\overline{B}(y,r)$ is connected since the space is geodesic. If $y\in F\cap B(x,3R)$ and $0<r\le R$, by Proposition \ref{prop:connected components} we know that $F\cap \overline{B}(y,r)$ consists of at most countably many components $F_1,F_2,\ldots$ and a set of measure zero $\widetilde{N}$. By Proposition \ref{prop:solutions are superminimizers} we know that $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F$ is a $1$-subminimizer in $B(x,4R)$, and then also in $B(y,r)\subset B(x,4R)$. Then each $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_{F_j}$ is a $1$-subminimizer in $B(y,r)$ by Proposition \ref{prop:components are subminimizers}. By Corollary \ref{cor:weak Harnack} we get for each $F_j$ with $\mu(B(y,r/2)\cap F_j)>0$ that \begin{equation}\label{eq:subminimizer component measure lower bound} \frac{\mu(F_j\cap B(y,r))}{\mu(B(y,r))}\ge (2^s C_1)^{-1}. \end{equation} Thus there are less than $2^s C_1+1$ such components, which we can relabel $F_1,\ldots,F_M$. Suppose \[ z\in B(y,r/2)\cap \widetilde{N}\setminus \bigcup_{j=1}^M F_j. \] This is at nonzero distance from all $F_1,\ldots,F_M$. Thus for small $\delta>0$, \[ \mu(B(z,\delta)\cap F)\le \mu(\widetilde{N})+\sum_{j=M+1}^{\infty}\mu(F_j\cap B(y,r/2))=0. \] As before, we have $\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F=\text{\raise 1.3pt \hbox{$\chi$}\kern-0.2pt}_F^{\vee}$. Thus in fact such $z$ cannot exist and \[ F\cap B(y,r/2)=B(y,r/2)\cap \bigcup_{j=1}^M F_j. \] Now Lemma \ref{lem:new metric lemma} gives that $(Z,d_M^V,\mu)$, with $Z=X\setminus V$, is a complete metric space, $d\le d_M^V$, the topologies induced by $d$ and $d_M^V$ are the same, and $(d_M^V)_M=d_M^V$. Note that $\mu$ restricted to the subsets of $X\setminus V$ is still a Borel regular outer measure, see \cite[Lemma 3.3.11]{HKST15}. Since the topologies induced by $d$ and $d_M^V$ are the same, $\mu$ remains a Borel regular outer measure in $Z$. (Note that as sets, we have $X\setminus V=F=Z$.) Denoting by $F_1$ the component of $F\cap \overline{B}(y,r)$ containing $y$, by \eqref{eq:subminimizer component measure lower bound} we have for $y\in F\cap B(x,3R)$ and $0<r\le R$ that \begin{equation}\label{eq:size of F1} \frac{\mu(B(y,r)\cap F_1)}{\mu(B(y,r))}\ge (2^s C_1)^{-1}. \end{equation} Recall that if $y\in F\setminus B(x,3R)$, then $F_1=\overline{B}(y,r)$ and so \eqref{eq:size of F1} holds. Eq. \eqref{eq:size of F1} is easily seen to hold also for all $x\in F$ and $r>R$ by \eqref{eq:measure of V}. It follows that for all $y\in F$ and $r>0$, we have \[ \frac{\mu(B_Z(y,2r))}{\mu(B(y,r))}\ge (2^s C_1)^{-1} \] and so in fact \[ \frac{\mu(B_Z(y,r))}{\mu(B(y,r))}\ge (2^s C_1 C_d)^{-1}\quad\textrm{for all }y\in Z \textrm{ and }r>0, \] as desired. Thus \[ \frac{\mu(B_Z(y,2r))}{\mu(B_Z(y,r))}\le 2^s C_1 C_d\frac{\mu(B(y,2r))}{\mu(B(y,r))} \le 2^s C_1 C_d^2. \] Thus in the space $(Z,d_M^V,\mu)$, the measure $\mu$ is doubling with constant $2^s C_1 C_d^2$. \end{proof} \section{Proof of the main result}\label{sec:proof of the main result} In this section we prove the main result of the paper, Theorem \ref{thm:main theorem}. First note that with the choice $\widehat{C}_d=2^s C_1 C_d^2$, the constant appearing in Corollary \ref{cor:density points} becomes \[ \frac{1}{4 \widehat{C}_d^{12}} =\frac{1}{4 (2^s C_1 C_d^2)^{12}}=:\beta_0. \] Recall from \eqref{eq:C1} that we can take $C_1=2^{(s+1)^2}(6\widetilde{C}_S C_d)^s$. Define \begin{equation}\label{eq:definition of beta} \begin{split} \beta:=\frac{\beta_0}{2^s C_1 C_d}=\frac{1}{2^{2+s}C_1 C_d (2^s C_1 C_d^2)^{12}} &=\frac{1}{2^{13s +2} (2^{(s+1)^2}(6\widetilde{C}_S C_d)^s)^{13} C_d^{25}}\\ &=\frac{1}{2^{13s^2+52s +15}3^{13s} \widetilde{C}_S^{13s} C_d^{13s+25}}. \end{split} \end{equation} Note that in the Euclidean space ${\mathbb R}^n$, $n\ge 2$, we can take $C_d=2^n$, $s=n$, and $\widetilde{C}_S=2^{-1}n^{-1/2}\omega_n^{1/n}$, where $\omega_n$ is the volume of the Euclidean unit ball, and then \begin{equation}\label{eq:choice of beta in Euclidean space} \beta = 2^{-26n^2-64n-15} 3^{-13n}n^{13n/2}\omega_n^{-13}. \end{equation} Recall the definition of the strong boundary from \eqref{eq:strong boundary}. \begin{theorem}\label{thm:comparison of boundaries} Let $\Omega\subset X$ be open and let $E\subset X$ be $\mu$-measurable with $\mathcal H(\Sigma_{\beta} E\cap \Omega)<\infty$. Then $\mathcal H((\partial^*E\setminus \Sigma_{\beta} E)\cap \Omega)=0$. \end{theorem} \begin{proof} By a standard covering argument (see e.g. the proof of \cite[Lemma 2.6]{KKST3}), we find that \[ \lim_{r\to 0}r\frac{\mathcal H(\Sigma_{\beta} E\cap B(x,r))}{\mu(B(x,r))}=0 \] for all $x\in \Omega\setminus (\Sigma_{\beta}E\cup N)$, with $\mathcal H(N)=0$. We will show that $\partial^*E\cap \Omega\subset (\Sigma_{\beta} E\cup N)\cap \Omega$ and thereby prove the claim. Suppose instead that there exists $x\in\Omega\cap \partial^*E\setminus (\Sigma_{\beta} E\cup N)$. Then \[ \lim_{r\to 0}r\frac{\mathcal H(\Sigma_{\beta} E\cap B(x,r))}{\mu(B(x,r))}=0 \] and \[ \limsup_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))}>0 \quad \textrm{and}\quad\limsup_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))}>0. \] Thus for some $0<a<(2C_d^2)^{-1}$ we have \[ \limsup_{r\to 0}\frac{\mu(B(x,r)\cap E)}{\mu(B(x,r))}>C_d a \quad \textrm{and}\quad\limsup_{r\to 0}\frac{\mu(B(x,r)\setminus E)}{\mu(B(x,r))}>C_d a. \] Now we can choose $0<R_0<\tfrac{1}{32} \diam X$ such that \[ \frac{\mu(B(x,40^{-1}R_0)\cap E)}{\mu(B(x,40^{-1}R_0))}>a \] and \[ r\frac{\mathcal H(\Sigma_{\beta} E\cap B(x,r))}{\mu(B(x,r))} <\frac{a}{24 \cdot 2^s C_P C_S C_1 C_2 C_r C_d^8} \] for all $0<r\le R_0$. Choose the smallest $j=0,1,\ldots$ such that for some $r\in (2^{-j-1}R_0,2^{-j}R_0]$ we have \[ \frac{\mu(B(x,40^{-1}r)\setminus E)}{\mu(B(x,40^{-1}r))}>C_d a \quad\textrm{and thus}\quad\frac{\mu(B(x,40^{-1}2^{-j}R_0)\setminus E)}{\mu(B(x,40^{-1}2^{-j}R_0))}>a. \] Let $R:=2^{-j}R_0$. If $j\ge 1$, then \[ \frac{\mu(B(x,20^{-1}R)\setminus E)}{\mu(B(x,20^{-1}R))} \le C_d a \] and so \begin{align*} \frac{\mu(B(x,40^{-1}R)\cap E)}{\mu(B(x,40^{-1}R))} &\ge \frac{\mu(B(x,40^{-1}R))-\mu(B(x,20^{-1}R)\setminus E)}{\mu(B(x,40^{-1}R))}\\ &\ge 1 -C_d \frac{\mu(B(x,20^{-1}R)\setminus E)}{\mu(B(x,20^{-1}R))}\\ &\ge 1 -C_d^2 a\ge 1 -C_d^2 \frac{1}{2 C_d^2}=\frac{1}{2}> a. \end{align*} Thus \begin{equation}\label{eq:portion of E} a<\frac{\mu(B(x,40^{-1}R)\cap E)}{\mu(B(x,40^{-1}R))}<1-a, \end{equation} which holds clearly also if $j=0$, and \[ R\frac{\mathcal H(\Sigma_{\beta} E\cap B(x,R))}{\mu(B(x,R))} <\frac{a}{24 \cdot 2^s C_P C_S C_1 C_2 C_r C_d^8}. \] Let $A:= \Sigma_{\beta} E\cap B(x,R)$. By Proposition \ref{prop:constructing the quasiconvex space} we find an open set $V$ with $A\subset V\subset B(x,2R)$ and such that denoting $Z=X\setminus V$, the space $(Z,d_M^V,\mu)$ is a complete metric space with $d\le d_M^V=(d_M^V)_M$ in $Z$, $\mu$ in $Z$ is a Borel regular outer measure and doubling with constant $\widehat{C}_d=2^s C_1 C_d^2$, and for every $y\in Z$ and $r>0$ we have \begin{equation}\label{eq:lower measure property} \frac{\mu(B_Z(y,r))}{\mu(B(y,r))}\ge (2^s C_1 C_d)^{-1}. \end{equation} Moreover, by choosing a suitably small $\varepsilon>0$, \begin{equation}\label{eq:perimeter of V} P(V,X)\le C_d\mathcal H(A)+\varepsilon < \frac{a}{2^{s+1}C_P C_S C_1 C_d^7}\frac{\mu(B(x,R))}{R}. \end{equation} Thus by the isoperimetric inequality \eqref{eq:isop inequality with zero boundary values}, \[ \mu(V)\le 2C_S RP(V,X)< \frac{1}{C_d^6}\mu(B(x,R)) \le \mu(B(x,40^{-1}R)). \] Thus we can choose $y\in B(x,40^{-1}R)\setminus V$. Denote $F:=X\setminus V$. Let $F_1$ be the component of $\overline{B}(y,20^{-1}R)\setminus V$ containing $y$. By \eqref{eq:size of F1} (and the comments after it) we know that \[ \mu(F_1)\ge (2^s C_1)^{-1}\mu(B(y,20^{-1}R)). \] Since $\mu(\{z\in X:\,d(z,y)=20^{-1}R\})=0$ (see \cite[Corollary 2.2]{Buc}), now also \[ \mu(B(y,20^{-1}R)\cap F_1)\ge (2^s C_1)^{-1}\mu(B(y,20^{-1}R)). \] Suppose that \[ \mu(B(y,20^{-1}R)\setminus F_1)\ge \frac{a}{2^s C_1 C_d^2}\mu(B(y,20^{-1}R)). \] Then \begin{align*} P(V,B(y,20^{-1}R)) &=P(F,B(y,20^{-1}R))\\ &\ge P(F_1,B(y,20^{-1}R))\quad\textrm{by Proposition }\ref{prop:connected components}\\ &\ge \frac{a}{2\cdot 2^s C_P C_1 C_d^2} \frac{\mu(B(y,20^{-1}R))}{20^{-1}R} \quad\textrm{by }\eqref{eq:relative isoperimetric inequality}\\ &\ge \frac{a}{2^{s+1} C_P C_1 C_d^7}\frac{\mu(B(x,R))}{R}. \end{align*} This contradicts \eqref{eq:perimeter of V}, and so necessarily \begin{equation}\label{eq:complement of F1 small measure} \mu(B(y,20^{-1}R)\setminus F_1)< \frac{a}{2^s C_1 C_d^2}\mu(B(y,20^{-1}R)) \le \frac{a}{C_d^2}\mu(B(y,20^{-1}R)). \end{equation} Now \begin{align*} C_d\frac{\mu(B_{Z}(y,10^{-1}R)\cap E)}{\mu(B(y,10^{-1}R))} &\ge \frac{\mu(B(y,20^{-1}R)\cap E\cap F_1)}{\mu(B(y,20^{-1}R))}\\ &\ge \frac{\mu(B(y,20^{-1}R)\cap E)}{\mu(B(y,20^{-1}R))}-\frac{a}{C_d^2}\quad\textrm{by }\eqref{eq:complement of F1 small measure}\\ &\ge \frac{1}{C_d^2}\frac{\mu(B(x,40^{-1}R)\cap E)}{\mu(B(x,40^{-1}R))}-\frac{a}{C_d^2}\\ &> \frac{a}{C_d^2}-\frac{a}{C_d^2}= 0\quad\textrm{by }\eqref{eq:portion of E}. \end{align*} The same string of inequalities holds with $E$ replaced by $X\setminus E$. It follows that \[ 0<\mu(B_{Z}(y,10^{-1}R)\cap E)<\mu(B_Z(y,10^{-1}R)). \] Denoting by $\Sigma_{\beta_0}^ZE$ the strong boundary defined in the space $(Z,d_M^V,\mu)$, by Corollary \ref{cor:density points} we find a point \[ z\in\Sigma_{\beta_0}^Z E\cap B_{Z}(y,9R/10) \subset \Sigma_{\beta_0}^Z E\cap B(y,9R/10)\setminus V \subset \Sigma_{\beta_0}^ZE\cap B(x,R)\setminus V. \] Now using \eqref{eq:lower measure property}, we get \[ \liminf_{r\to 0}\frac{\mu(B(z,r)\cap E)}{\mu(B(z,r))} \ge \liminf_{r\to 0}\frac{\mu(B_{Z}(z,r)\cap E)}{\mu(B_{Z}(z,r))} \frac{\mu(B_{Z}(z,r))}{\mu(B(z,r))} \ge \beta_0\frac{1}{2^s C_1 C_d}=\beta, \] and analogously for $X\setminus E$. Thus $z\in\Sigma_{\beta} E\cap B(x,R)\setminus V$, a contradiction. \end{proof} Recall the usual version of Federer's characterization in metric spaces. \begin{theorem}[{\cite[Theorem 1.1]{L-Fedchar}}]\label{thm:Federers characterization} Let $\Omega\subset X$ be an open set, let $E\subset X$ be a $\mu$-measurable set, and suppose that $\mathcal H(\partial^*E\cap \Omega)<\infty$. Then $P(E,\Omega)<\infty$. \end{theorem} Now we can prove our main result; recall from the discussion on page \pageref{quasiconvex and geodesic} that one can assume the space to be geodesic, as we have done in most of the paper. (However, the constant $\beta$, which is defined explicitly in geodesic spaces in \eqref{eq:definition of beta}, will have a different form in the original space considered in Theorem \ref{thm:main theorem}.) \begin{proof}[Proof of Theorem \ref{thm:main theorem}] By Theorem \ref{thm:comparison of boundaries} we get $\mathcal H(\partial^*E\cap \Omega)<\infty$, and then Theorem \ref{thm:Federers characterization} gives $P(E,\Omega)<\infty$. \end{proof}

https://arxiv.org/abs/1309.4564

Asymptotics of Landau constants with optimal error bounds

We study the asymptotic expansion for the Landau constants $G_n$ $$\pi G_n\sim \ln N + \gamma+4\ln 2 + \sum_{s=1}^\infty \frac {\beta_{2s}}{N^{2s}},~~n\rightarrow \infty, $$ where $N=n+3/4$, $\gamma=0.5772\cdots$ is Euler's constant, and $(-1)^{s+1}\beta_{2s}$ are positive rational numbers, given explicitly in an iterative manner. We show that the error due to truncation is bounded in absolute value by, and of the same sign as, the first neglected term for all nonnegative $n$. Consequently, we obtain optimal sharp bounds up to arbitrary orders of the form $$ \ln N+\gamma+4\ln 2+\sum_{s=1}^{2m}\frac{\beta_{2s}}{N^{2s}}< \pi G_n < \ln N+\gamma+4\ln 2+\sum_{s=1}^{2k-1}\frac{\beta_{2s}}{N^{2s}}$$ for all $n=0,1,2,\cdots$, $m=1,2,\cdots$, and $k=1,2,\cdots$.The results are proved by approximating the coefficients $\beta_{2s}$ with the Gauss hypergeometric functions involved, and by using the second order difference equation satisfied by $G_n$, as well as an integral representation of the constants $\rho_k=(-1)^{k+1}\beta_{2k}/(2k-1)!$.

\section{Introduction and statement of results} \indent\setcounter{section} {1} \setcounter{equation} {0} \label{sec:1} A century ago, it was shown by Landau \cite{Landau} that if a function $f(z)$ is analytic in the unit disc, such that $|f(z)|<1$, with the Maclaurin expansion $$f(z)=a_0+a_1 z+a_2 z^2+\cdots+ a_n z^n+\cdots ,~~~|z|<1,$$ then it holds $$|a_0+a_1+a_2+\cdots +a_n|\leq G_n,~~~n=0, 1,2,\cdots,$$ where $G_0=1$ and \begin{equation}\label{(1.1)} G_n=1+ \left(\frac 1 2\right)^2 +\left ( \frac {1\cdot 3}{2\cdot 4} \right )^2 +\cdots +\left ( \frac{1\cdot 3\cdots (2n-1)}{2\cdot 4\cdots (2n)}\right )^2 \end{equation} for $n=1,2,\cdots$, and the equal sign can be attained for each $n$. The constants $G_n$ are termed Landau's constants; see, e.g., Watson \cite{Watson}. Efforts have been made to approximate these constants from the very beginning. Indeed, Landau himself \cite{Landau} has worked out the large-$n$ behavior \begin{equation*}\label{(1.2)} G_n\sim \frac 1 \pi \ln n,~~~\mbox{as}~~n\rightarrow \infty; \end{equation*} see also Watson \cite{Watson}. Since then, the approximation of $G_n$ goes to two related directions. One is to find sharper bounds of $G_n$ for all positive integers $n$, and the other is to obtain large-$n$ asymptotic approximations for the constants. \subsection{Sharper bounds} Many authors have worked on the sharp bounds of $G_n$. For example, in 1982, Brutman \cite{Brutman} obtains \begin{equation*}\label{(1.3)} 1+ \pi^{-1} \ln(n+1) \leq G_n< 1.0663+ \pi^{-1} \ln(n+1),~~n=0,1,2,\cdots. \end{equation*} The result is improved in 1991 by Falaleev \cite{Falaleev} to give \begin{equation*}\label{(1.4)} 1.0662+\pi^{-1} \ln(n+0.75)< G_n \leq 1.0916 +\pi^{-1} \ln(n+0.75),~~n=0,1,2,\cdots. \end{equation*} In 2000, an attempt is made by Cvijovi\'c \& Klinowski \cite{CK} to use the digamma function $\psi=\Gamma'/\Gamma$ (see, e.g., \cite[p.136, (5.2.2)]{NIST}). They prove that \begin{equation*}\label{(1.5)} c_0 +\pi^{-1} \psi(n+ 5/4)< G_n < 1.0725+\pi^{-1} \psi(n+ 5/4) ,~~n=0,1,2,\cdots , \end{equation*} and \begin{equation*}\label{(1.6)} 0.9883+\pi^{-1} \psi(n+ 3 /2) < G_n < c_0+\pi^{-1} \psi(n+ 3 /2) ,~~n=0,1,2,\cdots , \end{equation*} where $c_0=(\gamma+4\ln 2)/\pi=1.0662\cdots$, $\gamma=0.5772\cdots$ is the Euler constant (\cite[(5.2.3)]{NIST}). Inequalities of this type are revisited in a 2002 paper \cite{Alzer} of Alzer. In that paper, the problem is turned into the following: to find the largest $\alpha$ and smallest $\beta$ such that \begin{equation*}\label{(1.7)}c_0+\pi^{-1} \psi(n+\alpha )\leq G_n\leq c_0+ \psi(n+\beta )~~\mbox{for~all}~ n\geq 0. \end{equation*} The answer is that $\alpha=5/4$ and $\beta=\psi^{-1}(\pi(1-c_0))=1.2662\cdots$, appealing to the complete monotonicity of $\Delta G_n$. In 2009, Zhao \cite{Zhao} starts seeking higher terms in the bounds. A formula in \cite{Zhao}, holding for all positive integer $n$, reads \begin{equation}\label{(1.8)} \ln (16n)+\gamma-\frac 1 {4n} +\frac 5 {192n^2} <\pi G_{n-1} < \ln (16n)+\gamma-\frac 1 {4n} +\frac 5 {192n^2} +\frac 3 {128n^3}. \end{equation} Several authors have made improvements. In a 2011 paper \cite{Mortici}, Mortici gives an inequality of the above type involving higher order term $1/n^{5}$. A $1/n^{7}$ term is brought in by Granath in a recent paper \cite{Granath} in 2012. It seems possible to obtain sharper bounds involving terms of higher and higher orders. Accordingly, difficulties may arise. The case by case process of taking more and more terms might be endless. \subsection{Asymptotic approximations} Most of the above inequalities can be used to derive asymptotic approximations for $G_n$. Such approximations can also be obtained by employing integral representations, generating functions and relations with hypergeometric functions; see, e.g., \cite{LLXZ}. Indeed, back to Watson \cite{Watson}, a formula of asymptotic nature is derived by using a certain integral representation: \begin{equation}\label{(1.9)} \begin{aligned} G_n& = \frac 1 \pi \ln(n+1) +\frac {\{\Gamma(n+3/2)\}^2}{\pi^2\Gamma(n+1)}\sum^{m-1}_{l=1} \frac{\{\Gamma(l+1/2)\}^2}{\Gamma(l+1)\Gamma(n+l+2)}\times \\ & \times \left\{ \psi(l+n+2) -\ln (n+1)+\psi(l+1)-2\psi(l+ 1/2)\right\}+ O\left \{ (n+1)^{1-m}\right \} \end{aligned} \end{equation}for large $n$ and positive integer $m$. Theoretically, an asymptotic expansion can be extracted from (\ref{(1.9)}) by substituting the large-$n$ expansions of $\Gamma$ and $\psi$ into it. In fact, Waston obtains \begin{equation*}\label{(1.10)} G_n\sim \frac 1 \pi\left [ \ln(n+1)+ \gamma+4\ln 2- \frac 1 {4 (n+1)} +\frac 5 {192 (n+1)^2} +\cdots\right ], \end{equation*} of which (\ref{(1.8)}) is an extended version. We skip to some very recent progress in this direction. In the manuscript \cite{Ismail}, Ismail, Li and Rahman derive a complete asymptotic expansion for the Landau constants $G_n$, using the asymptotic sequence $n!/(n + k)!$. The approach is based on a formula of Ramanujan, which connects the Landau constants with a hypergeometric function. Several relevant papers are worth mentioning. In \cite{NemesNemes}, Nemes and Nemes derive full asymptotic expansions using a formula in \cite{CK}. They also conjecture a symmetry property of the coefficients in the expansion. The conjecture has been proved by G. Nemes himself in \cite{Nemes}. \noindent \begin{prop}\label{Prop 1} (Nemes) Let $0 <h < 3/2$. The Landau constants $G_n$ have the following asymptotic expansions \begin{equation}\label{(1.11)} G_n\sim \frac 1 \pi \ln (n+h) +\frac 1 \pi (\gamma+4\ln 2 ) - \sum_{k\geq 1}\frac {g_k(h)}{(n+h)^k}\end{equation} as $n\rightarrow +\infty$, where the coefficients $g_k(h)$ are certain computable constants that satisfy $g_k(h)=(-1)^k g_k(3/2-h)$ for every $k\geq 1$.\end{prop} As an important special case, Nemes \cite{Nemes} has further proved that \begin{equation}\label{(1.12)} \pi G_n\sim \ln (n+3/4) + \gamma+4\ln 2 + \sum_{s=1}^\infty \frac { \beta_{2s}}{ (n+3/4)^{2s}},~~n\rightarrow \infty, \end{equation} where the coefficients $(-1)^{s+1}\beta_{2s}$ are positive rational numbers. The argument in \cite{Nemes} is based on an integral representation of $G_n$ involving a Gauss hypergeometric function in the integrand. While in \cite{LLXZ}, the authors of the present paper study this asymptotic problem by using an entirely different approach, starting from an obvious observation that the Landau constants satisfy a difference equation \begin{equation}\label{(1.13)} G_{n+1}-G_n=\left[\frac{2n+1}{2n+2}\right]^2 (G_n-G_{n-1}),~~~n=0,1,\cdots, \end{equation}as can be seen from the explicit formula (\ref{(1.1)}), where $G_{-1}:=0$. By applying the theory of Wong and Li for second-order linear difference equations \cite{WongLi1992a} to (\ref{(1.13)}), the general expansion in (\ref{(1.11)}) is obtained, and the conjecture of \cite{NemesNemes} is also confirmed. An advantage of this approach, compared with the previous ones, is that all coefficients in the expansion are given iteratively in an explicit manner. \subsection{A question and numerical evidences} As pointed out in \cite{LLXZ}, the case corresponding to (\ref{(1.12)}) is numerically efficient since all odd terms in the expansion vanish. We will find that this expansion in terms of $n+3/4$ is even more special, both from asymptotic and sharper bound points of view. From (\ref{(1.12)}), as suggested by the alternating signs and by numerical calculations, there is a natural question as follows: \noindent {\qe\label{question 1}{Is the error due to truncation of (\ref{(1.12)}) bounded in absolute value by, and of the same sign as, the first neglected term? Or, more precisely, do we have the following? \begin{equation}\label{(1.14)}\frac {\varepsilon_l(N)}{\beta_{2l}/N^{2l}} \in (0, 1) ~~\mbox{for}~~n=0,1,2,\cdots~~\mbox{and}~~l=1,2,\cdots, \end{equation} where $N=n+3/4$, and \begin{equation}\label{(1.15)}\varepsilon_l(N)=\pi G_n - \left \{ \ln N+\gamma+4\ln 2+\sum_{s=1}^{l-1} \frac{\beta_{2s}}{N^{2s}}\right\}.\end{equation} }}\vskip .5cm Recalling that $(-1)^{s+1}\beta_{2s}$ are positive, it is readily seen that a positive answer to (\ref{(1.14)}) is equivalent to \begin{equation}\label{(1.16)} \varepsilon_{2k}(N)< 0~~\mbox{and}~~\varepsilon_{2k-1}(N)>0 \end{equation}for all $k=1,2,3,\cdots$ and $n=0,1,2,\cdots$. The question reminds us of an earlier work of Shivakumar and Wong \cite{ShivakumarWong}, where an asymptotic expansion is obtained for the Lebesgue constants associated with the polynomial interpolation at the zeros of the Chebyshev polynomials, and the error in stopping the series at any time is shown to have the sign as, and is in absolute value less than, the first term neglected. Similar discussion can be found in, e.g., Olver \cite[p.285]{olver1974}, on the Euler-Maclaurin formula. Numerical experiments agree with (\ref{(1.14)}). The functions $\frac {\varepsilon_l(N)}{\beta_{2l}/N^{2l}}$, $N=n+\frac 3 4$, are depicted in Figure \ref{figure 1} for the first few $n$. \begin{figure}[h] \begin{center} \includegraphics[height=6cm]{figure1a.eps}\hskip 1cm \includegraphics[height=6.12cm]{figure1b.eps} \caption{The function $\frac {\varepsilon_l(N)}{\beta_{2l}/N^{2l}}$. Left: $l=2$, $n=0$-$30$. Right: $l=16$, $n=0$-$50$.} \label{figure 1} \end{center} \end{figure} \subsection{Statement of results } In the present paper, we will justify (\ref{(1.16)}). In fact, we will prove the following theorem. \noindent \begin{thm}\label{Thm 1}For $N=n+3/4$, it holds \begin{equation}\label{(1.17)} (-1)^{l+1} \varepsilon_l(N) >0\end{equation} for $l= 1,2,3,\cdots$ and $n=0,1,2,\cdots$, where $\varepsilon_l(N)$ is defined in (\ref{(1.15)}), the coefficients $\beta_{2s}$ are determined iteratively in (\ref{(2.4)}) below. \end{thm} The above theorem has direct applications both in asymptotics and sharp bounds. In asymptotic point of view, we can obtain error bounds which in a sense are optimal. To be precise, we have the following. \noindent \begin{thm}\label{Thm 2}The error due to truncation of (\ref{(1.12)}) is bounded in absolute value by, and of the same sign as, the first neglected term for all nonnegative $n$. That is, \begin{equation}\label{(1.18)} 0< (-1)^{l+1} \varepsilon_l(N) = \left | \varepsilon_l(N) \right | < \frac {\left | \beta_{2l}\right |} {N^{2l}} = \frac {(-1)^{l+1} \beta_{2l}} { N^{2l}}\end{equation} for $l=0,1,2\cdots$ and $n=0,1,2,\cdots$, where $N=n+3/4$. \end{thm} The error bound in (\ref{(1.18)}) is the first neglected term in the asymptotic expansion, and hence is optimal and can not be improved. The inequalities in (\ref{(1.18)}) can be derived from Theorem \ref{Thm 1} by noticing that $\varepsilon_l(N)= \beta_{2l}/N^{2l}+\varepsilon_{l+1}(N)$, as can be seen from (\ref{(1.15)}). Another application of Theorem \ref{Thm 1} is the construction of sharp bounds up to arbitrary orders. \noindent \begin{thm}\label{Thm 3}For $N=n+3/4$, it holds \begin{equation}\label{(1.19)} \ln N+\gamma+4\ln 2+\sum_{s=1}^{2m}\frac{\beta_{2s}}{N^{2s}}< \pi G_n < \ln N+\gamma+4\ln 2+\sum_{s=1}^{2k-1}\frac{\beta_{2s}}{N^{2s}} \end{equation} for all $n=0,1,2,\cdots$, $m=1,2,\cdots$, and $k=1,2,\cdots$. \end{thm} The inequalities in (\ref{(1.19)}) are understood as sharp bounds on both sides up to arbitrary orders. In a sense, the bounds are optimal and can not be improved. The first few coefficients $\beta_{2s}$ are listed in Table \ref{table1}, as can be evaluated via (\ref{(2.4)}). \noindent \begin{table}[h] \centering \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|} \hline $\beta_2$ & $\beta_4$ & $\beta_6$ & $\beta_8$ & $\beta_{10} $ &$\beta_{12}$ & $\beta_{14}$ \\[0.1cm] $\frac {11}{192}$ & $\frac{-1541}{122880}$ &$\frac{63433}{8257536}$&$\frac {-9199901}{1006632960}$&$\frac { 317959723}{17716740096}$&$\frac {-14849190321163}{281406257233920}$&$\frac {717209117969}{3298534883328}$ \\[0.1cm] \hline \end{tabular} \caption{The first few $\beta_{2s}$, $s=1,2,\cdots, 7$.}\label{table1} \end{table} \section { Proof of Theorem \ref{Thm 1} } \indent\setcounter{section} {2} \setcounter{equation} {0} \label{sec:2} The proof is based on the difference equation (\ref{(1.13)}) and an approximation of the coefficients $\beta_{2s}$. To justify Theorem \ref{Thm 1}, several lemmas are stated, and all, except one, are proved in the present section. While the validity of Lemma \ref{lem 2.2} is the objective of the next section. \subsection{The coefficients $\beta_s$ in (\ref{(1.12)})} Write the difference equation (\ref{(1.13)}) in the symmetrical form \begin{equation}\label{(2.1)} \left (1+\frac 1 {4N}\right )^2 w(N+1) -\left (2 +\frac 1 {8N^2}\right ) w(N)+\left (1-\frac 1 {4N}\right )^2 w(N-1)=0, \end{equation}in which $N=n+\frac 3 4$ for $n=0,1,2,\cdots$. As mentioned earlier and as in the previous paper \cite{LLXZ}, the Landau constants $w(N)=G_n$ solves (\ref{(2.1)}), having an asymptotic expansion \begin{equation}\label{(2.2)} \pi G_n\sim \ln N+ \gamma+4\ln 2 +\sum^\infty_{s=1} \frac {\beta_s} {N^s},\end{equation} and the coefficients are determined by a formal substitution of (\ref{(2.2)}) into (\ref{(2.1)}); see Wong and Li \cite{WongLi1992a}. The following result then follows: \noindent \begin{lem}\label{lem 2.1}For $N=n+\frac 3 4$, the coefficients $\beta_s$ in expansion (\ref{(2.2)}) fulfill \begin{equation}\label{(2.3)} \beta_{2k+1}=0,~~k=0,1,2,\cdots, \end{equation}and \begin{equation}\label{(2.4)} \beta_{2k}=\frac {-1} {4k^2} \left( d_{k-1, k+1} \beta_{2k-2} + d_{k-2, k+1} \beta_{2k-4}+\cdots + d_{1, k+1} \beta_2 -d_{0, k+1}\right ),~~k=1,2,\cdots, \end{equation} where $d_{j, j+1}= 4j^2$ for $j=1,2,\cdots$, \begin{equation}\label{(2.5)} d_{j, s}= \frac { (2s+2j-2) \;(2s-2)! }{(2s-2j)!\; (2j-1)!} + \frac { (2s-3)!}{8(2s-2j-2)!\; (2j-1)!}~~\mbox{for} ~~ s\geq j+2, \end{equation}and \begin{equation}\label{(2.6)} d_{0, s}= \left (\frac 1 s-\frac 1 {2s-1} \right ) +\frac 1 {16(s-1)},~~s=2,3,\cdots. \end{equation} In addition, it holds \begin{equation}\label{(2.7)} \beta_{2k}=(-1)^{k+1}\left | \beta_{2k}\right |,~~k=1,2,\cdots. \end{equation} \end{lem}\vskip .5cm Part of this lemma ( (\ref{(2.3)}) and (\ref{(2.7)}) ) has been proved in Nemes' recent paper \cite{Nemes}. Part of it, namely (\ref{(2.3)}), and an equivalent form of (\ref{(2.4)}), has been proved in our earlier paper \cite{LLXZ}. Following Wong and Li \cite{WongLi1992a}, (\ref{(2.3)}) and (\ref{(2.4)}) can be justified by substituting (\ref{(2.2)}) into (\ref{(2.1)}), expanding both sides in formal power series of $1/N$, and equalizing the coefficients of the same powers. It is readily seen that all $d_{j, s}>0$ for $l\geq 1$ and $s\geq l+1$, and $d_{0, s}>0$ for $s\geq 2$. \subsection{Analysis of $R_l(N)$} Here, \begin{equation}\label{(2.8)} R_l(N)= \left (1+\frac 1 {4N}\right )^2\varepsilon_l(N+1) -\left (2 +\frac 1 {8N^2}\right ) \varepsilon_l(N)+\left (1-\frac 1 {4N}\right )^2\varepsilon_l(N-1), \end{equation}with the error term $\varepsilon_l(N)$ being given in (\ref{(1.15)}), and $N=n+3/4$. There are several facts worth mentioning. It is readily seen from \eqref{(1.15)} that $\pi G_n-(\gamma+4\ln2)$ satisfies the difference equation \eqref{(2.1)}, and can then be removed from $R_l(N)$ in \eqref{(2.8)}. If we write $x=1/N$, then the logarithmic singularity at $x=0$ is also cancelled in $R_l(N)$. Therefore, each $R_l(N)$ is an analytic function in $x$ for $|x|<1$. Hence the asymptotic expansion for $R_l(N)$, in descending powers of $N$, is actually a convergent Taylor expansion in $x$, \begin{equation}\label{(2.9)} R_l(N)= \sum_{s=l+1}^\infty r_{l,s} x^{2s}, ~~ |x|<1, \end{equation}where \begin{equation}\label{(2.10)} r_{l,s} = -\left ( d_{l-1, s} \beta_{2l-2} + d_{l-2, s} \beta_{2l-4}+\cdots + d_{1, s} \beta_2 -d_{0, s}\right ) \end{equation}for $s\geq l+1$, with the leading coefficient $r_{l,l+1}=4l^2\beta_{2l}$; cf. (\ref{(2.4)}). For later use, we estimate the ratio of the consecutive coefficients $\beta_{2s}$. To this end, we introduce a sequence of positive constants \begin{equation}\label{(2.11)}\rho_0=1, ~~\mbox{and}~\rho_l=\frac {(-1)^{l+1} \beta_{2l}} {(2l-1)!},~~l=1,2,\cdots .\end{equation} We shall use the following lemma and leave its proof to Section \ref{sec:3} below. \noindent\begin{lem}\label{lem 2.2} It holds \begin{equation}\label{(2.12)} \rho_k/\rho_{k+1}\leq \frac {44} 9 \pi^2 \end{equation} for $k=0,1,2,\cdots$. \end{lem} Now we proceed to analyze $R_l(N)$ (sometimes denoted by $R_l(x)$, understanding that $x=1/N$), so as to show that $R_l(x)/\beta_{2l} \geq 0$ for $x\in [0, 1)$. More precisely, we prove a much stronger result, as follows: \noindent \begin{lem}\label{lem 2.3} For $l=1,2,\cdots$ and $N=n+3/4$ with $n=1,2,\cdots $, we have \begin{equation}\label{(2.13)} (-1)^{l+1} r_{l,s} >0, ~~ s\geq l+1 , \end{equation} where $r_{l,s}$ are given in (\ref{(2.9)}) and (\ref{(2.10)}). \end{lem} \noindent {\bf{Proof:}} The lemma can be proved by using induction with respect to $l$. Initially, we have \begin{equation*}\label{(2.14)} R_1(N)= \sum_{s=2}^\infty d_{0,s} x^{2s}. \end{equation*}Since $d_{0, s} >0$ for $s=2,3,\cdots$; cf. (\ref{(2.6)}), we see that (\ref{(2.13)}) holds for $l=1$. In view of the fact that $\beta_2=\frac {11}{192}$; cf. Table \ref{table1}, it is readily verified that \begin{equation*}\label{(2.15)} R_2(N)=\sum_{s=3}^\infty r_{2,s} x^{2s} =\sum_{s=3}^\infty \left ( -\frac {11}{192} d_{1, s}+d_{0,s}\right ) x^{2s}, \end{equation*}with all coefficients $r_{2,s}$ being negative. Indeed, in view of (\ref{(2.5)}) and (\ref{(2.6)}), we have $$ r_{2,s} =-\frac {11}{192}\left (2s+\frac {2s-3} 8\right ) +\left [ \left (\frac 1 s-\frac 1 {2s-1}\right )+\frac 1 {16(s-1)}\right ] <-\frac {11}{192} \cdot (2s)+\frac 1 s <0$$ for $s\geq 3$. Thus (\ref{(2.13)}) holds for $l=2$. Similarly, we can verify (\ref{(2.13)}) for $l=3$, recalling that $\beta_4=-\frac {1541}{122880}$; cf. Table \ref{table1}. Assume that for $l\leq k$, it holds $(-1)^{l+1} r_{l,s} >0$ for $s\geq l+1$. From (\ref{(2.10)}), we can write \begin{equation}\label{(2.16)}(-1)^{k+3} r_{k+2,s} =(-1)^{k+1}r_{k, s} +(-1)^k \left (d_{k+1, s} \beta_{2k+2}+d_{k,s}\beta_{2k}\right ) \end{equation}for $s\geq k+3 $. To show that (\ref{(2.13)}) is valid for $l=k+2$, it suffices to show that \begin{equation*}\label{(2.17)}(-1)^k \left (d_{k+1, s} \beta_{2k+2}+d_{k,s}\beta_{2k}\right )>0\end{equation*} for $s\geq k+4 $ since the validity of (\ref{(2.13)}) for $(l,s)=(k+2, k+3)$ is trivial. This is equivalent to show that \begin{equation}\label{(2.18)}c_{k+1, s}-c_{k, s} \rho_k/\rho_{k+1} = \frac { (-1)^k \left (d_{k+1, s} \beta_{2k+2}+d_{k,s}\beta_{2k}\right ) } {8(s-1)^2\; (2s-3)!\; \rho_{k+1}} >0,\end{equation} where $\rho_k>0$ is defined in \eqref{(2.11)}, and $c_{k,s}=\frac {(2k-1)! d_{k,s}}{8(s-1)^2 (2s-3)!}$; cf. (\ref{(3.2)}) below. In view of (\ref{(2.5)}), we may write $$ c_{k,s}=\left\{\frac 1 {2(2s-2k)!}+\frac k {2(s-1) (2s-2k)!} \right\} +\left\{ \frac 1 {64(s-1)^2(2s-2k-2)!}\right\}:=A+B.$$ For $k\geq 1$ and $s-k\geq 4$, we have $$c_{k+1,s}\geq (2s-2k-1)(2s-2k) A+(2s-2k-3)(2s-2k-2)B\geq 56A+30B >\frac {478} 9c_{k,s}.$$ The last inequality holds since $A>8B$. From (\ref{(2.12)}) in Lemma \ref{lem 2.2}, it is readily verified that $$ \frac {478} 9 > \frac {44} 9 \pi^2\geq \frac {\rho_k}{\rho_{k+1}}$$ for $k\geq 1$. Then (\ref{(2.18)}) holds for $s\geq k+4 $, and it follows that $(-1)^k \left (d_{k+1, s} \beta_{2k+2}+d_{k,s}\beta_{2k}\right )>0$ for $s\geq k+4 $. Accordingly, from (\ref{(2.16)}) we see that (\ref{(2.13)}) holds for $l=k+2$. This completes the proof of Lemma \ref{lem 2.3}. \subsection{Proof of Theorem \ref{Thm 1} } Now Lemma \ref{lem 2.3} implies that $(-1)^{l+1} R_l(N) >0$ for all $l$ and all $N=n+3/4$. To show that $\tilde\varepsilon_l(N):=(-1)^{l+1} \varepsilon_l (N)>0$ for all $N$, we note first that $\tilde\varepsilon_l(N)= \frac {|\beta_{2l}|}{N^{2l}}\left\{1+O\left (\frac 1 {N^2}\right )\right \}$ as $N\rightarrow +\infty$. Hence $\tilde\varepsilon_l (N)>0$ for $N$ large enough. Now assume that (\ref{(1.17)}) is not true. Then there exists a finite $M$ defined as $$M=\max\{N= n+3/4: ~n\in \mathbb{Z}~\mbox{and}~\tilde\varepsilon_l(N)\leq 0\},$$ so that $M-3/4$ is a positive integer and $\tilde\varepsilon_l(M)\leq 0$, while $\tilde\varepsilon_l(M+1), \tilde\varepsilon_l(M+2), \cdots >0$. For simplicity, we denote $a_N=(1+\frac 1 {4N} )^2$ and $b_N=(1-\frac 1 {4N} )^2$. From (\ref{(2.8)}) we have \begin{equation*}\label{(2.19)} a_{M+1}\tilde\varepsilon_l(M+2)=(a_{M+1}+b_{M+1}) \tilde\varepsilon_l(M+1) + b_{M+1} \left (- \tilde\varepsilon_l(M)\right ) + (-1)^{l+1} R_l(M+1). \end{equation*}The later terms on the right-hand side are non-negative, hence we obtain \begin{equation*}\label{(2.20)} a_{M+1}\tilde\varepsilon_l(M+2)\geq \left (a_{M+1}+b_{M+1}\right ) \tilde\varepsilon_l(M+1) , \end{equation*}which implies that \begin{equation}\label{(2.21)} \tilde\varepsilon_l(M+2)> \tilde\varepsilon_l(M+1). \end{equation} Using (\ref{(2.8)}) again for $N=M+2$, we have \begin{equation*}\label{(2.22)} a_{M+2}\tilde\varepsilon_l(M+3)\geq (a_{M+2}+b_{M+2}) \tilde\varepsilon_l(M+2) - b_{M+2} \tilde\varepsilon_l(M+1) . \end{equation*} A combination of the previous two inequalities gives \begin{equation}\label{(2.23)} \tilde\varepsilon_l(M+3)> \tilde\varepsilon_l(M+2). \end{equation} In general, we obtain \begin{equation}\label{(2.24)} \tilde\varepsilon_l(M+k+1)> \tilde\varepsilon_l(M+k) \end{equation}by induction. From the equalities in (\ref{(2.21)}), (\ref{(2.23)}) and (\ref{(2.24)}), we conclude that \begin{equation}\label{(2.25)} \tilde\varepsilon_l(M+k)> \tilde\varepsilon_l(M+1) \end{equation} for all $k\geq 2$. Recalling that $\tilde\varepsilon_l(N)= O(N^{-2l})$ for $N\rightarrow\infty$, letting $k\rightarrow \infty$ will give $\tilde\varepsilon_l(M+1) \leq 0$. This contradicts the definition of $M$. Thus we have proved Theorem \ref{Thm 1}. \section { Proof of Lemma \ref{lem 2.2}} \indent\setcounter{section} {3}\setcounter{subsection} {0} \setcounter{equation} {0} \label{sec:3} The idea is simple: to approximate the coefficients $\beta_{2s}$, and then to work out the ratio $\beta_{2s}/\beta_{2s+2}$. Yet the procedure is complicated. A brief outline of the proof is as follows: In Section \ref{sec:3.1}, we bring in an ordinary differential equation \eqref{(3.11)} with a specific analytic solution $v(z)$, of which $\rho_k=\frac {(-1)^{k+1} \beta_{2k}} {(2k-1)!}$ are coefficients of the Maclaurin expansion. The function $v(z)$ is then extended, in Section \ref{sec:3.2}, and via the hypergeometric functions, to a function analytic in the cut-strip $\{z ~|~ -4\pi<\mathop{\rm Re}\nolimits z< 4\pi, ~z\not \in \{(-\infty, -2\pi]\cup [2\pi, +\infty)\} \}$. An integral representation is then obtained by using the Cauchy integral formula, and the integration path is deformed based on the analytic continuation procedure. In Section \ref{sec:3.3}, the integral is spilt, approximated, and estimated, and hence bounds for $\rho_k$ on both sides are established in \eqref{(3.36)} for all $k\geq 10$. Eventually, in Section \ref{sec:3.4}, an upper bound for $\rho_k/\rho_{k+1}$ is obtained for all non-negative integer $k$. \subsection{Differential equation}\label{sec:3.1} In terms of $\rho_s$ defined in (\ref{(2.11)}), namely, $\rho_0=1$ and $\rho_l=\frac {(-1)^{l+1} \beta_{2l}} {(2l-1)!}$ for $l=1,2,\cdots$, formula (\ref{(2.4)}) can be written as \begin{equation}\label{(3.1)}c_{l,l+1} \rho_l -c_{l-1,l+1} \rho_{l-1} +\cdots + (-1)^{k} c_{l-k, l+1} \rho_{l-k} +\cdots + (-1)^{l-1} c_{1,l+1} \rho_1 +(-1)^l c_{0,l+1}\rho_0 =0 \end{equation} for $l=1,2,\cdots$, where $c_{l,l+1}= \frac 1 {2!}$, and \begin{equation}\label{(3.2)} c_{l-k,l+1}=\frac {(2l-2k-1)!}{(2l-1)!}\frac {d_{l-k, l+1}}{2d_{l, l+1}}= \frac {1}{2(2k+2)!} +\frac {l-k}{2l (2k+2)!}+\frac 1 {64 l^2 \cdot (2k)!}\end{equation} for $k=1,2,\cdots l-1$ and $l=1,2,\cdots$. It can be verified from (\ref{(2.6)}) that $c_{0, l+1}=\frac {d_{0, l+1}}{8l^2(2l-1)!}$ also takes the same form, that is, (\ref{(3.2)}) is also valid for $k=l$, $l=1,2,\cdots$. The idea now is to approximate $\rho_s$, and then to estimate the ratio $\rho_{s}/\rho_{s+1}$. Taking $l=1,2,\cdots$ in (\ref{(3.1)}), we have $$ \begin{array}{l} a_1:=c_{1,2}\rho_1-c_{0,2}\rho_0=0, \\[0.2cm] a_2:=c_{2,3}\rho_2-c_{1,3}\rho_1+c_{0,3}\rho_0=0,\\[0.2cm] a_3:=c_{3,4}\rho_3 - c_{2,4}\rho_2+c_{1,4}\rho_1-c_{0,4}\rho_0=0,\\[0.2cm] \cdots\cdots. \end{array}$$ Summing up $\sum^\infty_{s=1} a_s x^{2s}$ gives \begin{equation}\label{(3.3)}\rho_0\left (-c_{0,2} x^2+c_{0,3}x^4-c_{0,4} x^6+\cdots\right )+\sum^\infty_{k=1} \rho_k x^{2k} \sum^\infty_{s=1} (-1)^{s-1} c_{k,k+s} x^{2s-2}=0.\end{equation} In view of (\ref{(3.2)}), it is readily verified by summing up the series that \begin{equation}\label{(3.4)}-c_{0,2} x^2+c_{0,3}x^4-c_{0,4} x^6+\cdots=-\frac 1 4 +\frac {h(x)}{2} - \int^x_0 \frac {dt_1} {t_1}\int^{t_1}_0 \frac {t h(t) dt}{16}, \end{equation} where \begin{equation*}\label{(3.5)}h(x):=\frac {1-\cos x} {x^2}.\end{equation*} Also we have, for $k=1,2,\cdots$, \begin{equation}\label{(3.6)}\sum^\infty_{s=1} (-1)^{s-1} c_{k,k+s} x^{2s-2} =\frac { h(x)} 2+\frac k {x^{2k}} \int^x_0 t^{2k-1} h(t) dt - \frac 1 { x^{2k}}\int^x_0\frac {dt_1}{t_1}\int^{t_1}_0\frac{t^{2k+1}h(t)dt }{16}.\end{equation} Substituting (\ref{(3.4)}) and (\ref{(3.6)}) into (\ref{(3.3)}), we obtain an equation \begin{equation}\label{(3.7)}-\frac 1 4 +\frac 1 2 h(x)u(x)+\int^x_0\frac 1 2 h(t) u'(t) dt -\int^x_0\frac {dt_1}{t_1}\int^{t_1}_0\frac t{16} h(t) u(t) dt=0,\end{equation} where \begin{equation}\label{(3.8)}u(x):=\sum^\infty_{k=0}\rho_k x^{2k}.\end{equation} \noindent \begin{rem}\label{rem 1} The existence of $u(x)$ defined above and the validity of (\ref{(3.3)}) can be justified by showing that $ \left |\rho_k\right |\leq M_0/\delta^{2k}$ for all positive integers $k$, $M_0$ and $\delta$ being positive constants. Indeed, from (\ref{(3.2)}) it is readily seen that $\left |c_{l-k, l+1} \right |\leq \frac 1 {(2k)!}$ for $k=1,2,\cdots , l$. Now we assume that $ \left |\rho_k\right |\leq M_0/\delta^{2k}$ for $k<l$, where $\delta$ is small enough such that $2(\cosh \delta -1) <1$. Then, by using (\ref{(3.1)}) we have $$\frac 1 2 |\rho_l| \leq \sum^l_{k=1} \left |c_{l-k, l+1} \right | |\rho_{l-k}| \leq \sum^l_{k=1} \frac {M_0\delta^{2k-2l}} {(2k)!} \leq \frac {M_0}{\delta^{2l}} \left (\cosh\delta -1\right ).$$ Hence we have $ \left |\rho_l \right |\leq M_0/\delta^{2l}$ by induction. \end{rem} Applying the operator $\displaystyle{\frac d {dx} \left ( x \frac d {dx} \right )}$ to both sides of (\ref{(3.7)}), we see that $u(x)$ solves the second order differential equation \begin{equation}\label{(3.9)}\left [ x\left (\frac 1 2 h'(x) u(x)+h(x) u'(x)\right )\right ]' -\frac {xh(x) u(x)}{16} =0 \end{equation} in a neighborhood of $x=0$, with initial conditions $u(0)=1$ and $u'(0)=0$. In the next few steps we derive a representation of $u(x)$ for later use. First, substituting \begin{equation}\label{(3.10)}v(x)=\sqrt{h(x)} u(x)=\frac {\sqrt 2 \sin\frac x 2} x u(x)\end{equation} into equation (\ref{(3.9)}) yields \begin{equation}\label{(3.11)}\sin\frac x 2 v''(x)+\frac 1 2 \cos\frac x 2 v'(x) -\frac 1 {16} \sin \frac x 2 v(x)=0 \end{equation}in a neighborhood of $x=0$, with $v(0)=1/\sqrt 2$ and $v'(0)=0$. It is shown in Remark \ref{rem 1} that $u(x)$ is analytic at the origin. So is $v(x)$; cf. \eqref{(3.10)}. What is more, near $x=0$, the function $v(x)$ can be represented as a hypergeometric function. Indeed, a change of variable \begin{equation*}\label{(3.12)}t=\frac {1-\cos x} 2=\sin^2\frac x 2\end{equation*} turns the equation into the hypergeometric equation \begin{equation}\label{(3.13)} t(1-t)\frac {d^2 v}{d t^2}+\left (1-\frac 3 2 t\right )\frac {dv}{dt}-\frac 1 {16} v=0.\end{equation} Taking the initial conditions into account, it is easily verified that \begin{equation}\label{(3.14)}v(x)= \frac 1 {\sqrt 2} F\left (\frac 14, \frac 1 4; 1; \sin^2\frac x 2\right ) =\frac 1 {\sqrt 2} F\left (\frac 12, \frac 1 2; 1; \sin^2\frac x 4\right ) ,\end{equation} initially at $x=0$, and then analytically extended elsewhere; cf. \cite[(15.2.1)]{NIST}. The second equality follows from a quadratic hypergeometric transformation; see \cite[(15.8.18)]{NIST}. \subsection{Analytic continuation}\label{sec:3.2} Well-known formulas for hypergeometric functions include \begin{equation}\label{(3.15)} F(a,b;c;t)=\frac {\Gamma(c)}{\Gamma(b)\Gamma(c-b)} \int^1_0 s^{b-1}(1-s)^{c-b-1}(1-st)^{-a} ds\end{equation} for $t\in \mathbb{C}\backslash [1, +\infty)$ as $\mathop{\rm Re}\nolimits c>\mathop{\rm Re}\nolimits b >0$; cf. \cite[(15.6.1)]{NIST}, which extends the hypergeometric function to a single-valued analytic function in the cut-plane. Now we proceed to consider (\ref{(3.11)}) with complex variable $z$. It is worth noting that the solution $v(z)$ we seek is an even function. So we restrict ourselves to its analytic continuation on the right-half plane $\mathop{\rm Re}\nolimits z >0$. To this aim, we define \begin{equation}\label{f-def} f(z) =F\left (\frac 1 2,\frac 1 2;1;\sin^2\frac z 4\right ) ~~\mbox{for}~\mathop{\rm Re}\nolimits z\in (-2\pi, 2\pi)\cup (2\pi, 6\pi).\end{equation} We see that $t=\sin^2\frac z 4$ maps the strip $-2\pi< \mathop{\rm Re}\nolimits z <2\pi$ duplicately and analytically onto the cut-plane $t\in \mathbb{C}\backslash [1, +\infty)$. The same is true for $2\pi< \mathop{\rm Re}\nolimits z <6\pi$. Hence from (\ref{(3.14)}) we have the analytic continuation $v(z)=f(z)$ for $-2\pi<\mathop{\rm Re}\nolimits z<2\pi$. Moreover, from \eqref{f-def} it is readily seen that the function $f(z)$, defined and analytic in the disjoint strips, satisfies \begin{equation*}\label{f-periodic}f(z)=f(z-4\pi)~~\mbox{for}~~\mathop{\rm Re}\nolimits z\in (2\pi, 6\pi). \end{equation*} Careful treatment should be brought in here, since there is a logarithmic singularity of $v(z)$ at the boundary point $z= 2\pi$, as we will see later, and as can be seen from the equation (\ref{(3.11)}): $z=2\pi$ is one of the nearest regular singularities, and the indicial equation has a double root $0$ there. Next, we extend the domain of analyticity of $v(z)$ beyond the vertical line $\mathop{\rm Re}\nolimits z=2\pi$. To do so, we recall the jump along the branch cut of the hypergeometric function \begin{equation}\label{(3.21)}F(a,b;c;x+i0)- F(a,b;c;x-i0) = \frac {2\pi i \Gamma(c)}{\Gamma(a) \Gamma(b)} \frac { (x-1)^{c-a-b}} {\Gamma(c-a-b+1)} F(c-a, c-b; c-a-b+1; 1-x) \end{equation} for $x>1$; see \cite[(15.2.2)-(15.2.3)]{NIST}. Applying (\ref{(3.21)}) to $f(z)$ defined in \eqref{f-def}, a careful calculation yields \begin{equation*}\label{f-jump-upper}f(2\pi -0+iy)-f(2\pi +0+iy)=2i f(iy )~~\mbox{for}~~y=\mathop{\rm Im}\nolimits z >0, \end{equation*} where use has been made of the fact that $t=\sin^2 \frac {2\pi-0+iy} 4$, $y=\mathop{\rm Im}\nolimits z>0$ corresponds to the upper edge of $[1, \infty)$, while $t=\sin^2 \frac {2\pi+0+iy} 4=\sin^2 \frac {-2\pi+0+iy} 4$, $y=\mathop{\rm Im}\nolimits z>0$ corresponds to the lower edge, and that $t=\sin^2 \frac z 4$ maps the upper imaginary axis to $(-\infty, 0)$. Similarly we have \begin{equation*}\label{f-jump-lower}f(2\pi -0+iy)-f(2\pi +0+iy)=-2i f(iy )~~\mbox{for}~~y=\mathop{\rm Im}\nolimits z <0. \end{equation*} Summarizing the above, we have an analytic function in the cut-strip $\{z ~|~ 0\leq \mathop{\rm Re}\nolimits z< 4\pi, ~z\not \in [2\pi, +\infty) \}$, defined as follows: \begin{equation}\label{(3.25)} v(z)=\left\{ \begin{array}{ll} f(z), & 0\leq \mathop{\rm Re}\nolimits z<2\pi, \\[.1cm] f(z) +2if(z-2\pi), & 2\pi<\mathop{\rm Re}\nolimits z <4\pi,~ \mathop{\rm Im}\nolimits z >0, \\[.1cm] f(z) -2if(z-2\pi), & 2\pi<\mathop{\rm Re}\nolimits z <4\pi,~ \mathop{\rm Im}\nolimits z <0. \end{array}\right . \end{equation}The value at $\mathop{\rm Re}\nolimits z=2\pi$, $\mathop{\rm Im}\nolimits z\not=0$ is obtained by taking limit. At last, we confirm that there is a logarithmic singularity at $z=2\pi$, a regular singularity of (\ref{(3.11)}). Indeed, following the derivation from (\ref{(3.11)}) to (\ref{(3.13)}), we see that all solutions to (\ref{(3.11)}) takes the form $$A w_1(z) +B \left \{ w_1(z)\ln \left (\sin^2\frac z 2\right ) + w_2(z)\right\}=A w_1(z) +B \left \{2 w_1(z)\ln \left (z -2\pi \right ) + \tilde w_2(z)\right\};$$ cf. Wong \cite[(2.1.24)]{Wong2010}, where $w_2(z)$ and $\tilde w_2(z)$ are specific single-valued analytic functions at $z=2\pi$, $A$ and $B$ are constants, and $w_1(z)= f(z-2\pi)$ is an analytic solution of (\ref{(3.11)}) at $z=2\pi$. The function $v(z)$ in (\ref{(3.25)}) is such a solution, and, what is more, with $B=-\frac 1 \pi$, as can be seen by comparing the jumps along $(2\pi, 3\pi)$. Accordingly, we may write \begin{equation}\label{(3.26)} v(z)=v_A(z) -\frac 2 \pi f(z-2\pi)\ln \left (2\pi -z \right ) \end{equation} for $0< \mathop{\rm Re}\nolimits z <4\pi$, with $v_A(z)$ being analytic in the strip, and the branch of logarithm being chosen as $\arg (2\pi -z)\in (-\pi, \pi)$. \subsection{Approximation of $\rho_s$}\label{sec:3.3} Now we are in a position to approximate $\rho_s$. To begin, we mention several known facts. From (\ref{(3.25)}), and that $v(z)$ is even, we know that the function $v(z)$ is now extended analytically to the cut-strip $$\{z ~|~ -4\pi<\mathop{\rm Re}\nolimits z< 4\pi, ~z\not \in \{(-\infty, -2\pi]\cup [2\pi, +\infty)\} \}.$$ Again, from \eqref{f-def}, (\ref{(3.25)}) and (\ref{(3.26)}), and that the hypergeometric function $F(a,a;c;t)$ behaviors of the form $O(|t|^{-a}\ln |t|)$ at infinity; see \cite[(15.8.8)]{NIST}, we conclude that $$ v(z)=O\left ( e^{-\frac 1 4 |\mathop{\rm Im}\nolimits z|}\ln |\mathop{\rm Im}\nolimits z |\right )$$ as $\mathop{\rm Im}\nolimits z \rightarrow \infty$, holding uniformly in $|\mathop{\rm Re}\nolimits z |\leq 4\pi -\delta$ for arbitrary positive $\delta$. Thus $v(z)$ displays an exponentially small decay at infinity in each sub-strip. Hence, we can derive from (\ref{(3.8)}) and (\ref{(3.10)}) that \begin{equation}\label{(3.27)}\rho_k=\frac 1 {2\pi i} \oint z^{-2k-1} u(z)dz =\frac 1 {\sqrt 2\pi i} \oint \frac {\frac z 2}{\sin \frac z 2}\frac { v(z) dz} { z^{2k+1} } =\frac 1 {\sqrt 2\pi i} \int_\Gamma \frac {\frac z 2}{\sin \frac z 2}\frac { v(z) dz} { z^{2k+1} },\end{equation} where initially the integration path is a small circle encircling the origin, and then being deformed to the contour $\Gamma$ illustrated in Figure \ref{figure 2}. \begin{figure}[h] \begin{center} \includegraphics[height=6cm]{figure2.eps} \caption{The deformed contour $\Gamma$: the oriented curve.} \label{figure 2} \end{center} \end{figure} From the symmetry property of $v(z)$, we need only evaluate and estimate the integral on the right-half of $\Gamma$. We split the integral in (\ref{(3.27)}) into three integrals, namely, \begin{equation}\label{(3.28)}\frac {\pi i} {\sqrt 2}\rho_k= \int_{\Gamma_v} \frac {\frac z 2}{\sin \frac z 2}\frac { v(z) dz} { z^{2k+1} } + \int_{\Gamma_l} \frac {\frac z 2}{\sin \frac z 2}\frac { v_A(z) dz} { z^{2k+1} }- \int_{\Gamma_l} \frac {\frac z \pi f(z-2\pi)\ln (2\pi-z) dz}{\sin \frac z 2 ~ z^{2k+1} }:=I_v+I_a+I_l , \end{equation} where $\Gamma_v$ is the right-half vertical part $\mathop{\rm Re}\nolimits z=3\pi$, and $\Gamma_l$ is the remaining right-half $\Gamma$, consisting of an circular part around $z=2\pi$, and a pair of horizontal line segments joining the circle with the vertical line; compare Figure \ref{figure 2} for the curves and the orientation. We will show that the main contribution to $\rho_k$, when $k$ is not small, comes from $I_l$. Estimates will be obtained with full details. We do the calculation case by case. \vskip .5cm \noindent {\bf{Estimating $I_v$}}: First, we estimate $v(z)$ for $\mathop{\rm Re}\nolimits z=3\pi$. It is readily seen that $t=\sin^2\frac {3\pi+iy} 4=\frac {1-i\sinh y} 2$ and $\sin^2\frac {\pi+iy} 4=\frac {1+i\sinh y} 2$, where $y=\mathop{\rm Im}\nolimits z$. Hence $$\arg \left\{(1-st)^{-\frac 1 2}\right \}\in (- \pi/ 2, 0)~\mbox{for}~y\in (0, \infty),~~\mbox{and}~\left | (1-st)^{-\frac 1 2}\right |\leq \left (1-\frac s 2\right )^{-\frac 1 2}~\mbox{for}~0\leq s \leq 1.$$ Accordingly, from (\ref{(3.15)}) and \eqref{f-def} we have $$ \left | f(3\pi+iy)\right | \leq \frac 1 \pi \int^1_0 s^{-\frac 1 2} (1-s)^{-\frac 1 2} \left (1-s/2\right )^{-\frac 1 2} ds:=C_{f,v}=1.1803\cdots ~~\mbox{for}~~y> 0. $$ Noting that $\arg f(3\pi +iy) \in (-\pi/2, 0)$ for $y>0$, and that $f(\pi +iy)=\overline{f(3\pi +iy)}$, Substituting all above into (\ref{(3.25)}), we have $$|v( 3\pi +iy)| \leq \sqrt 5 \left | f(3\pi+iy)\right |\leq M_v:=\sqrt 5\; C_{f,v} =2.6393\cdots ,~~y>0.$$ Similar argument shows that the equality holds for $y\in \mathbb{R}$. Hence, we conclude from (\ref{(3.25)}) and (\ref{(3.28)}) that \begin{equation}\label{(3.29)} |I_v |\leq \frac {1} 2 \int^\infty_{-\infty} \frac {|v( 3\pi+iy)|\; dy}{\cosh\frac y 2 ~ | 3\pi+iy|^{2k}}\leq \frac {M_v} {2} \frac 1 {(3\pi)^{2k}}\int^\infty_{-\infty} \frac {dy} {\cosh\frac y 2} =\frac { \pi M_v} {(3\pi)^{2k}}=\frac { 8.2916\cdots } {(3\pi)^{2k}}. \end{equation} \vskip .5cm \noindent {\bf{Evaluating $I_a$}}: This time, $v_A(z)$ is analytic in a domain containing $\Gamma_l$. It is readily seen that only the simple pole $z=2\pi$ contributes to $I_a$. More precisely, applying the residue theorem we have $$ I_a =\frac {2\pi i v_A(2\pi)} {(2\pi)^{2k}}.$$ Here $v_A(2\pi)$ can be obtained by substituting $z=2\pi-\varepsilon$ into (\ref{(3.25)}) and (\ref{(3.26)}), and taking limit $$ v_A(2\pi)=\lim_{\varepsilon\rightarrow 0^+}\left\{ \frac 1 {\pi} \int^1_0 s^{-\frac 1 2} (1-s)^{-\frac 1 2} (1-st)^{-\frac 1 2} ds + \frac 2 \pi f(-\varepsilon) \ln\varepsilon\right \},$$where $t=\sin^2\frac z 4=\cos^2\frac \varepsilon 4$; cf. (\ref{(3.15)}) and \eqref{f-def}. A careful calculation leads to \begin{equation}\label{(3.30)} v_A(2\pi) =\frac 1 \pi\int^1_0 \frac {s^{-\frac 12 }-1}{1-s } ds + \frac 1 \pi\int^1_0 \frac {s^{-\frac 1 2 }-1}{1-s } ds+\frac 4 \pi \ln 2= \frac 8 \pi \ln 2 . \end{equation} Indeed, (\ref{(3.30)}) can be derived as follows: $$\frac 1 {\pi} \int^1_0\left ( s^{-\frac 1 2} -1\right )(1-s)^{-\frac 1 2} (1-st)^{-\frac 1 2} ds \longrightarrow \frac 1 \pi\int^1_0 \frac {s^{-\frac 12 }-1}{1-s } ds ~~\mbox{as}~~t\rightarrow 1^-$$ by applying the dominated convergence theorem. Also, $$\frac 1 {\pi} \int^1_0 (1-s)^{-\frac 1 2} (1-st)^{-\frac 1 2} ds = \frac 1 { \pi} \int^1_0 \frac {\tau ^{-\frac 1 2 } }{1-t\tau } d\tau $$ by making change of variable $\tau=\frac {1-s}{1-ts}$. Finally, $$\frac 1 {\pi} \int^1_0 \left\{s ^{-\frac 1 2 }-1 \right \} \frac 1 {1-ts } ds \longrightarrow \frac 1 { \pi} \int^1_0 \frac {s^{-\frac 1 2 }-1}{1-s } ds~~\mbox{as}~~t\rightarrow 1^- $$ by applying the dominated convergence theorem one more time. For $t<1$, it further holds $$\frac 1 {\pi} \int^1_0 \frac { 1 }{1-ts } ds=-\frac 1{\pi t} \ln (1-t).$$ Adding up all these gives (\ref{(3.30)}). Thus we have \begin{equation}\label{(3.31)} I_a = \frac {16\ln 2} {(2\pi)^{2k}}i = \frac {11.0903\cdots} {(2\pi)^{2k}}i . \end{equation} \vskip .5cm \noindent {\bf{Evaluating $I_l$}}: Now we turn to the last integral $I_l$ in (\ref{(3.28)}). First, we observe that $\displaystyle{\frac {f(z-2\pi)}{\sin\frac {z-2\pi} 2}- \frac {f(0)}{ \frac {z-2\pi} 2} } $ is analytic in the strip $\mathop{\rm Re}\nolimits z\in (0, 4\pi)$. Hence the circular part of $\Gamma_l$ collapses in the following integral to give $$\frac 1 \pi \int_{\Gamma_l} \left\{ \frac {f(z-2\pi)}{\sin\frac {z-2\pi} 2}- \frac {f(0)}{ \frac {z-2\pi} 2}\right \} \frac {\ln (2\pi -z)}{z^{2k}} dz= -2i \int_{2\pi}^{3\pi} \left\{ \frac {f(z-2\pi)}{\sin\frac {z-2\pi} 2}- \frac {f(0)}{ \frac {z-2\pi} 2}\right \} \frac {dz}{z^{2k}} . $$ Accordingly we have $$\left | \frac 1 \pi \int_{\Gamma_l} \left\{ \frac {f(z-2\pi)}{\sin\frac {z-2\pi} 2}- \frac {f(0)}{ \frac {z-2\pi} 2}\right \} \frac {\ln (2\pi -z)}{z^{2k}} dz\right |\leq \frac {4\pi M_f}{2k-1} \frac 1 {(2\pi)^{2k}},$$ where $$ M_f\geq \max_{2\pi \leq z \leq 3\pi} \left |\frac {f(z)}{\sin\frac {z-2\pi} 2}- \frac {f(2\pi)}{ \frac {z-2\pi} 2}\right |.$$ We give a rough estimate for this maximum value. Recalling that $f(z)=F\left (\frac 1 2, \frac 12; 1; \sin^2\frac {z} 4\right )$ and $f(0)=1$, noting that the Maclaurin expansion of $F\left (\frac 1 2, \frac 12; 1; t \right )$ has all positive coefficients (cf. \cite[(15.2.1)]{NIST}), and that $\frac {d}{dt} \left ( \frac {F\left (\frac 1 2, \frac 12; 1; t \right )-1}{\sqrt t}\right )>0$ for $t\in (0, 1)$ via term-by-term differentiation, we see that $\frac {f(z-2\pi)-1 }{\sin\frac {z-2\pi} 2}=\frac 1 {2\cos\frac {z-2\pi} 4} \frac {F\left (\frac 1 2, \frac 12; 1; t \right )-1}{\sqrt t}$ is monotone increasing for $ 2\pi< z \leq 3 \pi$, or correspondingly, $t= \sin^2\frac {z-2\pi} 4\in (0, 1/2]$. Therefore, we have $$ 0< \frac {f(z-2\pi)-1 }{\sin\frac {z-2\pi} 2} \leq F \left (\frac 1 2, \frac 12; 1; \frac 1 2 \right ) -1 = \frac {\sqrt\pi} {\Gamma(3/4)\Gamma(3/4)} -1; $$ see \cite[(15.4.28)]{NIST}. Also, one has $$0< \frac 1 {\sin x}-\frac 1 x =\frac {x-\sin x} {x\sin x}\leq \frac {x^3}{6x\sin x} \leq \frac {\pi^2} {24}~~\mbox{for}~~0<x\leq \frac \pi 2.$$ Hence an appropriate choice of $M_f$ is $$M_f=\frac {\sqrt\pi} {\Gamma(3/4)\Gamma(3/4)} -1 + \frac {\pi^2} {24}=0.5915\cdots.$$ We proceed to evaluate the crucial part $$\frac 2 \pi \int_{\Gamma_l} \frac {\ln (2\pi-z)}{z-2\pi} \frac {dz}{z^{2k}},$$ of which the integrand has a pole $z=2\pi$, coinciding with the logarithmic singularity. To treat such kind of singularities, we appeal to the idea of Wong and Wyman \cite{WongWyman}. For simplicity we re-scale the variable $\tau=\frac {z-2\pi} {2\pi}$, which turns the integral into $$\frac 2 {\pi (2\pi)^{2k}} \int_{\frac 1 2}^{(0-)} \frac {\ln (2\pi)+\ln(-\tau)}{\tau } \frac { d\tau} { e^{2k\ln (1+\tau)}} =-\frac {4i\ln 2\pi}{(2\pi)^{2k}} + \frac 2 {\pi(2\pi)^{2k}} \int_{\frac 1 2}^{(0-)} \frac { \ln(-\tau)}{\tau } \frac {d\tau }{e^{2k\ln (1+\tau)}} ,$$ where the integration path starts and ends at $\tau=\frac 1 2$, and encircles the origin clockwise: the loop is a re-scaled version of $\Gamma_l$. The branch of $\ln (1+\tau) $ is chosen so that $\ln (1+\tau)>0 $ for $\tau >0$. To extract the main contribution, we further split the exponent in the integrand. Indeed, we see that $$ \left | \frac 2 {\pi(2\pi)^{2k}} \int_{\frac 1 2}^{(0-)} \frac {e^{2k(\tau -\ln (1+\tau)) } -1}{\tau } e^{-2k\tau} \ln(-\tau)d\tau \right | \leq \frac 2 { (2\pi)^{2k}} \max_{0\leq \tau <1/2} \left |\varphi(\tau )\right |\leq \frac {2M_\varphi} { (2\pi)^{2k}},$$ where the function $$\varphi(\tau )=\frac {e^{2k(\tau -\ln (1+\tau)) } -1}{\tau } e^{-2k\tau}$$ is analytic in a neighborhood of $[0, 1/2]$, thus the integration path collapses to the lower and upper edges of $[0, 1/2]$, and its bound $M_\varphi$ can be obtained by noticing that $$0\leq \tau -\ln(1+\tau )\leq \frac {\tau^2} 2~~\mbox{for}~~\tau >0,~~~\mbox{and}~~-\ln(1+\tau)\leq -\frac 2 3\tau ~~\mbox{for}~~ 0\leq \tau \leq \frac 1 2.$$ Now for $0\leq \tau \leq \frac 1 {\sqrt k}$, one has $$0\leq \varphi(\tau)\leq \frac {e^{k \tau^2} -1}{\tau } e^{-2k\tau} \leq (e-1) (k\tau)e^{-2(k\tau)} \leq \frac {e-1}{2e}. $$ While for $ \frac 1 {\sqrt k}\leq \tau \leq \frac 1 2$, $$0\leq \varphi(\tau)\leq \frac {e^{-2k \ln (1+\tau) }}\tau \leq \frac {e^{-\frac 4 3 k\tau}}\tau \leq \left . \frac {e^{-\frac 4 3 k\tau}}\tau \right |_{\tau=\frac 1 {\sqrt k}}=\sqrt k e^{-\frac 4 3\sqrt k} \leq \frac 3 {4e} . $$ Thus we may chose $$M_\varphi=\max\left ( \frac {e-1}{2e}, \frac 3 {4e}\right ) =\frac {e-1}{2e},$$ which does not depend on $k$. The remaining piece would turn out to be of the most significance. Indeed, a change of variable $s=2k\tau$ makes \begin{equation*}\label{(3.32)} \begin{array}{rl} \displaystyle{ \frac 2 {\pi(2\pi)^{2k}} \int_{\frac 1 2}^{(0-)} \frac {\ln (-\tau) } \tau e^{-2k\tau} d\tau }&= \displaystyle{ \frac 2 {\pi(2\pi)^{2k}} \int_{k}^{(0-)} \frac {\ln(-s)-\ln 2k} s e^{-s} ds } \\[.4cm] & =\displaystyle{ \frac {4i\ln 2k} {(2\pi)^{2k}}+ \frac 2 {\pi(2\pi)^{2k}} \int_{k}^{(0-)} \frac {\ln(-s)} s e^{-s} ds .} \end{array} \end{equation*} The last integral can be approximated by extending the integration path to $\displaystyle{\int_{+\infty}^{(0-)}}$, with an error bounded by $\displaystyle{\frac {4 e^{-k}} {k(2\pi)^{2k}}}$. Recalling the integral representation $$\frac 1 {\Gamma(z)}=\frac 1 {2\pi i} \int^{(0-)}_{+\infty} (-s)^{-z} e^{-s}ds$$ for all finite $z$; cf. \cite[(5.9.2)]{NIST} or Wong and Wyman \cite{WongWyman}, taking derivative with respect to $z$ and setting $z=1$, we obtain $$ \int^{(0-)}_{+\infty} \frac {\ln(-s)} s e^{-s} ds= \frac {2\pi i\Gamma'(1)}{\Gamma^2(1)} =-2\pi i \gamma,$$ where $\gamma=0.5772\cdots$ is Euler's constant; see \cite[(5.4.12)]{NIST}. Hence we can write \begin{equation}\label{(3.33)} I_l= \frac {4i\ln 2k} {(2\pi)^{2k}}+ \frac {(-4\gamma-4\ln2\pi)i+\delta_{l,k}} { (2\pi)^{2k}}, \end{equation} with $$|\delta_{l,k}|\leq \frac {4\pi M_f}{2k-1} + 2M_\varphi + \frac {4e^{-k}} k .$$ \vskip .3cm Now we sum up the calculations and estimations made above. Substituting (\ref{(3.29)}), (\ref{(3.31)}) and (\ref{(3.33)}) into (\ref{(3.28)}), we have the approximation \begin{equation}\label{(3.34)} \frac \pi {\sqrt 2} \rho_k= \frac {4 \ln 2k} {(2\pi)^{2k}}+ \frac {\left (11.0903\cdots -4\gamma-4\ln2\pi\right )+\delta_{k}} { (2\pi)^{2k}}, \end{equation} with \begin{equation*}\label{(3.35)} |\delta_{k}| \leq \pi M_v \left ( \frac 2 3\right )^{2k} + \frac {4\pi M_f}{2k-1} +2M_\varphi + \frac {4e^{-k}} k < 1.0259 \end{equation*} for $k\geq 10$. Hence we obtain from (\ref{(3.34)}) that \begin{equation}\label{(3.36)} \frac {4 \ln 2k + 0.4041\cdots } {(2\pi)^{2k}} \leq \frac \pi {\sqrt 2} \rho_k\leq \frac {4 \ln 2k +2.4559\cdots } {(2\pi)^{2k}} \end{equation}for $k\geq 10$. \subsection{The ratio}\label{sec:3.4} For $k\geq 10$, it is readily verified that $$ \frac 8 9\ln(2k+2) \geq 2.7475\cdots >2.4559-\frac {11} 9 \times 0.4041\cdots.$$ Therefore, it follows from (\ref{(3.36)}) that \begin{equation*}\label{(3.37)} \rho_k/\rho_{k+1}\leq \frac {44} 9 \pi^2 \end{equation*}for $k\geq 10$. It is easily verified that the inequality holds for all $k\geq 0$ by numerical evaluation of the first few $\rho_k$ for $k=0,1,\cdots 9$, as can be seen from Table \ref{table2}. \begin{table}[h] \centering \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|} \hline $k$ & $0$ & $1$ & $2$ & $3$ &$4$ & $5$ & $6$ & $7$ & $8$ & $9$ \\ $\rho_k/\rho_{k+1}$ & 17.46 & 27.41 &32.65 &35.30 &36.67 &37.41 & 37.86 & 38.15 & 38.36 & 38.51 \\[0.1cm] $\frac {44} 9\pi^2$ & 48.25 & 48.25 &48.25 &48.25 &48.25 &48.25 &48.25 &48.25 &48.25 &48.25 \\ \hline \end{tabular} \caption{The first few $\rho_k$, $k=0,2,\cdots, 9$.}\label{table2} \end{table} A by-product is that $\rho_k/\rho_{k+1} \rightarrow 4\pi^2$ as $k\rightarrow \infty$. Another byproduct of (\ref{(3.36)}) and Table \ref{table2} is that $\rho_k>0$ for all $k$, and thus $\beta_{2k}$ having alternative signs, as stated in (\ref{(2.7)}). \section { Discussion } \indent\setcounter{section} {4} \setcounter{equation} {0} \label{sec:4} We discuss very briefly a conjecture of H. Granath, of which we were not aware until we almost finish writing the present paper. In \cite{Granath}, Granath derives an asymptotic expansion \begin{equation}\label{(4.1)} \pi G_{n-1}\sim \ln(16n)+\gamma +\sum_{k=1} ^\infty \frac{a_k}{(16n)^k}, ~~n\rightarrow\infty, \end{equation} where $a_k$ are `effectively computable' constants but not explicitly given, except for the first few. The author shows interest in seeking sharp bounds of arbitrary orders. Indeed, denoting \begin{equation}\label{(4.2)} A_m(n)=\ln(16n)+\gamma +\sum_{k=1} ^m \frac{a_k}{(16n)^k} , \end{equation} Granath proves that $A_5(n) <\pi G_{n-1} < A_7(n)$ and states that $A_9(n) <\pi G_{n-1} < A_{11}(n)$, for all positive $n$. It is then conjectured that (the sign in \cite[(10)]{Granath} is apparently wrong) \begin{equation}\label{(4.3)} (-1)^{\frac {m(m+1)} 2} \left ( \pi G_{n-1}-A_m(n)\right ) <0 \end{equation}for all $n=1,2,\cdots$ and $m=0,1,2,\cdots$. The existence of (\ref{(4.1)}) has also been justified in \cite{LLXZ} and \cite{NemesNemes}. It is easily seen that in terms of $\beta_{s}$ in (\ref{(2.3)}) and (\ref{(2.4)}), the coefficients can be written as \begin{equation}\label{(4.4)} a_k=4^k\left [ -\frac 1 k +\sum^k_{s=1} \frac { (k-1)! 4^s \beta_s }{(s-1)!(k-s)!} \right ], ~~k=1,2,\cdots; \end{equation} see also \cite{LLXZ} for an iterative formula. So far, numerical experiments agree with (\ref{(4.3)}). A proof of it might be found either by following the steps, or by using the results, of the present paper. Then, a natural question arises: \noindent {\qe\label{question 2}{Considering the general expansion in (\ref{(1.11)}), for what $h$ we have the ``best'' approximation in the sense of Theorem \ref{Thm 1} and (\ref{(4.3)})? }} The case $h=3/4$ is what we have been dealt with in the present paper. Very likely the expansion (\ref{(1.11)}) for $h=1/2$ and $h=1$ would turn out to be the ``best''. As mentioned earlier, the coefficients $\beta_{2s}$ of the expansion (\ref{(1.12)}) can be obtained iteratively via (\ref{(2.4)}). We complete the paper by giving a couple of alternative representations for $\beta_{2s}$. For example, it can be shown that \begin{equation}\label{(4.5)} \beta_{2l} = \frac { (-1)^{l+1}}{ 2^{2l} (l!)^2 } \left | \begin{array}{cccccc} d_{0,2} & d_{0,3} & \cdots & d_{0,l-1} & d_{0, l} & d_{0, l+1} \\ d_{1,2} & d_{1,3} & \cdots & d_{1,l-1} & d_{1, l} & d_{1, l+1} \\ 0 & d_{2,3} & \cdots & d_{2,l-1} & d_{2, l} & d_{2, l+1} \\ \vdots & \vdots & \ddots & \vdots & \vdots & \vdots \\ 0 & 0 & \cdots &d_{l-2,l-1}&d_{l-2,l} &d_{l-2,l+1} \\ 0 & 0 & \cdots & 0 &d_{l-1,l} &d_{l-1,l+1} \\ \end{array} \right |, ~~~l=1,2,\cdots.\end{equation}Indeed, taking $k=1,2,\cdots, l$ in (\ref{(2.4)}), we have a linear algebraic system with unknowns $\beta_2$, $\beta_4$, $\cdots$, $\beta_{2l}$. Solving the system gives (\ref{(4.5)}). Also, a combination of (\ref{(2.11)}), (\ref{(3.14)}) and (\ref{(3.27)}) yields an integral representation \begin{equation}\label{(4.6)} \beta_{2l}=\frac {(-1)^{l+1} (2l-1)!} {4\pi i} \oint \frac z {\sin\frac z 2} \frac {F\left ( 1/ 4 , 1/ 4 ; 1; \sin^2\frac z 2\right ) dz}{z^{2l+1}} , ~~l=1,2,\cdots,\end{equation} where the integration path is a small circle encircling the origin counterclockwise. Of course, (\ref{(3.34)}) can be interpreted as \begin{equation}\label{(4.7)} \beta_{2l}=\frac {(-1)^{l+1} (2l-1)!} {\pi (2\pi)^{2l}}\left\{4\sqrt 2\ln(2l) +O(1) \right\} \end{equation} for large $l$. Results can be obtained from such approximations. For example, the expansion (\ref{(1.12)}) is divergent; compare \cite[Theorem 3]{Granath}. \section*{Acknowledgements} The authors are grateful to Prof. R. Wong for bringing the problem into their attention. The authors thank the anonymous referees for their carefully reading of the manuscript and for the valuable suggestions and comments. One referee suggested the using of a quadratic transformation of the hypergeometric functions which makes the proof of Lemma \ref{lem 2.2} more natural and simplified. The other referee provided many constructive suggestions and corrections which have much improved the readability of the manuscript. The work of Yutian Li was supported in part by the HKBU Strategic Development Fund, a start-up grant from Hong Kong Baptist University, and a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. HKBU 201513). The work of Saiyu Liu was supported in part by Hunan Natural Science Foundation under grant number 14JJ6030, and by the National Natural Science Foundation of China under grant number 11326082. The work of Shuaixia Xu was supported in part by the National Natural Science Foundation of China under grant number 11201493, GuangDong Natural Science Foundation under grant number S2012040007824, and the Fundamental Research Funds for the Central Universities under grand number 13lgpy41. Yuqiu Zhao was supported in part by the National Natural Science Foundation of China under grant numbers 10471154 and 10871212.

https://arxiv.org/abs/0909.1859

Simplices with equiareal faces

We study simplices with equiareal faces in the Euclidean 3-space by means of elementary geometry. We present an unexpectedly simple proof of the fact that, if such a simplex is non-degenerate, than every two of its faces are congruent. We show also that this statement is wrong for degenerate simplices and find all degenerate simplices with equiareal faces.

\section{Introduction}\label{section1} \footnotetext{The first author was supported in part by the Russian State Program for Leading Scientific Schools, Grant~NSh--8526.2008.1.} This paper deals with simplices with equiareal faces in the Euclidean 3-space. A simplex is called a \textit{simplex with equiareal faces} if all its faces have the same area and is called \textit{degenerate} if all its vertices lie on a single plane. Our primary interest is the following \textbf{Problem 1.} \textit{Prove that all faces of any non-degenerate simplex with equiareal faces in the Euclidean 3-space are congruent to each other.} Problem 1 already appeared in several contexts. We mention here some of those appearances though we cannot say that our knowledge is complete. In the 1960s Professor Hans Vogler in Vienna used Problem 1 to convince his students how powerful the synthetic geometry is. As far as we know, his desciptive-geometrical solution was never published. Independently, in the 1970's and 1980's the following form of Problem 1 was used to prevent `undesirable applicants' from joining the Moscow State University: `The faces of a triangular pyramid have the same area. Show that they are congruent.' An elementary solution to the latter problem given in \cite{Vardi} shows that indeed the problem is far from being trivial. In \cite{McMullen}, among other results P. McMullen has proved that \textit{a non-degenerate simplex is a simplex with equiareal faces if an only if its opposite pairs of edges have the same lengths.} It is easy to see that this statement is equivalent to Problem 1. McMullen's proof is very short and natural, but it is not elementary since it rests on Minkowski's theorem on uniqueness and existence of a~closed convex polyhedron with given directions and areas of faces (see, e.\,g., \cite{Alexandrov}; the direction of a~face is determined by the outward unit normal to the face). In 2007 Professor Robert Connelly in Cornell University has brought our attention to the fact that it is reasonable to study degenerate simplices with equiareal faces. He argued that when we fix three vertices, say $A$, $B$, and $C$, and move arbitrarily the fourth vertex, say $D$, nothing special happens when $D$ occurs in the plane~$ABC$: the degenerate simplex is obtained as a limit of non-degenerate ones; the notions of the vertex, edge, and face are clear for it; the notion of the face area is well defined; from combinatorial point of view there is no difference between degenerate and non-degenerate simplices. Besides, the degenerate polyhedra play very important role in `advanced' study of convex polytopes, see, e.\,g., \cite{Alexandrov}. Also Professor Robert Connelly has brought our attention to the fact that, in contrast with Problem 1, there are degenerate simplices with non-congruent equiareal faces. In fact, every parallelogram equipped with its diagonals may be treated as a degenerate simplex with equiareal faces as well as every four points on a line may be treated as a vertex-set of a degenerate simplex with equiareal faces (with all areas equal zero). The problem is whether there are some other degenerate simplices with equiareal faces. In Section~\ref{section2} we give the shortest available for us elementary solution to Problem~1. As far as we know, its main idea should be attributed to Professor Hans Vogler who is now at the University of Innsbruck. In Section~\ref{section3} we use Heron's formula to study all simplices with equiareal faces, both degenerate and non-degenerate. \section{A short elementary solution to Problem 1}\label{section2} Note that in order to solve Problem 1 it is sufficient to solve the following problem~2 which is of independent interest. \textbf{Problem~2.} \textit{Let $ABCD$ be a non-degenerate simplex, let $a(\triangle ABC){=}a(\triangle ABD)$, and let $a(\triangle ACD)=a(\triangle BCD)$, where $a(\triangle XYZ)$ stands for the area of the triangle $\triangle XYZ$. Prove that $|AC|=|BD|$ and $|BC|=|AD|$, where $|XY|$ stands for the length of the straight line segment~$XY$} (see Fig.~1). \textbf{Solution} to Problem 2. Let $P$ be the plane which is parallel to the line~$AB$ and contains the line~$CD$ (see Fig.~2). We are going to study the orthogonal projection of the simplex $ABCD$ into the plane $P$. Denote by~$X^\perp$ the image of a point $X$ under that projection and denote by~$X^\ast$ the foot of the perpendicular to the line~$AB$ emanated from the point~$X$. \begin{figure} \includegraphics[width=0.45\textwidth]{alex_weiss_fig1.eps}\hfill \includegraphics[width=0.54\textwidth]{alex_weiss_fig2.eps}\\ \hskip-65mm\parbox[t]{0.47\textwidth}{\caption{}}\hskip10mm \parbox[t]{0.47\textwidth}{\caption{}} \end{figure} Since~$AB$ is parallel to~$P$ we have~$|D^\ast D^{\ast\perp}|=|C^\ast C^{\ast\perp}|$. Since the triangles~$\triangle ABC$ and~$\triangle ABD$ have equal areas and the common side~$AB$ we conclude that~$|DD^\ast |=|CC^\ast |$. Applying Pythagoras theorem to the right-angled triangles~$\triangle DD^\ast D^{\ast\perp}$ and $\triangle CC^\ast C^{\ast\perp}$ we get~$|DD^{\ast\perp}|^2 =|DD^{\ast\perp}|^2-|D^\ast D^{\ast\perp}|^2 =|CC^{\ast\perp}|^2-|C^\ast C^{\ast\perp}|^2 =|CC^{\ast\perp}|^2$. In terms of the quadrilateral $A^\perp CB^\perp D$ this means that the vertices~$D$ and~$C$ lie at the same distance from the diagonal~$A^\perp B^\perp$. Note also that the triangles $\triangle A^\perp B^\perp D$ and $\triangle A^\perp B^\perp C$ have the same area and the points~$D$ and~$C$ lie on the different sides of the line passing through the points~$A^\perp$ and~$B^\perp$. In fact, if they lie on the same side, the line through the points~$A^\perp$ and~$B^\perp$ must be parallel to the line through the points~$D$ and~$C$ and, thus, the points~$A$, $B$, $C$, and~$D$ should be coplanar. A contradiction. Similar arguments applied to the triangles~$\triangle ACD$ and~$\triangle BCD$ show that the vertices~$A^\perp$ and~$B^\perp$ of the quadrilateral~$A^\perp CB^\perp D$ lie at the same distance from the line through the points~$D$ and~$C$ and, moreover, lie on the different sides of that line. Hence, the quadrilateral~$A^\perp CB^\perp D$ is convex. Recall that if a convex planar quadrilateral is such that every two opposite vertices lie at the same distance from the diagonal that joins the rest two vertices then the quadrilateral is a parallelogram. This implies that the quadrilateral~$A^\perp CB^\perp D$ is a parallelogram and, thus, that $|A^\perp C|=|B^\perp D|$ and $|B^\perp C|=|A^\perp D|$. Applying Pythagoras theorem to the right-angled triangles $\triangle AA^\perp C$ and $\triangle BB^\perp D$ and taking into account that $|AA^\perp|=|BB^\perp|$ we get $|AC|^2=|AA^\perp|^2+|A^\perp C|^2 =|BB^\perp|^2+|B^\perp D|^2=|BD|^2$. Similar arguments applied to the triangles $\triangle AA^\perp D$ and $\triangle BB^\perp C$ show that $|BC|=|AD|$. \hfill Q.E.D. \section{A study of degenerate and non-degenerate simplices\\ with equiareal faces}\label{section3} From Section \ref{section1} we know the following three types of equiareal non-degenerate and degenerate simplices in the Euclidean 3-space: Type 1: non-degenerate simplex with all faces congruent; Type 2: parallelogram equipped with its diagonals; and Type 3: four points on a line treated as a vertex-set of a degenerate simplex with equiareal faces of zero area. In this Section we use Heron's formula to study the following \textbf{Problem~3.} \textit{Prove that there are no simplices with equiareal faces which do not belong to Types 1--3.} \textbf{Solution} to Problem 3. Let $T$ be a simplex with equiareal faces. Let the first face (I) of $T$ has edges of the lengths $a$, $b$, and $c$; the second face (II) has edges of the lengths $a$, $y$, and $z$; the third face (III) --- $b$, $x$, and $z$; and the forth face (IV) --- $c$, $x$, and $y$. (Equivalently, we can say that side $x$ is opposite to $a$; side $y$ --- to $b$; and $z$ --- to $c$). Let $S$ be common area of the faces of $T$. Heron's formula for the face (I) yields \begin{align} (4S)^2 &=(a+b+c)(-a+b+c)(a-b+c)(a+b-c)\notag\\ &= 2a^2b^2+2a^2c^2+2b^2c^2-a^4-b^4-c^4=-(a^2-b^2+c^2)^2+4a^2c^2. \end{align} Now let's use Heron's formula to express the fact that the faces (I) and (II) have the same areas \begin{align} (a+y+z)(-& a+y+z)(a-y+z)(a+y-z)-(4S)^2\notag\\ =& 2a^2y^2+2a^2z^2+2z^2y^2-a^4-y^4-z^4-(4S)^2\\ =&-(z^2-y^2-a^2)^2-(4S)^2+4y^2a^2=0.\notag \end{align} Solving this equation with respect to $z^2$ yields \begin{equation} z^2=y^2+a^2\pm\sqrt{4a^2y^2-(4S)^2}. \end{equation} Similarly, we use Heron's formula to express the fact that the faces (I) and (III) have the same areas \begin{align} (b+z+x)(-& b+z+x)(b-z+x)(b+z-x)-(4S)^2\notag\\ =& 2b^2z^2+2b^2x^2+2z^2x^2-b^4-z^4-x^4-(4S)^2\notag\\ =& -(z^2-x^2-b^2)^2-(4S)^2+4b^2x^2=0.\notag \end{align} Solving this equation with respect to $z^2$ yields \begin{equation} z^2=x^2+b^2\pm\sqrt{4b^2x^2-(4S)^2}. \end{equation} At last, we use Heron's formula to express the fact that the faces (I) and (IV) have the same areas \begin{align} (c+x+y)(-& c+x+y)(c-x+y)(c+x-y)-(4S)^2\notag\\ =& 2c^2x^2+2c^2y^2+2x^2y^2-c^4-x^4-y^4-(4S)^2\notag\\ =& -(y^2-x^2-c^2)^2-(4S)^2+4c^2x^2=0.\notag \end{align} Solving this equation with respect to $y^2$ \begin{equation} y^2=x^2+c^2\pm\sqrt{4c^2x^2-(4S)^2}. \end{equation} Eliminate $z^2$ from (3) and (4) $$ y^2+a^2\pm\sqrt{4a^2y^2-(4S)^2}=x^2+b^2\pm\sqrt{4b^2x^2-(4S)^2}, $$ then twice square this equation in order to eliminate square roots and use the formula $(a^2-b^2+c^2)^2=4a^2c^2-(4S)^2$ (which is a consequence of (1)) to obtain \begin{gather} 4a^4y^4+4b^4x^4+(x^2+b^2-y^2-a^2)^4-8a^2b^2x^2y^2-2a^2y^2(x^2+b^2-y^2-a^2)^2\notag\\ -2b^2x^2(x^2-y^2-a^2+b^2)^2=-4(x^2-y^2-a^2+b^2)^2. \end{gather} So, we arrive at the most computationally difficult, but still straightforward point of the solution: substitute $y^2$ in (6) by the right-hand side of (5). After simplifications and multiple usage of the formula $(a^2-b^2+c^2)^2=4a^2c^2-(4S)^2$ we get \begin{equation} 4(c^2x^2-4S^2)(x^2-a^2)^2 S^4= \bigl[x^4-x^2(a^2+b^2-c^2)-a^2(b^2-c^2)^2\bigr]S^4. \end{equation} We see that $S=0$ is a root of (7) which corresponds to simplices of Type 3. In order to find the other roots, cancel $S^4$ in (7), rearrange terms and use the formula $(a^2-b^2+c^2)^2=4a^2c^2-(4S)^2$ again to arrive at $$ (x^2-a^2)^2\bigl[x^4-2x^2(b^2+c^2)+a^2(2b^2+2c^2-a^2)\bigr]=0. $$ Note that the bi-quadratic expression in the brackets has two roots: $x^2=a^2$ and $x^2=2b^2+2c^2-a^2$. Note also that we can obtain similar equations for $y$ and $z$ just by permuting three pairs of symbols $(a,x)$, $(b,y)$, and $(z,c)$. As a result we find 8 solutions for $x, y,$ and $z$ that are accumulated as rows in the following table: \vskip10pt \begin{center} \renewcommand{\arraystretch}{1.5} \begin{tabular}{|c|c|c|c|} \hline & $x$ & $y$ & $z$ \\ \hline Solution 1 & $a$ & $b$ & $c$ \\ \hline Solution 2 & $a$ & $b$ & ${\sqrt{2a^2+2b^2-c^2}}^{}$ \\ \hline Solution 3 & $a$ & $\sqrt{2a^2-b^2+2c^2}$ & $c$ \\ \hline Solution 4 & $a$ & $\sqrt{2a^2-b^2+2c^2}$ & $\sqrt{2a^2+2b^2-c^2}$ \\ \hline Solution 5 & $\sqrt{2b^2+2c^2-a^2}$ & $b$ & $c$ \\ \hline Solution 6 & $\sqrt{2b^2+2c^2-a^2}$ & $b$ & $\sqrt{2a^2+2b^2-c^2}$ \\ \hline Solution 7 & $\sqrt{2b^2+2c^2-a^2}$ & $\sqrt{2a^2-b^2+2c^2}$ & $c$ \\ \hline Solution 8 & $\sqrt{2b^2+2c^2-a^2}$ & $\sqrt{2a^2-b^2+2c^2}$ & $\sqrt{2a^2+2b^2-c^2}$ \\ \hline \end{tabular} \end{center} \vskip10pt Solution 1, obviously, corresponds to the simplices of Type 1. Solutions 2, 3, and 5 correspond to simplices of Type 2. For example, a simplex $T$, corresponding to Solution 2, is degenerated into a parallelogram with the side lengths $a$ and $b$ and the diagonals $c$ and $\sqrt{2a^2+2b^2-c^2}$ (recall that in any parallelogram the sum of the squared lengths of all sides equals the sum of the squared lengths of the both diagonals). Solutions 4, 6, and 7 correspond to simplices of Type 2 again, but the face (I), with edge lengths $a$, $b$, and $c$, must be a right-angled triangle this time. For example, consider Solution 4. We have \begin{equation} x^2=a^2,\quad y^2=2a^2-b^2+2c^2, \quad\mbox{and}\quad z^2=2a^2+2b^2-c^2. \end{equation} Using the formula (2) we get $-(z^2-y^2-a^2)^2-(4S)^2+4a^2y^2=0$. Now we use the formula (1) and, after some simplifications, we get $a^4-(b^2-c^2)^2=0$. Without loss of generality, we may assume that $b\geqslant c$. This yields $a^2+c^2=b^2$ and, thus, the face (I) is a right-angled triangle. Moreover, now the formula (8) implies that $y^2= b^2$ and $z^2= 4a^2+c^2$. Hence, the simplex $T$, corresponding to Solution 4, is degenerated to a parallelogram with the side lengths $a$, $b$, $x=a$, and $y=b$ and the diagonals of the lengths $c$ and $z=\sqrt{4a^2+c^2}=\sqrt{2a^2+2b^2-c^2}$. Hence, the simplex $T$ is of Type 2. Solutions 6 and 7 are treated similarly. Solution 8 does not correspond to any simplex in the Euclidean 3-space (neither degenerated nor non-degenerated). In fact, we have \begin{equation} x^2=2b^2+2c^2-a^2, \quad y^2=2a^2-b^2+2c^2, \quad\mbox{and}\quad z^2=2a^2+2b^2-c^2. \end{equation} Using the formula (2) we get $-(z^2-y^2-a^2)^2-(4S)^2+4a^2y^2=0$. Now we use formula (1) and, after some simplifications, we get $a^4-(b^2-c^2)^2=0$ or \begin{equation} (a^2-b^2+c^2)(a^2+b^2-c^2)=0. \end{equation} The geometric meaning of the formula (10) is that either $b$ or $c$ is the hypotenuse of the right-angled triangle (I) with the sides $a$, $b$, and $c$. Similarly we can substitute (9) into formulas (3) and (4). Proceeding as above we get \begin{gather} (a^2+b^2-c^2)(-a^2+b^2+c^2)=0,\\ (a^2-b^2+c^2)(-a^2+b^2+c^2)=0. \end{gather} The geometric meaning of the formula (11) is that either $c$ or $a$ is the hypotenuse of the right-angled triangle (I) with the sides $a$, $b$, and $c$. Similarly, the formula (12) implies that either $b$ or $a$ is the hypotenuse of the right-angled triangle (I) with the sides $a$, $b$, and $c$. But the triangle (I) has only one hypotenuse! Hence the equations (10)--(12) can not hold true simultaneously. This means that Solution 8 does not correspond to any simplex. Now we can conclude that there is no simplices with equiareal faces which do not belong to Types 1--3. \hfill{Q.E.D.}

https://arxiv.org/abs/1303.4850

Regular graphs of odd degree are antimagic

An antimagic labeling of a graph $G$ with $m$ edges is a bijection from $E(G)$ to $\{1,2,\ldots,m\}$ such that for all vertices $u$ and $v$, the sum of labels on edges incident to $u$ differs from that for edges incident to $v$. Hartsfield and Ringel conjectured that every connected graph other than the single edge $K_2$ has an antimagic labeling. We prove this conjecture for regular graphs of odd degree.

\section{Introduction} A \emph{magic square} of order $n$ is a $n\times n$ arrangement of the integers $\{1,2,\ldots, n^2\}$ so that the sums of the entries in each row, each column, and along the two main diagonals are equal. These squares were known to the Chinese as early as the fourth century B.C.\ and have been widely studied in recreational mathematics~\cite{Gar88}. A \emph{labeling} of a graph $G$ with $m$ edges is a bijection from $E(G)$ to $\{1,2,\ldots,m\}$. Given a labeling of a graph, the \emph{vertex sum} at a vertex $v$ is the sum of the labels on edges incident to $v$. A labeling is \emph{magic} if all vertex sums are equal. Magic labelings take their name from their connection with magic squares, since a magic square of order $n$ naturally gives rise to a magic labeling of the complete bipartite graph $K_{n,n}$ (vertices in one part correspond to rows of the square, and vertices in the other correspond to columns). Finally, a labeling of a graph is \emph{antimagic} if all its vertex sums are different. We call a graph antimagic (magic) if it has an antimagic (magic) labeling. It is easy to find many graphs that are not magic (for example, forests). However, graphs that are not antimagic are rare. In fact, Hartsfield and Ringel conjectured the following. \begin{conj}[\cite{HR90}] Every connected graph other than $K_2$ is antimagic. \end{conj} Hartsfield and Ringel also explicitly conjectured that all trees other than $K_2$ are antimagic. Both conjectures remain wide open; however, much progress has been made. The first major result on antimagic labelings was due to Alon, Kaplan, Lev, Roditty, and Yuster~\cite{AKLRY04}. They showed that there exists a constant $c$ such that if $G$ is an $n$-vertex graph with minimum degree $\delta \ge c\log n$, then $G$ is antimagic. This proof relies on a combination of combinatorial ideas, probabilistic tools, and methods from analytic number theory. They also proved that graphs with maximum degree $\Delta\ge n-2$ are antimagic. Yilma~\cite{Yil13+} later extended this result to show that graphs with $\Delta\ge n-3$ are antimagic. His proof finds a breadth-first spanning tree $T$ rooted at a vertex of maximum degree; he labels all edges outside of $T$ first, then uses the largest $n-1$ labels on $T$ to guarantee an antimagic labeling. Hefetz~\cite{Hef05} used algebraic tools to show that a graph is antimagic if it has $3^k$ vertices and a $C_3$-factor. Hefetz, Saluz, and Tran~\cite{HST10} generalized this approach to show that a graph is antimagic if it has $p^k$ vertices and a $C_p$-factor (where $p$ is an odd prime). Cranston~\cite{Cra09} used Hall's marriage theorem to show that regular bipartite graphs are antimagic. Liang and Zhu~\cite{LZ13+} labeled edges in order of decreasing distance from a central vertex (breaking ties carefully) to show that 3-regular graphs are antimagic. Perhaps the most interesting result is that of Eccles~\cite{Ecc13+}, who recently improved on the work of Alon et.al. He showed that if a graph has no isolated edges or vertices and has average degree at least 4468, then it is antimagic. He conjectures that, under the same first condition, average degree at least $\sqrt{2}$ implies that a graph is antimagic. This much stronger conjecture immediately implies Conjecture~1, since a connected $n$-vertex graph has at least $n-1$ edges, and so for $n\ge 4$ has average degree at least $2(n-1)/n=2-2/n>\sqrt{2}$. In this note, we prove that every $k$-regular graph with $k$ odd and $k\ge 3$ is antimagic. \section{Main Result} A \emph{trail} is a walk in $G$ that may reuse vertices but may not reuse edges; a trail is \emph{open} if it starts and ends at distinct vertices, and is \emph{even (odd)} if its length is even (odd). For a set of vertices $U$ and a function $\sigma$ we write $\sigma(U)$ to denote $\{\sigma(u):u\in U\}$. For a subgraph or trail $H$, we write $d_H(v)$ for the degree of $v$ in $H$. We begin with an easy decomposition result for bipartite graphs. \begin{keylemma} \label{lemma1} Let $G$ be a bipartite graph with parts $U$ and $W$. There exists a function $\sigma: U\to E(G)$ and a set $\mathcal{T}=\{T_1,T_2,\ldots\}$ such that $\sigma(u)$ is incident to $u$ for all $u\in U$ and $\mathcal{T}$ is a collection of edge-disjoint open trails with at most one trail ending at each vertex and with $(\bigcup_{T\in\mathcal{T}}E(T))\cap \sigma(U)=\emptyset$ and $\bigcup_{T\in\mathcal{T}}E(T)\cup \sigma(U) = E(G)$. In other words, we can partition $E(G)$ into $\mathcal{T}$ and $\sigma(U)$. \end{keylemma} \begin{proof} We first choose $\sigma(U)$ arbitrarily, and let $\widehat{E}=E(G)\setminus \sigma(U)$. We form a greedy trail decomposition of $\widehat{E}$ as follows. Start at an arbitrary vertex and keep walking (using unused edges of $\widehat{E}$) as long as possible. When you reach a vertex with no unused edges, start a trail at another vertex. Repeat this process until all edges are used up. This gives a decomposition $\mathcal{T}$ of $\widehat{E}$, but it might contain a closed trail. Suppose that $\mathcal{T}$ contains a closed trail $T_1$. If any vertex $v$ of $T_1$ has an open trail $T_2$ that ends at $v$, then we splice $T_1$ and $T_2$ together, by starting at $v$, following all the edges of $T_1$, then following the edges of $T_2$. If no vertex of $T_1$ is the endpoint of an open trail in $\mathcal{T}$, then choose $u\in U\cap V(T_1)$ arbitrarily. Let $w$ be a successor of $u$ on $T_1$ and let $v$ be such that $\sigma(u)=uv$. We redefine $\sigma(u):=uw$, and redefine $T_1:=T_1-uw+uv$. Now $T_1$ is an open trail, since $d_{T_1}(w)$ is odd. By repeating this process for each closed trail in $\mathcal{T}$, we reach a collection of open trails. If any vertex $v$ is the endpoint of at least two open trails, then we merge them together, by walking along one to end at $v$, then walking along another starting from $v$. Merging two trails reduces the number of trails, and ``opening up'' a closed trail (as described above), does not increase this number. So iterating these merging and opening up steps gives the desired partition of $E(G)$ into $\mathcal{T}$ and $\sigma(U)$. \end{proof} Now we prove our main result. Our proof builds heavily on that of Liang and Zhu~\cite{LZ13+}, who showed that 3-regular graphs are antimagic. \begin{theoremA} Every $k$-regular graph with $k$ odd and $k\ge 3$ is antimagic. \end{theoremA} \begin{proof} Suppose that $G$ and $H$ are both antimagic $k$-regular graphs and that $|E(G)|=m$. Given antimagic labelings for $G$ and $H$, we get an antimagic labeling of $G\cup H$ by increasing the label on each edge of $H$ by $m$. Thus, we need only consider connected graphs. Choose an arbitrary vertex $v^*$ and let $V_i$ denote the set of vertices at distance exactly $i$ from $v^*$; let $p$ be the furthest distance of a vertex from $v^*$. Let $G[V_i]$ denote the subgraph induced by $V_i$ and $G[V_i,V_{i-1}]$ denote the induced bipartite subgraph with parts $V_i$ and $V_{i-1}$. For each $i$, we apply the \hyperref[lemma1]{Helpful Lemma} to $G[V_i,V_{i-1}]$ with $U=V_i$ and $W=V_{i-1}$ to get a partition of $E(G[V_i,V_{i-1}])$ into an edge set $\sigma(V_i)$ and a collection of edge-disjoint open trails. Let $G_{\sigma}[V_i,V_{i-1}] = G[V_i,V_{i-1}]\setminus \sigma(V_i)$. Let $E_i=E(G[V_i])$, let $E'_i=E(G_{\sigma}[V_i,V_{i-1}])$, and let $E''_i=\sigma(V_i)$; note that $E'_i$ and $E''_i$ partition $E(G[V_i,V_{i-1}])$. Given a labeling $f$ of the edges, we denote the total sum of labels on edges incident to vertex $v$ by $t(v)=\sum_{e\in E(v)}f(e)$, where $E(v)$ denotes the set of edges incident to $v$. Similarly, we denote the partial sum at $v$ (omitting the label on $\sigma(v)$) by $p(v)=\sum_{e\in E(v)\setminus\{\sigma(v)\}}f(e)=t(v)-f(\sigma(v))$. We now outline the proof. We will label the edges in the order $E_p, E'_p, E''_p, \ldots, E_1, E'_1, E''_1$, using the smallest unused labels on each edge set when we come to it. In other words, we use the $|E_p|$ smallest labels on $E_p$, the $|E'_p|$ next smallest labels on $E'_p$, the $|E''_p|$ next smallest labels after that on $E''_p$, etc. (Note that the labels assigned to each of these edge sets span an interval.) This label assignment immediately gives that if $i\ge j+2$ and $u\in V_i$ and $w\in V_j$, then $t(u)<t(w)$ since $G$ is regular and the edges incident to $u$ have smaller labels than the edges incident to $w$. Thus, we need only ensure that $t(u)\ne t(w)$ when either (i) $u,w\in V_i$ or (ii) $u\in V_i$ and $w\in V_{i-1}$. We handle these two cases by specifying more precisely how to assign the label to each edge of these $3p$ edge sets. We label the edges of each $E_i$ arbitrarily from its assigned labels. We now specify how to label each $E''_i$; in the process, we handle Case (i). Suppose that for some $i$, we have already labeled the edges of $E_p, E'_p, E''_p, \ldots, E_i, E'_i$. As a result, $p(u)$ is already determined for each $u\in V_i$. We may name the vertices of $V_i$ as $u_1, u_2, u_3, \ldots$ so that $p(u_1)\le p(u_2) \le p(u_3) \le \cdots$. Now we use the smallest label for $E''_i$ on $\sigma(u_1)$, the next smallest on $\sigma(u_2)$, etc. This ensures that $t(u_j)<t(u_{j+1})$ for all $u_j\in V_i$. Finally, we specify how to label each $E'_i$; in the process, we handle Case (ii). That is, we ensure that if $u\in V_i$ and $w\in V_{i-1}$, then $t(u)\net(w)$. Let $\{s,s+1,\ldots,\ell-1,\ell\}$ be the set of labels to be used on $E'_i$. Recall that $G$ is $k$-regular for odd $k\ge 3$, and let $t=(k-1)/2$. We will ensure that $p(u)\le t(s+\ell)$ and that $p(w)\ge t(s+\ell)$. Now since $f(\sigma(u))<f(\sigma(w))$, we get that $t(u)<t(w)$. The details follow. Let $\mathcal{T}$ be the set of open trails partitioning $E'_i$ (from the \hyperref[lemma1]{Helpful Lemma}). Again, let $\{s,s+1,\ldots,\ell-1,\ell\}$ be the labels assigned to $E'_i$. We label each trail so that every pair of successive labels (on a trail) incident to a vertex $u\in V_i$ has sum at most $s+\ell$ and each pair of successive labels incident to a vertex $w\in V_{i-1}$ has sum at least $s+\ell$. This ensures that $p(u)\le t(s+\ell)$ and $p(w)\ge t(s+\ell)$. We first label each even trail, then label the odd trails, taken together in pairs (possibly with a single odd trail last). Suppose that we have already labeled some even number $2r$ of edges in the set $E'_i$ and the remaining labels available for this edge set are $\{s+r, s+r+1,\ldots, \ell-r-1,\ell-r\}$. We have three possibilities. (1) Suppose first that $T\in \mathcal{T}$ is an even trail with both endpoints in $V_{i-1}$. We assign the labels: $s+r, \ell-r,s+r+1,\ell-r-1,\ldots$ successively along the trail. Now every two successive edges incident to $u\in V_i$ have sum $s+\ell$ and every two successive edges incident to $w\in V_{i-1}$ have sum $s+\ell+1$. (2) Suppose instead that $T\in \mathcal{T}$ is an even trail with both endpoints in $V_i$. Now we assign the labels: $\ell-r, s+r, \ell-r-1, s+r+1, \ldots$ successively along the trail. Now every two successive edges incident to $u\in V_i$ have sum $s+\ell-1$ and every two successive edges incident to $w\in V_{i-1}$ have sum $s+\ell$. (3) Finally, suppose that $T_1, T_2\in \mathcal{T}$ are odd trails with lengths $2a+1$ and $2b+1$. Beginning at a vertex in $V_i$, we label the edges of $T_1$ with $\ell-r, s+r, \ell-r-1, \ldots, s+r+a-1, \ell-r-a$. Here the successive pairs of labels incident to $u\in V_i$ sum to $s+\ell-1$ and the pairs incident to $w\in V_{i-1}$ sum to $s+\ell$. Finally, beginning at a vertex in $V_{i-1}$, we label the edges of $T_2$ with $s+r+a, \ell-r-a-1, s+r+a+1, \ldots, s+r+a+b$. Again the successive pairs incident to $u\in V_i$ sum to $s+\ell-1$ and the successive pairs incident to $w\in V_{i-1}$ sum to $s+\ell$. If we have a single odd trail left at the end, we treat it like a trail of length $2a+1$ above. All that remains is to verify that for $u\in V_i$ and $w\in V_{i-1}$ we have $p(u)\le t(s+\ell)$ and $p(w)\ge t(s+\ell)$. We consider $p(u)$, and the analysis for $p(w)$ is nearly identical. Recall that $d_G(u)=k=2t+1$. If $d_{E'_i}(u)=2t$, then the desired inequality holds, since each of the $t$ pairs of successive edges on trails through $u$ have label sum at most $s+\ell$. If $u$ is the end of some trail $T$ in $\mathcal{T}$, then let $e$ be the final edge of $T$ incident to $u$; note that $f(e)\le \ell$. But now, we have $d_{E'_i}(u)$ is odd, so $u$ has some incident edge (in fact, an odd number of them) in $E'_{i+1}\cup E''_{i+1}\cup E_i$; this edge has label less than $s$. Thus, the sum of this label and $f(e)$ is less than $s+\ell$. If $u$ has additional incident edges in $E'_{i+1}\cup E''_{i+1}\cup E_i$, then each edge has label less than $s$; thus, each pair of these edges has label sum less than $s+\ell$. So $p(u)\le t(s+\ell)$, as desired. For each $w\in V_{i-1}$, the analysis to show that $p(w)\ge t(s+\ell)$ is nearly identical to that above; the only difference is that all edges incident to $w$ that are not in $E'_i$ are in $E''_i\cup E_{i-1}\cup E'_{i-1}$, so each such edge has label larger than $\ell$. This completes the proof. \end{proof} We remark in closing that the proof easily translates to an efficient (polynomial time) algorithm to find an antimagic labeling. We thank Mike Barrus for his careful reading of this manuscript and detailed feedback.

https://arxiv.org/abs/1902.09146

Higher order Jacobians, Hessians and Milnor algebras

We introduce and study higher order Jacobian ideals, higher order and mixed Hessians, higher order polar maps, and higher order Milnor algebras associated to a reduced projective hypersurface. We relate these higher order objects to some standard graded Artinian Gorenstein algebras, and we study the corresponding Hilbert functions and Lefschetz properties.

\section{Introduction} In Algebraic Geometry and Commutative Algebra, the {\it Jacobian ideal} of a homogeneous reduced form $f \in R=\mathbb{C}[x_0,\ldots,x_n]$, denoted by $J(f )= (\frac{\partial f}{\partial x_0}, \ldots, \frac{\partial f}{\partial x_0})$, plays several key roles. Let $X=V(f) \subset \mathbb{P}^n$ be the associated hypersurface in the projective space. The linear system associated to the Jacobian ideal defines the {\it polar map} $\varphi_X:\mathbb{P}^n \dashrightarrow \mathbb{P}^n$, also called the {\it gradient map}, whose image is the {\it polar image} of $X$, denoted by $Z_X = \overline{\varphi(\mathbb{P}^n)}$. The restriction of the polar map to the hypersurface is the {\it Gauss map} of $X$, $\mathcal{G}_X=\varphi_X|X$, whose image is the {\it dual variety} of $X$. The base locus of these maps is the singular scheme of the hypersurface $X$, see \cite{Ru}. The {\it Milnor algebra} of $f$, also called the {\it Jacobian ring} of $f$, is the quotient $M(f )= R/J(f)$. This graded algebra is closely related to the Hodge filtration on the cohomology of $X$ and the period map, see \cite{DSa,Gr,Se}. {\it The aim of this paper is to construct higher order versions of these classical objects, explicit some relations among them and extend some classical results to this higher order context.} In the second section we give some definitions of Artinian Gorenstein algebras, Hessians, Lefschetz properties, Jacobian ideals and Milnor algebras, and review some known results. The {\it standard Artinian Gorenstein algebra} $A(f)$, associated to $f$, is given by Macaulay Matlis duality: the ring $Q = \mathbb{C}[X_0,\ldots,X_n]$ acts on $R$ via the identification $X_i =\frac{\partial }{\partial x_i}$, and we define $A(f )= Q/\operatorname{Ann}(f)$, see \cite{MW}. Since the Jacobian matrix associated to the polar map is the {\it Hessian matrix} of $f$, see \cite[Chapter 7]{Ru}, one gets that $\varphi_X$ is a dominant map, that is $Z _X = \mathbb{P}^n$, if and only if the {\it Hessian determinant} $\operatorname{hess}_f \ne 0$. A description of the forms $f$ with $\operatorname{hess}_f = 0$ is given by the Gordan-Noether criterion, and can be found, for example, in \cite{CRS, GR, Go, GRu, Ru}. The new results in the second sections are Proposition \ref{propSLP1} and Proposition \ref{propSLP2}, dealing with the Lefschetz properties of smooth cubic surfaces in $\mathbb{P}^3$ and smooth quartic curves in $\mathbb{P}^2$. In the third section we introduced the higher order Jacobian ideals $J^k(f) $ and the corresponding Milnor algebras $M^k(f)$. The Gauss map $\mathcal{G}$ of a smooth hypersurface is a birational morphism, see for instance \cite{GH,Z}. The natural $k$-th order version of smoothness is the hypothesis that all the points of $X$ have multiplicity at most $k$. In Theorem \ref{T1} we prove that this condition is equivalent to the $k$-th order Milnor algebra $M^k(f)$ being Artinian, generalizing a classical result for non singular hypersurfaces, and a second order result that can be found in \cite{DSt1}. We also discuss when the Hessian of the form $f$ belongs to the Jacobian ideal $J(f)$, see Proposition \ref{P1.5} and Question \ref{questionHESS}. In the fourth section we first show that the $k$-th order Milnor algebra $M^k(f)$ determines the hypersurface $V(f)$ up-to projective equivalence, for a generic $f$ and any $k\leq d/2-1$, see Theorem \ref{thmZW2}. In the rather long Example \ref{ex3}, we look at quartic curves in $\mathbb{P}^2$, both smooth and singular, and we compute the Hilbert functions for our graded algebras $M^k(f)$ and $A(f)$ as well as the minimal resolutions as a graded $Q$-module for $A(f)$ in many cases. In the fifth section, we construct the $k$-th polar map $\varphi^k_X$ and we prove, in Theorem \ref{thm:polarrankhess}, a higher order version of the Gordan-Noether criterion for the degeneracy of this $k$-th polar map. In this setting, we use the mixed Hessians developed in \cite{GZ}, generalizing higher order Hessians introduced in \cite{MW}. In Corollaries \ref{cor:polar1} and \ref{corHess10}, we give sufficient conditions for the non degeneracy of $\varphi^k_X$ and Theorem \ref{T2} give also some information about the degree of the $k$-th polar map. In \cite{D}, the author showed that the natural higher order related dual map, $\psi^k_X = \varphi^k_X|X$, is a finite map. In Theorem \ref{T2}, we assume that the $k$-th Milnor algebra is Artinian to prove that $\varphi^k_X$ is finite. \bigskip We would like to thank the referee for his very careful reading of our manuscript and for his suggestions which greatly improved the presentation of our results. \section{Preliminaries} \subsection{Artinian Gorenstein algebras and mixed Hessians} In this section we give a brief account of Artinian Gorenstein algebras and Macaulay-Matlis duality. \begin{defin}\rm Let $R = \mathbb{C}[x_0,\ldots,x_n] $ be a polynomial ring with the usual grading and $I \subset R$ be a homogeneous Artinian ideal and suppose, without loss of generality that $I_1=0$. Then the graded Artinian $\mathbb{C}$-algebra $A=R/I = \displaystyle\bigoplus_{i=0}^dA_i$ is standard, i.e. it is generated in degree $1$ as an algebra. Since $A$ is Artinian, under the hypothesis $I_1=0$, we call $n+1$ the codimension of $A$, by abuse of notation. Setting $h_i(A)=\dim_\mathbb{C} A_i$, the \emph{Hilbert vector} of $A$ is $\operatorname{Hilb}(A)=(1,h_1(A),\dots,h_d(A))$. The Hilbert vector is sometimes conveniently expressed as the \emph{Hilbert function} of $A$, given by the formula \begin{equation} \label{HF} H(A,t)=\sum_{k=0}^dh_k(A)t^k. \end{equation} \end{defin} The Hilbert vector $\operatorname{Hilb}(A)$ is said to be \emph{unimodal} if there exists an integer $t\ge 1$ such that $1\le h_1(A) \le \dots\le h_t(A)\ge h_{t+1}(A) \ge\dots\ge h_d(A).$ Moreover the Hilbert vector $\operatorname{Hilb}(A)=(1,h_1(A),\dots,h_d(A))$ is said to be \emph{symmetric} if $h_{d-i}(A)=h_i(A)$ for every $i=0,1,\dots,\lfloor\frac{d}{2}\rfloor$. The next Definition is based in a well known equivalence that can be found in \cite[Prop. 2.1]{MW}. \begin{defin}\rm A standard graded Artinian algebra $A$ as above is Gorenstein if and only if $h_d(A)= 1$ and the restriction of the multiplication of the algebra in complementary degree, that is $A_k \times A_{d-k} \to A_d$, is a perfect paring for $k =0,1,\ldots,d$, see \cite{MW}. If $A_j=0$ for $j >d$, then $d$ is called the \emph{socle degree} of $A$. \end{defin} It follows that the Hilbert vector $\operatorname{Hilb}(A)$ of a graded Artinian Gorenstein $\mathbb{C}-$algebra $A$ is symmetric. The converse is not true, and $\operatorname{Hilb}(A)$ is not always unimodal for $A$ Artinian Gorenstein. \begin{ex}\rm The first example of a non unimodal Hilbert vector $\operatorname{Hilb}(A)$ of a Gorenstein algebra $A$ was given by Stanley in \cite{St}, namely $$(1,13,12,13,1).$$ This algebra $A$ has codimension $13$ and socle degree $4$. In \cite{BI} we can find the first known example of a non unimodal Gorenstein Hilbert function in codimension $5$, namely $$(1,5, 12 , 22 , 35 , 51 , 70, 91 , 90 , 91 , 70 , 51 , 35 , 22 , 12 , 5 , 1).$$ All Gorenstein $h$-vectors are unimodal in codimension $\leq 3$, see \cite{St}. To the best of the authors knowledge, it is not known if there is a non unimodal Hilbert vector of a Gorenstein algebra in codimension $4$, see \cite{MN1}. \end{ex} Since our approach is algebro-geometric-differential, we recall a differentiable version of the Macaulay-Matlis duality which is equivalent to polarity in characteristic zero. We denote by $R_d=\mathbb{C}[x_0,\ldots,x_n]_d$ the $\mathbb{C}-$vector space of homogeneous polynomials of degree $d$. We denote by $Q=\mathbb{C}[X_0,\ldots,X_n]$ the ring of differential operators of $R$, where $X_i := \frac{\partial}{\partial x_i}$ for $i=0,\ldots,n.$ We denote by $Q_k=\mathbb{C}[X_0,\ldots,X_n]_k$ the $\mathbb{C}-$vector space of homogeneous differential operators of $R$ of degree $k$.\\ For each integer $k$, with $d\geq k\geq 0$ there exist natural $\mathbb{C}-$bilinear maps $R_d\times Q_k \to R_{d-k}$ defined by differentiation: $$(f,\alpha) \to f_\alpha := \alpha(f).$$ Let $f\in R$ be a homogeneous polynomial of degree $\deg f=d\geq 1$, we define the \emph{annihilator ideal of $f$} by $$\operatorname{Ann} (f) :=\left\{\alpha\in Q | \alpha(f)=0\right\}\subset Q.$$ Note that $\operatorname{Ann}(f)_1 \ne 0$ if and only if $X = V(f) \subset \mathbb{P}^n$ is a cone, that is, up-to a linear change of coordinates, the polynomial $f$ depends only of $x_1, \ldots,x_n$. {\it We assume from now on that $V(f)$ is not a cone, and hence that $\operatorname{Ann}(f)_1= 0.$} Since $\operatorname{Ann}(f)$ is a homogeneous ideal of $Q$, we can define $$A(f)=\frac{Q}{\operatorname{Ann}(f)}.$$ Then $A(f)$ is the standard graded Artinian Gorenstein $\mathbb{C}$-algebra associated to $f$, given by the Macaulay-Matlis duality, and it satisfies $$\begin{cases} A(f)_j=0 \mbox{ for } j >d\\ A(f)_d=\mathbb{C} \end{cases}.$$ A proof of this result can be found in \cite[Theorem 2.1]{MW}. \begin{ex} \label{exFer}\rm Take $f_F=x_0^d+...+x_n^d$, the Fermat type polynomial of degree $d$. In this case the ideal $\operatorname{Ann}(f)$ is generated by $X_iX_j$ for $0 \leq i <j \leq n$ and by $X_0^d-X_j^d$ for $j=1,2,...,n$. The graded part of degree $k$ of $A$ is $A_k = \langle X_0^k,X_1^k,\ldots,X_n^k\rangle$ for $k=1,\ldots,d-1$, and $A_j=\langle x_0^j\rangle $ for $j=0$ and $j=d$. This determines the Hilbert vector $$\operatorname{Hilb}(A(f_F))=(1,n+1, \ldots,n+1,1).$$ \end{ex} \begin{defin} Let $A=\displaystyle{\oplus_{i=0}^d}A_i$ be an Artinian graded $\mathbb{C}-$algebra with $A_d\neq 0$. \begin{enumerate} \item The algebra $A$ is said to have the Weak Lefschetz property, briefly WLP, if there exists an element $L\in A_1$ such that the multiplication map $\bullet L: A_i\to A_{i+1}$ is of maximal rank for $0\leq i \leq d-1$. \item The algebra $A$ is said to have the Strong Lefschetz property, briefly SLP, if there exists an element $L\in A_1$ such that the multiplication map $L^k: A_i\to A_{i+k}$ is of maximal rank for $0\leq i\leq d$ and $0\leq k\leq d-i$. \item We say that $A$ has the Strong Lefschetz property in the narrow sense, if there is $L \in A_1$ such that the linear map $\bullet L^{d-2k}: A_k \to A_{d-k}$ is an isomorphism for all $k\leq d/2$. \end{enumerate} \end{defin} \begin{rmk} In the case of standard graded Artinian Gorenstein algebra the two conditions SLP and SLP in the narrow sense are equivalent. \end{rmk} \begin{ex}\label{ex:monomial}\rm This example is due to Stanley \cite{St} and Watanabe \cite{Wa3}. It is considered to be the starting point of the research area of Lefschetz properties for graded algebras. Nowadays there are lots of different proofs for it. Consider the graded Artinian Gorenstein algebra $$A = \frac{\mathbb{C}[X_0,\ldots,X_n]}{(X_0^{a_0},\ldots,X_n^{a_n})} = \frac{\mathbb{C}[X_0]}{(X_0^{a_0})}\otimes \ldots \otimes \frac{\mathbb{C}[X_n]}{(X_n^{a_n})},$$ with integers $a_i>0$ for all $i=0,\ldots,n$. It is a monomial complete intersection. Since the cohomology of the complex projective space is $H^*(\mathbb{P}^m,\mathbb{C})=\mathbb{C}[x]/(x^{m+1})$, and the Segre product commutes with the tensor product by K\"unneth Theorem for cohomology, we have: $$H^*(\mathbb{P}^{a_0-1}\times \ldots \times \mathbb{P}^{a_n-1},\mathbb{C})=\frac{\mathbb{C}[X_0,\ldots,X_n]}{(X_0^{a_0},\ldots,X_n^{a_n})}. $$ By the Hard Lefschetz Theorem applied to the smooth projective variety $(\mathbb{P}^{a_0-1}\times \ldots \times \mathbb{P}^{a_n-1}$, we know that $A$ has the SLP. \end{ex} The standard graded Artinian Gorenstein algebra $A(f)$ associated to a form $f$ is a natural model for the cohomology algebras of spaces in several categories. For smooth projective varieties, the Hard Lefschetz theorem inspired what is now called Lefschetz properties for the algebra $A(f)$. As we show below, the geometric properties of the higher order objects introduced in this paper are intrinsicly connected with such Lefschetz properties, see Theorem \ref{thm:polarrankhess}. \subsection{Hessians and Lefschetz properties} We recall the following classical results involving the usual Hessian. Cones are trivial forms with vanishing Hessian and are characterized by the fact that $Z_X\subset H = \mathbb{P}^{n-1} \subset \mathbb{P}^n$ is a degenerate variety. Hesse claimed in \cite{He} that a reduced hypersurface has vanishing Hessian if and only if it is a cone. Gordan-Noether proved that the claim is true for $n \leq 3$ and false for $n \geq 4$ and this is part of the so called Gordan-Noether theory, see \cite{GN, CRS,Wa3, GR, Go, Ru}. More precisely, let $f \in R_d$ be a reduced form and let $X = V(f) \subset \mathbb{P}^n $ be the associated hypersurface. Consider the polar map associated to $f$: $$ \varphi_X:\mathbb{P}^n\dasharrow(\mathbb{P}^n)^{*}.$$ It is also called the gradient map of $X=V(f)\subset \mathbb{P}^n$, and it is defined by $$\varphi_X(p)=(f_{x_0}(p):\cdots :f_{x_n}(p)),$$ where $f_{x_i}= \frac{\partial f}{\partial x_i}$. The image $Z=Z_X$ of $\mathbb{P}^n$ under the polar map $\varphi_X$ is called the polar image of $X$. \begin{prop} \label{prop:GNcriteria} \cite{GN} Let $f\in \mathbb{C}[x_0,\ldots,x_n]$ be a reduced polynomial and consider $X = V(f) \subset \mathbb{P}^n$. Then \begin{enumerate} \item[(i)] $X$ is a cone if and only if $Z\subset H = \mathbb{P}^{n-1}$ is degenerated, which is equivalent to say that $ f_{x_0},\ldots,f_{x_n} $ are linearly dependent; \item[(ii)] $\operatorname{hess}_f=0$ if and only if $Z \subsetneq \mathbb{P}^n$, or equivalently $f_{x_0},\ldots,f_{x_n}$ are algebraically dependent. \end{enumerate} \end{prop} \begin{thm}\label{thm:GN} \cite{GN} Let $X = V(f) \subset \mathbb{P}^n$, $n \leq 3$, be a hypersurface such that $\operatorname{hess}_f=0$. Then $X$ is a cone. \end{thm} \begin{thm}\label{thm:GN2} \cite{GN} For each $n \geq 4$ and $d \geq 3$ there exist irreducible hypersurfaces $X = V(f) \subset \mathbb{P}^n$, of degree $\deg(f) = d$, not cones, such that $\operatorname{hess}_f=0$. \end{thm} Now we recall a generalization of a construction that can be found in \cite{MW}. Set $A=A(f)$, let $k\le l$ be two integers, take $L\in A_1$ and let us consider the linear map $$\bullet L^{l-k}: A_k\to A_l.$$ Let $\mathcal{B}_k=(\alpha_1,\ldots,\alpha_r)$ be a basis of the vector space $A_k$, and $\mathcal{B}_l=(\beta_1,\ldots,\beta_s)$ be a basis of the vector space $A_l.$ \begin{defin} \label{defMH} We call mixed Hessian of $f$ of mixed order $(k,l)$ with respect to the basis $\mathcal{B}_k$ and $\mathcal{B}_l$ the matrix: $$\operatorname{Hess}_f^{(k,l)}:=[ \alpha_i\beta_j(f)]$$ Moreover, we define $\operatorname{Hess}_f^k=\operatorname{Hess}_f^{(k,k)}$, $\operatorname{hess}_f^k = \det(\operatorname{Hess}_f^k)$ and $\operatorname{hess}_f=\operatorname{hess}_f^1$. \end{defin} Note that $A(f)_1=Q_1$ by our assumption, which implies that $\operatorname{Hess}_f=\operatorname{Hess}_f^1$ is the usual Hessian matrix of the polynomial $f$ and $\operatorname{hess}_f$ is the usual Hessian of $f$. Since $A$ is Gorenstein, there is an isomorphism $A_k^* \simeq A_{d-k} $. Therefore, given the basis $\mathcal{B}_k=(\alpha_1,\ldots,\alpha_r)$ of $A_k$ and a basis $\theta$ of $A_d \simeq \mathbb{C}$, we get the dual basis $\mathcal{B}^*_k=(\beta^*_1,\ldots,\beta^*_s),$ of $A_{d-k}$ in the following way $$\beta^*_i\beta_j(f)=\delta_{ij}\theta.$$ \begin{defin} We call dual mixed Hessian matrix the matrix $$\operatorname{Hess}^{(k^*,l)}(f):=[(\beta_i^*)\alpha_j(f)]$$ \end{defin} Note that $\operatorname{rk} \operatorname{Hess}^{(k^*,l)} = \operatorname{rk} \operatorname{Hess}^{(d-k,l)}$. If $L=a_0X_0+\ldots+a_nX_n \in Q_1,$ we set $L^{\perp}=(a_0,\ldots,a_n) \in \mathbb{C}^{n+1}.$ The next result can be found in \cite{GZ2} and it is a generalization of the main result of \cite{MW}. \begin{thm}\cite{GZ2} \label{thm:generalization} With the previous notation, let $M$ be the matrix associated to the map $\bullet L^{l-k}:A_k \to A_l$ with respect to the bases $\mathcal{B}_k$ and $\mathcal{B}_l.$ Then $$M=(l-k)!\operatorname{Hess}^{(l^*,k)}(f)(L^{\perp}).$$ \end{thm} \begin{cor} \label{cor1} For a generic $L$, one has the following. \begin{enumerate} \item The map $\bullet L^{d}:A_0 \to A_d$ is an isomorphism. \item The map $\bullet L^{d-2}:A_1 \to A_{d-1}$ is an isomorphism if and only if $\operatorname{hess}_f \ne 0$. \item If $d=2k$ is even, then $\operatorname{hess}_f^k \neq 0$. \item $A$ has the SLP if and only if $\operatorname{hess}^k_f \neq 0$ for all $k\leq d/2$. \end{enumerate} \end{cor} Using Theorem \ref{thm:GN} and Theorem \ref{thm:generalization} and we get the following. \begin{cor}\label{cor:lowdeg} All standard graded Artinian algebras $A$ of $\operatorname{codim} A \leq 4$ and of socle degree $=3,4$ have the SLP. \end{cor} \begin{cor}\cite{Go} For each pair $(n,d)\not\in\{(3,3), (3,4)\}$ with $N \geq 3$ and with $d \geq 3$, there exist standard graded Artinian Gorenstein algebras $A = \displaystyle \oplus_{i=0}^d A_i$ of codimension $\dim A_1=n+1 \geq 4$ and socle degree $d$ that do not satisfy the Strong Lefschetz Property. Furthermore, for each $L \in A_1$ we can choose arbitrarily the level $k$ where the map $$\bullet L^{d-2k} : A_k \to A_{d-k}$$ is not an isomorphism. \end{cor} \begin{rmk}\rm For algebras of codimension $2$, SLP hold in general. Therefore, for Gorenstein algebras, it means that the higher Hessians are not zero. This result is a first step in order to generalize Theorem \ref{thm:GN}. The issue is that in codimension $3$ the problem is open, that is, we do not know if there is an AG algebra failing SLP. A generalizaion of Theorem \ref{thm:GN2} can be found in \cite{Go}. In this work we give a generalization of Proposition \ref{prop:GNcriteria}, see Theorem \ref{thm:polarrankhess}. \end{rmk} \subsection{Jacobian ideals and Milnor algebras} Let $R=\mathbb{C}[x_0,\ldots,x_n]$ be the polynomial ring in $n+1$ variables with complex coefficients, endowed with the usual grading.\\ Let $f\in R_d$ be a homogeneous polynomial of degree $d$ such that the hypersurface $X=V(f)\subset \mathbb{P}^n$ is reduced. Let $J(f)$ be the Jacobian ideal of $f$, generated by the partial derivatives $f_{x_i}$, of $f$ with respect to $x_i$ for $i=0,\ldots,n$. If $X$ is smooth, then the ideal $J(f)$ is generated by a regular sequence, and $M(f)=R/J(f)$ is a Gorenstein Artinian algebra. Moreover we have $$\dim_{\mathbb{C}} M(f)<+\infty\Leftrightarrow V(f) \mbox{ is a smooth, }$$ and the corresponding Hilbert function is given by \begin{equation} \label{eq0} H(M(f);t)=\left( \frac{1-t^{d-1}}{1-t}\right)^{n+1}. \end{equation} In particular, the socle degree of $M(f) $ is $(d-2)(n+1)$.\\ Assume now that $X\subset \mathbb{P}^n$ is singular, but reduced. In this case the Jacobian algebra is not of finite length, in particular it is not Artinian. It contains information on the structure of the singularities and on the global geometry of $X$.The following results can be found in \cite{IG}. \begin{prop} \label{propIG1} Let $V:f = 0$ be a hypersurface in $\mathbb{P}^n$ of degree $d>2$, such that its singular locus $V_s$ has dimension at most $n-3$. Then $M(f)$ has the WLP in degree $d-2$. \end{prop} \begin{prop} \label{propIG2} Let $V:f = 0$ be a hypersurface in $\mathbb{P}^n$ of degree $d>2$, such that its singular locus $V_s$ has dimension at most $n-3$. Then for every positive integer $k<d-1$ $M(f)$ has the SLP in degree $d-k-1$ at range $k$. \end{prop} \begin{thm} \label{thmIG1} Let $V:f = 0$ be a general hypersurface, then $M(f)$ has the SLP. \end{thm} In view of the above result, it is natural to ask the following. \begin{question}\label{q1} \rm Is it true for any homogeneous polynomial $f$ with $V(f)$ smooth? \end{question} We have the following results in relation with this question. \begin{prop} \label{propSLP1} Let $V:f = 0$ be any smooth surface in $\mathbb{P}^3$ of degree $d=3$. Then $M(f)$ has the SLP. \end{prop} \begin{proof} Since $M(f)$ is Artinian Gorenstein, by \cite[Theorem 2.1]{MW} we have $$M(f) \cong Q/\operatorname{Ann}(g),$$ for some homogeneous polynomial $g$, where $$\deg(g)= \text{ socle degree of } M(f)= (n+1)(d-2)=4$$ and $\operatorname{hess}_g\neq 0,$ by Theorem \ref{thm:GN}. Indeed, otherwise $V(g)$ would be a cone, in contradiction with $\dim M(f)_1=4$. By Corollaries \ref{cor1} and \ref{cor:lowdeg}, $M(f)$ has the SLP. \end{proof} \begin{prop} \label{propSLP2} Let $V:f = 0$ be a smooth curve in $\mathbb{P}^2$ of even degree $d=2d'$. Then the multiplication by the square of a generic linear form $\ell \in R_1$ induces an isomorphism $$\ell^2: M(f)_{3d'-4} \to M(f)_{3d'-2}.$$ In particular, when $d=4$, the Milnor algebra $M(f)$ has the SLP. \end{prop} \begin{proof} Note that the socle degree of $M(f)$ is in this case $T=3(d-2)=6d'-6$. As explained in \cite[Remark 3.7]{DStJump}, a linear form $\ell$ such that the above map is not an isomorphism corresponds exactly to the fact that the associated line $L: \ell=0$ in $\mathbb{P}^2$ is a jumping line of the second kind for the rank two vector bundle $T\langle V \rangle $ on $\mathbb{P}^2$, where $T\langle V \rangle $ is the sheaf of logarithmic vector fields along $V$ as considered for instance in \cite{AD, DS14,MaVa}. Then a key result \cite[Theorem 3.2.2]{KH} of K. Hulek implies that the set of jumping lines of second kind is a curve in the dual projective plane $(\mathbb{P}^2)^{*}$ of all lines. When $d=4$, this yields an isomorphism $\ell^2: M(f)_{2} \to M(f)_{4}.$ The other isomorphisms necessary for the SLP follows from Corollary \ref{cor1}. \end{proof} \begin{rmk} \label{RkCurves} \rm For any smooth curve $V:f=0$ in $\mathbb{P}^2$, the associated Milnor algebra $M(f)$ has the WLP, as follows from the more general results in \cite{HMNW}. In addition, for a singular, reduced curve $V:f=0$ in $\mathbb{P}^2$, the associated Milnor algebra $M(f)$ is no longer Artinian or Gorenstein, but a partial WLP still holds, see \cite[Corollary 4.4]{DP}. \end{rmk}\rm \section{Higher order Jacobian ideals and Milnor algebras} Let us consider the $k$-th order Jacobian ideal of $f \in R$ to be $J^k =J^k(f)= (Q_k \ast f) = (A_k \ast f$), the ideal generated by the $k$-th order partial derivatives of $f$. Then $J^k$ is a homogeneous ideal and we define $M^k=M^k(f)=R/J^k$ to be the $k$-th order Milnor algebra of $f$. For $k=1$, the ideal $J^1$ is just the usual Jacobian ideal $J(f)$ of $f$ and $M^1$ is the usual Milnor algebra $M(f)$ as defined in the previous section. \begin{rmk} \label{R2} \rm For $k=2$, the ideal $J^2(f)$ is the ideal in $R$ spanned by all the second order partial derivatives of $f$. Euler formula implies that $J(f) \subset J^2(f)$, when $d=\deg(f) \geq 2$. It follows that $M^2(f)$ coincides with the graded {\it first Hessian algebra} $H_1(f)$ of the polynomial $f$, as defined in \cite{DSt1}. It follows from \cite[Theorem 1.1]{DSt1} and \cite[Example 2.7]{DSt1} that, for a hypersurface $V(f)$ having at most isolated singularities, the algebra $M^2(f)=H_1(f)$ is Artinian if and only if the multiplicity of the hypersurface $V(f)$ at any singular point is 2. \end{rmk}\rm The above remark can be extended to higher order Milnor algebras. First consider an isolated hypersurface singularity $(V,0):g=0$ at the origin of $\mathbb{C}^n$. Then we define the $k$-th order Tjurina ideal $TI^k(g)$ to by the ideal in the local ring ${\mathcal O}_n$ generated by all the partial derivatives $\partial^{\alpha}g$, for $0 \leq |\alpha|\leq k$. The $k$-th order Tjurina algebra of the germ $(V,0)$ is by definition the quotient $$T^k(V,0)=\frac{{\mathcal O}_n}{TI^k(g)}.$$ It can be shown that this algebra depends only on the isomorphism class of the germ $(V,0)$, and we define the $k$-th Tjurina number of $(V,0)$ to be the integer $$\tau^k(V,0)=\dim_{\mathbb{C}}T^k(V,0).$$ With this notation, we have the following result. \begin{thm} \label{T1} The $k$-th order Milnor algebra $M^k(f)$ of a reduced homogeneous polynomial $f$ is Artinian if and only if the multiplicity of the projective hypersurface $V(f)$ at any point $p \in V(f)$ is at most $k$. Moreover, if the hypersurface $V(f)$ has only isolated singularities, say at the points $p_1,\ldots,p_s$, then for any $k$ and for any large enough $m$ one has $$\dim_{\mathbb{C}} M^k(f)_m=\sum_{i=1}^s\tau^k(V,p_i).$$ \end{thm} \proof The algebra $M^k(f)$ is Artinian if and only if the zero set $Z(J^k(f))$ of the ideal $J^k(f)$ in $\mathbb{P}^n$ is empty. Note that a point $p \in Z(J^k(f))$ is a point on the hypersurface $V(f)$, by a repeated application of the Euler formula. If we choose the coordinates on $\mathbb{P}^n$ such that $p=(1:0:\ldots:0)$, then the local equation of the hypersurface germ $(V(f),p)$ is $$g(y_1,\ldots,y_n)=f(1,y_1,\ldots,y_n)=0,$$ exactly as in the proof of \cite[Theorem 1.1]{DSt1}. It follows that the localization of the ideal $J^k(f)$ at the point $p$ coincides with the ideal generated by all the partial derivatives $\partial^{\alpha}g$, for $0 \leq |\alpha|\leq k$. And these derivatives vanish all at $p$ exactly when the multiplicity of $V(f)$ at $p$ is $>k$. This proves the first claim. The proof of the second claim is completely similar. \endproof Note that \cite[Example 2.18]{DSt1} shows that even for a smooth curve $V(f)$, the algebra $M^2(f)=H_1(f)$ is not Gorenstein in general, since its Hilbert function, which depends on the choice of the smooth curve, is not symmetric and the dimension of the socle can be $>1$. However, there is a Zariski open subset $U_{d,k}$ in $R_d$ such that the Hilbert vector $\operatorname{Hilb}(M^k(f))$ is constant for $f \in U_{d,k}$. \begin{question}\label{q2} \rm Determine the value of the vector $\operatorname{Hilb}(M^k(f))$, or equivalently of the Hilbert function $H(M^k(f),t)$ for $f \in U_{d,k}$. \end{question} By semicontinuity, it follows that $$h_i(M^k(f))= \min \{ \dim(M^k(g)_i) \ : \ g \in R_d\}.$$ Similarly, there is a Zariski open subset $U'_{d}$ in $R_d$ such that the Hilbert vector $\operatorname{Hilb}(A(f))$ is constant for $f \in U'_{d}$. Using recent results by Zhenjian Wang, see \cite[Proposition 1.3]{ZW2}, we have the following. \begin{prop} \label{propZW} For a polynomial $f \in U'_d$, one has $h_k(A(f))=\dim Q_k={n+k \choose n}$ for $k \leq d/2$ and $h_k(A(f))=h_{d-k}(A(f))={n+d-k \choose n}$ for $d/2<k\leq d$. In particular, a Fermat type polynomial $f_F=x_0^d+ \ldots +x_n^d$ is not in $U'_d$, for $d \geq 4$. \end{prop} \proof Since $A(f)$ is Artinian Gorenstein with socle degree $d$, it is enough to prove only the claim for $k \leq d/2$. This claim is equivalent to $\operatorname{Ann}(f)_k=0$ for $k \leq d/2$, and also to $\dim J^k(f)_{d-k}=\dim Q_k$. This last equality is exactly the claim of \cite[Proposition 1.3]{ZW2}, where $J^k(f)_{d-k}$ is denoted by $E_k(f)$. The claim for the Fermat type polynomial follows from Example \ref{exFer}. \endproof \begin{cor} \label{cor2} For any $k \leq d/2$ and any polynomial $f \in U_{d,k}$ one has $$h_i(M^k(f))= {n+i \choose n}\text{ for } i<d-k$$ and $$h_{d-k}(M^k(f))=\dim R_{d-k}-\dim Q_k={n+d-k \choose n}-{n+k \choose n}.$$ In particular, a Fermat type polynomial $f=x_0^d+ \ldots +x_n^d$ is not in $U_{d,k}$, for $d \geq 2k \geq 4$. \end{cor} In conclusion, the introduction of higher order Milnor algebras is motivated by the desire to construct a larger class of Artin graded algebras starting with homogeneous polynomials. These new Artinian algebras may exhibit interesting examples with respect to Lefschetz properties. It is known that the Hessians are related to Lefschetz properties, and the Hessians of singular hypersurfaces behave in a different way from the ones of smooth hypersurfaces. As an example, we have the following. Recall first that $\operatorname{hess}_f$ denotes the Hessian of a homogeneous polynomial $f$ as in Definition \ref{defMH} or, more explicitly, $$\operatorname{hess}_f= \det \left( \frac{\partial ^2 f}{\partial x_i \partial x_j}\right)_{0\leq i,j \leq n}.$$ \begin{prop} \label{P1.5} Let $f$ be a homogeneous polynomial in $R$. \begin{enumerate} \item If the hypersurface $V(f)$ is smooth, then $\operatorname{hess}_f \notin J(f)$. \item If the hypersurface $V(f)$ is not smooth, but has isolated singularities, then $\operatorname{hess}_f \in J(f)$. \end{enumerate} \end{prop} \proof The first claim is well known, and it holds in fact for any isolated hypersurface singularity, not only for the cone over $V(f)$, see Theorem 1, section 5.11 in \cite{AGV}. The second claim is less known, and it follows from \cite[Proposition 1.4 (ii)]{DSt1}. Indeed, for the $n+1$-st Hessian algebra $H_{n+1}(f)$, one has the equalities $$H_{n+1}(f)=\frac{R}{J(f)+(\operatorname{hess}_f)}=\frac{M(f)}{({\overline \operatorname{hess}_f})},$$ where $(\operatorname{hess}_f)$ is the principal ideal in $R$ generated by the Hessian $\operatorname{hess}_f$ and $({\overline \operatorname{hess}_f})$ is the principal ideal in $M(f)$ generated by the class $ {\overline \operatorname{hess}_f}$ of the Hessian $\operatorname{hess}_f$ in $M(f)$. By \cite[Proposition 1.4 (ii)]{DSt1}, we know that the graded algebras $H_{n+1}(f)$ and $M(f)$ have the same Hilbert series when the hypersurface $V(f)$ is not smooth and has only isolated singularities. This proves our claim (2). \endproof \begin{question} \label{questionHESS} Is is true that $\operatorname{hess}_f \in J(f)$ for any reduced, singular hypersurface $V(f)$? \end{question} \section{The Hilbert functions of $A(f)$ and $M^k(f)$, and the geometry of $V(f)$} It is known that for two homogeneous polynomials $f, g \in R_d$, the corresponding Milnor algebras $M(f)$ and $M(g)$ are isomorphic as $\mathbb{C}$-algebras if and only if the associated hypersurfaces $V(f)$ and $V(g)$ in $\mathbb{P}^n$ are projectively equivalent. This claim follows from \cite{MY} when the hypersurfaces $V(f)$ and $V(f')$ are both smooth. However, the method of proof can be extended to cover all hypersurfaces. For a closely related result, see \cite{ZW1}. Note that a similar claim fails if we replace the Milnor algebra $M(f)=M^1(f)$ by the second order Milnor algebra $M^2(f)$. Indeed, it is enough to consider the family of complex plane cubics $f_a=x_0^3+x_1^3+x_2^3-3ax_0x_1x_2$, where $a\ne0$, $a^3\ne 1$. In this case $M^2({f_a})=R/(x_0,x_1,x_2)=\mathbb{C}$ does not detect the parameter $a$. However, the main result of \cite{ZW2} implies the following. \begin{thm} \label{thmZW2} The $k$-th order Milnor algebra $M^k(f)$ of a generic homogeneous polynomial $f$ determines the hypersurface $V(f)$ up-to projective equivalence, when $k \leq d/2-1$. \end{thm} \proof Let $f$ and $g$ be two generic, degree $d$ homogeneous polynomials in $R$, such that we have an isomorphism of graded algebras $$M^k(f)=R/J^k(f) \simeq R/J^k(g)=M^k(g).$$ Then there is a linear change of coordinates $\phi \in Gl_{n+1}(\mathbb{C})$, inducing the above isomorphism, and hence such that $\phi^*(J^k(f))=J^k(g)$. This last equality can be re-written as $J^k(f\circ \phi)=J^k(g)$, and Theorem 1.2 and Proposition 1.3 in \cite{ZW2} imply that the two hypersurfaces $V(f\circ \phi)$ and $V(g)$ coincide. \endproof Saying that the associated hypersurfaces $V(f)$ and $V(g)$ in $\mathbb{P}^n$ are projectively equivalent means that the two polynomials $f$ and $g$ belongs to the same $G$-orbit, where $G=Gl_{n+1}(\mathbb{C})$ is acting in the natural way on the space of polynomials $R_d$ by substitution. And the fact that $f$ and $g$ belongs to the same $G$-orbit implies immediately that the Milnor algebras $M(f)$ and $M(g)$ are isomorphic. Similarly, the fact that $f$ and $g$ belong to the same $G$-orbit implies immediately that the standard graded Artinian Gorenstein algebras $A(f)$ and $A(g)$ are isomorphic, see for instance \cite[Lemma 3.3]{DiPo}. In particular, the Hilbert function of $A(f)$ is determined by the $G$-orbit of $f$, and hopefully by the geometry of the corresponding hypersurface $V(f)$. However, the following example seems to suggest that it is a hard problem to relate the geometry of the hypersurface $V(f)$ to the properties of the algebras $A(f)$, $M(f)$ and $M^2(f)$. \begin{ex} \label{ex3}\rm In this example we look at quartic curves in $\mathbb{P}^2$, i.e. $(n,d)=(2,4)$. When $V(f)$ is not a cone, only dimension $h_2(A(f))$ has to be determined. It turns out that all the possible values $ \{3,4,5,6\}$ are obtained. All the computations below were done using CoCoA software, see \cite{CoCoA}, with the help of Gabriel Sticlaru. \medskip \noindent{\bf Case $V(f)$ smooth} \medskip All the smooth quartics $V(f)$ have the same Hilbert function $$H(M(f);t)=1+3t+6t^2+7t^3+6t^4+3t^5+t^6,$$ given by the formula \eqref{eq0}. But the other invariants may change, as the following examples show. \begin{enumerate} \item When $f_F=x_0^4+x_1^4+x_2^4$, the Fermat type polynomial of degree $4$, the corresponding Hilbert function is $$H(A(f_F);t)=1+3t+3t^2+3t^3+t^4,$$ by Example \ref{exFer}. The minimal resolution of $A(f_F)$ is given by $$0 \to Q(-7) \to Q(-3)^2 \oplus Q(-5)^3 \to Q(-2)^3 \oplus Q(-4)^2 \to Q,$$ in particular $A(f_F)$ is not a complete intersection. The second order Milnor algebra is $M^2({f_F})=R/(x_0^2,x_1^2,x_2^2)$, hence a complete intersection, with Hilbert function $$H(M^2({f_F});t)=1+3t+3t^2+t^3.$$ \item For the smooth Caporali quartic given by $f_{Ca}=x_0^4+x_1^4+x_2^4+(x_0+x_1+x_2)^4$, we get $$H(A(f_{Ca});t)=1+3t+4t^2+3t^3+t^4,$$ and the minimal resolution of $A(f_{Ca})$ is given by $$0 \to Q(-7) \to Q(-4) \oplus Q(-5)^2 \to Q(-2)^2 \oplus Q(-3) \to Q.$$ Hence $A(f_{Ca})$ is a complete intersection of multi-degree $(2,2,3)$. The second order Milnor algebra $M^2({f_{Ca}})$ has a Hilbert function given by $$H(M^2({f_{Ca}});t)=1+3t+2t^2,$$ in particular this algebra is not Gorenstein. \item For the smooth quartic given by $f_{Ca_1}=x_0^4+x_1^4+x_2^4+(x_0^2+x_1^2+x_2^2)^2$, we get $$H(Af_{Ca_1});t)=1+3t+6t^2+3t^3+t^4,$$ which coincides with the generic value given in Proposition \ref{propZW}, and the minimal resolution of $A(f_{Ca_1})$ is given by $$0 \to Q(-7) \to Q(-4)^7 \to Q(-3)^7 \to Q.$$ Hence $A(f_{Ca_1})$ is far from being a complete intersection. The second order Milnor algebra $M^2({f_{Ca_1}})$ has a Hilbert function given by $$H(M^2({f_{Ca_1}});t)=1+3t,$$ in particular this algebra is not Gorenstein. \item For the smooth quartic given by $f_{Ca_2}=x_0^4+x_1^4+x_2^4+(x_0^2+x_1^2)^2$, we get the same Hilbert function as for $A(f_{Ca})$, but the minimal resolution of $A(f_{Ca_2})$ is given by $$0 \to Q(-7) \to Q(-3)\oplus Q(-4)^2 \oplus Q(-5)^2 \to Q(-2)^2 \oplus Q(-3)^2\oplus Q(-4) \to Q.$$ The second order Milnor algebra $M^2({f_{Ca_2}})$ has a Hilbert function given by $$H(M^2({f_{Ca}});t)=1+3t+2t^2,$$ and hence this algebra is again not Gorenstein. \end{enumerate} \medskip \noindent{\bf Case $V(f)$ singular} \medskip \begin{enumerate} \item The rational quartic with an $E_6$-singularity, defined by $f_C=x_0^3x_1+x_2^4$ satisfies $H(A(f_{C});t)=H(A_{f_F};t)$ and the minimal resolution for $A_{f_{C}}$ is $$0 \to Q(-7) \to Q(-3)^2 \oplus Q(-5)^3 \to Q(-2)^3 \oplus Q(-4)^2 \to Q.$$ Hence the algebra $A(f_{C})$ has the same resolution as a graded $R$-module as the algebra $A(f_{F})$. But these two algebras are not isomorphic. Indeed, note that $$\operatorname{Ann}(f_F)_2=\langle X_0X_1,X_0X_2,X_1X_2\rangle \text{ and } \operatorname{Ann}(f_C)_2= \langle X_1^2,X_0X_2,X_1X_2\rangle.$$ An isomorphism $A(f_F) \simeq A(f_C)$ of $\mathbb{C}$-algebras would imply that the two nets of conics $$N_F: aX_0X_1+bX_0X_2+c X_1X_2 \text{ and } N_C: aX_1^2+b X_0X_2+c X_1X_2$$ are equivalent. This is not the case, since a conic in $N_F$ is singular if and only if it belongs to the union of three lines given by $abc=0$, while a conic in $N_C$ is singular if and only if it belongs to the union of two lines given by $ab=0$. For the associated Milnor algebras, one has $$H(M(f_C);t)=1+3t+6t^2+7t^3+7t^4+6 \frac{t^5}{1-t},$$ and $$H(M^2(f_C);t)=1+3t+3t^2+2 \frac{t^3}{1-t}.$$ Hence $M^2(f_C)$ is not Artinian, as predicted by Theorem \ref{T1}. Note that this curve has a unique $E_6$-singularity, with Tjurina numbers $\tau(E_6)=\tau^1(E_6)=6$ and $\tau^2(E_6)=2$, which explain the coefficients of the rational fractions in the above formulas, in view of Theorem \ref{T1}. \item For $f_{3A_2}=x_0^2x_1^2+x_1^2x_2^2+x_0^2x_2^2-2x_0x_1x_2(x_0+x_1+x_2)$, which defines a quartic curve with 3 cusps $A_2$, a direct computation shows that $$H(A(f_{3A_2});t)=1+3t+6t^2+3t^3+t^4,$$ which coincides with the generic value given in Proposition \ref{propZW}. The minimal resolution is $$0 \to Q(-7) \to Q(-4)^7 \to Q(-3)^7 \to Q.$$ Hence the algebra $A(f_{3A_2})$ has the same resolution as a graded $R$-module as the algebra $A(f_{Ca1})$. Does this imply that these two algebras are isomorphic? In this case $\operatorname{Ann}(f)_2=0$ and $\dim \operatorname{Ann}(f)_3=7$, hence it is more complicated to use the above method to distinguish these two algebras. Note also that the line arrangement $f=x_0x_1x_2(x_0+x_1+x_2)=0$ gives rise to an algebra $A(f)$ with exactly the same resolution as a graded $R$-module as the algebra $A(f_{Ca1})$. For the associated Milnor algebras, one has $$H(M(f_{3A_2});t)=1+3t+6t^2+7t^3+6 \frac{t^4}{1-t},$$ and $$H(M^2(f_{3A_2});t)=1+3t.$$ Hence $M^2(f_{3A_2})$ is Artinian, as predicted by Theorem \ref{T1}, but not Gorenstein. \item For $f_{2A_3}=x_0^2x_1^2+x_2^4$, which defines a quartic curve with $2$ singularities $A_3$, a direct computation shows that $$H(A(f_{2A_3});t)=1+3t+4t^2+3t^3+t^4$$ and the minimal resolution is $$0 \to Q(-7) \to Q(-3)\oplus Q(-4)^2 \oplus Q(-5)^2 \to Q(-2)^2 \oplus Q(-3)^2\oplus Q(-4) \to Q.$$ Hence the algebra $A(f_{2A_3})$ has the same resolution as a graded $R$-module as the algebra $A(f_{Ca_2})$. Does this imply that these two algebras are isomorphic? Note that $$\operatorname{Ann}(f_{2A_3})_2=\langle X_0X_2,X_1X_2\rangle =\operatorname{Ann}(f_{Ca_2})_2,$$ while $\operatorname{Ann}(f_{2A_3})_3$ and $\operatorname{Ann}(f_{Ca_2})_3$ have both dimension $7$. Hence again the above method to distinguish these two algebras is not easy to apply. For the associated Milnor algebras, one has $$H(M(f_{2A_3});t)=H(M(f_C);t)$$ and $$H(M^2(f_{2A_3});t)=1+3t+2t^2.$$ Hence $M^2(f_C)$ is Artinian, but not Gorenstein. \item For $f_{4A_1}=(x_0^2+x_1^2)^2+(x_1^2+x_2^2)^2$, which defines a quartic curve with $4$ singularities $A_1$ that is the union of two conics intersecting in the $4$ nodes, a direct computation shows that $$H(A(f_{4A_1});t)=1+3t+5t^2+3t^3+t^4$$ and the minimal resolution is $$0 \to Q(-7) \to Q(-4)^4 \oplus Q(-5) \to Q(-2) \oplus Q(-3)^4 \to Q.$$ \item For the line arrangement defined by $f=(x_0^3+x_1^3)x_2$, we get an algebra $A(f)$ with exactly the same resolution as a graded $R$-module as the algebra $A(f_{Ca})$, hence a complete intersection of multi-degree $(2,2,3)$. But these two algebras are not isomorphic. Indeed, note that $$\operatorname{Ann}(f_{Ca})_2=\langle X_0X_1-X_1X_2,X_0X_2-X_1X_2\rangle \text{ and } \operatorname{Ann}(f)_2=\langle X_0X_1,X_2^2\rangle.$$ An isomorphism $A(f_{Ca}) \simeq A(f)$ of $\mathbb{C}$-algebras would imply that the two pencils of conics $$P_{Ca}: a(y_0y_1-y_1y_2)+b(y_0y_2-y_1y_2) \text{ and } P_f: ay_0y_1+b y_2^2$$ are equivalent. This is not the case, since a conic in $P_{Ca}$ is singular if and only if it belongs to the union of three lines given by $ab(a+b)=0$, while a conic in $P_f$ is singular if and only if it belongs to the union of two lines given by $ab=0$. For the associated Milnor algebras, one has $$H(M(f_{4A_1});t)=1+3t+6t^2+7t^3+6t^4+4 \frac{t^5}{1-t},$$ and $$H(M^2(f_{4A_1});t)=1+3t+t^2.$$ Hence $M^2(f_{4A_1})$ is Artinian and Gorenstein. \end{enumerate} \end{ex} \begin{ex} \label{ex4}\rm We show here that for some smooth quartics $V(f)$ in $\mathbb{P}^2$ the multiplication by a generic linear form $\ell$ does not give rise to an injection $M^2(f)_1 \to M^2(f)_2$. Note that one has $$\dim M^2(f)_1=\dim R_1=3 \text{ and } \dim M^2(f)_2=\dim R_2-\dim A(f)_{2}= 6-\dim A(f)_{2},$$ using the general formula $$\dim J^k_{d-k}=\dim A(f)_{k}.$$ Hence, as soon as $\dim A(f)_{2}\geq 4$, the morphism $M^2(f)_1 \to M^2(f)_2$ cannot be injective. This happens for all the smooth quartics in Example \ref{ex3}, except for the Fermat one. \end{ex} \begin{question} \label{questionMY} \rm It would be interesting to find out whether Mather-Yau result extends to this setting, i.e. if an isomorphism $A(f) \simeq A(f')$ of $\mathbb{C}$-algebras implies that $f$ and $f'$ belongs to the same $G$-orbit. In the case when $(n,d)=(1,4)$ or $(n,d)=(2,3)$, one has $T=(n+1)(d-2)=d$ and the $G$-orbits of polynomials in $R_d$ corresponding to smooth hypersurfaces can be listed using 1-parameter families. In both cases, the correspondence $\operatorname{Mac}:f_t \to g_{u(t)}$ between a polynomial $f_t\in R_d$ with $V(f_t)$ smooth and a polynomial $g_{u(t)}=\operatorname{Mac}(f_t) \in R_d$, defined by the property that one has an isomorphism $$M(f_t) =A_{g_{u(t)}},$$ give rise to a bijection $u$ of the corresponding parameter space, see \cite{DiPo} for details. Hence the Mather-Yau result implies a positive answer to our question of the algebras $A_f$ in these two special cases. \end{question} \section{Higher Jacobians and Higher Polar mappings} Consider the following exact sequence $$0 \to I_k \to Q_k \to J^k_{d-k} \to 0,$$ where $I=\operatorname{Ann}(f)$ and $J^k=J^k(f)$. The map $Q_k \to J^k_{d-k}$ is given by evaluation $ \alpha \mapsto \alpha(f)$. \\ We have also the natural exact sequence: $$0 \to I_k \to Q_k \to A_k \to 0.$$ Note that the vector spaces $J^k_{d-k}$ and $A_k$ have the same dimension, \begin{equation}\label{eq1} \dim J^k_{d-k} = \dim Q_k -\dim I_k = \binom{n+k}{k} - \dim I_k = \dim A_k. \end{equation} \begin{defin}\rm The $k$-th polar mapping (or $k$-th gradient mapping) of the hypersurface $X=V(f) \subset \mathbb{P}^n$ is the rational map $\Phi^k_X: \mathbb{P}^n \dashrightarrow \mathbb{P}^{\binom{n+k}{k}-1}$ given by the $k$-th partial derivatives of $f$. The $k$-th polar image of $X$ is $\tilde{Z}_k = \overline{\Phi^k_X(\mathbb{P}^n)}\subseteq\mathbb{P}^{\binom{n+k}{k}-1}$, the closure of the image of the $k$-th polar map. \end{defin} For $\{\alpha_1, \ldots, \alpha_{a_k}\}$ a basis of the vector space $A_k$, we define the relative $k$-th polar map of $X$ to be the map $\varphi^k_X:\mathbb{P}^n \dashrightarrow \mathbb{P}^{a_k-1}$ given by the linear system $J^k_{d-k}$: $$\ \varphi^k_X(p)=(\alpha_1(f)(p): \ldots : \alpha_{a_k}(f)(p) ).$$ The $k$-th relative polar image of $X$ is $Z_k = \overline{\varphi^k_X(\mathbb{P}^n)}\subseteq \mathbb{P}^{a_k-1}$, the closure of the image of the relative $k$-th polar map.\\ It follows from Proposition \ref{propZW}, that for $k \leq d/2$ and $f$ generic, one has $a_k=\binom{n+k}{k}$ and hence $\Phi^k_X=\varphi^k_X$ in such cases. In general, the exact sequence $$0 \to I_k \to Q_k \to J^k_{d-k} \to 0$$ gives rise to a linear projection $\mathbb{P}^{\binom{n+k}{k}-1} \dashrightarrow \mathbb{P}^{a_k-1}$ making compatible these two polar maps, as in the diagram $$\begin{array}{ccc} \mathbb{P}^n & \rightarrow & \mathbb{P}^{\binom{n+k}{k}-1}\\ & \searrow & \downarrow \\ & & \mathbb{P}^{a_k-1}. \end{array} $$ Moreover, since $\tilde{Z}_k \subset \mathbb{P}(J^k_{d-k}) = \mathbb{P}^{a_k - 1} \subset \mathbb{P}^{\binom{n+k}{k}-1}$, the secant variety of $\tilde{Z}_k$ does not intersect the projection center, hence $Z_k \simeq \tilde{Z}_k$. The next result is a formalization of the intuitive idea that the mixed Hessian $\operatorname{Hess}_f^{(1,k)}$, introduced in Definition \ref{defMH}, is the Jacobian matrix of the gradient (or polar) map $\varphi^k$ of order $k$. \begin{thm}\label{thm:polarrankhess} With the above notations, we get: $$\dim Z_k = \dim \tilde{Z}_k = \operatorname{rk}(\operatorname{Hess}^{(1,k)}_X)-1.$$ In particular, the following conditions are equivalents: \begin{enumerate} \item[(i)] $\varphi^k$ is a degenerated map, that is, $\dim Z_k < n$; \item[(ii)] $\operatorname{rk} (\operatorname{Hess}^{(1,k)}_X) < n+1$; \item[(iii)] The map $\bullet L^{d-k-1}: A_1 \to A_{d-k}$ has not maximal rank for any $L \in A_1$. \end{enumerate} \end{thm} \begin{proof} If $p=[\mathbf v] \in \mathbb{P}^n$, then $T_p\mathbb{P}^n=\mathbb C^{n+1}/<\mathbf v>$ is the affine tangent space to $\mathbb{P}^n$ at $p$. Let $\operatorname{Hess}^{(1,k)}_X(p)$ be the equivalence class of the mixed Hessian matrix of $X=V(f)$ evaluated at $\mathbf v$. $\operatorname{Hess}^{(1,k)}_X(p)$ passes to the quotients and it induces the differential of the map $\varphi^k_X$ at $p$: \begin{equation}\label{eq:dfp} (d\varphi^k_X)_p: T_p\mathbb{P}^n\to T_{\varphi^k_X(p)}\mathbb{P}^{a_k-1}, \end{equation} whose image is exactly $T_{\varphi^k_X(p)}Z_k$, when $p$ is generic. From this we can describe explicitly the projective tangent space to $Z_k$ at $\varphi^k_X(p)$, obtaining \begin{equation}\label{eq:TpZ} T_{\varphi^k_X(p)}Z_k=\mathbb{P}(\Im(\operatorname{Hess}^{(1,k)}_X(v))\subseteq \mathbb{P}^{a_k-1}. \end{equation} Thus, there is an integer $\gamma_k \geq 0$ such that \begin{equation}\label{eq:dimZdiff} \dim Z_k = \operatorname{rk}(\operatorname{Hess}^{(1,k)}_X)-1=n-\gamma_k. \end{equation} The equivalence between $(i)$ and $(ii)$ is clear, since $\operatorname{rk} (\operatorname{Hess}^{(1,k)}_X) < n+1$ if and only if $\gamma_k > 0$ if and only if $\dim (Z_k) < n$. \\ The equivalence between $(ii)$ and $(iii)$ follows from Theorem \ref{thm:generalization}. \end{proof} Recall that for a standard graded Artinian Gorenstein algebra, the injectivity of $\bullet L: A_i \to A_{i+1}$ for a certain $L \in A_1$ implies the injectivity of $\bullet L: A_j \to A_{j+1}$ for all $j < i$, see \cite[Proposition 2.1]{MMN}. If $A(f)$ does not satisfy the WLP, then there is a minimal $k$ such that $\bullet L: A_k \to A_{k+1}$ is not injective for all $L \in A_1$. By Theorem \ref{thm:polarrankhess} the previous condition is equivalent to say that for all $j \leq k$ the matrix $\operatorname{Hess}_f^{(1,j)}$ has full rank. The following Corollary is a partial generalization of the Gordan-Noether Hessian criterion, recalled in Proposition \ref{prop:GNcriteria}, to the case of higher order Hessians. \begin{cor} \label{cor:polar1} Let $k\leq \lfloor \frac{d}{2}\rfloor$ be the greatest integer such that $\bullet L: A_{k-1} \to A_k$ is injective for some $L \in A_1$. For each $j \leq k$, we get that $\varphi^j$ degenerated implies $\operatorname{hess}^j_f = 0$. \end{cor} \begin{proof} The result follows from the commutative diagram, by Theorem \ref{thm:generalization} and by Theorem \ref{thm:polarrankhess}. $$\begin{array}{ccc} A_j & \to & A_{d-j}\\ \uparrow & \nearrow & \\ A_1 & \end{array} $$ Indeed, for $j\leq k$, we have that $\bullet L^{j-1}:A_1 \to A_j$ is injective, by composition. If $\bullet L^{d-2j}: A_j \to A_{d-j}$ is injective, then $\bullet L^{d-j-1} : A_1 \to A_{d-j}$ is injective. In other words, if $\operatorname{hess}^j_f \neq 0$, then $\varphi^j$ is not degenerated. \end{proof} The converse is not true, as one can see in the next example. \begin{ex}\rm Let $f=xu^3+yu^2v+zuv^2+v^4\in \mathbb{C}[x,y,z,u,v]_4$ as in \cite[Example 3]{Go}, and let $A=Q/\operatorname{Ann}(f)$. The map $\bullet L: A_1\to A_2$ is injective for $L=u+v$. For $j=2$, we have $$\operatorname{Hess}^2_f=\begin{pmatrix} 0&0&0&0&0&0&6&0\\ 0&0&0&0&0&2&0&0\\ 0&0&0&0&0&0&0&2\\ 0&0&0&0&0&0&2&0\\ 0&0&0&0&0&2&0&0\\ 0&2&0&0&2&0&0&0\\ 6&0&0&2&0&0&0&0\\ 0&0&2&0&0&0&0&24 \end{pmatrix}$$ and $\operatorname{hess}^2_f=\det\operatorname{Hess}^2_f=0$. Calculating the $\operatorname{Hess}_f^{(1,2)}$, we get: $$\operatorname{Hess}_f^{(1,2)}=\begin{pmatrix} 0&0&0&6u&0\\ 0&0&0&2v&2u\\ 0&0&0&0&2v\\ 0&0&0&2u&0\\ 0&0&0&2v&2u\\ 0&2u&2v&2y&2z\\ 6u&2v&0&6x&2y\\ 0&0&2u&2z&24v \end{pmatrix}.$$ The $\operatorname{rk}(\operatorname{Hess}_f^{(1,2)})=5$, hence $\varphi^2$ is not degenerated, by Theorem \ref{thm:polarrankhess}. \end{ex} \begin{cor} \label{corHess10} Let $f$ be a homogeneous form of degree $d$ and let $1<k<d-1$. Let $\varphi^k_f$ be the $k$-th polar map of $f$. If $\operatorname{hess}_f \neq 0$ then $\varphi^k_f$ is not degenerated, that is $\dim Z_k = n$. In particular, if $X = V(f)\subset \mathbb{P}^n$ with $n \leq 3$ is not a cone, then $\varphi^k_X$ is not degenerated. \end{cor} \begin{proof} By Theorem \ref{thm:polarrankhess} we have to prove that $rk(\operatorname{Hess}^{(1,k)}_f)=n+1$ is maximal. Let $L\in A_1$ be a general linear form. Since $\operatorname{hess}_f \neq 0$, the multiplication map $\bullet L^{d-2}: A_1 \to A_{d-1}$ is an isomorphism. Indeed, by Theorem \ref{thm:generalization}, after choosing basis the matrix of $\bullet L^{d-2}$ is $(d-2)!\operatorname{Hess}^{((d-1)^*,1)}$ whose rank is the same as the rank of $\operatorname{Hess}_f$. Since $\bullet L^{d-2}: A_1 \to A_{d-1}$ factors via $\bullet L^{d-k-1}: A_1 \to A_{d-k}$, the injectivity of $\bullet L^{d-2}$ implies the injectivity of $\bullet L^{d-k-1}$. On the other hand the rank of $\operatorname{Hess}^{((d-k)^*,1)}$ is equal to the rank of $\operatorname{Hess}_f^{(1,k)}$. The result now follows from Gordan-Noether theory. \end{proof} \begin{ex}\rm Let $X = V(f)\subset \mathbb{P}^3$ be Ikeda's surface given by $f=xuv^3+yu^3v+x^2y^3$ as in \cite{Ik, MW}. Since $X$ is not a cone, by the Gordan-Noether criterion, $\operatorname{hess}_f \neq 0$. On the other hand, $\operatorname{hess}^2_f=0$. By Corollary \ref{corHess10}, we know that the second polar map is not degenerated. Moreover, $\varphi_X = \Phi_X:\mathbb{P}^3 \dashrightarrow \mathbb{P}^9$ since $\dim A_2 = 10$. From an algebraic viewpoint we have $A_1 \to A_2 \to A_3$ and we know that $\bullet L: A_2 \to A_3$ is not an isomorfism, in particular it has non trivial kernel. On the other hand, $\bullet L^2: A_1 \to A_3$ is injective by Theorem \ref{thm:polarrankhess}. So the image of the first multiplication does not intersect the kernel of the second one. \end{ex} Consider the $k$-th polar map $\Phi^k_X:\mathbb{P}^n \to \mathbb{P}^N$ associated to a smooth hypersurface $X=V(f)$, where $N={n+k \choose k}-1$. Corollary \ref{corHess10} implies that the image of this map has dimension $n$ for any $k$ with $1<k<d-1$. There is a related result: consider the restriction \begin{equation} \label{eqPol} \psi^k_X=\Phi^k_X|X:X \to \mathbb{P}^N. \end{equation} Then it is shown in \cite{D} that the map $\psi^k_X$ is finite for $1\leq k<d$. The case $k=2$ is stated in \cite[Proposition 1.6]{D} and the general case in \cite[Remark 2.3 (ii)]{D}. We prove next that the map $\Phi^k_X$ is finite as well, which implies that $\dim Z_k=n$. Note that this map $\Phi^k_X$ is well defined as soon as $M^k(f)$ is Artinian. \begin{thm} \label{T2} Let $\Phi^k_X:\mathbb{P}^n \to \mathbb{P}^N$ be the $k$-th polar map associated to a hypersurface $X=V(f)\subset \mathbb{P}^n$ such that $M^k(f)$ is Artinian, where $N={n+k \choose k}-1$ and $0<k<d$. Then $\Phi^k_X$ is finite. In particular, $$\dim Z_k=\dim \Phi^k_X(\mathbb{P}^n)=n$$ and one has $$\deg \Phi^k_X \cdot \deg \Phi^k_X(\mathbb{P}^n)=(d-k)^n.$$ \end{thm} \proof Suppose there is a curve $C$ in $\mathbb{P}^n$ such that $\Phi^k_X$ is constant on $C$, say $$\Phi^k_f(x)=(b_{\alpha})_{|\alpha|=k} \in \mathbb{P}^N,$$ for any $x \in C$. There is at least one multi-index $\beta$ such that $b_{\beta}\ne 0$. But this leads to a contradiction, since either the partial derivative $\partial^{\beta}f$ is identically zero, or the hypersurface $\partial^{\beta}f=0$ meets the curve $C$, say at a point $p$. At such a point $p$, all the partial derivatives $\partial^{\alpha}f(p)=0$, for $|\alpha|=k$, in contradiction to the fact that $M^k(f)$ is Artinian. The claim about the degree follows by cutting $\Phi^k_X(\mathbb{P}^n)$ with $n$ generic hyperplanes $H_j$ in $\mathbb{P}^N$, where $j=1,...,n$ and using the fact that $(\Phi^k_X)^{-1}(H_j)$ is a family of $n$ hypersurfaces of degree $d-k$ meeting in $\deg \Phi^k_X \cdot \deg \Phi^k_X(\mathbb{P}^n)$ simple points. \endproof {\bf Acknowledgments}. The authors would like to thank the organizers and the participants of the workshop {\it Lefschetz Properties and Jordan Type in Algebra, Geometry and Combinatorics}, that took place in Levico Terme, Trento, Italy in June 2018. This work was started at this conference. \\ The second author would like to thank Francesco Russo and Giuseppe Zappalà for many conversations on this subject.

https://arxiv.org/abs/2003.10312

A termination criterion for stochastic gradient descent for binary classification

We propose a new, simple, and computationally inexpensive termination test for constant step-size stochastic gradient descent (SGD) applied to binary classification on the logistic and hinge loss with homogeneous linear predictors. Our theoretical results support the effectiveness of our stopping criterion when the data is Gaussian distributed. This presence of noise allows for the possibility of non-separable data. We show that our test terminates in a finite number of iterations and when the noise in the data is not too large, the expected classifier at termination nearly minimizes the probability of misclassification. Finally, numerical experiments indicate for both real and synthetic data sets that our termination test exhibits a good degree of predictability on accuracy and running time.

\section{Analysis of stopping criterion} \label{sec:Analysis} In this section, we present our analysis of the stopping criterion \eqref{eq: termination_test} proposed in Section~\ref{sec:termtest}. Here we introduce the first iteration at which the stopping criterion is satisfied, denoted by the random variable \begin{equation}\label{eq:stopping_criteria} T:=\inf\left\{k> 0: \hat{\bm{\xi}}_k^T\bm{\theta}_k \geq 1\right\}. \end{equation} By viewing the stopping criterion through the lens of stopping times, we are able to utilize probability theory to analyze the classifier at termination $\bm{\theta}_T$. Throughout this section, we work with the following filtration. \begin{equation}\label{eq:sigma} \mathcal{F}_0 = \sigma(\bm{\theta}_0) \quad \text{and} \quad \mathcal{F}_k:=\sigma(\bm{\theta}_0, \hat{\bm{\xi}_1}, \bm{\xi}_1, \hat{\bm{\xi}_2}, \bm{\xi}_2, \hdots, \hat{\bm{\xi}_k}, \bm{\xi}_k), \quad \text{for all $k\geq 1$} \end{equation} Clearly, the random variable $\bm{\theta}_k$ is $\mathcal{F}_k$-measurable. Our theoretical results are structured as follows. First, we show that SGD with our proposed termination test indeed stops after a finite number of iterations. To do so, we provide a bound on $\bE[T]$, \textit{i.e.} the expected number of iterations before termination. Yet, despite this guarantee, the resulting classifier at termination need not be optimal. Hence, our second result establishes that both $\bm{\theta}_T$ and $\bm{\theta}^*$ point in approximately the same direction; thereby ensuring that the classifier at termination, $\bm{\theta}_T$, is nearly optimal. We remark the worst-case bounds established throughout these sections are conservative; we observe in our experiments that the termination test stops sooner while also yielding good classification properties for Gaussian and non-Gaussian data sets. To bound $\bE[T]$, we identify subsets of $\R^d$ for which when an iterate enters the set, termination (\textit{i.e.} \eqref{eq: termination_test}) is \textit{highly likely} to succeed. Such sets $C$, we call \textit{target sets}. Precisely, for any $\bm{\theta} \in C$ and $\hat{\bm{\xi}}\sim N(\bm{\mu},\sigma^2I_d)$, the probability of terminating is at least $\delta>0$, \begin{equation}\label{eq:probability_delta} \exists \; \delta>0 \text{ such that } \mathbb{P}_{\hat{\bm{\xi}}}\left( \hat{\bm{\xi}}^T\bm{\theta}\geq 1 \right) \geq \delta. \end{equation} We guarantee the iterates generated by SGD enter the target set by way of a \textit{drift function}, $V:\R^d\rightarrow [0,+\infty)$. A drift function, on average, decreases each time the iterate fails to live in the target set. In other words, conditioned on the past iterates the following holds \begin{equation}\label{eq:driftequation} (\bE[V(\bm{\theta}_k)|\mathcal{F}_{k-1}]-V(\bm{\theta}_{k-1})])1_{\{\bm{\theta}_{k-1}\not\in C\}} \leq -b1_{\{\bm{\theta}_{k-1}\not\in C\}} \end{equation} for the target set $C$ and some positive constant $b$. Loosely speaking, the iterates in expectation \textit{drift} towards the target set. Target sets and drift functions in the context of drift analysis are well-studied in stochastic processes, see Lemma \ref{lem:drift_from_meyn} below. A natural choice for the target set is a neighborhood of the unique optimum solution of \eqref{optimization_problem}, $\bm{\theta}^*$, with the drift function $\Vert \bm{\theta}-\bm{\theta}^* \Vert^2$. Indeed, it is known the iterates of SGD converge to a neighborhood of $\bm{\theta}^*$ (\cite{Pflug}). However, an iterate may be nearly optimal well before it enters this neighborhood. In fact when $\sigma \ll \Vert \bm{\mu} \Vert$, we identify a target set where satisfying the stopping criterion occurs at least half the time and does not require the iterate to be near $\bm{\theta}^*$. We summarize below our target set and drift function. \begin{enumerate} \item Under the assumption $\sigma \leq c \Vert \bm{\mu} \Vert$ for some numerical constant $c$, which we call the \textit{Low Variance Regime}, we define the target set to be \begin{equation}\label{eq:targetset_lownoise} C = \{\bm{\theta}: \bm{\mu}^T\bm{\theta}\geq 1 \}, \end{equation} and the drift function by \begin{equation}\label{eq:driftfunction_lownoise} V(\bm{\theta})=\left(M-\bm{\mu}^T\bm{\theta}\right)^2, \end{equation} for some constant $M$, to be determined later. \item Under the assumption $c\Vert \bm{\mu} \Vert \leq \sigma$ where the constant $c$ is the same as in 1 above, which we call the \textit{High Variance Regime}, we define the target set to be \begin{equation}\label{eq:targetset_highnoise} C=\{\bm{\theta}: \vert \rho\sigma^2-1\vert <1 \text{ and } \sigma \Vert \tilde{\bm{\theta}}\Vert\leq c'\}, \end{equation} for some numerical constant $c'$. Here, we orthogonally decompose $\bm{\theta}=\rho \bm{\mu}+\tilde{\bm{\theta}}$ with $\bm{\mu}^T\tilde{\bm{\theta}}=0$. We use the following drift function \begin{equation}\label{eq:driftfunction_highnoise} V(\bm{\theta})=\frac{1}{2\alpha}\Vert \bm{\theta}-\bm{\theta}^*\Vert^2. \end{equation} \end{enumerate} In Section \ref{sec:proof_low} (resp. Section \ref{sec:proof_high}) we show that the pairs $(C,V)$ defined in \eqref{eq:targetset_lownoise} and \eqref{eq:driftfunction_lownoise} (resp. \eqref{eq:targetset_highnoise} and \eqref{eq:driftfunction_highnoise}) satisfies the drift equation \eqref{eq:driftequation} for any step-size $\alpha$ (resp. for any sufficiently small step-size $\alpha$). As mentioned above, the target set $C$ attracts the iterates generated by SGD. Each time an iterate enters $C$, the stopping criterion holds with probability at least $\delta>0$. Provided the iterates enters the set $C$ an infinite number of times, then after waiting a geometrically distributed many iterations, we expect the following condition to hold: \begin{equation}\label{eq:termination_inside_C} \hat{\bm{\xi}}_k^T\bm{\theta}_k\geq 1 \text{ and } \bm{\theta}_k \in C. \end{equation} The SGD algorithm does not know the value of $\bm{\theta}^*$; therefore at each iteration, it cannot check whether the condition \eqref{eq:termination_inside_C} occurs. Nevertheless, we are able to compute a bound on the average waiting time until \eqref{eq:termination_inside_C} holds and the first time \eqref{eq:termination_inside_C} holds is always an upper bound on $T$, our stopping criterion. This is summarized in Lemma \ref{lem:ETleqET_C}. Precisely, if we denote by \begin{equation}\label{eq:stoppingtime_TC} T_C:=\inf\{k>0:\hat{\bm{\xi}}_k^T\bm{\theta}_k\geq 1 \text{ and } \bm{\theta}_k \in C\}, \end{equation} then $T\leq T_C$, thus yielding $\bE[T]\leq \bE[T_C]$. We bound $\bE[T_C]$ by way of stopping times $\tau_m$ defined as the $m^{th}$ time the iterates of SGD enters $C$. Formally for any sequence $\{\bm{\theta}_k\}_{k=0}^\infty$ generated by SGD starting at $\bm{\theta}_0 = \bm{0}$, we set \begin{equation}\label{eq:tau1} \tau_1:=\inf\{k>0:\bm{\theta}_k\in C\} \end{equation} and inductively, for $m\geq 2$, \begin{equation}\label{eq:taum} \tau_m:=\inf\{k>\tau_{m-1}:\bm{\theta}_k\in C\}. \end{equation} The following lemma formalizes the discussion above. \begin{lemma}\label{lem:ETleqET_C} Let $\{\bm{\theta}_k\}_{k=0}^\infty$ be a sequence generated by SGD such that $\bm{\theta}_0 = \bm{0}$ and suppose that $\bE[\tau_m]<+\infty$ for all $m\geq 1$. Then the following holds \begin{equation} \bE[T]\leq \bE[T_C]\leq \sum_{m=1}^{\infty} \bE[\tau_m](1-\delta)^{m-1}, \end{equation} where $\delta$ satisfies \eqref{eq:probability_delta}. \end{lemma} \begin{proof} We first show that \begin{equation} \label{eq:analysis_1} \bE\left[1_{\{T_C\geq \tau_m\}}\right]\leq (1-\delta)^{m-1}.\end{equation} Define the $\sigma$-algebra $\mathcal{F'}=\sigma(\bm{\theta}_0,\bm{\xi}_1,\bm{\xi}_2,\cdots)$. From the independence between $\sigma(\hat{\bm{\xi}}_k)$'s and $\mathcal{F}'$ and also $\tau_i<+\infty$ a.s. for all $i\geq 1$, the following is obtained: \begin{align*} \bE\left[1_{\{ T_C\geq \tau_m\}}|\mathcal{F}'\right] &= \bE\left[1_{\{ \hat{\bm{\xi}}_{\tau_1}^T\bm{\theta}_{\tau_1}<1\}}\cdots1_{\{ \hat{\bm{\xi}}_{\tau_{m-1}}^T\bm{\theta}_{\tau_{m-1}}<1\}}|\mathcal{F}'\right]\\&=\prod_{i=1}^{m-1}\bE\left[1_{\{ \hat{\bm{\xi}}_{\tau_i}^T\bm{\theta}_{\tau_i}<1\}}|\mathcal{F}'\right]\\&\leq (1-\delta)^{m-1}. \end{align*} By taking expectations, we conclude \eqref{eq:analysis_1} holds. Now since $\bE[1_{\{T_C=+\infty\}}]\leq \bE[1_{\{T_C\geq \tau_m\}}]$ for all $m\geq 1$, it follows from \eqref{eq:analysis_1} that $T_C<\infty$ a.s. We next observe that \begin{align*} \bE\left[T_C1_{\{ T_C= \tau_m\}}|\mathcal{F}'\right]&=\bE\left[\tau_m1_{\{ T_C= \tau_m\}}|\mathcal{F}'\right]\\ &\le\tau_m\bE\left[1_{\{ \hat{\bm{\xi}}_{\tau_{1}}^T\bm{\theta}_{\tau_{1}}<1\}}\cdots1_{\{ \hat{\bm{\xi}}_{\tau_{m-1}}^T\bm{\theta}_{\tau_{m-1}}<1\}}|\mathcal{F}'\right]\\ &=\tau_m\prod_{i=1}^{m-1}\bE\left[1_{\{ \hat{\bm{\xi}}_{\tau_{i}}^T\bm{\theta}_{\tau_{i}}<1\}}|\mathcal{F}'\right]\\ &\leq \tau_m(1-\delta)^{m-1}. \end{align*} Taking expectations yields $\bE\left[T_C1_{\{T_C=\tau_m\}}\right]\leq \bE\left[\tau_m\right](1-\delta)^{m-1}$ for all $m\geq 1$. Now since $T_C<\infty$ a.s. we get $1=\sum_{m=1}^{+\infty}1_{\{T_C=\tau_m\}}$ a.s. This yields that \[ \bE[T]\leq \bE[T_C]=\sum_{m=1}^{\infty} \bE[T_C1_{\{T_C=\tau_m\}}]\leq \sum_{m=1}^{\infty}\bE[\tau_m](1-\delta)^{m-1}. \] The proof is complete. \end{proof} Now, in view of Lemma \ref{lem:ETleqET_C}, it suffices to bound $\bE[\tau_m]$ by a sequence which can not grow too fast in $m$. Indeed, we show that \eqref{eq:driftequation} implies the following \begin{equation} \bE[\tau_m]=\mathcal{O}(m). \end{equation} \begin{theorem}(Low Regime)\label{thm:low} Let $\{\bm{\theta}_k\}_{k=0}^\infty$ be a sequence generated by Algorithm~\ref{alg:SGD_termination} such that $\bm{\theta}_0=\bm{0}$. There exists positive constants $c,b$ and $M$ such that provided $\sigma \leq c\Vert \bm{\mu} \Vert$ the following holds. \begin{equation} \bE[T]\leq 2+\frac{2M^2}{b}\cdot\left(\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\frac{\alpha\sigma^3}{ \Vert \bm{\mu} \Vert }\cdot\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right)+1 \right). \end{equation} Here the constants $c,b$ and $M$ are defined as follows: \begin{enumerate} \item For the logistic loss, \begin{equation}\label{eq:para_log} c=0.33,\quad b=\alpha\Vert \bm{\mu} \Vert^2,\quad \text{ and } M=501+640\alpha \Vert \bm{\mu} \Vert^2. \end{equation} \item For the hinge loss, \begin{equation}\label{eq:para_hinge} c=1.25,\quad b=\alpha\Vert \bm{\mu} \Vert^2,\quad \text{ and } M=501+782\alpha \Vert \bm{\mu} \Vert^2. \end{equation} \end{enumerate} \end{theorem} Therefore, on relatively separable data (\textit{i.e.} in the low variance regime), the expected waiting time before termination exponentially decreases as the data becomes more separable (\textit{i.e.} $\sigma \to 0$). We prove Theorem \ref{thm:low} in Section \ref{sec:theorem_angle}. The next theorem shows that the expected value of the stopping time is finite provided that the $\sigma > c\Vert \bm{\mu} \Vert$ and the step-size is small enough. \begin{theorem}\label{thm:high}(High Regime) Suppose that $\sigma > c\Vert \bm{\mu} \Vert$ where $c$ is defined in \eqref{eq:para_log} and \eqref{eq:para_hinge}. Then there exists a universal positive constant $A$ such that if the step-size $\alpha$ satisfies \begin{equation} \alpha \leq A\cdot\frac{\Vert \bm{\mu} \Vert^2}{\sigma^2(\Vert \bm{\mu} \Vert^2+d\sigma^2)}, \end{equation} then it holds that $\bE[T]<+\infty$. In particular, the termination criterion occurs almost surely. \end{theorem} It remains to determine whether the classifier at termination $\bm{\theta}_T$, has desirable accuracy. The scale-invariance of optimal classifiers means a classifier yields a lower probability of misclassification the closer its direction aligns with any optimal classifier. In view of this, it suffices to bound the absolute value of the inner product of any unit vector that is perpendicular to $\bm{\theta}^*$, $\bm{v}$ with $\bm{\theta}_T$. The following theorem establishes a bound on $\bE[\vert \bm{v}^T\bm{\theta}_T\vert]$. \begin{theorem}\label{thm:angle_bound} Let $\bm{\theta}_0=\bm{0}$. Fix any unit vector $\bm{v}\in \R^d$ such that $\bm{v}^T\bm{\theta}^*=0$. Then the following estimate holds \begin{equation} \bE[\vert \bm{v}^T\bm{\theta}_T\vert]\leq \sigma \alpha \sqrt{\frac{2}{\pi}} \bE[T]. \end{equation} \end{theorem} In the low variance regime by combining Theorem \ref{thm:low} and \ref{thm:angle_bound} for a fixed step-size $\alpha$ it holds that $\bE[\vert \bm{v}^T\bm{\theta}\vert]\leq \mathcal{O}(\sigma)$. Thus, the more separable the data set is, the more accurate the classifier $\bm{\theta}_T$ is on average. In the high variance regime, Theorem \ref{thm:high} yields a very loose bound. Yet despite this, our numerical result in Section \ref{sec:Num_Experiment} show promising accuracy of \eqref{eq: termination_test} in this case as well. We conjecture that the inequality can be significantly strengthened. \input{arxiv_low_variance_regime.tex} \input{arxiv_high_variance_regime.tex} \input{arxiv_angle_bound.tex} \subsection{Angle bound, proof of Theorem \ref{thm:angle_bound}}\label{sec:theorem_angle} \begin{proof}[Proof of Theorem \ref{thm:angle_bound}] Recall the SGD algorithm for logistic regression uses the update \[ \bm{\theta}_k=\bm{\theta}_{k-1}+\frac{\alpha\bm{\xi}_k}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})} \] and for hinge regression \[ \bm{\theta}_k=\bm{\theta}_{k-1}+\alpha 1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}\bm{\xi}_{k-1} \] where $\bm{\theta}_0=\bm{0}$ and $\bm{\xi}_1,\bm{\xi}_2,\cdots \stackrel{i.i.d}{\sim} N(\bm{\mu},\sigma^2I_d)$. It clearly holds in both cases that \begin{equation} \left\vert\vert \bm{v}^T\bm{\theta}_{k}\vert-\vert \bm{v}^T\bm{\theta}_{k-1}\vert \right\vert\leq \alpha \vert \bm{v}^T\bm{\xi}_{k-1}\vert. \end{equation} We define a new random variable $X_k:=\vert \bm{v}^T\bm{\theta}_k\vert-k \sigma\alpha \sqrt{\frac{2}{\pi}} $. Observe that $\bE\left[\left|X_0\right|\right] = 0$ and for all $k \ge 1$, it holds that \begin{align*} \bE\left[\left\vert X_k\right\vert\right]&\leq \alpha\sum_{i=1}^k \bE\left[\left\vert \bm{v}^T\bm{\xi}_k \right\vert\right]+k\sigma\alpha \sqrt{\frac{2}{\pi}} <\infty, \end{align*} \textit{i.e.}, $X_k \in \mathcal{L}^1$ for all $k \geq 1$. Next, we have for any $k \ge 1$ \begin{align*} \bE\left[\left\vert X_{k}-X_{k-1}\right\vert\,|\,\mathcal{F}_{k-1} \right] \leq \bE\left[\left\vert \vert \bm{v}^T\bm{\theta}_k\vert -\vert \bm{v}^T\bm{\theta}_{k-1}\vert\right\vert\,|\,\mathcal{F}_{k-1} \right]+\sigma\alpha \sqrt{\frac{2}{\pi}} &\leq 2\sigma\alpha \sqrt{\frac{2}{\pi}} . \end{align*} Here we used that $\bm{v}^T\bm{\xi}_k \sim N(0,\sigma^2)$ along with \eqref{fact:norm_Gaussians}. We also see that \begin{align*} \bE\left[\vert \bm{v}^T\bm{\theta}_{k}\vert\,|\,\mathcal{F}_{k-1} \right]&\leq \vert \bm{v}^T\bm{\theta}_{k-1}\vert+ \sigma \alpha\sqrt{\frac{2}{\pi}} \quad \Rightarrow \quad \bE\left[X_{k}\,|\,\mathcal{F}_{k-1} \right]\leq X_{k-1}. \end{align*} Therefore, we have shown that $X_0,X_1,\cdots$ is a super-martingale. By Theorem \ref{thm:Durret_Martingale}, we have $\bE\left[X_{T}\right]\leq 0$. The result follows. \end{proof} \section{Appendix} In this section, we provide proofs of technical results that we used in the paper. \subsection{Drift Analysis Lemma} Here we state and prove a result on Drift Analysis that we used in the paper. Recall that we require the following equation to hold \begin{equation}\label{drift_equation1_appendix} \left( \bE\left[V(\bm{\theta}_k)|\mathcal{F}_{k-1} \right]-V(\bm{\theta}_{k-1}) \right)1_{\{\bm{\theta}_{k-1}\not\in C\}}\leq -1_{\{\bm{\theta}_{k-1}\not\in C\}}. \end{equation} However, when the iterate lies inside $C$, then we do not assume any bound on the expected increase in the drift function. Nevertheless, when the target set $C$ is compact then there exists a positive constant $b>0$ such that \begin{equation}\label{drift_equation2_appendix} \left( \bE\left[V(\bm{\theta}_k)|\mathcal{F}_{k-1} \right]-V(\bm{\theta}_{k-1}) \right)1_{\{\bm{\theta}_{k-1}\in C\}}\leq b1_{\{\bm{\theta}_{k-1}\in C\}}, \end{equation} The following lemma bounds the expected value of return times $\tau_m^C$. \begin{lemma}\label{lem:drift_lemma_appendix} Suppose that for some test function $V$ and a target set $C$, the drift equation \eqref{drift_equation1_appendix} holds. Let $\hat{\bm{\theta}} \in \R^d\backslash C$. The following is then true \begin{equation} \bE[\tau_1^C|\bm{\theta}_0=\hat{\bm{\theta}}]\leq V(\hat{\bm{\theta}}). \end{equation} In addition, if both equations \eqref{drift_equation1_appendix} and \eqref{drift_equation2_appendix} hold, then the following is true \begin{equation} \bE[\tau_m^C|\bm{\theta}_0=\hat{\bm{\theta}}]\leq V(\hat{\bm{\theta}})+(m-1)\sup_{\bm{\theta}\in C} V(\bm{\theta}). \end{equation} \end{lemma} \begin{proof} See \cite{meyn2012markov}. \end{proof} \subsection{Stopping Time Lemma}\label{stopping_time_lemma} In light of drift equation \eqref{drift_equation_appendix}, the iterates are inclined towards the target set $C$. With a positive probability $\delta$, the test will be activated once we are in $C$. Hence, the expected value of the stopping time $T$, i.e. $\bE[T]$, can be upper-bounded in terms of $\delta$ and expected return times $\bE[\tau_m^C]$. \begin{lemma}\label{lem:stopping_time_appendix} Suppose that $\bE[\tau_m^{C}]<\infty$ for all $m\geq 1$. Then it is true that $T_C<+\infty$. Moreover, the following holds. \begin{equation}\label{eq:ST_tctm} \bE[T]\leq \bE[T_C]\leq \sum_{m=1}^{\infty}\bE[\tau_m^{C}](1-\delta)^{m-1}, \end{equation} where for any fixed $\bm{\theta}\in C$ the constant $\delta$ satisfies \begin{equation} \delta= \min \mathbb{P}_{\bm{\xi} \sim \mathcal{P}_*}\left(\bm{\xi}^T\bm{\theta}\geq 1 \right). \end{equation} \end{lemma} \subsection{Basic Convex Analysis Lemma} Here we state and prove two basic results from convex analysis that we used in the paper. \begin{lemma}\label{lem:convex_analysis_lemma} Suppose that $g:\R^{\geq 0}\to \R$ is a convex function with a minimizer at $\rho^*>0$. Assume that $g$ is twice differentiable on the interval $[\tfrac{3}{4}\rho^*,\tfrac{5}{4}\rho^*]$. Moreover, assume that there exists a constant $r>0$ such that $g''(\rho)\geq r$ for all $\rho \in [\tfrac{3}{4}\rho^*,\tfrac{5}{4}\rho^*]$. Then it holds that \begin{equation}\label{eq0:cvx_lemma} g(\rho)-g(\rho^*)\geq \frac{\rho^*r}{8}\vert \rho-\rho^*\vert \quad \text{for all $\rho \not\in [\tfrac{1}{2}\rho^*,\tfrac{3}{2}\rho^*]$}. \end{equation} \end{lemma} \begin{proof} First, assume that $\rho > \tfrac{3}{2}\rho^*$ holds. There exists $\hat{\rho}\in [\rho^*,\tfrac{5\rho^*}{4}]$ such that \begin{equation}\label{eq1:cvx_lemma} g'(\rho^*+\frac{\rho^*}{4})=g'(\rho^*+\frac{\rho^*}{4})-g'(\rho^*)=\frac{\rho^*}{4}g''(\hat{\rho})\geq \frac{\rho^*r}{4}, \end{equation} where we used that $g'(\rho^*)=0$. With the convexity of $g$, for any $\rho>\frac{3\rho^*}{2}$, we have \[ g(\rho)\geq g(\rho^*+\frac{\rho^*}{4})+g'(\rho^*+\frac{\rho^*}{4})(\rho-\frac{5\rho^*}{4})\geq g(\rho^*)+\frac{\rho^*r}{4}(\rho-\frac{5\rho^*}{4}). \] Note that the second inequality follows from \eqref{eq1:cvx_lemma} and the optimal value of $g$ occurring at $g(\rho^*)$. Using the identity, $2(\rho-\frac{5\rho^*}{4})\geq \frac{\rho^*}{4}+\rho-\frac{5\rho^*}{4}=\rho-\rho^*$, we conclude \begin{equation}\label{eq:cvx_lemma1} g(\rho)-g(\rho^*)\geq \tfrac{1}{8}\rho^*r(\rho-\rho^*) \quad \text{for all $\rho > \frac{3\rho^*}{2}$}. \end{equation} We now consider the second case, $\rho<\frac{\rho^*}{2}$. Similarly as above, we get \begin{equation}\label{eq2:cvx_lemma} -g'(\rho^*-\frac{\rho^*}{4})= g'(\rho^*)-g'(\rho^*-\frac{\rho^*}{4})=\frac{\rho^*}{4}g''(\hat{\rho})\geq \frac{\rho^*r}{4}. \end{equation} for some $\hat{\rho}\in [\frac{3\rho^*}{4},\rho^*]$. Since $\rho<\frac{\rho^*}{2}$, we have \[ g(\rho)\geq g(\rho^*-\frac{\rho^*}{4})+g'(\rho^*-\frac{\rho^*}{4})(\rho-\frac{3\rho^*}{4})\geq g(\rho^*)+\frac{\rho^*r}{4}(\frac{3\rho^*}{4}-\rho). \] Here, again, the first inequality follows from convexity of $g$ and the second inequality from \eqref{eq2:cvx_lemma} and $g(\rho^*-\frac{\rho^*}{4})\geq g(\rho^*)$. We note that $2(\frac{3\rho^*}{4}-\rho)\geq \frac{\rho^*}{4}+\frac{3\rho^*}{4}-\rho=\rho^*-\rho$ and therefore \begin{equation}\label{eq:cvx_lemma2} g(\rho)-g(\rho^*) \geq \frac{\rho^*r}{8}(\rho^*-\rho)\quad \text{for all $\rho < \frac{\rho^*}{2}$}. \end{equation} Combining \eqref{eq:cvx_lemma1} and \eqref{eq:cvx_lemma2} the result follows. \end{proof} We used the following technical lemma in the paper as well. \begin{lemma}\label{lem:technical_convex_bound} Consider the following optimization problem \[ \min f(\bm{\theta}):=\bE_{(\bm{\xi},y)\sim \mathcal{P}_*}\left[\ell(\bm{\xi}^T\bm{\theta})\right]. \] where $\ell$ is a convex function. The sequences $\{\bm{\theta}_k, \bm{\xi}_k \}_{k = 0}^\infty$ generated by Algorithm~\ref{alg:SGD_termination} satisfy for any $\bm{\theta} \in \R^d$, the following \begin{equation} \label{eq:high_decrease} f(\bm{\theta}_{k-1})-f(\bm{\theta})\leq \frac{1}{2\alpha}\left(\Vert \bm{\theta}_{k-1}-\bm{\theta} \Vert^2-\bE\left[\Vert \bm{\theta}_{k}-\bm{\theta} \Vert^2\, |\mathcal{F}_{k-1}\right] \right)+\bE[\Vert\nabla_{\bm{\theta}}\ell\left( \bm{\xi}_k^T\bm{\theta}\right)\Vert^2 \, | \, \mathcal{F}_{k-1}], \end{equation} for all $k\geq 1$ where $f$ is defined in \eqref{optimization_problem} and the filtration $\{\mathcal{F}_k\}_{k=0}^{+\infty}$ in \eqref{eq:sigma}. \end{lemma} \begin{proof Define the quantity \begin{equation*} \bm{g_k}:=\frac{1}{\alpha}\left( \bm{\theta_{k-1}}-\bm{\theta_{k}}\right) = \nabla_{\bm{\theta}}\ell\left(\bm{\xi}_k^T\bm{\theta} \right). \end{equation*} Therefore, we obtain that \[ \bE_{\bm{\xi}_{k}}\left[\bm{g}_k|\mathcal{F}_{k-1} \right]=\nabla_{\bm{\theta}} f(\bm{\theta}_{k-1}) \] By convexity of the function $f$, we have for any $\bm{\theta} \in \R^d$ the following \begin{align*} \norm{\bm{\theta}_{k}-\bm{\theta}}^2 &= \norm{\bm{\theta}_{k-1}-\bm{\theta}}^2 -2 \alpha \bm{g}_k^T (\bm{\theta}_{k-1}-\bm{\theta}) + \alpha^2\norm{\bm{g}_k}^2\\ &= \norm{\bm{\theta}_{k-1}-\bm{\theta}}^2 - 2\alpha(\bm{g}_k-\bE_{\bm{\xi}_{k}}\left[\bm{g}_k|\mathcal{F}_{k-1} \right])^T(\bm{\theta}_{k-1}-\bm{\theta}) - 2 \alpha \bE_{\bm{\xi}_{k}}\left[\bm{g}_k|\mathcal{F}_{k-1} \right]^T (\bm{\theta}_{k-1}-\bm{\theta})+\alpha^2 \norm{\bm{g}_k}^2\\ &\le \norm{\bm{\theta}_{k-1}-\bm{\theta}}^2- 2\alpha(\bm{g}_k-\bE_{\bm{\xi}_{k}}\left[\bm{g}_k|\mathcal{F}_{k-1} \right])^T(\bm{\theta}_{k-1}-\bm{\theta})-2\alpha (f(\bm{\theta}_{k-1})-f(\bm{\theta})) + \alpha^2 \norm{\bm{g}_k}^2. \end{align*} By taking conditional expectations with respect to $\mathcal{F}_{k-1}$ and rearranging the above inequality, the result follows. \end{proof} \section{Background and preliminaries}\label{sec:defs} Throughout we consider a Euclidean space, denoted by $\R^d$, with an inner product and an induced norm $\norm{\cdot}$. The set of non-negative real numbers is denoted by $\R_{\geq 0}$. Bold-faced variables are vectors. Throughout, the matrix $I_d$ is the $d$ by $d$ identity matrix. All stochastic quantities defined hereafter live on a probability space denoted by $(\bP, \Omega, \mathcal{F})$, with probability measure $\bP$ and the $\sigma$-algebra $\mathcal{F}$ containing subsets of $\Omega$. Recall, a random variable (vector) is a measurable map from $\Omega$ to $\R$ ($\R^d$), respectively. An important example of a random variable is the \textit{indicator of the event $A \in \mathcal{F}$}: \[1_A(\omega) = \begin{cases} 1, & \omega \in A\\ 0, & \omega \not \in A. \end{cases}\] If $X$ is a measurable function and $t \in \R$, we often simplify the notation for the pull back of the function $X$, to simply $\{\omega \in \Omega \, :\, X(\omega) \le t\} =: \{X \le t\}$. As is often in probability theory, we will not explicitly define the space $\Omega$, but implicitly define it through random variables. For any sequence of random vectors $(\bm{X}_1, \bm{X}_2, \hdots, \bm{X}_k)$, we denote the \textit{$\sigma$-algebra generated by random vectors $\bm{X}_1, \bm{X}_2, \hdots, \bm{X}_k$} by the notation $\sigma(\bm{X}_1, \bm{X}_2, \bm{X}_3, \hdots, \bm{X}_k)$ and the \textit{expected value of $\bm{X}$} by $\bE[\bm{X}] := \int_{\Omega} \bm{X} \, d\bP$. Particularly, we are interested in random variables that are distributed from normal distributions. In the next section, we state some known results about normal distributions. \paragraph{Normal distributions} The \textit{probability density function of a univariate Gaussian} with mean $\mu$ and variance $\sigma^2$ is described by: \begin{equation*} \varphi(t) := \frac{1}{\sigma\sqrt{2\pi }}\exp\left(-\frac{(t-\mu)^2}{\sigma^2}\right). \end{equation*} In particular, we say a random variable $\xi$ is distributed as a Gaussian with mean $\mu$ and variance $\sigma^2$ by $\xi \sim N(\mu,\sigma^2)$ to mean $\bP(\xi \le t) = \int_{-\infty}^t \varphi(t) \, dt$. When the random variable $\xi \sim N(0,1)$, we denote its cumulative density function as \[\Phi(t) := \bP(\xi \le t) = \frac{1}{\sqrt{2\pi}} \int_{-\infty}^t \exp\left (-\xi^2 \right ) \, d\xi,\] and its complement by $\Phi^c(t) = 1-\Phi(t)$. The symmetry of a normal around its mean yields the identity, $\Phi(t) = \Phi^c(-t)$. One can, analogously, formulate a higher dimensional version of the univariate normal distribution called a \textit{multivariate normal distribution}. A random vector is a multivariate normal distribution if every linear combination of its component is a univariate normal distribution. We denote such multivariate normals by $\bm{\xi} \sim N(\bm{\mu},\Sigma)$ with $\bm{\mu}\in \R^d$ and $\Sigma$ is a symmetric positive semidefinite $d\times d$ matrix. Normal distributions have interesting properties which simplify our computations throughout the paper. We list those which we specifically rely on. See \cite{Famoye_univardist_book} for proofs. Below, $\bm{v},\bm{v}'\in \R^d$, $r\in \R$, $\bm{\xi} \sim N(\bm{\mu},\sigma^2I_d)$ and $\xi\sim N(\mu,\sigma^2)$. Also, $\psi \sim N(0,1)$. Throughout our analysis, we encounter random variables of the form $\bm{v}^T\bm{\xi}+r$, i.e. affine transformations of a given normal distribution. A fundamental property of normal distributions is that they stay in the same class of distributions after any such transformation. In other words, it holds that \begin{equation}\label{eq:fact_affine} \bm{v}^T\bm{\xi}+r\sim N(\bm{v}^T\bm{\mu}+r, \sigma^2\Vert \bm{v} \Vert^2). \end{equation} Working with independent random variables makes the analysis significantly easier. In particular, it is essential for us to know when the two random variables $\bm{v}^T\bm{\xi}$ and $\bm{v}'^T\bm{\xi}$ are independent. We will use the following simple fact below: The following is true \begin{equation} \label{eqn:fact_independence} \mbox{$\bm{v}^T\bm{\xi}$ and $\bm{v}'^T\bm{\xi}$ are independent} \quad \text{ if and only if } \quad \bm{v}^T\bm{v}'=0. \end{equation} We will also use the following simple fact about truncated normal distributions: \begin{equation} \label{eqn:fact_truncated} \bE_{\xi}[\xi1_{\{\xi\leq b\}}]=0 \Longrightarrow \Phi\left(\frac{b-\mu}{\sigma}\right)\cdot \exp\left(\frac{1}{2}\cdot\left(\frac{b-\mu}{\sigma}\right)^2 \right)=\frac{\sigma}{\mu}. \end{equation} We conclude our remarks on normal distributions with the statement of two facts about the expected value of their norm. The following hold: \begin{equation}\label{fact:norm_Gaussians} \bE\left[\Vert \bm{\xi} \Vert^2 \right]= \Vert \bm{\mu} \Vert^2+d\sigma^2, \quad \bE_{\xi}[\vert \xi \vert]\leq \sqrt{\frac{2}{\pi}}\cdot \sigma+\vert\mu\vert \quad \text{and} \quad \bE\left[\vert \psi \vert \right] = \sqrt{\frac{2}{\pi}}. \end{equation} \paragraph{Martingales and stopping times} Here we state some relevant definitions and theorems used in analyzing our stopping criteria in Section~\ref{sec:Analysis}. We refer the reader to \cite{Durrett_probability_book} for further details. For any probability space, $\left(\bP,\Omega,\mathcal{F}\right)$, we call a sequence of $\sigma$-algebras, $\{\mathcal{F}_k\}_{k=0}^\infty$, a \textit{filtration} provided that $\mathcal{F}_i \subset \mathcal{F}$ and $\mathcal{F}_0 \subseteq \mathcal{F}_1 \subseteq \mathcal{F}_2 \subseteq \cdots$ holds. Given a filtration, it is natural to define a sequence of random variables $\{X_k\}_{k=0}^\infty$ with respect to the filtration, namely $X_k$ is a $\mathcal{F}_k$-measurable function. If, in addition, the sequence satisfies \[\bE[|X_k|] < \infty \quad \text{and} \quad \bE[X_{k+1} | \mathcal{F}_k] \le X_k \quad \text{for all $k$},\] we say $\{X_k\}_{k=0}^\infty$ is a \textit{supermartingale}. In probability theory, we are often interested in the (random) time at which a given stochastic sequence exhibits a particular behavior. Such random variables are known as \textit{stopping times}. Precisely, a stopping time is a random variable $T: \Omega \to \mathbb{N} \cup \{0, \infty\}$ where the event $\{T=k\} \in \mathcal{F}_k$ for each $k$, i.e., the decision to stop at time $k$ must be measurable with respect to the information known at that time. Supermartingales and stopping times are closely tied together, as seen in the theorem below, which gives a bound on the expectation of a stopped supermartingale. \begin{theorem}[See \cite{Durrett_probability_book} Theorem 4.8.5]\label{thm:Durret_Martingale} Suppose that $\left\{X_k\right\}_{k=0}^{\infty}$ is a supermartingale w.r.t to the filtration $\left\{\mathcal{F}_k\right\}_{k=0}^{\infty}$ and let $T$ be any stopping time satisfying $\bE[T]<\infty$. Moreover if $\bE\left[\vert X_{k+1}-X_k\vert|\mathcal{F}_k\right]\leq B$ a.s. for some constant $B>0$, then it holds that $\bE[X_T]\leq \bE[X_0]$. \end{theorem} As we illustrate in Section~\ref{sec:Analysis}, a connection between stopping criteria (i.e. the decision to stop an algorithm) and stopping times naturally exists. \subsection{High regime, proof of Theorem \ref{thm:high}}\label{sec:proof_high} In this section, we consider the high variance regime. We consider the target set $C$ and the function $V$ defined in \eqref{eq:targetset_highnoise} and \eqref{eq:driftfunction_highnoise}, respectively, \textit{i.e.} \begin{equation}\label{eq:CV_high_recall} C:=\left\{\bm{\theta}: \vert \rho -\rho^*\vert<\tfrac{1}{2}\rho^* \text{ and } \sigma \Vert \tilde{\bm{\theta}}\Vert \leq c'\right\} \quad \text{and}\quad V(\bm{\theta}):=\frac{1}{2\alpha}\Vert \bm{\theta}-\bm{\theta}^*\Vert^2, \end{equation} where the minimizer $\bm{\theta}^*=\rho^*\bm{\mu}$ is defined in Lemma \ref{lem:minimizers} and the constant $c'$ is to be determined. We first aim to show that $V$ is a drift function with respect to the set $C$ under the high variance regime assumption, meaning $\sigma \geq c\Vert \bm{\mu} \Vert$. We next state a standard SGD convergence result applied to the logistic and hinge loss functions. \begin{lemma}\label{lem:technical_convex_bound} Consider the optimization problem \eqref{optimization_problem} where $\ell:\R\times \R \rightarrow \R$ is either the logistic or hinge loss function. Denote the vector $\bm{\theta}^*$ as the unique minimizer of $f$ in \eqref{optimization_problem}. Let $\bm{\theta}_0\in \R^d$. The sequence $\{\bm{\theta}_k \}_{k = 0}^\infty$ generated by SGD satisfies the following for all $k\geq 1$, \begin{equation} \label{eq:high_convex_tech_lemma} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)\leq \frac{1}{2\alpha}\left(\Vert \bm{\theta}_{k-1}-\bm{\theta} ^*\Vert^2-\bE\left[\Vert \bm{\theta}_{k}-\bm{\theta}^* \Vert^2\, |\mathcal{F}_{k-1}\right] \right)+\frac{\alpha}{2}\left(\Vert \bm{\mu} \Vert^2+d\sigma^2\right). \end{equation} \end{lemma} \begin{proof Define the quantity \begin{equation*} \bm{g_k}:=\frac{1}{\alpha}\left( \bm{\theta}_{k-1}-\bm{\theta}_{k}\right) = \nabla_{\bm{\theta}}\ell\left(\bm{\xi}_k^T\bm{\theta}_{k-1},1 \right). \end{equation*} Here, it is easy to check that the derivative with respect to $\bm{\theta}$ and the expectation over $\bm{\xi}_k$ are interchangeable, thus yielding \[ \bE_{\bm{\xi}_{k}}\left[\bm{g}_k|\mathcal{F}_{k-1} \right]=\nabla_{\bm{\theta}} f(\bm{\theta}_{k-1}). \] By convexity of the function $f$, we have the following \begin{align*} \norm{\bm{\theta}_{k}-\bm{\theta}^*}^2 &= \norm{\bm{\theta}_{k-1}-\bm{\theta}^*}^2 -2 \alpha \bm{g}_k^T (\bm{\theta}_{k-1}-\bm{\theta}^*) + \alpha^2\norm{\bm{g}_k}^2\\ &= \norm{\bm{\theta}_{k-1}-\bm{\theta}^*}^2 - 2\alpha(\bm{g}_k-\bE_{\bm{\xi}_{k}}\left[\bm{g}_k|\mathcal{F}_{k-1} \right])^T(\bm{\theta}_{k-1}-\bm{\theta}^*) - 2 \alpha \bE_{\bm{\xi}_{k}}\left[\bm{g}_k|\mathcal{F}_{k-1} \right]^T (\bm{\theta}_{k-1}-\bm{\theta}^*)+\alpha^2 \norm{\bm{g}_k}^2\\ &\le \norm{\bm{\theta}_{k-1}-\bm{\theta}^*}^2- 2\alpha(\bm{g}_k-\bE_{\bm{\xi}_{k}}\left[\bm{g}_k|\mathcal{F}_{k-1} \right])^T(\bm{\theta}_{k-1}-\bm{\theta}^*)-2\alpha (f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)) + \alpha^2 \norm{\bm{g}_k}^2. \end{align*} By taking conditional expectations with respect to $\mathcal{F}_{k-1}$ and rearranging the above inequality, we obtain that \begin{equation}\label{eq:label_ineq_blah1} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)\leq \frac{1}{2\alpha}\left(\Vert \bm{\theta}_{k-1}-\bm{\theta} ^*\Vert^2-\bE\left[\Vert \bm{\theta}_{k}-\bm{\theta}^* \Vert^2\, |\mathcal{F}_{k-1}\right] \right)+\frac{\alpha}{2}\bE_{\bm{\xi}_k}\left[\Vert \nabla_{\bm{\theta}} \ell(\bm{\xi}_k^T\bm{\theta}_{k-1},1)\Vert^2 \right]. \end{equation} We next observe the following bound \begin{equation}\label{eq:label_ineq_blah2} \bE_{\bm{\xi}_k}[\Vert\nabla_{\bm{\theta}}\ell\left( \bm{\xi}_k^T\bm{\theta}_{k-1},1\right)\Vert^2 \, | \, \mathcal{F}_{k-1}]\leq \bE_{\bm{\xi}_k}[\Vert \bm{\xi}_k \Vert^2|\mathcal{F}_{k-1}]=\Vert \bm{\mu} \Vert^2+d\sigma^2. \end{equation} Combining \eqref{eq:label_ineq_blah1} and \eqref{eq:label_ineq_blah2}, the result follows. \end{proof} By Lemma \ref{lem:technical_convex_bound} for each $k\geq 1$, we deduce \begin{equation}\label{eq:driftequation_high} \bE[V(\bm{\theta}_k)|\mathcal{F}_{k-1}]-V(\bm{\theta}_{k-1})\leq -\left(f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)\right)+\frac{\alpha}{2}(\Vert\bm{\mu}\Vert^2+d\sigma^2). \end{equation} Therefore, in order to show that the pair $(C,V)$ in \eqref{eq:CV_high_recall} satisfies the drift equation \eqref{eq:driftequation}, it suffices to lower bound the quantity $f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)$ whenever $\bm{\theta}_{k-1}\not\in C$. To do so, we orthogonally decompose $\bm{\theta}_{k-1}=\rho_{k-1}\bm{\mu}+\tilde{\bm{\theta}}_{k-1}$, \textit{i.e.} $\bm{\mu}^T\tilde{\bm{\theta}}_{k-1}=0$ and $\rho_{k-1}\in \R$ and write \begin{equation}\label{eq:quantity_split} \begin{aligned} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)= & \underbrace{f(\bm{\theta}_{k-1})-f(\rho_{k-1}\bm{\mu})}_{(a)}+\underbrace{f(\rho_{k-1}\bm{\mu})-f(\bm{\theta}^*)}_{(b)}. \end{aligned} \end{equation} The assumption $\bm{\theta}_{k-1}\not \in C$ yields that either $\sigma \Vert \tilde{\bm{\theta}}_{k-1}\Vert \geq c'$ or $\vert \rho_{k-1}-\rho^*\vert\geq \tfrac{1}{2}\rho^*$. In Lemma \ref{lem:lower_bound_1} (resp. \ref{lem:lower_bound_2}), we show that (a) (resp. (b)) in \eqref{eq:quantity_split} are both non-negative and they are lower bounded by some positive constant provided that $\sigma \Vert \tilde{\bm{\theta}}_{k-1}\Vert \geq c'$ and $\vert \rho_{k-1}-\rho^*\vert\leq \tfrac{1}{2}\rho^*$ (resp. $\vert \rho_{k-1}-\rho^*\vert\geq \tfrac{1}{2}\rho^*$). \begin{lemma}(Lower bound for (a) in \eqref{eq:quantity_split})\label{lem:lower_bound_1} Fix $\bm{\theta}\in \R^d$ and orthogonally decompose $\bm{\theta}=\rho\bm{\mu}+\tilde{\bm{\theta}}$ where $\bm{\mu}^T\tilde{\bm{\theta}}=0$ and $\rho \in \R$. Then the following are true \begin{enumerate} \item $f(\bm{\theta})-f(\rho\bm{\mu})\geq 0$. \item $f(\bm{\theta})-f(\rho\bm{\mu})\geq 1$ provided that $\vert \rho-\rho^*\vert\leq \tfrac{1}{2}\rho^*$, $\sigma \Vert \tilde{\bm{\theta}}\Vert \geq c'$ and $\sigma \geq c\Vert \bm{\mu} \Vert$ where $c$ is defined in \eqref{eq:para_log} and \eqref{eq:para_hinge}. Here $\rho^*$ is defined in Lemma \ref{lem:minimizers} and the constant $c'$ is defined by $436$ and $8+10\rho^*\sigma^2$ for the logistic and hinge loss respectively. \end{enumerate} \end{lemma} \begin{proof} We consider the logistic and hinge loss separately. \item \textbf{Logistic loss.} The two normal random variables, $\bm{\tilde{\theta}}^T\bm{\xi} \sim N(0,\sigma^2\Vert \bm{\tilde{\theta}} \Vert^2)$ and $\bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu} \Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$, are independent by \eqref{eqn:fact_independence}. Since we have $\bE_{\bm{\xi}}[\log(\exp(-\bm{\tilde{\theta}}^T\bm{\xi}))]=\bE_{\bm{\xi}}[\log(\exp(\bm{\tilde{\theta}}^T\bm{\xi}))]=0$, it holds \begin{align*} f(\bm{\theta})=\bE_{\bm{\xi}} \left[ \log\left (1+\exp(-\bm{\theta}^T\bm{\xi})\right)\right] &= \bE_{\bm{\xi}} \left[\log \left(1 + \exp(-\tilde{\bm{\theta}}^T\bm{\xi}) \exp(-\rho \bm{\mu}^T\bm{\xi}) \right ) \right ]\\ &=\bE_{\bm{\xi}} \left[ \log\left (\exp(\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)\right]\\ &=\bE_{\bm{\xi}} \left[ \log\left (\exp(-\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)\right], \end{align*} where the last equality is true because $\bm{\tilde{\theta}}^T\bm{\xi} \sim -\bm{\tilde{\theta}}^T\bm{\xi}$. Therefore we obtain \begin{align*} \bE_{\bm{\xi}} &\left[ \log\left (1+\exp(-\bm{\theta}^T\bm{\xi})\right)\right]\\ & \qquad \qquad= \frac{1}{2} \bE_{\bm{\xi}} \left[ \log\left (\exp(\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)\right] + \frac{1}{2}\bE_{\bm{\xi}} \left[ \log\left (\exp(-\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)\right] \\ & \qquad \qquad= \frac{1}{2} \bE_{\bm{\xi}} \left[ \log\left ( (\exp(\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi}))(\exp(-\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})) \right)\right]\\ &\qquad \qquad=\frac{1}{2}\bE_{\bm{\xi}} \left[\log\left(1+\exp(-\bm{\tilde{\theta}}^T\bm{\xi}-\rho \bm{\mu}^T\bm{\xi})+\exp(\bm{\tilde{\theta}}^T\bm{\xi}-\rho \bm{\mu}^T\bm{\xi})+\exp(-2\rho \bm{\mu}^T\bm{\xi})\right) \right]. \end{align*} By the equality $\exp(\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\bm{\tilde{\theta}}^T\bm{\xi})= 2+4\sinh^2(\tfrac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})$, we have \begin{align*} \bE_{\bm{\xi}} &\left[ \log\left (1+\exp(-\bm{\theta}^T\bm{\xi})\right)\right]\\ & \qquad \qquad \qquad =\frac{1}{2}\bE_{\bm{\xi}} \left[\log\left(1+2\exp(-\rho \bm{\mu}^T\bm{\xi})+\exp(-2\rho \bm{\mu}^T\bm{\xi})+4\sinh^2(\tfrac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})\exp(-\rho \bm{\mu}^T\bm{\xi})\right) \right]. \end{align*} Therefore, we have \begin{equation} \begin{aligned} \label{eqn: high_noise_blah_20} 2\bE_{\bm{\xi}}\left[\log\left( \frac{1+\exp(-\bm{\theta}^T\bm{\xi})}{1+\exp(-\rho \bm{\mu}^T\bm{\xi})}\right)\right]&=2\bE_{\bm{\xi}}\left[\log(1+\exp(-\bm{\theta}^T\bm{\xi}))\right]-\bE_{\bm{\xi}}\left[\log\left(1+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)^2\right]\\ \qquad \qquad &= \bE_{\bm{\xi}}\left[\log \left ( 1 + \frac{4\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})\exp(-\rho \bm{\mu}^T\bm{\xi})}{(1+ \exp(-\rho \bm{\mu}^T\bm{\xi}))^2} \right ) \right] \ge 0. \end{aligned} \end{equation} Thereby, we showed that $f(\bm{\theta})-f(\rho \bm{\mu})\geq 0$. Now we establish the positive lower bound. First, we note the following $1+ \exp(-\rho \bm{\mu}^T\bm{\xi}) = 2 \exp( -\tfrac{\rho \bm{\mu}^T\bm{\xi}}{2}) \cosh(\tfrac{\rho \bm{\mu}^T\bm{\xi}}{2})$. Fix a constant $r > 0$ and consider the set $\{\bm{\xi}: |\bm{\theta}^T\bm{\xi}| > r\}$. Applying the inequality $x^2+y^2\geq 2\vert xy \vert$ and \eqref{eqn: high_noise_blah_20}, we obtain that \begin{align} 2\bE_{\bm{\xi}}\left[\log\left( \frac{1+\exp(-\bm{\theta}^T\bm{\xi})}{1+\exp(-\rho \bm{\mu}^T\bm{\xi})}\right)\right] & = \bE_{\bm{\xi}}\left[\log\left(1 + \frac{4\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})\exp(-\rho \bm{\mu}^T\bm{\xi})}{(1+ \exp(-\rho \bm{\mu}^T\bm{\xi}))^2}\right) \right] \nonumber \\ &= \bE_{\bm{\xi}}\left[\log\left(1+\frac{\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})}{\cosh^2(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})} \right) \right] \nonumber \\ &\geq \bE_{\bm{\xi}}\left[\log\left(1+\frac{\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})}{\cosh^2(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})} \right)\cdot1_{\{\bm{\xi}:\vert \tilde{\bm{\theta}}^T\bm{\xi}\vert \geq r\}}\right] \label{eq:high_noise_blah_21} \\ &\geq \bE_{\bm{\xi}}\left[\left(\log2+\log\left( \frac{\vert\sinh(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})\vert}{\cosh(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})}\right) \right)\cdot1_{\{\bm{\xi}:\vert \tilde{\bm{\theta}}^T\bm{\xi}\vert \geq r\}}\right] \nonumber. \end{align} Here \eqref{eq:high_noise_blah_21} follows from $\log\left(1+\frac{\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})}{\cosh^2(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})} \right)$ is always positive. From \eqref{eq:fact_affine}, we have $\bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu} \Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$ and $\tilde{\bm{\theta}}^T\bm{\xi} \sim N(0,\sigma^2\Vert \tilde{\bm{\theta}} \Vert^2)$, so $\tilde{\bm{\theta}}^T\bm{\xi} = \sigma \| \tilde{\bm{\theta}} \| \psi$ where $\psi \sim N(0,1)$. Moreover, a simple computation shows that $-\log\left(\cosh(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})\right)1_{\{\vert \tilde{\bm{\theta}}^T\bm{\xi}\vert \geq r\}} \geq -\log\left(\cosh(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})\right) $ since $\cosh(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})\geq 1$ always holds. Using the inequality $\log\cosh(x)\leq \vert x\vert$ for $x$, the following bound holds \begin{equation} \begin{aligned} \label{eq: high_noise_22} &\bE_{\bm{\xi}}\left[\log\left( \frac{1+\exp(-\bm{\theta}^T\bm{\xi})}{1+\exp(-\rho \bm{\mu}^T\bm{\xi})}\right)\right]\\ & \quad \ge \tfrac{1}{2} \log(2) \cdot \bE_{\psi}\left[1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] + \tfrac{1}{2} \bE_{\psi}\left[\log\left\vert\sinh(\tfrac{\sigma\Vert \tilde{\bm{\theta}}\Vert\psi}{2})\right\vert1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] - \tfrac{1}{2}\bE_{\bm{\xi}} \left [ \log( \cosh(\tfrac{\rho}{2} \bm{\mu}^T\bm{\xi})) \right ]\\ & \quad \ge \tfrac{1}{2} \log(2) \cdot \bE_{\psi}\left[1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] + \tfrac{1}{2} \bE_{\psi}\left[\log\left\vert\sinh(\tfrac{\sigma\Vert \tilde{\bm{\theta}}\Vert\psi}{2})\right\vert1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] - \tfrac{1}{2} \bE_{\bm{\xi}} \left [ | \tfrac{\rho}{2} \bm{\mu}^T\bm{\xi} | \right ]\\ &\quad\geq \tfrac{1}{2} \log(2) \cdot \bE_{\psi}\left[1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] + \tfrac{1}{2} \bE_{\psi}\left[\log\left\vert\sinh(\tfrac{\sigma\Vert \tilde{\bm{\theta}}\Vert\psi}{2})\right\vert1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right]-\frac{3}{4}\left(\frac{\Vert \bm{\mu}\Vert^2}{\sigma^2}+ \sqrt{\frac{2}{\pi}}\cdot\frac{\Vert \bm{\mu} \Vert}{\sigma} \right), \end{aligned} \end{equation} where the last inequality uses \eqref{fact:norm_Gaussians} and $\rho\leq \frac{3}{\sigma^2}$. Using the inequality $\vert \sinh(x) \vert \geq \exp(\frac{\vert x\vert}{2})$ for all $\vert x \vert \geq 2\log(\sqrt{2}+1)$ and letting $r=4\log(\sqrt{2}+1)$, we obtain \begin{align} \tfrac{1}{2} \log(2) \cdot \bE_{\psi} \big [ 1_{\{|\psi| \ge \frac{4 \log( \sqrt{2}+ 1)\}}{\sigma \norm{\tilde{\bm{\theta}}} }} \big ] &+ \tfrac{1}{2}\bE_\psi\left[\log\left\vert\sinh(\tfrac{\sigma\Vert \tilde{\bm{\theta}}\Vert\psi}{2})\right\vert1_{\{|\psi| \geq \frac{4\log(\sqrt{2}+1)}{\sigma \Vert \tilde{\bm{\theta}}\Vert}\}}\right]\nonumber\\ &\geq \tfrac{1}{2} \log(2) \cdot \bE_{\psi} \big [ 1_{\{|\psi| \ge \frac{4 \log( \sqrt{2}+ 1)}{\sigma \norm{\tilde{\bm{\theta}}} } \}} \big ] + \tfrac{1}{2}\bE_\psi\left[\left\vert\tfrac{\sigma \Vert \tilde{\bm{\theta}}\Vert \psi}{4} \right\vert1_{\{|\psi| \geq \frac{4\log(\sqrt{2}+1)}{\sigma \Vert \tilde{\bm{\theta}}\Vert}\}}\right]\nonumber\\ &\geq \tfrac{1}{2} \log(2) \cdot \bE_{\psi}[1_{\{\vert\psi\vert \ge 1\}}] + \tfrac{1}{2} \bE_\psi\left[\left\vert\tfrac{\sigma \Vert \tilde{\bm{\theta}}\Vert \psi}{4} \right\vert1_{\vert\psi\vert \geq 1\}}\right] \label{eq:high_noise_blah_111} \\\nonumber &\geq \left ( \tfrac{1}{2} \log(2) + \frac{\sigma \Vert \tilde{\bm{\theta}}\Vert}{8} \right )\cdot\Phi^c(1). \end{align} Here \eqref{eq:high_noise_blah_111} follows from the assumption that $\sigma \Vert \tilde{\bm{\theta}}\Vert\geq 436 $. Combining \eqref{eq: high_noise_22}, \eqref{eq:high_noise_blah_111} and the bounds $\sigma \geq 0.33\Vert \bm{\mu} \Vert$ and $\sigma\Vert \tilde{\bm{\theta}}\Vert \geq 436$ the result follows. \item \textbf{Hinge loss.} We begin by denoting $\xi_1:=\bm{\xi}^T\tilde{\bm{\theta}}$ and $\xi_2:=\bm{\xi}^T\bm{\mu}$. Notice that $\xi_1$ and $\xi_2$ are independent random variables. Recall that $\ell(t):=\ell(t,1)=\max(0,1-t)$. We have that \begin{align*} f(\bm{\theta})-f(\rho \bm{\mu})&=\bE_{\bm{\xi}}\left[\ell(\bm{\xi}^T\bm{\theta})-\ell(\rho\bm{\xi}^T\bm{\mu}) \right]\\&=\bE_{\xi_1,\xi_2}\left[\ell(\xi_1+\rho\xi_2)-\ell(\rho\xi_2) \right]\\&=\bE_{\xi_1,\xi_2}\left[\ell(-\xi_1+\rho\xi_2)-\ell(\rho\xi_2) \right]. \end{align*} The second equality follows since $\xi_1 \sim -\xi_1$. We define the function \[ \kappa(\xi_1,\xi_2):=\ell(\xi_1+\rho\xi_2)+\ell(-\xi_1+\rho\xi_2)-2\ell(\rho\xi_2). \] We therefore obtain that \[ 2\left(f(\bm{\theta})-f(\rho\bm{\mu})\right)=\bE_{\xi_1,\xi_2}\left[\kappa(\xi_1,\xi_2)\right]. \] Next we claim that \begin{equation}\label{eq:hinge_lemma_claim1} \kappa(\xi_1,\xi_2)=0 \text{ whenever } \vert \xi_1\vert\leq \vert 1-\rho\xi_2\vert. \end{equation} To see this, suppose that $\vert \xi_1\vert\leq \vert 1-\rho\xi_2\vert$ holds. We consider two cases. First, assume that $0\leq 1-\rho\xi_2$ which yields that $\rho\xi_2-\xi_1\leq 1$ and $\rho\xi_2+\xi_1\leq 1$. We therefore have $\kappa(\xi_1,\xi_2)=1-\xi_1-\rho\xi_2+1+\xi_1-\rho\xi_2-2(1-\rho\xi_2)=0$. Second, assume that $1-\rho\xi_2\leq 0$. It thus holds that $1\leq \rho\xi_2-\xi_1$ and $1\leq \rho\xi_2+\xi_1$. Now it immediately follows that $\kappa(\xi_1,\xi_2)=0$ and equation \eqref{eq:hinge_lemma_claim1} is established. We claim the following \begin{equation}\label{eq:hinge_lemma_claim2} \kappa(\xi_1,\xi_2)=\vert \xi_1\vert-\vert 1-\rho\xi_2\vert \text{ whenever } \vert \xi_1\vert\geq \vert 1-\rho\xi_2\vert. \end{equation} To this end, we again consider two cases. First, assume that $\xi_1\leq -\vert 1-\rho\xi_2\vert$. This yields that $1\leq -\xi_1+\rho\xi_2$ and $\xi_1+\rho\xi_2\leq 1$, so it holds that $ \kappa(\xi_1,\xi_2)= 1-\xi_1-\rho\xi_2-2\ell(\rho\xi_2)$. The claim \eqref{eq:hinge_lemma_claim2} follows from the following simple identity \begin{equation}\label{eq:hinge_identity} 2\ell(t)= 1-t+\vert 1 - t \vert, \quad \forall t \in \R. \end{equation} Second, assume that $\xi_1\geq \vert 1-\rho\xi_2\vert$. It then holds that $\xi_1+\rho\xi_2\geq 1$ and $-\xi_1+\rho\xi_2\leq 1$ and therefore $ \kappa(\xi_1,\xi_2)=1+\xi_1-\rho\xi_2-2\ell(\rho \xi_2). $ The claim \eqref{eq:hinge_lemma_claim2} follows from the identity \eqref{eq:hinge_identity}. We therefore obtain \begin{align} \bE_{\xi_1,\xi_2}[\kappa(\xi_1,\xi_2)]&=2\bE_{\xi_1,\xi_2}[(\ell(\xi_1+\rho\xi_2)+\ell(-\xi_1+\rho\xi_2)-2\ell(\rho\xi_2))1_{\{ \xi_1>0\}}]\label{eq:hinge_lemma_it1}\\&=2\bE_{\xi_1,\xi_2}[(\ell(\xi_1+\rho\xi_2)+\ell(-\xi_1+\rho\xi_2)-2\ell(\rho\xi_2))1_{\{ \xi_1\geq\vert 1-\rho \xi_2\vert\}}]\label{eq:hinge_lemma_it2}\\&=\bE_{\xi_1,\xi_2}[(\xi_1-\vert 1-\rho \xi_2\vert)1_{\{ \xi_1\geq \vert 1-\rho\xi_2\vert\}} ]. \label{eq:hinge_lemma_it3} \end{align} Here equation \eqref{eq:hinge_lemma_it1} holds because $\xi_1\sim -\xi_1$ and $\kappa(\xi_1,\xi_2)=\kappa(-\xi_1,\xi_2)$. Equation \eqref{eq:hinge_lemma_it2} is true because of claim \eqref{eq:hinge_lemma_claim1} and \eqref{eq:hinge_lemma_it3} follows from claim \eqref{eq:hinge_lemma_claim2}. From \eqref{eq:hinge_lemma_it3}, we conclude that $\bE_{\xi_1,\xi_2}[\kappa(\xi_1,\xi_2)]\geq 0$. We then observe the bound \begin{equation} \begin{aligned} \bE_{\xi_1,\xi_2}[(\xi_1-\vert 1-\rho \xi_2\vert)1_{\{ \xi_1\geq \vert 1-\rho\xi_2\vert\}} ] &= \tfrac{1}{2}\bE_{\xi_1,\xi_2}[\xi_1-\vert 1-\rho \xi_2\vert+\vert \xi_1-\vert 1-\rho \xi_2\vert \vert]\\&\geq -\tfrac{1}{2}\bE_{\xi_2}[\vert 1-\rho \xi_2\vert]+\tfrac{1}{2}\bE_{\xi_1,\xi_2}[\vert \xi_1\vert-\vert 1-\rho \xi_2\vert]\\&=\tfrac{1}{2}\bE_{\xi_1}[\vert \xi_1\vert]-\bE_{\xi_2}[\vert1-\rho\xi_2\vert]. \label{eq:thrid_last_equation} \end{aligned} \end{equation} The second inequality follows from $\bE_{\xi_1}[\xi_1]=0$ and the triangle inequality $\vert x\vert-\vert y\vert \leq \vert \vert x\vert-y\vert$. On the other hand, it holds that \begin{equation}\label{first_last_equation} \bE_{\xi_1}[\vert \xi_1\vert]=\sqrt{\frac{2}{\pi}}\cdot\sigma \Vert \tilde{\bm{\theta}}\Vert, \end{equation} and \begin{equation}\label{second_last_equation} \bE_{\bm{\xi}}[\vert 1-\rho \bm{\mu}^T\bm{\xi}\vert]\leq 1+\rho \bE_{\bm{\xi}}[\vert \bm{\mu}^T\bm{\xi}\vert]\leq 1+\rho\Vert \bm{\mu} \Vert \left(\sqrt{\frac{2}{\pi}}\cdot\sigma +\Vert \bm{\mu} \Vert \right). \end{equation} Combing equations \eqref{eq:hinge_lemma_claim1}, \eqref{eq:hinge_lemma_claim2}, \eqref{eq:thrid_last_equation}, \eqref{first_last_equation}, and \eqref{second_last_equation}, we deduce \begin{equation}\label{eq:last_hinge_high} f(\bm{\theta})-f(\rho \bm{\mu}) \geq \frac{1}{2}\left(\sqrt{\frac{1}{2\pi}}\cdot\sigma\Vert\tilde{\bm{\theta}}\Vert-1-\rho\Vert \bm{\mu} \Vert \left(\sqrt{\frac{2}{\pi}}\cdot\sigma +\Vert \bm{\mu} \Vert \right) \right). \end{equation} Using the bounds $\sigma\Vert \tilde{\bm{\theta}}\Vert \geq 8+10\rho^*\sigma^2$, $\sigma \geq 0.62\Vert \bm{\mu} \Vert$ and $\rho\leq \tfrac{3}{2}\rho^*$, the result follows from \eqref{eq:last_hinge_high}. \end{proof} We next derive a lower bound \eqref{eq:quantity_split}, Part (b). But, first we need a basic lemma from convex analysis. \begin{lemma}\label{lem:convex_analysis_lemma} Suppose that $g:\R_{\geq 0}\to \R$ is a convex function with a minimizer at $\rho^*>0$. Assume that $g$ is twice differentiable on the interval $[\tfrac{3}{4}\rho^*,\tfrac{5}{4}\rho^*]$ and there exists a constant $B>0$ such that $g''(\rho)\geq B$ for all $\rho \in [\tfrac{3}{4}\rho^*,\tfrac{5}{4}\rho^*]$. Then it holds that \begin{equation}\label{eq0:cvx_lemma} g(\rho)-g(\rho^*)\geq \frac{\rho^*B}{8}\vert \rho-\rho^*\vert \quad \text{for all $\rho \not\in [\tfrac{1}{2}\rho^*,\tfrac{3}{2}\rho^*]$}. \end{equation} \end{lemma} \begin{proof} The proof follows by considering the second order Taylor series expansion of the function $g$. \end{proof} \begin{lemma}(Lower bound for (b) in \eqref{eq:quantity_split})\label{lem:lower_bound_2} Fix $\bm{\theta}\in \R^d$ and orthogonally decompose $\bm{\theta}=\rho\bm{\mu}+\tilde{\bm{\theta}}$. Suppose that $\vert \rho-\rho^*\vert \geq \tfrac{1}{2}\rho^*$. Then provided that $\sigma \geq c\Vert \bm{\mu} \Vert$ where the constant $c$ is defined in \eqref{eq:para_log} and \eqref{eq:para_hinge}, there exists a positive constant $A$ such that the following is true \begin{equation}\label{eq:lemma9} f(\rho \bm{\mu})-f(\bm{\theta^*}) \geq A\cdot\frac{\Vert \bm{\mu} \Vert^2}{\sigma^2}. \end{equation} \begin{proof} We consider the logistic and hinge loss separately. \item \textbf{Logistic loss.} Define the function \[ g(\rho):=\bE_{\bm{\xi}}\left[\log\left(1+\exp(-\rho \bm{\mu}^T\bm{\xi})\right) \right], \quad \bm{\xi} \sim N(\bm{\mu},\sigma^2I_d). \] By Lemma~\ref{lem:minimizers}, we know that $g$ is a convex function with a unique minimizer at $\rho^*:=\frac{2}{\sigma^2}$. Observe that $f(\rho \bm{\mu})-f(\bm{\theta}^*) = g(\rho)-g(\rho^*)$; hence in order to prove \eqref{eq:lemma9}, we instead aim to bound this difference in the function $g$. From \eqref{eq:fact_affine}, we have $\bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu} \Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$. It thus holds \[ 4g''(\rho)=\bE\left(\frac{(\bm{\mu}^T\bm{\xi})^2}{\cosh(\frac{\rho}{2} \bm{\mu}^T\bm{\xi})^2} \right)=\frac{1}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi }}\int_{-\infty}^{\infty}\frac{z^2}{\cosh^2(\frac{\rho z}{2})}\exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2}\right)dz. \] Upper bounding $\cosh^2(\frac{\rho z}{2})$ by $\exp(\vert \rho z\vert)$, we next obtain \begin{align*} 4g''(\rho)&\geq \frac{1}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi}} \int_{-\infty}^{\infty} z^2\exp\left(-\vert \rho z \vert \right)\exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu}\Vert^2}\right)dz & \\ &=\frac{1}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi }} \int_{-\infty}^{\infty} z^2\exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2+2\sigma^2\Vert \bm{\mu}\Vert^2\vert\rho z\vert}{2\sigma^2\Vert \bm{\mu}\Vert^2}\right)dz ,&\\ &=\frac{1}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi }}\cdot\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right) \int_{-\infty}^{\infty} z^2\exp\left(-\frac{z^2-2\Vert \bm{\mu} \Vert^2 z+2\sigma^2\Vert \bm{\mu}\Vert^2\vert\rho z\vert}{2\sigma^2\Vert \bm{\mu}\Vert^2}\right)dz \\&= \frac{\sigma^2\Vert \bm{\mu}\Vert^2}{\sqrt{2\pi }}\cdot\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right)\int_{-\infty}^{+\infty} z^2\exp\left(-\frac{z^2-2\frac{\Vert \bm{\mu} \Vert}{\sigma}z+2\vert \rho z \vert \sigma\Vert\bm{\mu}\Vert }{2} \right) dz\\&\geq \frac{\sigma^2\Vert \bm{\mu}\Vert^2}{\sqrt{2\pi }}\cdot\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right)\int_{0}^{+\infty} z^2\exp\left(-\frac{z^2}{2}\right)\exp\left(z\left(\frac{\Vert\bm{\mu}\Vert}{\sigma}-\rho \sigma \Vert \bm{\mu} \Vert\right) \right)dz\\&\geq \frac{\sigma^2\Vert \bm{\mu}\Vert^2}{\sqrt{2\pi }}\cdot\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}-\frac{1}{2}-\left \vert \frac{\Vert \bm{\mu} \Vert}{\sigma}-\rho \sigma \Vert \bm{\mu} \Vert \right\vert \right)\int_{0}^1 z^2 dz. \end{align*} Here the second to last inequality follows from the change of variables $z\rightarrow z\sigma\Vert \bm{\mu}\Vert$. The last inequality follows from restricting the integral's domain to $[0,1]$ and also lower bounding $-\frac{z^2}{2}$ and $z\left(\frac{\Vert\bm{\mu}\Vert}{\sigma}-\rho \sigma \Vert \bm{\mu} \Vert\right)$ by $\frac{-1}{2}$ and $-\left\vert\frac{\Vert\bm{\mu}\Vert}{\sigma}-\rho \sigma \Vert \bm{\mu} \Vert\right\vert$ respectively. We see that, for $\rho \in [\tfrac{3}{4}\rho^*,\tfrac{5}{4}\rho^*]$, the term $\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}-\frac{1}{2}-\left \vert \frac{\Vert \bm{\mu} \Vert}{\sigma}-\rho \sigma \Vert \bm{\mu} \Vert \right\vert \right)$ is lower bounded by $\exp\left(-\frac{1}{2c^2}-\frac{1}{4c}-\frac{1}{2} \right)$. By Lemma \ref{lem:convex_analysis_lemma}, the result follows with the constant $A$ computed as follows \begin{equation*}\label{eq:constant_A_log} A=\frac{1}{12\sqrt{2\pi}}\cdot\exp\left(-\frac{1}{2c^2}-\frac{1}{4c}-\frac{1}{2} \right). \end{equation*} \item \textbf{Hinge loss.} We begin by defining the function $h(\rho)=f(\rho\bm{\mu})$. Therefore \[ f(\rho \bm{\mu})=\bE_{\bm{\xi}}[\ell(\rho\bm{\xi}^T\bm{\mu})]=\bE_{\bm{\xi}}[(1-\rho \bm{\xi}^T\bm{\mu})1_{\{\rho\bm{\xi}^T\bm{\mu}\leq 1 \}}]. \] Hence, it holds that \[ h'(\rho)=\bm{\mu}^T\nabla f(\rho \bm{\mu})=-\bE_{\bm{\xi}}[\bm{\xi}^T\bmu1_{\{ \rho\bm{\xi}^T\bm{\mu}\leq 1\}}]. \] From \eqref{eq:fact_affine}, we obtain that $\bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu} \Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$. For $\rho>0$, therefore, it holds that \begin{equation}\label{der_phi_pos} h'(\rho)=\frac{-1}{\sigma\Vert \bm{\mu} \Vert\sqrt{2\pi}}\int_{-\infty}^{\frac{1}{\rho}}z\exp\left(-\frac{1}{2}\cdot\left(\frac{z}{\sigma\Vert \bm{\mu} \Vert}-\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)^2\right)dz. \end{equation} Applying chain rule thus yields \[ h''(\rho)=\frac{1}{\rho^3\sigma\Vert \bm{\mu} \Vert\sqrt{2\pi}}\exp\left(-\frac{1}{2}\cdot \left(\frac{1}{\rho\sigma\Vert \bm{\mu} \Vert}-\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)^2 \right)\quad \text{for all $\rho>0$}. \] Hence, for all $\rho \in [\tfrac{3}{4}\rho^*,\tfrac{5}{4}\rho^*]$ it holds that \[ h''(\rho)\geq \frac{64}{125{\rho^*}^3\sigma\Vert \bm{\mu} \Vert\sqrt{2\pi}}\exp\left(-\frac{1}{2}\cdot \Gamma^2 \right), \] where $\Gamma:=\max\left\{\left|\frac{4}{3\rho^*\sigma\Vert \bm{\mu} \Vert}-\frac{\Vert \bm{\mu} \Vert}{\sigma} \right|,\left|\frac{4}{5\rho^*\sigma\Vert \bm{\mu} \Vert}-\frac{\Vert \bm{\mu} \Vert}{\sigma} \right| \right\}$. Therefore, by Lemma \ref{lem:convex_analysis_lemma} and $\vert \rho-\rho^*\vert \geq \tfrac{1}{2}\rho^*$, it holds that \begin{equation}\label{eq:constant_r_hinge_loss} \begin{aligned} f(\rho\bm{\mu})-f(\bm{\theta}^*)&\geq \frac{4}{125\sqrt{2\pi}}\cdot\frac{\sigma}{r\Vert \bm{\mu} \Vert}\cdot\exp\left(-\frac{1}{2}\cdot \Gamma^2 \right). \end{aligned} \end{equation} Here $r=\rho^*\sigma^2$. Note that $r>0$ by Lemma \ref{lem:minimizers}. We aim to lower bound the right-hand side of \eqref{eq:constant_r_hinge_loss}. We denote by $w=\frac{\sigma}{r\Vert \bm{\mu} \Vert}-\frac{\Vert \bm{\mu} \Vert}{\sigma}$ the quantity defined in Lemma \ref{lem:minimizers}. In particular, by Lemma \ref{lem:minimizers}, the following holds \begin{equation}\label{eq:w_hinge_later} \frac{1}{\sqrt{2\pi}}\cdot\frac{\sigma}{\Vert \bm{\mu} \Vert}=\Phi(w)\cdot \exp(\tfrac{1}{2}w^2). \end{equation} We consider two cases. First suppose that $w\geq \frac{1}{(3\sqrt{2}-4)c}$. Along with the assumption $\frac{\sigma}{\Vert \bm{\mu} \Vert}\geq c$ this implies that $w\geq \frac{1}{3\sqrt{2}-4}\cdot \frac{\Vert \bm{\mu} \Vert}{\sigma}$. A simple computation shows that $w^2\geq \frac{1}{2}\cdot \Gamma^2$ for all $w\geq \frac{1}{3\sqrt{2}-4}\cdot \frac{\Vert \bm{\mu} \Vert}{\sigma}$. On the other hand, by \eqref{eq:w_hinge_later} for $w\geq 0$, we obtain that $\frac{2}{\pi}\cdot \frac{\sigma^2}{\Vert \bm{\mu} \Vert^2}\geq \exp(w^2)$. Plugging in the bounds $w^2\geq \frac{1}{2}\cdot \Gamma^2 $, $\exp(-w^2)\geq \frac{\pi}{2}\cdot\frac{\Vert \bm{\mu} \Vert^2}{\sigma^2}$, and $\frac{\sigma}{r\Vert \bm{\mu} \Vert}\geq w\geq \frac{1}{(3\sqrt{2}-4)c}$ into the right-hand-side of \eqref{eq:constant_r_hinge_loss}, we obtain that \begin{equation*} f(\rho\bm{\mu})-f(\bm{\theta}^*)\geq \frac{\sqrt{2\pi}}{125(3\sqrt{2}-4)c}\cdot\frac{\Vert \bm{\mu} \Vert^2}{\sigma^2}. \end{equation*} Next, suppose that $w<\frac{1}{(3\sqrt{2}-4)c}$. In this case, the two factors $\frac{\sigma}{r\Vert \bm{\mu} \Vert}$ and $\exp\left(-\frac{1}{2}\cdot\Gamma^2 \right)$ in \eqref{eq:constant_r_hinge_loss} are lower bounded separately. Note that it always holds that $w\geq -\frac{\Vert \bm{\mu} \Vert}{\sigma}$ as $r>0$. Therefore, it is easy to see that the latter factor is lower bounded by $\exp\left(-\frac{1}{2} \left(\frac{4}{3(3\sqrt{2}-4)c}+\frac{1}{3c} \right)^2\right)$. Hence, it remains to bound the factor $\frac{\sigma}{r\Vert \bm{\mu} \Vert}$ in \eqref{eq:constant_r_hinge_loss}. To this end, we show that $w\geq -\frac{\Vert \bm{\mu} \Vert}{2\sigma}$ for all $\frac{\sigma}{\Vert \bm{\mu} \Vert}\geq c$. Note that a chain of change of variables gives \begin{equation*}\label{eq:myeq_blah} \begin{aligned} \Phi(w)\cdot\exp\left(\frac{w^2}{2}\right)&=\frac{1}{\sqrt{2\pi}}\cdot\int_{0}^{+\infty}\exp(-\frac{1}{2}t^2)\cdot\exp(wt) \, dt. \end{aligned}\end{equation*} The right-hand side of \eqref{eq:w_hinge_later} is an increasing function with respect to $w$. Therefore it suffices to show that the following holds \begin{equation}\label{eq:sigma/bmu>c} \frac{1}{\sqrt{2\pi}}\cdot \frac{\sigma}{\Vert \bm{\mu} \Vert} \geq \Phi\left(-\frac{\Vert \bm{\mu} \Vert}{2\sigma} \right)\cdot \exp\left(\frac{\Vert \bm{\mu} \Vert^2}{8\sigma^2} \right) \quad \text{whenever} \quad \frac{\sigma}{\Vert \bm{\mu} \Vert}\geq c. \end{equation} However, it can be verified by a plot that $\tfrac{1}{\sqrt{2\pi}}\geq t\cdot \Phi\left(-\tfrac{t}{2} \right)\cdot \exp\left(\frac{t^2}{8} \right)$ holds for all $t\in (0,\frac{1}{c})$. Therefore, we have shown that $w\geq -\frac{\Vert \bm{\mu} \Vert}{2\sigma}$ which implies that $\frac{\sigma}{r\Vert \bm{\mu} \Vert}\geq \frac{\Vert \bm{\mu} \Vert}{2\sigma}$. Finally we lower bound the quantity $\frac{\sigma}{r\Vert\bm{\mu}\Vert}$ by $c\cdot\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}$. We have concluded \eqref{eq:lemma9} in case of hinge loss function where the constant $A$ can be computed as follows \begin{equation*}\label{eq:constant_A_hinge} A=\min\left\{\frac{c}{2}\cdot\exp\left(-\frac{1}{2} \left(\frac{4}{3(3\sqrt{2}-4)c}+\frac{1}{3c} \right)^2\right), \frac{\sqrt{2\pi}}{125(3\sqrt{2}-4)c} \right\}. \end{equation*} \end{proof} \end{lemma} We now have the ingredients to prove Theorem \ref{thm:high}. \begin{proof}[Proof of Theorem \ref{thm:high}] Consider the set $C$ and function $V$ defined in \eqref{eq:CV_high_recall}: \begin{equation}\label{eq:CV_high_proofthm3} C:=\left\{\bm{\theta}: \vert \rho -\rho^*\vert<\tfrac{1}{2}\rho^* \text{ and } \sigma \Vert \tilde{\bm{\theta}}\Vert \leq c'\right\} \quad \text{ and }\quad V(\bm{\theta})=\frac{1}{2\alpha}\Vert \bm{\theta}-\bm{\theta}^*\Vert^2. \end{equation} We let $c'$ to be defined as in Lemma \ref{lem:lower_bound_1}. This means that $c'$ equals to $436$ and $8+10\rho^*\sigma^2$ in case of logistic and hinge loss respectively. We next show that there exists a positive constant $\delta$ such that the following is true \begin{equation} \bP_{\bm{\xi}}\left(\bm{\xi}^T\bm{\theta}\geq 1\right)\geq \delta \quad \text{for all}\quad \bm{\theta} \in C. \end{equation} Let $\bm{\theta}\in C$ and orthogonally decompose it into $\bm{\theta}=\rho\bm{\mu}+\tilde{\bm{\theta}}$. We have that $\bm{\xi}^T\bm{\theta}=\rho\bm{\xi}^T\bm{\mu}+\bm{\xi}^T\tilde{\bm{\theta}}$. Note that $\rho>0$ as $\bm{\theta}\in C$. By \eqref{eqn:fact_independence}, we see that $\bm{\xi}^T\bm{\theta}$ and $\bm{\xi}^T\tilde{\bm{\theta}}$ are independent normal random variables. It thus holds that \begin{equation} \bP_{\bm{\xi}}\left(\bm{\xi}^T\bm{\theta}\geq 1\right)\geq \bP_{\bm{\xi}}\left(\rho\bm{\xi}^T\bm{\mu}\geq 1\right)\cdot \bP_{\bm{\xi}}\left(\bm{\xi}^T\tilde{\bm{\theta}}\geq 0\right) =\frac{1}{2}\cdot\bP_{\bm{\xi}}\left(\bm{\xi}^T\bm{\mu}\geq \frac{1}{\rho}\right). \end{equation} Rewrite the inequality $\bm{\xi}^T\bm{\mu}\geq \frac{1}{\rho}$ by $z:=\frac{\bm{\xi}^T\bm{\mu}-\Vert \bm{\mu} \Vert^2}{\sigma\Vert \bm{\mu} \Vert}\geq \frac{\frac{1}{\rho}-\Vert \bm{\mu} \Vert^2}{\sigma\Vert \bm{\mu} \Vert}$. Noting that $z\sim N(0,1)$ and using the inequality $\frac{2}{\rho^*}\geq\frac{1}{\rho}$, we obtain that \begin{equation}\label{eq:delta_proof} \bP_{\bm{\xi}}\left(\bm{\xi}^T\bm{\theta}\geq 1\right)\geq \delta:= \frac{1}{2}\cdot\Phi^c\left(\frac{\frac{2}{\rho^*}-\Vert\bm{\mu} \Vert^2}{\sigma\Vert \bm{\mu} \Vert}\right). \end{equation} We next show that the pair $(C,V)$ satisfies the drift equation \eqref{eq:driftequation}. Let us rewrite \eqref{eq:quantity_split}: \begin{equation}\label{eq:quantity_split2} \begin{aligned} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)= & \underbrace{f(\bm{\theta}_{k-1})-f(\rho_{k-1}\bm{\mu})}_{(a)}+\underbrace{f(\rho_{k-1}\bm{\mu})-f(\bm{\theta}^*)}_{(b)}. \end{aligned} \end{equation} By Lemmas \ref{lem:lower_bound_1} and \ref{lem:lower_bound_2}, both terms in $(a)$ and $(b)$ in \eqref{eq:quantity_split2} are non-negative . Assume that $\bm{\theta}_{k-1}\not\in C$. Therefore, either $\sigma \Vert \tilde{\bm{\theta}}_{k-1}\Vert\geq c'$ or $\vert\rho_{k-1}-\rho^*\vert\geq \tfrac{1}{2}\rho^*$; this implies that the quantity $(a)$ is at least 1 or the quantity $(b)$ is at least $A\cdot\frac{\Vert \bm{\mu} \Vert^2}{\sigma^2}$ respectively. The constant $A$ in Lemma \ref{lem:lower_bound_2} satisfies $1\geq A\cdot \frac{\Vert \bm{\mu} \Vert^2}{\sigma^2}$ for all $\frac{\sigma}{\Vert \bm{\mu} \Vert}\geq c$. Hence it holds that \begin{equation} A\cdot\frac{\Vert \bm{\mu} \Vert^2}{\sigma^2}\leq f(\bm{\theta}_{k-1})-f(\bm{\theta}^*) \quad \text{for all } \bm{\theta}_{k-1}\not\in C. \end{equation} We use \eqref{eq:high_convex_tech_lemma} next to establish the drift equation \eqref{eq:driftequation}. Recall that the following holds \begin{equation} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)\leq \frac{1}{2\alpha}\left(\Vert \bm{\theta}_{k-1}-\bm{\theta} ^*\Vert^2-\bE\left[\Vert \bm{\theta}_{k}-\bm{\theta}^* \Vert^2\, |\mathcal{F}_{k-1}\right] \right)+\frac{\alpha}{2}\left(\Vert \bm{\mu} \Vert^2+d\sigma^2\right). \end{equation} Combining the last two displayed inequalities and using the definition of function $V$, we obtain that \begin{equation} \left(\bE\left[V(\bm{\theta}_k)|\mathcal{F}_{k-1}\right]-V(\bm{\theta}_{k-1})\right)\cdot 1_{\{\bm{\theta}_{k-1}\not\in C\}} \leq \left(\frac{\alpha}{2}(\Vert \bm{\mu} \Vert^2+d\sigma^2)-A\cdot\frac{\Vert \bm{\mu}\Vert^2}{\sigma^2}\right)\cdot 1_{\{\bm{\theta}_{k-1}\not\in C\}}. \end{equation} Therefore, by choosing $\alpha<A\cdot\frac{\Vert \bm{\mu} \Vert^2}{\sigma^2(\Vert \bm{\mu} \Vert^2+d\sigma^2)}$, we obtain the drift equation \eqref{eq:driftequation} holds with $b:=\frac{A}{2}\cdot\frac{\Vert \bm{\mu} \Vert^2}{\sigma^2}$. Next, we obtain bounds on $\bE[\tau_m]$ for $m\geq 1$. By Lemma \ref{lem:drift_from_meyn} and a simple induction, we obtain that \begin{equation} \bE[\tau_m]\leq \tfrac{1}{b}V(0)+\tfrac{1}{b}(m-1)\sup_{\bm{\theta}\in C} V(\bm{\theta}). \end{equation} Compactness of set $C$ yields that, $\sup_{\bm{\theta}\in C} V(\bm{\theta})<+\infty$. Therefore, for some constant $\gamma$, the following is true \begin{equation}\label{eq:bound_on_tau_m_high} \bE[\tau_m] \leq \gamma\cdot m. \end{equation} Combining \eqref{eq:bound_on_tau_m_high}, \eqref{eq:delta_proof} and Lemma \ref{lem:ETleqET_C}, the proof immediately follows. \end{proof} \subsection{Hinge Regression} In this section, we study the following optimization problem \begin{align*} \min f(\bm{\theta}):=\bE_{\bm{\xi} \sim \mathcal{P}_*} [\ell(\bm{\xi}^T\bm{\theta})], \end{align*} where $\ell(x):=\ell(x,1)$ with the hinge loss function defined in \eqref{eq:hinge_loss_definition}. We state the main theorem of this section below. We defer the proof to the end of this subsection. \begin{theorem}\label{theorem_hinge} Let $\bm{\theta}_0=\bm{0}$. The following are true. \begin{enumerate} \item Fix $\epsilon\in (0,\tfrac{1}{2})$ and let $M=1+\frac{\alpha(\Vert \bm{\mu} \Vert^2+\sigma^2)}{2\epsilon}$. Suppose that $\sigma \leq \left(\tfrac{1}{2}-\epsilon\right)\sqrt{\tfrac{\pi}{2}}\Vert \bm{\mu} \Vert$ (low variance regime). The following then holds. \begin{equation}\begin{aligned} \label{eq:low_Et_proof} \bE[T]&\leq \left(2+\frac{\left(M-1\right)^2}{\alpha\Vert \bm{\mu} \Vert^2}+\frac{\sigma}{\Vert \bm{\mu} \Vert}\left(\frac{1.6M^2}{\alpha \Vert \bm{\mu} \Vert^2}+7,612\sigma^2\alpha^4\Vert\bm{\mu} \Vert^4\right) \right)+\frac{2M^2}{\alpha \Vert\bm{\mu}\Vert^2}. \end{aligned} \end{equation} \item Suppose that $\sqrt{\frac{\pi}{8}}\Vert \bm{\mu} \Vert \leq \sigma$ (high variance regime). It holds that $\bE[T]<+\infty$ provided the step-size $\alpha$ satisfies \begin{equation} \alpha< \frac{\min\{1,\tfrac{1}{2}r\rho^*\}}{\Vert \bm{\mu} \Vert^2+d\sigma^2} \end{equation} where the constant $r$ is defined as \begin{equation*} r:= \frac{64}{125 \rho^*\sigma \Vert \bm{\mu} \Vert \sqrt{2\pi}}\exp\left(-\frac{(\tfrac{4}{3}\frac{1}{\rho^*}-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2} \right). \end{equation*} \end{enumerate} \end{theorem} \subsubsection{Low Variance Regime}\label{sec:low_hinge} Throughout this section, we fix a positive constant $0<\epsilon<\tfrac{1}{2}$ and consider the following drift function. \[ V(\bm{\theta}):=\frac{1}{\alpha \Vert \bm{\mu}\Vert^2}(M-\bm{\mu}^T\bm{\theta})^2, \] where $M=1+\frac{\alpha(\Vert \bm{\mu} \Vert^2+\sigma^2)}{2\epsilon}$. In addition, we consider the following target set. \begin{equation}\label{eq:hinge_low_C} C:=\{\bm{\theta}:\bm{\mu}^T\bm{\theta} \geq 1\}. \end{equation} \begin{proposition}\label{prop:low_noise_hinge_Etau1} Fix $\epsilon\in(0,\tfrac{1}{2})$ and suppose that $\sigma \leq \left(\tfrac{1}{2}-\epsilon\right)\sqrt{\tfrac{\pi}{2}}\Vert \bm{\mu} \Vert$. Let $\bm{\theta}\in \R^d$ such that $\bm{\mu}^T\bm{\theta}\leq 1$. The following is then true \begin{equation} \bE[\tau_1^C|\bm{\theta}_0=\bm{\theta}]\leq V(\bm{\theta}). \end{equation} \end{proposition} \begin{proof} Suppose that $\bm{\mu}^T\bm{\theta}_{k-1}\leq 1$. It, then, holds that \begin{align} \bE[(M-\bm{\mu}^T\bm{\theta}_k)^2|\mathcal{F}_{k-1}]1_{\{\tau_1^C\wedge n\geq k \}}&= \bE_{\bm{\xi}_k}[(M-\bm{\mu}^T\bm{\theta}_{k-1}-\alpha 1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}\bm{\mu}^T\bm{\xi}_k)^2|\mathcal{F}_{k-1}]1_{\{\tau_1^C\wedge n\geq k \}}\\&=(M-\bm{\mu}^T\bm{\theta}_{k-1})^2-2\alpha(M-\bm{\mu}^T\bm{\theta}_{k-1})\bE_{\bm{\xi}_k}[1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}\bm{\mu}^T\bm{\xi}_k|\mathcal{F}_{k-1}]\\&+\alpha^2\bE_{\bm{\xi}_k}[1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}(\bm{\mu}^T\bm{\xi}_k)^2|\mathcal{F}_{k-1}]. \end{align} We need to lower bound the quantity $\bE_{\bm{\xi}_k}[1_{\{ \bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}\bm{\mu}^T\bm{\xi}_k|\mathcal{F}_{k-1}]$. We have \begin{align} \bE_{\bm{\xi}_k}[1_{\{ \bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}\bm{\mu}^T\bm{\xi}_k|\mathcal{F}_{k-1}]&=\Vert \bm{\mu}\Vert^2\bE_{\bm{\xi}_k}[1_{\{ \bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}|\mathcal{F}_{k-1}]+\bE_{\bm{\xi}_k}[1_{\{ \bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}\bm{\mu}^T(\bm{\xi}_k-\bm{\mu})|\mathcal{F}_{k-1}]\\&\geq \tfrac{1}{2}\Vert \bm{\mu} \Vert^2-\bE_{\bm{\xi}_k}[\vert \bm{\mu}^T(\bm{\xi}_k-\bm{\mu})]\vert]\\&\geq \tfrac{1}{2}\Vert \bm{\mu} \Vert^2-\sqrt{\tfrac{2}{\pi}}\sigma\Vert \bm{\mu} \Vert. \end{align} Here we used that $1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq \bm{\mu}^T\bm{\theta}_{k-1}\}}\leq 1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}$ and $\bE_{\bm{\xi}_k}[1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq \bm{\mu}^T\bm{\theta}_{k-1}\}}]=\tfrac{1}{2}$. Continuing, we obtain that \begin{align*} \bE[(M-\bm{\mu}^T\bm{\theta}_k)^2-(M-\bm{\mu}^T\bm{\theta}_{k-1})^2&|\mathcal{F}_{k-1}]1_{\{\tau_1^C\wedge n\geq k \}}\\ &\leq \alpha^2 \bE_{\bm{\xi}_k}[(\bm{\mu}^T\bm{\xi}_k)^2]-2\alpha(M-\bm{\mu}^T\bm{\theta}_{k-1})(\tfrac{1}{2}\Vert \bm{\mu} \Vert^2-\sqrt{\tfrac{2}{\pi}}\sigma\Vert \bm{\mu} \Vert)\\&\leq \alpha^2\Vert \bm{\mu}\Vert^2(\Vert \bm{\mu} \Vert^2+\sigma^2)-2\alpha(M-\bm{\mu}^T\bm{\theta}_{k-1})\left(\tfrac{1}{2}\Vert \bm{\mu} \Vert^2-\sqrt{\tfrac{2}{\pi}}\sigma\Vert \bm{\mu} \Vert\right) \end{align*} Hence, using the bounds $ \sigma\leq \left(\tfrac{1}{2}-\epsilon\right)\sqrt{\tfrac{\pi}{2}}\Vert \bm{\mu} \Vert$ and $M\geq 1+\frac{\alpha(\Vert \bm{\mu} \Vert^2+\sigma^2)}{2\epsilon}$, we obtain that \begin{align*} \bE[V(\bm{\theta}_k)-V(\bm{\theta}_{k-1})|\mathcal{F}_{k-1}]1_{\{\tau_1^C\wedge n\geq k \}}\leq \alpha (\Vert \bm{\mu} \Vert^2+\sigma^2)-2(M-\bm{\mu}^T\bm{\theta}_{k-1})\left(\tfrac{1}{2}-\sqrt{\tfrac{2}{\pi}}\frac{\sigma}{\Vert \bm{\mu} \Vert}\right)\leq -1_{\{\tau_1^C\wedge n\geq k \}}. \end{align*} Using Appendix Lemma \ref{lem:drift_lemma_appendix}, we obtain that $\bE[\tau_1^C\wedge n|\mathcal{F}_{-1}] \leq V(\bm{\theta})$. Monotone convergence theorem completes the proof. \end{proof} We next upper bound the expected value of $\tau_m^C$ when we initialize with $\bm{\theta}_0=\bm{0}$. \textcolor{red}{Is the proof of this proposition the same as Proposition 2 in logistic} \textcolor{blue}{Sina:They have a lot in common! but not entirely the same. It would be best if we make this one shorter.} \begin{proposition}\label{prp:hinge_low_etm} Let $\bm{\theta}_0=\bm{0}$ and suppose that $\sigma\leq \left(\tfrac{1}{2}-\epsilon\right)\sqrt{\tfrac{\pi}{2}}\Vert \bm{\mu} \Vert$. Consider the set $C$ defined in \eqref{eq:hinge_low_C}. Then the following bound holds for all $m \ge 1$ \begin{equation} \bE[\tau_m^C]\leq (m-1)\left(2+\frac{\left(M-1\right)^2}{\alpha\Vert \bm{\mu} \Vert^2}+\frac{\sigma}{\Vert \bm{\mu} \Vert}\left(\frac{1.6M^2}{\alpha \Vert \bm{\mu} \Vert^2}+7,612\sigma^2\alpha^4\Vert\bm{\mu} \Vert^4\right) \right)+\frac{M^2}{\alpha \Vert\bm{\mu}\Vert^2}, \end{equation} where $M=1+\frac{\alpha(\Vert \bm{\mu} \Vert^2+\sigma^2)}{2\epsilon}$. \end{proposition} \begin{proof} Similar to \eqref{eq: low_noise_bound}, we obtain that \begin{equation}\label{eq:low_noise_hinge1} \bE\left[(\tau_m^{C}-\tau_{m-1}^{C})\wedge n|\mathcal{F}_{\tau_{m-1}^{C}+1}\right]\leq 2+\sum_{i=1}^{\infty}\bE\left[\tau_1^{C}\wedge n|\bm{\theta}_0=\bm{\theta}_{\tau_{m-1}^{C}+1}\right]1_{\{i-1\leq 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}<i\}}. \end{equation} For each $i \ge 2$, we observe the bound \begin{align*} i-1 \leq 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1} &=1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}}-\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^C+1}1_{\{\bm{\xi}_{\tau_{m-1}^C+1}^T\bm{\theta}_{\tau_{m-1}^C}\leq 1\}}\\&\leq -\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^C+1}1_{\{\bm{\xi}_{\tau_{m-1}^C+1}^T\bm{\theta}_{\tau_{m-1}^C}\leq 1\}}\\&\leq -\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^C+1}. \end{align*} Here the first inequality follows since $1\leq \bm{\mu}^T\bm{\theta}_{\tau_{m-1}^C}$ and the second inequality from the fact that $0\leq -\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^C+1}$ as $i-1> 0$. We next, similar to our computation in \eqref{eq: low_noise_blah_3} and \eqref{eq: low_noise_blah_4}, have that \begin{equation} \begin{aligned} \label{eq: low_noise_hinge2} \bE\left[\tau_1^{C}\wedge n|\bm{\theta}_0=\bm{\theta}_{\tau_{m-1}^{C}+1}\right]1_{\{i-1\leq 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}<i\}}&\leq \frac{(M+i-1)^2}{\alpha \Vert \bm{\mu}\Vert^2}1_{\{ \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^{C}+1}\leq \frac{1-i}{\alpha}\}} \quad \forall i\geq 2. \end{aligned} \end{equation} Moreover, \begin{equation} \begin{aligned} \label{eq: low_noise_hinge_3} \bE\left[1_{\{ \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^{C}+1}\leq \frac{1-i}{\alpha}\}} \right]&=\Phi\left(\frac{\frac{1-i}{\alpha}-\Vert \bm{\mu}\Vert^2}{\sigma\Vert \bm{\mu}\Vert } \right). \end{aligned} \end{equation} Combining \eqref{eq:low_noise_hinge1}, \eqref{eq: low_noise_hinge2} and \eqref{eq: low_noise_hinge_3} we obtain that \begin{equation}\label{eq:low_noise_hinge_blah12} \bE\left[(\tau_m^{C}-\tau_{m-1}^{C})\wedge n|\mathcal{F}_{\tau_{m-1}^{C}+1}\right]\leq 2+\frac{\left(M-1\right)^2}{\alpha\Vert \bm{\mu} \Vert^2}+\frac{\sigma}{\Vert \bm{\mu} \Vert\sqrt{2\pi}}\sum_{i=2}^{+\infty}\frac{\left(M+i-1\right)^2}{\alpha \Vert \bm{\mu} \Vert^2+i-1}\exp\left(-\tfrac{1}{2}\left(\frac{\Vert \bm{\mu} \Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu} \Vert^2}\right)^2 \right). \end{equation} Note that we used the bound $\bE\left[\tau_1^{C}\wedge n|\bm{\theta}_0=\bm{\theta}_{\tau_{m-1}^{C}+1}\right]1_{\{0\leq 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}<1\}}\leq \frac{(M-1)^2}{\alpha \Vert\bm{\mu}\Vert^2}$. Continuing, similar to \eqref{eq: low_noise_blah_6} and \eqref{eq: low_noise_blah_7}, we have \begin{equation} \sum_{i=3}^{+\infty}\frac{\left(M+i-1\right)^2}{\alpha \Vert \bm{\mu} \Vert^2+i-1}\exp\left(-\tfrac{1}{2}\left(\frac{\Vert \bm{\mu} \Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu} \Vert^2}\right)^2 \right) \leq 19,080 \alpha^4 \sigma^4 \Vert \bm{\mu} \Vert^4. \end{equation} We upper bound the term in \eqref{eq:low_noise_hinge_blah12} for $i=2$ by $\frac{4M^2}{\alpha \Vert \bm{\mu} \Vert^2}$. Finally we conclude that \begin{equation} \bE\left[\tau_m^C-\tau_{m-1}^C\wedge n \right] \leq 2+\frac{\left(M-1\right)^2}{\alpha\Vert \bm{\mu} \Vert^2}+\frac{\sigma}{\Vert \bm{\mu} \Vert}\left(\frac{1.6M^2}{\alpha \Vert \bm{\mu} \Vert^2}+7,612\sigma^2\alpha^4\Vert\bm{\mu} \Vert^4\right). \end{equation} We, therefore, obtain that \begin{equation} \bE\left[\tau_m^C\right] \leq (m-1)\left(2+\frac{\left(M-1\right)^2}{\alpha\Vert \bm{\mu} \Vert^2}+\frac{\sigma}{\Vert \bm{\mu} \Vert}\left(\frac{1.6M^2}{\alpha \Vert \bm{\mu} \Vert^2}+7,612\sigma^2\alpha^4\Vert\bm{\mu} \Vert^4\right) \right)+\bE\left[\tau_1^C\right] \end{equation} The result follows after applying Proposition \ref{prop:low_noise_hinge_Etau1} with $\bm{\theta}_0=\bm{0}$. \end{proof} \begin{proof}[Proof of Theorem 3.1] Using Proposition \ref{prp:hinge_low_etm} and Lemma \ref{lem:stopping_time_appendix}, Theorem \ref{theorem_hinge}.1 immediately follows. \end{proof} \subsubsection{High Variance Regime}\label{sec:high_hinge} The main theorem of this section is as follows. \begin{theorem}\label{thm:hinge_loss_high_noise} Suppose that the step-size $\alpha$ satisfies \begin{equation} \alpha< \frac{\min\{1,\tfrac{1}{2}r\rho^*\}}{\Vert \bm{\mu} \Vert^2+d\sigma^2} \end{equation} where $\rho^*$ is defined in Lemma \ref{lem:hinge_minimizer} and the constant $r$ is defined as \begin{equation*} r:= \frac{64}{125 \rho^*\sigma \Vert \bm{\mu} \Vert \sqrt{2\pi}}\exp\left(-\frac{(\tfrac{4}{3}\frac{1}{\rho^*}-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2} \right). \end{equation*} Then it holds that $\bE[T]<+\infty$. \end{theorem} We refer the proof to the end of this section. We first need a technical lemma. \begin{lemma}\label{lem:hinge_loss_high_noise_fundamental_lemma} For any $\bm{\theta}$ the following holds \begin{equation}\label{eq:lemma_for_high_noise_hinge} f(\bm{\theta}_{k-1})-f(\bm{\theta}) \leq \frac{1}{2\alpha} \left(\Vert \bm{\theta}_{k-1}-\bm{\theta}\Vert^2-\bE\left[\Vert \bm{\theta}_k-\bm{\theta}\Vert^2 |\mathcal{F}_{k-1}\right]\right)+\frac{\alpha}{2}\left(\Vert \bm{\mu} \Vert^2+d\sigma^2\right). \end{equation} \end{lemma} \begin{proof} See Appendix, Lemma \ref{lem:technical_convex_bound}. Note that we are using the inequality $\Vert \nabla_{\bm{\theta}}\ell(\bm{\xi}^T\bm{\theta})\Vert^2\leq \Vert \bm{\xi} \Vert^2$ and then applying \eqref{fact:norm_Gaussians}. \end{proof} We write the orthogonal complement of $\bm{\theta}\in \R^d$ as $\bm{\theta}=\rho\bm{\mu}+\tilde{\bm{\theta}}$. As before, the idea is to lower bound the left hand side of \eqref{eq:lemma_for_high_noise_hinge} for $\bm{\theta}=\bm{\theta}^*:=\rho^* \bm{\mu}$. To do so, we write $f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)=f(\bm{\theta}_{k-1})-f(\rho\bm{\mu})+f(\rho \bm{\mu})-f(\bm{\theta}^*)$ and then provide bounds for the both terms $f(\bm{\theta}_{k-1})-f(\rho_{k-1} \bm{\mu})$ and $f(\rho_{k-1}\bm{\mu})-f(\rho^*\bm{\mu})$. The following lemma bounds $f(\bm{\theta}_{k-1})-f(\rho \bm{\mu})$. \begin{lemma}\label{lem:hinge_high_noise_first_lemma} For each $\bm{\theta}\in \R^d$, the following bound holds \begin{equation} f(\bm{\theta})-f(\rho\bm{\mu})\geq \max\left\{\tfrac{1}{\sqrt{2\pi}}\sigma \Vert \tilde{\bm{\theta}}\Vert-\rho \Vert \bm{\mu} \Vert\left(\sigma\sqrt{\tfrac{2}{\pi}}+\Vert \bm{\mu} \Vert \right)-1,0 \right\}. \end{equation} \end{lemma} \begin{proof} We begin by denoting $\xi_1:=\bm{\xi}^T\tilde{\bm{\theta}}$ and $\xi_2:=\bm{\xi}^T\bm{\mu}$. Notice that $\xi_1$ and $\xi_2$ are independent random variables. We have that \begin{align*} f(\bm{\theta})-f(\rho \bm{\mu})&=\bE_{\bm{\xi}}\left[\ell(\bm{\xi}^T\bm{\theta})-\ell(\rho\bm{\xi}^T\bm{\mu}) \right]\\&=\bE_{\bm{\xi}}\left[\ell(\xi_1+\rho\xi_2)-\ell(\rho\xi_2) \right]\\&=\bE_{\bm{\xi}}\left[\ell(-\xi_1+\rho\xi_2)-\ell(\rho\xi_2) \right]. \end{align*} The second equality follows since $\xi_1 \sim -\xi_1$. We define the function \[ \Phi(\xi_1,\xi_2):=\ell(\xi_1+\rho\xi_2)+\ell(-\xi_1+\rho\xi_2)-2\ell(\rho\xi_2). \] We, therefore, obtain that \[ 2\left(f(\bm{\theta})-f(\rho\bm{\mu})\right)=\bE_{\xi_1,\xi_2}\left[\Phi(\xi_1,\xi_2)\right]. \] We, next, claim that \begin{equation}\label{eq:hinge_lemma_claim1} \Phi(\xi_1,\xi_2)=0 \text{ whenever } \vert \xi_1\vert\leq \vert 1-\rho\xi_2\vert. \end{equation} To see this, suppose that $\vert \xi_1\vert\leq \vert 1-\rho\xi_2\vert$ holds. We consider two cases. First, assume that $0\leq 1-\rho\xi_2$. Then, it holds that $\rho\xi_2-\xi_1\leq 1$ and $\rho\xi_2+\xi_1\leq 1$. We then have $\Phi(\xi_1,\xi_2)=1-\xi_1-\rho\xi_2+1+\xi_1-\rho\xi_2-2(1-\rho\xi_2)=0$. Second, assume that $1-\rho\xi_2\leq 0$. Then, it holds that $1\leq \rho\xi_2-\xi_1$ and $1\leq \rho\xi_2+\xi_1$. Now it immediately follows that $\Phi(\xi_1,\xi_2)=0$. We have thus established \eqref{eq:hinge_lemma_claim1}. We, next, claim the following \begin{equation}\label{eq:hinge_lemma_claim2} \Phi(\xi_1,\xi_2)=\vert \xi_1\vert-\vert 1-\rho\xi_2\vert \text{ whenever } \vert \xi_1\vert\geq \vert 1-\rho\xi_2\vert. \end{equation} To this end, we again consider two cases. First, assume that $\xi_1\leq -\vert 1-\rho\xi_2\vert$. It, then, holds that $1\leq -\xi_1+\rho\xi_2$ and $\xi_1+\rho\xi_2\leq 1$. We, therefore, have $ \Phi(\xi_1,\xi_2)= 1-\xi_1-\rho\xi_2-2\ell(\rho\xi_2)$. The claim \eqref{eq:hinge_lemma_claim2} follows from the following simple identity \begin{equation}\label{eq:hinge_identity} 2\ell(t)= 1-t+\vert 1 - t \vert, \quad \forall t \in \R. \end{equation} Second, assume that $\xi_1\geq \vert 1-\rho\xi_2\vert$. It, then, holds that $\xi_1+\rho\xi_2\geq 1$ and $-\xi_1+\rho\xi_2\leq 1$. We, therefore, have $ \Phi(\xi_1,\xi_2)=1+\xi_1-\rho\xi_2-2\ell(\rho \xi_2). $ The claim \eqref{eq:hinge_lemma_claim2} follows from the identity \eqref{eq:hinge_identity}. We, therefore, obtain \begin{align} \bE_{\xi_1,\xi_2}[\Phi(\xi_1,\xi_2)]&=2\bE_{\xi_1,\xi_2}[(\ell(\xi_1+\rho\xi_2)+\ell(-\xi_1+\rho\xi_2)-2\ell(\rho\xi_2))1_{\{ \xi_1>0\}}]\label{eq:hinge_lemma_it1}\\&=2\bE_{\xi_1,\xi_2}[(\ell(\xi_1+\rho\xi_2)+\ell(-\xi_1+\rho\xi_2)-2\ell(\rho\xi_2))1_{\{ \xi_1\geq\vert 1-\rho \xi_2\vert\}}]\label{eq:hinge_lemma_it2}\\&=\bE_{\xi_1,\xi_2}[(\xi_1-\vert 1-\rho \xi_2\vert)1_{\{ \xi_1\geq \vert 1-\rho\xi_2\vert\}} ]. \label{eq:hinge_lemma_it3} \end{align} Here, \eqref{eq:hinge_lemma_it1} holds because $\xi_1\sim -\xi_1$ and $\Phi(\xi_1,\xi_2)=\Phi(-\xi_1,\xi_2)$. Equation \eqref{eq:hinge_lemma_it2} is true because of claim \eqref{eq:hinge_lemma_claim1} and \eqref{eq:hinge_lemma_it3} follows from claim \eqref{eq:hinge_lemma_claim2}. From \eqref{eq:hinge_lemma_it3}, we clearly conclude that $\bE_{\xi_1,\xi_2}[\Phi(\xi_1,\xi_2)]\geq 0$. We, next, observe the bound \begin{align} \bE_{\xi_1,\xi_2}[(\xi_1-\vert 1-\rho \xi_2\vert)1_{\{ \xi_1\geq \vert 1-\rho\xi_2\vert\}} ] &= \tfrac{1}{2}\bE_{\xi_1,\xi_2}[\xi_1-\vert 1-\rho \xi_2\vert+\vert \xi_1-\vert 1-\rho \xi_2\vert \vert]\nonumber\\&\geq -\tfrac{1}{2}\bE_{\xi_2}[\vert 1-\rho \xi_2\vert]+\tfrac{1}{2}\bE_{\xi_1,\xi_2}[\vert \xi_1\vert-\vert 1-\rho \xi_2\vert]\nonumber\\&=\tfrac{1}{2}\bE_{\xi_1}[\vert \xi_1\vert]-\bE_{\xi_2}[\vert1-\rho\xi_2\vert]. \label{eq:thrid_last_equation} \end{align} The second inequality follows from $\bE_{\xi_1}[\xi_1]=0$ and the triangle inequality $\vert x\vert-\vert y\vert \leq \vert \vert x\vert-y\vert$. On the other hand, it holds that \begin{equation}\label{first_last_equation} \bE_{\xi_1}[\vert \xi_1\vert]=\sqrt{\tfrac{2}{\pi}}\sigma \Vert \tilde{\bm{\theta}}\Vert, \end{equation} and \begin{equation}\label{second_last_equation} \bE_{\bm{\xi}}[\vert 1-\rho \bm{\mu}^T\bm{\xi}\vert]\leq 1+\rho \bE_{\bm{\xi}}[\vert \bm{\mu}^T\bm{\xi}\vert]\leq 1+\rho\Vert \bm{\mu} \Vert \left(\sigma \sqrt{\tfrac{2}{\pi}}+\Vert \bm{\mu} \Vert \right). \end{equation} Combing equations \eqref{eq:thrid_last_equation} , \eqref{first_last_equation} and \eqref{second_last_equation} concludes the proof. \end{proof} In turn, we bound $f(\rho \bm{\mu})-f(\rho^*\bm{\mu})$ in the next lemma. \begin{lemma}\label{lem:hinge_loss_second_lemma} There exists a constant $r>0$ such that the following holds \begin{equation} f(\rho \bm{\mu})-f(\rho^*\bm{\mu}) \geq r\vert \rho -\rho^*\vert \quad \text{for any $\rho$ satisfying $\vert \rho-\rho^*\vert >\tfrac{1}{2}\rho^*$ }. \end{equation} \end{lemma} \begin{proof} We begin by defining the function $h(\rho)=f(\rho\bm{\mu})$. Recall that \[ f(\rho \bm{\mu})=\bE_{\bm{\xi} \sim \mathcal{P}_*}[\ell(\rho\bm{\xi}^T\bm{\mu})]=\bE_{\bm{\xi} \sim \mathcal{P}_*}[(1-\rho \bm{\xi}^T\bm{\mu})1_{\{\rho\bm{\xi}^T\bm{\mu}\leq 1 \}}] \] Hence, it holds that \[ h'(\rho)=\bm{\mu}^T\nabla f(\rho \bm{\mu})=-\bE_{\bm{\xi} \sim \mathcal{P}_*}[\bm{\xi}^T\bmu1_{\{ \rho\bm{\xi}^T\bm{\mu}\leq 1\}}]. \] From \eqref{eq:fact_affine}, we obtain that $\bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu} \Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$. For $\rho>0$, therefore, it holds that \begin{equation}\label{der_phi_pos} h'(\rho)=\frac{-1}{\sigma\Vert \bm{\mu} \Vert\sqrt{2\pi}}\int_{-\infty}^{\frac{1}{\rho}}z\exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2)}{2\sigma^2\Vert \bm{\mu} \Vert^2} \right)dz. \end{equation} Applying chain rule thus yields \[ h''(\rho)=\frac{1}{\rho^3\sigma\Vert \bm{\mu} \Vert\sqrt{2\pi}}\exp\left(-\frac{(\frac{1}{\rho}-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2} \right) \quad \text{for all $\rho>0$}. \] Clearly, there exists $r>0$ such that for $\rho\in [\tfrac{3}{4}\rho^*,\tfrac{5}{4}\rho^*]$, it holds that $h''(\rho)\geq r$. Applying Lemma \ref{lem:convex_analysis_lemma} the result follows with the constant $r$ defined as follows \begin{equation}\label{eq:constant_r_hinge_loss} r:= \frac{64}{125 \rho^*\sigma \Vert \bm{\mu} \Vert \sqrt{2\pi}}\exp\left(-\frac{(\tfrac{4}{3}\frac{1}{\rho^*}-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2} \right). \end{equation} \end{proof} We define the target set $C$ as follows: \begin{equation}\label{target_set_for_hinge} C:=\left\{\bm{\theta}:\vert \rho-\rho^*\vert \leq \tfrac{1}{2}\rho^* \text{ and } \Vert \tilde{\bm{\theta}}\Vert \leq c:=\frac{\sqrt{2\pi}}{\sigma}\left(2+\tfrac{3}{2}\rho^*\Vert \bm{\mu} \Vert\left(\sigma\sqrt{\tfrac{2}{\pi}}+\Vert \bm{\mu} \Vert \right) \right) \right\}. \end{equation} Also, we consider the following drift function \begin{equation} V(\bm{\theta}):=\Vert \bm{\theta} -\bm{\theta}^*\Vert^2. \end{equation} The following proposition establishes the drift equation. \begin{proposition} Suppose that the step-size $\alpha$ satisfies the bound \begin{equation}\label{bound_for_alpha_hinge_loss} \alpha< \frac{\min\{1,\tfrac{1}{2}r\rho^*\}}{\Vert \bm{\mu} \Vert^2+d\sigma^2} \end{equation} where $r$ is defined in \eqref{eq:constant_r_hinge_loss}. The following is then true. \begin{equation} \left(\bE\left[V(\bm{\theta}_k)|\mathcal{F}_{k-1} \right]-V(\bm{\theta}_{k-1})\right)1_{\{\bm{\theta}_{k-1}\not\in C\}}\leq -\tfrac{1}{2}\min\{1,\tfrac{1}{2}r\rho^*\}1_{\{\bm{\theta}_{k-1}\not\in C\}}. \end{equation} \end{proposition} \begin{proof} Suppose that $\bm{\theta}_{k-1}\not\in C$. Lemma \ref{lem:hinge_loss_second_lemma} yields that \begin{equation}\label{eq:hinge_lemma_proposition_eq2} f(\rho_{k-1}\bm{\mu})-f(\rho^*\bm{\mu}) \geq \tfrac{1}{2}r\rho^* \quad \text{ whenever } \vert \rho-\rho^* \vert \geq \tfrac{1}{2}\rho^* \end{equation} where constant $r$ is defined in \eqref{eq:constant_r_hinge_loss}. Moreover, from Lemma \ref{lem:hinge_high_noise_first_lemma}, we obtain that \begin{equation}\label{eq:hinge_lemma_proposition_eq1} f(\bm{\theta}_{k-1})-f(\rho_{k-1} \bm{\mu})\geq 1 \quad \text{ whenever } \sigma \Vert \tilde{\bm{\theta}}_{k-1}\Vert \geq c \text{ and } \vert \rho-\rho^* \vert < \tfrac{1}{2}\rho^*. \end{equation} Moreover, Combining \eqref{eq:hinge_lemma_proposition_eq1}, \eqref{eq:hinge_lemma_proposition_eq2} yields that \begin{equation} f(\bm{\theta}_{k-1})-f(\rho^*\bm{\mu}) \geq \min\{1,\tfrac{1}{2}r\rho^*\} \quad \text{ whenever } \bm{\theta}_{k-1} \not\in C. \end{equation} Lemma \eqref{lem:hinge_loss_high_noise_fundamental_lemma} and the bound on $\alpha$ in \eqref{bound_for_alpha_hinge_loss} along with the above displayed bound gives that \begin{equation} \bE\left[V(\bm{\theta}_k)|\mathcal{F}_{k-1} \right]-V(\bm{\theta}_{k-1}) \leq -\tfrac{1}{2}\min\{1,\tfrac{1}{2}r\rho^*\}. \end{equation} The proof is complete. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:hinge_loss_high_noise}] The target set $C$ defined in \eqref{target_set_for_hinge} is compact and thus there exists $b>0$ such that \begin{equation} \left( \bE\left[V(\bm{\theta}_k)|\mathcal{F}_{k-1} \right]-V(\bm{\theta}_{k-1})\right)1_{\{\bm{\theta}_{k-1}\in C\}} \leq b1_{\{\bm{\theta}_{k-1}\in C\}}. \end{equation} Therefore, it holds that \begin{equation}\label{drift_for_hinge_loss} \bE\left[V(\bm{\theta}_k)|\mathcal{F}_{k-1} \right]-V(\bm{\theta}_{k-1}) \leq -\tfrac{1}{2}\min\{1,\tfrac{1}{2}r\rho^*\}1_{\{\bm{\theta}_{k-1}\not\in C\}}+b1_{\{\bm{\theta}_{k-1}\in C\}}. \end{equation} Furthermore, since $C$ is a compact set which does not contain $\bm{0}$, there exists a positive constant $\delta$ such that for any fixed $\bm{\theta}\in C$ it holds that \begin{equation} \delta:= \min \mathbb{P}_{\bm{\xi} \sim \mathcal{P}_*}\left(\bm{\xi}^T\bm{\theta}\geq 1 \right)>0. \end{equation} The result now follows from Appendix Lemma \ref{lem:drift_lemma_appendix} and Lemma \ref{lem:stopping_time_appendix}. \end{proof} \section{Introduction} Minimization of an expected loss objective function using linear predictors, \begin{equation} \label{eq:intro_expected_loss} \min_{\bm{\theta} \in \R^d} f(\bm{\theta}) := \bE_{(\bm{\zeta},y) \sim \mathcal{P}} \ell(\bm{\zeta}^T\bm{\theta}, y), \end{equation} is a central task in machine learning. Here the loss function $\ell: \R \times \R \to \R$, the probability distribution $\mathcal{P}$ is unknown, and the data sample $(\bm{\zeta}, y) \in \R^d \times \R$ is a random vector distributed as $\mathcal{P}$. The most prevalent algorithm employed for solving \eqref{eq:intro_expected_loss} is \textit{stochastic gradient descent} (SGD). Whereas a significant amount of work has been devoted to the convergence analysis of SGD (see, {\em e.g.}, \cite{RM1951,Curtis_SGD,Bubeck_Convex_book, Pflug}), leading, in particular, to learning rate schedules, the question of how to terminate the algorithm when one is near an optimal classifier remains largely unaddressed. Yet, inexpensive stopping criteria are of utmost interest in machine learning. For instance, if one could produce a low cost test to determine near-optimality, then without sacrificing the quality of the solution or efficiency of the SGD algorithm, needless computational time would be eliminated. Secondly, early termination tests impose a degree of predictability on accuracy and running times-- a useful quality when SGD occurs as a subproblem of a larger computation. Several works show that early termination of SGD can prevent overfitting, speed up learning procedures, and/or improve generalization properties \citep{Prechelt2012, Hardt:2016:TFG:3045390.3045520,Yao2007}. Motivated by these facts, we sought to address from stochastic optimization the following question: \begin{center} How to design a test to terminate SGD with a fixed learning rate that is inexpensive without sacrificing quality of the solution? \end{center} To do so, we simplified our setting to binary classification, one of the fundamental examples of supervised machine learning \citep{ShalevShwartzBenDavid}. In binary classification, the learning algorithm is given a sequence of training examples $(\bm{\zeta}_1,y_1),(\bm{\zeta}_2,y_2),\ldots$, often noisy, where $\bm{\zeta}_i\in\R^d$ and $y_i\in\{0,1\}$ for each $i$. The job of the algorithm is to develop a rule for distinguishing future, unseen $\bm{\zeta}$'s that are classified as $1$ from those classified as $0$. In this work, we limit attention to linear classifiers. This means that the learning algorithm must determine a vector $\bm{\theta}$ such that the classification of $\bm{\zeta}$ is $1$ when $\bm{\zeta}^T\bm{\theta}>0$ else it is $0$. Note that any algorithm for linear classification can be extended to one for nonlinear classification via the construction of ``kernels''; see, {\em e.g.}, \cite{ShalevShwartzBenDavid}. This extension is not pursued; we leave it for later work. The usual technique for determining $\bm{\theta}$, which is also adopted herein, is to define a loss function that turns the discrete problem of computing a $1$ or $0$ for $\bm{\zeta}$ to a continuous quantity. Common choices of loss functions include logistic and hinge. For simplicity, we consider only the unregularized logistic and hinge loss in this work. Our theoretical results assume that our data comes from a Gaussian mixture model (GMM). The GMM is attributed to \cite{article}. The problem of identifying GMM parameters given random samples has attracted considerable attention in the literature; see, \textit{e.g}., the recent work of \cite{Ashtiani} and earlier references therein. Another common use of GMMs in the literature, similar to our application here, is as test-cases for a learning algorithm intended to solve a more general problem. Examples include clustering; see, \textit{e.g.,} \cite{DBLP:journals/corr/abs-1902-07137} and \cite{pmlr-v70-panahi17a} and tensor factorization; see, \textit{e.g.,} \cite{sherman2019estimating}. Ordinarily in deterministic first-order optimization methods, one terminates when the norm of the gradient falls below a predefined tolerance. In the case of SGD for binary classification, this is unsuitable for two reasons. First, the true gradient is generally inaccessible to the algorithm or it is computationally expensive to generate even a sufficient approximation of the gradient. Second, even if the computations were possible, an `optimal' classifier $\bm{\theta}$ for the classification task is not necessarily the minimizer of the loss function since the loss function is merely a surrogate for correct classification of the data. \paragraph{Our contributions.} In this paper, we introduce a new and simple termination criterion for stochastic gradient descent (SGD) applied to binary classification using logistic regression and hinge loss with constant step-size $\alpha>0$. Notably, our proposed criterion adds no additional computational cost to the SGD algorithm. We analyze the behavior of the classifier at termination, where we sample from a normal distribution with unknown means $\bm{\mu}_0,\bm{\mu}_1\in \R^d$ and variances $\sigma^2I_d$. Here $\sigma>0$ and $I_d$ is the $d \times d$ identity matrix. As such, we make no assumptions on the separability of the data set. When the variance is not too large, we have the following results: \begin{enumerate} \item The test will be activated for any fixed positive step-size. In particular, we establish an upper bound for the expected number of iterations before the activation occurs. This upper bound tends to a numeric constant when $\sigma$ converges to zero. In fact, we show that the expected time until termination decreases linearly as the data becomes more separable ({\em i.e.}, as the noise $\sigma \to 0$). \item We prove that the accuracy of the classifier at termination nearly matches the accuracy of an optimal classifier. Accuracy is the fraction of predictions that a classification model got right while an optimal classifier minimizes the probability of misclassification when the sample is drawn from the same distribution as the training data. \end{enumerate} When the variance is large, we show that the test will be activated for a sufficiently small step-size. We empirically evaluate the performance of our stopping criterion versus a baseline competitor. We compare performances on both synthetic (Gaussian and heavy-tailed $t$-distribution) as well as real data sets (MNIST \citep{MNIST} and CIFAR-10 \citep{cifar10}). In our experiments, we observe that our test yields relatively accurate classifiers with small variation across multiple runs. \paragraph{Related works.} To the best of our knowledge, the earliest comprehensive numerical testing of a stopping termination test for SGD in neural networks was introduced by \cite{Prechelt2012}. His stopping criteria, which we denote as \textit{small validation set} (SVS), periodically checks the iterate on a validation set. Theoretical guarantees for SVS were established in the works of \citep{Early_stopping_Lin,Yao2007}. \cite{Hardt:2016:TFG:3045390.3045520} shows that SGD is uniformly stable and thus solutions with low training error found quickly generalize well. These results support exploring new computationally inexpensive termination tests-- the spirit of this paper. In a related topic, the relationship between generalization and optimization is an active area of research in machine learning. Much of the pioneering work in this area focused on understanding how early termination of algorithms, such as conjugate gradient, gradient descent, and SGD, can act as an implicit regularizer and thus exhibit better generalization properties \citep{Prechelt2012,Early_stopping_Lin,Yao2007,CG_implicit_regularization,Multipass_Lin}. The use of early stopping as a tool for improving generalization is not studied herein because our experiments indicate that for the problem under consideration, binary classification with a linear separator, the accuracy increases as SGD proceeds and ultimately reaches a steady value but does not decrease, meaning that there is no opportunity to improve generalization by stopping early. See also \cite{Nemirovski_Robust_Stochastic_1}. Instead of using a validation set to stop early, \cite{Variational_Duvenaud} employs an estimate of the marginal likelihood as a stopping criteria. Another termination test based upon a Wald-type statistic developed for solving least squares with reproducing kernels guarantees a minimax optimal testing \citep{Liu_stopping}. However it is unclear the practical benefits of such procedures over a validation set. Several works have introduced validation procedures to check the accuracy of solutions generated from stochastic algorithms based upon finding a point $\bm{\theta}_\varepsilon$ that satisfies a high confidence bound $\bP(f(\bm{\theta}_\varepsilon) - \min f \le \varepsilon) \ge 1-p$, in essence, using this as a stopping criteria (\textit{e.g.}, see \cite{Dima_Robust_Stochastic,Ghadimi_Strongly_cvx_validation_2,Ghadimi_Strongly_cvx_validation_1, Juditsky_robust_stochastic,Nemirovski_Robust_Stochastic_1}). Yet, notably, all these procedures produce points with small function values. For binary classification, however, this could be quite expensive and a good classifier need not necessarily be the minimizer of the loss function. Ideally, one should terminate when the classifier's direction aligns with the optimal direction-- the approach we pursue herein. \subsection{Low regime, proof of Theorem \ref{thm:low}}\label{sec:proof_low} In this section, we investigate the low variance regime. We consider the target set $C$ and function $V$ defined in \eqref{eq:targetset_lownoise} and \eqref{eq:driftfunction_lownoise} respectively, \textit{i.e.} \begin{equation}\label{eq:CV_low_recall} C = \{\bm{\theta}: \bm{\mu}^T\bm{\theta}\geq 1 \}, \quad V(\bm{\theta})=\left(M-\bm{\mu}^T\bm{\theta}\right)^2, \end{equation} where $M$ is a constant to be determined. Next lemma shows that the drift equation \eqref{eq:driftequation} holds for the pair $(C,V)$. \begin{lemma}[Drift equation]\label{lem:driftequation} Consider the SGD algorithm and let the set $C$ and the function $V$ be as in \eqref{eq:CV_low_recall}. Define the constants $c,b,M$ as in \eqref{eq:para_log} and \eqref{eq:para_hinge}. Then provided that $\sigma \leq c\Vert \bm{\mu} \Vert$, the function $V$ is a drift function with respect to the set $C$ and it satisfies the drift equation \eqref{eq:driftequation} with the constant $b$. \end{lemma} \begin{proof}For simplicity we write $\mathcal{F}_{-1}:=\sigma\left(\{\bm{\theta}_0=\bm{\theta}\} \right)$. Fix $k\geq 1$ and write $\bm{\xi}_k=\bm{\mu}+\sigma\bm{\psi}_k$ with $\bm{\psi}_k \sim N(0,I_d)$. Denote $\psi_k:=\frac{\bm{\mu}^T\bm{\psi}_k}{\Vert\bm{\mu} \Vert}$, thus $\psi_k\sim N(0,1)$. In order to show that the function $V$ satisfies the drift equation \eqref{eq:driftequation}, it suffices to assume $\bm{\theta}_{k-1}\not\in C$; in particular, this means $\bm{\theta}_{k-1}^T\bm{\mu}<1$. \item \textbf{Logistic loss.} By expanding out the term using the update formula, we get the following \begin{align} \label{eq: low_noise_blah_1} V(\bm{\theta}_{k})= V(\bm{\theta}_{k-1})-\frac{2\alpha \bm{\mu}^T\bm{\xi}_k(M-\bm{\mu}^T\bm{\theta}_{k-1})}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}+\frac{\alpha^2(\bm{\mu}^T\bm{\xi}_k)^2}{(1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1}))^2}. \end{align} We have \begin{align*} &\bE_{\bm{\xi}_k}\left[ \frac{\bm{\mu}^T\bm{\xi}_k}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}|\mathcal{F}_{k-1}\right]\\ & \qquad =\Vert\bm{\mu}\Vert^2\bE_{\bm{\xi}_k}\left[ \frac{1}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}|\mathcal{F}_{k-1}\right]+\sigma \Vert \bm{\mu}\Vert\bE_{\bm{\xi}_k,\psi_k}\left[ \frac{\psi_k}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}|\mathcal{F}_{k-1}\right]\\ & \qquad \geq \Vert \bm{\mu} \Vert^2 \bE_{\bm{\xi}_k}\left[ \frac{1}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}|\mathcal{F}_{k-1}\right]+\sigma \Vert \bm{\mu}\Vert\bE_{\psi_k}\left[\psi_k1_{\{\psi_k<0\}} \right]\\ & \qquad = \Vert \bm{\mu} \Vert^2 \bE_{\bm{\xi}_k}\left[ \frac{1}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})} \left ( 1_{\{\bm{\mu}^T \bm{\theta}_{k-1} \geq \bm{\xi}_k^T \bm{\theta}_{k-1}\}} + 1_{\{\bm{\mu}^T \bm{\theta}_{k-1} < \bm{\xi}_k^T \bm{\theta}_{k-1}\}} \right ) |\mathcal{F}_{k-1}\right] -\sigma \Vert \bm{\mu}\Vert \sqrt{\frac{1}{2\pi}}\\ & \qquad \geq \frac{\Vert \bm{\mu}\Vert^2}{1+\exp(\bm{\mu}^T\bm{\theta}_{k-1})}\bE_{\bm{\xi}_k}\left[ 1_{\{\bm{\mu}^T \bm{\theta}_{k-1} \geq \bm{\xi}_k^T \bm{\theta}_{k-1}\}} |\mathcal{F}_{k-1} \right]-\sigma \Vert \bm{\mu}\Vert \sqrt{\frac{1}{2\pi}}\\ &\qquad \geq \frac{\Vert \bm{\mu}\Vert^2}{2(1+e)}-\sigma \Vert \bm{\mu}\Vert \sqrt{\frac{1}{2\pi}}\\ & \qquad \geq 0.001\Vert \bm{\mu} \Vert^2 . \end{align*} Here the first inequality follows from $\bE[X] \ge \bE[X1_{\{X<0\}}]$ and $1+ \exp(\bm{\xi}_k^T\bm{\theta}_{k-1}) \ge 1$, the second equation from \eqref{fact:norm_Gaussians}, and the second to last from the observation that for any $X$ normally distributed, $\bP(\bE[X] \ge X) = 1/2$ and $\bm{\xi}_k^T\bm{\theta}_{k-1} \sim N(\bm{\mu}^T\bm{\theta}_{k-1}, \sigma^2 \norm{\bm{\theta}_{k-1}}^2)$ and $\bm{\mu}^T\bm{\theta}_{k-1} < 1$. The last inequality uses the assumption $\sigma \le 0.33 \norm{\bm{\mu}}$. By taking the conditional expectations of \eqref{eq: low_noise_blah_1} combined with the above sequence of inequalities, we deduce the following bound \begin{align*} &\bE\left[V(\bm{\theta}_{k})-V(\bm{\theta}_{k-1})|\mathcal{F}_{k-1}\right]\\ & \quad \qquad = \bE_{\bm{\xi}_k} \left [-\frac{2\alpha \bm{\mu}^T\bm{\xi}_k(M-\bm{\mu}^T\bm{\theta}_{k-1})}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})} | \mathcal{F}_{k-1} \right] + \bE_{\bm{\xi}_k} \left [\frac{\alpha^2(\bm{\mu}^T\bm{\xi}_k)^2}{(1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1}))^2} | \mathcal{F}_{k-1} \right]\\ & \quad \qquad \leq -0.002(M-1)\alpha \Vert \bm{\mu} \Vert^2+ \alpha^2\Vert \bm{\mu} \Vert^2\left(\Vert \bm{\mu}\Vert^2+\sigma^2 \right)\\ & \quad \qquad = \alpha \Vert \bm{\mu} \Vert^2\left[ -0.002(M-1)+\alpha\left(\Vert \bm{\mu} \Vert^2+\sigma^2\right)\right]. \end{align*} Here the first inequality follows from $\bm{\mu}^T\bm{\theta}_{k-1} < 1$ and by upper bounding $\frac{(\bm{\mu}^T\bm{\xi}_k)^2}{(1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1}))^2}$ with $(\bm{\mu}^T\bm{\xi}_k)^2$ and then applying \eqref{fact:norm_Gaussians}. A quick computation after plugging in the value of $M$ and the bound $\sigma\leq 0.33\Vert \bm{\mu} \Vert$ from \eqref{eq:para_log} yields the drift equation \eqref{eq:driftequation} with $b=\alpha \Vert \bm{\mu} \Vert^2$. \item \textbf{Hinge loss.} By expanding out the term using the update formula, we get the following \begin{equation}\label{eq:Vequation_hinge} V(\bm{\theta}_k)=V(\bm{\theta}_{k-1})-2\alpha (M-\bm{\mu}^T\bm{\theta}_{k-1})\bm{\mu}^T\bm{\xi}_k1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}+\alpha^2(\bm{\mu}^T\bm{\xi}_k)^21_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}. \end{equation} We have \begin{align*} \bE_{\bm{\xi}_k}[1_{\{ \bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}\bm{\mu}^T\bm{\xi}_k|\mathcal{F}_{k-1}] &=\Vert \bm{\mu}\Vert^2\bE_{\bm{\xi}_k}[1_{\{ \bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}|\mathcal{F}_{k-1}]+\sigma\Vert \bm{\mu} \Vert\bE_{\bm{\xi}_k,\psi_k}[1_{\{ \bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}\psi_k|\mathcal{F}_{k-1}]\\ & \geq \frac{1}{2}\Vert \bm{\mu} \Vert^2+\sigma\Vert \bm{\mu} \Vert\bE_{\psi_k}[\psi_k1_{\{\psi_k<0 \}}]\\ & = \frac{1}{2}\Vert \bm{\mu} \Vert^2-\sigma\Vert \bm{\mu} \Vert\sqrt{\frac{1}{2\pi}} \\& \geq 0.001 \Vert \bm{\mu} \Vert^2. \end{align*} Here the first inequality follows from $1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq \bm{\mu}^T\bm{\theta}_{k-1}\}}\leq 1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}$ and $\bE_{\bm{\xi}_k}[1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq \bm{\mu}^T\bm{\theta}_{k-1}\}}]=\tfrac{1}{2}$, and the second from \eqref{fact:norm_Gaussians}. The last inequality uses the assumption $\sigma \leq 1.25 \Vert \bm{\mu} \Vert$. By taking conditional expectations of \eqref{eq:Vequation_hinge} combined with the above sequence of inequalities, we deduce the bound \begin{equation*} \begin{aligned} \bE[V(\bm{\theta}_k)-V(\bm{\theta}_{k-1})|\mathcal{F}_{k-1}]&=\bE_{\bm{\xi}_k}\left[-2\alpha(M-\bm{\mu}^T\bm{\theta}_{k-1})1_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}|\mathcal{F}_{k-1}\right]+\bE_{\bm{\xi}_k}\left[\alpha^2(\bm{\mu}^T\bm{\xi}_k)^21_{\{\bm{\xi}_k^T\bm{\theta}_{k-1}\leq 1\}}|\mathcal{F}_{k-1}\right] \\ &\leq \alpha \Vert \bm{\mu} \Vert^2\left[ -0.002(M-1)+\alpha\left(\Vert \bm{\mu} \Vert^2+\sigma^2\right)\right]. \end{aligned} \end{equation*} A quick computation after plugging in the value of $M$ and the bound $\sigma \leq 1.25 \Vert \bm{\mu} \Vert$ yields the desired result. \end{proof} Recall, the stopping times $\tau_m$ denote the $m^{th}$ time that the SGD iterates enter the target set $C$. We show that $\bE[\tau_m]=\mathcal{O}(m)$. To do so, we begin by stating a lemma that gives a bound on the stopping time $\tilde{\tau}_1$ starting from any $\bm{\theta}_0$. In other words, for an arbitrary starting $\bm{\theta}_0$, we define \[ \tilde{\tau}_1:=\inf\{k>0:\bm{\theta}_k\in C\}. \] \begin{lemma}[\cite{meyn2012markov}, Theorem 11.3.4]\label{lem:drift_from_meyn} Suppose that $V:\R^d\rightarrow [0,+\infty)$ is a drift function with respect to some target set $C$ \textit{i.e.} for some constant $b\in (0,+\infty)$ the drift equation \eqref{eq:driftequation} holds. The following is true \begin{equation} \bE[\tilde{\tau}_1|\bm{\theta}_0=\bm{\theta}]\leq \tfrac{1}{b}V(\bm{\theta}). \end{equation} \end{lemma} We establish upper bounds on $\bE[\tau_m]$ for $m\geq 1$ in the following proposition. \begin{proposition}(Bound on $\bE[\tau_m]$)\label{prp:Etaum_lownoise} Let $\bm{\theta}_0=\bm{0}$ and assume the notation and assumptions of Lemma \ref{lem:driftequation} hold. The following is true for all $m\geq 1$ \begin{align} \bE[\tau_m]\leq (m-1) \left ( 1+\frac{M^2}{b}\cdot\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\frac{\alpha\sigma^3 M^2}{ \Vert \bm{\mu} \Vert b}\cdot\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right) \right ) +\frac{M^2}{b}. \end{align} \end{proposition} \begin{proof}First, the result for $m=1$ follows immediately by combining Lemmas \ref{lem:driftequation} and \ref{lem:drift_from_meyn} with $\bm{\theta}_0 = \bm{0}$. We now assume that $\tau_{m-1}<\infty$ a.s. for some $m\geq 2$. Fix an integer $n \ge 1$. We decompose the space to yield the following bounds \begin{equation} \begin{aligned} \label{eq: low_noise_bound} \bE\big[(\tau_m-\tau_{m-1})\wedge& n|\mathcal{F}_{\tau_{m-1}+1}\big]=\bE\left[((\tau_m-\tau_{m-1})\wedge n)|\mathcal{F}_{\tau_{m-1}+1}\right]1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\geq 1\}}\\ & \qquad \qquad \qquad \qquad \qquad +\bE\left[((\tau_m-\tau_{m-1})\wedge n)|\mathcal{F}_{\tau_{m-1}+1}\right]1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}<1\}}\\ &= 1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\geq 1\}}+ \bE\left[((\tau_m-\tau_{m-1})\wedge n)|\mathcal{F}_{\tau_{m-1}+1}\right]1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}<1\}}\\ &=1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\geq 1\}}+\sum_{i=1}^{\infty}\bE\left[(\tau_m-\tau_{m-1})\wedge n|\mathcal{F}_{\tau_{m-1}+1}\right]1_{\{i-1< 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\leq i\}}\\ &= 1+\sum_{i=1}^{\infty}\bE\left[\tilde{\tau}_1\wedge n|\bm{\theta}_0=\bm{\theta}_{\tau_{m-1}+1}\right]1_{\{i-1< 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\leq i\}}. \end{aligned} \end{equation} Here the first equality follows because $((\tau_m-\tau_{m-1})\wedge n) 1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1} \ge 1\}} = 1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\geq 1\}}$ and the last equality by the strong Markov property. We consider the logistic and hinge loss case separately to show that the following is true \begin{equation}\label{eq:ineq:indicators} \begin{aligned} 1_{\{i-1< 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\leq i\}} &\leq 1_{\{\bm{\mu}^T\bm{\xi}_{\tau_{m-1}+1}<\frac{1-i}{\alpha}\}}. \end{aligned} \end{equation} For clarity, in the next few inequalities, we write $1\{.\}$ instead of $1_{\{.\}}$. In case of logistic loss, for each $i \ge 1$, we observe the bound \begin{equation} \begin{aligned} 1\{i-1< 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\leq i\}&\leq 1\{i-1< 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\}\nonumber\\&=1\left\{i-1< 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}}-\frac{\alpha \bm{\mu}^T\bm{\xi}_{{\tau_{m-1}}+1}}{1+\exp(\bm{\xi}_{{\tau_{m-1}}+1}^T\bm{\theta}_{\tau_{m-1}})}\right\}\nonumber\\ &\leq 1\left\{i-1<-\frac{\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}+1}}{1+\exp(\bm{\xi}_{\tau_{m-1}+1}^T\bm{\theta}_{\tau_{m-1}})}\right\} \label{eq:i-1_calculus}\nonumber\\ & \leq 1\left\{i-1<-\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}+1}\right\},\nonumber \end{aligned} \end{equation} where the second inequality follows because $\bm{\mu}^T\bm{\theta}_{\tau_{m-1}} \geq 1$ and the last inequality because $-\alpha \bm{\mu}^T \bm{\xi}_{\tau_{m-1}+1}$ is positive since $i-1\geq 0$. \item In case of hinge loss, for each $i\geq 1$, similar as above, we observe the bound \begin{equation} \begin{aligned} 1\left\{i-1<1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\leq i\right\}&\leq 1\left\{i-1<1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\right\}\\&\leq 1\left\{i-1<1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}}-\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}+1}1_{\{\bm{\xi}_{\tau_{m-1}+1}^T\bm{\theta}_{\tau_{m-1}}\leq 1\}}\right\}\\&\leq 1\left\{i-1<-\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}+1}1_{\{\bm{\xi}_{\tau_{m-1}+1}^T\bm{\theta}_{\tau_{m-1}}\leq 1\}} \right\}\\&= 1\left\{i-1<-\alpha\bm{\mu}^T\bm{\xi}_{\tau_{m-1}+1}\right\}. \end{aligned} \end{equation} Therefore we have shown that \eqref{eq:ineq:indicators} holds. Setting $\bm{\theta}_0=\bm{\theta}_{\tau_{m-1}+1}$ , by Lemma \ref{lem:drift_from_meyn} for each $i\geq 1$, we deduce \begin{equation} \begin{aligned} \label{eq: low_noise_blah_3} \bE\left[\tilde{\tau}_1\wedge n|\bm{\theta}_0=\bm{\theta}_{\tau_{m-1}+1}\right]1_{\{i-1< 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\leq i\}}&\leq \frac{(M-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1})^2}{b}1_{\{i-1< 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}+1}\leq i\}}\\&\leq \frac{(M+i-1)^2}{b}1_{\{ \bm{\mu}^T\bm{\xi}_{\tau_{m-1}+1}< \frac{1-i}{\alpha}\}}. \end{aligned} \end{equation} Finally we observe that \begin{equation} \begin{aligned} \label{eq: low_noise_blah_4} \bE\left[1_{\{ \bm{\mu}^T\bm{\xi}_{\tau_{m-1}+1}< \frac{1-i}{\alpha}\}} \right]&=\bE\left[\sum_{k=1}^{\infty}1_{\{ \bm{\mu}^T\bm{\xi}_{k+1}< \frac{1-i}{\alpha}\}}1_{\{\tau_{m-1}=k\}}\right]\\ &=\sum_{k=1}^{\infty}\bE\left[1_{\{ \bm{\mu}^T\bm{\xi}_{k+1}< \frac{1-i}{\alpha}\}} \right]\bE\left[1_{\{\tau_{m-1}=k\}} \right]\\ &= \Phi\left(\frac{\frac{1-i}{\alpha}-\Vert \bm{\mu}\Vert^2}{\sigma\Vert \bm{\mu}\Vert } \right)\sum_{k=1}^{\infty}\bE\left[1_{\{\tau_{m-1}=k\}} \right]\\ &=\Phi\left(\frac{\frac{1-i}{\alpha}-\Vert \bm{\mu}\Vert^2}{\sigma\Vert \bm{\mu}\Vert } \right). \end{aligned} \end{equation} The second equality is by independence and the third equality because $\bm{\mu}^T \bm{\xi}_{k+1} \sim N(\norm{\bm{\mu}}^2, \sigma^2 \norm{\bm{\mu}}^2)$. By combining \eqref{eq: low_noise_bound}, \eqref{eq: low_noise_blah_3}, and \eqref{eq: low_noise_blah_4}, we obtain the following \begin{equation} \begin{aligned} \label{eq: low_noise_blah_5} \bE\big[(\tau_m-&\tau_{m-1})\wedge n\big]\leq 1+\frac{M^2}{b}\cdot\Phi\left(-\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\sum_{i=2}^{\infty} \frac{(M+i-1)^2}{b}\cdot\Phi\left(\frac{\frac{1-i}{\alpha}-\Vert \bm{\mu}\Vert^2}{\sigma\Vert \bm{\mu}\Vert } \right)\\ &=1+\frac{M^2}{b}\cdot\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\sum_{i=2}^{\infty} \frac{(M+i-1)^2}{b}\cdot\Phi^{c}\left(\frac{\Vert \bm{\mu}\Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu}\Vert } \right)\\ &\le1+\frac{M^2}{b}\cdot\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\frac{\alpha\sigma\Vert \bm{\mu} \Vert}{b\sqrt{2\pi}}\cdot\sum_{i=2}^{\infty} \frac{(M+i-1)^2}{\alpha\Vert \bm{\mu}\Vert^2+i-1}\cdot\exp\left(-\frac{1}{2}\left(\frac{\Vert \bm{\mu}\Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu}\Vert}\right)^2 \right), \end{aligned} \end{equation} where we used the inequality $\Phi^{c}(t)<\frac{1}{t\sqrt{2\pi}}\exp(-\frac{t^2}{2})$ for all $t>0$. Next, note that $\frac{M+i-1}{\alpha \Vert \bm{\mu} \Vert^2+i-1}\leq \frac{M}{\alpha \Vert\bm{\mu}\Vert^2}$ holds for all $i\geq 2$. Using this we obtain the following bound \begin{equation} \begin{aligned}\label{eq:exp_low_noise} \sum_{i=2}^{\infty} \frac{(M+i-1)^2}{\alpha\Vert \bm{\mu}\Vert^2+i-1}\cdot\exp&\left(-\frac{1}{2}\left(\frac{\Vert \bm{\mu}\Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu}\Vert}\right)^2 \right)\\ &\leq \frac{\sigma M^2}{\alpha \Vert \bm{\mu} \Vert^3}\cdot\sum_{i=2}^{\infty} \frac{\alpha\Vert \bm{\mu}\Vert^2+i-1}{\alpha \sigma \Vert \bm{\mu} \Vert}\cdot\exp\left(-\frac{1}{2}\left(\frac{\alpha\Vert \bm{\mu}\Vert^2+i-1}{\alpha\sigma\Vert \bm{\mu}\Vert}\right)^2 \right)\\&\leq \frac{\sigma M^2}{\alpha \Vert \bm{\mu} \Vert^3}\cdot \alpha\sigma \Vert \bm{\mu} \Vert\cdot\int_{\frac{\Vert \bm{\mu} \Vert}{\sigma}}^{+\infty} t\exp\left(-\frac{t^2}{2}\right)dt \\&=\frac{\sigma^2 M^2}{\Vert \bm{\mu} \Vert^2}\cdot\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right). \end{aligned} \end{equation} Here we have used that $t \mapsto t\exp(-\frac{t^2}{2})$ is decreasing over $[1,+\infty)$. Combining \eqref{eq: low_noise_blah_5} and \eqref{eq:exp_low_noise}, we obtain that \begin{equation}\begin{aligned} \bE\left[(\tau_m-\tau_{m-1})\wedge n\right]&\leq 1+\frac{M^2}{b}\cdot\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\frac{\alpha\sigma^3 M^2}{ \Vert \bm{\mu} \Vert b}\cdot\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right). \end{aligned} \end{equation} Taking the limit as $n \to +\infty$, we observe that \[\bE[\tau_m]\leq 1+\frac{M^2}{b}\cdot\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\frac{\alpha\sigma^3 M^2}{ \Vert \bm{\mu} \Vert b}\cdot\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right)+\bE[\tau_{m-1}]. \] We then iterate the above inequality yielding \[ \bE[\tau_m] \le (m-1) \left ( 1+\frac{M^2}{b}\cdot\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\frac{\alpha\sigma^3 M^2}{ \Vert \bm{\mu} \Vert b}\cdot\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right)\right ) + \bE[\tau_1]. \] The result follows by plugging in the bound from Lemma \ref{lem:drift_from_meyn} for the base case $m=1$. \end{proof} We are now ready to prove Theorem~\ref{thm:low}. \begin{proof}[Proof of Theorem \ref{thm:low}] In order to simplify the subsequent argument, we define the quantity, \[ M' := 1+\frac{M^2}{b}\cdot\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma} \right)+\frac{\alpha\sigma^3 M^2}{ \Vert \bm{\mu} \Vert b}\cdot\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right). \] It is easy to see that $\mathbb{P}_{\hat{\bm{\xi}}\sim N(\bm{\mu},\sigma^2I_d)}\left( \hat{\bm{\xi}}^T\bm{\theta}\geq 1 \right) \geq \frac{1}{2}$ for any $\bm{\theta}\in C$. Therefore $\delta=\frac{1}{2}$ satisfies \eqref{eq:probability_delta}. By Proposition~\ref{prp:Etaum_lownoise} with Lemma \ref{lem:ETleqET_C}, we conclude that \begin{align*} \bE[T]&\leq \bE[T_C]=\sum_{m=1}^{\infty} \bE[T_C1_{\{T_C=\tau_m\}}]\leq \sum_{m=1}^{\infty}\frac{\bE[\tau_m]}{2^{m-1}} \le \sum_{m=1}^\infty \frac{(m-1)M' + \frac{M^2}{b} }{2^{m-1}} = 2M' + \frac{2M^2}{b}. \end{align*} \end{proof} \section{Numerical Experiments} \label{sec:Num_Experiment} We investigate the performance of our termination test on two popular data sets, MNIST \citep{MNIST} and CIFAR-10 \citep{cifar10}, as well as synthetic data generated from Gaussians and heavy-tailed student t-distributions. All tests were performed using our zero overhead stopping criteria outlined in \eqref{eq: practical_termination_test}; experiments using our test which required an extra sample \eqref{eq: termination_test} are not presented since the behaviors of the two criteria were indistinguishable on all data sets. \paragraph{Comparison with a popular stopping criterion.} We include as a baseline a popular termination test, the small validation set (SVS) \citep{Prechelt2012}. The SVS termination test is as follows. One fixes a validation set of $p$ instances $(\bm{\zeta}^{\rm V}_1,y^{\rm V}_1)$, \ldots, $(\bm{\zeta}^{\rm V}_p,y^{\rm V}_p)$ drawn from the same distribution as the training data. Then for $m = 1, 2, \ldots$, one checks the fraction correct of the current classifier $\bm{\theta}_{ml}$, where $ml$ is the iteration index, on the $p$ instances. In other words, the SVS test is run once every $l$ iterations. If the fraction correct fails to increase compared to the last run of the SVS, then the SGD iterations are terminated. Note the computational overhead of running the small validation set is about $p$ times the cost of one SGD iteration. Therefore, in order to make the overhead only a constant factor, we choose $l=2p$, meaning an approximately 50\% overhead for SVS. In contrast, the overhead for \eqref{eq: practical_termination_test} is 0. The value of $p$ is a tuning parameter for SVS; we exhibit results for three different $p$ values (see Figs.~\ref{fig:normal_scatter}, \ref{fig:ht_scatter}, \ref{fig:mnist_scatter}, \ref{fig:cifar_scatter} ). \begin{figure*}[htp!] \begin{center} \subfigure[]{\includegraphics[height=3cm]{it_vs_sigma_l.eps}} \quad \subfigure[\label{fig:black}]{\includegraphics[height=3cm]{acc_vs_sigma_l.eps}} \quad \subfigure[]{\includegraphics[height=3cm]{it_vs_sigma_h.eps}} \quad \subfigure[]{\includegraphics[height=3cm]{acc_vs_sigma_h.eps}} \end{center} \caption{Performance of stopping criterion \eqref{eq: practical_termination_test} on a mixture of Gaussians as $\sigma$ is varied. Plots $(a),(b)$ are logistic and $(c),(d)$ are hinge. All plots show tests for values of $\sigma$ equally spaced from 0.05 to 2.0. For each value of $\sigma$, ten trials were run. Plots $(a),(c)$ show the relationship between $\sigma$ and $k$, the iteration number when \eqref{eq: practical_termination_test} first holds. Plots $(b),(d)$ show the accuracy as red asterisks. The green asterisks show the accuracy of the optimal classifier. The black curve on the right is the ratio of the average accuracy (over 10 trials) of the classifier when \eqref{eq: practical_termination_test} holds to the accuracy of the optimal classifier.} \label{fig:acctime} \end{figure*} \paragraph{Measuring the accuracy.} In all the experiments, we measure the performance of a method with a score, generally known as ``accuracy," that is the fraction correct on a large validation set drawn from the same distribution as the training data. Thus, 1.0 is perfect accuracy, while 0.5 means that $\bm{\theta}_k$ is no better at classifying than random guessing. It is important to note that even on data for which the means $\bm{\mu}_0,\bm{\mu}_1$ are known a priori ({\em e.g.}, synthetic data), the score of the optimal $\bm{\theta}^*$ will not be 1.0 because the large validation set itself is noisy. We center the data so that the linear classifier is homogeneous. In a preliminary phase, 100 samples are drawn from the training set. From this, $\bm{\mu}_0$ and $\bm{\mu}_1$ are estimated, and then the average of these estimates is used to offset training instances during SGD. \paragraph{Parameter settings.} After centering, the vectors $\bm{\theta}$ and $\bm{\xi}$ scale inversely, so the step-size parameter $\alpha$ should scale as $1/\sigma^2$. Therefore, we take the step-size to be $\tilde\alpha/\tilde\sigma^2$. Here, $\tilde\sigma^2$ is the average of $\norm{\bm{\zeta}_j-\tilde\bm{\mu}_{y_j}}^2$, and $\tilde\bm{\mu}_i$ ($i=0$ or $i=1$) is the estimate of $\bm{\mu}_i$, averaged over the two classes. We compute the quantities $\tilde\sigma^2$ and $\tilde\bm{\mu}_i$ using the 100 samples described in the preceding paragraph. Note that for the Gaussian mixture model, the expected value of $\tilde\sigma^2$ is $\sigma^2d$. For the synthetic data, the means and variances are known exactly a priori, so the estimation procedures described in the previous two paragraphs are unnecessary. However, we used them anyway in order to be consistent with the tests on the realistic data. The parameter $\tilde\alpha$ described in the last paragraph is a scale-free tuning parameter. It is known (see, e.g., \cite{Nemirovski_Robust_Stochastic_1}) that a smaller $\tilde\alpha$ corresponds to more iterations but greater ultimate accuracy under a reasonable model of the data. Our termination test is obviously sensitive to the choice of $\tilde\alpha$: the condition $\bm{\xi}_{k+1}^T\bm{\theta}_k\ge 1$ cannot hold unless $\norm{\bm{\theta}_k}\ge 1/\norm{\bm{\xi}_{k+1}}$, but $\bE\left[\norm{\bm{\theta}_k}\right]\le O(\alpha k)$. See also Theorems~\ref{thm:low} and \ref{thm:high}. On the other hand, SVS is only mildly sensitive to $\tilde\alpha$, according to our testing. Indeed, there is an upper bound of $pl$ on the total number of iterations possible before termination using the SVS condition, independent of $\tilde\alpha$ and of all other aspects of the problem. The dependence of the termination test on $\tilde\alpha$ is evidently desirable because the user is presumably seeking greater accuracy when a smaller value of $\tilde\alpha$ is selected. \subsection{Experiments with synthetic data} \label{sec:comp-sim} \begin{figure*}[htp!] \centering \includegraphics[height=2.5cm]{gm_scatter99_500_l_10.eps} \qquad \includegraphics[height=2.5cm]{gm_scatter99_500_l_200.eps} \qquad \includegraphics[height=2.5cm]{gm_scatter99_500_h_10.eps}\qquad \includegraphics[height=2.5cm]{gm_scatter99_500_h_200.eps} \\ \vspace{2 mm} \includegraphics[height=2.5cm]{gm_scatter100_500_l_10.eps}\qquad \includegraphics[height=2.5cm]{gm_scatter100_500_l_200.eps} \qquad \includegraphics[height=2.5cm]{gm_scatter100_500_h_10.eps}\qquad \includegraphics[height=2.5cm]{gm_scatter100_500_h_200.eps} \\ \vspace{2 mm} \includegraphics[height=2.5cm]{gm_scatter101_500_l_10.eps}\qquad \includegraphics[height=2.5cm]{gm_scatter101_500_l_200.eps} \qquad \includegraphics[height=2.5cm]{gm_scatter101_500_h_10.eps}\qquad \includegraphics[height=2.5cm]{gm_scatter101_500_h_200.eps} \caption{Each plot shows 10 random runs of SGD applied to normally distributed data with indicated values of $\sigma$ and for a fixed dimension $d=500$. For each of the ten runs, five termination tests corresponding to five colors were applied. SVS was tried with $p=32, 128, 512$, depicted as red, magenta and cyan circles respectively. Test \eqref{eq: practical_termination_test} is indicated with a blue asterisk. A green `+' corresponds to termination after $1.5k$ iterations, where $k$ is the iteration index that \eqref{eq: practical_termination_test} first holds. The notation $(l/200)$ means logistic loss with $\tilde\alpha=1/200$; simillarly $(h/10)$ means hinge loss with $\tilde\alpha = 1/10$, and so on.} \label{fig:normal_scatter} \end{figure*} \paragraph{Normal distribution.} We generated test and training data using a mixture of Gaussians given by $N(\bz,\sigma^2I)$ for the 0-class and $N(\bm{e}_1,\sigma^2I)$ for the 1-class, where $\bm{e}_1 = (1,0,\hdots, 0)^T \in \R^d$. In Fig.~\ref{fig:acctime}, we present the running time and accuracy (fraction correct) of our termination test for a fixed dimension $d=500$ and $\sigma$ ranging from $0.05$ to $2$. We record 10 runs for each value of $\sigma$. The performance of the classifier when our termination test \eqref{eq: practical_termination_test} holds almost matches the optimal classifier; in particular, the averaged accuracy of our classifier/accuracy of the optimal classifier over the 10 runs, black curve in Fig.~\ref{fig:black}, never dips below $0.95$. In Fig.~\ref{fig:normal_scatter}, we compare performance of \eqref{eq: practical_termination_test} against SVS termination. One axis shows accuracy while the other shows iteration count. We continued to run SGD for an additional $1.5k$ iterations where $k$ is the first iteration at which \eqref{eq: practical_termination_test} holds (green '+') to test whether accuracy improves after termination. The tests (for several values of $\sigma$, both hinge and logistic, and two values of $\tilde\alpha$) in Fig.~\ref{fig:normal_scatter} indicate that \eqref{eq: practical_termination_test} is more accurate than SVS, more predictable (i.e., there is less spread in the scatter plot), and that running until $1.5k$ iterations does not significantly improve the solution. As expected, for a large $\tilde\alpha$, \eqref{eq: practical_termination_test} requires fewer iterations than SVS with $p=512$, while the opposite relationship holds for a small $\tilde\alpha.$ \begin{figure} \begin{center} \includegraphics[height=2.5cm]{ht_scatter98_2_l_10.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter98_2_l_200.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter98_2_h_10.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter98_2_h_200.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter99_2_l_10.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter99_2_l_200.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter99_2_h_10.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter99_2_h_200.eps} \\ \vspace{2 mm} \includegraphics[height=2.5cm]{ht_scatter100_2_l_10.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter100_2_l_200.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter100_2_h_10.eps} \qquad \includegraphics[height=2.5cm]{ht_scatter100_2_h_200.eps} \\ \end{center} \caption{Tests on the student-t distribution (heavy tailed) with two degrees of freedom and the indicated value of parameter $\beta$. See the caption of Fig.~\ref{fig:normal_scatter} for explanation of the plots.} \label{fig:ht_scatter} \end{figure} \paragraph{Heavy-tailed distribution.} We consider the student t-distribution with two degrees of freedom. This distribution is heavy-tailed since some of its higher moments are infinite. The two classes were generated as follows. For $\bm{\zeta}$ in the 0-class, each of the $d$ entries of $\bm{\zeta}$ is chosen as $\beta\eta$, where $\beta$ is varied in the experiments and $\eta$ is drawn from the student t-distribution with two degrees of freedom. For the 1-class, $\bm{\zeta}$ is chosen in the same way except that the first entry is incremented by 1. Fig.~\ref{fig:ht_scatter} shows our performance against SVS. The results in this table show similar trends as in the normally distributed case. One difference is that the accuracy achieved by our termination test \eqref{eq: practical_termination_test} is more spread out presumably because of the heavy-tailed nature of the data set. \subsection{Experiments with real data} \label{sec:comp-realistic} \paragraph{MNIST handwritten digits.} We compared our termination test on the MNIST handwritten digit set \citep{MNIST} ($d=784$, no preprocessing of the data other than centering between the two means). Two trials are shown: distinguishing 1 from 8 (easy case) and distinguishing 7 from 9 (more difficult case). The test runs are obtained by running through the training data in different randomized orders. The plots in Fig.~\ref{fig:mnist_scatter} show similar trends as before. As expected, the accuracy is overall higher for $\tilde\alpha=1/200$ than for $\tilde\alpha=1/10$. \begin{figure} \centering \includegraphics[height=2.5cm]{mnist_scatter1_8_l_10.eps} \qquad \includegraphics[height=2.5cm]{mnist_scatter1_8_l_200.eps} \qquad \includegraphics[height=2.5cm]{mnist_scatter1_8_h_10.eps} \qquad \includegraphics[height=2.5cm]{mnist_scatter1_8_h_200.eps} \\ \vspace{2 mm} \includegraphics[height=2.5cm]{mnist_scatter7_9_l_10.eps} \qquad \includegraphics[height=2.5cm]{mnist_scatter7_9_l_200.eps} \qquad \includegraphics[height=2.5cm]{mnist_scatter7_9_h_10.eps} \qquad \includegraphics[height=2.5cm]{mnist_scatter7_9_h_200.eps} \\ \caption{Tests on the MNIST handwritten digit data set for discerning ``1'' from ``8'' and ``7'' from ``9'' for both hinge and logistic, and for both $\tilde\alpha=1/10$ and $\tilde\alpha=1/200$. Refer to the caption of Fig.~\ref{fig:normal_scatter} for the key to the plots.} \label{fig:mnist_scatter} \end{figure} \paragraph{CIFAR-10 image set.} We compared our termination test on the CIFAR-10 \citep{cifar10} ($d=3072$, no preprocessing of the data other than centering between the two means as described earlier). Two trials are shown: distinguishing deer from airplanes and frogs from trucks. As in MNIST, test runs are obtained by running through the training data in different randomized orders. \begin{figure} \centering \includegraphics[height=2.5cm]{cifar_scatter_1_5_l_10.eps} \qquad \includegraphics[height=2.5cm]{cifar_scatter_1_5_l_200.eps} \qquad \includegraphics[height=2.5cm]{cifar_scatter_1_5_h_10.eps} \qquad \includegraphics[height=2.5cm]{cifar_scatter_1_5_h_200.eps} \\ \vspace{2 mm} \includegraphics[height=2.5cm]{cifar_scatter_7_10_l_10} \qquad \includegraphics[height=2.5cm]{cifar_scatter_7_10_l_200.eps} \qquad \includegraphics[height=2.5cm]{cifar_scatter_7_10_h_10.eps} \qquad \includegraphics[height=2.5cm]{cifar_scatter_7_10_h_200.eps} \caption{Tests on the CIFAR-10 image set for two tasks, for logistic and hinge losses, and for $\tilde\alpha=1/10$ and $\tilde\alpha=1/200$. Refer to the caption of Fig.~\ref{fig:normal_scatter} for the key to the plots. The plot in the first row, right, does not include cyan circles because the training data was exhausted before the SVS test could activate for $p=512$.} \label{fig:cifar_scatter} \end{figure} \section{Conclusions} We have proposed a simple and computationally free termination test for SGD for binary classification, supported by both theoretical and experimental results. The theoretical results show that the test will stop SGD after a finite time with a bound on the expected accuracy of the resulting classifier. The bounds that we proved are weaker than what we observed in our experiments. Therefore, the first obvious question left open by this work is whether the theoretical bounds can be improved. In our experimental results, the plots in Figs.~\ref{fig:normal_scatter} through \ref{fig:cifar_scatter} show a consistent pattern that \eqref{eq: practical_termination_test} achieves low accuracy but is faster than SVS for $\tilde\alpha=1/10$, while it achieves higher accuracy with more iterations when $\tilde\alpha=1/200$. This is useful behavior in practice, compared to SVS, since it puts the accuracy/iterations tradeoff in the hands of the user who selects the stepsize $\tilde\alpha$. Another benefit of \eqref{eq: practical_termination_test} apparent from all plots is that the number of iterations is more consistent across random trials, which is beneficial in the case that SGD is used as a subproblem of a larger computation. This work did not explore regularization via early stopping. As mentioned in the introduction, experiments showed that as SGD iterations continued, the accuracy on the test set eventually levels off but does not decrease significantly, {\em i.e.}, SGD for binary classification is not prone to overfitting. Because the test accuracy never shows marked decline, there is no opportunity for early stopping to regularize. However, we know of other settings in which early stopping has a strong regularizing effect ({\em e.g.}, conjugate gradient iterations for image deconvolution, already known in \cite{CG_implicit_regularization}), so if \eqref{eq: practical_termination_test} is extended beyond binary classification in future work, there will likely also be an opportunity to explore regularization. \section{Stopping criterion for stochastic gradient descent}\label{sec:termtest} We analyze learning by minimizing an expected loss problem of homogeneous linear predictors ({\em i.e.}, without bias) of the form \begin{align*} \bE_{(\bm{\zeta},y) \sim \mathcal{P}} [\ell(\bm{\zeta}^T\bm{\theta}, y)] \end{align*} using logistic and hinge regression. Here the samples $(\bm{\zeta}, y) \in \R^d \times \{0,1\}$. We recall that in logistic regression the loss function is defined as follows \begin{equation}\label{eq:log_loss_definition} \ell(x,y):=-yx+\log\left(1+\exp(x)\right). \end{equation} Also, the hinge loss is defined as the following \begin{equation}\label{eq:hinge_loss_definition} \begin{aligned} \ell(x,y):=\begin{cases} \max(1-x,0)& \quad y=1, \\\max(1+x,0)& \quad y=0. \end{cases} \end{aligned} \end{equation} The data comes from a mixture model, that is, flip a coin to determine whether an item is in the $y=0$ or $y=1$ class, then generate the sample $\bm{\zeta}$ from either the distribution $\mathcal{P}_0$ (if $y=0$ was selected) or $\mathcal{P}_1$ (if $y=1$ was selected). We denote the mean of the $\mathcal{P}_0$ (resp. $\mathcal{P}_1$) distribution by $\bm{\mu}_0$ (resp. $\bm{\mu}_1$). The homogeneity of the linear classifier is without loss of much generality because we can assume $\bm{\mu}_0 = - \bm{\mu}_1$. We enforce this assumption, with minimal loss in accuracy, by recentering the data using a preliminary round of sampling (see Sec.~\ref{sec:Num_Experiment}). Because of the homogeneity, we can simplify the notation by redefining our training examples to be $\bm{\xi}_k:=(2y_k-1)\bm{\zeta}_k$ and then assuming that for all $k \ge 0$, $y_k = 1$. Then the new samples $\bm{\xi}$ can be drawn from a \textit{single}, mixed distribution $\mathcal{P}_*$ with mean $\bm{\mu}:=\bm{\mu}_1$ where sampling $\bm{\xi}\sim \mathcal{P}_1$ occurs with probability 0.5 and $-\bm{\xi}\sim\mathcal{P}_0$ occurs with probability 0.5. We make this simplification and, from this point on, we analyze the following optimization problem: \begin{equation}\label{optimization_problem} \min_{\bm{\theta} \in \R^d} f(\bm{\theta}):= \bE_{\bm{\xi} \sim \mathcal{P}_*}[ \ell(\bm{\xi}^T\bm{\theta},1)] \end{equation} Let us remark that the right-hand side of \eqref{optimization_problem} is differentiable with respect to $\bm{\theta}$ in either cases of logistic and hinge loss functions. Indeed, in case of hinge loss, note that for any $\bm{\theta}_{k-1}$, the function $\bm{\xi}_k \mapsto \ell(\bm{\xi}_k^T\bm{\theta}_{k-1},1)$ is almost surely differentiable as $\bP_{\bm{\xi}_k}\left(\bm{\xi}_k^T\bm{\theta}_{k-1}=1\right)=0$. Hence, we consider the expectation in \eqref{optimization_problem} to be over $\mathbb{R}^d\backslash \left\{\bm{\xi}_k:\bm{\xi}_k^T\bm{\theta}_{k-1}=1 \right\}$ on which the argument is differentiable with respect to $\bm{\theta}_{k-1}$. The most widely used method to solve \eqref{optimization_problem} is SGD. Unlike gradient descent which uses the entire data to compute the gradient of the objective function, the SGD algorithm, at each iteration, generates a sample from the probability distribution and updates the iterate based only on this sample, \begin{align} \label{eq:derivative_formula} \bm{\te}_{k}=\bm{\theta}_{k-1} -\alpha \nabla_{\bm{\theta}} \ell(\bm{\xi}_{k}^T\bm{\theta}_{k-1},1), \end{align} where $\bm{\xi}_k \sim \mathcal{P}_*$. Our presentation of SGD assumes a constant step-size $\alpha>0$. Constant step-size is commonly used in machine learning implementations despite the decreasing step-size often assumed to prove convergence (see, \textit{e.g.}, \cite{RM1951}). \cite{Nemirovski_Robust_Stochastic_1} explain in more detail the theoretical basis for both constant and decreasing step-size and provide an explanation as well as workarounds for the poor practical performance of decreasing step-size. However, in practice, constant step-size is still widely used. With constant step-size, SGD is known to asymptotically converge to a neighborhood of the minimizer (see, {\em e.g.}, \cite{Pflug}). Yet, for binary classification, one does not require convergence to a minimizer in order to obtain good classifiers. For homogeneous linear classifiers applied to the hinge loss function, it has been shown (\cite{molitor2019bias}) that the homotopic sub-gradient method converges to a maximal margin solution on linearly separable data. In (\cite{Srebro_logistic}), SGD applied to the logistic loss on linearly separable data will produce a sequence of $\bm{\theta}_k$ that diverge to infinity, but when normalized also converge to the $L_2$-max margin solution. Little is known about the behavior of constant step-size SGD when the linear separability assumption on the data is removed (see, \textit{e.g.}, \citep{Telgarsky_logistic}). The assumption of zero-noise in our context would mean that $\mathcal{P}_0$, $\mathcal{P}_1$ each reduce to a single point, a trivial example of separable data. Since there is often noise in the sample procedure, the data \textit{may not necessarily be linearly separable}. Understanding the behavior of SGD in the presence of noise is, therefore, important. \subsection{Stopping criterion} A common stopping criterion from deterministic first-order optimization methods is to terminate at an iterate satisfying $\norm{\nabla f(\bm{\theta})}^2 < \varepsilon$ for a predetermined $\varepsilon > 0$. Yet, in stochastic optimization, the full gradient is inaccessible or it is simply too expensive to compute. Several works \citep{Dima_Robust_Stochastic,Ghadimi_Strongly_cvx_validation_2,Ghadimi_Strongly_cvx_validation_1, Juditsky_robust_stochastic,Nemirovski_Robust_Stochastic_1} have suggested an alternative for the stochastic setting-- terminate when $\bP(f(\bm{\theta})-\min~f \le \varepsilon) \ge 1-p$ for some chosen small $\varepsilon >0$ and probability $p$. However, for binary classification, the minimizer of the loss function and a perfect classifier may not be the same or one may find a suitable substitute, at a lower cost, without having to compute the exact minimizer. \paragraph{Optimal classifiers.} In classification, we call a classifier, $\bm{\theta}^*$, \textit{optimal} if it has the property that \begin{equation} \label{eq:stop_criteria_observation} \bm{\theta}^*\in \Argmax_{\bm{\theta}} \bP\left(\bm{\xi}^T\bm{\theta}>0\,|\,\bm{\xi}\sim \mathcal{P}_*\right), \end{equation} \textit{i.e.}, the classifier, $\bm{\theta}^*$, minimizes the probability of misclassifying. Note there exist many optimal classifiers, in fact, the condition \eqref{eq:stop_criteria_observation} is scale-invariant; hence, for any $\lambda > 0$, $\lambda \cdot \bm{\xi}^T \bm{\theta}^* > 0 \Longleftrightarrow \bm{\xi}^T \bm{\theta}^* > 0$. Even though the binary classifier is scale-free, the logistic and hinge regression loss is not. It transitions from flat to unit-slope when $\bm{\xi}^T\bm{\theta}=O(1)$. This suggests that when $\bm{\theta}$ reaches this region, a classification has been made. \paragraph{Termination test.} Motivated by the above property of optimal classifiers, we propose the following termination test: Sample $\hat{\bm{\xi}}_k \sim \mathcal{P}_*$ and \begin{align} \label{eq: termination_test} \text{Terminate when \;$ \hat{\bm{\xi}}_k^T\bm{\theta}_k \ge 1$}. \end{align} A second motivation for this termination test comes from support vector machine (SVM) theory \citep{ShalevShwartzBenDavid} in which the scaling of the optimizing classifier is constrained so that the margin between classes is $O(1)$. Therefore, our termination test blends an SVM notion with SGD. Algorithm~\ref{alg:SGD_termination} describes the termination criteria \eqref{eq: termination_test} as applied with the update rule governed by SGD. The termination test \eqref{eq: termination_test} requires an additional sample and an additional inner product per iteration and, as such, imposes a small additional cost. To reduce this cost, in all our numerical experiments (Sec.~\ref{sec:Num_Experiment}), we use the following termination test. \begin{equation}\label{eq: practical_termination_test} \mbox{Terminate when \;$ \bm{\xi}_{k+1}^T\bm{\theta}_k\ge 1,$} \end{equation} which imposes no computational overhead as SGD already computes $\bm{\xi}_{k+1}^T\bm{\theta}_k$. Unfortunately, we could not perform a straightforward analysis of \eqref{eq: practical_termination_test} because it introduces additional dependencies in the sequences $\{\bm{\xi}_k\}_{k=1}^\infty$ and $\{\bm{\theta}_k\}_{k=0}^\infty$. After testing both \eqref{eq: termination_test} and \eqref{eq: practical_termination_test}, we found that up to the noise from the randomness, their behaviors in numerical experiments were identical. \begin{algorithm}[htp!] \textbf{initialize:} $\bm{\theta}_0 \in \R^d$, $\alpha > 0$, $\hat{\bm{\xi}}_0 \sim \mathcal{P}_*$, $k = 0$\\ \textbf{while $\hat{\bm{\xi}}_k^T \bm{\theta}_k < 1$}\\ \qquad \qquad Pick data point $\bm{\xi}_{k+1} \sim \mathcal{P}_*$.\\ \qquad \qquad Compute $\nabla_{\bm{\theta}} \ell(\bm{\xi_{k+1}}^T\bm{\theta}_k,1)$ as in \eqref{eq:derivative_formula} \\ \qquad \qquad Update $\bm{\theta}$ by setting \begin{equation}\label{eq: SGD_update_expected_loss} \bm{\theta}_{k+1} \leftarrow \bm{\theta}_k - \alpha \nabla_{\bm{\theta}} \ell(\bm{\xi_{k+1}}^T\bm{\theta}_k,1) \end{equation}\\ \qquad \qquad Sample $\hat{\bm{\xi}}_{k+1} \sim \mathcal{P}_*$\\ \qquad \qquad $k \leftarrow k+1$\\ \textbf{end} \caption{SGD with termination test} \label{alg:SGD_termination} \end{algorithm} \begin{assumption}\label{Gaussian_Assumption}[The distribution $\mathcal{P}_{*}$ is Gaussian] \rm{Our theoretical analysis makes a further assumption on the distribution $\mathcal{P}_*$. For the rest of this section and Sec.~\ref{sec:Analysis}, $\mathcal{P}_0=N(\bm{\mu}_0,\sigma^2I_d)$, $\mathcal{P}_1=N(\bm{\mu}_1,\sigma^2I_d)$, and therefore $\mathcal{P}_*=N(\bm{\mu},\sigma^2I_d)$, a Gaussian with unknown mean $\bm{\mu}\; (=\bm{\mu}_1=-\bm{\mu}_0)$ and variance $\sigma^2I_d$. This assumption allows for non-separable data provided $\sigma > 0$.} \end{assumption} \paragraph{The minimizer of logistic and hinge regression} In \eqref{eq:stop_criteria_observation} we defined $\bm{\theta}^*$ to be any member of the set of optimal classifiers. For the remainder of this section, we provide an exact characterization of this set. In the next lemma, we redefine $\bm{\theta}^*$ to the minimizer of the expected loss function for either hinge or logistic and show that it is a positive scalar multiple of $\bm{\mu}$. We will continue to use $\bm{\theta}^*$ with this meaning for the remainder of the paper. In the lemma after that, we show that the set of optimal classifiers are exactly positive scalar multiples of $\bm{\mu}$ (or of $\bm{\theta}^*$). \begin{lemma}[Minimizer of the logistic and hinge loss] \label{lem:minimizers} The function $f$ defined in \eqref{optimization_problem} with $\ell$ defined in \eqref{eq:log_loss_definition} or \eqref{eq:hinge_loss_definition} has a unique minimizer at $\bm{\theta}^* = \rho^*\bm{\mu}$ for some $\rho^*\in (0,+\infty)$. Moreover, let $r=\rho^*\sigma^2$. Then in the case of logistic regression, it holds that $r=2$ and in the case of hinge loss, $w=\frac{\sigma}{r\Vert \bm{\mu} \Vert}-\frac{\Vert \bm{\mu} \Vert}{\sigma}$ satisfies \begin{equation}\label{eq:w_hinge} \frac{1}{\sqrt{2\pi}}\cdot\frac{\sigma}{\Vert \bm{\mu} \Vert}=\Phi(w)\cdot \exp(\tfrac{1}{2}w^2). \end{equation} \end{lemma} \begin{proof} We consider the logistic and hinge loss case separately. \item \textbf{Logistic loss.} We have \[ f(\bm{\theta})=\bE_{\bm{\xi} \sim N(\bm{\mu},\sigma^2I_d)}[ -\bm{\theta}^T\bm{\xi} + \log(1+ \exp(\bm{\theta}^T\bm{\xi}))]. \] Clearly, $f$ is a convex function. We next observe that for any $\bm{v},\bm{\theta}\in \R^d$ with $\bm{v}^T\bm{\theta}=0$, it holds that \begin{equation}\label{app.eq:min_eq101} \bm{v}^T\nabla f\left(\bm{\theta}\right)=\bE_{\bm{\xi}}\left[\frac{\bm{\xi}^T\bm{v}}{1+\exp(\bm{\xi}^T\bm{\theta})}\right]=\bE_{\bm{\xi}}[\bm{\xi}^T\bm{v}]\bE_{\bm{\xi}}\left[\frac{1}{1+\exp(\bm{\xi}^T\bm{\theta})}\right]=\bm{v}^T\bm{\mu}\cdot\bE_{\bm{\xi}}\left[\frac{1}{1+\exp(\bm{\xi}^T\bm{\theta})}\right]. \end{equation} Here we used that $\bm{\xi}^T\bm{v}$ and $\bm{\xi}^T\bm{\theta}$ are independent random variables and the expectation of the product of two uncorrelated random variables is the product of the expectations. Now note that for any $\bm{\theta}$, the quantity $\bE_{\bm{\xi}}\left[\frac{1}{1+\exp(\bm{\xi}^T\bm{\theta})}\right]$ is strictly positive. Therefore, if $\bm{v}^T\bm{\theta}=0$ and $\nabla f(\bm{\theta})=\bm{0}$ then, using \eqref{app.eq:min_eq101}, we obtain that $\bm{v}^T\bm{\mu}=0$. Hence, we established that $\nabla f(\bm{\theta})=\bm{0}$ implies $\bm{\theta}=\rho \bm{\mu}$ for some $\rho \in \R$. On the other hand, using \eqref{app.eq:min_eq101} again, we have that $\nabla f(\rho \bm{\mu})=0$ if and only if $\bm{\mu}^T\nabla f(\rho \bm{\mu})=0$. To see the only if direction, suppose $\bm{\mu}^T\nabla f(\rho \bm{\mu}) = 0$ and $\nabla f(\rho \bm{\mu}) \neq 0$. Then we have $\nabla f(\rho \bm{\mu}) = \bm{v}$ where the vector $\bm{v}$ is nonzero such that $\bm{v}^T\bm{\mu} = 0$. By \eqref{app.eq:min_eq101}, we deduce $\norm{\bm{v}}^2 = \bm{v}^T \nabla f(\rho \bm{\mu}) = 0$ yielding a contradiction. Next, we consider the function, \begin{equation*} g(\rho):=-\bE_{\bm{\xi}}\left[\frac{\bm{\mu}^T\bm{\xi}}{1+\exp(\rho{\bm{\mu}}^T\bm{\xi})} \right]. \end{equation*} Observe that $g(\rho) = \bm{\mu}^T \nabla f(\rho \bm{\mu})$. Therefore, if we can show $g(\rho)$ has a unique zero at $\rho = \tfrac{2}{\sigma^2} =: \rho^*$, we can conclude that $\bm{\mu}^T\nabla f(\rho^* \bm{\mu}) = 0$ which, in turn, gives us that $\rho^* \bm{\mu}$ is the unique solution to $\nabla f(\rho^* \bm{\mu}) = 0$. It remains to show that $\rho^*$ is the unique zero of $g$. By \eqref{eq:fact_affine}, $z: = \bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu}\Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$. Therefore, this yields \begin{equation*} g(\rho)=\frac{1}{\sigma\Vert \bm{\mu} \Vert \sqrt{2\pi}} \int_{-\infty}^{\infty} \frac{z}{1+\exp(\rho z)} \exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2} \right)dz. \end{equation*} Expanding out the term inside the integral, we conclude \begin{align} \frac{z}{1+\exp(\rho z)} \exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2} \right)&=\frac{z}{2\cosh\left(\frac{\rho z}{2}\right)}\exp\left(-\frac{\rho z}{2}-\frac{(z-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2} \right)\nonumber\\ &=\frac{z}{2\cosh\left(\frac{\rho z}{2}\right)}\exp\left(-\frac{z^2+\left(\rho\sigma^2\Vert \bm{\mu} \Vert^2-2\Vert \bm{\mu} \Vert^2 \right)z+\Vert \bm{\mu} \Vert^4}{2\sigma^2\Vert \bm{\mu} \label{app.eq:integrho^*} \Vert^2}\right). \end{align} When $\rho = \rho^*$, we observe that equation \eqref{app.eq:integrho^*} is an odd function of $z$. Therefore, the function $g(\rho^*) = 0$, i.e. the integral of \eqref{app.eq:integrho^*} is $0$. To see that $\rho^*$ is the only zero of $g$, we note that \[ g'(\rho)=\bE_{\bm{\xi}}\left[\frac{\left(\bm{\mu}^T\bm{\xi} \right)^2\exp(\rho \bm{\mu}^T\bm{\xi})}{(1+\exp(\rho \bm{\mu}^T\bm{\xi}))^2} \right] >0. \] Here, $g'(\rho)=0$ implies that $\bm{\mu}^T\bm{\xi}=0$ a.s. which is not true. As a result, the function $g(\rho)$ is strictly decreasing with a zero at $\rho^*$. The result follows. \item \textbf{Hinge loss.} We begin by noting that $f$ is differentiable and it holds that \[ \nabla f(\bm{\theta})= -\bE_{\bm{\xi}}[\bxi1_{\{\bm{\xi}^T\bm{\theta}\leq 1\}}]. \] We next observe that for any $\bm{v},\bm{\theta}\in \R^d$ such that $\bm{v}^T\bm{\theta}=0$, it holds that \begin{equation}\label{eq:hinge_min_lemma} -\bm{v}^T\nabla f(\bm{\theta})=\bE_{\bm{\xi}}[\bm{v}^T\bxi1_{\{\bm{\xi}^T\bm{\theta}\leq 1\}}]=\bE_{\bm{\xi}}[\bm{v}^T\bm{\xi}]\bE_{\bm{\xi}}[1_{\{\bm{\xi}^T\bm{\theta}\leq 1\}}]=\bm{v}^T\bm{\mu}\cdot\bE_{\bm{\xi}}[1_{\{\bm{\xi}^T\bm{\theta}\leq 1\}}]. \end{equation} Here we used that $\bm{\xi}^T\bm{v}$ and $\bm{\xi}^T\bm{\theta}$ are independent random variables and the expectation of the product of two uncorrelated random variables is the product of the expectations. Now note that for any $\bm{\theta}$, the quantity $\bE_{\bm{\xi}}[1_{\{\bm{\xi}^T\bm{\theta}\leq 1\}}]$ is strictly positive. Therefore, if $\bm{v}^T\bm{\theta}=0$ and $\nabla f(\bm{\theta})=\bm{0}$ then, using \eqref{eq:hinge_min_lemma}, we obtain that $\bm{v}^T\bm{\mu}=0$. Hence, we established that $\nabla f(\bm{\theta})=\bm{0}$ implies $\bm{\theta}=\rho \bm{\mu}$ for some $\rho \in \R$. On the other hand, using \eqref{eq:hinge_min_lemma} again, we have that $\nabla f(\rho \bm{\mu})=0$ if and only if $\bm{\mu}^T\nabla f(\rho \bm{\mu})=0$. To see the only if direction, suppose $\bm{\mu}^T\nabla f(\rho \bm{\mu}) = 0$ and $\nabla f(\rho \bm{\mu}) \neq 0$. Then we have $\nabla f(\rho \bm{\mu}) = \bm{v}$ where the vector $\bm{v}$ is nonzero such that $\bm{v}^T\bm{\mu} = 0$. By \eqref{eq:hinge_min_lemma}, we deduce $\norm{\bm{v}}^2 = \bm{v}^T \nabla f(\rho \bm{\mu}) = 0$ yielding a contradiction. Next, consider the function \begin{equation}\label{eq:derivative_hinge} g(\rho)=\bE_{\bm{\xi}}[ \bm{\mu}^T\bm{\xi} 1_{\{\rho \bm{\xi}^T\bm{\mu}\leq 1 \}}]. \end{equation} Observe that $g(\rho) = \bm{\mu}^T \nabla f(\rho \bm{\mu})$. Dominated Convergence Theorem yields that \[ \lim_{\rho \to +\infty}g(\rho)=\bE_{\bm{\xi}}[\bm{\mu}^T\bxi1_{\{\bm{\mu}^T\bm{\xi}\leq 0 \}}], \quad \lim_{\rho \to -\infty}g(\rho)=\bE_{\bm{\xi}}[\bm{\mu}^T\bxi1_{\{\bm{\mu}^T\bm{\xi}\geq 0 \}}]. \] It, therefore, holds that $\lim_{\rho \to +\infty}g(\rho)<0$ and $\lim_{\rho \to -\infty}g(\rho)>0$. Since $g(0)=\bE_{\bm{\xi}}[\bm{\mu}^T\bm{\xi}]>0$, it remains to show that $g$ is a strictly decreasing function. To this end, we note that for any fixed $\rho_1<\rho_2$, it holds that \begin{equation}\label{eq:hinge_g_decreasing} \bm{\mu}^T\bm{\xi}\left(1_{\{\rho_1\bm{\mu}^T\bm{\xi}\leq 1 \}}-1_{\{\rho_2\bm{\mu}^T\bm{\xi}\leq 1 \}} \right) \geq 0 \quad \text{for any value of $\bm{\xi}$}. \end{equation} Indeed, if $\bm{\mu}^T\bm{\xi}\geq 0$, then $\rho_1\bm{\mu}^T\bm{\xi}\leq \rho_2\bm{\mu}^T\bm{\xi}$; thus ensuring $1_{\{\rho_1\bm{\mu}^T\bm{\xi}\leq 1 \}}\geq 1_{\{\rho_2\bm{\mu}^T\bm{\xi}\leq 1 \}}$. The case $\bm{\mu}^T\bm{\xi}\leq 0$ follows similarly. We, therefore, conclude that $g(\rho_1)\geq g(\rho_2)$. Finally, note that $g(\rho_1)=g(\rho_2)$, implies that \eqref{eq:hinge_g_decreasing} holds with equality, almost surely. Clearly, this yields a contradiction. It remains to show \eqref{eq:w_hinge}. By \eqref{eq:derivative_hinge}, we have that $g'(\rho^*)=\bE_{\bm{\xi}}[\bm{\mu}^T\bxi1_{\{\bm{\mu}^T\bm{\xi}\leq \frac{1}{\rho^*}\}}]$. Using \eqref{eq:fact_affine} and \eqref{eqn:fact_truncated}, we obtain that \begin{equation}\label{myeq2} \Phi\left(\frac{1-\rho^*\Vert \bm{\mu} \Vert^2}{\rho^*\sigma\Vert\bm{\mu}\Vert} \right)\cdot \exp\left(\frac{1}{2}\cdot\left(\frac{1-\rho^*\Vert \bm{\mu} \Vert^2}{\rho^*\sigma\Vert\bm{\mu}\Vert} \right)^2\right)=\frac{1}{\sqrt{2\pi}}\cdot\frac{\sigma}{\Vert\mu\Vert}. \end{equation} The result immediately follows. \end{proof} The previous lemma has defined $\bm{\theta}^*$ to be the minimizer of the loss function and showed that it is a positive multiple of $\bm{\mu}$. We now show that this $\bm{\theta}^*$ and its positive scalar multiples are exactly the set of optimal classifiers in the sense of \eqref{eq:stop_criteria_observation}, i.e., we give an exact characterization of that set. \begin{lemma}[Characterization of the optimal classifier] The following is true \begin{equation} \Argmax_{\bm{\theta}} \bP\left(\bm{\xi}^T\bm{\theta}>0\right) = \{ \lambda\cdot \bm{\theta}^*: \lambda>0\}.\end{equation} \end{lemma} \begin{proof} Observe that the following simple fact holds. \begin{equation}\label{app.fact:PhiProb} \bP_{\hat{\bm{\xi}}}\left(\hat{\bm{\xi}}^T\bm{\theta}\geq t\right)=\Phi^c\left( \frac{\bm{\mu}^T\bm{\theta}-t}{\sigma \Vert \bm{\theta}\Vert}\right), \quad \text{for all $\bm{\theta}\in \R^d$, $t\in \R$ and $\hat{\bm{\xi}} \sim N(\bm{\mu},\sigma^2I_d)$}. \end{equation} Therefore we have that $\bP_{\bm{\xi}}(\bm{\xi}^T\bm{\theta}>0)=\Phi^c\left(\frac{\Vert \bm{\mu} \Vert}{\sigma}\cdot\cos(w_{\bm{\theta}}) \right)$ where $\bm{\xi} \sim N(\bm{\mu},\sigma^2I_d)$ and $w_{\bm{\theta}}$ denotes the angle between the two vectors $\bm{\theta}$ and $\bm{\mu}$. On the other hand a classifier $\bm{\theta}$ is optimal if and only if $\bm{\theta}=\rho \bm{\mu}$ for some $\rho > 0$, i.e. $\cos(w_{\bm{\theta}})=0$. The proof is complete after noting that $\Phi$ is an increasing function. \end{proof} \subsubsection{Low Variance Regime}\label{sec:low_log} The main goal of this subsection is to give a proof of Theorem~\ref{theorem_logistic}.1. We restate it below. \begin{manualtheorem}{2.1 }[]\label{thm.a} Let $\bm{\theta}_0=\bm{0}$ and suppose that $\sigma \leq 0.16\Vert \bm{\mu} \Vert$. The following is then true. \begin{equation}\begin{aligned} \bE[T]&\leq 4+\frac{\sigma}{\Vert \bm{\mu}\Vert}\left(\frac{29,584 \cdot (1+\alpha \Vert \bm{\mu}\Vert^2)^2}{\alpha \Vert \bm{\mu}\Vert^2}+15,224 \cdot \sigma^4\alpha^4\Vert \bm{\mu}\Vert^4\right)+\frac{14,792 \cdot (1+\alpha \Vert \bm{\mu}\Vert^2)^2}{\alpha\Vert\bm{\mu}\Vert^2}. \end{aligned} \end{equation} \end{manualtheorem} Recall that in the low variance regime (\textit{i.e.}, $\sigma \leq 0.16\Vert \bm{\mu} \Vert$), we define the target set by \begin{equation}\label{eq:C_low_proof} C=\{\bm{\theta}:\bm{\theta}^T\bm{\mu}\geq 1\} \end{equation} and the test function by \begin{equation}\label{eq:V_low_proof} V=\left(M-\bm{\mu}^T\bm{\theta}\right)^2 \end{equation} where $M=86+86\alpha \Vert \bm{\mu} \Vert^2$. As we discussed earlier, our aim is to show that \[ \bE[\tau_m^C] \underset{\sim}{<} m. \] The following proposition bounds the expected value of the stopping time $\tau_1^C$ provided we initialize with $\bm{\theta}_0=\bm{\theta}$ such that $\bm{\mu}^T\bm{\theta}<1$. \begin{proposition}\label{prp:low_et1} Fix a vector $\bm{\theta} \in \R^d$ such that $\bm{\mu}^T \bm{\theta} < 1$. Suppose that $\sigma\leq 0.16 \Vert \bm{\mu} \Vert$. Consider the set $C$ and the function $V$ defined in \eqref{eq:C_low_proof} and \eqref{eq:V_low_proof} respectively. Then the following bound holds \begin{equation} \bE[\tau_1^{C}|\bm{\theta}_0=\bm{\theta}] \leq \frac{V(\bm{\theta})}{\alpha \Vert \bm{\mu} \Vert^2}. \end{equation} \end{proposition} \begin{proof}[Proof of Proposition \ref{prp:low_et1}] For simplicity we write $\mathcal{F}_{-1}:=\sigma\left(\{\bm{\theta}_0=\bm{\theta}\} \right)$. By expanding out the term, we get for any $k \ge 1$ the following \begin{align} \label{eq: low_noise_blah_1} V(\bm{\theta}_{k})= V(\bm{\theta}_{k-1})-\frac{2\alpha \bm{\mu}^T\bm{\xi}_k(M-\bm{\mu}^T\bm{\theta}_{k-1})}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}+\frac{\alpha^2(\bm{\mu}^T\bm{\xi}_k)^2}{(1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1}))^2}. \end{align} We fix an integer $n \ge 1$ and write the quantity $\bm{\mu}^T\bm{\xi}_k=\Vert\bm{\mu}\Vert^2+\sigma\Vert \bm{\mu}\Vert \psi_k$ with $\psi_k \sim N(0,1)$. We have \begin{align*} &\bE_{\bm{\xi}_k}\left[ \frac{\bm{\mu}^T\bm{\xi}_k}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}|\mathcal{F}_{k-1}\right]1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \qquad =\bE_{\bm{\xi}_k}\left[ \frac{\Vert\bm{\mu}\Vert^2}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}|\mathcal{F}_{k-1}\right]1_{\{ \tau_1^{C}\wedge n \geq k\}}+\sigma \Vert \bm{\mu}\Vert\bE_{\bm{\xi}_k}\left[ \frac{\psi_k}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}|\mathcal{F}_{k-1}\right]1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \qquad \geq \Vert \bm{\mu} \Vert^2 \bE_{\bm{\xi}_k}\left[ \frac{1}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})}|\mathcal{F}_{k-1}\right]1_{\{ \tau_1^{C}\wedge n \geq k\}}-\sigma \Vert \bm{\mu}\Vert\bE_{\psi_k}\left[ \vert \psi_k \vert\right]1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \qquad \ge \Vert \bm{\mu} \Vert^2 \bE_{\bm{\xi}_k}\left[ \frac{1}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})} \left ( 1_{\{\bm{\mu}^T \bm{\theta}_{k-1} \geq \bm{\xi}_k^T \bm{\theta}_{k-1}\}} + 1_{\{\bm{\mu}^T \bm{\theta}_{k-1} < \bm{\xi}_k^T \bm{\theta}_{k-1}\}} \right ) |\mathcal{F}_{k-1}\right]1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \qquad \qquad \qquad -\sigma \Vert \bm{\mu}\Vert \sqrt{\frac{2}{\pi}}1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \qquad \geq \frac{\Vert \bm{\mu}\Vert^2}{1+\exp(\bm{\mu}^T\bm{\theta}_{k-1})}\bE_{\bm{\xi}_k}\left[ 1_{\{\bm{\mu}^T \bm{\theta}_{k-1} \geq \bm{\xi}_k^T \bm{\theta}_{k-1}\}} |\mathcal{F}_{k-1} \right]1_{\{ \tau_1^{C}\wedge n \geq k\}}-\sigma \Vert \bm{\mu}\Vert \sqrt{\frac{2}{\pi}}1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ &\qquad \geq \frac{\Vert \bm{\mu}\Vert^2}{2(1+e)}1_{\{ \tau_1^{C}\wedge n \geq k\}}-\sigma \Vert \bm{\mu}\Vert \sqrt{\frac{2}{\pi}}1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \qquad \geq 0.006\Vert \bm{\mu} \Vert^2 1_{\{ \tau_1^{C}\wedge n \geq k\}}. \end{align*} Here the first inequality follows from $-\bE[X] \le \bE[|X|]$ and $1+ \exp(\bm{\xi}_k^T\bm{\theta}_{k-1}) \ge 1$, the second from ~\eqref{fact:norm_Gaussians}, and the second to last from the observation that for any $X$ normally distributed, $\bP(\bE[X] \ge X) = 1/2$ and $\bm{\xi}_k^T\bm{\theta}_{k-1} \sim N(\bm{\mu}^T\bm{\theta}_{k-1}, \sigma^2 \norm{\bm{\theta}_{k-1}}^2)$ and $\bm{\mu}^T\bm{\theta}_{k-1} < 1$ on the set $\tau_1^{C} \wedge n \ge k$. The last inequality uses the assumption $\sigma \le 0.16 \norm{\bm{\mu}}$. By taking the conditional expectations of \eqref{eq: low_noise_blah_1} combined with the above sequence of inequalities, we deduce the following bound \begin{align*} &\bE\left[V(\bm{\theta}_{k})-V(\bm{\theta}_{k-1})|\mathcal{F}_{k-1}\right]1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \quad \qquad = \bE_{\bm{\xi}_k} \left [-\frac{2\alpha \bm{\mu}^T\bm{\xi}_k(M-\bm{\mu}^T\bm{\theta}_{k-1})}{1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1})} | \mathcal{F}_{k-1} \right] 1_{\{ \tau_1^{C}\wedge n \geq k\}} + \bE_{\bm{\xi}_k} \left [\frac{\alpha^2(\bm{\mu}^T\bm{\xi}_k)^2}{(1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1}))^2} | \mathcal{F}_{k-1} \right] 1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \quad \qquad \leq -0.012(M-1)\alpha \Vert \bm{\mu} \Vert^21_{\{ \tau_1^{C}\wedge n \geq k\}}+ \alpha^2\Vert \bm{\mu} \Vert^2\left(\Vert \bm{\mu}\Vert^2+\sigma^2 \right)1_{\{ \tau_1^{C}\wedge n \geq k\}}\\ & \quad \qquad = \alpha \Vert \bm{\mu} \Vert^2\left[ \alpha\left(\Vert \bm{\mu} \Vert^2+\sigma^2\right)-0.012(M-1)\right]1_{\{ \tau_1^{C}\wedge n \geq k\}}. \end{align*} Here the first inequality follows from $\bm{\mu}^T\bm{\theta}_{k-1} 1_{\{\tau_1^{C} \wedge n \ge k\}} \le 1_{\{\tau_1^{C}\wedge n \geq k\}}$ and upper bounding $\frac{(\bm{\mu}^T\bm{\xi}_k)^2}{(1+\exp(\bm{\xi}_k^T\bm{\theta}_{k-1}))^2}$ by $(\bm{\mu}^T\bm{\xi}_k)^2$ and then applying \eqref{fact:norm_Gaussians}. A quick computation after plugging in the value of $M$ and the bound $\sigma\leq 0.16\Vert \bm{\mu} \Vert$ yields \begin{equation} \label{eq: low_noise_blah_10} \bE\left[V(\bm{\theta}_{k})-V(\bm{\theta}_{k-1})|\mathcal{F}_{k-1}\right]1_{\{ \tau_1^{C}\wedge n \geq k\}} \leq -\alpha \Vert \bm{\mu} \Vert^21_{\{ \tau_1^{C}\wedge n \geq k\}}. \end{equation} Using Appendix Lemma \ref{lem:drift_lemma_appendix}, we obtain that $\bE[\tau_1^{C}\wedge n|\mathcal{F}_{-1}] \leq \frac{V(\bm{\theta})}{\alpha \Vert \bm{\mu} \Vert^2}$. By monotone convergence theorem the result follows. \end{proof} We next upper bound $\bE[\tau_m^C]$ for $m\geq 1$ in the following proposition. \begin{proposition}\label{prp:low_etm} Let $\bm{\theta}_0=\bm{0}$ and suppose that $\sigma\leq 0.16 \Vert \bm{\mu} \Vert$. Consider the set $C$ defined in \eqref{eq:C_low_proof}. Then the following bound holds for all $m \ge 1$ \begin{equation} \bE[\tau_m^C]\leq (m-1)\left(2+\frac{\sigma}{\Vert \bm{\mu}\Vert}\left(\frac{2M^2}{\alpha \Vert \bm{\mu}\Vert^2}+7,612 \cdot \sigma^4\alpha^4\Vert \bm{\mu}\Vert^4\right)\right)+\frac{M^2}{\alpha \Vert \bm{\mu}\Vert^2}, \end{equation} where $M=86+86\alpha \Vert \bm{\mu} \Vert^2$. \end{proposition} \begin{proof}[Proof of Proposition \ref{prp:low_etm}] First, it is clear when $m = 1$, the result holds by Proposition~\ref{prp:low_et1} with $\bm{\theta}_0 = \bm{0}$. We now assume that $\tau_{m-1}^{C}<\infty$ a.s. for some $m\geq 2$ and fix an integer $n \ge 1$. We decompose the space to yield the following bounds \begin{equation} \begin{aligned} \label{eq: low_noise_bound} \bE\left[(\tau_m^{C}-\tau_{m-1}^{C})\wedge n|\mathcal{F}_{\tau_{m-1}^{C}+1}\right]&=\bE\left[((\tau_m^{C}-\tau_{m-1}^{C})\wedge n)|\mathcal{F}_{\tau_{m-1}^{C}+1}\right]1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}\geq 1\}}\\ &+\bE\left[((\tau_m^{C}-\tau_{m-1}^{C})\wedge n)|\mathcal{F}_{\tau_{m-1}^{C}+1}\right]1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}<1\}}\\ &\leq 1+ \bE\left[((\tau_m^{C}-\tau_{m-1}^{C})\wedge n)|\mathcal{F}_{\tau_{m-1}^{C}+1}\right]1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}<1\}}\\ &=1+\sum_{i=1}^{\infty}\bE\left[(\tau_m^{C}-\tau_{m-1}^{C})\wedge n|\mathcal{F}_{\tau_{m-1}^{C}+1}\right]1_{\{i-1\leq 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}<i\}}\\ &\leq 2+\sum_{i=1}^{\infty}\bE\left[\tau_1^{C}\wedge n|\bm{\theta}_0=\bm{\theta}_{\tau_{m-1}^{C}+1}\right]1_{\{i-1\leq 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}<i\}}. \end{aligned} \end{equation} Here the first inequality follows because $((\tau_m^{C}-\tau_{m-1}^{C})\wedge n) 1_{\{\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1} \ge 1\}} \le 1$ by definition of C and the last inequality by the strong Markov property. For each $i \ge 1$, we observe the bound \begin{align} 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1} &=1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}}-\frac{\alpha \bm{\mu}^T\bm{\xi}_{{\tau_{m-1}^{C}}+1}}{1+\exp(\bm{\xi}_{{\tau_{m-1}^{C}}+1}^T\bm{\theta}_{\tau_{m-1}^{C}})}\nonumber\\ &\leq -\frac{\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^{C}+1}}{1+\exp(\bm{\xi}_{\tau_{m-1}^{C}+1}^T\bm{\theta}_{\tau_{m-1}^{C}})} \label{eq:i-1_calculus}\\ & \leq -\alpha \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^{C}+1},\nonumber \end{align} where the second inequality follows because $1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}} \le 0$ and the last inequality because $-\alpha \bm{\mu}^T \bm{\xi}_{\tau_{m-1}^{C}+1}$ is positive. By Proposition~\ref{prp:low_et1} and the above, we deduce \begin{equation} \begin{aligned} \label{eq: low_noise_blah_3} \bE\left[\tau_1^{C}\wedge n|\bm{\theta}_0=\bm{\theta}_{\tau_{m-1}^{C}+1}\right]1_{\{i-1\leq 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}<i\}}&\leq \frac{(M-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1})^2}{\alpha \Vert \bm{\mu} \Vert^2}1_{\{i-1\leq 1-\bm{\mu}^T\bm{\theta}_{\tau_{m-1}^{C}+1}\leq i\}}\\&\leq \frac{(M+i-1)^2}{\alpha \Vert \bm{\mu}\Vert^2}1_{\{ \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^{C}+1}\leq \frac{1-i}{\alpha}\}}. \end{aligned} \end{equation} Finally we observe that \begin{equation} \begin{aligned} \label{eq: low_noise_blah_4} \bE\left[1_{\{ \bm{\mu}^T\bm{\xi}_{\tau_{m-1}^{C}+1}\leq \frac{1-i}{\alpha}\}} \right]&=\bE\left[\sum_{k=1}^{\infty}1_{\{ \bm{\mu}^T\bm{\xi}_{k+1}\leq \frac{1-i}{\alpha}\}}1_{\{\tau_{m-1}^{C}=k\}}\right]\\ &=\sum_{k=1}^{\infty}\bE\left[1_{\{ \bm{\mu}^T\bm{\xi}_{k+1}\leq \frac{1-i}{\alpha}\}} \right]\bE\left[1_{\{\tau_{m-1}^{C}=k\}} \right]\\ &= \Phi\left(\frac{\frac{1-i}{\alpha}-\Vert \bm{\mu}\Vert^2}{\sigma\Vert \bm{\mu}\Vert } \right)\sum_{k=1}^{\infty}\bE\left[1_{\{\tau_{m-1}^{C}=k\}} \right]\\ &=\Phi\left(\frac{\frac{1-i}{\alpha}-\Vert \bm{\mu}\Vert^2}{\sigma\Vert \bm{\mu}\Vert } \right). \end{aligned} \end{equation} The first equality is by independence and the second equality because $\bm{\mu}^T \bm{\xi}_{k+1} \sim N(\norm{\bm{\mu}}^2, \sigma^2 \norm{\bm{\mu}}^2)$. Indeed, the same argument shows $\bm{\xi}_{\tau_{m-1}^{C}+1} \sim N(\bm{\mu},\sigma^2I_d)$, see \cite[Theorem 4.3.1]{Durrett_probability_book}. By combining \eqref{eq: low_noise_bound}, \eqref{eq: low_noise_blah_3}, and \eqref{eq: low_noise_blah_4}, we obtain the following \begin{equation} \begin{aligned} \label{eq: low_noise_blah_5} \bE\left[(\tau_m^{C}-\tau_{m-1}^{C})\wedge n\right]&\leq 2+\sum_{i=1}^{\infty} \frac{(M+i-1)^2}{\alpha \Vert \bm{\mu}\Vert^2}\Phi\left(\frac{\frac{1-i}{\alpha}-\Vert \bm{\mu}\Vert^2}{\sigma\Vert \bm{\mu}\Vert } \right)\\&=2+\sum_{i=1}^{\infty} \frac{(M+i-1)^2}{\alpha \Vert \bm{\mu}\Vert^2}\Phi^{c}\left(\frac{\Vert \bm{\mu}\Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu}\Vert } \right)\\&\le2+\frac{\sigma}{\Vert \bm{\mu}\Vert\sqrt{2\pi}}\sum_{i=1}^{\infty} \frac{(M+i-1)^2}{\alpha\Vert \bm{\mu}\Vert^2+i-1}\exp\left(-\frac{1}{2}\left(\frac{\Vert \bm{\mu}\Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu}\Vert}\right)^2 \right), \end{aligned} \end{equation} where we used the inequality $\Phi^{c}(t)<\frac{1}{t\sqrt{2\pi}}\exp(-\frac{t^2}{2})$. We bound the sum by \begin{equation} \begin{aligned} \label{eq: low_noise_blah_6} \sum_{i=3}^{\infty} &\frac{(M+i-1)^2}{\alpha\Vert \bm{\mu}\Vert^2+i-1}\exp\left(-\frac{1}{2}\left(\frac{\Vert \bm{\mu}\Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu}\Vert}\right)^2 \right)\\ &\qquad = \sum_{i=3}^{\infty} \frac{(M+i-1)^2}{\alpha\Vert \bm{\mu}\Vert^2+i-1} \cdot \frac{4\sigma^4 \norm{\bm{\mu}}^2 \alpha^4}{(\alpha \norm{\bm{\mu}}^2 + i-1)^4} \cdot \frac{(\alpha \norm{\bm{\mu}}^2 + i-1)^4}{4\sigma^4 \norm{\bm{\mu}}^2 \alpha^4} \exp\left(-\frac{1}{2}\left(\frac{\Vert \bm{\mu}\Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu}\Vert}\right)^2 \right)\\ &\qquad \leq 4\alpha^4\sigma^4\Vert \bm{\mu} \Vert^4 \sum_{i=3}^{\infty}\frac{(M+i-1)^2}{(\alpha \Vert \bm{\mu}\Vert^2+i-1)^4}\\ &\qquad \leq 29,584 \cdot \alpha^4\sigma^4\Vert \bm{\mu}\Vert^4 \sum_{i=3}^{\infty}\frac{1}{(i-1)^2} \\ &\qquad \leq 29,584 \cdot \left ( \frac{\pi^2}{6}-1 \right ) \cdot \alpha^4\sigma^4\Vert \bm{\mu}\Vert^4\\ & \qquad \leq 19,080 \cdot \alpha^4\sigma^4 \Vert \bm{\mu}\Vert^4. \end{aligned} \end{equation} Here the first inequality uses $x^2\leq e^x$ for all $x>0$ and $\alpha \norm{\bm{\mu}}^2 + i-1 \ge 1$ for all $i \ge 2$, the second uses that $\frac{M+i-1}{\alpha \Vert \bm{\mu}\Vert^2+i-1}\leq 86$ for all $i \ge 3$ and $\frac{1}{\alpha \Vert \bm{\mu}\Vert^2+i-1}\leq \frac{1}{i-1}$ for $i\geq 2$. We also upper bound \begin{align} \label{eq: low_noise_blah_7} \sum_{i=1}^{2} \frac{(M+i-1)^2}{\alpha\Vert \bm{\mu}\Vert^2+i-1}\exp\left(-\frac{1}{2}\left(\frac{\Vert \bm{\mu}\Vert^2+\frac{i-1}{\alpha}}{\sigma\Vert \bm{\mu}\Vert}\right)^2 \right) &\leq \frac{M^2}{\alpha \Vert \bm{\mu} \Vert^2}+\frac{(M+1)^2}{\alpha \Vert \bm{\mu} \Vert^2+1} \le \frac{5M^2}{\alpha \norm{\bm{\mu}}^2}, \end{align} where the first inequality we use $e^x \le 1$ for all $x<0$ and the second inequality because $\alpha \norm{\bm{\mu}}^2 + 1 \ge \alpha \norm{\bm{\mu}}^2$ and $M +1 \le 2M$. By combining \eqref{eq: low_noise_blah_5}, \eqref{eq: low_noise_blah_6}, and \eqref{eq: low_noise_blah_7}, we obtain \begin{align*} \bE\left[(\tau_m^{C}-\tau_{m-1}^{C})\wedge n\right]&\leq 2+\frac{\sigma}{\Vert \bm{\mu}\Vert}\left(\frac{2M^2}{\alpha \Vert \bm{\mu}\Vert^2}+7,612 \cdot \sigma^4\alpha^4\Vert \bm{\mu}\Vert^4\right). \end{align*} Taking the limit as $n \to +\infty$, we observe that \[\bE[\tau_m^C]\leq 2+\frac{\sigma}{\Vert \bm{\mu}\Vert}\left(\frac{2M^2}{\alpha \Vert \bm{\mu}\Vert^2}+7,612 \cdot \sigma^4\alpha^4\Vert \bm{\mu}\Vert^4\right)+\bE[\tau_{m-1}^{C}]. \] We then iterate the above inequality yielding \[ \bE[\tau_m^C] \le (m-1) \left ( 2+\frac{\sigma}{\Vert \bm{\mu}\Vert}\left(\frac{2M^2}{\alpha \Vert \bm{\mu}\Vert^2}+7,612 \cdot \sigma^4\alpha^4\Vert \bm{\mu}\Vert^4\right) \right ) + \bE[\tau_1^{C}]. \] The result follows after applying Proposition~\ref{prp:low_et1} with $\bm{\theta}_0 = \bm{0}$. \end{proof} We are now ready to prove Theorem~\ref{thm.a}. \begin{proof}[Proof of Theorem \ref{thm.a}] In order to simplify the subsequent argument, we define the quantity, \[ M' := 2+\frac{\sigma}{\Vert \bm{\mu}\Vert}\left(\frac{2M^2}{\alpha \Vert \bm{\mu}\Vert^2}+7,612\sigma^4\alpha^4\Vert \bm{\mu}\Vert^4\right) \] where $M = 86 + 86 \alpha \norm{\bm{\mu}}^2$, we see in Proposition~\ref{prp:low_etm}. By Lemma \ref{lem:st_ETC}, $T_C<\infty$ a.s. and, therefore we have $\sum_{m=1}^{\infty}1_{\{T_C=\tau_m^{C}\}} = 1$ a.s. We next apply Proposition~\ref{prp:low_etm} with Lemma \ref{lem:st_ETC}. Note that $\delta:=\tfrac{1}{2}$ with the target set $C=\{\bm{\theta}:\bm{\mu}^T\bm{\theta}\geq 1 \}$ satisfy \eqref{eq:ST_delta} \begin{align*} \bE[T]&\leq \bE[T_C]=\sum_{m=1}^{\infty} \bE[\tau_C1_{\{T_C=\tau_m^{C}\}}]\leq \sum_{m=1}^{\infty}\frac{\bE[\tau_m^C]}{2^{m-1}} \le \sum_{m=1}^\infty \frac{(m-1)M' + \frac{M^2}{\alpha \norm{\bm{\mu}}^2} }{2^{m-1}} = 2M' + \frac{2M^2}{\alpha \norm{\bm{\mu}}^2}. \end{align*} The result follows after plugging in the value of $M = 86 + 86 \alpha \norm{\bm{\mu}}^2$. \end{proof} \subsubsection{High Variance Regime}\label{sec:high_logistic} The main goal of this subsection is to give a proof of Theorem~\ref{theorem_logistic}.2. We restate it below. \begin{manualtheorem}{2.2}[]\label{thm.b} Let $\bm{\theta}_0=\bm{0}$ and suppose that $ 0.16\Vert \bm{\mu} \Vert \leq \sigma$. Then it holds that $\bE[T]<+\infty$ provided the step-size $\alpha$ satisfies \begin{equation}\label{eq:high_alpha} \alpha < \tfrac{\min\{1,B\}}{\Vert \bm{\mu} \Vert^2+d\sigma^2} \end{equation} where the constant $B$ is defined by \begin{equation} B:=\tfrac{\Vert \bm{\mu}\Vert^4}{4\sigma^4\sqrt{\pi}}\exp\left(-\tfrac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}\right) \Phi^{c}\left( \tfrac{7\sqrt{2}\Vert \bm{\mu} \Vert}{\sigma}\right). \end{equation} \end{manualtheorem} The proof is deferred to the end of this subsection. We first need some preliminaries. Throughout this subsection, the following notation is used \[ \bm{\theta} = \rho\bm{\mu}+ \bm{\tilde{\theta}} \text{ where } \bm{\mu}^T \bm{\tilde{\theta}} = 0. \] We also express $\bm{\theta}_k=\rho_k\bm{\mu}+\tilde{\bm{\theta}}_k$, etc. Recall that in the high variance regime (\textit{i.e.}, $0.16\Vert \bm{\mu} \Vert\leq \sigma$), we define the target set by \begin{equation}\label{eq:C_high_proof_1} C=\left\{\bm{\theta}: \vert \rho\sigma^2-2\vert < 1 \text{ and } \sigma \Vert \tilde{\bm{\theta}} \Vert \leq 3404.46 \right\}.` \end{equation} We first establish \eqref{eq:ST_delta}. \begin{lemma}\label{lem: high_noise_delta} Consider the set defined in \eqref{eq:C_high_proof_1}. Fix $\bm{\theta}\in {C}$ and let $\hat{\bm{\xi}} \sim N(\bm{\mu},\sigma^2 I_d)$. The following bound holds. \begin{equation}\label{eq:deltaP} \bP_{\hat{\bm{\xi}}}\left(\bm{\theta}^T\hat{\bm{\xi}}\geq 1\right)\geq \Phi\left(\frac{-\sigma}{\Vert \bm{\mu} \Vert}\right). \end{equation} \end{lemma} \begin{proof} It is easy to verify that $\bP_{\hat{\bm{\xi}}}\left(\bm{\theta}^T\hat{\bm{\xi}}\geq 1\right)=\Phi\left(\frac{\bm{\theta}^T\bm{\mu}-1}{\sigma \Vert \bm{\theta}\Vert} \right)$. We are interested in lower bounding the term $\frac{\bm{\theta}^T\bm{\mu}-1}{\sigma \Vert \bm{\theta}\Vert}$ independent of $\bm{\theta}$. As $\bm{\theta} \in {C}$, we have $\vert \rho\sigma^2-2\vert < 1$ which yields $\rho \geq \frac{1}{\sigma^2}$. Combining this with $\Vert \bm{\theta} \Vert^2 = \| \rho \bm{\mu} + \tilde{\bm{\theta}} \|^2 \geq \rho^2\Vert \bm{\mu} \Vert^2$, we obtain that $\sigma \Vert \bm{\theta} \Vert\geq \frac{\Vert \bm{\mu} \Vert}{\sigma}$. This gives \begin{equation} \frac{\bm{\theta}^T\bm{\mu}-1}{\sigma \Vert \bm{\theta}\Vert} = \frac{(\rho \bm{\mu} + \tilde{\bm{\theta}})^T\bm{\mu}-1}{\sigma \norm{\bm{\theta}}} \geq \frac{-1}{\sigma \Vert \bm{\theta}\Vert}\geq \frac{-\sigma}{\Vert \bm{\mu} \Vert}. \end{equation} Here the first inequality follows from $\rho>0$ and the second inequality from $\sigma \Vert \bm{\theta} \Vert\geq \frac{\Vert \bm{\mu} \Vert}{\sigma}$. The proof is complete after noting that $\Phi$ is an increasing function. \end{proof} We need the following technical lemma below. The proof is deferred to Appendix Sec. ???. \begin{lemma}\label{lem:high_tech_not_appendix} Consider the following optimization problem \[ \min f(\bm{\theta}):=\bE_{(\bm{\xi},y)\sim \mathcal{P}_*}\left[\ell(\bm{\xi}^T\bm{\theta},y)\right]. \] where $\ell$ is a convex function. The sequences $\{\bm{\theta}_k, \bm{\xi}_k \}_{k = 0}^\infty$ generated by Algorithm~\ref{alg:SGD_termination} satisfy for any $\bm{\theta} \in \R^d$, the following \begin{equation} \label{eq:tech_lemma_inside} f(\bm{\theta}_{k-1})-f(\bm{\theta})\leq \frac{1}{2\alpha}\left(\Vert \bm{\theta}_{k-1}-\bm{\theta} \Vert^2-\bE\left[\Vert \bm{\theta}_{k}-\bm{\theta} \Vert^2\, |\mathcal{F}_{k-1}\right] \right)+\bE[\Vert\nabla_{\bm{\theta}}\ell\left( \bm{\xi}_k^T\bm{\theta}\right)\Vert^2 \, | \, \mathcal{F}_{k-1}], \end{equation} for all $k\geq 1$ where $f$ is defined in \eqref{optimization_problem} and the filtration $\{\mathcal{F}_k\}_{k=0}^{+\infty}$ in \eqref{eq:sigma}. \end{lemma} \begin{remark} Note that in \eqref{eq:tech_lemma_inside}, we are not assuming that $\ell$ is differentiable. We have two choices for the $\ell$ function: logistic loss and hinge loss. Nevertheless, hinge loss is not differentiable, since for any $\bm{\theta}$ w.p. 1 $\bm{\xi}_k^T\bm{\theta}\neq 1$, we conclude that $\ell(\bm{\xi}_k^T\bm{\theta})$ as a function of $\bm{\xi}_k$ is almost everywhere differentiable. \end{remark} Let $\bm{\theta}$ be $\bm{\theta}^*$ in Lemma~\ref{lem:high_tech_not_appendix} and write \begin{equation}\label{eq:high_ftheta_ktheta} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)=f(\bm{\theta}_{k-1})-f(\rho\bm{\mu})+f(\rho\bm{\mu})-f(\bm{\theta}^*). \end{equation} In Lemmas~\ref{lem: logistic_case_2} and \ref{lem: logistic_case_1} below, we will show that both terms $f(\bm{\theta}_{k-1})-f(\rho \bm{\mu})$ and $f(\rho\bm{\mu})-f(\bm{\theta}^*)$ are non-negative. Moreover, we will establish that whenever $\bm{\theta}_{k-1} \not\in C$, either the first or second term can be lower bounded by some positive constant. We first lower bound $f(\bm{\theta}_{k-1})-f(\rho \bm{\mu})$. Let us recall that $\bm{\theta}=\rho \bm{\mu}+\tilde{\bm{\theta}}$ where $\hat{\bm{\theta}}^T\bm{\mu}=0$. \begin{lemma} \label{lem: logistic_case_2} Fix a vector $\bm{\theta} \in \R^d$ and let $\bm{\xi} \sim N(\bm{\mu},\sigma^2I_d)$. Provided that $\sigma \Vert \tilde{\bm{\theta}}\Vert\geq 4\log(\sqrt{2}+1)$, the following holds \begin{equation}\label{eq:high_lem_sigmatheta} \bE_{\bm{\xi}} \left[\log\left( \frac{1+\exp(-\bm{\theta}^T\bm{\xi})}{1+\exp(-\rho \bm{\mu}^T\bm{\xi})}\right)\right] \geq \max\left\{\left ( \tfrac{1}{2} \log(2) + \tfrac{\sigma \Vert \tilde{\bm{\theta}}\Vert}{8} \right )\Phi^c(1)-\left | \frac{\rho}{4}\right | \left(\Vert \bm{\mu}\Vert^2+\sigma \Vert \bm{\mu} \Vert \sqrt{\frac{2}{\pi}} \right) ,0\right\}. \end{equation} \end{lemma} \begin{proof} The two normal random variables, $\bm{\tilde{\theta}}^T\bm{\xi} \sim N(0,\sigma^2\Vert \bm{\tilde{\theta}} \Vert^2)$ and $\bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu} \Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$, are independent by \eqref{eqn:fact_independence}. Since we have $\bE_{\bm{\xi}}[\log(\exp(-\bm{\tilde{\theta}}^T\bm{\xi}))]=\bE_{\bm{\xi}}[\log(\exp(\bm{\tilde{\theta}}^T\bm{\xi}))]=0$, it holds \begin{align*} \bE_{\bm{\xi}} \left[ \log\left (1+\exp(-\bm{\theta}^T\bm{\xi})\right)\right] &= \bE_{\bm{\xi}} \left[\log \left(1 + \exp(-\tilde{\bm{\theta}}^T\bm{\xi}) \exp(-\rho \bm{\mu}^T\bm{\xi}) \right ) \right ]\\ &=\bE_{\bm{\xi}} \left[ \log\left (\exp(\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)\right]\\ &=\bE_{\bm{\xi}} \left[ \log\left (\exp(-\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)\right], \end{align*} where the last equality is true because $\bm{\tilde{\theta}}^T\bm{\xi} \sim -\bm{\tilde{\theta}}^T\bm{\xi}$. Therefore we obtain \begin{align*} \bE_{\bm{\xi}} &\left[ \log\left (1+\exp(-\bm{\theta}^T\bm{\xi})\right)\right]\\ & \qquad \qquad= \frac{1}{2} \bE_{\bm{\xi}} \left[ \log\left (\exp(\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)\right] + \frac{1}{2}\bE_{\bm{\xi}} \left[ \log\left (\exp(-\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)\right] \\ & \qquad \qquad= \frac{1}{2} \bE_{\bm{\xi}} \left[ \log\left ( (\exp(\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi}))(\exp(-\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\rho \bm{\mu}^T\bm{\xi})) \right)\right]\\ &\qquad \qquad=\frac{1}{2}\bE_{\bm{\xi}} \left[\log\left(1+\exp(-\bm{\tilde{\theta}}^T\bm{\xi}-\rho \bm{\mu}^T\bm{\xi})+\exp(\bm{\tilde{\theta}}^T\bm{\xi}-\rho \bm{\mu}^T\bm{\xi})+\exp(-2\rho \bm{\mu}^T\bm{\xi})\right) \right]. \end{align*} By the equality $\exp(\bm{\tilde{\theta}}^T\bm{\xi})+\exp(-\bm{\tilde{\theta}}^T\bm{\xi})= 2+4\sinh^2(\tfrac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})$, we have \begin{align*} \bE_{\bm{\xi}} &\left[ \log\left (1+\exp(-\bm{\theta}^T\bm{\xi})\right)\right]\\ & \qquad \qquad \qquad =\frac{1}{2}\bE_{\bm{\xi}} \left[\log\left(1+2\exp(-\rho \bm{\mu}^T\bm{\xi})+\exp(-2\rho \bm{\mu}^T\bm{\xi})+4\sinh^2(\tfrac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})\exp(-\rho \bm{\mu}^T\bm{\xi})\right) \right]. \end{align*} Therefore, we have \begin{equation} \begin{aligned} \label{eqn: high_noise_blah_20} 2\bE_{\bm{\xi}}\left[\log\left( \frac{1+\exp(-\bm{\theta}^T\bm{\xi})}{1+\exp(-\rho \bm{\mu}^T\bm{\xi})}\right)\right]&=2\bE_{\bm{\xi}}\left[\log(1+\exp(-\bm{\theta}^T\bm{\xi}))\right]-\bE_{\bm{\xi}}\left[\log\left(1+\exp(-\rho \bm{\mu}^T\bm{\xi})\right)^2\right]\\ \qquad \qquad &= \bE_{\bm{\xi}}\left[\log \left ( 1 + \frac{4\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})\exp(-\rho \bm{\mu}^T\bm{\xi})}{(1+ \exp(-\rho \bm{\mu}^T\bm{\xi}))^2} \right ) \right] \ge 0. \end{aligned} \end{equation} Thus it remains to establish the other lower bound in \eqref{eq:high_lem_sigmatheta}. First, we note the following $1+ \exp(-\rho \bm{\mu}^T\bm{\xi}) = 2 \exp( -\tfrac{\rho \bm{\mu}^T\bm{\xi}}{2}) \cosh(\tfrac{\rho \bm{\mu}^T\bm{\xi}}{2})$. Let the constant $r > 0$ and consider the set $\{\bm{\xi}: |\bm{\theta}^T\bm{\xi}| > r\}$. Applying the inequality $x^2+y^2\geq 2\vert xy \vert$ and \eqref{eqn: high_noise_blah_20}, we obtain that \begin{align} 2\bE_{\bm{\xi}}\left[\log\left( \frac{1+\exp(-\bm{\theta}^T\bm{\xi})}{1+\exp(-\rho \bm{\mu}^T\bm{\xi})}\right)\right] & = \bE_{\bm{\xi}}\left[\log\left(1 + \frac{4\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})\exp(-\rho \bm{\mu}^T\bm{\xi})}{(1+ \exp(-\rho \bm{\mu}^T\bm{\xi}))^2}\right) \right] \nonumber \\ &= \bE_{\bm{\xi}}\left[\log\left(1+\frac{\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})}{\cosh^2(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})} \right) \right] \nonumber \\ &\geq \bE_{\bm{\xi}}\left[\log\left(1+\frac{\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})}{\cosh^2(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})} \right) 1_{\{\bm{\xi}:\vert \tilde{\bm{\theta}}^T\bm{\xi}\vert \geq r\}}\right] \label{eq:high_noise_blah_21} \\ &\geq \bE_{\bm{\xi}}\left[\left(\log2+\log\left( \frac{\vert\sinh(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})\vert}{\cosh(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})}\right) \right)1_{\{\bm{\xi}:\vert \tilde{\bm{\theta}}^T\bm{\xi}\vert \geq r\}}\right] \nonumber. \end{align} Here \eqref{eq:high_noise_blah_21} follows from $\log\left(1+\frac{\sinh^2(\frac{\tilde{\bm{\theta}}^T\bm{\xi}}{2})}{\cosh^2(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})} \right)$ is always positive. From \eqref{eq:fact_affine}, we have $\bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu} \Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$ and $\tilde{\bm{\theta}}^T\bm{\xi} \sim N(0,\sigma^2\Vert \tilde{\bm{\theta}} \Vert^2)$. Thus, we write $\tilde{\bm{\theta}}^T\bm{\xi} = \sigma \| \tilde{\bm{\theta}} \| \psi$ where $\psi \sim N(0,1)$. Moreover, we have that $-\log\left(\cosh(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})\right)1_{\{\vert \tilde{\bm{\theta}}^T\bm{\xi}\vert \geq r\}} \geq -\log\left(\cosh(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})\right) $ since $\cosh(\frac{\rho}{2}\bm{\mu}^T\bm{\xi})\geq 1$ always holds. Using the inequality $\log\cosh(x)\leq \vert x\vert$ for all $x$, the following bound holds \begin{equation} \begin{aligned} \label{eq: high_noise_22} &\bE_{\bm{\xi}}\left[\log\left( \frac{1+\exp(-\bm{\theta}^T\bm{\xi})}{1+\exp(-\rho \bm{\mu}^T\bm{\xi})}\right)\right]\\ & \quad \ge \tfrac{1}{2} \log(2) \cdot \bE_{\psi}\left[1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] + \tfrac{1}{2} \bE_{\psi}\left[\log\left\vert\sinh(\tfrac{\sigma\Vert \tilde{\bm{\theta}}\Vert\psi}{2})\right\vert1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] - \tfrac{1}{2}\bE_{\bm{\xi}} \left [ \log( \cosh(\tfrac{\rho}{2} \bm{\mu}^T\bm{\xi})) \right ]\\ & \quad \ge \tfrac{1}{2} \log(2) \cdot \bE_{\psi}\left[1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] + \tfrac{1}{2} \bE_{\psi}\left[\log\left\vert\sinh(\tfrac{\sigma\Vert \tilde{\bm{\theta}}\Vert\psi}{2})\right\vert1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] - \tfrac{1}{2} \bE_{\bm{\xi}} \left [ | \tfrac{\rho}{2} \bm{\mu}^T\bm{\xi} | \right ]\\ &\quad\geq \tfrac{1}{2} \log(2) \cdot \bE_{\psi}\left[1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right] + \tfrac{1}{2} \bE_{\psi}\left[\log\left\vert\sinh(\tfrac{\sigma\Vert \tilde{\bm{\theta}}\Vert\psi}{2})\right\vert1_{\{\vert \psi\vert \geq \frac{r}{\sigma \Vert\tilde{\bm{\theta}}\Vert}\}}\right]-\left | \frac{\rho}{4}\right | \left(\Vert \bm{\mu}\Vert^2+\sigma \Vert \bm{\mu} \Vert \sqrt{\frac{2}{\pi}} \right), \end{aligned} \end{equation} where the last inequality uses \eqref{fact:norm_Gaussians}. Using the inequality $\vert \sinh(x) \vert \geq \exp(\frac{\vert x\vert}{2})$ for $\vert x \vert \geq 2\log(\sqrt{2}+1)$, letting $r=4\log(\sqrt{2}+1)$, we obtain \begin{align} \tfrac{1}{2} \log(2) \cdot \bE_{\psi} \big [ 1_{\{|\psi| \ge \frac{4 \log( \sqrt{2}+ 1)\}}{\sigma \norm{\tilde{\bm{\theta}}} }} \big ] &+ \tfrac{1}{2}\bE_\psi\left[\log\left\vert\sinh(\tfrac{\sigma\Vert \tilde{\bm{\theta}}\Vert\psi}{2})\right\vert1_{\{|\psi| \geq \frac{4\log(\sqrt{2}+1)}{\sigma \Vert \tilde{\bm{\theta}}\Vert}\}}\right]\nonumber\\ &\geq \tfrac{1}{2} \log(2) \cdot \bE_{\psi} \big [ 1_{\{|\psi| \ge \frac{4 \log( \sqrt{2}+ 1)}{\sigma \norm{\tilde{\bm{\theta}}} } \}} \big ] + \tfrac{1}{2}\bE_\psi\left[\left\vert\tfrac{\sigma \Vert \tilde{\bm{\theta}}\Vert \psi}{4} \right\vert1_{\{|\psi| \geq \frac{4\log(\sqrt{2}+1)}{\sigma \Vert \tilde{\bm{\theta}}\Vert}\}}\right]\nonumber\\ &\geq \tfrac{1}{2} \log(2) \cdot \bE_{\psi}[1_{\{\vert\psi\vert \ge 1\}}] + \tfrac{1}{2} \bE_\psi\left[\left\vert\tfrac{\sigma \Vert \tilde{\bm{\theta}}\Vert \psi}{4} \right\vert1_{\vert\psi\vert \geq 1\}}\right] \label{eq:high_noise_blah_111} \\\nonumber &\geq \left ( \tfrac{1}{2} \log(2) + \frac{\sigma \Vert \tilde{\bm{\theta}}\Vert}{8} \right )\Phi^c(1). \end{align} Here \eqref{eq:high_noise_blah_111} follows from $1\geq \frac{4\log(\sqrt{2}+1)}{\sigma \Vert \tilde{\bm{\theta}}\Vert} $. By combining the above inequality with \eqref{eq: high_noise_22}, the desired result holds. \end{proof} We now turn to lower bound the second summand in the right-hand side of \eqref{eq:high_ftheta_ktheta}, namely $f(\rho \bm{\mu})-f(\bm{\theta^*})$. \begin{lemma} \label{lem: logistic_case_1} Provided that $\vert \rho\sigma^2-2 \vert> 1$, the following estimate holds \begin{equation}\label{eq:high_lemhrho} f(\rho \bm{\mu})-f(\bm{\theta^*}) \geq \frac{\Vert \bm{\mu}\Vert^4}{4\sigma^4\sqrt{\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}\right) \Phi^{c}\left( \frac{7\sqrt{2}\Vert \bm{\mu} \Vert}{\sigma}\right). \end{equation} \end{lemma} \begin{proof} Define the function \[ g(\rho):=\bE_{\bm{\xi}}\left[\log\left(1+\exp(-\rho \bm{\mu}^T\bm{\xi})\right) \right], \quad \bm{\xi} \sim N(\bm{\mu},\sigma^2I_d). \] By Lemma~\ref{lem:minimizer_logistic}, we know that $g$ is a convex function with a unique minimizer at $\rho^*:=\frac{2}{\sigma^2}$. Observe that $f(\rho \bm{\mu})-f(\bm{\theta}^*) = g(\rho)-g(\rho^*)$; hence in order to prove \eqref{eq:high_lemhrho}, we instead aim to bound this difference in the function $g$. From \eqref{eq:fact_affine}, we have $\bm{\mu}^T\bm{\xi} \sim N(\Vert \bm{\mu} \Vert^2,\sigma^2\Vert \bm{\mu} \Vert^2)$. It thus holds \[ 4g''(\rho)=\bE\left(\frac{(\bm{\mu}^T\bm{\xi})^2}{\cosh(\frac{\rho}{2} \bm{\mu}^T\bm{\xi})^2} \right)=\frac{1}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi }}\int_{-\infty}^{\infty}\frac{z^2}{\cosh^2(\frac{\rho z}{2})}\exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu} \Vert^2}\right)dz. \] Upper bounding $\cosh^2(\frac{\rho z}{2})$ by $\exp(\vert \rho z\vert)$, we next obtain \begin{align*} 4g''(\rho)&\geq \frac{1}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi}} \int_{-\infty}^{\infty} z^2\exp\left(-\vert \rho z \vert \right)\exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2}{2\sigma^2\Vert \bm{\mu}\Vert^2}\right)dz & \\ &=\frac{1}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi }} \int_{-\infty}^{\infty} z^2\exp\left(-\frac{(z-\Vert \bm{\mu} \Vert^2)^2+2\sigma^2\Vert \bm{\mu}\Vert^2\vert\rho z\vert}{2\sigma^2\Vert \bm{\mu}\Vert^2}\right)dz ,&\\ &=\frac{1}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi }}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right) \int_{-\infty}^{\infty} z^2\exp\left(-\frac{z^2-2\Vert \bm{\mu} \Vert^2 z+2\sigma^2\Vert \bm{\mu}\Vert^2\vert\rho z\vert}{2\sigma^2\Vert \bm{\mu}\Vert^2}\right)dz \end{align*} One can easily verify that whenever $\vert z \vert \geq 2\Vert \bm{\mu} \Vert^2 +2\sigma^2\Vert \bm{\mu}\Vert^2 \vert \rho \vert$ yields \begin{equation*} \exp\left(-\frac{z^2-2\Vert \bm{\mu} \Vert^2 z+2\sigma^2\Vert \bm{\mu}\Vert^2\vert\rho z\vert}{2\sigma^2\Vert \bm{\mu}\Vert^2}\right) \geq \exp\left(-\frac{z^2}{\sigma^2\Vert \bm{\mu}\Vert^2} \right). \end{equation*} Therefore, we have \begin{equation*} 4g''(\rho) \geq \frac{2}{\sigma\Vert \bm{\mu}\Vert\sqrt{2\pi }}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2} \right) \int_{2\Vert \bm{\mu} \Vert^2 +2\sigma^2\Vert \bm{\mu}\Vert^2 \vert \rho \vert}^{\infty} z^2\exp\left(-\frac{z^2}{\sigma^2\Vert \bm{\mu}\Vert^2} \right)dz. \end{equation*} The change of variables $z \mapsto \frac{\sigma\Vert\bm{\mu}\Vert}{\sqrt{2}} z$ yields \begin{align*} g''(\rho) &\geq \frac{\sigma^2\Vert \bm{\mu}\Vert^2}{8\sqrt{\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}\right) \int_{ 2\sqrt{2}(1+2\vert \frac{\rho}{\rho^*}\vert)\frac{\Vert \bm{\mu} \Vert}{\sigma}}^{\infty} z^2\exp\left(-\frac{z^2}{2}\right)dz\\ &\geq \frac{\Vert \bm{\mu} \Vert^4}{\sqrt{\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}\right) \int_{ 2\sqrt{2}(1+2\vert \frac{\rho}{\rho^*}\vert)\frac{\Vert \bm{\mu} \Vert}{\sigma}}^{\infty} \exp\left(-\frac{z^2}{2}\right)dz, \end{align*} where the last inequality holds because $z^2 \ge \frac{8\norm{\bm{\mu}}^2}{\sigma^2}$ on the interval $[2\sqrt{2}(1+2 |\tfrac{\rho}{\rho^*}|) \tfrac{\norm{\bm{\mu}}}{\sigma}, \infty)$. Let $r(\rho) :=2\sqrt{2}(1+2\vert \frac{\rho}{\rho^*}\vert)$. Continuing, we get \begin{equation*} g''(\rho) \geq\frac{\Vert \bm{\mu} \Vert^4}{\sqrt{\pi}}\exp\left(-\frac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}\right) \Phi^{c}\left( \frac{r(\rho)\Vert \bm{\mu} \Vert}{\sigma}\right). \end{equation*} The proof now follows from Lemma \ref{lem:high_tech_not_appendix}. \end{proof} Let us recall that we define the target set $C$ by \begin{equation}\label{eq:C_high_proof} C=\left\{\bm{\theta}: \vert \rho\sigma^2-2\vert < 1 \text{ and } \sigma \Vert \tilde{\bm{\theta}} \Vert \leq c \right\}, \end{equation} where the constant $c$ is \begin{equation}\label{eq:c_high_proof} c=8\left(\frac{\tfrac{3}{2}\left(\frac{1}{0.16^2}+\sqrt{\frac{2}{\pi}}\frac{1}{0.16} \right)+1}{\Phi^c(1)}-\tfrac{1}{2}\log(2)\right) \approx 3404.46. \end{equation} The next proposition establishes the desired bound $\bE[\tau_m^C] \underset{\sim}{<} m.$ Note that the set $C$ is compact. This implies that the value of the test function at the next iteration increases, in expectation, by at most a fixed constant whenever the current iterate is inside $C$. This facilitates establishing bounds for $\bE[\tau_C^m]$ as we will in the proof of following proposition. \begin{proposition}\label{prp:high_V} Let $\bm{\theta}_0=\bm{0}$ and suppose that $0.16\Vert \bm{\mu} \Vert \leq\sigma$. Consider the set $C$ defined in \eqref{eq:C_high_proof}. Let the step-size $\alpha > 0$ satisfy \begin{equation}\label{eq:high_noise_alpha_assumption} \alpha < \frac{\min\{1,B\}}{\Vert \bm{\mu} \Vert^2+d\sigma^2}, \end{equation} where the constant $B$ is \begin{equation} B:=\tfrac{\Vert \bm{\mu}\Vert^4}{4\sigma^4\sqrt{\pi}}\exp\left(-\tfrac{\Vert \bm{\mu} \Vert^2}{2\sigma^2}\right) \Phi^{c}\left( \tfrac{7\sqrt{2}\Vert \bm{\mu} \Vert}{\sigma}\right). \end{equation} Then the following is true \begin{equation}\label{eq:high_Tm} \bE\left[\tau_m^{C}\right]\leq \frac{\Vert \bm{\mu} \Vert^2}{4\alpha\sigma^4\min\{ 1,B\}}+(m-1)\left(2+\frac{\Vert \bm{\mu} \Vert^2+c\sigma^2}{\alpha \sigma^4\min\{1,B\}}\right). \end{equation} \end{proposition} \begin{proof}[Proof of Proposition \ref{prp:high_V}] To simplify the subsequent argument, we define the test function $\overline{V}$ by \[ \overline{V}(\bm{\theta})=\frac{\Vert \bm{\theta}-\bm{\theta}^*\Vert^2}{\alpha\min\{1,B\}}. \] We first establish the drift equation \eqref{eq:st_VoutC} for the test function $\overline{V}$ and the target set \eqref{eq:C_high_proof}. To do so, we first show that \begin{equation}\label{eq:high_thetak} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*) \geq \min\{1,B\}1_{\{\bm{\theta}_{k-1}\not\in {C}\}}. \end{equation} Clearly \eqref{eq:high_thetak} holds when $\bm{\theta}_{k-1}\in {C}$ so suppose that $\bm{\theta}_{k-1}\not\in {C}$. Therefore, either we have $\vert\rho_{k-1}\sigma^2-2\vert>1$ or $\sigma \Vert \tilde{\bm{\theta}}_{k-1}\Vert \geq c$. If $\vert\rho_{k-1}\sigma^2-2\vert>1$, then Lemma \ref{lem: logistic_case_1} yields that $ f(\rho_{k-1}\bm{\mu})-f(\bm{\theta}^*)\geq B$. Then we write that \begin{equation} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)=\underbrace{f(\bm{\theta}_{k-1})-f(\rho_{k-1}\bm{\mu})}_{\geq 0 \text{ by Lemma~\ref{lem: logistic_case_2} }}+f(\rho_{k-1}\bm{\mu})-f(\bm{\theta}^*) \geq B, \end{equation} establishing \eqref{eq:high_thetak}. So assume $\vert \rho_{k-1} \sigma^2-2\vert<1$ and hence $\sigma \Vert \tilde{\bm{\theta}}_{k-1}\Vert \geq c$. We obtain that \begin{align} \bE_{\bm{\xi}_k} \left[\log\left( \frac{1+\exp(-\bm{\theta}_{k-1}^T\bm{\xi}_k)}{1+\exp(-\rho_{k-1} \bm{\mu}^T\bm{\xi}_k)}\right)\right]&\geq \left ( \tfrac{1}{2} \log(2) + \tfrac{\sigma \Vert \tilde{\bm{\theta}}\Vert}{8} \right )\Phi^c(1)-\left | \frac{\rho_{k-1}}{4}\right | \left(\Vert \bm{\mu}\Vert^2+\sigma \Vert \bm{\mu} \Vert \sqrt{\frac{2}{\pi}} \right)\nonumber\\&\geq \left ( \tfrac{1}{2} \log(2) + \tfrac{\sigma \Vert \tilde{\bm{\theta}}\Vert}{8} \right )\Phi^c(1) -\frac{3}{2}\left(\frac{1}{0.16^2}+\frac{1}{0.16} \sqrt{\frac{2}{\pi}} \right)\geq 1\label{eq:high_noise_lastprop_101}, \end{align} where the first inequality follows from Lemma~\ref{lem: logistic_case_2} (Note: $\bm{\xi}_k$ and $\bm{\theta}_{k-1}$ are independent). In the second inequality, we used that $\rho_{k-1}<\tfrac{3}{\sigma^2}$ and also the assumption that $0.16\Vert \bm{\mu} \Vert \leq \sigma$. The last inequality follows from the assumption that $\sigma \Vert\tilde{\bm{\theta}}_{k-1}\Vert\geq c$. We therefore have \[ f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)=\underbrace{f(\bm{\theta}_{k-1})-f(\rho_{k-1}\bm{\mu})}_{\geq 1 \text{ by \eqref{eq:high_noise_lastprop_101}}}+\underbrace{f(\rho_{k-1}\bm{\mu})-f(\bm{\theta}^*)}_{\geq 0 \text{ by Lemma~\ref{lem: logistic_case_1}}} \geq 1. \] We have thus shown \eqref{eq:high_thetak}. By Lemma~\ref{lem:technical_convex_bound}, we have \begin{align} f(\bm{\theta}_{k-1})-f(\bm{\theta}^*)&\leq \tfrac{1}{2\alpha}\left(\Vert \bm{\theta}_{k-1}-\bm{\theta}^* \Vert^2-\bE\left[\Vert \bm{\theta}_{k}-\bm{\theta}^* \Vert^2\, |\mathcal{F}_{k-1}\right] \right)+\tfrac{\alpha}{2}\left (\norm{\bm{\mu}}^2 + d \sigma^2 \right ) \nonumber\\ &\leq \tfrac{1}{2\alpha}\left(\Vert \bm{\theta}_{k-1}-\bm{\theta}^* \Vert^2-\bE\left[\Vert \bm{\theta}_{k}-\bm{\theta}^* \Vert^2\, |\mathcal{F}_{k-1}\right] \right)+\tfrac{1}{2}\min\{1,B\}, \label{eq:high_decrease_v2} \end{align} where in the last inequality we used the assumption on $\alpha$, namely \eqref{eq:high_noise_alpha_assumption}. By \eqref{eq:high_thetak}, we obtain that \[ \min\{1,B\}\left(2\cdot 1_{\{\bm{\theta}_{k-1}\notin C\}}-1\right)\leq \tfrac{1}{\alpha}\left(\Vert \bm{\theta}_{k-1}-\bm{\theta}^* \Vert^2-\bE\left(\Vert \bm{\theta}_{k}-\bm{\theta}^* \Vert^2\, |\mathcal{F}_{k-1}\right] \right). \] A simple manipulation then yields \begin{align} \bE\left[\overline{V}(\bm{\theta}_k)|\mathcal{F}_{k-1}\right] &\leq \overline{V}(\bm{\theta}_{k-1})-2\cdot 1_{\{\bm{\theta}_{k-1}\not\in {C} \}}+1 \nonumber \\&=\overline{V}(\bm{\theta}_{k-1})-1+2\cdot 1_{\{\bm{\theta}_{k-1}\in {C} \}}. \label{eq:high_noise_drift_eq} \end{align} From Appendix Lemma \ref{lem:drift_lemma_appendix}, we thus obtain that \begin{equation}\label{eq:high_T_v2} \bE\left[\tau_m^{C}\right]\leq \overline{V}(\bm{0})+(m-1)\left(2+\sup_{\bm{\theta}\in {C}}\overline{V}(\bm{\theta})\right). \end{equation} It remains to upper bound the quantity $\sup_{\bm{\theta}\in {c}}\overline{V}(\bm{\theta})$. For any $\bm{\theta} \in C$, it holds that $\vert \rho\sigma^2-2\vert<1$. From this, we obtain that \begin{align*} \Vert \bm{\theta}-\bm{\theta}^*\Vert^2&=\left( \rho \sigma^2-\rho^*\sigma^2 \right )^2\frac{\Vert \bm{\mu} \Vert^2}{\sigma^4}+\Vert \tilde{\bm{\theta}}\Vert^2 \\&\leq \frac{\Vert \bm{\mu} \Vert^2}{\sigma^4}+\frac{c}{\sigma^2}, \end{align*} We, therefore, have \begin{equation} \sup_{\bm{\theta}\in {C}}\overline{V}(\bm{\theta})\leq \frac{\Vert \bm{\mu} \Vert^2+c\sigma^2}{\alpha \sigma^4\min\{1,B\}}. \end{equation} The proof is complete. \end{proof} We are now ready to prove Theorem~\ref{thm.b}. \begin{proof}[Proof of Theorem \ref{thm.b}] It follows from Proposition~\ref{prp:high_V} that there exists some constant $r>0$ such that $\bE\left[\tau_m^{C}\right]\leq rm$ for all $m\geq 1$. Taken Lemma~\ref{lem:st_ETC} and Lemma~\ref{lem: high_noise_delta} together, we obtain that \[ \bE[T]\leq \bE[T_C]\leq \sum_{m=1}^{\infty}\bE[\tau_m^{C}]\Phi^{c}\left(\frac{-\sigma}{\Vert \bm{\mu} \Vert}\right)^{m-1}. \] Therefore, \[ \bE[T]\leq \bE[T_C]\leq r\sum_{m=1}^{\infty}m\Phi^{c}\left(\frac{-\sigma}{\Vert \bm{\mu} \Vert}\right)^{m-1}<+\infty. \] The proof is complete. \end{proof}

https://arxiv.org/abs/1604.00699

Anticommutator Norm Formula for Projection Operators

We prove that for any two projection operators $f,g$ on Hilbert space, their anticommutator norm is given by the formula \[\|fg + gf\| = \|fg\| + \|fg\|^2.\] The result demonstrates an interesting contrast between the commutator and anticommutator of two projection operators on Hilbert space. Specifically, the norm of the anticommutator $\|fg + gf\|$ is a simple quadratic function of the norm $\|fg\|$ while the commutator norm $\|fg - gf\|$ is not a function of $\|fg\|$. Nevertheless, the result gives the following bounds that are functions of $\|fg\|$ on the commutator norm: $\|fg\| - \|fg\|^2 \le \|fg - gf\| \le \|fg\|$.

\section{\bf The Main Result}}} The main result of this paper is proving the following norm formula. \medskip \begin{thm}\label{fggf} For any two projection operators $f,g$ on Hilbert space, \begin{align} \|fg + gf\| \ &= \ \|fg\| + \|fg\|^2. \end{align} \end{thm} In particular, the anticommutator norm of projection operators is a simple quadratic function of the norm of their product. This is quite different from the commutator $fg - gf$ of projections since its norm is not a function of the norm of $fg$ (see remark below), nor is $\|fg\|$ a function of $\|fg-gf\|$. In view of Theorem \ref{fggf} we can nevertheless give bounds on the commutator norm that are functions of the norm $\|fg\|$. In this connection, the above theorem then has the following the consequence. \begin{cor}\label{cor} For any two projection operators $f,g$ on Hilbert space, one has \[ \|fg\| - \|fg\|^2 \ \le \ \|fg - gf\| \ \le\ \|fg\| . \] \end{cor} \begin{proof} By Theorem 1.1 we have \[ \|fg-gf\| = \|2gf - (fg+gf)\| \ge 2\|gf\| - \|fg+gf\| \ge \|fg\| - \|fg\|^2. \] Lemma \ref{lemma} below gives the inequality $\|fg - gf\| \le \|fg\|$. \end{proof} \medskip (Lemma \ref{lemma} and its proof are given at the end of the paper.) \medskip We made an application of Theorem \ref{fggf} in \ccite{SWnearlyorthog} in order to obtain sharp upper bound estimates for projection operator on Hilbert space that are nearly orthogonal to their (unitary) symmetries. Specifically, Theorem \ref{fggf} has lead us to obtain the relatively larger bound of $0.455$ that would guarantee that if a projection operator $e$ and a Hermitian unitary operator $w$ on Hilbert space satisfy $\|ewe\| < 0.455$, then a projection operator $q$ exists such that\footnote{The condition $qwq=0$, of course, just means that $q$ is orthogonal to its symmetric image $wqw^*$ under $w$. And so the condition that $\|ewe\|$ is ``small" means that $e$ is nearly orthogonal to its symmetry.} $qwq=0$ and $\|e-q\| \le \frac12\|ewe\| + 4\|ewe\|^2$. (Further, $q$ lies in the C*-subalgebra of $\mathcal B(\mathcal H)$ generated by $e$ and $wew^*$.) \medskip \begin{rmk} It is not hard to see that the commutator norm of projections $\|fg - gf\|$ is generally not a function of the norm $\|fg\|$. For instance, if $f=g\not=0$, then $\|fg\|=1$ and their commutator is zero, while if $p,q$ are the two generating projections of the universal C*-algebra generated by two projections, then $\|pq\|=1$ and $\|pq-qp\| = \frac12$. Conversely, neither is the norm $\|fg \|$ a function of the norm $\|fg-gf\|$, since for the 2 by 2 matrix projections $a = [\smallmatrix 1 & 0 \\ 0 & 0 \endsmallmatrix], \ b = \frac12[\smallmatrix 1 & 1 \\ 1 & 1 \endsmallmatrix]$, one has $\|ab-ba\| = \frac12 = \|pq-qp\|$ while $\|ab\| = \frac1{\sqrt2} \not= 1 = \|pq\|$. \end{rmk} \begin{rmk} We caution that it is not enough to check equalities such as that in Theorem \ref{fggf} for 2 by 2 matrices over the complex numbers and expect that they generally hold for all projections. For example, one can check that the equation \begin{equation}\label{2by2} \|fg - gf\|^2 = \|fg\|^2(1-\|fg\|^2) \end{equation} holds for all projections in $M_2(\mathbb C)$. One can, however, give simple examples of 4 by 4 projections for which equation \eqref{2by2} does not hold. Equation \eqref{2by2} also does not hold for the two projections $p,q$ of the universal C*-algebra generated by two projections since they do not commute and $\|pq\|=1$. \end{rmk} \noindent{\bf Acknowledgments.} This research was partly supported by a grant from the Natural Science and Engineering Council of Canada. \bigskip {\Large{\section{\bf Proof of $\|fg+gf\| \le \|fg\| + \|fg\|^2$}}} \medskip We shall use the following lemma. \medskip \begin{lem}\label{fg} For any two projections $f,g$ and $m\ge1$ we have $\|(fg)^m\| \le \|fg\|^{2m-1}$. \end{lem} \begin{proof} By induction, one checks the equality $(fg)^m = (fgf)^{m-1} (fg)$ (for $m\ge1$). Using $\|fgf\| = \|fg\|^2$, we have \[ \|(fg)^m\| \le \|(fgf)^{m-1}\| \|fg\| \le \|fg\|^{2m-2} \|fg\| = \|fg\|^{2m-1} \] as required. \end{proof} We begin by observing and establishing the following formula for the powers of the anticommutator in terms of polynomials in the operators $fg, gf,$ $fgf$, and $gfg$: \begin{equation}\label{powersplus} (fg + gf)^n = P_n(fg) + P_n(gf) + Q_n(fgf) + Q_n(gfg) \end{equation} where $P_n, Q_n$ ($n=1,2,\dots$) are polynomials given recursively according to the dynamics \begin{align}\label{polyPQ} P_{n+1}(x) &= xP_n(x) + xQ_n(x), \\ Q_{n+1}(x) &= P_n(x) + xQ_n(x) \notag \end{align} with initial data $P_1(x) = x, \ Q_1(x) = 0$.\footnote{Interestingly, these polynomials turn out to be similar to Fibonacci polynomials as they will be given in very similar closed forms in terms of $\sqrt x$.} The equation \eqref{powersplus} can be checked by induction by making strong use of the fact that $f,g$ are projections. In order to find these polynomials in explicit form we express \eqref{polyPQ} in matrix form \[ \bmatrix P_{n+1} \\ Q_{n+1} \endbmatrix = \bmatrix x & x \\ 1 & x \endbmatrix \bmatrix P_{n} \\ Q_{n} \endbmatrix. \] In order to telescope this expression we diagonalize the matrix here as follows: \[ \bmatrix x & x \\ 1 & x \endbmatrix = S \bmatrix x+\sqrt{x} & 0 \\ 0 & x-\sqrt{x} \endbmatrix S^{-1}, \qquad S := \bmatrix \sqrt{x} & -\sqrt{x} \\ 1 & 1 \endbmatrix \] (which is easily checked). Therefore, we calculate the polynomials as follows \begin{align*} \bmatrix P_{n+1} \\ Q_{n+1} \endbmatrix &= \bmatrix x & x \\ 1 & x \endbmatrix^n \bmatrix P_1 \\ Q_1 \endbmatrix = S \bmatrix (x+\sqrt{x})^n & 0 \\ 0 & (x-\sqrt{x})^n \endbmatrix S^{-1} \bmatrix x \\ 0 \endbmatrix \\ \\ &= \frac{1}{2\sqrt{x}} \bmatrix \sqrt{x} & -\sqrt{x} \\ 1 & 1 \endbmatrix \bmatrix (x+\sqrt{x})^n & 0 \\ 0 & (x-\sqrt{x})^n \endbmatrix \bmatrix 1 & \sqrt{x} \\ -1 & \sqrt{x} \endbmatrix \bmatrix x \\ 0 \endbmatrix \\ \\ &= \frac{1}{2\sqrt{x}} \bmatrix \sqrt{x} & -\sqrt{x} \\ 1 & 1 \endbmatrix \bmatrix (x+\sqrt{x})^n & 0 \\ 0 & (x-\sqrt{x})^n \endbmatrix \bmatrix x \\ -x \endbmatrix \\ \\ &= \frac{1}{2\sqrt{x}} \bmatrix \sqrt{x} & -\sqrt{x} \\ 1 & 1 \endbmatrix \bmatrix x(x+\sqrt{x})^n \\ -x(x-\sqrt{x})^n \endbmatrix \end{align*} yielding the closed forms \begin{align*} P_{n+1}(x) &= \frac{x}2 \Big[(x+\sqrt{x})^n + (x-\sqrt{x})^n\Big], \\ Q_{n+1}(x) &= \frac{\sqrt{x}}{2} \Big[(x+\sqrt{x})^n - (x-\sqrt{x})^n\Big]. \end{align*} Next, we express these using their binomial expansions: \begin{align*} (x+\sqrt{x})^n &= \sum_{j=0}^n {n\choose j} x^j x^{\frac12(n-j)} \\ (x-\sqrt{x})^n &= \sum_{j=0}^n {n\choose j} x^j (-1)^{n-j} x^{\frac12(n-j)} \end{align*} Using the notation $\delta_2^k = \frac12 (1+(-1)^k)$ which is 1 when $k$ is even and 0 when $k$ is odd, we can write \[ P_{n+1}(x) = \frac12 x \sum_{j=0}^n {n\choose j} x^j (1 + (-1)^{n-j}) x^{\frac12(n-j)} = x \sum_{j=0}^n {n\choose j} x^j \delta_2^{n-j} x^{\frac12(n-j)}. \] Let us choose odd $n = 2N-1$ so that \begin{align*} P_{2N}(x) &= x \sum_{j=0}^{2N-1} {2N-1\choose j} x^j \delta_2^{2N-1-j} x^{\frac12(2N-1-j)}. \\ &= \sum_{j=0}^{2N-1} {2N-1\choose j} \delta_2^{j-1} x^{N+1 + \frac12(j-1)} \intertext{Now put $j = 2\ell-1$ where $\ell=1,2,\dots, N$ to get} P_{2N}(x) &= \sum_{\ell=1}^{N} {2N-1\choose 2\ell-1} x^{N+\ell}. \end{align*} Now we can compute the norm estimate at $fg$ (or $gf$) as follows: \[ \|P_{2N}(fg)\| \le \sum_{\ell=1}^{N} {2N-1\choose 2\ell-1} \|(fg)^{N+\ell}\|. \] Here we use the inequality $\|(fg)^m\| \le \|fg\|^{2m-1}$ from Lemma \ref{fg} to get \[ \|P_{2N}(fg)\| \le \sum_{\ell=1}^{N} {2N-1\choose 2\ell-1} \|fg\|^{2N+2\ell - 1} = \|fg\|^{2N-1} \sum_{\ell=1}^{N} {2N-1\choose 2\ell-1} \|fg\|^{2\ell} \] we note that the same bound is the same number for $\|P_{2N}(gf)\|$. At this juncture we make use of the identity \[ A_N(a) := \sum_{\ell=1}^{N} {2N-1\choose 2\ell-1} a^{2\ell} = \frac{a}{2}\big[ (1+a)^{2N-1} - (1-a)^{2N-1} \big] \] which we shall call $A_N(a)$ for simplicity, where $a \ge 0$. We can then write \[ \|P_{2N}(fg)\| \le \|fg\|^{2N-1} A_N(\|fg\|) \] and we obtain \[ \|P_{2N}(fg)\| + \|P_{2N}(gf)\| \ \le \ 2 \|fg\|^{2N-1} A_N(\|fg\|). \] Similarly we work out the norms $\|Q_{2N}(fgf)\|$ and $\|Q_{2N}(gfg)\|$. \[ Q_{n+1}(x) = \frac{\sqrt{x}}{2} \sum_{j=0}^n {n\choose j} x^j (1- (-1)^{n-j}) x^{\frac12(n-j)} = \sqrt{x} \sum_{j=0}^n {n\choose j} x^j \delta_2^{n-j-1} x^{\frac12(n-j)} \] and again inserting $n = 2N-1$: \begin{align*} Q_{2N}(x) &= \sqrt{x} \sum_{j=0}^{2N-1} {2N-1\choose j} x^j \delta_2^{2N-1-j-1} x^{\frac12(2N-1-j)} \\ &= \sum_{j=0}^{2N-1} {2N-1\choose j} \delta_2^{j} x^{N + \frac{j}2} \intertext{put $j = 2\ell$ where $\ell = 0, 1, 2, \dots, N-1$:} Q_{2N}(x) &= \sum_{\ell=0}^{N-1} {2N-1\choose 2\ell} x^{N + \ell}. \end{align*} The norm becomes (using $\|(fgf)^m\| \le \|fg\|^{2m}$) \[ \|Q_{2N}(fgf)\| \le \sum_{\ell=0}^{N-1} {2N-1\choose 2\ell} \|(fgf)^{N + \ell}\| \le \sum_{\ell=0}^{N-1} {2N-1\choose 2\ell} \|fg\|^{2N + 2\ell} \] or \[ \|Q_{2N}(fgf)\| \le \|fg\|^{2N} \sum_{\ell=0}^{N-1} {2N-1\choose 2\ell} \|fg\|^{2\ell} = \|fg\|^{2N} B_N(a) \] where we use the identity \[ B_N(a) := \sum_{\ell=0}^{N-1} {2N-1\choose 2\ell} a^{2\ell} = \frac{1}{2}\big[ (1+a)^{2N-1} + (1-a)^{2N-1} \big] \] which we shall call $B_N(a)$ for convenience. Let's write $a = \|fg\|$. Then we get \begin{align*} \|(fg+gf)^{2N}\| &= \|P_n(fg) + P_n(gf) + Q_n(fgf) + Q_n(gfg)\| \\ &\le 2 \|P_n(fg)\| + 2\|Q_n(fgf)\| \\ &\le 2 \|fg\|^{2N-1} A_N(\|fg\|) + 2\|fg\|^{2N} B_N(\|fg\|) \\ &= 2 a^{2N-1} A_N(a) + 2 a^{2N} B_N(a) \end{align*} \begin{align*} \ \ &= 2 a^{2N-1} \cdot \frac{a}{2}\big[ (1+a)^{2N-1} - (1-a)^{2N-1} \big] + 2 a^{2N} \cdot \frac{1}{2}\big[ (1+a)^{2N-1} + (1-a)^{2N-1} \big] \\ &= a^{2N} \cdot \big[ (1+a)^{2N-1} - (1-a)^{2N-1} \big] + a^{2N} \cdot \big[ (1+a)^{2N-1} + (1-a)^{2N-1} \big] \\ &= 2 a^{2N} \cdot (1+a)^{2N-1}. \end{align*} Taking $2N$-th roots, \[ \|fg+gf\| \le 2^{1/2N} a (1+a)^{1-\frac1{2N}} \] which in the limit as $N \to\infty$ gives \[ \|fg+gf\| \le \|fg\| + \|fg\|^2. \] \bigskip {\Large{\section{\bf Proof of $\|fg+gf\| \ge \|fg\| + \|fg\|^2$}}} Any two projection operators $f,g$ on Hilbert space $\mathcal H$ may be represented in matrix block forms as \[ f = \bmatrix I & 0 \\ 0 & 0 \endbmatrix, \quad g = \bmatrix D & V \\ V^* & D' \endbmatrix \] with respect to the orthogonal decomposition $\mathcal H = \mathcal M \oplus \mathcal M^\perp$ where $\mathcal M$ is the range of $f$ (and $\mathcal M^\perp$ that of $1-f$). Here, $D, D'$ are positive operators on $\mathcal M$ and $\mathcal M^\perp$, respectively, and $V:\mathcal M^\perp \to \mathcal M$, satisfying the relations \[ D - D^2 = VV^*, \quad DV + VD' = V, \quad D' - D'^2 = V^*V. \] (in view of $g$ being a projection). Since \[ fg(fg)^* = \bmatrix D & V \\ 0 & 0 \endbmatrix \bmatrix D & 0 \\ V^* & 0 \endbmatrix = \bmatrix D^2+VV^* & 0 \\ 0 & 0 \endbmatrix = \bmatrix D & 0 \\ 0 & 0 \endbmatrix \] we see that $\|fg\|^2 = \|D\|$. We now work out the powers of the anticommutator as follows. First, we have \[ fg + gf = \bmatrix D & V \\ 0 & 0 \endbmatrix + \bmatrix D & 0 \\ V^* & 0 \endbmatrix = \bmatrix 2D & V \\ V^* & 0 \endbmatrix. \] We observe that the powers of this anticommutator have the form \[ (fg + gf)^n = \bmatrix F_n(D) & F_{n-1}(D)V \\ \star & \star \endbmatrix \] where $F_n(x)$ is a certain sequence of polynomials with integer coefficients -- with initial data $F_1(x) = 2x, F_0(x) = 1$. We do not need to know the $\star$ entries at the bottom of the matrix because we are interested in the northwestern\footnote{It is rather interesting in this case that the information regarding the anticommutator norm is contained in this block for large $n$.} block $F_n(D)$ of $(fg + gf)^n$. Multiplying \[ \bmatrix F_n(D) & F_{n-1}(D)V \\ \star & \star \endbmatrix \bmatrix 2D & V \\ V^* & 0 \endbmatrix = \bmatrix 2DF_n(D) + F_{n-1}(D)VV^* & F_n(D)V \\ \star & \star \endbmatrix \] or \[ (fg + gf)^{n+1} = \bmatrix 2DF_n(D) + (D - D^2)F_{n-1}(D) & F_n(D)V \\ \star & \star \endbmatrix \] from which we see that the polynomial sequence has the Fibonacci-type recursion relation \[ F_{n+1}(x) = 2xF_n(x) + (x - x^2)F_{n-1}(x). \] Telescoping this as one would with ordinary Fibonacci numbers, one eventually gets \[ F_n(x) = \frac12 x^{n/2} \left[ (\sqrt x + 1)^{n+1} - (\sqrt x - 1)^{n+1} \right]. \] (Indeed, one can check this by induction.) This polynomial function is (for each $n$) an increasing function over the interval $[0,\infty)$. Therefore, since $D$ is a positive operator, we have \[ \|(fg + gf)^n\| \ge \|F_n(D)\| = F_n(\|D\|) = F_n(\|fg\|^2) = \frac12 \|fg\|^n \left[ (\|fg\| + 1)^{n+1} - (\|fg\| - 1)^{n+1} \right] \] \[ = \frac12 \|fg\|^n (\|fg\| + 1)^{n+1} \left[ 1 - \left(\frac{\|fg\| - 1}{\|fg\| + 1}\right)^{n+1} \right]. \] Taking $n$-th roots (noting that the anticommutator $fg+gf$ is a Hermitian operator) \[ \|fg + gf\| \ge \frac1{2^{1/n}} \|fg\|\, (\|fg\| + 1)^{1+\frac1n} \left[ 1 - \left(\frac{\|fg\| - 1}{\|fg\| + 1}\right)^{n+1} \right]^{1/n}. \] Letting $n\to\infty$ the right side converges to $\|fg\| + \|fg\|^2$ (since $(1-c^n)^{1/n} \to 1$ for\footnote{If $-1<c<1$ then $-1< c^n \le |c| < 1$ for each $n\ge1$, which gives $0<1-|c| \le 1-c^n < 2$ and the result follows by taking $n$-th roots.} any $-1 < c < 1$). This completes the proof of Theorem \ref{fggf}. \bigskip We end the paper with the proof of the lemma used by Corollary \ref{cor}. \medskip \begin{lem}\label{lemma} For any two projections $f,g$ on Hilbert space, $\|fg - gf\| \le \|fg\|$. Further, $\|fg - gf\| = \|fg-fgf\|$. \end{lem} \begin{proof} Write \[ \|fg - gf\|^2 = \|(fg-gf)^*(fg-gf)\| = \| fgf + gfg - fgfg - gfgf \| \] and note that the operator in the last norm can be written as the sum of two orthogonal positive operators: \[ fgf + gfg - fgfg - gfgf = fg(1-f)gf + (1-f)gfg(1-f) = uu^* + u^*u \] where $u = fg(1-f)$. So its norm is the max of the norms of each term, both of which are equal to $\|u\|^2$. Thus, \[ \|fg - gf\| = \|u\| = \|fg-fgf\| = \|fg(1-f)\| \le \|fg\| \] as needed. \end{proof} \medskip \noindent{\bf Note added in proof.} A generous colleague pointed out to the author that with a bit more work, one can deduce the anticommutator norm formula from a theorem of Halmos in \ccite{Halmos}. Our proof, however, is self-contained and independent of this -- and the formula (simple as it is) seems to be unknown. \bigskip

https://arxiv.org/abs/0801.0120

Combinatorics of the change-making problem

We investigate the structure of the currencies (systems of coins) for which the greedy change-making algorithm always finds an optimal solution (that is, a one with minimum number of coins). We present a series of necessary conditions that must be satisfied by the values of coins in such systems. We also uncover some relations between such currencies and their sub-currencies.

\section{Introduction} In the change-making problem we are given a set of coins and we wish to determine, for a given amount $c$, what is the minimal number of coins needed to pay $c$. For instance, given the coins $1,2,5,10,20,50$, the minimal representation of $c=19$ requires $4$ coins ($10+5+2+2$). This problem is a special case of the general knapsack problem with all coins of unit weights. In some cases the solution may be found by a greedy strategy that uses as many of the largest coin as possible, then as many of the next one as possible and so on. This greedy solution is optimal for the set of coins given above, but fails to be optimal in general. For instance, if we have coins $1,5,9,16$, then the amount $18$ will be paid greedily as $16+1+1$, while the optimal solution ($9+9$) requires just two coins. In this paper we shall concentrate on the combinatorial properties of those sets of coins for which the greedy solution is always optimal. The sequence $A=(a_0,a_1,\ldots,a_k)$, where $1=a_0<a_1<\ldots<a_k$ will be called a {\it currency} or {\it coinage system}. We always assume $a_0=1$ is the smallest coin to avoid problems with non-representability of certain amounts. For any amount $c$ by $\textrm{opt}_A(c)$ and $\textrm{grd}_A(c)$ we denote, respectively, the minimal number of coins needed to pay $c$ and the number of coins used when paying $c$ greedily (for example, if $A=(1,5,9,16)$ then $\textrm{opt}_A(18)=2$ and $\textrm{grd}_A(18)=3$). The currency $A$ will be called {\it orderly}\footnote{Various authors have used the terms orderly \cite{Jones,Maurer}, canonical \cite{KoZa,Pear}, standard \cite{TienHu} or greedy \cite{CCS}.} if for all amounts $c>0$ we have $\textrm{opt}_A(c)=\textrm{grd}_A(c)$. If a coinage system $A$ is not orderly then any amount $c$ for which $\textrm{opt}_A(c)<\textrm{grd}_A(c)$ will be called a {\it counterexample}. Let us briefly summarize related work. Magazine, Nemhauser and Trotter \cite{MNT} gave a necessary and sufficient condition to decide whether $A=(1,a_1,\ldots,a_{k+1})$ is orderly provided we know in advance that $A'=(1,a_1,\ldots,a_k)$ is orderly, the so-called one-point theorem (see section \ref{section2} of the present paper). Kozen and Zaks \cite{KoZa} proved, among other things, that the smallest counterexample (if exists), does not exceed the sum of the two highest coins. They also asked if there is a polynomial-time algorithm that tests if a coinage system is orderly. Such an algorithm was presented by Pearson \cite{Pear}. It produces a set of $O(k^2)$ ``candidates for counterexamples'', which is guaranteed to contain the smallest counterexample if one exists. The rest of the algorithm is just testing these potential candidates, and the overall complexity is $O(k^3)$. A similar set of possible counterexamples (perhaps not containing the smallest one), but of size $O(k^3)$, was given by Tien and Hu in \cite{TienHu} (see formula (4.20) and Theorem 4.1 of that paper). It leads to an $O(k^4)$ algorithm. The authors of \cite{MNT} and \cite{TienHu} were concentrating mainly on the error analysis between the greedy and optimal solutions. Apparently Jones \cite{Jones} was the only one who attempted to give a neat combinatorial condition characterizing orderly currencies, but his theorem suffered from a major error, soon pointed out by Maurer \cite{Maurer}. Our paper has been paralleled by an independent work of Cowen, Cowen and Steinberg \cite{CCS} about currencies whose all prefixes are orderly and about non-orderly currencies which cannot be ``fixed'' by appending extra coins. The results contained in section \ref{section4} and a special case ($l=2$) of Theorem \ref{theoremprefixorderly} of this paper have also been proved in \cite{CCS}. The aim of this paper is to study some orderly coinage systems from a combinatorial viewpoint, motivated by the need to have some nice characterization. One of the motivations was the observation that \emph{if $A=(1,a_1,a_2,\ldots,a_k)$ is an orderly currency, then the currency $(1,a_1,a_2)$ is also orderly}, that will be generalized and proved in Theorem \ref{theoremprefixorderly}. Going further, one may start with an orderly currency, take out some of its coins and ask if the remaining coins again form an orderly currency. The precise answer to this question, given in section \ref{section7}, will be a consequence of the results of sections \ref{section3} and \ref{section5}, where we prove some properties of the distances $a_j-a_i$ between the coins of an orderly currency. In section \ref{section4} these results will be used to give a complete description of orderly currencies with less than 6 coins. In section \ref{section6} we study the behaviour of the currencies obtained as prefixes of an orderly currency. Some closing remarks and open problems are included in section \ref{section8}. \section{Preliminary results} \label{section2} If $A=(1,a_1,\ldots,a_k)$ is a currency it will often be convenient to set $a_{k+1}=\infty$. This will be especially useful whenever we want to choose, say, the first interval $[a_m,a_{m+1}]$ of length at least $d$ for some $d$. The reader will see that in all applications the infinite interval will be as useful as proper intervals. There are three standard arguments that will be used repeatedly throughout this paper, so we quote them now to avoid excessive repetitions in the future. All the time we assume $A=(1,a_1,\ldots,a_k)$ is orderly. First, suppose we have $a_m$, $a_i$ and $a_l$, such that $a_l<a_m+a_i<a_{l+1}$. Then $a_m+a_i$ has a representation that uses 2 coins. Since $a_m+a_i$ is strictly between $a_l$ and $a_{l+1}$ its greedy representation must start with $a_l$, followed (since $A$ is orderly) by just one other coin $a_r$. It follows that there exists $r$ such that $a_m+a_i=a_l+a_r$. The second argument is a slight modification of the first one; namely, if $a_l<a_m+a_i$ and the number $a_m+a_i-a_l$ is not one of the coins, then $a_{l+1}\leq a_m+a_{i}$. The third argument is a bit more complicated. Suppose that for some $j>i\geq 1$ we have $a_j-a_i=d$. Let us choose the largest $m$ for which $a_m-a_{m-1}<a_i$ (such $m$'s exist, for instance $a_i-a_{i-1}<a_i$). Then $a_{m+1}-a_m\geq a_i$ (here it is possible that $a_{m+1}=\infty$), so we have $a_m<a_{m-1}+a_i<a_{m+1}$. If we also have $a_m<a_{m-1}+a_j<a_{m+1}$ then, as before, there exist numbers $r<i$ and $s<j$ such that: \begin{center} \begin{tabular}{c} $a_{m-1}+a_i=a_m+a_r,$\\ $a_{m-1}+a_j=a_m+a_s.$ \end{tabular} \end{center} \begin{figure} \epsfbox{figs.1} \caption{Illustration of a standard argument} \end{figure} Then $a_s-a_r=a_j-a_i=d$, so we decreased the coins' indicies from $(j,i)$ to $(s,r)$, keeping the difference $d$ unchanged. Therefore, if additionally $(j,i)$ was the {\it smallest} pair of indices for which $a_j-a_i=d$, we would have a contradiction, hence we may assume that in such case $a_{m-1}+a_j\geq a_{m+1}$. We shall frequently make use of the following famous result: \begin{theorem}[{\bf One-point theorem, \cite{MNT,HuLen,CCS}}] \label{onept} Suppose $A'=(1,a_1,\ldots,a_k)$ is orderly and $a_{k+1}>a_k$. Let $m=\lceil a_{k+1}/a_{k}\rceil$. Then $A=(1,a_1,\ldots,a_k,a_{k+1})$ is orderly if and only if $\textrm{opt}_A(ma_k)=\textrm{grd}_A(ma_k)$. \end{theorem} {\bf Remark.} According to this theorem, if the shorter currency $A'$ is orderly, then the optimality of the greedy solution for $A$ needs to be checked only for the single value $ma_k$. This justifies the name {\it one-point theorem}. Note, that although in general it is NP-hard to compute $\textrm{opt}_A(c)$ for arbitrary $A$ and $c$ (see \cite{Shallit} and \cite{KoZa} for a discussion), the one-point theorem test $\textrm{opt}_A(ma_k)=\textrm{grd}_A(ma_k)$ runs in polynomial time, since it is equivalent to $\textrm{grd}_{A'}(ma_k-a_{k+1})\leq m-1$. For the sake of completeness we decided to include a short proof of the one-point theorem. {\bf Proof.} One of the implications is trivial. Now suppose that $\textrm{opt}_A(ma_k)=\textrm{grd}_A(ma_k)$. We have $$(m-1)a_k+1\leq a_{k+1}\leq ma_k.$$ For all values $c< a_{k+1}$ all the payments $\textrm{grd}_A(c)$, $\textrm{grd}_{A'}(c)$, $\textrm{opt}_A(c)$, $\textrm{opt}_{A'}(c)$ coincide, so $\textrm{grd}_A(c)=\textrm{opt}_A(c)$. All other $c$ will be split in two groups: $c\in[a_{k+1}, ma_k)$ and $c\geq ma_k$. {\bf 1. $a_{k+1}\leq c<ma_k$.} For every such $c$ we have $c<2a_{k+1}$, therefore any payment of $c$ contains either $0$ or $1$ copies of $a_{k+1}$. Together with the orderliness of $A'$ this implies $$\textrm{opt}_A(c)=\min\{1+\textrm{grd}_{A'}(c-a_{k+1}), \textrm{grd}_{A'}(c)\}.$$ At the same time $1+\textrm{grd}_{A'}(c-a_{k+1})=\textrm{grd}_A(c)$, so in order to prove $\textrm{opt}_A(c)=\textrm{grd}_A(c)$ it suffices to show the inequality $$\textrm{grd}_A(c)\leq \textrm{grd}_{A'}(c).$$ Observe that $$\textrm{grd}_{A'}(c)=(m-1)+\textrm{grd}_{A'}(c-(m-1)a_k).$$ The function $\textrm{grd}_{A'}=\textrm{opt}_{A'}$ satisfies the triangle inequality, so $$\textrm{grd}_{A'}(ma_k-a_{k+1})+\textrm{grd}_{A'}(c-(m-1)a_k)\geq \textrm{grd}_{A'}(c-a_{k+1}+a_k)=1+\textrm{grd}_{A'}(c-a_{k+1}).$$ Finally $$\textrm{grd}_{A'}(c)-\textrm{grd}_{A}(c)=(m-1)+\textrm{grd}_{A'}(c-(m-1)a_k)-(1+\textrm{grd}_{A'}(c-a_{k+1} ))\geq$$ $$\geq m-2 + 1 -\textrm{grd}_{A'}(ma_k-a_{k+1})=m-1-\textrm{grd}_{A'}(ma_k-a_{k+1}).$$ However $$1+\textrm{grd}_{A'}(ma_k-a_{k+1})=\textrm{grd}_A(ma_k)=\textrm{opt}_A(ma_k)\leq m,$$ which eventually implies the desired inequality $$\textrm{grd}_{A'}(c)-\textrm{grd}_{A}(c)\geq 0.$$ {\bf 2. $c\geq ma_k$.} Denote by $\mathcal{OPT}(c)$ the set of optimal payments for $c$: $$\mathcal{OPT}(c) = \{(x_0,\ldots,x_{k+1}): \sum_{i=0}^{k+1}x_ia_i=c \textrm{ and } \sum_{i=0}^{k+1} x_i \textrm{ is minimal}\}$$ It is sufficient to exhibit a payment $(x_i)\in\mathcal{OPT}(c)$ with $x_{k+1}>0$. Consider any optimal payment $(x_i)$. We may apply to it the following two operations: \begin{itemize} \item if $x_k\geq m$ then replace $m$ coins $a_k$ with the greedy decomposition of $ma_k$. This way the number of coins in the payment does not increase (since $\textrm{opt}_A(ma_k)=\textrm{grd}_A(ma_k)$), while the multiplicity of $a_k$ in the payment decreases. \item if $\sum_{i=0}^{k-1}x_ia_i\geq a_k$ then instead of the coins needed to pay $\sum_{i=0}^{k-1}x_ia_i$ insert the greedy decomposition of this amount with respect to $A'$. This will not increase the overall number of coins (since $A'$ was orderly), but it will decrease the amount paid with $1,a_1,\ldots,a_{k-1}$. \end{itemize} It is clear that repeating these two steps sufficiently many times we will finally end up with an optimal payment $(x_i)$ satisfying $\sum_{i=0}^{k-1}x_ia_i<a_k$ and $x_k<m$. Then $$\sum_{i=0}^{k}x_ia_i\leq a_k-1+(m-1)a_k=ma_k-1<c$$ hence $x_{k+1}>0$ in this payment. \qed It is obvious that the one-coin currency $A=(1)$ is orderly, as well as all the two-coin currencies $A=(1,a_1)$. The reader may now wish to solve the easy problem of when a three-coin currency $A=(1,a_1,a_2)$ is orderly. For reasons which will become clear later we shall express the solution in terms of the following set: \begin{definition} \label{defA} For any $a>0$ we define: $$\mathcal{A}(a)=\bigcup_{m=1}^{\infty} \bigcup_{l=0}^m \{ma-l\}=$$ $$=\{a-1,a\}\cup\{2a-2,2a-1,2a\}\cup\ldots\cup\{ma-m,\ldots,ma\}\cup\ldots$$ \end{definition} \begin{proposition} \label{lemma3orderly} The currency $A=(1,a_1,a_2)$ is orderly if and only if $a_2-a_1\in\mathcal{A}(a_1)$. \end{proposition} {\bf Proof.} Let $m=\lceil a_2/a_1\rceil$. By the one-point theorem $A$ is orderly if and only if the greedy algorithm is optimal for $ma_1$, which is equivalent to $$\textrm{grd}_A(ma_1)\leq m$$ or $$ma_1-a_2\leq m-1$$ which means that $a_2-a_1 = (m-1)a_1 - (ma_1-a_2) \in \mathcal{A}(a_1)$ (more precisely, $a_2-a_1$ belongs to the $(m-1)$-st summand of $\mathcal{A}(a_1)$). On the other hand, if $m$ is the least number for which $a_2-a_1$ belongs to the $(m-1)$-th summand of $\mathcal{A}(a_1)$, then $\lceil a_2/a_1\rceil=m$ and $a_2-a_1=(m-1)a_1-l$ for some $l\leq m-1$. Then $$\textrm{grd}_A(ma_1)=1+(ma_1-a_2) = 1+l\leq m$$ as required.\qed \section{Investigating differences, part I} \label{section3} In this section we begin investigating distances between the coins of an orderly coinage system, followed by an application of these results. \begin{proposition} \label{lemmanot1} If $A=(1,a_1,\ldots,a_k)$ is orderly and $a_1\geq 3$, then $$a_{i}-a_{i-1}\neq 1$$ for all $i=1,\ldots,k$. \end{proposition} {\bf Proof.} Suppose, on the contrary, that $a_j-a_{j-1}=1$ and let $j$ be the least index with this property. Since $a_1\geq 3$, we have $j\geq 2$. Let us choose the largest index $m$ for which $a_m-a_{m-1}<a_{j-1}$. Then $a_{m+1}-a_m\geq a_{j-1}$, and if $a_{m-1}+a_j<a_{m+1}$ then we would have a contradiction by the third standard argument from section \ref{section2}. Therefore $a_{m+1}\leq a_{m-1}+a_j$. Since $a_{m+1}-a_m\geq a_{j-1}$, we have $$a_j=a_{j-1}+1\leq (a_{m+1}-a_m)+(a_m-a_{m-1})=a_{m+1}-a_{m-1}\leq a_j$$ meaning that $a_m-a_{m-1}=1$ and $a_{m+1}-a_m=a_{j-1}$. It follows that $$a_m<a_{m-1}+a_{j-1}<a_{m+1}$$ which means that $a_{m-1}+a_{j-1}-a_m=a_{j-1}-1$ must be one of the coins, contradicting the minimality of $j$. This ends the proof.\qed The previous proposition can be sharpened as follows: \begin{proposition} \label{lemmadiffa1-1} If $A=(1,a_1,\ldots,a_k)$ is orderly then $$a_{i}-a_{i-1} \geq a_1-1$$ for all $i=1,\ldots,k$. \end{proposition} {\bf Proof.} This is obviously true if $a_1=2$, so let $a_1\geq 3$. Let $j$ be the largest index for which $a_j-a_{j-1}\leq a_1-2$. By Proposition \ref{lemmanot1} we have $a_j-a_{j-1}\geq 2$. From the maximality of $j$ we have $a_{j+1}-a_j\geq a_1-1$ (it is possible that $a_{j+1}=\infty$). Now consider the amount $$c=a_{j-1}+a_1.$$ \begin{figure} \epsfbox{figs.2} \caption{Illustration of the proof of Prop. \ref{lemmadiffa1-1}} \end{figure} It satisfies $a_j+2\leq c \leq a_j+a_1-2 < a_{j+1}$, hence $\textrm{opt}_A(c)=2$. When paid greedily, the amount $c$ is decomposed to $a_j$ and $c-a_j$ copies of the coin $1$, which makes $$1+(c-a_j)\geq 1+2=3$$ coins altogether. This contradicts the fact $A$ is orderly, thus completing the proof of this proposition.\qed Proposition \ref{lemmadiffa1-1} imposes certain restrictions on the possible differences $a_i-a_{i-1}$. In the next theorem we shall generalize this restriction, but first let us state without proof some obvious properties of the sets $\mathcal{A}(a)$ that will be useful in the proof: \begin{fact} Let $a\geq 2$ be an integer. Then: \label{factabouta} \begin{itemize} \item[(1)] if $x,y\in\mathcal{A}(a)$ then $x+y\in\mathcal{A}(a)$. \item[(2)] an integer $x\geq 2$ does not belong to $\mathcal{A}(a)$ if and only if there exists an integer $p\geq 0$ such that $$pa+1\leq x\leq (p+1)a-(p+2).$$ \item[(3)] if $p_1<p_2<\ldots<p_m$ and $p_m-p_{1}\not\in\mathcal{A}(a)$ then $p_j-p_{j-1}\not\in\mathcal{A}(a)$ for some $2\leq j\leq m$ (this follows from (1)). \item[(4)] if $pa<x$ and $x=(p+1)a-c$ for some $c$ (possibly negative), then $x\in\mathcal{A}(a)$ implies $c\leq p+1$. \end{itemize} \end{fact} Now we can state the main theorem of this section: \begin{theorem} \label{theoremdiffaa} If $A=(1,a_1,\ldots,a_k)$ is orderly then $$a_{j}-a_{i} \in \mathcal{A}(a_1)$$ for all $0\leq i<j\leq k$. \end{theorem} {\bf Proof.} If not then by property (3) above there exists a number $j$ for which $a_j-a_{j-1}\not\in\mathcal{A}(a_1)$, which is equivalent to $$pa_1+1\leq a_j-a_{j-1}\leq (p+1)a_1-(p+2)$$ for some $p$. Among all pairs $(p,j)$ for which these inequalities hold let us choose the lexicographically smallest one. Comparing the leftmost and rightmost expressions in this double inequality yields $a_1\geq p+3$, hence $a_1\geq 4$ and $1\leq p \leq a_1-3$. We have $\textrm{opt}_A(a_{j-1}+(p+1)a_1)\leq p+2$ and $$a_j+(p+2)\leq a_{j-1}+(p+1)a_1\leq a_j+a_1-1$$ hence $\textrm{grd}_{(1,\ldots,a_j)}(a_{j-1}+(p+1)a_1)\geq p+3$. It follows that $a_{j+1}\leq a_{j-1}+(p+1)a_1$. Then $$a_{j+1}-a_j\leq a_{j-1}+(p+1)a_1 -a_j \leq a_1-1.$$ By Proposition \ref{lemmadiffa1-1} all these inequalities must in fact be equalities. In other words: \begin{center} \begin{tabular}{c} $a_{j+1}=a_{j-1}+(p+1)a_1,$\\ $a_j=a_{j-1}+pa_1+1.$ \end{tabular} \end{center} Choose the largest $l$ for which $a_{l+1}-a_l\leq a_{j-1}+(p-1)a_1+2$ (such $l$'s exist, for instance $a_{j+1}-a_{j}=a_1-1$ is sufficiently small). By maximality of $l$ we have $a_{l+2}-a_{l+1}\geq a_{j-1}+(p-1)a_1+3$ (it is possible that $a_{l+2}=\infty$). Observe that $$a_{l+2}-a_l = (a_{l+2}-a_{l+1})+(a_{l+1}-a_l)\geq a_{j-1}+(p-1)a_1+3 + a_1-1=a_{j-1}+pa_1+2=a_j+1$$ and $$a_{l+1}-a_l\leq a_{j-1}+(p-1)a_1+3 = a_j-a_1+2$$ which means that $$a_{l+1}+a_1-2\leq a_l+a_j<a_{l+2}.$$ This eventually implies that $a_l+a_j=a_{l+1}+a_r$ for some $1\leq r<j$. The rest of the proof depends on the possible locations of $a_l+a_{j-1}$. If $a_l+a_{j-1}>a_{l+1}$ then the same argument yields an index $s<j-1$ for which $a_l+a_{j-1}=a_{l+1}+a_s$. In that case $a_r-a_s=a_j-a_{j-1}\not\in\mathcal{A}(a_1)$. By properties (3) and (2) of $\mathcal{A}(a_1)$ there exist numbers $s<r'\leq r$ and $p'$ for which $$p'a_1+1\leq a_{r'}-a_{r'-1}\leq (p'+1)a_1-(p'+2).$$ The inequality $a_{r'}-a_{r'-1}\leq a_j-a_{j-1}$ implies $p'\leq p$. The pair $(p',r')$ is lexicographically smaller that $(p,j)$, which is a contradiction since the latter was chosen to be minimal. If, on the other hand, $a_l+a_{j-1}=a_{l+1}$, then $a_l+a_j=a_l+a_{j-1}+pa_1+1=a_{l+1}+pa_1+1$, which means that $a_r=pa_1+1$. Then $a_r-a_1=(p-1)a_1+1\not\in \mathcal{A}(a_1)$, contradicting the minimality of $(p,j)$ be the same argument as above. Therefore we are left with the case $a_l+a_{j-1}<a_{l+1}$. The number $a_r$ satisfies $$a_r=a_l+a_j-a_{l+1}<a_l+a_j-(a_l+a_{j-1})=a_j-a_{j-1}=pa_1+1.$$ Since $a_r-a_1<(p-1)a_1+1$, the minimality of $p$ implies that $a_r-a_1\in\mathcal{A}(a_1)$. It means that $a_r=qa_1-q'$ for some $q\leq p$ and $0\leq q'<q$. Next we are going to show that $a_{l+1}-(a_l+a_{j-1})\not\in\mathcal{A}(a_1)$. Observe that $$a_{l+1}-(a_l+a_{j-1})=(a_l+a_j-a_r)-a_l-a_{j-1}=pa_1+1-a_r=(p-q)a_1+(1+q').$$ which is more than $(p-q)a_1$, while at the same time it equals: $$(p-q+1)a_1-(a_1-1-q')$$ with $a_1-1-q'>(p+2)-1-q'=p+1-q'>p-q+1$. By property (2) the number $a_{l+1}-(a_l+a_{j-1})$ does not belong to $\mathcal{A}(a_1)$. \begin{figure}[!h] \epsfbox{figs.3} \caption{The last case of the proof. The length of the bold interval is not in $\mathcal{A}(a_1)$.} \end{figure} Now let us choose the least $p'$ for which $a_l+a_{j-1}+p'a_1\geq a_{l+1}$. In this case $$a_l+a_{j-1}+p'a_1<a_{l+1}+a_1\leq a_{l+1}+a_r=a_l+a_j.$$ Obviously $\textrm{opt}_A(a_l+a_{j-1}+p'a_1)\leq p'+2$. On the other hand, the greedy decomposition of $a_l+a_{j-1}+p'a_1$ is $a_{l+1}+s\cdot 1$, where $s=a_l+a_{j-1}+p'a_1-a_{l+1}$. By optimality $$s+1\leq p'+2$$ so $s\leq p'+1$. On the other hand, we have already proved that $p'a_1-s=a_{l+1}-(a_l+a_{j-1})\not\in\mathcal{A}(a_1)$, so $s\geq p'+1$. Finally we have $s=p'+1$. To end the proof we compute $a_r-a_1$ in terms of $p,p'$ and $a_1$: \begin{center} \begin{tabular}{c} $a_r-a_1=a_l+a_j-a_{l+1}-a_1=a_l+(a_{j-1}+pa_1+1)-(a_l+a_{j-1} +p'a_1-s)-a_1=$\\ $=(p-p'-1)a_1+(s+1)=(p-p'-1)a_1+(p'+2)=$\\ $=(p-p')a_1-(a_1-p'-2)$ \end{tabular} \end{center} Since $a_r-a_1\in\mathcal{A}(a_1)$, by property (4) we obtain \begin{center} \begin{tabular}{c} $a_1-p'-2\leq p-p',$\\ $a_1<p+3.$ \end{tabular} \end{center} This contradiction ends the proof.\qed As an immediate corollary we obtain the theorem announced in the introduction: \begin{theorem} \label{theoremprefixorderly} If $A=(1,a_1,\ldots,a_k)$ is orderly then for any $2\leq l\leq k$ the currency $(1,a_1,a_l)$ is also orderly. In particular the currency $(1,a_1,a_2)$ is orderly. \end{theorem} {\bf Proof}. If $A$ is orderly then by Theorem \ref{theoremdiffaa} we have $a_l-a_1\in\mathcal{A}(a_1)$. By Proposition \ref{lemma3orderly} this is sufficient for $(1,a_1,a_l)$ to be orderly.\qed \section{Short currencies} \label{section4} Theorems \ref{theoremprefixorderly} and \ref{onept} allow us to give a complete characterization of all orderly currencies with at most 5 coins. The currencies with 1, 2 and 3 coins have already been discussed. Here we concentrate on the cases of 4 and 5 coins. Following \cite{CCS} call a currency $A=(1,a_1,\ldots,a_k)$ \emph{totally orderly}\footnote{Also called normal in \cite{TienHu}.} if every prefix sub-currency of the form $(1,a_1,\ldots,a_l)$ is orderly for $l=0,\ldots,k$. \begin{proposition} \label{lemma4} The currency $A=(1,a_1,a_2,a_3)$ is orderly if and only if it is totally orderly. \end{proposition} \begin{proposition} \label{lemma5} The currency $A=(1,a_1,a_2,a_3,a_4)$ is orderly if and only if \begin{itemize} \item (1) either $(1,a_1,a_2,a_3,a_4)=(1,2,a,a+1,2a)$ for some $a\geq 4$, in which case $(1,a_1,a_2,a_3)$ is not orderly, \item (2) or $A$ is totally orderly. \end{itemize} \end{proposition} {\bf Remark.} The conditions given in the above propositions are efficiently computable, since it can be quickly checked if a currency is totally orderly (as opposed to checking whether it is just orderly). One simply repeats the one-point theorem test with for longer and longer prefixes; see also \cite{CCS}. {\bf Proof of Propositions \ref{lemma4} and \ref{lemma5}.} The one-point theorem, together with Theorem \ref{theoremprefixorderly} covers Proposition \ref{lemma4} and case (2) of Proposition \ref{lemma5}. It remains to show that all orderly currencies $(1,a_1,a_2,a_3,a_4)$ in which the sub-currency $(1,a_1,a_2,a_3)$ is disorderly are of the form (1) from Proposition \ref{lemma5}. Let $m=\lceil a_3/a_2\rceil$. The triple $(1,a_1,a_2)$ is orderly by Theorem \ref{theoremprefixorderly}. By the one-point theorem $ma_2$ is a counterexample for $(1,a_1,a_2,a_3)$, hence $a_4\leq ma_2$. Both values $a_3+a_3$ and $a_3+a_2$ exceed $ma_2$, so they exceed $a_4$, so by optimality there must exist $i<j\leq 2$ for which: \begin{center} \begin{tabular}{c} $a_3+a_2=a_4+a_i,$\\ $a_3+a_3=a_4+a_j.$ \end{tabular} \end{center} Subtracting these equations we get $$a_3-a_2=a_j-a_i<a_j\leq a_2$$ which in turn gives $a_3<2a_2$. That means $m=2$. There are two cases to consider: {\bf $j=2$.} Then $a_3-a_2=a_2-a_i$, so $2a_2=a_3+a_i$, which contradicts the fact that $(1,a_1,a_2,a_3)$ is disorderly. {\bf $j=1$.} Then $i=0$ and previous equations take the form: \begin{center} \begin{tabular}{c} $a_3+a_2=a_4+1,$\\ $a_3+a_3=a_4+a_1.$ \end{tabular} \end{center} The following computation $$a_4+1=a_3+a_2>2a_2=ma_2\geq a_4$$ implies $$a_4+1=a_3+a_2=2a_2+1.$$ Setting $a_2=a$ we get $a_3=a+1$, $a_4=2a$ and $a_1=2a_3-a_4=2$. The routine check that $(1,2,a,a+1,2a)$ is orderly resembles the technique used in the proof of case 2 of Theorem \ref{onept} and is left to the reader. For $a\geq 4$ the sub-currency $(1,2,a,a+1)$ is disorderly.\qed Attempts to continue similar reasoning with longer coinage systems encounter a serious problem, because the applicability of the one-point theorem is limited. More precisely, the ``intermediate'' currencies may not be orderly even if $A$ is orderly as we see from part (1) of Proposition \ref{lemma5}. We shall return to these matters in section \ref{section6}. \section{Investigating differences, part II} \label{section5} In the previous sections we were discussing relation of the distances $a_j-a_i$ and the value of $a_1$. Here we shall extend some of this to further coins. Note that Proposition \ref{lemmadiffa1-1} may be interpreted as follows: if some difference $a_j-a_i$ belongs to the interval $(1,a_1)$, then it must be necessarily equal $a_1-1$. We are interested in the possible values of $a_j-a_i$ in the cases when this difference belongs to $(a_{m-1},a_m)$. {\bf Throughout this section we always assume that $A=(1,a_1,\ldots,a_k)$ is orderly.} The key results of this sections are Corollary \ref{possibletoa2} and Theorem \ref{theorembigdiff}. \begin{lemma} \label{aux1} If $$a_m-a_{l+1}<a_j-a_i<a_m-a_l$$ for some $i<j$, $l<m$, then $$a_{j+1}\leq a_i+a_m.$$ \end{lemma} {\bf Proof.} We have $a_j+a_l<a_i+a_m<a_j+a_{l+1}$. If there was no new coin between $a_j$ and $a_i+a_m$ then there would be no greedy decomposition of $a_i+a_m$ in two steps.\qed \begin{lemma} \label{aux2} There are no numbers $0\leq i<j\leq k$ and $1\leq m\leq k$ that satisfy $a_{m-1}\leq a_j-a_i < a_m-a_{m-1}$. \end{lemma} {\bf Proof.} Suppose the contrary and let $(j,i,m)$ be some triple satisfying the above inequalities, such that $j$ is the least possible. If $i=0$ then $a_{m-1}\leq a_j-1<a_m-a_{m-1}<a_m$, hence $j=m$, which in turn implies $a_{m-1}<1$, but this is not possible. Therefore $i\geq 1$ and we are free to choose the largest index $l$ for which $a_l-a_{l-1}<a_i$. If $a_{l-1}+a_j<a_{l+1}$ then by the standard argument we obtain a contradiction with the minimality of $j$. Hence $a_{l-1}+a_j\geq a_{l+1}$. It follows that $$a_l-a_{l-1}=(a_{l+1}-a_{l-1})-(a_{l+1}-a_l)\leq a_j-a_i<a_m-a_{m-1}.$$ By Lemma \ref{aux1} it follows that $a_{l+1}-a_{l-1}\leq a_m$, so $a_i<a_{l+1}-a_{l-1}\leq a_m$. In effect $i\leq m-1$. At the same time we also have $a_j>a_{m-1}$, so $j\geq m$. All this implies $$a_j-a_i\geq a_m-a_{m-1}.$$ This contradiction ends the proof.\qed \begin{lemma} \label{aux3} Let $m\geq 2$. If the difference $a_j-a_i$ belongs to the interval $[a_m-a_1,a_m-1]$ then it can only be one of the numbers $a_m-a_1$, $a_m-a_1+1$ and $a_m-1$. \end{lemma} {\bf Proof.} Suppose that $$a_m-a_1<a_j-a_i<a_m-1.$$ From Lemma \ref{aux1} we get $a_{j+1}\leq a_i+a_m<a_j+a_1$. In this case Proposition \ref{lemmadiffa1-1} implies $a_{j+1}=a_j+a_1-1$. Moreover, we have $$a_j+2\leq a_i+a_m\leq a_j+a_1-1=a_{j+1}.$$ Hence $a_i+a_m=a_{j+1}$ (otherwise the amount $a_i+a_m$ would not have a greedy decomposition in two steps). Eventually we get $$a_j-a_i=(a_{j+1}-a_1+1)-(a_{j+1}-a_m)=a_m-a_1+1.$$\qed \begin{lemma} \label{aux4} If $a_1<a_2-a_1+1<a_2-1$ then the value $a_2-a_1+1$ cannot be attained by any of the differences $a_j-a_i$. \end{lemma} {\bf Proof.} First note that the given inequalities imply $a_1\geq 3$. Suppose that $j$ is the minimal number for which there exists an $i$ such that $a_j-a_i=a_2-a_1+1$. Clearly $i\geq 2$. From the proof of Lemma \ref{aux3} we know that $$a_{j+1}=a_i+a_2=a_j+a_1-1.$$ Let $m$ be the maximal index for which $a_m-a_{m-1}<a_i$. Then $a_{m+1}-a_m\geq a_i$. If $a_{m-1}+a_j<a_{m+1}$ then considering the amounts $a_{m-1}+a_i$ and $a_{m-1}+a_j$ and their greedy decompositions we obtain a contradiction with the minimality of $j$ in the usual way. Hence we may assume that $$a_{m-1}+a_j\geq a_{m+1}.$$ If $a_{m-1}+a_j=a_{m+1}$ then consider the amount $a_{m-1}+a_{j+1}$. It satisfies $$a_{m-1}+a_{j+1}=a_{m+1}+a_1-1<a_{m+1}+a_2\leq a_{m+1}+a_i\leq a_{m+2}.$$ Since $a_1-1\geq 2$ this amount cannot be greedily decomposed in two steps, so we have a contradiction, which means that $$a_{m-1}+a_j\geq a_{m+1}+1$$ which in turn implies $$a_m-a_{m-1}=(a_{m+1}-a_{m-1})-(a_{m+1}-a_m)\leq a_j-1-a_i=a_2-a_1.$$ We know from the previous lemmas that in this case the only possible values of the difference $a_m-a_{m-1}$ are $a_2-a_1$, $a_1$ and $a_1-1$. Let us investigate these cases separately. {\bf Case 1.} $a_m-a_{m-1}=a_2-a_1$. Then $a_{m+1}\geq a_m+a_i$ and $$a_{m+1}\leq a_{m-1}+a_j-1=a_m-a_2+a_1+a_j-1=a_m+a_i$$ hence $a_{m+1}=a_m+a_i=a_{m-1}+a_j-1$. Now consider the amount $a_m+a_j$. It satisfies $$a_m+a_j=a_m+a_i+a_2-a_1+1=a_{m+1}+a_2-a_1+1<a_{m+1}+a_2\leq a_{m+1}+a_i\leq a_{m+2}$$ so it could be decomposed greedily in two steps only if $a_2-a_1+1$ was one of the coins, which is not true by the assumptions of the lemma. {\bf Case 2.} $a_m-a_{m-1}=a_1$. Now consider the amount $a_{m-1}+a_2$: $$a_m<a_{m-1}+a_2=a_m+(a_2-a_1)<a_m+a_i\leq a_{m+1}.$$ This amount can only be decomposed optimally if $a_2-a_1$ is a coin. Since $a_1\geq 3$, by Proposition \ref{lemmanot1} we have $a_2-a_1\neq 1$. Therefore $a_2-a_1=a_1$ and we have $$a_m-a_{m-1}=a_1=a_2-a_1$$ and the argument from case 1 can be repeated. {\bf Case 3.} $a_m-a_{m-1}=a_1-1$. An exact repetition of case 2 shows that in this case $a_2-a_1+1$ would have to be one of the coins. However, this possibility is excluded by the assumptions of our lemma. \qed The results from this section, together with Proposition \ref{lemmadiffa1-1} can be used to characterize the set of possible values of $a_j-a_i$ which fit in the interval $(1,a_2)$. For a currency $A$ let $S(A)=\{a_j-a_i: 0\leq i<j\leq k\}$. \begin{corollary} \label{possibletoa2} For an orderly currency $A=(1,a_1,a_2,\ldots,a_k)$ \begin{itemize} \item[(a)] we always have $$S(A)\cap (1,a_1)\subset\{a_1-1\}$$ $$S(A)\cap (a_1,a_2)\subset\{a_2-a_1,a_2-1\}$$ \item[(b)] if $a_2=2a_1-1$ or $a_2=2a_1$ then $S(A)\cap (1,a_2)\subset\{a_1-1,a_1,a_2-1\}$ \item[(c)] if $a_2>2a_1$ then $S(A)\cap (1,a_2)=\{a_1-1,a_2-a_1,a_2-1\}$ \end{itemize} \end{corollary} {\bf Proof.} Property (a) is just a restatement of Proposition \ref{lemmadiffa1-1} and Lemmas \ref{aux2}, \ref{aux3} and \ref{aux4}. By Theorem \ref{theoremdiffaa} there are no other possible values of $a_2$ except of those in (b) and (c). In both cases, if $a_j-a_i < a_1$, then Proposition \ref{lemmadiffa1-1} applies. In case (c) $a_1<a_2-a_1<a_2-a_1+1\leq a_2-1$ and an application of Lemmas \ref{aux3} and \ref{aux4} proves that our theorem enumerates all possible elements of $S(A)\cap (1,a_2)$. Of course all the given values are attained, so in (c) we are free to use equality rather than inclusion. In case (b) the difference $a_1$ may or may not be attained (consult the currencies $(1,3,5)$ and $(1,3,5,8,10,15)$). Once again one needs to combine the before-mentioned lemmas; we omit the details.\qed {\bf Remark.} Corollary \ref{possibletoa2} and Theorem \ref{theoremdiffaa} give two independent conditions that must be satisfied by orderly currencies. For instance every three-coin currency satisfies Corollary \ref{possibletoa2}, but not necessarily Theorem \ref{theoremdiffaa} (it is also easy to imagine more complicated examples of this kind). On the other hand, the currency $(1,3,7,12)$ satisfies Theorem \ref{theoremdiffaa}, but $12-7=5\not\in\{7-3,7-1\}$, so part (a) of Corollary \ref{possibletoa2} is violated. Our last theorem in this section will be important in section \ref{section7}. It can roughly be stated as ``if some two consecutive differences are large, then the subsequent differences must also be large''. \begin{theorem} \label{theorembigdiff} Suppose $(1,a_1,\ldots,a_k)$ is orderly, $m\geq 2$ and $$a_{m-1}>2a_{m-2}, \ a_{m}>2a_{m-1}.$$ Then for every $t\geq m$ we have $a_{t+1}-a_t\geq a_m-a_{m-1}$. \end{theorem} {\bf Proof.} Suppose, on the contrary, that $a_{t+1}-a_t< a_m-a_{m-1}$ for some $t\geq m$, and let $t$ be the smallest index with these properties. Choose $s$ as the largest index for which $$a_{s+1}-a_s<a_t-a_{m-2}$$ (such numbers $s$ exist; for instance $s=t-1$ satisfies this inequality). Note that by maximality of $s$ we have $a_{s+2}-a_{s+1}\geq a_t-a_{m-2}$ (possibly $a_{s+2}=\infty$) and $a_{s+3}-a_{s+2}\geq a_t-a_{m-2}$ (if $a_{s+2}<\infty$). The proof is split into two cases. {\bf Case 1.} $a_s+a_{t+1}<a_{s+2}$. With this assumption we have $$a_{s+1}<a_s+a_t<a_s+a_{t+1}<a_{s+2}$$ so there exist indices $r,l$ such that \begin{center} \begin{tabular}{c} $a_s+a_t=a_{s+1}+a_r,$\\ $a_s+a_{t+1}=a_{s+1}+a_l,$ \end{tabular} \end{center} with $r<l\leq t$. This implies $$a_l-a_{l-1}\leq a_l-a_r=a_{t+1}-a_t<a_m-a_{m-1}.$$ Since $l-1< t$ and $t$ was chosen to be minimal with respect to the condition $t\geq m$ and the above inequality, we obtain $l-1<m$. Since $l=m$ does not satisfy the above inequality, we have $l\leq m-1$ and $r\leq m-2$, but then $$a_{s+1}-a_s=a_t-a_r\geq a_t-a_{m-2}$$ contradicting the choice of $s$. This completes the first case of the proof. {\bf Case 2.} Now suppose $a_s+a_{t+1}\geq a_{s+2}$. We are going to prove the following sequence of inequalities: \begin{center} \begin{tabular}{lc} (1) & $a_{s+1}-a_s>a_{m-2}$\\ (2) & $a_{s+2}-a_s>a_m$\\ (3) & $a_{s+1}-a_s<a_m$\\ (4) & $a_{s+1}-a_s\geq a_m-a_{m-1}$\\ (5) & $a_{s+2}<a_{s+1}+a_t<a_{s+1}+a_{t+1}<a_{s+3}$ \end{tabular} \end{center} (1): We always have $$a_{s+2}-a_s>a_{s+2}-a_{s+1}\geq a_t-a_{m-2}\geq a_m-a_{m-2}>a_{m-1}.$$ If we also had $a_{s+1}-a_s\leq a_{m-2}$ then $$a_{s+1}\leq a_s+a_{m-2}<a_s+a_{m-1}<a_{s+2}.$$ As usually, it means that $a_{m-1}-a_{m-2}=a_l-a_r$ for some $r<l\leq m-2$ or $a_{m-1}-a_{m-2}=a_l$ for $l\leq m-2$. In either case $a_{m-1}-a_{m-2}\leq a_{m-2}$, contradicting the assumptions of the theorem. Therefore $a_{s+1}-a_s>a_{m-2}$. (2): This follows straight from (1) and the maximality of $s$: $$a_{s+2}-a_s=(a_{s+2}-a_{s+1})+(a_{s+1}-a_s)>a_t-a_{m-2}+a_{m-2}=a_t\geq a_m.$$ (3): Since we assumed $a_{s+2}-a_s\leq a_{t+1}$ for this case, we obtain, using the properties of $s$ and $t$, that $$a_{s+1}-a_s=(a_{s+2}-a_s)-(a_{s+2}-a_{s+1})\leq a_{t+1}-(a_t-a_{m-2})<a_m-a_{m-1}+a_{m-2}<a_m.$$ (4): By (2) and (3) we have $a_{s+1}<a_s+a_m<a_{s+2}$, therefore $a_s+a_m=a_{s+1}+a_r$ for some $r\leq m-1$. Finally $$a_{s+1}-a_s=a_m-a_r\geq a_m-a_{m-1}.$$ (5): First note that by $a_t\geq a_m$ and $2a_{m-2}<a_{m-1}$ we obtain $$a_{s+3}-a_{s+1}\geq 2(a_t-a_{m-2})=a_t+a_t-2a_{m-2}> a_t+a_m-a_{m-1}>a_{t+1}.$$ Moreover, by (4) and the assumption $a_{s+2}-a_s\leq a_{t+1}$ we get $$a_{s+2}-a_{s+1}=(a_{s+2}-a_s)-(a_{s+1}-a_s)\leq a_{t+1}-(a_m-a_{m-1})<a_t.$$ This ends the proof of (1)--(5). \begin{figure} \epsfbox{figs.4} \caption{The situation in case 2. in Theorem \ref{theorembigdiff}.} \end{figure} Now (5) implies the existence of $r<l\leq t$ such that \begin{center} \begin{tabular}{c} $a_{s+1}+a_t=a_{s+2}+a_r,$\\ $a_{s+1}+a_{t+1}=a_{s+2}+a_l.$ \end{tabular} \end{center} As a consequence of these formulae we obtain the inequality $$a_r=a_t-(a_{s+2}-a_{s+1})\leq a_t-(a_t-a_{m-2})=a_{m-2}, \ \textrm{hence } r\leq m-2,$$ which in turn implies $$a_l=(a_{t+1}-a_t)+a_r<a_m-a_{m-1}+a_{m-2}<a_m, \ \textrm{hence } l\leq m-1.$$ Combining this, we get $$a_{s+1}-a_s=a_{s+2}-a_s+a_l-a_{t+1}\leq a_{t+1}+a_l-a_{t+1}=a_l\leq a_{m-1}.$$ However, by (4) $a_{s+1}-a_s\geq a_m-a_{m-1}>a_{m-1}$, so we have a contradiction which ends the proof of case 2, and the whole theorem.\qed \section{$+/-$-classes} \label{section6} If $A=(1,a_1,\ldots,a_k)$ is orderly then some prefix sub-currency, i.e. a currency of the form $A'=(1,a_1,\ldots,a_l)$ with $l<k$ might not be orderly (for instance, $(1,2,a,a+1,2a)$ is orderly, but $(1,2,a,a+1)$ is not for $a\geq 4$, as in Proposition \ref{lemma5}). This situation was still quite manageable in the case of 5 coins, but it gets more and more complicated as the number of coin increases, thus making inductive analysis (possibly using the one-point theorem) impossible. To describe the prefix currencies we introduce the notion of {\it $+/-$-classes}. To every currency $A=(1,a_1,\ldots,a_k)$ we may assign a pattern of $k+1$ signs {\tt +} and {\tt -}, defined as follows: the $l$-th symbol of the pattern ($l=0,\ldots,k$) is {\tt +} if the prefix currency $(1,a_1,\ldots,a_l)$ is orderly and {\tt -} in the opposite case. A {\it $+/-$-class} is the set of all currencies corresponding to a given $+/-$-pattern. For instance, the pattern \verb?++++?$\ldots$\verb?+++? corresponds to totally orderly currencies. Another well-described example is the $+/-$-class given by the pattern \verb?+++-+? --- it consists precisely of the currencies $(1,2,a,a+1,2a)$ with $a\geq 4$ (this is the consequence of Proposition \ref{lemma5}, since an orderly $5$-coin currency which is not totally orderly satisfies part (1) of that proposition). The $+/-$-patterns that correspond to non-empty classes cannot be completely arbitrary, for instance, if a pattern ends with a {\tt +} then it must begin with {\tt +++} -- this is a consequence of Theorem \ref{theoremprefixorderly}. The patterns beginning with {\tt +++} and ending with {\tt +} will be called {\it proper}. Mysteriously, some proper patterns describe empty classes. Here is a sample proposition of this sort: \begin{proposition} \label{emptyclass} The $+/-$-class described by the pattern \verb?+++-+-+? is empty. \end{proposition} {\bf Proof.} Suppose that $A=(1,a_1,a_2,a_3,a_4,a_5,a_6)$ is a coinage system in the class \verb?+++-+-+?. By case (1) of Proposition \ref{lemma5} we know that in fact $A$ is of the form $$(1,2,a,a+1,2a,a_5,a_6)$$ for some $4\leq a$, $2a< a_5< a_6$. By the one-point theorem some multiple of $2a$ is a counterexample for $(1,2,a,a+1,2a,a_5)$. Extending this by $a_6$ must fix this problem, hence $$a_6-a_5<2a.$$ Since $A$ is orderly, there exist numbers $r, s$ such that: \begin{center} \begin{tabular}{c} $a_5+a_5=a_6+a_r,$\\ $a_5+2a=a_6+a_s,$ \end{tabular} \end{center} with $a_r\leq 2a$, $a_s\leq a+1$, $1 \leq a_s<a_r$. Subtracting the two equations yields $a_5-2a=a_r-a_s$. Possible differences $a_r-a_s$ ($0\leq s<r\leq 4$) form the set $$\{1,a-2,a-1,a,2a-2,2a-1,2a\}$$ so the possible values of $a_5$ are $2a+1,3a-2,3a-1,3a,4a-2,4a-1,4a$. The values $3a-1,3a,4a-2,4a-1,4a$ can be excluded from this set, since then $(1,2,a,a+1,2a,a_5)$ would be orderly, which can be checked easily by the one-point theorem (the ``suspected'' amount to be tested for optimality is $4a$). Therefore we are left with $a_5\in\{2a+1,3a-2\}$. If $a_5=2a+1$ then the greedy algorithm for $(1,\ldots,a_5)$ fails to be optimal already for $3a=2a+a$, hence $a_6\leq 3a$. On the other hand, all three numbers $2a+2a$, $2a+(2a+1)$ and $(2a+1)+(2a+1)$ can be obtained with two coins, hence $4a-a_6$, $4a+1-a_6$ and $4a+2-a_6$ must be three consecutive integers which are coins, all less than $a_6$. This is only possible if $a_6=4a$, contradiction. Now suppose that $a_5=3a-2$. Then for the number $2a+(a+1)=3a+1$ not to be a counterexample we must have $3a-1\leq a_6\leq 3a+1$. If $a_6=3a-1$ then $4a-2=(3a-2)+a=(3a-1)+(a-1)$ is a counterexample ($a-1$ is not a coin). If $a_6=3a$ then the counterexample is $4a-1=(3a-2)+(a+1)=3a+(a-1)$ (reason as before). Finally, if $a_6=3a+1$ then $4a=2a+2a=(3a+1)+(a-1)$ is the counterexample.\qed Of course, given a currency, we may recover its $+/-$-class in $O(k^4)$ time simply by repeating Pearson's algorithm \cite{Pear} for each prefix sub-currency. The reverse problem, to determine whether a given proper $+/-$-pattern describes a non-empty $+/-$-class is actually much harder and we have not been able to find any algorithm solving it. From this point of view the most ``messy'' orderly currencies are those which belong to the class determined by \verb?+++----?$\ldots$\verb?--+?. These classes are indeed non-empty for $k=0,2\pmod{3}$. Their representatives for $k=3l$ and $k=3l-1$, respectively, are \begin{center} \begin{tabular}{c} $(1,2,\ 4,5,\ 7,8,\ldots,3l-2,3l-1,\ 3l+1,\ 3l+4,\ldots,6l-2),$\\ $(1,2,\ 4,5,\ 7,8,\ldots,3l-2,3l-1,\ 3l+2,\ 3l+5,\ldots,6l-4).$ \end{tabular} \end{center} On the other hand, there seem to be no coinage systems of type \verb?+++----?$\ldots$\verb?--+? for $k=1\pmod{3}$, but we have not been able to prove this. \section{Classification of orderly sub-currencies} \label{section7} Every set $P=\{i_0,i_1,\ldots,i_l\}\subset\{0,1,\ldots,k\}$, where $0=i_0<i_1<\ldots<i_l$ determines a sub-currency $(a_{i_0},a_{i_1},\ldots,a_{i_l})$ of any currency $A=(1,a_1,\ldots,a_k)$. From Theorem \ref{theoremprefixorderly} we know that if $A$ is orderly then the sub-currency determined by $P=\{0,1,l\}$ ($2\leq l\leq k$) is also orderly. Is this just a lonely phenomenon, or could a similar theorem be proved for some other sets $P$? \begin{definition} \label{hered} The set $P$ of the form given above will be called \emph{hereditary} if the following is true: \begin{center} for every orderly currency $A=(1,a_1,\ldots,a_k)$\\ the sub-currency determined by $P$ is also orderly \end{center} \end{definition} Let us enumerate some interesting classes of subsets of $\{0,1,\ldots,k\}$: \begin{itemize} \item[type 1:] the singleton set $\{0\}$ \item[type 2:] the sets $\{0,l\}$ for $1\leq l\leq k$ \item[type 3:] the sets $\{0,1,l\}$ for $2\leq l\leq k$ \item[type 4:] the sets $\{0,1,2,l\}$ for $4\leq l\leq k$ \item[type 5:] the full set $\{0,1,\ldots,k\}$ \end{itemize} Note that $\{0,1,2,3\}$ is a peculiar exception: it is \emph{not} of type 4 (an immediate example is $(1,2,a,a+1,2a)$ for $a\geq 4$ and its non-orderly sub-currency $(1,2,a,a+1)$ determined by $\{0,1,2,3\}$). We already know that sets $P$ of type 1, 2, 3 or 5 are hereditary. In this section we shall prove that sets that are not specified in types 1--5 are not hereditary.\footnote{To be precise, every set $P$ should always be thought of as a subset of $\{0,1,\ldots,k\}$ for a certain $k$. In most cases $k$ will be implicit, but to improve clarity we shall sometimes stress this connection by writing $P\subset\{0,1,\ldots,k\}$.} We also conjecture that all sets $P$ of type 4 are hereditary, and we prove this conjecture under some mild additional assumptions. The general case remains open. Before proceeding with the elimination of non-hereditary subsets $P$ let us make a few observations. \begin{lemma} For any $l\geq 3$ let $B_l$ denote the currency $$B_l=(1,2,3,\ldots,l-1,2l-2,2l-1,4l-4)$$ where $a_l=2l-1$. Then $B_l$ is orderly of type \verb?+++?$\ldots$\verb?+-+?. \end{lemma} {\bf Proof.} The prefix currency $(1,2,3\ldots,l-1)$ is clearly of type \verb?+++?$\ldots$\verb?++?. Extending this by $2(l-1)$ we get an orderly currency by the one-point theorem. The next prefix, ending in $2l-1$ is not orderly since $2\cdot 2(l-1)=4l-4$ is the smallest counterexample. The complete currency is orderly which can be proved easily by the techniques from the proof of Theorem \ref{onept}.\qed \begin{lemma} For any $m>l\geq 2$ and $p\geq 1$ let $A_{l,m}(p)$ denote the currency: $$A_{l,m}(p) = (a_0,a_1,a_2,\ldots,a_{l-1},a_l,a_{l+1},a_{l+2},\ldots,a_{m})=$$ $$(1,2,3,\ldots,l,pl,(2p-1)l,(3p-2)l,\ldots,((m-l+1)p-(m-l))l)$$ where $a_l=pl$. This currency is orderly. Moreover, if $p>m-l$ then $\lceil a_m/a_l \rceil = m-l+1$. \end{lemma} {\bf Proof.} The given currency is in fact of type \verb?++++?$\ldots$\verb?+++?, which can be verified inductively by the one-point theorem: to check that $(1,a_1,\ldots,a_{l+i})$ is orderly for $i \geq 1$ it suffices to observe that $$2a_{l+i-1} = 2(pi-(i-1))l= (p(i+1)-i)l + (p(i-1)-(i-2))l=a_{l+i}+a_{l+i-2}.$$ To prove the last statement note that $$a_{m}=((m-l+1)p-(m-l))l<(m-l+1)pl=(m-l+1)a_l$$ and, if $p>m-l$: $$a_{m}=((m-l+1)p-(m-l))l = (m-l)pl + l(p-(m-l)) > (m-l)a_l.$$\qed \begin{lemma} \label{obs3} An orderly currency may be extended by any multiple of its highest coin and the resulting currency will be orderly. \end{lemma} {\bf Proof.} A trivial consequence of the one-point theorem.\qed The last observation will be used in the following way: suppose we want to prove that some set $P=\{i_0,i_1,\ldots,i_l\}\subset\{0,1,\ldots,k\}$ is not hereditary. First we find a shorter orderly currency $A'=(1,a_1,\ldots,a_r)$, such that the sub-currency determined by $P'=\{i_0,i_1,\ldots,i_{r'}\}\subset\{0,1,\ldots,r\}$ is not orderly (here $r'< r\leq k$) and $i_{r'+1}>r$ or $r'=l$. Let $c$ be any counterexample for this sub-currency and let $m$ be any number for which $ma_{r}>c$. Then the currency $$A=(1,a_1,\ldots,a_r,ma_r,2ma_r,\ldots,(k-r)ma_r)$$ is orderly (Lemma \ref{obs3}) and its sub-currency determined by $P$ is not, since all the added coins are too large to fix the problem with $c$ (the exact form of $P\setminus P'$ is actually immaterial, it is important that its smallest element is at least $r+1$). \begin{theorem} The sets $P$ not of the form 1, 2, 3, 4 or 5 are not hereditary. \end{theorem} {\bf Proof.} Let $P=\{i_0,\ldots,i_s\}\subset\{0,\ldots,k\}$, $i_0=0$, be such a set. Let $r$ be the largest index for which $i_r=r$ (i.e. $\{0,\ldots,r\}\subset P$, $r+1\not\in P$). We shall consider a few cases: {\bf Case $3\leq r<k$}. Here we employ the orderly currency $B_r$. Its sub-currency $(1,a_1,\ldots,a_r)$ is not orderly. If $r=k-1$ then we are done, while for $r<k-1$ we must expand $B_r$ to an orderly currency with $k+1$ coins in the standard way described earlier. The resulting currency will have a disorderly sub-currency determined by $P$.\ {\bf Case $r=2$}. In this case $|P|\geq 5$, since otherwise $P$ would be of the form $\{0,1,2\}$ or $\{0,1,2,l\}$ for some $l\geq 4$ and these sets are of type 3 and 4, respectively. Denote $l=i_3\geq 4$, $m=i_4$ and consider the currency $A_{l,m}(p)$ with $p>m-l$. Its sub-currency $$(1,2,3,a_l,a_m)$$ is not orderly since the amount $$\lceil a_m/a_l\rceil a_l = (m-l+1)a_l$$ paid greedily splits into the coin $a_m$ and some of the coins $1,2,3$, thus requiring at least $$1+\frac{(m-l)l}{3} > 1+(m-l)$$ coins, which is more than if it was paid with $m-l+1$ copies of $a_l$. Now it suffices to expand this currency to a currency with $k+1$ coins as previously. {\bf Case $r=1$}. Then $|P|\geq 4$, since otherwise $P$ would be of the form $\{0,1,l\}$, which is of type 3. Let $l=i_2\geq 3$ and $m=i_3$ and consider the currency $A_{l,m}(p)$ with $p>m-l$. The sub-currency $(1,2,a_l,a_m)$ is not orderly for the same reason as previously: the amount $\lceil a_m/a_l\rceil a_l = (m-l+1)a_l$ must be paid greedily with at least $1+\frac{(m-l)l}{2} > 1+(m-l)$ coins and the proof follows. {\bf Case $r=0$}. Clearly $|P|\geq 3$, since sets of the form $\{0\}$ and $\{0,l\}$ are of type 1 and 2. Let $l=i_1\geq 2$ and $m=i_2$. Repeat the same arguments with the currency $A_{l,m}(p)$ ($p>m-l$) and its sub-currency $(1,a_l,a_m)$: this time the amount $\lceil a_m/a_l\rceil a_l = (m-l+1)a_l$ must be paid greedily with at least $1+\frac{(m-l)l}{1} > 1+(m-l)$ coins.\qed Sets $P$ of type 4 are the most peculiar ones. We believe they are also hereditary; that is, we have the following: \begin{conjecture} \label{contype4} If $A=(1,a_1,\ldots,a_k)$ is orderly, then the currency $(1,a_1,a_2,a_l)$ is also orderly for every $4\leq l\leq k$. \end{conjecture} While this is not known to be true in general, we can prove this conjecture under some mild additional conditions. \begin{theorem} \label{theorempartialconjecture} Conjecture \ref{contype4} is true if we additionally assume that $a_2>2a_1$ and $a_3>2a_2$. \end{theorem} {\bf Proof.} We shall verify that $(1,a_1,a_2,a_l)$ is orderly by Proposition \ref{lemma4}. Let $m=\lceil a_l/a_2\rceil$. By Theorem \ref{theorembigdiff} for every $l\geq 3$ we have the first of the following inequalities: $$a_{l+1}-a_l\geq a_3-a_2>a_2>ma_2-a_l.$$ It means that $a_{l+1}>ma_2$, so there is no new coin between $a_l$ and $ma_2$, and the greedy decomposition of $ma_2$ with respect to $A$ involves only the coins $1,a_1,a_l$. This justifies the first equality in the following comparison: $$\textrm{grd}_{(1,a_1,a_2,a_l)}(ma_2)=\textrm{grd}_{A}(ma_2)=\textrm{opt}_{A}(ma_2)\leq \textrm{opt}_{(1,a_1,a_2,a_l)}(ma_2)$$ and by Proposition \ref{lemma4} the proof is complete.\qed \section{Closing remarks and open problems} \label{section8} Throughout this paper we have proposed some possible approaches to the problem of describing orderly coinage systems and their interesting properties. Some of these techniques have enabled us to prove the most important results of this paper, namely the structural theorems, like Theorem \ref{theoremdiffaa} and Corollary \ref{possibletoa2}, or to give concise descriptions of small systems. There is still quite a lot of work to be done in the following areas: \begin{itemize} \item {\bf sub-currencies}: prove Conjecture \ref{contype4}, thus completing the classification of orderly sub-currencies. \item {\bf prefix sub-currencies}: invent an algorithm to decide whether a given $+/-$--pattern describes a non-empty class or devise some other properties of such $+/-$--patterns. Another interesting conjecture, to which we have not found a counterexample, is: \begin{conjecture} If a $+/-$--class is non-empty, then it has a representative $A=(1,a_1,\ldots,a_k)$ with $a_1=2$. \end{conjecture} \item {\bf differences}: can Corollary \ref{possibletoa2} be generalized? In other words, what can be said about the differences $a_j-a_i$ that belong to $(a_{m-1},a_m)$ for some $m$? Is it true that in general $$S(A)\cap (a_{m-1},a_m)\subset \{a_m-a_{m-1},a_m-a_{m-2},\ldots,a_m-1\},$$ where $S(A)=\{a_j-a_i: 0\leq i<j\leq k\}$ for an orderly currency $A$? We already know this is true for $m=1,2$. The lemmas from section \ref{section6} provide some partial results in the general case as well. \item {\bf extending}: Theorem \ref{theoremdiffaa}, Corollary \ref{possibletoa2} and Conjecture \ref{contype4} can be thought of as {\it obstructions} against extending: if a currency does not satisfy one of these conditions then it cannot be extended to an orderly currency by appending new coins of high denominations (higher than all the existing coins). What are the other invariants of this sort? Is there an algorithm that decides if a currency can be extended to an orderly one? Problems related to obstructions and extending can also be found in \cite{CCS}. \end{itemize} {\bf Acknowledgements.} We are indebted to the referee, whose valuable suggestions improved both the presentation and some technical aspects of our paper. We also thank Lenore Cowen for pointing us to \cite{CCS}.

https://arxiv.org/abs/1311.2657

Random perturbation of low rank matrices: Improving classical bounds

Matrix perturbation inequalities, such as Weyl's theorem (concerning the singular values) and the Davis-Kahan theorem (concerning the singular vectors), play essential roles in quantitative science; in particular, these bounds have found application in data analysis as well as related areas of engineering and computer science. In many situations, the perturbation is assumed to be random, and the original matrix has certain structural properties (such as having low rank). We show that, in this scenario, classical perturbation results, such as Weyl and Davis-Kahan, can be improved significantly. We believe many of our new bounds are close to optimal and also discuss some applications.

\section{Introduction} The singular value decomposition of a real $m \times n$ matrix $A$ is a factorization of the form $A = U \Sigma V^\mathrm{T}$, where $U$ is a $m \times m$ orthogonal matrix, $\Sigma$ is a $m \times n$ rectangular diagonal matrix with non-negative real numbers on the diagonal, and $V^\mathrm{T}$ is an $n \times n$ orthogonal matrix. The diagonal entries of $\Sigma$ are known as the \emph{singular values} of $A$. The $m$ columns of $U$ are the \emph{left-singular vectors} of $A$, while the $n$ columns of $V$ are the \emph{right-singular vectors} of $A$. If $A$ is symmetric, the singular values are given by the absolute value of the eigenvalues, and the singular vectors are just the eigenvectors of $A$. Here, and in the sequel, whenever we write \emph{singular vectors}, the reader is free to interpret this as left-singular vectors or right-singular vectors provided the same choice is made throughout the paper. Consider a real (deterministic) $m \times n$ matrix $A$ with singular values $$\sigma_1 \geq \sigma_2 \geq \cdots \geq \sigma_{\min\{m,n\}} \geq 0$$ and corresponding singular vectors $v_1, v_2, \ldots, v_{\min\{m,n\}}.$ We will call $A$ the data matrix. In general, the vector $v_i$ is not unique. However, if $\sigma_i$ has multiplicity one, then $v_i$ is determined up to sign. An important problem in statistics and numerical analysis is to compute the first $k$ singular values and vectors of $A$. In particular, the largest few singular values and corresponding singular vectors are typically the most important. Among others, this problem lies at the heart of Principal Component Analysis (PCA), which has a very wide range of applications (for many examples, see \cite{KVbook, LR} and the references therein) and in the closely related low rank approximation procedure often used in theoretical computer science and combinatorics. In application, $m,n$ are typically large and $k$ is small, often a fixed constant. A problem of fundamental importance in quantitative science (including pure and applied mathematics, statistics, engineering, and computer science) is to estimate how a small perturbation to the data effects the spectrum. This problem has been discussed in virtually every text book on quantitative linear algebra and numerical analysis (see, for instance, \cite{BT, Hig1, Hig2, SS}). A basic model is as follows. Instead of $A$, one needs to work with $A+E$, where $E$ represents the perturbation matrix. Let $$ \sigma_1' \geq \cdots \geq \sigma_{\min\{m,n\}}' \geq 0 $$ denote the singular values of $A+E$ with corresponding singular vectors $v_1', \ldots, v_{\min\{m,n\}}'$. We consider the following natural questions. \begin{question} Is $v_i'$ a good approximation of $v_i$? \end{question} \begin{question} \label{quest:weyl} Is $\sigma_i'$ a good approximation of $\sigma_i$? \end{question} These two questions are addressed by the Davis-Kahan-Wedin sine theorem and Weyl's inequality. Let us begin with the first question in the case when $i=1$. A canonical way (coming from the numerical analysis literature; see for instance \cite{GVL}) to measure the distance between two unit vectors $v$ and $v'$ is to look at $ \sin \angle(v,v')$, where $\angle(v,v')$ is the angle between $v$ and $v'$ taken in $[0,\pi/2]$. It has been observed by numerical analysts (in the setting where $E$ is deterministic) for quite some time that the key parameter to consider in the bound is the gap (or separation) \begin{equation} \label{def:delta} \delta := \sigma_1 - \sigma_2, \end{equation} between the first and second singular values of $A$. The first result in this direction is the famous Davis-Kahan sine $\theta$ theorem \cite{DK} for Hermitian matrices. The non-Hermitian version was proved later by Wedin \cite{W}. Throughout the paper, we use $\|M\|$ to denote the spectral norm of a matrix $M$. That is, $\|M\|$ is the largest singular value of $M$. \begin{theorem}[Davis-Kahan, Wedin; sine theorem] \label{wedin} $$ \sin \angle(v_1, v_1') \leq 2 \frac{\|E\|}{\delta}. $$ \end{theorem} \begin{remark} Theorem \ref{wedin} is trivially true when $\delta \leq 2 \|E\|$ since sine is always bounded above by one. In other words, even if the vector $v_1'$ is not uniquely determined, the bound is still true for any choice of $v_1'$. On the other hand, when $\delta > 2 \|E\|$, the proof of Theorem \ref{wedin} reveals that the vector $v_1'$ is uniquely determined up to sign. \end{remark} Theorem \ref{wedin} is a simple corollary of \cite[Theorem V.4.4]{SS} which is originally due to Wedin \cite{W}; we present a proof below for completeness. More generally, one can consider approximating the $i$-th singular vector $v_i$ or the space spanned by the first $i$ singular vectors $\mathrm{Span}\{v_1, \ldots, v_i \}$. Naturally, in these cases, one must consider the gaps $$ \delta_i := \sigma_i - \sigma_{i+1}. $$ Question \ref{quest:weyl} is addressed by Weyl's inequality. In particular, Weyl's perturbation theorem \cite{Wy} gives the following deterministic bound for the singular values (see \cite[Theorem IV.4.11]{SS} for a more general perturbation bound due to Mirsky \cite{M}). \begin{theorem} [Weyl's bound] \label{theorem:Weyl} \begin{equation*} \max_{1 \leq i\leq \min\{m,n\}} | \sigma_i -\sigma_i'| \le \|E \|. \end{equation*} \end{theorem} For more discussions concerning general perturbation bounds, we refer the reader to \cite{B, SS} and references therein. We now pause for a moment to prove Theorem \ref{wedin}. \begin{proof}[Proof of Theorem \ref{wedin}] If $\delta \leq 2 \|E\|$, the theorem is trivially true since sine is always bounded above by one. Thus, assume $\delta > 2 \|E\|$. By Theorem \ref{theorem:Weyl}, we have $$ \sigma_1' - \sigma_2' \geq \delta - 2 \|E\| > 0, $$ and hence the singular vectors $v_1$ and $v_1'$ are uniquely determined up to sign. By another application of Theorem \ref{theorem:Weyl}, we obtain $$ \delta = \sigma_1 - \sigma_2 \leq \sigma_1 - \sigma_2' + \|E\|. $$ Rearranging the inequalities, we have $$ \sigma_1 - \sigma_2' \geq \delta - \|E\| \geq \frac{1}{2} \delta > 0. $$ Therefore, by \cite[Theorem V.4.4]{SS}, we conclude that $$ \sin \angle(v_1, v_1') \leq \frac{\|E\|}{ \sigma_1 - \sigma_2'} \leq 2 \frac{\|E\|}{ \delta}, $$ and the proof is complete. \end{proof} Let us now focus on the matrices $A$ and $E$. It has become common practice to assume that the perturbation matrix $E$ is random. Furthermore, researchers have observed that data matrices are usually not arbitrary. They often possess certain structural properties. Among these properties, one of the most frequently seen is having low rank (see, for instance, \cite{CP, CR, CRT, CS, TK} and references therein). The goal in this paper is to show that in this situation, one can significantly improve classical results like Theorems \ref{wedin} and \ref{theorem:Weyl}. To give a quick example, let us assume that $A$ and $E$ are $n \times n$ matrices and that the entries of $E$ are independent and identically distributed (iid) random variables with zero mean, unit variance (which is just matter of normalization), and bounded fourth moment. It is well known that in this case $\|E\|= (2+o(1)) \sqrt n $ with high probability\footnote{We use asymptotic notation under the assumption that $n \to \infty$. Here we use $o(1)$ to denote a term which tends to zero as $n$ tends to infinity.} \cite[Chapter 5]{BS}. Thus, the above two theorems imply \begin{corollary} \label{wedin-cor} For any $\eta > 0$, with probability $1-o(1)$, \begin{equation*} |\sigma_1 -\sigma _1'| \le (2 + \eta) \sqrt n, \end{equation*} and \begin{equation} \label{bound0} \sin \angle(v_1, v_1') \leq 2 (2+\eta) \frac{\sqrt n }{\delta}. \end{equation} \end{corollary} Among others, this shows that if one wants accuracy $\varepsilon$ in the first singular vector computation, $A$ needs to satisfy \begin{equation} \label{bound1} \delta \ge 2 (2 + \eta) \varepsilon^{-1} \sqrt n. \end{equation} We present the results of a numerical simulation for $A$ being a $n \times n$ matrix of rank 2 when $n=400$, $\delta=8$, and where $E$ is a random Bernoulli matrix (its entries are iid random variables that take values $\pm 1$ with probability $1/2$). The results, shown in Figure \ref{young}, turn out to be very different from what \eqref{bound1} predicts. It is easy to see that for the parameters $n=400$ and $\delta =8$, Corollary \ref{wedin-cor} does not give a useful bound (since $\frac{\sqrt{n}}{\delta} = 2.5 >1$). However, Figure \ref{young} shows that, with high probability, $\sin \angle(v_1,v_1') \leq 0.2$, which means $v_1'$ approximates $v_1$ with a relatively small error. \begin{figure}[!Ht] \begin{center} \includegraphics[width=12cm]{noise_n=400r=2.pdf} \includegraphics[width=12cm]{noise_n=1000r=2.pdf} \caption{The cumulative distribution functions of $\sin \angle(v_1, v_1')$ where $A$ is a $n \times n$ deterministic matrix with rank $2$ ($n=400$ for the figure on top and $n=1000$ for the one below) and the noise $E$ is a Bernoulli random matrix, evaluated from $400$ samples (top figure) and $300$ samples (bottom figure). In both figures, the largest singular value of $A$ is taken to be $200$.} \label{young} \end{center} \end{figure} \section{ The \emph{real dimension} and new results} Trying to explain the inefficiency of the Davis-Kahan-Wedin bound in the above example, the second author was led to the following intuition. \begin{quote} If $A$ has rank $r$, all actions of $A$ focus on an $r$ dimensional subspace; intuitively then, $E$ must act like an $r$ dimensional random matrix rather than an $n$ dimensional one. \end{quote} This means that the {\it real dimension} of the problem is $r$, not $n$. While it is clear that one cannot automatically ignore the (rather wild) action of $E$ outside the range of $A$, this intuition, if true, would show that what really matters in \eqref{bound0} or \eqref{bound1} is $r$, the rank of $A$, rather than its size $n$. If this is indeed the case, one may hope to obtain a bound of the form \begin{equation} \label{bound2} \sin \angle(v_1, v_1') \leq C \frac{\sqrt r}{\delta}, \end{equation} for some constant $C$ (with some possible corrections). This is much better than \eqref{bound0} when $A$ has low rank and explains the phenomenon arising from Figure \ref{young}. In \cite{V}, the second author managed to prove \begin{equation*} \sin ^2 \angle(v_1, v'_1) \le C \frac{ \sqrt{r \log n} }{\delta } \end{equation*} under certain conditions. While the right-hand side is quite close to the optimal form in \eqref{bound2}, the main problem here is that in the left-hand side one needs to square the sine function. The bound for $\sin \angle (v_i, v_i') $ with $i \ge 2$ was done by an inductive argument and was rather complicated. Finally, the problem of estimating the singular values was not addressed at all in \cite{V}. In this paper, by using an entirely different (and simpler) argument, we are going to remove the unwanted squaring effect. This enables us to obtain a near optimal improvement of the Davis-Kahan-Wedin theorem. One can easily extend the proof to give a (again near optimal) bound on the angle between two subspaces spanned by the first few singular vectors of $A$ and their counterparts of $A+E$. (This is the space one often actually cares about in PCA and low rank approximation procedures.) Finally, as a co-product, we obtain an improved version of Weyl's bound, which also supports our {\it real dimension} intuition. Our results hold under very mild assumptions on $A$ and $E$. As a matter of fact, in the strongest results, we will not even need the entries of $E$ to be independent. As an illustration, let us first state a result in the case that $A$ is a $n \times n$ matrix and $E$ is a Bernoulli matrix (the entries are iid Bernoulli random variables, taking values $\pm 1$ with probability $1/2$). \begin{theorem} \label{theorem:main1} Let $ E$ be a $n \times n$ Bernoulli random matrix and fix $\varepsilon > 0$. Then there exists constants $C_0, \delta_0 > 0$ (depending only on $\varepsilon$) such that the following holds. Let $A$ be a $n \times n$ matrix with rank $r$ satisfying $\delta \geq \delta_0$ and $\sigma_1 \geq \max\{n,\sqrt{n}\delta\}$. Then, with probability at least $1-\varepsilon$, \begin{equation*} \sin \angle (v_1, v_1') \leq C \frac{\sqrt{r}} {\delta} . \end{equation*} \end{theorem} Notice that the assumptions on $E$ are normalized (as we assume that the variance of the entries in $E$ is one). If the error entries have variance $\sigma^2$, then we need to scale accordingly by replacing $A +E$ by $\frac{1}{\sigma} A + \frac{1}{\sigma} E $; thus, the assumptions become weaker as $\sigma$ decreases. For the singular values, a good toy result is the following \begin{theorem} \label{thm:probweyl0} Let $E$ be an $n \times n$ Bernoulli random matrix and fix $ \varepsilon > 0$. Then there exists a constant $C_0>0$ (depending only on $\varepsilon$) such that the following holds. Let $A$ be an $n \times n$ matrix with rank $r$ satisfying $ \sigma_1 \geq n$. Then with probability at least $1-\varepsilon$ \begin{equation*} \label{eq:probweylbnd0} \sigma_1 - C \leq \sigma_1' \leq \sigma_1 + C \sqrt{r}. \end{equation*} \end{theorem} It may be useful for the reader to compare these new bounds with the bounds obtained directly from the Davis-Kahan-Wedin sine theorem and Weyl's inequality (see Corollary \ref{wedin-cor}). Both theorems above are corollaries of much more general statements, which we describe in the next sections. \section{Models of random noise} In the literature, there are many models of random matrices. We can capture almost all natural models by focusing on a common property. \begin{definition} \label{def:concentration} We say the $m \times n$ random matrix $E$ is $(C_1,c_1, \gamma)$-concentrated if for all unit vectors $u \in \mathbb{R}^m, v \in \mathbb{R}^n$, and every $t>0$, \begin{equation} \label{eq:concentration} \mathbb{P}( |u^T E v| > t ) \leq C_1 \exp(-c_1 t^\gamma). \end{equation} \end{definition} The key parameter is $\gamma$. It is easy to verify the following fact, which asserts that the concentration property is closed under addition. \begin{fact} \label{fact1} If $E_1$ is $(C_1,c_1, \gamma)$-concentrated and $E_2$ is $(C_2,c_2, \gamma)$-concentrated, then $E_3 =E_1+E_2$ is $(C_3, c_3, \gamma)$-concentrated for some $C_3, c_3$ depending on $C_1,c_1, C_2, c_2$. \end{fact} Furthermore, the concentration property guarantees a bound on $\| E \|$. A standard net argument (see Lemma \ref{lemma:net}) shows \begin{fact} \label{fact2} If $E$ is $(C_1,c_1, \gamma)$-concentrated then there are constants $C', c' >0$ such that $\mathbb{P} (\| E \| \ge C' n^{1/\gamma} ) \le C_1 \exp (-c'n) $. \end{fact} For readers not familiar with random matrix theory, let us point out why the concentration property is expected to hold for any natural model. If $E$ is random and $v$ is fixed, then the vector $Ev$ must look random. It is well known that in a high dimensional space, a random vector, with very high probability, is nearly orthogonal to any fixed vector. Thus, one expects that very likely, the inner product of $u$ and $E v$ is small. Definition \ref{def:concentration} is a way to express this observation quantitatively. It turns out that all random matrices with independent entries satisfying a mild condition have the concentration property. This class covers virtually all examples one sees in practice. In particular, Lemma \ref{lemma:bernoulli} shows that if $E$ is a $n \times n$ Bernoulli random matrix, then $E$ is $\left(2, \frac{1}{2}, 2 \right)$-concentrated, and $\|E\| \leq 3 \sqrt{n}$ with high probability \cite{V,Vnorm}. A convenient feature of the definition is that independence between the entries is not a requirement. For instance, it is easy to show that a random orthogonal matrix satisfies the concentration property. We continue the discussion of the $(C_1,c_1,\gamma)$-concentration property (Definition \ref{def:concentration}) in Section \ref{sec:concentration}. \vskip2mm Let us state an extension of Theorem \ref{theorem:main1}. \begin{theorem} \label{thm:main} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$, and suppose $A$ has rank $r$. Then, for any $t>0$, $$ \sin \angle(v_1,v_1') \leq 4 \sqrt{2} \left( \frac{t r^{1/\gamma}} {\delta} + \frac{ \|E\|}{\sigma_1} + \frac{ \|E\|^2}{\sigma_1 \delta} \right) $$ with probability at least $$ 1 - 54 C_1 \exp\left(-c_1\frac{\delta^\gamma}{8^\gamma} \right) - 2C_1 9^{2r} \exp \left( -c_1 r \frac{t^\gamma}{4^{\gamma}} \right). $$ \end{theorem} \begin{remark} \label{remark:boundonE} Using Fact \ref{fact2}, one can replace $\| E\|$ on the right-hand side by $C' n^{1/\gamma }$, which yields that $$ \sin \angle(v_1,v_1') \leq 4 \sqrt{2} \left( \frac{t r^{1/\gamma}} {\delta} + \frac{ C' n^{1/\gamma} }{\sigma_1} + \frac{ C'^2 n^{2/\gamma} }{\sigma_1 \delta} \right)$$ with probability at least $$ 1 - 54 C_1 \exp\left(-c_1\frac{\delta^\gamma}{8^\gamma} \right) - 2C_1 9^{2r} \exp \left( -c_1 r \frac{t^\gamma}{4^{\gamma}} \right) - C_1 \exp( -c' n). $$ However, we prefer to state our theorems in the form of Theorem \ref{thm:main}, as the bound $C' n^{1/\gamma}$, in many cases, may not be optimal. \end{remark} \begin{remark} Another useful corollary of Theorem \ref{thm:main} is the following. For any constant $\varepsilon >0$ there are constants $C_0 = C_0(\varepsilon, C_1, c_1, \gamma) >0$ and $\delta_0 = \delta_0(\varepsilon, C_1, c_1, \gamma)$ such that if $\delta \geq \delta_0$, then \begin{equation*} \sin \angle(v_1,v_1') \leq C_0 \left( \frac{ r^{1/\gamma}} {\delta} + \frac{ \|E\|}{\sigma_1} + \frac{ \|E\|^2}{\sigma_1 \delta} \right) \end{equation*} with probability at least $1-\varepsilon$. The first term $\frac{r ^{1/\gamma}} {\delta} $ on the right-hand side corresponds to the conjectured optimal bound \eqref{bound2}. The second term $\frac{\|E\|}{ \sigma_1}$ is necessary. If $\|E \| \gg \sigma_1$, then the intensity of the noise is much stronger than the strongest signal in the data matrix, so $E$ would corrupt $A$ completely. Thus in order to retain crucial information about $A$, it seems necessary to assume $\|E\| < \sigma_1$. We are not absolutely sure about the necessity of the third term $ \frac{ \|E\|^2}{\sigma_1 \delta}$, but under the condition $\|E\| \ll \sigma_1 $, this term is superior to the Davis-Kahan-Wedin bound $\frac{\|E\| }{ \delta}$. \end{remark} We are able to extend Theorem \ref{thm:main} in two different ways. First, we can bound the angle between $v_j$ and $v_j'$ for any index $j$. Second, and more importantly, we can bound the angle between the subspaces spanned by $\{v_1, \dots, v_j \}$ and $\{v_1', \dots, v_j' \}$, respectively. As the projection onto the subspaces spanned by the first few singular vectors (i.e. low rank approximation) plays an important role in a vast collection of problems, this result potentially has a large number of applications. We are going to present these two results in the next section. To conclude this section, let us mention that related results have been obtained in the case where the random matrix $E$ contains Gaussian entries. In \cite{RRW}, R.~Wang estimates the non-asymptotic distribution of the singular vectors when the entries of $E$ are iid standard normal random variables. Recently, Allez and Bouchaud have studied the eigenvector dynamics of $A+E$ when $A$ is a real symmetric matrix and $E$ is a symmetric Brownian motion (that is, $E$ is a diffusive matrix process constructed from a family of independent real Brownian motions) \cite{AB}. Our results also seems to have a close tie to the study of spiked covariance matrices, where a different kind of perturbation has been considered; see \cite{Ma, John,Nadler} for details. It would be interesting to find a common generalization for these problems. \section{General theorems} First, we consider the problem of approximating the $j$-th singular vector $v_j$ for any $j$. In light of the Davis-Kahan-Wedin result and Theorem \ref{thm:main}, it is natural to consider the gap $$ \delta_j := \sigma_j - \sigma_{j+1}. $$ \begin{theorem} \label{thm:general} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Suppose $A$ has rank $r$, and let $1 \leq j \leq r$ be an integer. Then, for any $t>0$, $$ \sin \angle (v_j, v_j') \leq 4 \sqrt{2} \left( \left( \sum_{i=1}^{j-1} \sin^2 \angle (v_i, v_i') \right)^{1/2} + \frac{t r^{1/\gamma}}{\delta_j} + \frac{\|E\|^2}{\sigma_j \delta_j} + \frac{\|E\|}{\sigma_j} \right) $$ with probability at least $$ 1 - 6C_1 9^j \exp \left( -c_1 \frac{\delta_j^\gamma}{8^\gamma} \right) - 2C_1 9^{2r} \exp \left( -c_1 r \frac{t^\gamma}{4^\gamma} \right). $$ \end{theorem} In the next theorem, we bound the largest principal angle between \begin{equation} \label{eq:uspan} V := \mathrm{Span}\{v_1, \ldots, v_j\} \quad \text{and}\quad V' := \mathrm{Span}\{v_1', \ldots, v_j'\} \end{equation} for some integer $1 \leq j \leq r$, where $r$ is the rank of $A$. Let us recall that if $U$ and $V$ are two subspaces of the same dimension, then the (principal) angle between them is defined as \begin{equation} \label{eq:ssad} \sin \angle(U,V) := \max_{u \in U; u \neq 0} \min_{v \in V; v \neq 0} \sin \angle(u,v) = \|P_U - P_V \| = \|P_{U^\perp} P_{V} \|, \end{equation} where $P_W$ denotes the orthogonal projection onto subspace $W$. \begin{theorem} \label{thm:subspace} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Suppose $A$ has rank $r$, and let $1 \leq j \leq r$ be an integer. Then, for any $t>0$, $$ \sin \angle(V,V') \leq 4 \sqrt{2j} \left( \frac{t r^{1/\gamma}}{\delta_j} + \frac{\|E\|^2}{\sigma_j \delta_j} + \frac{ \|E\|}{\sigma_j} \right), $$ with probability at least $$ 1 - 6C_1 9^j \exp \left( -c_1 \frac{\delta_j^\gamma}{8^\gamma} \right) - 2C_1 9^{2r} \exp \left( -c_1 r \frac{t^\gamma}{4^\gamma} \right), $$ where $V$ and $V'$ are the $j$-dimensional subspaces defined in \eqref{eq:uspan}. \end{theorem} It remains an open question to give an efficient bound for subspaces corresponding to an arbitrary set of singular values. However, we can use Theorem \ref{thm:subspace} repeatedly to obtain bounds for the case when one considers a few intervals of singular values. For instance, by applying Theorem \ref{thm:subspace} twice, we obtain \begin{corollary} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Suppose $A$ has rank $r$, and let $1 < j \leq l \leq r$ be integers. Then, for any $t>0$, $$ \sin \angle(V,V') \leq 8 \sqrt{2l} \left( \frac{t r^{1/\gamma}}{\delta_{j-1}} + \frac{t r^{1/\gamma}}{\delta_l} + \frac{\|E\|^2}{\sigma_{j-1} \delta_{j-1}} + \frac{\|E\|^2}{\sigma_l \delta_l} +\frac{ \|E\|}{\sigma_l} \right), $$ with probability at least $$ 1 - 6C_1 9^{j-1} \exp \left( -c_1 \frac{\delta_{j-1}^\gamma}{8^\gamma} \right) - 6C_1 9^l \exp \left( -c_1 \frac{\delta_l^\gamma}{8^\gamma} \right) - 4C_1 9^{2r} \exp \left( -c_1 r \frac{t^\gamma}{4^\gamma} \right), $$ where $$ V:= \mathrm{Span}\{v_j,\ldots, v_l\} \quad \text{and}\quad V':=\mathrm{Span}\{v_j',\ldots,v_l'\}. $$ \end{corollary} \begin{proof} Let \begin{align*} V_1 &:= \mathrm{Span}\{v_1,\ldots,v_l\}, \quad V_1' := \mathrm{Span}\{v_1',\ldots,v_l'\}, \\ V_2 &:= \mathrm{Span}\{v_1,\ldots,v_{j-1}\}, \quad V_2' := \mathrm{Span}\{v_1',\ldots,v_{j-1}'\}. \end{align*} For any subspace $W$, let $P_W$ denote the orthogonal projection onto $W$. It follows that $P_{W^\perp} = I - P_{W}$, where $I$ denotes the identity matrix. By definition of the subspaces $V,V'$, we have $$ P_V = P_{V_1} P_{V_2^\perp} \quad \text{and}\quad P_{V'} = P_{V_1'} P_{V_2'^\perp}. $$ Thus, by \eqref{eq:ssad}, we obtain \begin{align*} \sin \angle(V,V') &= \| P_{V_1} P_{V_2^\perp} - P_{V_1'} P_{V_2'^\perp} \| \\ &\leq \| P_{V_1} P_{V_2^\perp} - P_{V_1'} P_{V_2^\perp} \| + \| P_{V_1'} P_{V_2^\perp} - P_{V_1'} P_{V_2'^\perp} \| \\ &\leq \| P_{V_1} - P_{V_1'} \| + \| P_{V_2} - P_{V_2'} \| \\ &= \sin \angle (V_1,V_1') + \sin \angle(V_2,V_2'). \end{align*} Theorem \ref{thm:subspace} can now be invoked to bound $\sin\angle(V_1,V_1')$ and $\sin\angle(V_2,V_2')$, and the claim follows. \end{proof} \noindent Finally, let us present the general form of Theorem \ref{thm:probweyl0} for singular values. \begin{theorem} \label{thm:probweyl} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Suppose $A$ has rank $r$, and let $1 \leq j \leq r$ be an integer. Then, for any $t>0$, \begin{equation} \label{eq:probweylbndlower} \sigma_j' \geq \sigma_j - t \end{equation} with probability at least $$ 1 - 2C_1 9^j \exp \left( -c_1 \frac{t^\gamma}{4^\gamma} \right), $$ and \begin{equation} \label{eq:probweylbndupper} \sigma_j' \leq \sigma_j + t r^{1/\gamma} + 2\sqrt{j} \frac{ \|E\|^2}{\sigma_j'} + j \frac{\|E\|^3}{{\sigma_j'}^2} \end{equation} with probability at least $$ 1 - 2C_1 9^{2r} \exp \left( -c_1 r \frac{t^\gamma}{4^\gamma} \right). $$ \end{theorem} \begin{remark} Notice that the upper bound for $\sigma_j'$ given in \eqref{eq:probweylbndupper} involves $1/\sigma_j'$. In many situations, the lower bound in \eqref{eq:probweylbndlower} can be used to provide an upper bound for $1/\sigma_j'$. \end{remark} \section{Overview and outline} We now briefly give an overview of the paper and discuss some of the key ideas behind the proof of our main results. For simplicity, let us assume that $A$ and $E$ are $n \times n$ real symmetric matrices. (In fact, we will symmetrize the problem in Section \ref{sec:prelim} below.) Let $\sigma_1 \geq \cdots \geq \sigma_n$ be the eigenvalues of $A$ with corresponding (orthonormal) eigenvectors $v_1, \ldots, v_n$. Let $\sigma_1'$ be the largest eigenvalue of $A+E$ with corresponding (unit) eigenvector $v_1'$. Suppose we wish to bound $\sin \angle(v_1, v_1')$ (from Theorem \ref{thm:main}). Since $$ \sin^2 \angle (v_1, v_1') = 1- \cos^2 \angle(v_1, v_1') = \sum_{k=2}^n | v_k \cdot v_1' |^2, $$ it suffices to bound $|v_k \cdot v_1'|$ for $k=2, \ldots, n$. Let us consider the case when $k=2, \ldots, r$. In this case, we have $$ v_k^\mathrm{T} (A+E) v_1' - v_k^\mathrm{T} A v_1' = v_k^\mathrm{T} E v_1'. $$ Since $(A+E) v_1' = \sigma_1' v_1'$ and $v_k^\mathrm{T} A = \sigma_k v_k$, we obtain $$ |\sigma_1' - \sigma_k| |v_k \cdot v_1'| \leq | v_k^\mathrm{T}E v_1' |. $$ Thus, the problem of bounding $|v_k \cdot v_1'|$ reduces to obtaining an upper bound for $| v_k^\mathrm{T}E v_1' |$ and a lower bound for the gap $|\sigma_1' - \sigma_k|$. We will obtain bounds for both of these terms by using the concentration property (Definition \ref{def:concentration}). More generally, in Section \ref{sec:prelim}, we will apply the concentration property to obtain lower bounds for the gaps $\sigma_j' - \sigma_k$ when $j < k$, which will hold with high probability. Let us illustrate this by now considering the gap $\sigma_1' - \sigma_2$. Indeed, we note that $$ \sigma_1' = \| A + E \| \geq v_1^\mathrm{T} (A+E) v_1 = \sigma_1 + v_1^\mathrm{T} E v_1. $$ Applying the concentration property \eqref{eq:concentration}, we see that $\sigma_1' > \sigma_1 - t$ with probability at least $1 - C_1 \exp(- c_1 t^{\gamma})$. As $\delta := \sigma_1 - \sigma_2$, we in fact observe that $$ \sigma_1' - \sigma_2 = \sigma_1' - \sigma_1 + \delta > \delta - t. $$ Thus, if $\delta$ is sufficiently large, we have (say) $\sigma_1' - \sigma_2 \geq \delta/2$ with high probability. In Section \ref{sec:proof}, we will again apply the concentration property to obtain upper bounds for terms of the form $v_k E v_j'$. At the end of Section \ref{sec:proof}, we combine these bounds to complete the proof of Theorems \ref{thm:main}, \ref{thm:general}, \ref{thm:subspace} and \ref{thm:probweyl}. In Section \ref{sec:concentration}, we discuss the $(C_1,c_1,\gamma)$-concentration property (Definition \ref{def:concentration}). In particular, we generalize some previous results obtained by the second author in \cite{V}. Finally, in Section \ref{section:app}, we present some applications of our main results. Among others, our results seem useful for matrix recovery problems. The general matrix recovery problem is the following. $A$ is a large matrix. However, the matrix $A$ is unknown to us. We can only observe its noisy perturbation $A+E$, or in some cases just a small portion of the perturbation. Our goal is to reconstruct $A$ or estimate an important parameter as accurately as possible from this observation. Furthermore, several problems from combinatorics and theoretical computer science can also be formulated in this setting. Special instances of the matrix recovery problem have been investigated by many researchers using spectral techniques and combinatorial arguments in ingenious ways \cite{AM, AK, AKS,AzarMc,CCS,CP,CR,CRT,CT,Cest, DGP, KMO,KMO2,Krank,KLT,Kucera,MHT,Mc,NW,RVsamp}. We propose the following simple analysis: if $A$ has rank $r$ and $1 \le j \le r$, then the projection of $A+E$ on the subspace $V'$ spanned by the first $j$ singular vectors of $A+E$ is close to the projection of $A+E$ onto the subspace $V$ spanned by the first $j$ singular vectors of $A$, as our new results show that $V$ and $V'$ are very close. Moreover, we can also show that the projection of $E$ onto $V$ is typically small. Thus, by projecting $A+E$ onto $V'$, we obtain a good approximation of the rank $j$ approximation of $A$. In certain cases, we can repeat the above operation a few times to obtain sufficient information to recover $A$ completely or to estimate the required parameter with high accuracy and certainty. \section{Preliminary tools} \label{sec:prelim} In this section, we present some of the preliminary tools we will need to prove Theorems \ref{thm:main}, \ref{thm:general}, \ref{thm:subspace}, and \ref{thm:probweyl}. To begin, we define the $(m+n) \times (m+n)$ symmetric block matrices \begin{equation} \label{eq:def:tildeA} \tilde{A} := \begin{bmatrix} 0 & A \\ A^\mathrm{T} & 0 \end{bmatrix} \end{equation} and $$ \tilde{E} := \begin{bmatrix} 0 & E \\ E^\mathrm{T} & 0 \end{bmatrix}. $$ We will work with the matrices $\tilde{A}$ and $\tilde{E}$ instead of $A$ and $E$. In particular, the non-zero eigenvalues of $\tilde{A}$ are $\pm \sigma_1, \ldots, \pm \sigma_r$ and the eigenvectors are formed from the left and right singular vectors of $A$. Similarly, the non-trivial eigenvalues of $\tilde{A} + \tilde{E}$ are $\pm \sigma_1', \ldots, \pm \sigma_{\min\{m,n\}}'$ (some of which may be zero) and the eigenvectors are formed from the left and right singular vectors of $A+E$. Along these lines, we introduce the following notation, which differs from the notation used above. The non-zero eigenvalues of $\tilde{A}$ will be denoted by $\pm \sigma_1, \ldots, \pm \sigma_r$ with orthonormal eigenvectors $u_k$, $k=\pm 1, \ldots, \pm r$ such that $$ \tilde{A} u_k = \sigma_k u_k, \qquad \tilde{A} u_{-k} = - \sigma_k u_{-k}, \qquad k = 1, \ldots, r. $$ Let $v_1, \ldots, v_j$ be the orthonormal eigenvectors of $\tilde{A}+\tilde{E}$ corresponding to the $j$-largest eigenvalues $\lambda_1 \geq \cdots \geq \lambda_j$. In order to prove Theorems \ref{thm:main}, \ref{thm:general}, \ref{thm:subspace}, and \ref{thm:probweyl}, it suffices to work with the eigenvectors and eigenvalues of the matrices $\tilde{A}$ and $\tilde{A}+\tilde{E}$. Indeed, Proposition \ref{prop:sine} will bound the angle between the singular vectors of $A$ and $A+E$ by the angle between the corresponding eigenvectors of $\tilde{A}$ and $\tilde{A} + \tilde{E}$. \begin{proposition} \label{prop:sine} Let $u_1, v_1 \in \mathbb{R}^m$ and $u_2, v_2 \in \mathbb{R}^n$ be unit vectors. Let $u, v \in \mathbb{R}^{m+n}$ be given by $$ u = \begin{bmatrix} u_1 \\ u_2 \end{bmatrix}, \quad v = \begin{bmatrix} v_1 \\ v_2 \end{bmatrix}. $$ Then $$ \sin^2 \angle(u_1, v_1) + \sin^2 \angle(u_2, v_2) \leq 2 \sin^2 \angle(u, v). $$ \end{proposition} \begin{proof} Since $\|u\|^2 = \|v\|^2 = 2$, we have \begin{align*} \cos^2 \angle(u,v) = \frac{1}{4} |u \cdot v|^2 \leq \frac{1}{2} |u_1 \cdot v_1|^2 + \frac{1}{2} |u_2 \cdot v_2|^2. \end{align*} Thus, \begin{align*} \sin^2 \angle(u,v) = 1 - \cos^2 \angle(u,v) \geq \frac{1}{2} \sin^2 \angle(u_1, v_1) + \frac{1}{2} \sin^2\angle(u_2, v_2), \end{align*} and the claim follows. \end{proof} We now introduce some useful lemmas. The first lemma below, states that if $E$ is $(C_1,c_1,\gamma)$-concentrated, then $\tilde{E}$ is $(\tilde{C}_1,\tilde{c}_1,\gamma)$-concentrated, for some new constants $\tilde{C}_1 := 2C_1$ and $\tilde{c}_1:= c_1/2^{\gamma}$. \begin{lemma} \label{lemma:tilde} Assume that $E$ is $(C_1, c_1,\gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Let $\tilde{C}_1 := 2C_1$ and $\tilde{c}_1:= c_1/2^{\gamma}$. Then for all unit vectors $u,v \in \mathbb{R}^{n+m}$, and every $t>0$, \begin{equation} \label{eq:tilde-concentration} \mathbb{P}( |u^t \tilde{E} v| > t ) \leq \tilde{C}_1 \exp(-\tilde{c}_1 t^\gamma ). \end{equation} \end{lemma} \begin{proof} Let $$ u = \begin{bmatrix} u_1 \\ u_2 \end{bmatrix}, \quad v = \begin{bmatrix} v_1 \\ v_2 \end{bmatrix} $$ be unit vectors in $\mathbb{R}^{m+n}$, where $u_1, v_1 \in \mathbb{R}^m$ and $u_2,v_2 \in \mathbb{R}^n$. We note that $$ u^{\mathrm{T}} \tilde{E} v = u_1^\mathrm{T} E v_2 + u_2^\mathrm{T} E^\mathrm{T} v_1. $$ Thus, if any of the vectors $u_1, u_2, v_1, v_2$ are zero, \eqref{eq:tilde-concentration} follows immediately from \eqref{eq:concentration}. Assume all the vectors $u_1, u_2, v_1, v_2$ are nonzero. Then $$ | u^{\mathrm{T}} \tilde{E} v | = |u_1^\mathrm{T} E v_2 + u_2^\mathrm{T} E^\mathrm{T} v_1| \leq \frac{|u_1^\mathrm{T} E v_2|}{\|u_1\| \|v_2\|} + \frac{|v_1^\mathrm{T} E u_2|}{\|u_2\|\|v_1\|}. $$ Thus, by \eqref{eq:concentration}, we have \begin{align*} \mathbb{P}( |u^\mathrm{T} \tilde{E} v| > t) &\leq \mathbb{P} \left( \frac{|u_1^\mathrm{T} E v_2|}{\|u_1\| \|v_2\|} > \frac{t}{2} \right) + \mathbb{P} \left( \frac{|v_1^\mathrm{T} E u_2|}{\|u_2\|\|v_1\|} > \frac{t}{2} \right) \\ &\leq 2C_1 \exp \left( -c_1 \frac{ t^\gamma }{ 2^\gamma } \right), \end{align*} and the proof of the lemma is complete. \end{proof} We will also consider the spectral norm of $\tilde{E}$. Since $\tilde{E}$ is a symmetric matrix whose eigenvalues in absolute value are given by the singular values of $E$, it follows that \begin{equation} \label{eq:normE} \| \tilde{E} \| = \|E \|. \end{equation} We introduce $\varepsilon$-nets as a convenient way to discretize a compact set. Let $\varepsilon > 0$. A set $X$ is an $\varepsilon$-net of a set $Y$ if for any $y \in Y$, there exists $x \in X$ such that $\|x-y\| \leq \varepsilon$. The following estimate for the maximum size of an $\varepsilon$-net of a sphere is well-known (see for instance \cite{RV}). \begin{lemma} \label{lemma:net} A unit sphere in $d$ dimensions admits an $\varepsilon$-net of size at most $$ \left(1+\frac{2}{\varepsilon} \right)^d.$$ \end{lemma} Lemmas \ref{lemma:r-norm}, \ref{lemma:largest}, and \ref{lemma:j-largest} below are consequences of the concentration property \eqref{eq:tilde-concentration}. \begin{lemma} \label{lemma:r-norm} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Let $A$ be a $m \times n$ matrix with rank $r$. Let $U$ be the $(m+n) \times 2r$ matrix whose columns are the vectors $u_1, \ldots, u_r, u_{-1}, \ldots, u_{-r}$. Then, for any $t > 0$, $$ \mathbb{P} \left( \|U^\mathrm{T} \tilde{E}U\| > t r^{1/\gamma} \right) \leq \tilde{C}_1 9^{2r} \exp\left( - \tilde{c}_1 r \frac{t^\gamma}{2^\gamma} \right). $$ \end{lemma} \begin{proof} Clearly $U^\mathrm{T} \tilde{E} U$ is a symmetric $2r \times 2r$ matrix. Let $S$ be the unit sphere in $\mathbb{R}^{2r}$. Let $\mathcal{N}$ be a $1/4$-net of $S$. It is easy to verify (see for instance \cite{RV}) that for any $2r \times 2r$ symmetric matrix $B$, $$ \|B\| \leq 2 \max_{x \in \mathcal{N}} |x^\ast B x| . $$ For any fixed $x \in \mathcal{N}$, we have $$ \mathbb{P}( |x^\mathrm{T} U^\mathrm{T} \tilde{E} U x | > t) \leq \tilde{C}_1 \exp(-\tilde{c}_1 t^\gamma) $$ by Lemma \ref{lemma:tilde}. Since $|\mathcal{N}| \leq 9^{2r}$, we obtain \begin{align*} \mathbb{P} ( \| U^\mathrm{T} \tilde{E} U \| > t r^{1/\gamma}) &\leq \sum_{x \in \mathcal{N}} \mathbb{P}\left( |x^\mathrm{T} U^\mathrm{T} \tilde{E} U x | > \frac{1}{2} t r^{1/\gamma} \right) \\ & \leq \tilde{C}_1 9^{2r} \exp\left(-\tilde{c}_1 r \frac{t^\gamma}{2^{\gamma}} \right). \end{align*} \end{proof} \begin{lemma} \label{lemma:largest} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Suppose $A$ has rank $r$. Then, for any $t > 0$, \begin{equation} \label{eq:lambda1bnd} \lambda_1 \geq \sigma_1 - t \end{equation} with probability at least $1 - \tilde{C}_1 \exp(-\tilde{c}_1 t^{\gamma})$. In particular, if $\sigma_1 > 0$, then $\lambda_1 \geq \frac{\sigma_1}{2}$ with probability at least $1 - \tilde{C}_1 \exp \left( -\tilde{c}_1 \frac{\sigma_1^\gamma}{2^\gamma} \right)$. If, in addition, $\delta > 0$, then $$ \lambda_1 - \sigma_k \geq \frac{1}{2} \delta $$ for $k = 2,\ldots, r $ with probability at least $1 - \tilde{C}_1 \exp \left( -\tilde{c}_1 \frac{\delta^\gamma}{2^\gamma} \right)$. \end{lemma} \begin{proof} We observe that $$ \lambda_1 = \|\tilde{A} + \tilde{E}\| \geq u_1^\mathrm{T} (\tilde{A} + \tilde{E}) u_1 = \sigma_1 + u_1^\mathrm{T} \tilde{E} u_1. $$ By Lemma \ref{lemma:tilde}, we have $$ \mathbb{P}( |u_1^\mathrm{T} \tilde{E} u_1| > t) \leq \tilde{C}_1 \exp( -\tilde{c}_1 t^\gamma) $$ for every $t > 0$, and \eqref{eq:lambda1bnd} follows. If $\sigma_1 > 0$, then the bound $\lambda_1 \geq \frac{\sigma_1}{2}$ can be obtained by taking $t = \sigma_1/2$ in \eqref{eq:lambda1bnd}. Assume $\delta > 0$. Taking $t=\delta/2$ in \eqref{eq:lambda1bnd} yields $$ \lambda_1 - \sigma_k \geq \lambda_1 - \sigma_2 = \lambda_1 - \sigma_1 + \delta \geq \frac{\delta}{2} $$ for $k=2, \ldots, r$ with probability at least $1 - \tilde{C}_1 \exp \left( -\tilde{c}_1 \frac{\delta^\gamma}{2^\gamma} \right)$. \end{proof} Using the Courant minimax principle, Lemma \ref{lemma:largest} can be generalized to the following. \begin{lemma} \label{lemma:j-largest} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Suppose $A$ has rank $r$, and let $1 \leq j \leq r$ be an integer. Then, for any $t > 0$, \begin{equation} \label{eq:lambdajbnd1} \lambda_j \geq \sigma_j - t \end{equation} with probability at least $1 - \tilde{C}_1 9^j \exp\left( - \tilde{c}_1 \frac{t^\gamma}{2^\gamma} \right)$. In particular, $\lambda_j \geq \frac{\sigma_j}{2}$ with probability at least $1 - \tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\sigma_j^\gamma}{4^\gamma} \right)$. In addition, if $\delta_j > 0$, then \begin{equation} \label{eq:lambdajbnd2} \lambda_j - \sigma_k \geq \frac{\delta_j}{2} \end{equation} for $k=j+1, \ldots, r$ with probability at least $1 - \tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\delta_j^\gamma}{4^\gamma} \right)$. \end{lemma} \begin{proof} It suffices to prove \eqref{eq:lambdajbnd1}. Indeed, the bound $\lambda_j \geq \frac{\sigma_j}{2}$ follows from \eqref{eq:lambdajbnd1} by taking $t = \sigma_j/2$, and \eqref{eq:lambdajbnd2} follows by taking $t = \delta_j/2$. Let $S$ be the unit sphere in $\mathrm{Span}\{u_1,\ldots,u_j\}$. By the Courant minimax principle, \begin{align*} \lambda_j &= \max_{\dim(V)=j} \min_{\|v\|=1;v\in V} v^\mathrm{T} (\tilde{A}+\tilde{E})v \\ & \geq \min_{v \in S} v^\mathrm{T} (\tilde{A}+\tilde{E})v \\ & \geq \sigma_j + \min_{v \in S} v^\mathrm{T} \tilde{E} v. \end{align*} Thus, it suffices to show $$ \mathbb{P}\left( \sup_{v \in S} |v^\mathrm{T} \tilde{E} v| > t \right) \leq \tilde{C}_1 9^j \exp\left( -\tilde{c}_1 \frac{t^\gamma}{2^\gamma} \right) $$ for all $t > 0$. Let $\mathcal{N}$ be a $1/4$-net of $S$. By Lemma \ref{lemma:net}, $|\mathcal{N}| \leq 9^{j}$. We now claim that \begin{equation} \label{eq:supmaxnet} T := \sup_{v \in S} | v^\mathrm{T} \tilde{E} v| \leq 2 \max_{ u \in \mathcal{N}} |u^\mathrm{T} \tilde{E} u|. \end{equation} Indeed, fix a realization of $\tilde{E}$. Since $S$ is compact, there exists $v \in S$ such that $T = |v^\mathrm{T} \tilde{E} v|$. Moreover, there exists $x \in \mathcal{N}$ such that $\|x - v\| \leq 1/4$. Clearly the claim is true when $x = v$; assume $x \neq v$. Then, by the triangle inequality, we have \begin{align*} T &\leq |v^\mathrm{T} \tilde{E} v - v^\mathrm{T} \tilde{E} x| + |v^\mathrm{T} \tilde{E} x - x^\mathrm{T} \tilde{E} x| + |x^\mathrm{T} \tilde{E} x| \\ &\leq \frac{1}{4} \frac{ |v^\mathrm{T} \tilde{E} (v-x)| }{\| v-x\|} + \frac{1}{4} \frac{ |(v-x)^\mathrm{T} \tilde{E} x |}{\|v-x\| } + \sup_{u \in \mathcal{N}} |u^\mathrm{T} \tilde{E} u| \\ & \leq \frac{T}{2} + \sup_{u \in \mathcal{N}} |u^\mathrm{T} \tilde{E} u|, \end{align*} and \eqref{eq:supmaxnet} follows. Applying \eqref{eq:supmaxnet} and Lemma \ref{lemma:tilde}, we have \begin{align*} \mathbb{P} \left( \sup_{v \in S} |v^\mathrm{T} \tilde{E} v| > t \right) &\leq \sum_{u \in \mathcal{N}} \mathbb{P} \left( |u^\mathrm{T} \tilde{E} u| > \frac{t}{2} \right) \leq 9^j \tilde{C}_1 \exp \left( -\tilde{c}_1 \frac{t^\gamma}{2^\gamma} \right), \end{align*} and the proof of the lemma is complete. \end{proof} We will continually make use of the following simple fact: \begin{equation} \label{eq:aea} (\tilde{A} + \tilde{E}) - \tilde{A} = \tilde{E}. \end{equation} \section{Proof of Theorems \ref{thm:main}, \ref{thm:general}, \ref{thm:subspace}, and \ref{thm:probweyl}} \label{sec:proof} This section is devoted to Theorems \ref{thm:main}, \ref{thm:general}, \ref{thm:subspace}, and \ref{thm:probweyl}. To begin, define the subspace $$ W:= \mathrm{Span}\{u_1, \ldots, u_r, u_{-1}, \ldots, u_{-r} \}.$$ Let $P$ be the orthogonal projection onto $W^\perp$. \begin{lemma} \label{lemma:proj_bound} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Suppose $A$ has rank $r$, and let $1 \leq j \leq r$ be an integer. Then \begin{equation} \label{eq:suppvi} \sup_{1 \leq i \leq j} \|P v_i \| \leq 2 \frac{\|E\|}{\sigma_j} \end{equation} with probability at least $1 - \tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\sigma_j^\gamma}{4^\gamma} \right)$. \end{lemma} \begin{proof} Consider the event $$ \Omega_j := \left\{ \lambda_j \geq \frac{1}{2} \sigma_j \right\}. $$ By Lemma \ref{lemma:j-largest} (or Lemma \ref{lemma:largest} in the case $j=1$), $\Omega_j$ holds with probability at least $1 - \tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\sigma_j^\gamma}{4^\gamma} \right)$. Fix $1 \leq i \leq j$. By multiplying \eqref{eq:aea} on the left by $(P v_i)^\mathrm{T}$ and on the right by $v_i$, we obtain $$ | \lambda_i (P v_i)^\mathrm{T} v_i | \leq \| P v_i\| \|\tilde{E}\| $$ since $(P v_i)^\mathrm{T} \tilde{A} = 0$. Thus, on the event $\Omega_j$, we have $$ \| P v_i \|^2 = |(P v_i)^\mathrm{T} v_i | \leq \frac{1}{\lambda_j}\|P v_i\| \|\tilde{E}\| \leq \frac{2}{\sigma_j} \| P v_i \|\|\tilde{E}\|. $$ We conclude that, on the event $\Omega_j$, $$ \sup_{1 \leq i \leq j} \|P v_i \| \leq 2 \frac{\|E\|}{\sigma_j}, $$ and the proof is complete. \end{proof} \begin{lemma} \label{lemma:uproj} Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Suppose $A$ has rank $r$, and let $1 \leq j \leq r$ be an integer. Define $U_j$ to be the $(m+n) \times (2r-j)$ matrix with columns $u_{j+1}, \ldots, u_r, u_{-1}, \ldots, u_{-r}$. Then, for any $t>0$, \begin{equation} \label{eq:sujtv} \sup_{1 \leq i \leq j} \| U_j^\mathrm{T} v_i \| \leq 4 \left( \frac{t r^{1/\gamma}}{\delta_j} + \frac{\|E\|^2}{\delta_j \sigma_j} \right) \end{equation} with probability at least $$ 1 - 2\tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\delta_j^\gamma}{4^\gamma}\right) - \tilde{C}_1 9^{2r} \exp \left( - \tilde{c}_1r \frac{t^\gamma}{2^\gamma} \right). $$ \end{lemma} \begin{proof} Define the event \begin{align*} \Omega_j &:= \left\{ \sup_{1 \leq i \leq j} \|P v_i \| \leq 2 \frac{\|E\|}{\sigma_j} \right\} \bigcap \left\{ \| U^\mathrm{T} \tilde{E} U \| \leq t r^{1/\gamma} \right\} \bigcap \left\{ \lambda_j - \sigma_{j+1} \geq \frac{\delta_j}{2} \right\}. \end{align*} By Lemmas \ref{lemma:r-norm}, \ref{lemma:j-largest}, and \ref{lemma:proj_bound}, it follows that $$ \mathbb{P}( \Omega_j ) \geq 1 - 2\tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\delta_j^\gamma}{4^\gamma}\right) - \tilde{C}_1 9^{2r} \exp \left( - \tilde{c}_1r \frac{t^\gamma}{2^\gamma} \right). $$ Fix $1 \leq i \leq j$. We multiply \eqref{eq:aea} on the left by $U_j^\mathrm{T}$ and on the right by $v_i$ to obtain \begin{equation} \label{eq:ujae} U_j^\mathrm{T} (\tilde{A} + \tilde{E}) v_i - U_j^\mathrm{T} \tilde{A} v_i = U_j^\mathrm{T} \tilde{E} v_i. \end{equation} We note that $$ U_j^\mathrm{T} (\tilde{A} + \tilde{E}) v_i = \lambda_i U_j^\mathrm{T} v_i $$ and $$ U_j^\mathrm{T} \tilde{A} v_i = D_j U_j^\mathrm{T} v_i, $$ where $D_j$ is the diagonal matrix with the values $\sigma_{j+1}, \ldots, \sigma_r, -\sigma_{1}, \ldots, -\sigma_{r}$ on the diagonal. For the right-hand side of \eqref{eq:ujae}, we write $v_i = U U^\mathrm{T} v_i + P v_i$, where $U$ is the matrix with columns $u_1, \ldots, u_r, u_{-1}, \ldots, u_{-r}$ and $P$ is the orthogonal projection onto $W^\perp$. Thus, on the event $\Omega_j$, we have \begin{align*} \|U_j^\mathrm{T} \tilde{E} v_i\| \leq \|U_j^\mathrm{T} \tilde{E} U\| + \|\tilde{E}\| \|P v_i\| \leq t r^{1/\gamma} + 2 \frac{\|E\|^2}{\sigma_j}. \end{align*} Here we used the fact that $U_j^\mathrm{T} \tilde{E} U$ is a sub-matrix of $U^\mathrm{T} \tilde{E} U$ and hence $$ \|U_j^\mathrm{T} \tilde{E} U\| \leq \| U^\mathrm{T} \tilde{E} U\|. $$ Combining the above computations and bound yields $$ \| (\lambda_i I - D_j) U_j^\mathrm{T} v_i \| \leq 2 \left( t r^{1/\gamma} + \frac{\|E\|^2}{\sigma_j} \right) $$ on the event $\Omega_j$. We now consider the entries of the diagonal matrix $\lambda_i I - D_j$. On $\Omega_j$, we have that, for any $k \geq j+1$, $$ \lambda_i - \sigma_k \geq \lambda_j - \sigma_{j+1} \geq \frac{\delta_j}{2}. $$ By writing the elements of the vector $U_j^\mathrm{T} v_i$ in component form, it follows that $$ \|(\lambda_i I - D_j) U_j^\mathrm{T} v_i \| \geq \frac{\delta_j}{2} \| U_j^\mathrm{T} v_i \| $$ and hence $$ \| U_j^\mathrm{T} v_i \| \leq 4 \left( \frac{t r^{1/\gamma}}{\delta_j} + \frac{\|E\|^2}{\sigma_j \delta_j} \right) $$ on the event $\Omega_j$. Since this holds for each $1 \leq i \leq j$, the proof is complete. \end{proof} With Lemmas \ref{lemma:proj_bound} and \ref{lemma:uproj} in hand, we now prove Theorems \ref{thm:main}, \ref{thm:general}, \ref{thm:subspace}, and \ref{thm:probweyl}. By Proposition \ref{prop:sine}, in order to prove Theorems \ref{thm:main} and \ref{thm:general}, it suffices to bound $\sin \angle (u_j, v_j)$ because $u_j, v_j$ are formed from the left and right singular vectors of $A$ and $A+E$. \begin{proof}[Proof of Theorem \ref{thm:main}] We write $$ v_1 = \sum_{k=1}^r \alpha_k u_k + \sum_{k=1}^r \alpha_{-k} u_{-k} + P v_1, $$ where $P$ is the orthogonal projection onto $W^\perp$. Then $$ \sin^2 \angle (u_1, v_1) = 1- \cos^2 \angle (u_1, v_1) = \sum_{k=2}^r |\alpha_k|^2 + \sum_{k=1}^r |\alpha_{-k}|^2 + \| P v_1 \|^2. $$ Applying the bounds obtained from Lemmas \ref{lemma:proj_bound} and \ref{lemma:uproj} (with $j=1$), we obtain $$ \sin^2 \angle (u_1, v_1) \leq 16 \left( \frac{t r^{1/\gamma}}{\delta} + \frac{\|E\|^2}{\sigma_1 \delta} \right)^2 + 4 \frac{\|E\|^2}{ \sigma_1^2} $$ with probability at least \begin{equation} \label{eq:probholdmain} 1 - 27 \tilde{C}_1 \exp \left( -\tilde{c}_1 \frac{\delta^\gamma}{4^\gamma} \right) - \tilde{C}_1 9^{2r} \exp \left(- \tilde{c}_1 r \frac{t^\gamma}{2^\gamma} \right). \end{equation} We now note that \begin{align*} 16 \left( \frac{t r^{1/\gamma}}{\delta} + \frac{\|E\|^2}{\sigma_1 \delta} \right)^2 + 4 \frac{\|E\|^2}{ \sigma_1^2} &\leq 16 \left( \frac{t r^{1/\gamma}}{\delta} + \frac{\|E\|^2}{\sigma_1 \delta} + \frac{\|E\|}{ \sigma_1} \right)^2. \end{align*} The correct absolute constant in front can now be deduced from the bound above and Proposition \ref{prop:sine}. The lower bound on the probability given in \eqref{eq:probholdmain} can be written in terms of the constants $C_1, c_1, \gamma$ by recalling the definitions of $\tilde{C}_1$ and $\tilde{c}_1$ given in Lemma \ref{lemma:tilde}. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:general}] We again write \begin{equation} \label{eq:vjproj} v_j = \sum_{k=1}^r \alpha_k u_k + \sum_{k=1}^r \alpha_{-k} u_{-k} + P v_j, \end{equation} where $P$ is the orthogonal projection onto $W^\perp$. Then we have that \begin{align*} \sin^2 \angle (u_j, v_j) &= 1- \cos^2 \angle (u_j, v_j) \\ &= \sum_{k=1}^{j-1} |\alpha_k|^2 + \sum_{k=j+1}^r |\alpha_k|^2 + \sum_{k=1}^r |\alpha_{-k}|^2 + \|P v_j\|^2. \end{align*} For any $1 \leq k \leq j-1$, we have that $$ |\alpha_k|^2 = | v_j \cdot (u_k - v_k) |^2 \leq \|v_k - u_k\|^2 \leq 2 (1 - \cos \angle (v_k,u_k)) \leq 2 \sin^2 \angle(v_k,u_k). $$ Moreover, from Lemmas \ref{lemma:proj_bound} and \ref{lemma:uproj}, we have $$ \sum_{k=j+1}^r |\alpha_k|^2 + \sum_{k=1}^r |\alpha_{-k}|^2 \leq 16 \left( \frac{t r^{1/\gamma}}{\delta_j} + \frac{\|E\|^2}{\sigma_j \delta_j} \right)^2 $$ with probability at least $$ 1 - 2\tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\delta_j^\gamma}{4^\gamma}\right) - \tilde{C}_1 9^{2r} \exp \left( - \tilde{c}_1r \frac{t^\gamma}{2^\gamma} \right). $$ and $$ \| P v_j \|^2 \leq 4 \frac{\|E\|^2}{\sigma_j^2} $$ with probability at least $1 - \tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\sigma_j^\gamma}{4^\gamma} \right)$. The proof of Theorem \ref{thm:general} is complete by combining the bounds above\footnote{Here the bounds are given in terms of $\sin^2 \angle(v_k, u_k)$ for $1 \leq k \leq j-1$. However, $u_k$ and $v_k$ are formed from the left and right singular vectors of $A$ and $A+E$. To avoid the dependence on both the left and right singular vectors, one can begin with \eqref{eq:vjproj} and consider only the coordinates of $v_j$ which correspond to the left (alternatively right) singular vectors. By then following the proof for only these coordinates, one can bound the left (right) singular vectors by terms which only depend on the previous left (right) singular vectors.}. As in the proof of Theorem \ref{thm:main}, the correct constant factor in front can be deduced from Proposition \ref{prop:sine}. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:subspace}] Define the subspaces $$ \tilde{U}:= \mathrm{Span}\{u_1, \ldots, u_j \} \quad \text{and} \quad \tilde{V}:= \mathrm{Span}\{v_1, \ldots, v_j\}. $$ By Proposition \ref{prop:sine}, it suffices to bound $\sin \angle(\tilde{U}, \tilde{V})$. Let $Q$ be the orthogonal projection onto $\tilde{U}^\perp$. By Lemmas \ref{lemma:proj_bound} and \ref{lemma:uproj}, it follows that \begin{equation} \label{eq:supqvi} \sup_{1 \leq i \leq j} \|Q v_i \| \leq 4 \left( \frac{t r^{1/\gamma} }{\delta_j} + \frac{\|E\|^2}{\sigma_j \delta_j} + \frac{\|E\|}{\sigma_j} \right) \end{equation} with probability at least $$ 1 - 3\tilde{C}_1 9^j \exp \left( - \tilde{c}_1 \frac{\delta_j^\gamma}{4^\gamma}\right) - \tilde{C}_1 9^{2r} \exp \left( - \tilde{c}_1r \frac{t^\gamma}{2^\gamma} \right). $$ On the event where \eqref{eq:supqvi} holds, we have $$ \sup_{v \in \tilde{V}, \|v\| = 1} \| Q v \| \leq 4 \sqrt{j} \left( \frac{t r^{1/\gamma} }{\delta_j} + \frac{\|E\|^2}{\sigma_j \delta_j} + \frac{\|E\|}{\sigma_j} \right) $$ by the triangle inequality and the Cauchy-Schwarz inequality. Thus, by \eqref{eq:ssad}, we conclude that \begin{align*} \sin \angle(\tilde{U}, \tilde{V}) &\leq 4 \sqrt{j} \left( \frac{t r^{1/\gamma} }{\delta_j} + \frac{\|E\|^2}{\sigma_j \delta_j} + \frac{\|E\|}{\sigma_j} \right) \end{align*} on the event where \eqref{eq:supqvi} holds. The claim now follows from Proposition \ref{prop:sine}. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:probweyl}] The lower bound \eqref{eq:probweylbndlower} follows from Lemma \ref{lemma:j-largest}; it remains to prove \eqref{eq:probweylbndupper}. Let $U$ be the $(m+n) \times 2r$ matrix whose columns are given by the vectors $u_1, \ldots, u_r, u_{-1}, \ldots, u_{-r}$, and recall that $P$ is the orthogonal projection onto $W^\perp$. Let $S$ denote the unit sphere in $\mathrm{Span}\{v_1, \ldots, v_j\}$. Then for $1 \leq i \leq j$, we multiply \eqref{eq:aea} on the left by $v_i^\mathrm{T} P$ and on the right by $v_i$ to obtain $$ \lambda_i \|P v_i \|^2 \leq \| v_i^\mathrm{T} P \tilde{E} v_i \| \leq \|P v_i \| \|E\|. $$ Here we used \eqref{eq:normE} and the fact that $P \tilde{A} = 0$. Therefore, we have the deterministic bound $$ \sup_{1 \leq i \leq j} \| P v_i \| \leq \frac{ \| E\|}{\lambda_j}. $$ By the Cauchy-Schwarz inequality, it follows that \begin{equation} \label{eq:detsupbnd} \sup_{v \in S} \| P v \| \leq \sqrt{j} \frac{ \|E \| }{\lambda_j}. \end{equation} By the Courant minimax principle, we have \begin{align*} \sigma_j = \max_{\dim(V)=j} \min_{v\in V,\|v\|=1} v^\mathrm{T} \tilde{A}v \geq \min_{v \in S} v^\mathrm{T} \tilde{A} v \geq \lambda_j - \max_{v \in S} |v^\mathrm{T} \tilde{E} v|. \end{align*} Thus, it suffices to show that \begin{equation*} \max_{v \in S} |v^\mathrm{T} \tilde{E} v| \leq t r^{1/\gamma} + 2\sqrt{j} \frac{ \|E\|^2}{\lambda_j} + j \frac{\|E\|^3}{{\lambda_j}^2} \end{equation*} with probability at least $1-\tilde{C}_1 9^{2r} \exp \left( - \tilde{c}_1 r \frac{t^\gamma}{2^\gamma} \right)$. We decompose $v = Pv + U U^\mathrm{T} v$ and obtain \begin{align*} \max_{v \in S} |v^\mathrm{T} \tilde{E} v| \leq \max_{v \in S} \|Pv\|^2 \|\tilde{E}\| + 2 \max_{v \in S} \| Pv \| \|\tilde{E}\| + \| U^\mathrm{T} \tilde{E} U \|. \end{align*} Thus, by Lemma \ref{lemma:r-norm} and \eqref{eq:detsupbnd}, we have $$ \max_{v \in S} |v^\mathrm{T} \tilde{E} v| \leq j \frac{ \|E\|^3}{\lambda_j^2} + 2 \sqrt{j} \frac{ \|E\|^2}{\lambda_j} + t r^{1/\gamma} $$ with probability at least $1-\tilde{C}_1 9^{2r} \exp \left( - \tilde{c}_1 r \frac{t^\gamma}{2^\gamma} \right)$, and the proof is complete. \end{proof} \section{The concentration property} \label{sec:concentration} In this section, we give examples of random matrix models satisfying Definition \ref{def:concentration}. \begin{lemma} \label{lemma:bernoulli} There exists a constant $C_1$ such that the following holds. Let $E$ be a random $n \times n$ Bernoulli matrix. Then $$ \mathbb{P} ( \|E\| > 3 \sqrt{n} ) \leq \exp(-C_1 n), $$ and for any fixed unit vectors $u,v$ and positive number $t$, $$ \mathbb{P} (|u^\mathrm{T} E v| \geq t ) \leq 2 \exp(-t^2/2). $$ \end{lemma} The bounds in Lemma \ref{lemma:bernoulli} also hold for the case where the noise is Gaussian (instead of Bernoulli). Indeed, when the entries of $E$ are iid standard normal random variables, $u^{\mathrm{T}} E v$ has the standard normal distribution. The first bound is a corollary of a general concentration result from \cite{V}. It can also be proved directly using a net argument. The second bound follows from martingale different sequence inequality \cite{McDiarmid}; see also \cite{V} for a direct proof with a more generous constant. We now verify the $(C_1,c_1,\gamma)$-concentration property for slightly more general random matrix models. We will discuss these matrix models further in Section \ref{section:app}. In the lemmas below, we consider both the case where $E$ is a real symmetric random matrix with independent entries and when $E$ is a non-symmetric random matrix with independent entries. \begin{lemma} \label{lemma:conc-sym} Let $E = (\xi_{ij})_{i,j=1}^n$ be a $n \times n$ real symmetric random matrix where $$ \{ \xi_{ij} : 1 \leq i \leq j \leq n\} $$ is a collection of independent random variables each with mean zero. Further assume $$ \sup_{1 \leq i \leq j \leq n} |\xi_{ij}| \leq K $$ with probability $1$, for some $K \geq 1$. Then for any fixed unit vectors $u,v$ and every $t > 0$ $$ \mathbb{P} (|u^\mathrm{T} E v| \geq t) \leq 2 \exp \left( \frac{-t^2}{8 K^2 } \right). $$ \end{lemma} \begin{proof} We write $$ u^\mathrm{T} E v = \sum_{1 \leq i < j \leq n} (u_i v_j + v_i u_j) \xi_{ij} + \sum_{i=1}^n u_i v_i \xi_{ii}. $$ As the right side is a sum of independent, bounded random variables, we apply Hoeffding's inequality (\cite[Theorem 2]{H}) to obtain $$ \mathbb{P} (|u^\mathrm{T} E v - \mathbb{E} u^\mathrm{T} E v| \geq t) \leq 2 \exp \left( \frac{-t^2}{8 K^2 } \right). $$ Here we used the fact that $$ \sum_{1 \leq i < j \leq n} (|u_i| |v_j| + |v_i| |u_j|)^2 + \sum_{i=1}^n |u_i|^2 |v_i|^2 \leq 4 \sum_{i,j=1}^n |u_i|^2 |v_j|^2 \leq 4 $$ because $u,v$ are unit vectors. Since each $\xi_{ij}$ has mean zero, it follows that $\mathbb{E} u^\mathrm{T} E v = 0$, and the proof is complete. \end{proof} \begin{lemma} \label{lemma:conc-nonsym} Let $E = (\xi_{ij})_{1 \leq i \leq m, 1 \leq j \leq n}$ be a $m \times n$ real random matrix where $$ \{ \xi_{ij} : 1 \leq i \leq m, 1 \leq j \leq n \} $$ is a collection of independent random variables each with mean zero. Further assume $$ \sup_{1 \leq i \leq m, 1 \leq j \leq n} |\xi_{ij}| \leq K $$ with probability $1$, for some $K \geq 1$. Then for any fixed unit vectors $u \in \mathbb{R}^m, v \in \mathbb{R}^n$, and every $t > 0$ \begin{equation} \label{eq:concprop-nonsym} \mathbb{P}( |u^\mathrm{T} E v| \geq t) \leq 2 \exp \left( \frac{-t^2}{2 K^2} \right). \end{equation} \end{lemma} The proof of Lemma \ref{lemma:conc-nonsym} is nearly identical to the proof of lemma \ref{lemma:conc-sym}. Indeed, \eqref{eq:concprop-nonsym} follows from Hoeffding's inequality since $u^\mathrm{T}E v$ can be the written as the sum of independent random variables; we omit the details. Many other models of random matrices satisfy Definition \ref{def:concentration}. If the entries of $E$ are independent and have a rapidly decaying tail, then $E$ will be $(C_1,c_1,\gamma)$-concentrated for some constants $C_1,c_1,\gamma>0$. One can achieve this by standard truncation arguments. For many arguments of this type, see for instance \cite{VW}. As an example, we present a concentration result from \cite{RV} when the entries of $E$ are iid sub-exponential random variables. \begin{lemma}[Proposition 5.16 of \cite{RV}] \label{lemma:conc-sub} Let $E = (\xi_{ij})_{1 \leq i \leq m, 1 \leq j \leq n}$ be a $m \times n$ real random matrix whose entries $\xi_{ij} $ are iid copies of a sub-exponential random variable $\xi$ with constant $K$, i.e. $\mathbb{P}(|\xi| > t) \le \exp(1-t/K)$ for all $t>0$. Assume $\xi$ has mean 0 and variance 1. Then there are constants $C_1, c_1>0$ (depending only on $K$) such that for any fixed unit vectors $u \in \mathbb{R}^m, v \in \mathbb{R}^n$ and any $t > 0$, one has \begin{equation*} \mathbb{P}( |u^\mathrm{T} E v| \geq t) \leq C_1 \exp \left( -c_1 t \right). \end{equation*} \end{lemma} Finally, let us point out that the assumption that the entries are independent is not necessary. As an example, we mention random orthogonal matrices. For another example, one can consider the elliptic ensembles; this can be verified using standard truncation and concentration results, see for instance \cite{KS, LT, McDiarmid, RV} and \cite[Chapter 5]{BS}. \section{An application: The matrix recovery problem} \label{section:app} The matrix recovery problem is the following: $A$ is a large unknown matrix. We can only observe its noisy image $A+E$, or in some cases just a small part of it. We would like to reconstruct $A$ or estimate an important parameter as accurately as possible from this observation. Consider a deterministic $m \times n$ matrix $$A = (a_{ij})_{1 \leq i \leq m, 1 \leq j \leq n.}$$ Let $Z$ be a random matrix of the same size whose entries $\{z_{ij} : 1 \leq i \leq m, 1 \leq j \leq n \}$ are independent random variables with mean zero and unit variance. For convenience, we will assume that $\| Z \| _{\infty} := \max_{i,j} |z_{ij}| \le K$, for some fixed $K >0$, with probability $1$. Suppose that we have only partial access to the noisy data $A+Z$. Each entry of this matrix is observed with probability $p$ and unobserved with probability $1-p$ for some small $p$. We will write $0$ if the entry is not observed. Given this sparse observable data matrix $B$, the task is to reconstruct $A$. The matrix completion problem is a central one in data analysis, and there is a large collection of literature focusing on the low rank case; see \cite{AM,CCS,CP,CR,CRT,CT,Cest,KMO,KMO2,Krank,KLT,MHT,NW,RVsamp} and references therein. A representative example here is the Netflix problem, where $A$ is the matrix of ratings (the rows are viewers, the columns are movie titles, and entries are ratings). In this section, we are going to use our new results to study this problem. The main novel feature here is that our analysis allows us to approximate {\it any given column (or row)} with high probability. For instance, in the Netflix problem, one can figure out the ratings of any given individual, or any given movie. In earlier algorithms we know of, the approximation was mostly done for the Frobenius norm of the whole matrix. Such a result is equivalent to saying that a {\it random} row or column is well approximated, but cannot guarantee anything about a specific row or column. Finally, let us mention that there are algorithms which can recover $A$ precisely, but these work only if $A$ satisfies certain structural assumptions \cite{CCS,CP,CR,CRT,CT}. Without loss of generality, we assume $A$ is a square $n \times n$ matrix. The rectangular case follows by applying the analysis below to the matrix $\tilde{A}$ defined in \eqref{eq:def:tildeA}. We assume that $n$ is large and asymptotic notation such as $o, O, \Omega, \Theta$ will be used under the assumption that $n \rightarrow \infty$. Let $A$ be a $n \times n$ deterministic matrix with rank $r$ where $\sigma_1 \geq \cdots \geq \sigma_r > 0$ are the singular values with corresponding singular vectors $u_1, \ldots, u_r$. Let $\chi_{ij}$ be iid indicator random variables with $\mathbb{P} (\chi_{ij}=1)=p$. The entries of the sparse matrix $B$ can be written as \begin{equation*} b_{ij} = (a_{ij } +z_{ij} ) \chi_{ij} = p a_{ij} + a_{ij} (\chi_{ij} -p) + z_{ij} \chi_{ij} = pa_{ij} + f_{ij}, \end{equation*} where \begin{equation*} f_{ij} := a_{ij} (\chi_{ij} -p) + z_{ij} \chi_{ij} . \end{equation*} It is clear that the $f_{ij}$ are independent random variables with mean 0 and variance $\sigma_{ij}^2 = a_{ij}^2 p(1-p) + p $. This way, we can write $\frac{1}{p} B$ in the form $A + E$, where $E$ is the random matrix with independent entries $e_{ij} := p^{-1} f_{ij}$. We assume $p \le 1/2$; in fact, our result works for $p$ being a negative power of $n$. Let $1 \le j \le r$ and consider the subspace $U$ spanned by $u_1, \dots, u_j$ and $V$ spanned by $v_1, \dots, v_j$, where $u_i$ (alternatively $v_i)$ is the $i$-th singular vector of $A$ (alternatively $B$). Fix any $1 \le m \le n$ and consider the $m$-th columns of $A$ and $A+E$. Denote them by $x$ and $\tilde x $, respectively. We have \begin{equation*} \| x- P_{V} \tilde x\| \le \| x - P_U x\| + \| P_U x - P_U \tilde x\| + \| P_U \tilde x - P_V \tilde x \|. \end{equation*} Notice that $P_V \tilde x $ is efficiently computable given $B$ and $p$. (In fact, we can estimate $p$ very well by the density of $B$, so we don't even need to know $p$.) In the remaining part of the analysis, we will estimate the three error terms on the right-hand side. We will make use of the following lemma, which is a variant of \cite[Lemma 2.2]{TVdet}; see also \cite{VW} where results of this type are discussed in depth. \begin{lemma} \label{lemma:projection} Let $X$ be a random vector in $\mathbb{R}^n$ whose coordinates $x_i, 1\le i \le n$ are independent random variables with mean 0, variance at most $\sigma^2$, and are bounded in absolute value by $1$. Let $H$ be a fixed subspace of dimension $d$ and $P_H (X)$ be the projection of $X$ onto $H$. Then \begin{equation} \label{eqn:distance} \mathbb{P} \left( \| P_H (X) \| \ge \sigma d^{1/2} + t \right) \le C \exp (-c t^2 ) ,\end{equation} where $c, C>0$ are absolute constants. \end{lemma} The first term $\| x -P_U x \|$ is bounded from above by $\sigma_{j+1}$. The second term has the form $\| P_U X \|$, where $X:=x -\tilde x$ is the random vector with independent entries, which is the $m$-th column of $E$. Notice that entries of $X$ are bounded (in absolute value) by $\alpha : = p^{-1} (\|x\| _{\infty}+ K)$ with probability $1$. Applying Lemma \ref{lemma:projection} (with the proper normalization), we obtain \begin{equation} \label{recovery3} \mathbb{P} \left( \| P_U X \| \geq j^{1/2} \sqrt{ \frac{ \|x\|_{\infty}^2 + 1}{p} }+ t \right) \le C \exp( -c t^2 \alpha^{-2} ) \end{equation} since $\sigma_{im}^2 \leq p^{-1} ( \|x\|_{\infty}^2 + 1)$. By setting $t := c^{-1/2} \alpha \lambda $, \eqref{recovery3} implies that, for any $\lambda > 0$, \begin{equation*} \| P_U X \| \le j^{1/2} \sqrt{ \frac{ \|x\|_{\infty}^2 + 1}{p} }+ c^{-1/2} \lambda \alpha \end{equation*} with probability at least $1- C \exp(-\lambda^2 ) $. To bound $\| P_U \tilde x - P_V \tilde x \|$, we appeal to Theorem \ref{thm:subspace}. Assume for a moment that $E$ is $(C_1, c_1, \gamma)$-concentrated for some constants $C_1, c_1, \gamma > 0$. Let $\delta_j := \sigma_j - \sigma_{j+1}$. Then it follows that, for any $\lambda > 0$, \begin{equation*} \| P_U - P_V\| \le C \sqrt{j} \left(\frac{ \lambda^{2/\gamma} r^{1/\gamma} }{ \delta_{j} } + \frac{\|E \| }{\sigma_j} +\frac{\| E\|^2 }{ \sigma_j \delta_j } \right), \end{equation*} with probability at least $$ 1 - 6C_1 9^j \exp \left( -c_1 \frac{\delta_j^\gamma}{8^\gamma} \right) - 2C_1 9^{2r} \exp \left( -c_1 r \frac{\lambda^2}{4^\gamma} \right), $$ where $C$ is an absolute constant. Since $$ \| P_U \tilde{x} - P_V \tilde{x} \| \leq \|P_U - P_{V} \| \|\tilde{x} \|, $$ it remains to bound $\| \tilde{x} \|$. We first note that $\|\tilde{x}\| \leq \|x\| + \|X \|$. By Talagrand's inequality (see \cite{Tconc} or \cite[Theorem 2.1.13]{Tbook}) , we have $$ \mathbb{P} \left( \|X \| \geq \mathbb{E} \|X\| + t \right) \leq C \exp(-c t^2 \alpha^{-2}). $$ In addition, $$ \mathbb{E} \|X\|^2 = \frac{1}{p^2} \sum_{i=1}^n \sigma_{im}^2 \leq \frac{1}{p} \left( \|x\|^2 + n \right). $$ Thus, we conclude that $$ \|X\| \leq \sqrt{ \frac{\|x\|^2 + n }{p} } + c^{-1/2} \lambda \alpha $$ with probability at least $1 - C \exp(-\lambda^2)$. Putting the bounds together, we obtain Theorem \ref{theorem:recovery} below. \begin{theorem} \label{theorem:recovery} Assume that $A$ has rank $r$ and $\| Z\|_{\infty} \le K $ with probability $1$. Assume that $E$ is $(C_1, c_1, \gamma)$-concentrated for a trio of constants $C_1, c_1, \gamma >0$. Let $m$ be an arbitrary index between $1$ and $n$, and let $x$ and $\tilde x$ be the $m$-th columns of $A$ and $\frac{1}{p} B$. Let $1 \leq j \leq r$ be an integer, and let $V$ be the subspace spanned by the first $j$ singular vectors of $B$. Let $\sigma_1 \ge \dots \ge \sigma_r > 0$ be the singular values of $A$. Assume $\delta_j := \sigma_j - \sigma_{j+1}$. Then, for any $\lambda >0$, \begin{equation*} \| x - P_V (\tilde x) \| \le \sigma_{j+1} + j^{1/2} \sqrt{ \frac{ \|x\|_{\infty}^2 + 1}{p} } + \mu \left( \sqrt{ \frac{ \|x\|^2 + n}{p} } + C \lambda \alpha \right) + C \lambda \alpha, \end{equation*} with probability at least $$ 1 - C \exp(-\lambda^2) - 6C_1 9^j \exp \left( -c_1 \frac{\delta_j^\gamma}{8^\gamma} \right) - 2C_1 9^{2r} \exp \left( -c_1 r \frac{\lambda^2}{4^\gamma} \right), $$ where $$ \alpha := p^{-1} (\| x\| _{\infty} + K) \quad \text{and}\quad \mu:= C \sqrt{j} \left(\frac{\lambda^{2/\gamma} r^{1/\gamma} }{ \delta_{j} } + \frac{\|E \| }{\sigma_j} +\frac{\| E\|^2 }{ \sigma_j \delta_j } \right), $$ and $C$ is an absolute constant. \end{theorem} As this theorem is a bit technical, let us consider a special, simpler case. Assume that all entries of $A$ are of order $\Theta(1)$ and $p=\Theta(1)$. Thus, any column $x$ has length $ \Theta (n^{1/2})$. Assume furthermore that $j=r=\Theta(1)$ and $\sigma_r = \Omega(n^{1/2+\varepsilon})$ for some $\varepsilon > 0$. Then our analysis yields \begin{corollary} There exists $c_0 > 0$ (depending only on $\varepsilon$) such that, for any given column $x$, $$ \| x - P_V (\tilde x) \| = O( n^{-c_0} \|x\|) $$ with probability $1-o(1)$. \end{corollary} \subsection*{Acknowledgements} The authors would like to thank Nicholas Cook and David Renfrew for useful comments.

https://arxiv.org/abs/1412.2851

Minimum Local Distance Density Estimation

We present a local density estimator based on first order statistics. To estimate the density at a point, $x$, the original sample is divided into subsets and the average minimum sample distance to $x$ over all such subsets is used to define the density estimate at $x$. The tuning parameter is thus the number of subsets instead of the typical bandwidth of kernel or histogram-based density estimators. The proposed method is similar to nearest-neighbor density estimators but it provides smoother estimates. We derive the asymptotic distribution of this minimum sample distance statistic to study globally optimal values for the number and size of the subsets. Simulations are used to illustrate and compare the convergence properties of the estimator. The results show that the method provides good estimates of a wide variety of densities without changes of the tuning parameter, and that it offers competitive convergence performance.

\section{Introduction} Nonparametric density estimation is a classic problem that continues to play an important role in applied statistics and data analysis. More recently, it has also become a topic of much interest in computational mathematics, especially in the uncertainty quantification community where one is interested in, for example, densities of a large number of coefficients of a random function in terms of a fixed set of deterministic functions (e.g., truncated Karhunen-Lo\`eve expansions). The method we present here was motivated by such applications. Among the most popular techniques for density estimation are the histogram \cite{scott1979optimal,scottmv}, kernel~\cite{parzen1962estimation,scottmv,wand} and orthogonal series~\cite{efrom,silverman1986density} estimators. For the one-dimensional case, histogram methods remain in widespread use due to their simplicity and intuitive nature, but kernel density estimation has emerged as a method of choice thanks, in part, to recent adaptive bandwidth-selection methods providing fast and accurate results~\cite{botev2010kernel}. However, these kernel density estimators can fail to converge in some cases (e.g., recovering a Cauchy density with Gaussian kernels)~\cite{buch2005kernel} and can be computationally expensive with large samples ($\orderof{N^2}$, for a sample size $N$). Note that histogram estimators are typically implemented using equal-sized bins, and nearest-neighbor density estimators can be roughly thought as histograms whose bins adapt to the local density of the data. More precisely, let $X_1,\ldots, X_N$ be iid variables from a distribution with density, $f$, and let $X_{(1)},\ldots,X_{(N)}$ be the corresponding order statistics. For any $x$, define $Y_i = |X_i-x|$ and $D_j(x) = Y_{(j)}$. The $k$-nearest-neighbor estimate of $f$ is defined as (see \cite{silverman1986density} for an overview): \( \widehat{f}_N(x) = (\,C_N/N\,)/[\,2 D_k(x)\,], \) where $C_N$ is a constant that may depend on the sample size. We may think of $2D_k(x)$ as the width of the bin around $x$. The value of $C_N$ is often chosen as $C_N\approx N^{1/2}$ but some effort has been directed towards its optimal selection~\cite{fukunaga1973optimization,hall2008choice,li1984consistency}, with some recent work involving the use of order statistics~\cite{kung2012optimal}. One of the disadvantages of nearest-neighbor estimators is that their derivative has discontinuities at the points $(X_{(j)}+X_{(j+k)})/2$, which is caused by the discontinuities of the derivative of the function $D_k(x)$ at these points. This is clear in Figure \ref{fig:smoothdist}, which shows plots of $D_k(x)$ for a sample of size $N=125$ from a Cauchy$(0,1)$ distribution with $k=1$ and \edit{$k=round(\sqrt{N})$}. One way to obtain smoother densities is using a combination of kernel and nearest-neighbor density estimation where the nearest-neighbors technique is used to choose the kernel bandwidth~\cite{silverman1986density}. We introduce an alternative averaging method that improves smoothness and can still be used to obtain local density estimates. The main idea of this paper may be summarized as follows: Instead of using the $k$th nearest-neighbor to provide an estimate of the density at a point, $x$, we use a subset-average of first order statistics of $|X_i-x|$. So, the original sample of size $N$ is split into $m$ subsets of size $s$ each; this decomposition into subsets allows the control of the asymptotic mean squared error (MSE) of the density estimate. Thus, the problem of bandwidth selection is transformed into that of choosing an optimal number of subsets. This density estimator is naturally parellelizable with complexity $\orderof{N^{{1}/{3}}}$ for parallel systems. The rest of this article is organized as follows. In Section~\ref{sec:theory} we develop the theory that underlies the estimator and describe asymptotic results. In Sections~\ref{sec:algo} and \ref{sec:numexp} we describe the actual estimator and study its performance using numerical experiments. A variety of densities are used to reveal the strengths and weaknesses of the estimator. We provide concluding remarks and generalizations in Section~\ref{sec:conc}. Proofs and other auxiliary results are collected in Appendix \ref{sec:proofs}. \edit{From here on, when we refer to the size of a sample set being the power of the total number of samples, we assume that it represents a rounded value, for example, $k = \sqrt{N} \Rightarrow k = round(\sqrt{N})$.} \section{Theoretical framework \label{sec:theory}} Let $X_1,\ldots,X_N$ be iid random variables from a distribution with invertible CDF, $F$, and PDF $f$. Our goal is to estimate the value of $f$ at a point, $x_*$, where $f(x_*)>0$, and where $f$ is either continuous or has a jump discontinuity. The non-negative random variables $Y_i = |X_i-x_*|$ are iid with PDF: \( g(y) = f(y+x_*) + f(x_*-y). \) In particular, $f(x_*)=g(0)/2$. Thus, an estimate of $g(0)$ leads to an estimate of $f(x_*)$. Furthermore, $g$ is more regular than $f$ in a sense described by the following lemma (its proof and those of the other results in this section are collected in Appendix \ref{sec:proofs}). \begin{lemma} \label{lemma:gprop} Let $f$ and $g$ be as defined above. Then: \vspace{-.3cm} \begin{itemize} \item[(i)] If $f$ has left and right limits at $x_*$ (i.e., it is either continuous or has a jump discontinuity at $x_*$), then $g$ is continuous at zero. \vspace{-.2cm} \item[(ii)] If $f$ has left and right derivatives at $x_*$, then $g$ has a right derivative at zero. Furthermore, if $f$ is differentiable at $x_*$, then $g'(0)=0$. \end{itemize} \end{lemma} The original question is thus reduced to the following problem: Let $X_1,\ldots,X_N$ be iid non-negative random variables from a distribution with invertible CDF, $G$, and PDF $g$. The goal is to estimate $g(0)>0$ assuming that $g$ is right continuous at zero. The continuity at zero comes from Lemma \ref{lemma:gprop}(i). For some asymptotic results we also assume that $g$ is right-differentiable with $g'(0)=0$. The zero derivative is justified by Lemma \ref{lemma:gprop}(ii). We estimate $g(0)$ using a subset-average of first order statistics. There is a natural connection between density estimation and first order statistics: If ${X_{(1),N}}$ is the first order statistic of $X_1,\ldots,X_N$, then (under regularity conditions) $\mathbb{E} {X_{(1),N}} \sim Q(1/(N+1))$ as $N\to \infty$, where $Q = G^{-1}$ is the quantile function, and therefore $(N+1)\,\mathbb{E} {X_{(1),N}} \to 1/g(0)$. This shows that one should be able to estimate $g(0)$ provided $N$ is large and we have a consistent estimate of $\mathbb{E} {X_{(1),N}} $. In the next section we provide conditions for the limit to be valid and derive a similar limit for the second moment of ${X_{(1),N}}$; we then define the estimator and study its asymptotics. \subsection{Limits of first order statistics} We start by finding a representation of the first two moments of the first-order statistic in terms of functions that allow us to determine the limits of the moments as $N\to \infty$. \begin{lemma}\label{lemma:ordmom} Let $X_1,\ldots,X_N$ be iid non-negative random variables with PDF $g$, invertible CDF $G$ and quantile function $Q$. Assume that $g(0)>0$, and define the sequence of functions \( \delta_N(z) = (N+1)(1-z)^N \) on $z\in [0,1]$, $N\in \mathbb{N}$. Then: \vspace{-.3cm} \begin{itemize} \item[(i)] \begin{eqnarray} (N+1)\,\mathbb{E} {X_{(1),N}} &=& \int_0^1 \frac{\delta_N(z)}{g(Q(z))}\,dz=\int_0^1 Q'(z)\, \delta_N(z)\,dz \label{eq:EXone}\\ &=& \frac{1}{g(0)} + \frac{1}{(N+2)}\int_0^1 Q''(z)\,\delta_{N+1}(z)\,dz.\label{eq:EXonev1} \end{eqnarray} Furthermore, if $g$ is twice differentiable with $g'(0)=0$, then \begin{equation} \label{eq:EXonev2} (N+1)\,\mathbb{E} {X_{(1),N}} = \frac{1}{g(0)} + \frac{1}{(N+2)(N+3)}\int_0^1 Q'''(z)\,\delta_{N+2}(z)\,dz. \end{equation} \vspace{-.7cm} \item[(ii)] If $g$ is differentiable a.e., then \begin{equation}\label{eq:EXone2} (N+1)^2\,\mathbb{E}[\,{X_{(1),N}}^2\,] = \left(\frac{N+1}{N+2}\right)\int_0^1 (Q^2(z))''\,\delta_{N+1}(z)\,dz. \end{equation} \end{itemize} \end{lemma} We use the following result to evaluate the limits of the moments as $N\to \infty$. \begin{prop}\label{prop:convlim} Let $H$ be a function defined on $[0,1]$ that is continuous at zero, and assume there is an integer $m>0$ and a constant $C>0$ such that \begin{equation}\label{eq:growth} |H(x)| \leq {C}/{(1-x)^m} \end{equation} a.e. on $[0,1]$. Then,\, \( \lim_{N\to \infty} \int_0^1 H(x)\,\delta_N(x)\,dx = H(0). \) \end{prop} This proposition allows us to compute the limits of \eqref{eq:EXone}-\eqref{eq:EXone2} provided the quantile functions satisfy appropriate regularity conditions. When a function $H$ satisfies \eqref{eq:growth}, we shall say that $H$ satisfies a tail condition for some $C>0$ and integer $m>0$. The following corollary follows from Lemma \ref{lemma:ordmom} and Proposition \ref{prop:convlim}: \begin{corollary}\label{cor:momlims} Let $X_1,\ldots,X_N$ be iid non-negative random variables with PDF $g$, invertible CDF $G$ and quantile function $Q$. Assume that $g(0)>0$. Then: \vspace{-.2cm} \item[(i)] If $g$ is continuous at zero and $Q'$ satisfies a tail condition, then \begin{equation}\label{eq:limfirst} \lim_{N\to\infty} (N+1)\,\mathbb{E} {X_{(1),N}} = Q'(0) = {1}/{g(0)}. \end{equation} If $g$ is differentiable and $Q''$ satisfies a tail condition, then \begin{equation}\label{eq:limfirstv1} (N+1)\,\mathbb{E} {X_{(1),N}} = {1}/{g(0)} + \orderof{1/N}. \end{equation} \item[(ii)] If $g$ is twice differentiable with $g'(0)=0$, $g''$ is continuous at zero and $Q'''$ satisfies a tail condition, then \begin{equation}\label{eq:limfirstv2} (N+1)\,\mathbb{E} {X_{(1),N}} = {1}/{g(0)} + \orderof{1/N^2}. \end{equation} \item[(iii)] If $g$ is differentiable a.e., $g'$ and $g$ are continuous at zero, and $Q''$ satisfies a tail condition, then \begin{eqnarray} \lim_{N\to \infty} (N+1)^2\,\mathbb{E}[{X_{(1),N}}^2] &=& 2\,Q'(0)^2 = {2}/{g(0)^2}\label{eq:limsec}\\ \lim_{N\to\infty} {\mathbb{V}{\rm ar}}\left[\,(N+1)\,{X_{(1),N}}\,\right] & = & {1}/{g(0)^2}.\label{eq:limvar} \end{eqnarray} \end{corollary} We now provide examples of distributions that satisfy the hypotheses of Corollary \ref{cor:momlims}. \edit{For these examples, we temporarily return to the notations $X_i$ (iid random variables) and $Y_i= |X_i-x_*|$ used before Lemma~\ref{lemma:gprop}}. \begin{example}{\rm Let $X_1,\ldots,X_N$ be iid with exponential distribution $\mathcal{E}(\lambda)$ and fix $x_*>0$. The PDF, CDF and quantile function of $Y_i$ are, respectively, \begin{eqnarray*} g(y) &=& 2\lambda\,e^{-\lambda x_*}\cosh(\lambda y)\,I_{y\leq x_*} + \lambda\,e^{-\lambda(x_*+y)}\,I_{y> x_*}\\ G(y) &=& 2e^{-\lambda x_*}\sinh(\lambda y) \,I_{y\leq x_*} + (1-e^{-\lambda(x_*+y)}) \,I_{y> x_*}\\ Q(z) &=& \lambda^{-1}\mathrm{arcsinh} (ze^{\lambda x_*}/2)\,I_{z\leq z_*}- [\,x_* + \lambda^{-1}\,\log(1-z)\,]\,I_{z>z_*} \end{eqnarray*} for $y\geq 0$, $z\in [0,1)$ and $z_*=1-e^{-2\lambda x_*}$. As expected, $g'(0)=0$. In addition, $Q$ and its derivatives are continuous at zero. Furthermore, since $|\log(1-z)|\leq z/(1-z)$ on $(0,1)$, we see that $Q$ and its derivatives satisfy tail conditions. } \end{example} \begin{example}{\rm Let $X_1,\ldots,X_N$ be iid with Cauchy distribution and fix $x_*\in{\mathbb{R}}$. The PDF and CDF of $Y_i$ are: \begin{eqnarray*} g(y) &=& \frac{1}{\pi[1+(y+x_*)^2]} + \frac{1}{\pi[1+(x_*-y)^2]}\\ G(y) &=& \arctan(y+x_*)/\pi - \arctan(x_*-y)/\pi. \end{eqnarray*} Again, $g'(0)=0$. To verify the conditions on the quantile function, $Q$, note that \[ Q(z) = -\cot(\pi z) + \cot(\pi z)\sqrt{1 + (1+x_*^2)\tan^2(\pi z)}, \] in a neighborhood of zero, while for $z$ in a neighborhood of 1, $Q$ is given by \[ Q(z) = -\cot(\pi z) - \cot(\pi z)\sqrt{1 + (1+x_*^2)\tan^2(\pi z)}. \] Since $Q(z)\to 0$ as $z\to 0^+$ and $g$ is smooth, it follows that $Q$ and its derivatives are continuous at zero. It is easy to see that the tail conditions for $Q'$, $Q'''$ and $(Q^2)''$ are determined by the tail condition of $\csc(\pi z)$, which in turn follows from the inequality $|\csc(\pi z)|\leq 1/[\,\pi z(1-z)]$ on $(0,1)$. } \end{example} It is also easy to check that the Gaussian and beta distributions satisfy appropriate tail conditions for Corollary \ref{cor:momlims}. \subsection{Estimators and their properties} Let $X_1,\ldots,X_N$ be iid non-negative random variables whose PDF $g$, CDF $G$ and quantile function $Q$ satisfy appropriate regularity conditions for Corollary \ref{cor:momlims}. We randomly split the sample into $m_N$ independent subsets of size $s_N$. Both sequences, $(m_N)$ and $(s_N)$, tend to infinity as $N\to \infty$ and satisfy $m_N s_N = N$. Let $X^{(1)}_{(1),s_N},\ldots,X^{(m_N)}_{(1),s_N}$ be the first-order statistics for each of the $m_N$ subsets, and let ${\overline{X}_{m_N,s_N}}$ be their average, \begin{equation}\label{eq:meanX} {\overline{X}_{m_N,s_N}} = \frac{1}{m_N}\sum_{k=1}^{m_N} X^{(k)}_{(1),s_N}. \end{equation} The estimators of $1/g(0)$ and $g(0)$ are defined, respectively, as: \begin{equation}\label{eq:estdefs} \widehat{f^{-1}(0)}_{N} = (s_N+1){\overline{X}_{m_N,s_N}},\quad \widehat{f(0)}_{N} = 1/\widehat{f^{-1}(0)}_{N}. \end{equation} \begin{prop}\label{prop:mselim} Let $N$, $m_N$ and $s_N$ be as defined above. Then: \vspace{-.1cm} \begin{itemize} \item[(i)] If $g$ is differentiable a.e., $g'$ and $g$ are continuous at zero, and $Q''$ satisfies a tail condition, then \begin{equation}\label{eq:mselim} \lim_{N\to \infty} {\rm MSE}(\,\widehat{f^{-1}(0)}_{N}\,) = 0, \end{equation} and therefore \( \widehat{f(0)}_{N} \stackrel{{P}}{\longrightarrow} g(0) \) as $N\to \infty$. \item[(ii)] Let $g$ be twice differentiable with $g'(0)=0$, $g''$ be continuous at zero, and let $Q'''$ satisfy a tail condition. If $\sqrt{m_N}/s_N\to \infty$ and $\sqrt{m_N}/s_N^2\to 0$ as $N\to \infty$, then \begin{equation}\label{eq:fidistlim} \sqrt{m_N}\left(\,\widehat{f^{-1}(0)}_{N}-1/g(0)\,\right) \stackrel{{\cal L}}{\longrightarrow} N(0,1/g(0)^2), \end{equation} which leads to \begin{equation}\label{eq:fdistlim} \sqrt{m_N}\left(\,\widehat{f(0)}_{N}-g(0)\,\right) \stackrel{{\cal L}}{\longrightarrow} N(0,g(0)^2). \end{equation} Furthermore, ${\rm MSE}(\widehat{f^{-1}(0)}_{N})$ and ${\rm MSE}(\widehat{f(0)}_{N})$ are $\orderof{1/m_N}$. In particular, \eqref{eq:fidistlim} and \eqref{eq:fdistlim} are satisfied when $s_N = N^\alpha$ and $m_N = N^{1-\alpha}$ for some $\alpha\in (1/5,1/3)$. This leads to the MSE optimal rate $\orderof{N^{-4/5-\varepsilon}}$ for any $\varepsilon>0$. \end{itemize} \end{prop} By (ii), we need a balance between the sample size, $s_N$, and the number of samples, $m_N$: $m_N$ should grow faster than $s_N$ but not much faster. For comparison, the optimal rate of the MSE is $\orderof{N^{-2/3}}$ for the smoothed histogram, and $\orderof{N^{-4/5}}$ for the kernel density estimator \cite{dasgupta}. \subsubsection*{Distance function} We return to the original sample $X_1,\ldots,X_N$ from a density $f$ before the transformation to $Y_1 = |X_1-x|,\ldots,Y_N = |X_N-x|$. The sample is split into $m_N$ subsets. Let $D_1(x;m)$ be the distance from $x$ to its nearest-neighbor in the $m$th subset. The mean ${\overline{X}_{m_N,s_N}}$ in \eqref{eq:meanX} is the average of $D_1(x;m)$ over all the subsets; we call this average the distance function, $D_{\rm MLD}$, of the MLD density estimator. That is, \[ D_{\rm MLD}(x) = {\overline{X}_{m_N,s_N}}=\frac{1}{m_N}\sum_{m=1}^{m_N} D_1(x;m). \] The estimators in \eqref{eq:estdefs} can then be written in terms of $D_{\rm MLD}(x)$. This distance function tends to be smoother than the usual distance function used by $k$-nearest-neighbor density estimators. For example, Figure \ref{fig:smoothdist} shows the different distance functions $D_{\rm MLD}(x)$, $D_1(x)$ and $D_k(x)$ (the latter as defined in the introduction) for a sample of $N=125$ variables from a Cauchy$(0,1)$. Note that $D_{\rm MLD}$ is an average of first-order statistics for samples of size $s_N$, while $D_1$ is a first-order statistic for a samples of size $N$, so $D_{\rm MLD}>D_1$. On the other hand, $D_{\sqrt{N}}$ is a $N^{1/2}$th-order statistic based on a sample of size $N$; hence the order $D_{\rm MLD}>D_{\sqrt{N}}>D_1$. \begin{figure}[!h] \begin{center} \includegraphics[keepaspectratio,width=0.5\textwidth]{smoothdist.jpg} \caption{Distance function $D_k(x)$ for $k$-nearest-neighbor (for $k=1$ and $k=11\approx \sqrt{N}$) and the distance function $D_{\rm MLD}(x)$ \edit{(with $m_N = N^{\frac{2}{3}} = 25$ subsets)} for \edit{125 samples taken from a Cauchy(0,1) distribution.}\label{fig:smoothdist}} \end{center} \end{figure} \section{Minimum local distance density estimator \label{sec:algo}} We now describe the local distance density estimator (MLD-DE). The inputs are: a sample, a set of points where the density is to be estimated and the parameter $\alpha$ whose default is set to $\alpha=1/3$. The basic steps to obtain the density estimate at a point $x$ are: (1) Start with a sample of $N$ iid variables from the unknown density, $f$; (2) Randomly split the sample into $m_N = N^{1-\alpha}$ disjoint subsets of size $s_N = N^{\alpha}$ each; (3) Find the nearest sample distance to $x$ in each subset; (4) Compute the density estimate by inverting the average nearest distance across the subsets and scaling it (see~Eq.\eqref{eq:estdefs}). This is summarized in Algorithm \ref{alg:MLD}. \begin{algorithm}[!h] \caption{\label{alg:MLD} Returns density estimates at the points \edit{of evaluation $\{x_{l}\}_{l=1}^{M}$} given the sample $X_1,\ldots,X_N$ from the unknown density $f$.} \label{alg:nnj} \begin{algorithmic}[1] \State $m_N \leftarrow$ round($N^{1-a}$) \State $s_N \leftarrow$ round(${N}/{m_N}$) \State Create an $s_N \times m_N$ matrix $M_{ij}$ with the $m_N$ subsets with $s_N$ variables each \State Create a vector $\widehat{f}=(\widehat{f}_\ell)$ to hold the density estimates at the points $\{x_{l}\}_{l=1}^{M}$ \For{$l = 1 \to M$} \For{$k = 1 \to s_N$} \State Find the nearest distance $d_{lk}$ to the current point $x_\ell$ within the $k$th subset \EndFor \State Compute the subset average of distances to $x_\ell$: $d_\ell = (1/m_N) \sum\limits_{k=1}^{m_N}d_{\ell k}$ \State Compute the density estimate at $x_l$: $\widehat{f}_\ell = {1}/{2 d_\ell}$ \EndFor \\ \Return $\widehat{f}$ \end{algorithmic} \end{algorithm} Note that for each of the $M$ points where the density is to be estimated, the algorithm loops over $N^{1-\alpha}$ subsets, and within each it does a nearest-neighbor search over $N^{\alpha}$ points. The computational complexity is therefore $\orderof{M N^{1-\alpha} N^{\alpha}} = \orderof{MN}$, which is of the same order as the $\orderof{N^2}$ complexity of kernel density estimators~\cite{raykar2010fast} when $M \sim N$. However, MLD-DE displays multiple levels of parallelism. The first level is the highly parallelizable evaluation of the density at the $M$ specified points. The second level arises from the the nearest-neighbor distances that can be computed independently in each subset. Thus, for parallel systems the effective computational complexity of the algorithm is $\orderof{MN^{\alpha}}$, which is the same as that of histogram methods if $\alpha = {1}/{3}$. \section{Numerical examples \label{sec:numexp}} An extensive suite of numerical experiments was used to test the MLD-DE method. We now summarize the results to show that they are consistent with the theory derived in Section~\ref{sec:theory}, and illustrate some salient features of the estimator. We also compare MLD-DE to the adaptive kernel density estimator (KDE) introduced by Botev et al.~\cite{botev2010kernel} and to the histogram method based on Scott's normal reference rule~\cite{scott1979optimal}. We first discuss experiments for density estimation at a fixed point and show the effects of changing the number of subsets for a fixed sample size. We then estimate the integrated mean square error for various densities, and compare the convergence of MLD-DE to that of other density estimators. Next, we present numerical experiments that show the spatial variation of the bias and variance of MLD-DE, and relate them to the theory derived in Section~\ref{sec:theory}. Finally, we check the impact of changing the tuning parameter $\alpha$ (see Proposition~\ref{prop:mselim}). \subsection{Pointwise estimation of a density} We use MLD-DE to estimate values of the beta$(1,4)$ and $N(0,1)$ densities at a single point and analyze its convergence performance. Starting with a sample size $N=100$, $N$ was progressively increased to three million. For each $N$, 1000 trials were performed to estimate the MSE of the density estimate. The parameter $\alpha$ was also changed; it was set to ${1}/{3}$ for one set of experiments anticipating a bias of $\orderof{{1}/{N}}$, and to ${1}/{5}$ for another set, anticipating a bias of $\orderof{{1}/{N^2}}$. The results are shown in Figure~\ref{fig:pointwise_conv}. \begin{figure}[!h] \begin{center} \includegraphics[keepaspectratio,width=0.45\textwidth]{beta_pointwise.jpg}\hspace{.2cm} \includegraphics[keepaspectratio,width=0.45\textwidth]{gaussian_pointwise.jpg} \caption{Convergence plots of the density estimates at $x=1/2$ for the distribution beta$(1,4)$ (left), and at $x=1$ for $N(0,1)$ (right).\label{fig:pointwise_conv}} \end{center} \end{figure} We see the contrasting convergence behavior for the beta$(1,4)$ and $N(0,1)$ distributions. For the former, the convergence is faster when $\alpha = {1}/{3}$, while for the Gaussian it is faster with $\alpha= {1}/{5}$. We recall from Section \ref{sec:theory} that the asymptotic bias of the density estimate at a point is $\orderof{{1}/{N^2}}$. However, reaching the asymptotic regime depends on the convergence of $\int_{0}^{1} Q''(z) \, \delta_N(z) \, dz $ to zero, which can be quite slow, depending on the behavior of the density at the chosen point. Hence, the effective bias in simulations can be $\orderof{{1}/{N}}$. The numerical experiments thus indicate that the quantile function derivative of the Gaussian decays to zero much faster than that of the beta distribution, and hence the optimal value of $\alpha$ for $N(0,1)$ is ${1}/{5}$, while that for beta$(1,4)$ is ${1}/{3}$. However, in either case the order of the decay in the figure is close to $N^{-3/4}$. \subsection{$L^2$-convergence} We now summarize simulation results regarding the $L^2$-error (i.e., integrated MSE) of estimates of a beta$(1,4)$, a Gaussian mixture and the Cauchy$(0,1)$ density. The Gaussian mixture used is (see ~\cite{wasserman2006all}): \( \label{eq:gm_pdf} 0.5\, N(0,1) + 0.1 \sum_{i=0}^{4} N(\,i/2-1,\,1/100^2\,). \) For comparison, these densities were estimated using MLD-DE, the Scott's rule-based histogram, and the adaptive KDE proposed by~\cite{botev2010kernel}. Both, the Scott's rule-based histograms and KDE method fail to recover the Cauchy$(0,1)$ density. For the histogram method, this limitation was overcome using an interquartile range (IQR) based approach for the Cauchy density that uses a bandwidth, $h_N$, based on the Freedman-Diaconis rule~\cite{freedman1981histogram}: \begin{equation} \label{eq:iqr_bwith} h_N =2\, N^{-1/3}\,\mathrm{IQR}_N, \end{equation} where IQR$_N$ is the sample interquartile range for a sample of size $N$. For the KDE, there is no clear method that enables us to estimate a Cauchy density, thus KDE was only used for the Gaussian mixture and beta densities. \begin{figure}[!h] \begin{center} \subfigure[Beta(1,4) distribution] { \includegraphics[width=0.35\textwidth]{beta_comparison.pdf} \label{fig:beta_comparison} } \subfigure[Gaussian Mixture] { \includegraphics[width=0.35\textwidth]{gm_comparison.pdf} \label{fig:gm_comparison} }\\ \subfigure[Gaussian Mixture: optimal $\alpha$] { \includegraphics[width=0.35\textwidth]{gm_comparison_opt.pdf} \label{fig:gm_comparison_optimal_a} } \subfigure[Cauchy(0,1) distribution] { \includegraphics[width=0.35\textwidth]{cauchy_comparison_iqr.pdf} \label{fig:cauchy_comparison} } \caption{Density estimates using MLD-DE, KDE and histogram approaches for the beta$(1,4)$, Gaussian mixture and Cauchy$(0,1)$ distributions.} \label{fig:comparison} \end{center} \end{figure} For the MLD-DE and histogram-based estimators, estimates were obtained for 256 points in specified intervals. The interval used for each distribution is shown in the figures as the range over which the densities are plotted. Once the pointwise density estimates were calculated, interpolated density estimates were obtained using nearest-neighbor interpolation. For example, Figure~\ref{fig:comparison} shows density estimates from a single sample using $\alpha=1/3$ for the beta (Figure~\ref{fig:beta_comparison}), Gaussian Mixture (Figure~\ref{fig:gm_comparison}) and Cauchy (Figure~\ref{fig:cauchy_comparison}), and with an optimal $\alpha$ for the Gaussian mixture (Figure~\ref{fig:gm_comparison_optimal_a}) obtained by simulation. The sample size was again increased progressively starting with $N=125$ up to a maximum sample size $N=8000$. The MSE was calculated at every point of estimation, and then numerically integrated to obtain an estimate of the $L^2$-error. A total of 1000 trials were performed at each sample size to obtain the expected $L^2$-error for such sample size. Figure~\ref{fig:L2_conv} shows the convergence plots obtained for the three densities using the various density estimation methods (the error bars are the size of the plotting symbols). We see that the performance of MLD-DE is comparable to that of the histogram method for the beta and Gaussian mixture densities, and KDE performs better with both these densities. \begin{figure}[!h] \begin{center} \subfigure[Beta(1,4) distribution] { \includegraphics[width=0.35\textwidth]{beta_L2_zero.pdf} \label{fig:beta_L2}} \subfigure[Gaussian Mixture] { \includegraphics[width=0.35\textwidth]{gmixture_L2_zero.pdf} \label{fig:gmixture_L2}} \subfigure[Gaussian Mixture: optimal $\alpha$] { \includegraphics[width=0.35\textwidth]{gmixture_L2_zero_opt_a.pdf} \label{fig:gmixture_L2_opt_a}} \subfigure[Cauchy(0,1) density] { \includegraphics[width=0.35\textwidth]{cauchy_L2_zero.pdf} \label{fig:cauchy_L2}} \caption{$L^{2}$-convergence plots for various densities.} \label{fig:L2_conv} \end{center} \end{figure} For the Cauchy density, both the histogram based on Scott's rule and the KDE approach fail to converge. This is because Scott's rule requires a finite second moment, whereas the kernel used in the KDE estimator is a Gaussian kernel, which has finite moments. But MLD-DE produces convergent estimates of the Cauchy density without any need to change the parameters from those used with the other densities. Furthermore, it also performs better than the IQR-based histogram, which is designed to be less sensitive to outliers in the data. Thus, MLD-DE provides a robust alternative to the histogram and kernel density estimation methods, while offering competitive convergence performance. \subsection{Spatial variation of the pointwise error} We now consider the pointwise bias and variance of MLD-DE. Given a fixed sample size, $N$, the bias and variance are estimated by simulations over 1000 trials. Figure~\ref{fig:std_bias} shows the results; it shows pointwise estimates of the mean and the standard error of the density estimates plotted alongside the true densities. We see that the pointwise variance increases with the value of the true density, while the bias is larger towards the corners of the estimation region. For comparison, Figure~\ref{fig:std_bias_others} shows analogous plots for the KDE and IQR histogram methods. \begin{figure}[!h] \begin{center} \subfigure[Beta(1,4) distribution] { \includegraphics[width=0.35\textwidth]{beta_std.pdf} \label{fig:beta_std}} \subfigure[Gaussian Mixture] { \includegraphics[width=0.35\textwidth]{gm_std.pdf} \label{fig:gmixture_std}} \subfigure[Gaussian mixture: optimal $\alpha$] { \includegraphics[width=0.35\textwidth]{gm_std_opt.pdf} \label{fig:gmixture_std_opt_a}} \subfigure[Cauchy(0,1) distribution] { \includegraphics[width=0.35\textwidth]{cauchy_std.pdf} \label{fig:cauchy_std}} \caption{Pointwise mean and variance of the MLD-DE estimates for various densities.} \label{fig:std_bias} \end{center} \end{figure} In particular, for the beta density (Figure~\ref{fig:beta_std}), the bias is smaller in the middle regions of the support of the density. However, the bias is large near the boundary point $x=0$, where the density has a discontinuity. Figure~\ref{fig:gmixture_std} shows the corresponding results for the Gaussian mixture. Again, we see a smaller variance in the tails of the density, but a larger bias in the tails. As the variance increases with the density, we see larger variances near the peaks than at the troughs. The results improve considerably with the optimal choice of $\alpha$ (Figure~\ref{fig:gmixture_std_opt_a}), with a significant decrease in the bias. Figure~\ref{fig:cauchy_std} shows the results for the Cauchy density; these show a small bias in the tails but very low variance. \begin{figure}[!h] \begin{center} \subfigure[Beta(1,4) distribution] { \includegraphics[width=0.35\textwidth]{beta_std_kde.pdf} \label{fig:beta_std_kde}} \subfigure[Gaussian Mixture] { \includegraphics[width=0.35\textwidth]{gm_std_kde.pdf} \label{fig:gmixture_std_kde}} \subfigure[Cauchy(0,1) distribution] { \includegraphics[width=0.35\textwidth]{cauchy_std_iqr.pdf} \label{fig:cauchy_std_iqr}} \caption{Pointwise mean and variance of the adaptive KDE and IQR based histogram methods.} \label{fig:std_bias_others} \end{center} \end{figure} \subsection{Effect of varying the tuning parameter $\alpha$} The MLD-DE method depends on the parameter, $\alpha$, that controls the ratio of number of subsets, $m_N$, to size, $s_N$, of each subset. This is similar to the dependence of histogram and KDE methods on a bandwidth parameter. However, MLD-DE allows the use of different $\alpha$ at each point of estimation without affecting the estimates at other points. This opens the possibility of flexible adaptive density estimation. To evaluate the effect of $\alpha$ on the $L^{2}$-error, simulations were performed using values of $\alpha$ that increased from zero to one, with the total number of samples fixed to $N=1000$. The simulations were done for the beta$(1,4)$, Gaussian mixture and Cauchy$(0,1)$ distributions. Figure~\ref{fig:error_vs_a} shows plots of the estimated $L^{2}$-error as a function of $\alpha$ for the different densities. \begin{figure}[!h] \begin{center} \subfigure[Beta(1,4) distribution] { \includegraphics[width=0.35\textwidth]{beta_vs_a_zero.pdf} \label{fig:beta_a}} \subfigure[Gaussian Mixture] { \includegraphics[width=0.35\textwidth]{gm_vs_a_zero.pdf} \label{fig:gmixture_a}} \subfigure[Cauchy(0,1) distribution] { \includegraphics[width=0.35\textwidth]{cauchy_vs_a_zero.pdf} \label{fig:cauchy_a}} \caption{$L^{2}$-error versus the parameter $\alpha$ for various densities. The sample size was fixed to $N = 1000$. A large $\alpha$ implies a small number of subsets $m_N$, but a large number of samples $s_N$ in each subset, while a smaller $\alpha$ implies the converse.} \label{fig:error_vs_a} \end{center} \end{figure} All the curves have a similar profile, with the error increasing sharply for $\alpha \geq 0.7$; so the plots only show the errors for $\alpha \leq 0.8$. This indicates that, as we saw in Section \ref{sec:theory}, the number of subsets must be larger than their size. As we decrease $\alpha$ (i.e., increase the number of subsets), we see that the error is less sensitive to changes in the parameter. Decreasing $\alpha$ increases the bias, but keeps the variance low. In general, the `optimal' value of $\alpha$ lies in between 0.2 and 0.6 for these simulations, which further restricts the search range of any optimization problem for $\alpha$. \subsubsection*{An example of adaptive implementation} An adaptive approach was used to improve MLD-DE estimates of the Cauchy distribution. The numerical results in Figure~\ref{fig:cauchy_std} indicate that there is a larger bias in the tails of the distribution, while the theory indicates that the bias can be reduced by decreasing the number of subsets (correspondingly increasing the number of samples in each subset). The adaptive procedure used is as follows: (1) A pilot density was first computed using MLD-DE with $\alpha = {1}/{3}$; (2) The points of estimation where the pilot density was within a fifth of the gap between the maximum and minimum density values from the minimum value (i.e., where the density was relatively small) were identified; (3) The MLD-DE procedure was repeated with the value $\alpha={1}/{2}$ for those points of estimation. \begin{figure}[!h] \begin{center} \subfigure[Mean Absolute Error] { \includegraphics[width=0.35\textwidth]{mae_vs_x_cauchy_adaptive.pdf} \label{fig:mae_cauchy_adaptive}} \subfigure[Bias and Variance Error] { \includegraphics[width=0.35\textwidth]{cauchy_std_adaptive.pdf} \label{fig:std_cauchy_adaptive}} \subfigure[Adaptive MLD Density] { \includegraphics[width=0.35\textwidth]{cauchy_comparison_adaptive.pdf} \label{fig:comparison_cauchy_adaptive}} \caption{Cauchy density estimation with the adaptive MLD-DE} \label{fig:cauchy_adaptive} \end{center} \end{figure} Figure~\ref{fig:cauchy_adaptive} shows the results of this adaptive approach. We see that the bias has decreased significantly compared to that shown in the earlier plots for the non-adaptive approach. More sophisticated adaptive strategies can be employed with MLD-DE on account of its naturally adaptive nature, however a discussion of them is beyond the scope of this paper. \section{Discussion and generalizations \label{sec:conc}} We have presented a simple, robust and easily parallelizable method for one-dimensional density estimation. Like nearest-neighbor density estimators, the method is based on nearest-neighbors but it offers the advantage of providing smoother density estimates, and has parallel complexity $\orderof{N^{{1}/{3}}}$ . Its tuning parameter is the number of subsets in which the original sample is divided. Theoretical results concerning the asymptotic distribution of the estimator were developed and its MSE was analyzed to determine a globally optimal split of the original sample into subsets. Numerical experiments illustrate that the method can recover different types of densities, including the Cauchy density, without the need for special kernels or bandwidth selections. Based on a heuristic analysis of high bias in low-density regions, an adaptive implementation that reduces the bias was also presented. Further work will be focused on more sophisticated adaptive schemes for one-dimensional density estimation and extensions to higher dimensions. We present here a brief overview of a higher dimensional extension of MLD-DE. Its generalization is straightforward but its convergence is usually not better than that of histogram methods. To see why, we consider the bivariate case. Let $(X,Y)$ be a random vector with PDF $f(x,y)$, and let $h(x,y)$ and $H(x,y)$ be the PDF and CDF of $(|X|,|Y|)$. It is easy to see that \( h(0,0) = 4 f(0,0). \) In addition, let $q(t)=H(t,t)$, then \( q'(t) = \int_0^t h(t,y)\,dy + \int_0^t h(x,t)\,dx. \) It follows that (assuming continuity at $(0,0)$), \( q''(0)=\lim_{t\to 0} q'(t)/t = 2\, h(0,0). \) Let ${\bf{X}}_1=(X_1,Y_1),\ldots,{\bf{X}}_N=(X_N,Y_N)$ be iid vectors and define $U_i$ to be the product norm of ${\bf{X}}_i$: \( U_i = \|{\bf{X}}_i\|_\otimes = \max\{\,|X_i|,\,|Y_i|\,\}, \) and $Z = U_{(1)}$. Then \begin{eqnarray*} \mathbb{P}(\,Z>t\,) &=& \mathbb{P}( \,\|{\bf{X}}_1\|_\otimes >t,\ldots, \|{\bf{X}}_N\|_\otimes >t\,)=\mathbb{P}( \,\|{\bf{X}}_1\|_\otimes >t\,)^N\\ &=& [\,1-\mathbb{P}(\,\|{\bf{X}}_1\|_\otimes \leq t\,)\,]^N = [\,1-\mathbb{P}(\,|X_1| \leq t,|Y_1| \leq t\,)\,]^N\\ &=& [\,1-q(\,t\,)\,]^N. \end{eqnarray*} Let $Q$ be the inverse of the function $q$. It is easy to check that \( Q'(z) = 1/q'(Q(z)). \) Proceeding as in the 1D case, we have \begin{eqnarray*} \mathbb{E} (\,Z^2\,) &=& 2\!\!\int_0^\infty \hspace{-.3cm}t\,\mathbb{P}(\,Z>t\,)\,dt = 2\!\!\int_0^\infty\hspace{-.3cm} t\,[\,1-q(\,t\,)\,]^N\,dt = \int_0^1 (Q^2(z))' \,(1-z)^N\,dz. \end{eqnarray*} Therefore \( \mathbb{E} [\,(N+1)\,Z^2\,] = \int_0^1 (Q^2(z))' \,\delta_N(z)\,dz, \) and by the results in Section \ref{sec:theory}, \[ \lim_n \mathbb{E}[\,(N+1)\,Z^2\,] = (Q^2(z))'|_{z=0}={1}/{h(0,0)} ={1}/{(4\,f(0,0))}. \] Furthermore, \[ \mathbb{E} [\,(N+1)\,Z^2\,] = \frac{1}{h(0,0)} + \frac{1}{N+2}\int_0^1 (Q^2(z))'' \,\delta_{N+1}(z)\,dz. \] But, unlike in the 1D case, this time we have \( \lim_{z\to 0}\, (Q^2(z))'' = {q^{(4)}(0)}/{(3 q''(0))}\neq 0, \) and this makes the convergence rates closer to those of histogram methods. \smallskip \noindent{\bf Acknowledgments.} We thank Y. Marzouk for reviewing our proofs, and D. Allaire, L. Ng, C. Lieberman and R. Stogner for helpful discussions. The first and third authors acknowledge the support of the the DOE Applied Mathematics Program, Awards DE-FG02-08ER2585 and DE-SC0009297, as part of the DiaMonD Multifaceted Mathematics Integrated Capability Center. \bibliographystyle{apalike}

https://arxiv.org/abs/2206.06971

Conditioning of linear systems arising from penalty methods

Penalizing incompressibility in the Stokes problem leads, under mild assumptions, to matrices with condition numbers $\kappa =\mathcal{O} (\varepsilon ^{-1}h^{-2})$, $\varepsilon =$ penalty parameter $<<1$, and $ h= $ mesh width $<1$. Although $\kappa =\mathcal{O}(\varepsilon ^{-1}h^{-2}) $ is large, practical tests seldom report difficulty in solving these systems. In the SPD case, using the conjugate gradient method, this is usually explained by spectral gaps occurring in the penalized coefficient matrix. Herein we point out a second contributing factor. Since the solution is approximately incompressible, solution components in the eigenspaces associated with the penalty terms can be small. As a result, the effective condition number can be much smaller than the standard condition number.

\section{Estimating conditioning} \begin{center} \textit{We dedicate this paper to Professor Owe Axelsson.} \end{center} Penalty methods have advantages and disadvantages. Disadvantages include the need to pick a specific value of the penalty parameter $\varepsilon $ and ill-conditioning of the associated linear system. We show herein that, measured by the system's effective condition number, this ill-conditioning is not severe, Theorem 2.1 below.\ This observation is developed herein for the standard penalty approximation to the Stokes problem% \begin{equation*} -\triangle u+\nabla p=f(x)\ \ \text{and} \ \ \nabla \cdot u=0, \end{equation*} in a polygonal domain $\Omega $ with boundary condition $u=0$ on $\partial \Omega $ and normalization $\int_{\Omega }pdx=0$. Let $(\cdot ,\cdot ),||\cdot ||$ denote the $L^{2}(\Omega )$ inner product and norm and $|\cdot |$\ the usual euclidean norm of a vector and matrix. $X^{h}$ denotes a standard, conforming, velocity finite element space of continuous, piecewise polynomials that vanish on $\partial \Omega $. The penalty approximation results from replacing $\nabla \cdot u=0$\ by $\nabla \cdot u^{\varepsilon }=-\varepsilon p^{\varepsilon }$ and eliminating $p^{\varepsilon }$. It is: find $u^{\varepsilon }\in X^{h}$ satisfying% \begin{equation} (\nabla u^{\varepsilon },\nabla v)+\varepsilon ^{-1}(\nabla \cdot u^{\varepsilon },\nabla \cdot v)=(f,v)\text{ for all }v\in X^{h}. \label{eq:PenaltyStokeseqns} \end{equation}% Picking a basis $\left\{ \phi _{1},\cdot \cdot \cdot ,\phi _{N}\right\} $ for $X^{h}$ leads to a linear system with coefficient matrix% \begin{equation*} A_{ij}=(\nabla \phi _{i},\nabla \phi _{j})+\varepsilon ^{-1}(\nabla \cdot \phi _{i},\nabla \cdot \phi _{j}),i,j=1,\cdot \cdot \cdot ,N. \end{equation*}% Standard condition number estimates for this system yield bounds like $% \kappa \leq C\varepsilon ^{-1}h^{-2}$. Recall that the standard condition number, \cite{A96,W63}, introduced in \cite{NG47,T48} and still of interest, e.g., \cite% {TV10}, measuring the correlation between relative error and relative residual, the distance to the nearest singular matrix and the difficulty, cost and worse-case accuracy in solving a linear system with $A$, is $\kappa :=|A||A^{-1}|$. However, estimates of the above with $\kappa $\ are often (but not always, \cite{B98}) too rough because they do not consider the size of solution components in the matrix eigenspaces. Thus, generalizations exist, such as the Kaporin condition number \cite{K94} and extensions based on generalized inverses \cite{D86}, that can be used to obtain better predictions. One important, easy to compute and interpret generalization is the \textit{condition number at the solution}, also called the \textit{% effective condition number}. \begin{definition} Let $Ac=b$. Then, $\kappa (c)$, the condition number at the solution $c$, is% \begin{equation*} \kappa (c):=|A^{-1}|\frac{|Ac|}{|c|}. \end{equation*} \end{definition} Clearly $\kappa (c)\leq \kappa $ and $\kappa (c)$ takes into account both the spectrum and the magnitude of solution components across eigenspaces. It is also known, e.g. \cite{A96}, that the relative error in approximations is bounded by $\kappa (c)$\ times the relative residual. To our knowledge, this extension is due to Chan and Foulser \cite{C88}. It was soon thereafter developed by Axelsson \cite{A95,A96} and has been further developed in important work in \cite{AK00,AK01,B99,LH08,LHH08,CH94}. Section 2 gives the proof that $\kappa (u^{\varepsilon })<<\kappa $. Section 3 presents consistent numerical tests. \section{Analysis of $\protect\kappa (u^{\protect\varepsilon })$} We assume $X^{h}$\ satisfies the following two assumptions typical, e.g. \cite{BS08,C02}, of finite element spaces on quasi-uniform meshes. A1: [Inverse estimate] For all $v\in X^{h},$ $||\nabla v||\leq Ch^{-1}||v||.$ A2: [Norm equivalence] Let $v=\sum_{i=1}^{N}a_{i}\phi _{i}(x),N=Ch^{-d},d=\dim (\Omega )=2$ or $3$. Then $||v||$ and $\sqrt{\frac{1% }{N}\sum_{i=1}^{N}a_{i}^{2}}$ are uniform-in-$h$ equivalent norms. \begin{theorem} Let A1 and A2 hold. Select $|\cdot |$ to be the euclidean norm. Let $f^{h}$ be the projection of $f$ into the finite element space and $u^{\varepsilon }$% \ the solution of (\ref{eq:PenaltyStokeseqns}). Then,% \begin{eqnarray*} \max_{f^{h}}\kappa (u^{\varepsilon }) &=&\kappa \leq C(h^{-2}+\varepsilon ^{-1}h^{-2}) \\ \kappa (u^{\varepsilon }) &\leq &C\frac{||f^{h}||}{||u^{\varepsilon }||}% \text{, and} \\ \kappa (u^{\varepsilon }) &\leq &Ch^{-2}\left( 1+\frac{h}{\varepsilon }\frac{% ||\nabla \cdot u^{\varepsilon }||}{||u^{\varepsilon }||}\right) . \end{eqnarray*} \end{theorem} \begin{proof} We first estimate $|A^{-1}|^{2}=\max_{b}|A^{-1}b|^{2}/|b|^{2}.$ Given an arbitrary right-hand side $\overrightarrow{b}$, let the linear system for the undetermined coefficients $(c_{1},c_{2},\cdot \cdot \cdot ,c_{N})^{T}=% \overrightarrow{c}$ be denoted $A\overrightarrow{c}=\overrightarrow{b}.$ We convert in a standard way this linear system to an equivalent formulation similar to (\ref{eq:PenaltyStokeseqns}). Recall that the mass matrix associated with this basis is \begin{equation*} M_{ij}=(\phi _{i},\phi _{j}),i,j=1,\cdot \cdot \cdot ,N. \end{equation*}% The matrix $M$ is symmetric and positive definite. Under A1 and A2, its eigenvalues satisfy $0<C_{1}N\leq \lambda (M)\leq C_{2}N$. Given the vector $% \overrightarrow{b}$ let $\overrightarrow{a}=M^{-1}\overrightarrow{b}$. By construction \begin{equation*} g(x)=\sum_{i=1}^{N}a_{i}\phi _{i}(x)\text{ satisfies }(g(x),\phi _{j})=b_{j},j=1,\cdot \cdot \cdot ,N. \end{equation*}% By A2, $||g||$ and $(N^{-1}\sum a_{i}^{2})^{1/2}$\ are uniformly equivalent norms. The bounds on $\lambda (M)$ (also from A2) imply $(N^{-1}\sum a_{i}^{2})^{1/2}$\ and $(N^{-1}\sum b_{i}^{2})^{1/2}$\ are also uniformly equivalent norms. Next define \begin{equation*} w=\sum_{i=1}^{N}c_{i}\phi _{i}(x). \end{equation*}% By A2 again $||w||$, $(N^{-1}\sum c_{i}^{2})^{1/2}$are equivalent norms. Thus,% \begin{equation*} |A^{-1}|^{2}=\max_{b}\frac{|A^{-1}b|^{2}}{|b|^{2}}\leq C\max_{g\in X^{h}}% \frac{||w||^{2}}{||g||^{2}}. \end{equation*}% By construction $w,g\in X^{h}$\ satisfy% \begin{equation*} (\nabla w,\nabla v)+\varepsilon ^{-1}(\nabla \cdot w,\nabla \cdot v)=(g,v)% \text{ for all }v\in X^{h}. \end{equation*}% Setting $v=w$ and using simple inequalities gives $||w||^{2}\leq C||g||^{2}$% . Indeed,% \begin{equation*} C||w||^{2}\leq ||\nabla w||^{2}+\varepsilon ^{-1}||\nabla \cdot w||^{2}=(g,w)% \text{ }\leq \frac{C}{2}||w||^{2}+C||g||^{2}. \end{equation*}% Thus $||w||^{2}\leq C||\nabla w||^{2}\leq C||g||^{2}$ and $% ||w||^{2}/||g||^{2}\leq C$. This implies $|A^{-1}|\leq C$, uniformly of $h,$ $\varepsilon $. Next we estimate $|A\overrightarrow{c}|/|\overrightarrow{c}|$ where $% u^{\varepsilon }=\sum_{i=1}^{N}c_{i}\phi _{i}(x)$ is the solution of (\ref% {eq:PenaltyStokeseqns}). By norm equivalence, as above, and (\ref% {eq:PenaltyStokeseqns}) this is equivalent to $||f^{h}||/||u^{\varepsilon }|| $ where $f^{h}$\ is the $L^{2}$ projection of $f(x)$ into the finite element space. This gives the first estimate $\kappa (u^{\varepsilon })\leq C||f^{h}||/||u^{\varepsilon }||$. For the second estimate, we have% \begin{eqnarray*} ||f^{h}|| &=&\max_{v\in X^{h}}\frac{(f^{h},v)}{||v||}=\max_{v\in X^{h}}\frac{% (\nabla u^{\varepsilon },\nabla v)+\varepsilon ^{-1}(\nabla \cdot u^{\varepsilon },\nabla \cdot v)}{||v||} \\ &\leq &\max_{v\in X^{h}}\frac{||\nabla u^{\varepsilon }||||\nabla v||+\varepsilon ^{-1}||\nabla \cdot u^{\varepsilon }||||\nabla \cdot v||}{% ||v||} \\ \text{(using A1)} &\leq &\max_{v\in X^{h}}\frac{Ch^{-2}||u^{\varepsilon }||||v||+\varepsilon ^{-1}Ch^{-1}||\nabla \cdot u^{\varepsilon }||||v||}{% ||v||} \\ &\leq &Ch^{-2}||u^{\varepsilon }||+\varepsilon ^{-1}Ch^{-1}||\nabla \cdot u^{\varepsilon }||\text{, which implies} \\ \frac{||f^{h}||}{||u^{\varepsilon }||} &\leq &Ch^{-2}+C\varepsilon ^{-1}h^{-1}\frac{||\nabla \cdot u^{\varepsilon }||}{||u^{\varepsilon }||}. \end{eqnarray*}% This implies% \begin{equation*} \kappa (u^{\varepsilon })\leq C\frac{||f^{h}||}{||u^{\varepsilon }||}\leq C\left( h^{-2}+\varepsilon ^{-1}h^{-1}\frac{||\nabla \cdot u^{\varepsilon }||% }{||u^{\varepsilon }||}\right) . \end{equation*}% Using $||\nabla \cdot u^{\varepsilon }||\leq C||\nabla u||\leq Ch^{-1}|||u^{\varepsilon }||$ and A1 yields the standard estimate $\kappa \leq C(h^{-2}+\varepsilon ^{-1}h^{-2}).$ \end{proof} \section{An Illustration} The result in Section 2 gives 3 estimates of conditioning. We explore if the 3 estimates% \begin{gather*} \kappa \leq C(h^{-2}+\varepsilon ^{-1}h^{-2}),\text{ \ \ }\kappa (u^{\varepsilon })\leq C\frac{||f^{h}||}{||u^{\varepsilon }||},\text{ and} \\ \kappa (u^{\varepsilon })\leq C\left( h^{-2}+\varepsilon ^{-1}h^{-1}\frac{% ||\nabla \cdot u^{\varepsilon }||}{||u^{\varepsilon }||}\right) \end{gather*}% yield significantly different predictions. We investigate this question for 2 test problems and 2 finite element spaces. The first test problem has a smooth solution and the second has an $f(x,y)$ oscillating as fast as possible on the given mesh. The domain is the unit square and the mesh is a uniform mesh of squares each divided into 2 right triangles. The first finite element space is P1, conforming linear elements. The second is P2, conforming quadratics. The space of conforming linears does not contain a divergence-free subspace. Thus the coefficient matrix $A$ is expected to show greater ill-conditioning as $\varepsilon \rightarrow 0$\ than with conforming quadratics. The above constants "$C$" are $\mathcal{O}(1)$ constants, independent of $\varepsilon $ and $h$. Thus, we compute and compare the RHS' below% \begin{eqnarray*} Est.1 &\simeq &h^{-2}+\varepsilon ^{-1}h^{-2}, \\ Est.2 &\simeq &\text{ }\frac{||f^{h}||}{||u^{\varepsilon }||}\text{ and } \\ Est.3 &\simeq &\varepsilon ^{-1}h^{-1}\frac{||\nabla \cdot u^{\varepsilon }||% }{||u^{\varepsilon }||}. \end{eqnarray*}% We computed these values starting with $\varepsilon =1,h=1$ and halving each until $\varepsilon =5.96E-08$ and $h=0.00195312.$ We present the results below selected from equi-spaced $\varepsilon $-values in this data. \textbf{Test problem 1:} We solve this test problem with P1, conforming linear elements, and then with P2, conforming, quadratic elements. Choose $f(x,y)=\left( \sin \left( x+y\right) ,\cos \left( x+y\right) \right) ^{T}$. For P1 elements we have the following results for $Est.2$. \begin{gather*} \begin{array}{c|cccccc} {\small \varepsilon \Downarrow \setminus h\Rightarrow }\text{ \ } & {\small % 0.125} & {\small 6.2E-2} & {\small 3.1E-2} & \text{ }{\small 1.6E-2} & {\small 7.8E-3} & {\small 3.9E-3} \\ \hline {\small 1} & {\small 39} & {\small 38} & {\small 37} & {\small 37} & {\small % 37} & {\small 37} \\ {\small 6.3E-2} & {\small 1.9E2} & {\small 1.7E2} & {\small 1.7E2} & {\small % 1.7E2} & {\small 1.7E2} & {\small 1.7E2} \\ {\small 3.9E-3} & {\small 1.8E3} & {\small 7.8E2} & {\small 4.3E2} & {\small % 3.3E2} & {\small 3.0E2} & {\small 2.9E2} \\ {\small 2.4E-4} & {\small 2.5E4} & {\small 7.5E3} & {\small 2.2E3} & {\small % 8.0E2} & {\small 4.3E2} & {\small 3.3E2} \\ {\small 1.5E-5} & {\small 4.0E5} & {\small 1.1E5} & {\small 2.9E4} & {\small % 7.5E3} & {\small 2.1E3} & {\small 8.0E2} \\ {\small 9.5E-7} & {\small 6.4E6} & {\small 1.8E6} & {\small 4.5E5} & {\small % 1.1E5} & {\small 2.9E4} & {\small 7.4E3} \\ {\small 6.0E-8} & {\small 1.0E8} & {\small 2.9E7} & {\small 7.2E6} & {\small % 1.8E6} & {\small 4.5E5} & {\small 1.1E5}\\ \end{array} \\ \text{ Table 1: Values of }Est.2\simeq \text{ }\frac{||f^{h}||}{% ||u^{\varepsilon }||}\text{, P1 elements} \end{gather*}% Down the columns (fixing $h$ and decreasing $\varepsilon $), the data for\ $% Est.2$ shows that this quantity grows as $\varepsilon \downarrow 0$, roughly like $\varepsilon ^{-1}$\ for fixed $h$. Across the rows, for fixed $% \varepsilon $\ and $h\downarrow 0$, $Est.2$ decreases. Diagonals (necessary if we choose $\varepsilon \sim h$) show mild growth. Comparing the last row of the table with the row of $Est.1$ values shows that $Est.2$\ consistently yields a lower estimate of ill-conditioning than $Est.1$. Concerning the growth of $Est.2$ as $\varepsilon \downarrow 0$ for fixed $h$, since $h,||f|| $ do not change as $\varepsilon $ varies, this growth is due to $% ||u^{\varepsilon }||\rightarrow 0$ as $\varepsilon \downarrow 0$. The P1 finite element space does not have a non-trivial divergence free subspace, \cite{JLMNR17}. Thus, $\varepsilon \downarrow 0$\ forces $||\nabla \cdot% u||\rightarrow 0$\ which forces $||u||\rightarrow 0$. This indicates that ill-conditioning of penalty methods \textit{with P1 elements} as $% \varepsilon \downarrow 0$ is an essential feature caused by the \textit{lack of a divergence-free subspace}. For comparison with the last row in Table 1, the values of\ $Est.1\simeq h^{-2}+\varepsilon ^{-1}h^{-2}$ for $\varepsilon =6.0E-8$ are \begin{equation*} \begin{array}{cccccc} \text{ }{\small 1.0E9} & \text{ }{\small 4.3E9}\text{ } & {\small 1.7E10} & {\small 6.9E10} & {\small 2.7E11} & {\small 1.1E12}% \end{array}% . \end{equation*}% Clearly $Est.2$\ provides a smaller estimate than $Est.1$. Another interpretation is that a poor choice of finite element space (made to accentuate ill-conditioning) makes the problem of selecting $\varepsilon $ acute. Table 2 presents the analogous results for $Est.3$ and P1 elements.% \begin{gather*} \begin{array}{c|cccccc} {\small \varepsilon \Downarrow \setminus h\Rightarrow } & {\small 0.125} & {\small 6.2E-2} & {\small 3.1E-2} & \text{ }{\small 1.6E-2} & {\small 7.8E-3} & {\small 3.9E-3} \\ \hline {\small 1} & {\small 25} & {\small 50} & {\small 1.0E2} & {\small 2.0E2} & {\small 4.0E2} & {\small 8.0E2} \\ {\small 6.3E-2} & {\small 3.5E2} & {\small 6.6E2} & {\small 1.3E3} & {\small % 2.5E3} & {\small 5.0E3} & {\small 1.0E4} \\ {\small 3.9E-3} & {\small 4.2E3} & {\small 4.1E3} & {\small 4.5E3} & {\small % 6.3E3} & {\small 1.1E4} & {\small 2.0E4} \\ {\small 2.4E-4} & {\small 6.2E4} & {\small 4.9E4} & {\small 4.2E4} & {\small % 4.5E4} & {\small 3.5E2} & {\small 5.1E4} \\ {\small 1.5E-5} & {\small 9.8E5} & {\small 7.6E5} & {\small 6.3E5} & {\small % 4.2E4} & {\small 6.0E5} & {\small 6.4E5} \\ {\small 9.5E-7} & {\small 1.5E7} & {\small 1.2E7} & {\small 1.0E7} & {\small % 9.5E6} & {\small 9.4E6} & {\small 9.4E6} \\ {\small 6.0E-8} & {\small 2.5E8} & {\small 1.9E8} & {\small 1.6E8} & {\small % 1.5E8} & {\small 1.5E8} & {\small 1.5E8}% \end{array} \\ \text{Table 2: Values of }Est.3\simeq \varepsilon ^{-1}h^{-1}\frac{||\nabla \cdot u^{\varepsilon }||}{||u^{\varepsilon }||}\text{ , P1 elements} \end{gather*}% For $Est.3$\ the data shows similar behavior to $Est.2$ except for fixed $% \varepsilon $ and $h\downarrow 0$. In this case, ill-conditioning for small $% \varepsilon $\ and $h\downarrow 0$\ is over-estimated compared with Table 1. This is expected because the data comes from P1 elements and $||\nabla \cdot u^{\varepsilon }||$\ occurs in $Est.3$.\ Again, $Est.3$\ still provides a smaller estimate of ill-conditioning than $Est.1$. We have attributed the ill-conditioning observed above as $\varepsilon \rightarrow 0$ to the use of P1 elements. To test if this explanation is plausible we repeated this test with P2 elements. Since this finite element space contains a non-trivial divergence free subspace, \cite{JLMNR17}, our intuition is that ill-conditioning would be reduced. Table 3 below presents $% Est.3\simeq \varepsilon ^{-1}h^{-1}\frac{||\nabla \cdot u^{\varepsilon }||}{% ||u^{\varepsilon }||}$ values. $Est.2\simeq ||f^{h}||/||u^{\varepsilon }||$ values, not presented, were significantly smaller than $Est.3$ and converged to $1.13$\ as $\varepsilon ,h\rightarrow 0$.% \begin{gather*} \begin{array}{c|cccccc} {\small \varepsilon \Downarrow \setminus h\Rightarrow } & {\small 0.125} & {\small 6.2E-2} & {\small 3.1E-2} & \text{ }{\small 1.6E-2} & {\small 7.8E-3} & {\small 3.9E-3} \\ \hline {\small 1} & {\small 25} & {\small 50} & {\small 1.0E2} & {\small 2.0E2} & {\small 4.0E2} & {\small 8.0E2} \\ {\small 6.3E-2} & {\small 3.1E2} & {\small 6.2E2} & {\small 1.2E3} & {\small % 2.5E3} & {\small 5.0E3} & {\small 1.0E4} \\ {\small 3.9E-3} & {\small 6.7E2} & {\small 1.3E3} & {\small 2.5E3} & {\small % 5.0E3} & {\small 1.0E4} & {\small 2.0E4} \\ {\small 2.4E-4} & {\small 7.1E2} & {\small 1.3E3} & {\small 2.6E3} & {\small % 5.1E3} & {\small 1.0E4} & {\small 2.0E4} \\ {\small 1.5E-5} & {\small 7.1E2} & {\small 1.3E3} & {\small 2.6E3} & {\small % 5.2E3} & {\small 1.0E4} & {\small 2.0E4} \\ {\small 9.5E-7} & {\small 7.1E2} & {\small 1.3E3} & {\small 2.7E3} & {\small % 5.3E3} & {\small 1.1E4} & {\small 2.1E4} \\ {\small 6.0E-8} & {\small 7.1E2} & {\small 1.3E3} & {\small 2.7E3} & {\small % 5.3E3} & {\small 1.1E4} & {\small 2.1E4}% \end{array} \\ \text{Table 3: Values of }Est.3\simeq \varepsilon ^{-1}h^{-1}\frac{||\nabla \cdot u^{\varepsilon }||}{||u^{\varepsilon }||}\text{, P2 elements} \end{gather*}% The above data indicates that with P2 elements and a problem with a smooth solution, the contribution of the penalty term to ill-conditioning is small. \textbf{Test problem 2:} We take \begin{equation*} f(x,y)=\left( 10+50\sin \left( \frac{2\pi x}{h}+\frac{2\pi y}{h}\right) ,10+50\sin \left( \frac{2\pi x}{h}+\frac{2\pi y}{h}\right) \right) ^{T}. \end{equation*}% Each component oscillates as rapidly as the given mesh allows, see Figure % \ref{FigF(x,y)} below for $h=0.125$. \begin{figure}[htp] \begin{center} \includegraphics[height=2.5381in, width=3.0933in]{Figure1.png} \end{center} \caption{Plot of $10+50\sin(\frac{2\pi x}{0.125}+\frac{2\pi y}{0.125})$.}% \label{FigF(x,y)}% \end{figure} The forcing function and thus the solution change with the mesh size. \begin{gather*} \begin{array}{c|cccccc} {\small \varepsilon \Downarrow \setminus h\Rightarrow } & {\small 0.125} & {\small 6.2E-2} & {\small 3.1E-2} & \text{ }{\small 1.6E-2} & {\small 7.8E-3} & {\small 3.9E-3} \\ \hline {\small 1} & {\small 1.1E2} & {\small 1.0E2} & {\small 1.0E2} & {\small 1.0E2% } & {\small 1.0E2} & {\small 1.0E2} \\ {\small 6.3E-2} & {\small 5.6E2} & {\small 5.6E2} & {\small 5.6E2} & {\small % 5.6E2} & {\small 5.6E2} & {\small 5.6E2} \\ {\small 3.9E-3} & {\small 7.4E3} & {\small 7.5E3} & {\small 7.6E3} & {\small % 7.6E3} & {\small 7.7E3} & {\small 7.7E3} \\ {\small 2.4E-4} & {\small 1.1E5} & {\small 1.1E5} & {\small 1.2E5} & {\small % 1.2E5} & {\small 1.2E5} & {\small 1.2E5} \\ {\small 1.5E-5} & {\small 1.8E6} & {\small 1.8E6} & {\small 1.8E6} & {\small % 1.8E6} & {\small 1.8E6} & {\small 1.9E6} \\ {\small 9.5E-7} & {\small 2.9E7} & {\small 2.9E7} & {\small 2.9E7} & {\small % 2.9E7} & {\small 2.9E7} & {\small 2.9E7} \\ {\small 6.0E-8} & {\small 4.7E8} & {\small 4.6E8} & {\small 4.6E8} & {\small % 4.6E8} & {\small 4.6E8} & {\small 4.6E8}% \end{array}\\ \text{Table 3: Values of }Est.2\simeq \frac{||f^{h}||}{||u^{\varepsilon }||} \text{, P1 elements} \end{gather*} For comparison with the last row in Table 3, the values of\ $Est.1\simeq h^{-2}+\varepsilon ^{-1}h^{-2}$ for $\varepsilon =6.0E-8$ are% \begin{equation*} \begin{array}{cccccc} {\small 1.1E9} & {\small 4.3E9} & {\small 1.7E10} & {\small 6.9E10} & {\small 2.7E11} & {\small 1.1E12}% \end{array}. \end{equation*} In Table 3, as $\varepsilon \downarrow 0,$ $Est.2$\ grows roughly like $% \varepsilon ^{-1}$. The stable behavior of $Est.2$ as $h\downarrow 0$\ is unexpected. The behavior of $Est.3\simeq \varepsilon ^{-1}h^{-1}\frac{% ||\nabla \cdot u^{\varepsilon }||}{||u^{\varepsilon }||}$\ for P1 elements was similar, so not presented. We now present tests of P2 elements for this problem. In this test $Est.3$ was larger than $Est.2,$ so we present only $Est.3$ data.% \begin{gather*} \begin{array}{c|cccccc} {\small \varepsilon \Downarrow \setminus h\Rightarrow } & {\small 0.125} & {\small 6.2E-2} & {\small 3.1E-2} & \text{ }{\small 1.6E-2} & {\small 7.8E-3} & {\small 3.9E-3} \\ \hline {\small 1} & {\small 25} & {\small 49} & {\small 1.0E2} & {\small 2.0E2} & {\small 4.0E2} & {\small 7.9E2} \\ {\small 6.3E-2} & {\small 3.8E2} & {\small 7.5E2} & {\small 1.5E3} & {\small % 3.0E3} & {\small 6.0E3} & {\small 1.2E4} \\ {\small 3.9E-3} & {\small 6.0E3} & {\small 1.2E4} & {\small 2.4E4} & {\small % 4.8E4} & {\small 9.6E4} & {\small 1.9E5} \\ {\small 2.4E-4} & {\small 9.6E4} & {\small 1.9E5} & {\small 3.8E5} & {\small % 7.7E5} & {\small 1.5E6} & {\small 3.1E6} \\ {\small 1.5E-5} & {\small 1.5E6} & {\small 3.1E6} & {\small 6.2E6} & {\small % 1.2E7} & {\small 2.5E7} & {\small 4.9E7} \\ {\small 9.5E-7} & {\small 2.5E7} & {\small 4.9E7} & {\small 9.8E7} & {\small % 2.0E8} & {\small 3.9E8} & {\small 7.9E8} \\ {\small 6.0E-8} & {\small 3.9E8} & {\small 7.9E8} & {\small 1.6E9} & {\small % 3.2E9} & {\small 6.3E9} & {\small 1.3E10}% \end{array} \\ \text{Table 4: Values of }Est.3\simeq \varepsilon ^{-1}h^{-1}\frac{||\nabla \cdot u^{\varepsilon }||}{||u^{\varepsilon }||}\text{ },\text{ P2 elements} \end{gather*}% For comparison with the last row in Table 4,\ $Est.1$ values for $% \varepsilon =6.0E-8$ are% \begin{equation*} \begin{array}{cccccc} {\small 1.1E9} & {\small 4.3E9} & {\small 1.7E10} & {\small 6.9E10} & {\small 2.7E11} & {\small 1.1E12}% \end{array}% . \end{equation*}% For this problem, down the columns (fixed $h$, $\varepsilon \downarrow 0$ ) the computed estimate of $\kappa (u^{\varepsilon })$\ grows roughly like $% \varepsilon ^{-1}$. Reading across the columns, the values in each row grow like $h^{-1}$\ (not $h^{-2}$). In all cases, $Est.3$ was smaller than $Est.1$% . \section{Conclusions} The classical view is that two significant issues are impediments to penalty methods giving low cost and highly accurate velocity approximates. The first is ill-conditioning of the resulting system matrix. We have shown that the effective condition number $\kappa (u^{\varepsilon })$ is much smaller than the usual condition number due to the magnitude of the components in the penalized eigenspaces being small in a precise sense. Motivated by this theoretical result, we then compared the derived estimates of ill-conditioning on two test problems and for two elements. With P1 elements, ill-conditioning was not as severe as $\mathcal{O}(\varepsilon ^{-1}h^{-2})$\ but followed $\varepsilon ^{-1}$ growth as $\varepsilon \rightarrow 0$. For P2 elements and approximating a smooth solution, $\kappa (u^{\varepsilon })$\ was smaller the expected $\kappa \sim h^{-2}$ for the discrete Laplacian. With both P1 and P2 elements, and an academic test problem with data oscillating as fast as the mesh allows, ill-conditioning was not as severe as the expected $\mathcal{O}(\varepsilon ^{-1}h^{-2})$\ but also followed the $\varepsilon ^{-1}$ pattern as $\varepsilon \rightarrow 0$. The second significant issue is the difficulty in the selection of an effective value of $\varepsilon $. While the most commonly recommended, e.g. \cite{CK82,HV95}, choices are $\varepsilon =$ \textit{time step, mesh width}, and (\textit{machine epsilon)}$^{1/2}$, none have proven reliably effective. Recent work of \cite{KXX21, X21} may resolve this impediment by an algorithmic, self-adaptive selection of $\varepsilon $ based on some indicator of violation of incompressibility. The estimates in Theorem 2.1 give insight into the resulting conditioning when this is done. Let $TOL$ denote a specified tolerance. If $\varepsilon $ is adapted so that the penalized solution $u^{\varepsilon }$\ satisfies $\frac{||\nabla \cdot u^{\varepsilon }||}{||u^{\varepsilon }||}\leq TOL,$ then, Theorem 2.1 immediately implies $\kappa (u^{\varepsilon })$ satisfies% \begin{equation} \kappa (u^{\varepsilon })\leq Ch^{-2}\left( 1+h\frac{TOL}{\varepsilon }% \right) . \label{EstAdaptive1} \end{equation}% If $\varepsilon $ is adapted so that the penalized solution $u^{\varepsilon } $\ satisfies $\frac{||\nabla \cdot u^{\varepsilon }||}{||\nabla u^{\varepsilon }||}\leq TOL,$ then, similarly, Theorem 2.1 implies $\kappa (u^{\varepsilon })$ satisfies% \begin{equation} \kappa (u^{\varepsilon })\leq Ch^{-2}\left( 1+\frac{TOL}{\varepsilon }% \right) . \label{EstAdaptive2} \end{equation}% For (\ref{EstAdaptive2}), rewrite $||\nabla \cdot u^{\varepsilon }||/||u^{\varepsilon }||$ as $\left( ||\nabla \cdot u^{\varepsilon }||/||\nabla u^{\varepsilon }||\right) \left( ||\nabla u^{\varepsilon }||/||u^{\varepsilon }||\right) $. The inverse estimate, A1, implies $% ||\nabla u^{\varepsilon }||/||u^{\varepsilon }||\leq Ch^{-1}$, yielding (\ref% {EstAdaptive2}). The numerical tests suggest that two factors not considered in Theorem 2.1 are significant. The first is whether the finite element space has a nontrivial, divergence-free subspace. The second is the influence of smoothness of the sought solution or its problem data on effective conditioning. In addition, highly refined meshes are used in practical flow simulations and the linear system often has large skew symmetric part. The extension of the analysis herein to include these effects is an open problem. \bibliographystyle{siam}

https://arxiv.org/abs/0710.2357

Overhang

How far off the edge of the table can we reach by stacking $n$ identical, homogeneous, frictionless blocks of length 1? A classical solution achieves an overhang of $1/2 H_n$, where $H_n ~ \ln n$ is the $n$th harmonic number. This solution is widely believed to be optimal. We show, however, that it is, in fact, exponentially far from optimality by constructing simple $n$-block stacks that achieve an overhang of $c n^{1/3}$, for some constant $c>0$.

\section{Introduction} \label{sec:intro} How far off the edge of the table can we reach by stacking $n$ identical, homogeneous, frictionless blocks of length~1? A classical solution achieves an overhang asymptotic to $\frac{1}{2} \ln n$. This solution is widely believed to be optimal. We show, however, that it is exponentially far from optimality by constructing simple $n$-block stacks that achieve an overhang of~$cn^{1/3}$, for some constant $c>0$. The problem of stacking a set of objects, such as bricks, books, or cards, on a tabletop to maximize the overhang is an attractive problem with a long history. J. G. Coffin~\cite{C23} posed the problem in the ``Problems and Solutions" section of this {\sc Monthly}, but no solution was given there. The problem recurred from time to time over subsequent years, e.g., \cite{S53,S54}, \cite{J55}, \cite{E59}. Either deliberately or inadvertently, these authors all seem to have introduced the further restriction that there can be at most one object resting on top of another. Under this restriction, the \emph{harmonic stacks}, described below, are easily seen to be optimal. The classical harmonic stack of size~$n$ is composed of $n$ blocks stacked one on top of the other, with the $i$th block from the top extending by $\frac{1}{2i}$ beyond the block below it. (We assume that the length of each block is~$1$.) The overhang achieved by the construction is clearly $\frac{1}{2} H_n$, where $H_n=\sum_{i=1}^n\frac{1}{i}\sim \ln n$ is the $n$th harmonic number. Both a 3D and a 2D view of the harmonic stack of size~10 are given in Figure~\ref{fig:h10}. The harmonic stack of size~$n$ is balanced since, for every $i< n$, the center of mass of the topmost $i$ blocks lies exactly above the right-hand edge of the $(i+1)$st block, as can be easily verified by induction. Similarly, the center of mass of all the $n$ blocks lies exactly above the right edge of the table. A formal definition of ``balanced'' is given in Definition~\ref{def:balance}. A perhaps surprising and counterintuitive consequence of the harmonic stacks construction is that, given sufficiently many blocks, it is possible to obtain an arbitrarily large overhang! Harmonic stacks became widely known in the recreational math community as a result of their appearance in the \emph{Puzzle-Math} book of Gamow and Stern \cite{GS58} (Building-Blocks, pp.~90--93) and in Martin Gardner's ``Mathematical Games'' section of the November 1964 issue of Scientific American~\cite{G64} (see also~\cite{G71}, Chapter~17: Limits of Infinite Series, p.~167). Gardner refers to the fact that an arbitrarily large overhang can be achieved, using sufficiently many blocks, as the \emph{infinite-offset paradox}. Harmonic stacks were subsequently used by countless authors as an introduction to recurrence relations, the harmonic series, and simple optimization problems; see, e.g.,~\cite{GKP88} pp.~258--260. Hall~\cite{H05} notes that harmonic stacks started to appear in textbooks on physics and engineering mechanics as early as the mid-19th century (see, e.g.,~\cite{M07} p.~341, \cite{P50} pp.~140--141, \cite{W55} p.~183). It is perhaps surprising that none of the sources cited above realizes how limiting is the one-on-one restriction under which the harmonic stacks are optimal. Without this restriction, blocks can be used as counterweights, to balance other blocks. The problem then becomes vastly more interesting, and an exponentially larger overhang can be obtained. \begin{figure}[t] \begin{center} \includegraphics[height=50mm]{06-0595Fig2-1.pdf}\hspace*{1cm} \includegraphics[height=50mm]{06-0595Fig2-2.pdf} \caption{Optimal stacks with 3 and 4 blocks compared with the corresponding harmonic stacks.} \label{fig:opt34} \end{center}\vspace{-5mm} \end{figure} Stacks with a specific small number of blocks that do not satisfy the one-on-one restriction were considered before by several other authors. Sutton \cite{S55}, for example, considered the case of three blocks. One of us set a stacking problem with three uniform thin planks of lengths~2,~3, and~4 for the Archimedeans Problems Drive in 1964 \cite{HP64}. Ainley~\cite{A79} found the maximum overhang achievable with four blocks to be $\frac{15-4\sqrt 2}{8}\sim 1.16789$. The optimal stacks with~3 and~4 blocks are shown, together with the corresponding harmonic stacks, in Figure~\ref{fig:opt34}. Very recently, and independently of our work, Hall \cite{H05} explicitly raises the problem of finding stacks of blocks that maximize the overhang without the one-on-one restriction. (Hall calls such stacks \emph{multiwide stacks}.) Hall gives a sequence of stacks which he claims, without proof, to be optimal. We show, however, that the stacks suggested by him are optimal only for $n\leqslant 19$. The stacks claimed by Hall to be optimal fall into a natural class that we call \emph{spinal stacks}. We show in Section~\ref{sec:spinal} that the maximum overhang achievable using such stacks is only $\ln n+O(1)$. Thus, although spinal stacks achieve, asymptotically, an overhang which is roughly twice the overhang achieved by harmonic stacks, they are still exponentially far from being optimal. Optimal stacks with up to 19 blocks are shown in Figures~\ref{fig:2-10} and~\ref{fig:11-19}. The lightly shaded blocks in these stacks form the \emph{support set}, while the darker blocks form the \emph{balancing set}. The \emph{principal block} of a stack is defined to be the block which achieves the maximum overhang. (If several blocks achieve the maximum overhang, the lowest one is chosen.) The \emph{support set} of a stack is defined recursively as follows: the principal block is in the support set, and if a block is in the support set then any block on which this block rests is also in the support set. The \emph{balancing set} consists of all the blocks that do not belong to the support set. A stack is said to be \emph{spinal} if its support set has a single block in each level, up to the level of the principal block. All the stacks shown in Figures~\ref{fig:2-10} and~\ref{fig:11-19} are thus spinal. \begin{figure*}[t] \centerline{ \includegraphics[height=3cm]{06-0595Fig3-1.pdf}\hspace*{5mm} \includegraphics[height=3cm]{06-0595Fig3-2.pdf}\hspace*{5mm} \includegraphics[height=3cm]{06-0595Fig3-3.pdf}\hspace*{5mm} } \centerline{ \includegraphics[height=3cm]{06-0595Fig3-4.pdf}\hspace*{5mm} \includegraphics[height=3cm]{06-0595Fig3-5.pdf}\hspace*{5mm} \includegraphics[height=3cm]{06-0595Fig3-6.pdf}\hspace*{5mm} } \vspace*{0.5cm} \centerline{ \includegraphics[height=3cm]{06-0595Fig3-7.pdf} \hspace*{3mm} \includegraphics[height=3cm]{06-0595Fig3-8.pdf} \hspace*{3mm} \includegraphics[height=3cm]{06-0595Fig3-9.pdf} \hspace*{3mm} } \vspace*{0.5cm} \caption{Optimal stacks with 2 up to 10 blocks.} \label{fig:2-10} \end{figure*} \begin{figure*}[t] \centerline{ \includegraphics[height=4cm]{06-0595Fig4-1.pdf}\hspace*{5mm} \includegraphics[height=4cm]{06-0595Fig4-2.pdf}\hspace*{5mm} \includegraphics[height=4cm]{06-0595Fig4-3.pdf}\hspace*{5mm} } \centerline{ \includegraphics[height=4cm]{06-0595Fig4-4.pdf}\hspace*{5mm} \includegraphics[height=4cm]{06-0595Fig4-5.pdf}\hspace*{5mm} \includegraphics[height=4cm]{06-0595Fig4-6.pdf}\hspace*{5mm} } \vspace*{0.7cm} \centerline{ \includegraphics[height=4cm]{06-0595Fig4-7.pdf} \hspace*{3mm} \includegraphics[height=4cm]{06-0595Fig4-8.pdf} \hspace*{3mm} \includegraphics[height=4cm]{06-0595Fig4-9.pdf} \hspace*{3mm} } \vspace*{0.5cm} \caption{Optimal stacks with 11 up to 19 blocks.} \label{fig:11-19} \end{figure*} It is very tempting to conclude, as done by Hall \cite{H05}, that the optimal stacks are spinal. Surprisingly, the optimal stacks for $n\geqslant 20$ are not spinal! Optimal stacks containing 20 and 30 blocks are shown in Figure~\ref{fig:20-30}. Note that the right-hand contours of these stacks are not monotone, which is somewhat counterintuitive. For all $n\leqslant 30$, we have searched exhaustively through all combinatorially distinct arrangements of $n$ blocks and found optimal displacements numerically for each of these. The resulting stacks, for $2\leqslant n\leqslant 19$ are shown in Figures~\ref{fig:2-10} and~\ref{fig:11-19}. Optimal stacks with 20 and 30 blocks are shown in Figure~\ref{fig:20-30}. We are confident of their optimality, though we have no formal optimality proofs, as numerical techniques were used. While there seems to be a unique optimal placement of the blocks that belong to the support set of an optimal stack, there is usually a lot of freedom in the placement of the balancing blocks. Optimal stacks seem not to be unique for $n\geqslant 4$. \begin{figure}[t] \centerline{ \includegraphics[height=6.25cm]{06-0595Fig5-1.pdf} \hspace*{3mm} \includegraphics[height=6.25cm]{06-0595Fig5-2.pdf} } \vspace*{0.5cm} \caption{Optimal stacks with 20 and 30 blocks.} \label{fig:20-30} \end{figure} In view of the non-uniqueness and added complications caused by balancing blocks, it is natural to consider \emph{loaded stacks}, which consist only of a support set with some \emph{external forces} (or \emph{point weights}) attached to some of their blocks. We will take the weight of each block to be~$1$; the size, or weight, of a loaded stack is defined to be the number of blocks contained in it plus the sum of all the point weights attached to it. The point weights are not required to be integral. Loaded stacks of weight $40,60,80,$ and $100$, which are believed to be close to optimal, are shown in Figure~\ref{fig:p40-100}. The stack of weight~$100$, for example, contains $49$ blocks in its support set. The sum of all the external forces applied to these blocks is $51$. As can be seen, the stacks become more and more non-spinal. It is also interesting to note that the stacks of Figure~\ref{fig:p40-100} contain small gaps that seem to occur at irregular positions. (There is also a scarcely visible gap between the two blocks at the second level of the 20-block stack of Figure~\ref{fig:20-30}.) \begin{figure*}[t] \centerline{ \includegraphics[height=6.25cm]{06-0595Fig6-1.pdf}\hspace*{3mm} \includegraphics[height=6.25cm]{06-0595Fig6-2.pdf} } \vspace*{0.3cm} \centerline{ \includegraphics[height=6.25cm]{06-0595Fig6-3.pdf}\hspace*{3mm} \includegraphics[height=6.25cm]{06-0595Fig6-4.pdf} } \vspace*{0.4cm} \caption{Loaded stacks, believed to be close to optimal, of weight $40,60,80,$ and $100$.} \label{fig:p40-100} \end{figure*} That harmonic stacks are balanced can be verified using simple center-of-mass considerations. These considerations, however, are not enough to verify the balance of more complicated stacks, such as those in Figures~\ref{fig:2-10}, \ref{fig:11-19}, \ref{fig:20-30}, and~\ref{fig:p40-100}. A formal mathematical definition of ``balanced'' is given in the next section. Briefly, a stack is said to be balanced if there is an appropriate set of forces acting between the blocks of the stacks, and between the blocks at the lowest level and the table, under which all blocks are \emph{in equilibrium}. A block is in equilibrium if the sum of the forces and the sum of the moments acting upon it are both 0. As shown in the next section, the balance of a given stack can be determined by checking whether a given set of \emph{linear inequalities} has a feasible solution. Given the fact that the 3-block stack that achieves the maximum overhang is an \emph{inverted 2-triangle} (see Figure~\ref{fig:opt34}), it is natural to enquire whether larger inverted triangles are also balanced. Unfortunately, the next inverted triangle is already unbalanced and would collapse in the way indicated in Figure~\ref{fig:p23}. Inverted triangles show that simple center-of-mass considerations are not enough to determine the balance of stacks. As an indication that balance issues are not always intuitive, we note that inverted triangles are falsely claimed by Jargodzki and Potter~\cite{JP01} (Challenge~271: A staircase to infinity, p.~246) to be balanced. Another appealing structure, the \emph{$m$-diamond}, illustrated for $m=4$ and~$5$ in Figure~\ref{fig:d45}, consists of a symmetric diamond shape with rows of length $1,2,\ldots,m-1,m,m-1,\ldots,2,1$. Small diamonds were considered by Drummond \cite{D81}. The $m$-diamond uses~$m^2$ blocks and would give an overhang of $m/2$, but unfortunately it is unbalanced for $m\geqslant 5$. A $5$-diamond would collapse in the way indicated in the figure. An $m$-diamond could be made balanced by adding a column of sufficiently many blocks resting on the top block. The methodology introduced in Section~\ref{sec:spinal} can be used to show that, for $m\geqslant 5$, a column of at least $2^m-m^2-1$ blocks would be needed. We can show that this number of blocks is also sufficient, giving a stack of $2^m-1$ blocks with an overhang of $m/2$. It is interesting to note that these stacks are already better than the classical harmonic stacks, as with $n=2^m-1$ blocks they give an overhang of $\frac{1}{2}\log_2(n+1)\simeq 0.693\ln n$. Determining the \emph{exact} overhang achievable using $n$ blocks, for large values of~$n$, seems to be a formidable task. Our main goal in this paper is to determine the \emph{asymptotic growth} of this quantity. Our main result is that there exists a constant $c>0$ such that an overhang of $cn^{1/3}$ is achievable using $n$ blocks. Note that this is an exponential improvement over the $\frac{1}{2}\ln n+O(1)$ overhang of harmonic stacks and the $\ln n+O(1)$ overhang of the best spinal stacks! In a subsequent paper \cite{PPTWZ07}, with three additional coauthors, we show that our improved stacks are asymptotically optimal, i.e., there exists a constant $C>0$ such that the overhang achievable using $n$ blocks is at most $Cn^{1/3}$. Our stacks that achieve an asymptotic overhang of $cn^{1/3}$, for some $c>0$, are quite simple. We construct an explicit sequence of stacks, called \emph{parabolic stacks}, with the $r$th stack in the sequence containing about $2r^3/3$ blocks and achieving an overhang of $r/2$. One stack in this sequence is shown in Figure~\ref{fig:6stack}. The balance of the parabolic stacks is established using an inductive argument. \begin{figure}[t] \begin{center} \includegraphics[height=20mm]{06-0595Fig7-1.pdf}\hspace*{1cm} \includegraphics[height=20mm]{06-0595Fig7-2.pdf} \caption{The balanced inverted 2-triangle and the unbalanced inverted 3-triangle.} \label{fig:p23} \end{center} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[height=30mm]{06-0595Fig8-1.pdf}\hspace*{1cm} \includegraphics[height=30mm]{06-0595Fig8-2.pdf} \caption{The balanced 4-diamond and the unbalanced 5-diamond.} \label{fig:d45} \end{center} \end{figure} \begin{figure}[t] \vspace*{0.25cm} \begin{center} \includegraphics[width=90mm]{06-0595Fig9.pdf} \caption{A parabolic stack consisting of 111 blocks and giving an overhang of $3$.} \label{fig:6stack} \end{center}\vspace{-5mm} \end{figure} The remainder of this paper is organized as follows. In the next section we give formal definitions of all the notions used in this paper. In Section~\ref{sec:spinal} we analyze \emph{spinal stacks}. In Section~\ref{sec:parabolic}, which contains our main results, we introduce and analyze our parabolic stacks. In Section~\ref{sec:general} we describe some experimental results with stacks that seem to improve, by a constant factor, the overhang achieved by parabolic stacks. We end in Section~\ref{sec:conc} with some open problems. \section{Stacks and their balance} \label{sec:prelim} As the maximum overhang problem is physical in nature, our first task is to formulate it mathematically. We consider a 2-dimensional version of the problem. This version captures essentially all the interesting features of the overhang problem. A \emph{block} is a rectangle of length~$1$ and height~$h$ with uniform density and unit weight. (We shall see shortly that the height~$h$ is unimportant.) We assume that the table occupies the quadrant $x,y\leqslant 0$ of the 2-dimensional plane. A \emph{stack} is a collection of blocks. We consider only \emph{orthogonal} stacks in which the sides of the blocks are parallel to the axes, with the length of each block parallel to the $x$-axis. The position of a block is then determined by the coordinate $(x,y)$ of its lower left corner. Such a block occupies the box $[x,x+1]\times[y,y+h]$. A stack composed of~$n$ blocks is specified by the sequence $(x_1,y_1),\ldots,(x_n,y_n)$ of the coordinates of the lower left corners of its blocks. We require each $y_i$ to be a nonnegative integral multiple of~$h$, the height of the blocks. Blocks can touch each other but are not allowed to overlap. The \emph{overhang} of the stack is $1+\max_{i=1}^n x_i$. A block at position $(x_1,y_1)$ \emph{rests on} a block in position $(x_2,y_2)$ if $|x_1-x_2| < 1$ and $y_1-y_2=h$. The \emph{interval of contact} between the two blocks is then $[\max\{x_1,x_2\},1+\min\{x_1,x_2\}]\times \{y_1\}$. A block placed at position $(x,0)$ \emph{rests on the table} if $x < 0$. The interval of contact between the block and the table is $[x,\min\{x+1,0\}]\times\{0\}$. When block~$A$ rests on block~$B$, the two blocks may exert a (possibly infinitesimal) force on each other at every point along their interval of contact. A force is a vector acting at a specified point. By Newton's third law, forces come in opposing pairs. If a force~$f$ is exerted on block~$A$ by block~$B$, at $(x,y)$, then a force~$-f$ is exerted on block~$B$ by block~$A$, again at $(x,y)$. We assume that edges of all the blocks are completely smooth, so that there is no \emph{friction} between them. All the forces exerted on block~$A$ by block~$B$, and vice versa, are therefore \emph{vertical} forces. Furthermore, as there is nothing that holds the blocks together, blocks $A$ and $B$ can \emph{push}, but not pull, one another. Thus, if block~$A$ rests on block~$B$, then all the forces applied on block~$A$ by block~$B$ point upward, while all the forces applied on block~$B$ by block~$A$ point downward, as shown on the left in Figure~\ref{fig:forces}. Similar forces are exerted between the table and the blocks that rest on it. \begin{figure*}[t] \begin{center} \includegraphics[height=35mm]{06-0595Fig10-1.pdf}\hspace*{1cm} \includegraphics[height=35mm]{06-0595Fig10-2.pdf}\hspace*{1cm} \includegraphics[height=35mm]{06-0595Fig10-3.pdf} \caption{Equivalent sets of forces acting between two blocks.} \vspace*{-0.2cm} \label{fig:forces} \end{center} \end{figure*} The distribution of forces acting between two blocks may be hard to describe explicitly. Since all these forces point in the same direction, they can always be replaced by a single \emph{resultant} force acting at a some point within their interval of contact, as shown in the middle drawing of Figure~\ref{fig:forces}. As an alternative, they may be replaced by two resultant forces that act at the endpoints of the contact interval, as shown on the right in Figure~\ref{fig:forces}. Forces acting between blocks and between the blocks and the table are said to be \emph{internal} forces. Each block is also subjected to a downward \emph{gravitational force} of unit size, acting at its center of mass. As the blocks are assumed to be of uniform density, the center of mass of a block whose lower left corner is at $(x,y)$ is at $(x+\frac{1}{2},y+\frac{h}{2})$. A rigid body is said to be in \emph{equilibrium} if the sum of the forces acting on it, and the sum of the \emph{moments} they apply on it, are both zero. A 2-dimensional rigid body acted upon by~$k$ vertical forces $f_1,f_2,\ldots,f_k$ at $(x_1,y_1),\ldots,(x_k,y_k)$ is in equilibrium if and only if $\sum_{i=1}^k {f}_i=0$ and $\sum_{i=1}^k x_i f_i=0$. (Note that $f_1,f_2,\ldots,f_k$ are \emph{scalars} that represent the magnitudes of vertical forces.) A collection of internal forces acting between the blocks of a stack, and between the blocks and the table, is said to be a \emph{balancing} set of forces if the forces in this collection satisfy the requirements mentioned above (i.e., all the forces are vertical, they come in opposite pairs, and they act only between blocks that rest on each other) and if, taking into account the gravitational forces acting on the blocks, all the blocks are in equilibrium under this collection of forces. We are now ready for a formal definition of balance. \begin{definition}[Balance]\label{def:balance} A stack of blocks is \emph{balanced} if and only if it admits a balancing set of forces. \end{definition} Static balance problems of the kind considered here are often \emph{under-determined}, so that the resultants of balancing forces acting between the blocks are usually not uniquely determined. It was the consideration by one of us of balance issues that arise in the game of \emph{Jenga}~\cite{Z02} which stimulated this current work. The following theorem shows that the balance of a given stack can be checked efficiently. \begin{theorem} \label{thm:LP} The balance of a stack containing $n$ blocks can be decided by checking the feasibility of a collection of linear equations and inequalities with $O(n)$ variables and constraints. \end{theorem} \begin{proof} Let $(x_1,y_1),\ldots,(x_n,y_n)$ be the coordinates of the lower left corners of the blocks in the stack. Let $B_i$, for $1\leqslant i\leqslant n$, denote the $i$th block of the stack, and let $B_0$ denote the table. Let $B_i/B_j$, where $0\leqslant i,j\leqslant n$, signify that $B_i$ rests on~$B_j$. If $B_i/B_j$, we let $a_{ij}=\max\{x_i,x_j\}$ and $b_{ij}=\min\{x_i,x_j\}+1$ be the $x$-coordinates of the endpoints of the interval of contact between blocks~$i$ and~$j$. (If $j=0$, then $a_{i0}=x_i$ and $b_{i0}=\min\{x_i+1,0\}$.) For all $i$ and $j$ such that $B_i/B_j$, we introduce two variables $f^0_{ij}$ and $f^1_{ij}$ that represent the resultant forces that act between $B_i$ and $B_j$ at $a_{ij}$ and $b_{ij}$. By Definition~\ref{def:balance} and the discussion preceding it, the stack is balanced if and only if there is a feasible solution to the following set of linear equalities and inequalities: $$\begin{array}{cl} \displaystyle\sum_{j\,:\,B_i/B_j} (f^0_{ij} + f^1_{ij}) \; - \sum_{k\,:\, B_k/B_i} (f^0_{ki} + f^1_{ki}) \;=\; 1 \ , & {\rm for} \ 1\leqslant i\leqslant n ;\\ \displaystyle\sum_{j\,:\,B_i/B_j} (a_{ij}f^0_{ij} + b_{ij}f^1_{ij}) \; - \sum_{k\,:\, B_k/B_i} (a_{ki}f^0_{ki} + b_{ki}f^1_{ki}) \;=\; x_i+\frac{1}{2} \ , & {\rm for} \ 1\leqslant i\leqslant n ;\\[17pt] \displaystyle f^0_{ij},f^1_{ij}\;\geqslant\; 0 \ , & \mbox{for $i,j$ such that $B_i / B_j$.} \end{array}$$ The first $2n$ equations require the forces applied on the blocks to exactly cancel the forces and moments exerted on the blocks by the gravitational forces. (Note that the table is not required to be in equilibrium.) The inequalities $f^0_{ij},f^1_{ij}\geqslant 0$, for every $i$ and $j$ such that~$B_i/B_j$, require the forces applied on $B_i$ by $B_j$ to point upward. As a unit length block can rest on at most two other unit length blocks, the number of variables is at most~$4n$ and the number of constraints is therefore at most~$6n$. The feasibility of such a system of linear equations and inequalities can be checked using \emph{linear programming} techniques. (See, e.g., Schrijver \cite{S98}.) \end{proof} \begin{definition}[Maximum overhang, the function $D(n)$] The maximum overhang that can be achieved using a balanced stack comprising $n$ blocks of length~1 is denoted by $D(n)$. \end{definition} We now repeat the definitions of the principal block, the support set, and the balancing set sketched in the introduction. \begin{definitions}[Principal block, support set, balancing set] The block of a stack that achieves the maximum overhang is the \emph{principal block} of the stack. If several blocks achieve the maximum overhang, the lowest one is chosen. The \emph{support set} of a stack is defined recursively as follows: the principal block is in the support set, and if a block is in the support set then any block on which this block rests is also in the support set. The \emph{balancing set} consists of any blocks that do not belong to the support set. \end{definitions} The blocks of the support sets of the stacks in Figures~\ref{fig:2-10}, \ref{fig:11-19}, and \ref{fig:20-30} are shown in light gray while the blocks in the balancing sets are shown in dark gray. The purpose of blocks in the support set, as its name indicates, is to support the principal block. The blocks in the balancing set, on the other hand, are used to counterbalance the blocks in the support set. As already mentioned, there is usually a lot of freedom in the placement of the blocks of the balancing set. To concentrate on the more important issue of where to place the blocks of the support set, it is useful to introduce the notion of loaded stacks. \begin{definitions}[Loaded stacks, the function $D^*(w)$] A \emph{loaded stack} consists of a set of blocks with some \emph{point weights} attached to them. The \emph{weight} of a loaded stack is the sum of the weights of all the blocks and point weights that participate in it, where the weight of each block is taken to be~$1$. A loaded stack is said to be balanced if it admits a balancing set of forces, as for unloaded stacks, but now also taking into account the point weights. The maximum overhang that can be achieved using a balanced loaded stack of weight~$w$ is denoted by~$D^*(w)$. \end{definitions} Clearly $D^*(n)\geqslant D(n)$, as a standard stack is a trivially loaded stack with no point weights. When drawing loaded stacks, as in Figure~\ref{fig:p40-100}, we depict point weights as external forces acting on the blocks of the stack, with the length of the arrow representing the force proportional to the weight of the point weight. (Since forces can be transmitted vertically downwards through any block, we may assume that point weights are applied only to upper edges of blocks outside any interval of contact.) As the next lemma shows, balancing blocks can always be replaced by point weights, yielding loaded stacks in which all blocks belong to the support set. \begin{lemma} For every balanced stack that contains $k$ blocks in its support set and $n-k$ blocks in its balancing set, there is a balanced loaded stack composed of $k$ blocks, all in the support set, and additional point weights of total weight $n-k$ that achieves the same overhang. \end{lemma} \begin{proof} Consider the set of forces exerted on the support set of the stack by the set of balancing blocks. From the definition of the support set, no block of the support set can rest on any balancing block, therefore the effect of the balancing set can be represented by a set of \emph{downward} vertical forces on the support set, or equivalently by a finite set of point weights attached to the support set with the same total weight as the set of balancing blocks. \end{proof} \begin{figure*}[t] \begin{center} \includegraphics[height=36mm]{06-0595Fig11-1.pdf} \hspace*{7mm} {\includegraphics[height=36mm]{06-0595Fig11-2.pdf}} \caption{Optimal loaded stacks of weight 3 and 5.} \label{fig:l3} \end{center} \end{figure*} Given a loaded stack of integral weight, it is in many cases possible to replace the set of point weights by a set of appropriately placed balancing blocks. In some cases, however, such a conversion of a loaded stack into a standard stack is not possible. The optimal loaded stacks of weight~3,~5, and~7 cannot be converted into standard stacks without decreasing the overhang, as the number of point weights needed is larger than the number of blocks remaining. (The cases of weights~3 and~5 are shown in Figure~\ref{fig:l3}.) In particular, we get that $D^*(3)=\frac{11-2\sqrt{6}}{6}>D(3)=1$. Experiments with optimal loaded stacks lead us, however, to conjecture that the difference $D^*(n)-D(n)$ tends to~$0$ as $n$ tends to infinity. \begin{conjecture}\label{con:D} $D(n)= D^*(n) - o(1)$. \end{conjecture} \section{Spinal stacks} \label{sec:spinal} In this section we focus on a restricted, but quite natural, class of stacks which admits a fairly simple analysis. \begin{definitions}[Spinal stacks, spine] A stack is \emph{spinal} if its support set has just a single block at each level. The support set of a spinal stack is referred to as its \emph{spine}. \end{definitions} The optimal stacks with up to 19 blocks, depicted in Figures~\ref{fig:2-10} and~\ref{fig:11-19}, are spinal. The stacks of Figure~\ref{fig:20-30} are \emph{not} spinal. A stack is said to be \emph{monotone} if the $x$-coordinates of the rightmost blocks in the various levels, starting from the bottom, form an increasing sequence. It is easy to see that every monotone stack is spinal. \begin{definitions}[The functions $S(n)$, $S^*(w)$ and $S^*_k(w)$] Let $S(n)$ be the maximum overhang achievable using a spinal stack of size~$n$. Similarly, let $S^*(w)$ be the maximum overhang achievable using a loaded spinal stack of weight~$w$, and let $S^*_k(w)$ be the maximum overhang achievable using a spinal stack of weight~$w$ with exactly~$k$ blocks in its spine. \end{definitions} \begin{figure*}[t] \begin{center} \includegraphics[height=80mm]{06-0595Fig12.pdf} \caption{A generic loaded spinal stack.} \label{fig:spinal} \end{center} \end{figure*} It is tempting to make the (false) assumption that optimal stacks are spinal. (As mentioned in the introduction, this assumption is implicit in \cite{H05}.) The assumption holds, however, only for $n\leqslant 19$. (See the discussion following Theorem~\ref{thm:lower}.) As spinal stacks form a very natural class of stacks, it is still interesting to investigate the maximum overhang achievable using stacks of this class. A generic loaded spinal stack with $k$ blocks in its spine is shown in Figure~\ref{fig:spinal}. We denote the blocks from top to bottom as $B_1,B_2, \ldots ,B_k$, with $B_1$ being the principal block. We regard the tabletop as $B_{k+1}$. For $1\leqslant i\leqslant k$, the weight attached to the left edge of~$B_i$ is denoted by~$w_i$, and the relative overhang of~$B_i$ beyond~$B_{i+1}$ is denoted by~$d_i$. We define $t_i = \sum_{j=1}^i (1+w_j)$, the total downward force exerted upon $B_{i+1}$ by block $B_i$. We also define $t_0=0$. Note that $t_i=t_{i-1}+w_i+1$, for $1\leqslant i\leqslant k$, and that $t_k=w=k+\sum_{i=1}^k w_i$, the total weight of the loaded stack. The assumptions made in Figure~\ref{fig:spinal}, that each block is supported by a force that acts along the right-hand edge of the block underneath it and that all point weights are attached to the left-hand ends of blocks, are justified by the following lemma. \begin{lemma}\label{lem:L1} In an optimal loaded spinal stack: $(i)$ Each block is supported by a force acting along the right-hand edge of the block underneath it. In particular, the stack is monotone. $(ii)$ All point weights are attached to the left-hand ends of blocks. \end{lemma} \begin{proof} For~$(i)$, suppose there were some block $B_{i+1}$ ($1\leqslant i\leqslant k$) where the resultant force exerted on it from $B_i$ does not go through its right-hand end. If $i<k$ then $B_{i+1}$ could be shifted some distance to the left and $B_i$ together with all the blocks above it shifted to the right in such a way that the resultant force from $B_{i+1}$ on $B_{i+2}$ remains unchanged in position and the stack is still balanced. In the case of $i=k$ (where $B_{k+1}$ is the tabletop), the whole stack could be moved to the right. The result of any such change is a balanced spinal stack with an increased overhang, a contradiction. As an immediate consequence, we get that optimal spinal stacks are monotone. For~$(ii)$, suppose that some block has weights attached other than at its left-hand end. We may replace all such weights by the same total weight concentrated at the left end. The result will be to move the resultant force transmitted to any lower block somewhat to the left. Since the stack is monotone, this change cannot unbalance the stack, and indeed would then allow the overhang to be increased by slightly shifting all blocks to the right; again a contradiction. \end{proof} We next note that for any nonnegative point weights $w_1,w_2,\ldots,w_k\geqslant 0$, there are appropriate positive displacements $d_1,d_2,\ldots,d_k> 0$ for which the generic spinal stack of Figure~\ref{fig:spinal} is balanced. \begin{lemma}\label{lem:L2} A loaded spinal stack with $k$ blocks in its spine that satisfies the two conditions of Lemma~\ref{lem:L1} is balanced if and only if $d_i = \frac{w_i+\frac{1}{2}}{t_i} = 1- \frac{t_{i-1}+\frac{1}{2}}{t_i} $, for $1\leqslant i\leqslant k$. \end{lemma} \begin{proof} The lemma is verified using a simple calculation. The net downward force acting on~$B_i$ is $(w_i+t_{i-1}+1)-t_i=0$, by the definition of~$t_i$. (Recall that $t_i = \sum_{j=1}^i (1+w_j)$.) The net moment acting on~$B_i$, computed relative to the right-hand edge of~$B_i$, is $d_it_i-(\frac{1}{2}+w_i)$, which vanishes if and only if $d_i=\frac{\frac{1}{2}+w_i}{t_i} = 1- \frac{t_{i-1}+\frac{1}{2}}{t_i} $, as required. \end{proof} Note, in particular, that if $w_i=0$ for $1\leqslant i\leqslant k$, then $t_i=i$ and $d_i=\frac{1}{2i}$, and we are back to the classic harmonic stacks. We can now also justify the claim made in the introduction concerning the instability of diamond stacks. Consider the spine of an $m$-diamond. In this case, $d_i=\frac12$ for all~$i$ and so the balance conditions give the equations $t_i= 2t_{i-1}+1$ for $1\leqslant i\leqslant m$. As $t_0 = 0$, we have $t_i\geqslant 2^i -1$ for all~$i$ and hence $t_m\geqslant 2^m -1$. Since $t_m$ is the total weight of the stack, the number of extra blocks required to be added for stability is at least $2^m -1-m^2$, which is positive for $m\geqslant 5$. Next, we characterize the choice of the weights $w_1,w_2,\ldots,w_k$, or alternatively of the total loads $t_1,t_2,\ldots,t_k$, that maximizes the overhang achieved by a spinal stack of total weight~$w$. (Note that $w_i=t_i-t_{i-1}-1$, for $1\leqslant i\leqslant k$.) \begin{lemma} \label{lem:L3} If a loaded spinal stack with total weight $w$ and with $k$ blocks in its spine achieves the maximal overhang of $S^*_k(w)$, then for some $j$ $(1\leqslant j\leqslant k)$ we have $t_i^2=(t_{i-1}+\frac{1}{2})t_{i+1}$, for $1\leqslant i<j$, and $w_i=0$, for $j< i\leqslant k$. \end{lemma} \begin{proof} Let $w_1,w_2,\ldots,w_k$ be the point weights attached to the blocks of an optimal spinal stack with overhang $S^*_k(w)$. For some $i$ satisfying $1\leqslant i<k$ and a small $x$, consider the stack obtained by increasing the point weight at the left-hand end of block~$B_i$ from $w_i$ to $w_i+x$, and decreasing the point weight on~$B_{i+1}$ from $w_{i+1}$ to $w_{i+1}-x$, assuming that $w_{i+1}\geqslant x$. Note that this small perturbation does not change the total weight of the stack. The overhang of the perturbed stack is $$V(x) \;=\; \left(1-\frac{t_{i-1}+\frac12}{t_i+x}\right) + \frac{w_{i+1}-x+\frac12}{t_{i+1}} + \sum_{j\ne i,i+1} \frac{w_j+\frac12}{t_j}\;.$$ The first two terms in the expression above are the new displacements $d_i(x)$ and $d_{i+1}(x)$. Note that all other displacements are unchanged. Differentiating $V(x)$ we get $$V'(x) \;=\; \frac{t_{i-1}+\frac12}{(t_i+x)^2} - \frac{1}{t_{i+1}}\quad{\rm\ and}\quad V'(0) \;=\; \frac{t_{i-1}+\frac12}{t_i^2} - \frac{1}{t_{i+1}}\;.$$ If $w_i=0$ while $w_{i+1}>0$, then $t_{i-1}=t_i-1$ and $t_{i+1}>t_i+1$, which in conjunction with $t_i\geqslant 1$ implies that $V'(0)>0$, contradicting the optimality of the stack. Thus, if in an optimal stack we have $w_i=0$, then also $w_{i+1}=w_{i+2}=\ldots=w_k=0$. If $w_i,w_{i+1}>0$, then we must have $V'(0)=0$, or equivalently $t_i^2=(t_{i-1}+\frac{1}{2})t_{i+1}$, as claimed. \end{proof} The optimality equations given in Lemma~\ref{lem:L3} can be solved numerically to obtain the values of~$S^*_k(w)$ for specific values of $w$ and $k$. The value of $S^*(w)$ is then found by optimizing over~$k$. The optimal loaded spinal stacks of weight~3 and 5, which also turn out to be the optimal loaded stacks of these weights, are shown in Figure~\ref{fig:l3}. The optimality equations of Lemma~\ref{lem:L3} were also used to compute the spines of the optimal stacks with up to 19 blocks shown in Figures~\ref{fig:2-10} and~\ref{fig:11-19}. The spines of the stacks with~3 and~5 blocks were obtained by adding the requirement that no point weight be attached to the topmost block of the spine. A somewhat larger example is given on the top left of Figure~\ref{fig:s100} where the optimal loaded spinal stack of weight~100 is shown. It is interesting to note that the point weights in optimal spinal stacks form an almost arithmetical progression. This observation is used in the proof of Theorem~\ref{thm:spinal-lower}. Numerical experiments suggest that for every $w\geqslant 1$, all the point weights in the spinal stacks with overhang~$S^*(w)$ are nonzero. There are, however, non-optimal values of~$k$ for which some of the bottom blocks in the stack that achieves an overhang of $S^*_k(w)$ have no point weights attached to them. We next show, without explicitly using the optimality conditions of Lemma~\ref{lem:L3}, that $S^*(w)=\ln w + \Theta(1)$. \begin{theorem} \label{thm:spinal-upper} $S^*(w) < \ln w +1$. \end{theorem} \begin{proof} For fixed total weight $w=t_k$ and fixed $k$, the largest possible overhang $S^*_k(w)= \sum_{i=1}^k d_i$ is attained when the conditions of Lemmas~\ref{lem:L1} and~\ref{lem:L2} (and~\ref{lem:L3}) hold. Thus, as $t_{0}=0$, $$\sum_{i=1}^k d_i \;=\; \sum_{i=1}^{k}\left(1- \frac{t_{i-1}+\frac{1}{2}}{t_i} \right) \;<\; k - \sum_{i=2}^{k} \frac{t_{i-1}}{t_i}\;.$$ Putting $x_i=\frac{t_{i-1}}{t_i}$, we see that $$S^*_k(w) \;<\; k - \sum_{i=2}^{k} x_i \;\; {\rm\ and\;\; } \prod_{i=2}^{k}x_i \;=\; \frac{t_1}{t_k} \;\geqslant\; \frac{1}{w}\;.$$ The minimum sum for a finite set of positive real numbers with fixed product is attained when the numbers are equal, hence $$S^*_k(w) \;<\; k - (k-1)w^{-\frac{1}{k-1}}\;.$$ Let $z=\frac{k-1}{\ln w}$, so that $k-1=z\ln w$ and $w^{-\frac{1}{k-1}} = e^{-1/z}$. Then $$S^*_k(w) \;<\; 1 + z\ln w(1-e^{-1/z}) < 1 + \ln w\;,$$ as $z(1-e^{-1/z})\leqslant 1$, for every $z>0$. \end{proof} \begin{corollary} \label{cor:spinal-upper} $S(n) < \ln n + 1$. \end{corollary} We can now describe a construction of loaded spinal stacks which achieves an overhang agreeing asymptotically with the upper bound proved in Theorem~\ref{thm:spinal-upper}. \begin{theorem}\label{thm:spinal-lower} $S^*(w) > \ln w - 1.313$. \end{theorem} \begin{proof} We construct a spine with $k=\lfloor \sqrt{w}\rfloor$ blocks in it with $w_i=2(i-1)$, for $1\leqslant i\leqslant k$. It follows easily by induction that $t_i=i^2$, for $1\leqslant i\leqslant k$. In particular, the total weight of the stack is $t_k=k^2\leqslant w$, as required. By Lemma~\ref{lem:L2}, we get that $$d_i \;=\; \frac{w_i+\frac{1}{2}}{t_i} \;=\; \frac{2(i-1)+\frac{1}{2}}{i^2} \;=\; \frac{2}{i} - \frac{3}{2i^2}\;.$$ Thus, $$S^*(w) \;\geqslant\; \sum_{i=1}^k d_i \;=\; 2\sum_{i=1}^k\frac{1}{i} -\frac{3}{2}\sum_{i=1}^k\frac{1}{i^2}\;=\; 2H_{\lfloor\sqrt{w}\rfloor}-\frac{3}{2}\sum_{i=1}^{\lfloor\sqrt{w}\rfloor}\frac{1}{i^2}\;>\; \ln w +2\gamma - \frac{\pi^2}{4} \;>\; \ln w - 1.313\;.$$ In the above inequality, $\gamma\simeq 0.5772156$ is Euler's gamma. \end{proof} \begin{figure*}[t] \begin{center} \includegraphics[height=80mm]{06-0595Fig13.pdf} \caption{A spinal stack with a shield.} \label{fig:shadow} \end{center} \end{figure*} We next discuss a technique that can be used to convert loaded spinal stacks into standard stacks. This is of course done by constructing balancing sets that apply the required forces on the left-hand edges of the spine blocks. The first step is the placement of \emph{shield} blocks on top of the spine blocks, as shown in Figure~\ref{fig:shadow}. We let $B'_i$, for $0\leqslant i\leqslant k-1$, be the shield block placed on top of spine block $B_{i+1}$ and alongside spine block $B_i$ for $i>0$. We let $y_i$ be the $x$-coordinate of the left edge of~$B'_i$, for $1\leqslant i\leqslant k-1$. Note that $x_{i+1}-1 < y_i\leqslant x_i-1$, where $x_i$ is the $x$-coordinate of the left edge of~$B_i$. Shield block~$B'_i$ applies a downward force of $w_{i+1}$ on $B_{i+1}$. The force is applied at $x_{i+1}$, i.e., at the left edge of $B_{i+1}$. Block~$B'_i$ also applies a downward force of $u_{i+1}$ on $B'_{i+1}$ at $z_{i+1}$, where $y_i\leqslant z_{i+1}\leqslant y_{i+1}+1$. Similarly, block $B'_{i-1}$ applies a downward force of $u_i$ on $B'_i$ at $z_i$. Finally a downward external force of $v_i$ is applied on the left edge of $B'_i$. The goal of the shield blocks is to aggregate the forces that should be applied on the spine blocks and to replace them by a set of fewer \emph{integral} forces that are to be applied on the shield blocks. We will therefore place the shield blocks and choose the forces $u_i$ and their positions in such a way that most of the $v_i$ will be~$0$. (This is why we use dashed arrows to represent the $v_i$ forces in Figure~\ref{fig:shadow}.) The shield blocks are in equilibrium provided that the following balance conditions are satisfied: $$\begin{array}{c} u_i+v_i+1 \;=\; u_{i+1}+w_{i+1}\;,\\ z_iu_i+y_iv_i+(y_i+\frac12) \;=\; z_{i+1}u_{i+1}+x_{i+1}w_{i+1}\;,\\ \end{array}$$ for $1\leqslant i\leqslant k-1$. (We define $u_k=0$.) It is easy to see that if $u_{i+1},w_{i+1},x_{i+1},y_{i+1},$ and $z_{i+1}$ are set, then any choice of $v_i$ uniquely determines $u_i$ and $z_i$. The choice is feasible if $u_i,v_i\geqslant 0$ and $y_{i-1}\leqslant z_i$. \begin{figure}[t] \begin{center} \includegraphics[width=8cm]{06-0595Fig14-1.pdf} \includegraphics[width=8cm]{06-0595Fig14-2.pdf}\\[0.8cm] \includegraphics[width=8cm]{06-0595Fig14-3.pdf} \caption{Optimal loaded spinal stack of weight 100 (top left), with shield added (top right) and with a complete balancing set added (bottom).} \label{fig:s100} \end{center}\vspace{-5mm} \end{figure} In our constructions, we used the following heuristic to place the shield blocks and specify the forces between them. We start placing the shield blocks from the bottom up. In most cases, we choose $y_i=x_i-1$ and $v_i=0$, i.e., $B'_i$ is adjacent to $B_i$ and no external force is applied to it. Eventually, however, it may happen that $z_{i+1}<x_{i}-1$, which makes it impossible to place~$B'_{i}$ adjacent to~$B_{i}$ and still apply the force~$u_{i+1}$ down on~$B_{i+1}$ at~$z_{i+1}$. In that case we choose $y_i=z_{i+1}$. A more significant event, that usually occurs soon after the previous event, is when $z_{i+1}\leqslant x_{i+1}-1$, in which case no placement of $B'_i$ allows it to apply the forces $u_{i+1}$ and $v_{i+1}$ on $B'_{i+1}$ and $B_{i+1}$ at the required positions, as they are at least a unit distance apart. In this case, we introduce a nonzero, integral, external force~$v_{i+1}$ as follows. We let $v_{i+1}=\lfloor (1-z_{i+1}+y_{i+1})u_{i+1} \rfloor$ and then recompute $u_{i+1}$ and $z_{i+1}$. It is easy to check that $u_{i+1},v_{i+1}\geqslant 0$ and that $y_i\leqslant z_{i+1}\leqslant y_{i+1} +1$. If we now have $z_{i+1}>x_{i+1}-1$, then the process can continue. Otherwise we stop. In our experience, we were always able to use this process to place all the shield blocks, except for a very few top ones. The~$v_i$ forces left behind tend to be few and far apart. When this process is applied, for example, on the optimal loaded spinal stack of weight~$100$, only one such external force is needed, as shown in the second diagram of Figure~\ref{fig:s100}. The nonzero $v_i$'s can be easily realized by erecting appropriate towers, as shown at the bottom of Figure~\ref{fig:s100}. The top part of the balancing set is then designed by solving a small linear program. We omit the fairly straightforward details. The overhang achieved by the spinal stack shown at the bottom of Figure~\ref{fig:s100} is about $3.6979$, which is a considerable improvement on the $2.5937$ overhang of a 100-block harmonic stack, but is also substantially less than the $4.23897$ overhang of the non-spinal loaded stack of weight 100 given in Figure~\ref{fig:p40-100}. Using the heuristic described above we were able to fit appropriate balancing sets for all optimal loaded spinal stacks of integer weight~$n$, for every $n\leqslant 1000$, with the exception of $n=3,5,7$. We conjecture that the process succeeds for every $n\ne 3,5,7$. \begin{conjecture} $S(n)= S^*(n)$ for $n\neq 3,5,{\rm or\ }7$. \end{conjecture} \begin{figure}[t] \begin{center} \includegraphics[width=85mm]{06-0595Fig15.pdf} \caption{A $6$-stack composed of $r$-slabs, for $r=2,3,\ldots,6$, and an additional block.} \label{fig:slabs2-6} \end{center}\vspace{-5mm} \end{figure} \section{Parabolic stacks} \label{sec:parabolic} We now give a simple explicit construction of $n$-block stacks with an overhang of about $(3n/16)^{1/3}$, an \emph{exponential} improvement over the $O(\log n)$ overhang achievable using spinal stacks in general and the harmonic stacks in particular. Though the stacks of this sequence are not optimal (see the empirical results of the next section), they are within a constant factor of optimality, as will be shown in a subsequent paper \cite{PPTWZ07}. The stacks constructed in this section are what we term \emph{brick-wall} stacks. The blocks in each row are contiguous, and each is centered over the ends of blocks in the row beneath. This resembles the simple ``stretcher-bond'' pattern in real-life bricklaying. Overall the stacks have a symmetric roughly parabolic shape, hence the name, with vertical axis at the table edge and a brick-wall structure. An illustration of a $111$-block \emph{parabolic} $6$-stack with overhang~$3$ was given in Figure~\ref{fig:6stack}. \begin{figure}[t] \begin{center} \vspace*{0.3cm} \includegraphics[scale=0.6]{06-0595Fig16.pdf} \caption{A $6$-slab with a grey $5$-slab contained in it.} \label{fig:slab6} \end{center} \end{figure} An \emph{$r$-row} is a row of $r$ adjacent blocks, symmetrically placed with respect to $x=0$. An \emph{$r$-slab}, for $r\geqslant 2$, has height $2r-3$ and consists of alternating $r$-rows and $(r-1)$-rows, starting and finishing with $r$-rows. An $r$-slab therefore contains $r(r-1)+(r-1)(r-2)=2(r-1)^2$ blocks. Figure~\ref{fig:slabs2-6} shows $r$-slabs, for $r=2,3,\ldots,6$. A \emph{parabolic $d$-stack}, or just \emph{$d$-stack}, for short, is a $d$-slab on a $(d-1)$-slab on \ldots\ on a 2-slab on a single block. The slabs shown in Figure~\ref{fig:slabs2-6} thus compose a $6$-stack. \begin{lemma} A parabolic $d$-stack contains $\frac{d(d-1)(2d-1)}{3} + 1$ blocks and, if balanced, has an overhang of~$\frac{d}{2}$. \end{lemma} \begin{proof} The number of blocks contained in a $d$-stack is $1+\sum_{r=2}^d 2(r-1)^2 = 1 + \frac{d(d-1)(2d-1)}{3}$. The overhang achieved, if the stack is balanced, is half the width of the top row, i.e., $\frac{d}{2}$. \end{proof} In preparation for proving the balance of parabolic stacks, we show in the next lemma that a slab can concentrate a set of forces acting on its top together with the weights of its own blocks down into a narrower set of forces acting on the row below it. The lemma is illustrated in Figure~\ref{fig:slab6}. \begin{lemma} \label{lem:slab} For any $g\geqslant 0$, an $r$-slab with forces of $g,2g,2g,\ldots ,2g,g$ acting downwards onto its top row at positions $-\frac{r}{2},-\frac{r-2}{2},-\frac{r-4}{2}, \dots ,\frac{r-2}{2},\frac{r}{2}$, respectively, can be stabilized by applying a set of upward forces $g',2g',2g',\ldots ,2g',g'$, where $g'=\frac{r}{r-1}g+r-1$, on its bottom row at positions $-\frac{r-1}{2},-\frac{r-3}{2}, \dots ,\frac{r-3}{2},\frac{r-1}{2}$, respectively. \end{lemma} \begin{proof} The proof is by induction on $r$. For $r=2$, a $2$-slab is just a $2$-row, which is clearly balanced with downward forces of $g,2g,g$ at $-1,0,1$ and upward forces of $2g+1,2g+1$ at $-\frac{1}{2},\frac{1}{2}$, when half of the downward force $2g$ acting at $x=0$ is applied on the right-hand edge of the left block and the other half applied on the left-hand edge of the right block. For the induction step, we first observe that for any $r\geqslant 2$ an $(r+1)$-slab can be regarded as an $r$-slab with an $(r+1)$-row added above and below and with an extra block added at each end of the $r-2$ rows of length $r-1$ of the $r$-slab. The 5-slab (shaded) contained in a 6-slab together with the added blocks is shown in Figure~\ref{fig:slab6}. \begin{figure}[t] \begin{center} \includegraphics[scale=0.7]{06-0595Fig17.pdf} \caption{The proof of Lemma~\ref{lem:slab}.} \label{fig:slab-induction} \end{center} \end{figure} Suppose the statement of the lemma holds for $r$-slabs and consider an $(r+1)$-slab with the supposed forces acting on its top row. Let $f=g/r$, so that $g=rf$. As in the basis of the induction, the top row can be balanced by $r+1$ equal forces of $2rf+1$ from below (the 1 is for the weight of the blocks in the top row) acting at positions $-\frac{r}{2},-\frac{r-2}{2},\ldots ,\frac{r-2}{2},\frac{r}{2}$. As $$2rf+1 \;=\; (r-1)f+((r+1)f+1) \;=\; 2(r-1)f+2f+1\;,$$ we can express this constant sequence of $r+1$ forces as the sum of the following two force sequences: $$\begin{array}{ccccccccccccc} (r-1)f &,& 2(r-1)f &,& 2(r-1)f &,& \ldots &,& 2(r-1)f &,& 2(r-1)f &,& (r-1)f\\ (r+1)f+1 &,& 2f+1 &,& 2f+1 &,& \ldots &,& 2f+1 &,& 2f+1 &,& (r+1)f+1\\ \end{array}$$ The forces in the first sequence can be regarded as acting on the $r$-slab contained in the $(r+1)$-slab, which then, by the induction hypothesis, yield downward forces on the bottom row of $$rf+r-1\;\;,\;\;2rf+2(r-1)\;\;,\;\;\ldots \;\;,\;\;2rf+2(r-1)\;\;,\;\;rf+r-1$$ at positions $-\frac{r-1}{2},-\frac{r-3}{2},\dots\ ,\frac{r-3}{2},\frac{r-1}{2}$. The forces of the second sequence, together with the weights of the outermost blocks of the $(r+1)$-rows, are passed straight down through the rigid structure of the $r$-slab to the bottom row. The combined forces acting down on the bottom row are now $$(r+1)f+r{-}1\;,\;rf+r{-}1\;,\;2f+1\;,\;2rf+2(r{-}1)\;,\;2f+1\;,\;\ldots \;,\;2f+1\;,\;rf+r{-}1\;,\;(r+1)f+r{-}1$$ at positions $-\frac{r}{2},-\frac{r-1}{2},\dots\ ,\frac{r-1}{2},\frac{r}{2}$. The bottom row is in equilibrium when the sequence of upward forces $$(r+1)f+r\;\;,\;\;2(r+1)f+2r\;\;,\;\;2(r+1)f+2r\;\;,\;\;\ldots \;\;,\;\;2(r+1)f+2r\;\;,\;\;2(r+1)f+2r\;\;,\;\;(r+1)f+r$$ is applied on the bottom row at positions $-\frac{r}{2},-\frac{r-2}{2},\dots\ ,\frac{r-2}{2},\frac{r}{2}$, as required. \end{proof} \begin{theorem} \label{thm:construction} For any $d\geqslant 2$, a parabolic $d$-stack is balanced, contains $\frac{d(d-1)(2d-1)}{3} + 1$ blocks, and has an overhang of $\frac{d}{2}$. \end{theorem} \begin{proof} The balance of a parabolic $d$-stack follows by a repeated application of Lemma~\ref{lem:slab}. For $2\leq r\leq d$, let $g(r)$ denote the value of $g$ in Lemma~\ref{lem:slab} for the $r$-slab in the $d$-stack. Although the argument does not rely on the specific values that $g(r)$ assumes, it can be verified that $g(r)=\frac{1}{r}\sum_{i=r}^{d-1} i^2$. Note that $g(d)=0$, as no downward forces are exerted on the top row of the $d$-slab, which is also the top row of the $d$-stack, and that $g(r-1)=\frac{r}{r-1}g(r)+r-1$, as required by Lemma~\ref{lem:slab}. \end{proof} \begin{theorem} \label{thm:lower} $D(n)\geqslant (\frac{3n}{16})^{1/3} - \frac14$ for all $n$. \end{theorem} \begin{proof} Choose $d$ so that $\frac{(d-1)d(2d-1)}{3}+1 \leqslant n \leqslant \frac{d(d+1)(2d+1)}{3}$. Then Theorem~\ref{thm:construction} shows that a $d$-stack yields an overhang of $d/2$ and can be constructed using $n$ or fewer blocks. Any extra blocks can be just placed in a vertical pile in the center on top of the stack without disturbing balance (or arbitrarily scattered on the table). Hence $$n < \frac{2(d+\frac{1}{2})^3}{3}\ {\rm\quad and\ so\quad } \ D(n) \geqslant d/2> \left(\frac{3n}{16}\right)^{1/3}-\frac{1}{4}\;. \vspace*{-5pt}$$ \end{proof} In Section~\ref{sec:spinal} we claimed that optimal stacks are spinal \emph{only} for $n\leqslant 19$. We can justify this claim for $n\leqslant 30$ by exhaustive search, while comparison of the lower bound from Theorem~\ref{thm:lower} with the upper bound of $S(n) < 1+\ln n$ from Corollary~\ref{cor:spinal-upper} deals with the range $n\geqslant 5000$. The intermediate values of $n$ can be covered by combining a few explicit constructions, such as the stack shown in Figure~\ref{fig:vase95}, with numerical bounds using Lemma~\ref{lem:L3}. \begin{figure}[t] \begin{center} \includegraphics[scale=0.3]{06-0595Fig18.pdf} \caption{Incremental block-by-block construction of modified parabolic stacks.} \label{fig:modified} \end{center} \end{figure} Can parabolic $d$-stacks be built incrementally by laying one brick at a time? The answer is no, as the bottom three rows of a parabolic stack form an unbalanced inverted 3-triangle. The inverted 3-triangle remains unbalanced when the first block of the fourth row is laid down. Furthermore, the bottom six rows, on their own, are also not balanced. These, however, are the only obstacles to an incremental row-by-row and block-by-block construction of parabolic stacks and they can be overcome by the modified parabolic stacks shown in Figure~\ref{fig:modified}. We simply omit the lowest block and move the whole stack half a block length to the left. The bricks can now be laid row by row, going in each row from the center outward, alternating between the left and right sides, with the left side, which is over the table, taking precedence. The numbers in Figure~\ref{fig:modified} indicate the order in which the blocks are laid. Thus, unlike with harmonic stacks, it is possible to construct an arbitrarily large overhang using sufficiently many blocks, \emph{without} knowing the desired overhang in advance. \section{General stacks} \label{sec:general} We saw in Section~\ref{sec:prelim} that the problem of checking whether a given stack is balanced reduces to checking the feasibility of a system of linear equations and inequalities. Similarly, the minimum total weight of the point weights that are needed to stabilize a given loaded stack can be found by solving a linear program. Finding a stack with a given number of blocks, or a loaded stack with a given total weight, that achieves maximum overhang seems, however, to be a much harder computational task. To do so, one should, at least in principle, consider all possible combinatorial stack structures and for each of them find an optimal placement of the blocks. The \emph{combinatorial structure} of a stack specifies the contacts between the blocks of the stack, i.e., which blocks rest on which, and in what order (from left to right), and which rest on the table. The problem of finding a (loaded) stack with a given combinatorial structure with maximum overhang is again not an easy problem. As both the forces and their locations are now unknowns, the problem is not linear, but rather a constrained quadratic programming problem. Though there is no general algorithm for efficiently finding the global optimum of such constrained quadratic programs, such problems can still be solved in practice using nonlinear optimization techniques. For stacks with a small number of blocks, we enumerated all possible combinatorial stack structures and numerically optimized each of them. For larger numbers of blocks this approach is clearly not feasible and we had to use various heuristics to cut down the number of combinatorial structures considered. The stacks of Figures~\ref{fig:2-10}, \ref{fig:11-19}, \ref{fig:20-30}, and~\ref{fig:p40-100} were found using extensive numerical experimentation. The stacks of Figures~\ref{fig:2-10}, \ref{fig:11-19}, and \ref{fig:20-30} are optimal, while the stacks of Figure~\ref{fig:p40-100} are either optimal or very close to being so. The collections of forces that stabilize the loaded stacks of Figure~\ref{fig:p40-100} (and the loaded stacks contained in the stacks of Figures~\ref{fig:2-10}, \ref{fig:11-19}, and \ref{fig:20-30}) have the following interesting properties. First, the stabilizing collections of forces of these stacks are unique. Second, almost all downward forces in these collections are applied at the edges of blocks. The only exceptions occur when a downward force is applied on a \emph{right-protruding} block, i.e., a rightmost block in a level that protrudes beyond the rightmost block of the level above it. In addition, all point weights are placed on the left-hand edges of \emph{left-protruding} blocks, where left-protruding blocks are defined in an analogous way. The table, of course, supports the (only) block that rests on it at its right-hand edge. A collection of stabilizing forces that satisfies these conditions is said to be \emph{well-behaved}. A schematic description of well-behaved collections of stabilizing forces is given in Figure~\ref{fig:schematic}. The two right-protruding blocks are shown with a slightly lighter shading. A right-protruding block is always adjacent to the block on its left. We conjecture that forces that balance optimal loaded stacks are always well-behaved. \begin{figure}[t] \centerline{ \includegraphics[height=80mm]{06-0595Fig19.pdf}} \caption{A schematic description of a well-behaved set of stabilizing forces.} \label{fig:schematic} \end{figure} A useful property of well-behaved collections of stabilizing forces is that the total weight of the stack and the positions of its blocks uniquely determine all the forces in the collection. This follows from the fact that each block has either two downward forces acting upon it at specified positions, namely at its two edges, or just a single force in an unspecified position. Given the upward forces acting on a block, the downward force or forces acting upon it can be obtained by solving the force and moment equations of the block. All the forces in the collection can therefore be determined in a bottom-up fashion. We conducted most of our experiments, on blocks with more than 30 blocks, on loaded stacks balanced by well-behaved sets of balancing forces. We saw in Section~\ref{sec:prelim} that loaded stacks of total weight~3, 5, and~7 achieve a larger overhang than the corresponding unloaded stacks, simply because the number of blocks available for use in their balancing sets is smaller than the number of point weights to be applied. The loaded stacks of Figure~\ref{fig:p40-100} exhibit another trivial impediment to the conversion of loaded stacks into standard ones: the point weight to be applied in the lowest position has magnitude less than~$1$. Thus, these stacks can be converted into standard ones only after making some small adjustments. These adjustments have only a very small effect on the overhang achieved. Thus, although we believe that the difference between the maximum overhangs achieved by loaded and unloaded stacks is bounded by a small universal constant, we also believe that for most sizes, loaded stacks yield slightly larger overhangs. Although the placements of the blocks in the optimal, or close to optimal, stacks of Figure~\ref{fig:p40-100} are somewhat irregular, with some small (essential) gaps between blocks of the same layer, at a high level, these stacks seem to resemble brick-wall stacks, as defined in Section~\ref{sec:parabolic}. This, and the fact that brick-wall stacks were used to obtain the $\Omega(n^{1/3})$ lower bound on the maximum overhang, indicate that it might be interesting to investigate the maximum overhang that can be achieved using brick-wall stacks. The parabolic brick-wall stacks of Section~\ref{sec:parabolic} were designed to enable a simple inductive proof of their balance. Parabolic stacks, however, are far from being optimal brick-wall stacks. The balanced 95-block symmetric brick-wall stack with an overhang of 4 depicted in Figure~\ref{fig:vase95}, for example, contains fewer blocks and achieves a larger overhang than that achieved by the 111-block overhang-3 parabolic stack of Figure~\ref{fig:6stack}. \begin{figure}[t] \centerline{ \includegraphics[height=80mm]{06-0595Fig20.pdf}} \caption{A 95-block symmetric brick-wall stack with overhang 4.} \label{fig:vase95} \end{figure} \begin{figure}[t] \centerline{ \includegraphics[height=80mm]{06-0595Fig21-1.pdf}\hspace*{1cm} \includegraphics[height=80mm]{06-0595Fig21-2.pdf}} \caption{A schematic description of well-behaved collections of forces that stabilize symmetric and asymmetric brick-wall stacks.} \label{fig:symfor} \end{figure} Loaded brick-wall stacks are especially easy to experiment with. Empirically, we have again discovered that the minimum weight collections of forces that balance them turn out to be well-behaved, in the formal sense defined above. When the brick-wall stacks are \emph{symmetric} with respect to the $x=0$ axis, and have a flat top, point weights are attached only to blocks at the top layer of the stack. Protruding blocks, both on the left and on the right, then simply serve as \emph{props}, while all other blocks are perfect \emph{splitters}, i.e., they are supported at the center of their lower edge and they support other blocks at the two ends of their upper edge. In non-symmetric brick-wall stacks it is usually profitable to use the left-protruding blocks as splitters and not as props, attaching point weights to their left ends. A schematic description of well-behaved forces that stabilize symmetric and asymmetric brick-wall stacks is shown in Figure~\ref{fig:symfor}. As can be seen, all forces in such well-behaved collections are linear functions of~$w$, the total weight of the stack. This allows us, in particular, to find the minimum total weight needed to stabilize a brick-wall loaded stack \emph{without} solving a linear program. We simply choose the smallest total weight~$w$ for which all forces are nonnegative. This observation enabled us to experiment with huge symmetric and asymmetric brick-wall stacks. The best symmetric loaded brick-wall stacks with overhangs 10 and 50 that we have found are shown in Figures~\ref{fig:s10} and~\ref{fig:width100}. Their total weights are about 1151.76 and 115,467, respectively. The blocks in the larger stack are so small that they are not shown individually. We again believe that these stacks are close to being the optimal stacks of their kind. They were found using a local search approach. In particular, these stacks cannot be improved by widening or narrowing layers, or by adding or removing single layers. Essentially the same symmetric stacks were obtained by starting from almost any initial stack and repeatedly improving it by widening, narrowing, adding, and removing layers. As can be seen from Figures~\ref{fig:s10} and~\ref{fig:width100}, the shapes of optimal symmetric loaded stacks, after suitable scaling, seem to tend to a limiting curve. This curve, which we have termed the \emph{vase}, is similar to but different from that of an inverted normal distribution. We have as yet no conjecture for its equation. We have conducted similar experiments with asymmetric loaded brick-wall stacks. The best such stack with overhang~10 that we have found is shown in Figure~\ref{fig:a10}. Its total weight of about 1128.84 is about $3.38\%$ less than the weight of the symmetric stack of Figure~\ref{fig:s10}. The scaled shapes of optimal asymmetric loaded brick-wall stacks seem again to tend to a limiting curve which we have termed the \emph{oil lamp}. We again have no conjecture for its equation. \begin{figure}[t] \centerline{ \includegraphics[height=80mm]{06-0595Fig22.pdf}} \caption{A symmetric loaded brick-wall stack with an overhang of 10.} \label{fig:s10} \end{figure} \begin{figure}[t] \centerline{ \includegraphics[height=8cm]{06-0595Fig23.pdf}} \caption{A scaled outline of a loaded brick-wall stack with an overhang of 50.} \label{fig:width100} \end{figure} \begin{figure}[t] \centerline{ \includegraphics[height=80mm]{06-0595Fig24.pdf}} \caption{An asymmetric loaded brick-wall stack with an overhang of 10.} \label{fig:a10} \end{figure} \section{Open problems} \label{sec:conc} \ignore{ We have revisited the classic overhang problem and answered some of the questions that were latent there. We have shown that the overhang achievable with~$n$ blocks is exponentially larger than was previously supposed. } Some intriguing problems still remain open. In a subsequent paper \cite{PPTWZ07}, we show that the $\Omega(n^{1/3})$ overhang lower bound presented here is optimal, up to a constant factor, but it would be interesting to determine the largest constant $c_{over}$ for which overhangs of $(c_{over}-o(1))n^{1/3}$ are possible. Can this constant $c_{over}$ be achieved using stacks that are simple to describe, e.g., brick-wall stacks, or simple modifications of them, such as brick-wall stacks with adjacent levels having a displacement other than~$\frac12$, or small gaps left between the blocks of the same level? What are the limiting vase and oil lamp curves? Do they yield, asymptotically, the maximum overhangs achievable using symmetric and general stacks? Another open problem is the relation between the maximum overhangs achievable using loaded and unloaded stacks. We believe, as expressed in Conjecture~\ref{con:D}, that the difference between these two quantities tends to~$0$ as the size of the stacks tends to infinity. We also conjecture that $D^*(n)-D(n)\leqslant D^*(3)-D(3)=\frac{5-2\sqrt{6}}{6}\simeq 0.017$ , for every $n\geqslant 1$. Our notion of balance, as defined formally in Section~\ref{sec:prelim}, allows stacks to be precarious: stacks that achieve maximum overhang are always on the verge of collapse. It is not difficult, however, to define more robust notions of balance, where there is some \emph{stability}. In one such natural definition, a stack is \emph{stable} if there is a balancing set of forces in which none of the forces acts at the edge of any block. We note in passing that Farkas' lemma, or the theory of linear programming duality (see \cite{S98}), can be used to derive an equivalent definition of stability: a stack is stable if and only if every feasible infinitesimal motion of the blocks of the stack increases the total potential energy of the system. This requirement of stability raises some technical difficulties but does not substantially change the nature of the overhang problem. Our parabolic $d$-stacks, for example, can be made stable by adding a $(d-1)$-row symmetrically placed on top. The proof of this is straightforward but not trivial. We believe that for any $n\neq 3$, the loss in the overhang due to this stricter definition is infinitesimal. Our analysis of the overhang problem was made under the \emph{no friction} assumption. All the forces considered were therefore vertical. The presence of friction introduces horizontal forces and thus changes the picture completely, as also observed by Hall \cite{H05}. We can show that there is a fixed coefficient of friction such that the inverted triangles are all balanced, and so achieve overhang of order $n^{1/2}$. \bibliographystyle{plain}

https://arxiv.org/abs/1109.4676

A note on heavy cycles in weighted digraphs

A weighted digraph is a digraph such that every arc is assigned a nonnegative number, called the weight of the arc. The weighted outdegree of a vertex $v$ in a weighted digraph $D$ is the sum of the weights of the arcs with $v$ as their tail, and the weight of a directed cycle $C$ in $D$ is the sum of the weights of the arcs of $C$. In this note we prove that if every vertex of a weighted digraph $D$ with order $n$ has weighted outdegree at least 1, then there exists a directed cycle in $D$ with weight at least $1/\log_2 n$. This proves a conjecture of Bollobás and Scott up to a constant factor.

\section{Introduction} We use Bondy and Murty \cite{Bondy_Murty} for terminology and notation not defined here, and consider digraphs containing no multiple arcs only. Let $D$ be a digraph. The number of vertices and loops of $D$ are denoted by $n(D)$ and $r(D)$, respectively. We call $D$ a {\em weighted digraph} if each arc $a$ of $D$ is assigned a nonnegative number $w_D(a)$, called the {\em weight} of $a$. For a subdigraph $H$ of $D$, $V(H)$ and $A(H)$ are used to denote the {set} of vertices and arcs of $H$, respectively. The {\em weight} of $H$ is defined by $$ w_D(H)=\sum_{a\in A(H)}w_D(a). $$ For a vertex $v\in V(D)$, $N_H^+(v)$ denotes the set, and $d_H^+(v)$ the number, of vertices in $H$ to which there is an arc from $v$. We define the {\em weighted outdegree} of $v$ in $H$ by $$ d_H^{w+}(v)=\sum_{h\in N_H^+(v)}w_D(vh). $$ When no confusion occurs, we will denote $w_D(a)$, $w_D(H)$, $N_D^+(v)$, $d_D^+(v)$ and $d_D^{w+}(v)$ by $w(a)$, $w(H)$, $N^+(v)$, $d^+(v)$ and $d^{w+}(v)$, respectively. An unweighted digraph $D$ can be regarded as a weighted digraph in which each arc $a$ is assigned weight $w(a)=1$. Thus, in an unweighted digraph, $d^{w+}(v)=d^+(v)$ for every vertex $v$, and the weight of a subdigraph is simply the number of its arcs. A loopless digraph is one containing no loops. Let $D$ be a loopless digraph such that every vertex of $D$ has outdegree at least $d$. {It is easy to see} that $D$ contains a {directed} cycle with length at least $d+1$. For weighted digraphs, Bondy {\cite{Bondy}} conjectured that if every vertex in a weighted loopless digraph has weighted outdegree at least 1, then the digraph contains a {directed} cycle of weight at least 1. This conjecture was disproved by T. Spencer of Nebraska {(See \cite{Bollobas_Scott})}. Bollob\'{a}s and Scott \cite{Bollobas_Scott} gave a lower bound {on} the weight of heaviest directed cycles in a weighted loopless digraph under the weighted outdegree condition. \begin{theorem}[Bollob\'{a}s and Scott \cite{Bollobas_Scott}] Let $D$ be a weighted loopless digraph with {$n\geq 2$} vertices. If $d^{w+}(v)\geq 1$ for every vertex $v\in V(D)$, then $D$ contains a {directed} cycle $C$ such that $w(C)\geq(24n)^{-1/3}$. \end{theorem} For an upper bound, Bollob\'{a}s and Scott constructed a class of digraphs with minimum weighted outdegree at least 1 such that the maximum weight of cycles in these digraphs is at most $c\log_2\log_2n/\log_2n$, where $c$ is a constant and $n$ is the order of the digraph. {As remarked} in \cite{Bollobas_Scott}, it seems likely that $n^{-1/3}$ is much too small. Bollob\'{a}s and Scott proposed the following conjecture. \begin{conjecture}[Bollob\'{a}s and Scott \cite{Bollobas_Scott}] Let $D$ be a weighted loopless digraph with {$n\geq 2$} vertices. If $d^{w+}(v)\geq1$ for every vertex $v\in V(D)$, then $D$ contains a {directed} cycle $C$ such that $w(C)\geq 2/\log_2 n$. \end{conjecture} In this paper, we {prove the conjecture up to a constant factor}. \begin{theorem} Let $D$ be a weighted loopless digraph with {$n\geq 2$} vertices. If $d^{w+}(v)\geq 1$ for every vertex $v\in V(D)$, then $D$ contains a {directed} cycle $C$ such that $w(C)\geq 1/\log_2 n$. \end{theorem} {In fact, we can prove the following stronger assertion.} \begin{theorem} Let $D$ be a weighted digraph with {$n\geq 1$} vertices and $r$ loops. If $d^{w+}(v)\geq 1$ for every vertex $v\in V(D)$, then $D$ contains a {directed} cycle $C$ such that $w(C)\geq 1/\log_2 (n+r)$. \end{theorem} We postpone the proof of Theorem 3 to the next section. \section{Proof of Theorem 3} We use induction on $n$. If $D$ {has} only one vertex, denote it by $v$. By $d^{w+}(v)\geq 1$, we have $A(D)=\{vv\}$, $w(vv)\geq 1$ and $r=1$. Thus $C=vv$ is a {directed} cycle with weight at least 1. The result is true. Now, we suppose that $D$ has $n\geq 2$ vertices and $r$ loops. \begin{case} $D$ is not {strongly} connected. \end{case} Let $D'$ be a {strongly} connected component of $D$ such that there are no arcs from $V(D')$ to $V(D)\backslash V(D')$. It is easy to know that $d_{D'}^{w+}(v)=d_{D}^{w+}(v)\geq 1$ for all $v\in V(D')$. By the induction hypothesis, there exists a {directed} cycle $C$ in $D'$ (and then, in $D$) such that {$w(C)\geq {1}/{\log_2(n(D')+r(D'))}$}. Clearly $n(D')\leq n$ and $r(D')\leq r$. Thus, we have {$w(C)\geq {1}/{\log_2(n+r)}$}, and complete the proof. \begin{case} $D$ is {strongly} connected. \end{case} \begin{subcase} There exists a vertex $z$ such that $zz\notin A(D)$. \end{subcase} By the {strongly-connectedness} of $D$, there exists at least one arc with head $z$. Let $y$ be a vertex such that $yz\in A(D)$ and $w(yz)=\max\{w(vz): vz\in A(D)\}$. Consider the digraph $D'$ such that {$V(D')=V(D)\backslash\{y\}$, $A(D')=A(D-y)\cup\{vz: vy\in A(D)\}$,} and $$ w_{D'}(uv)=\left\{ \begin{array}{ll} w_D(uy)+w_D(yz), & \mbox{if\ } uy\in A(D) \mbox{\ and\ } v=z;\\ w_D(uv), & \mbox{otherwise}. \end{array} \right. $$ Note that if $zy\in A(D)$, then $zz\in A(D')$, and $w_{D'}(zz)=w_D(zy)+w_D(yz)$. For every vertex $v\in V(D')$, its weighted outdegree is not less than that in $D$. Thus, we have $d_{D'}^{w+}(v)\geq 1$ for all $v\in V(D')$. By the induction hypothesis, there exists a {directed} cycle $C'$ in $D'$ such that {$w_{D'}(C')\geq {1}/{\log_2(n(D')+r(D'))}$}. {Since $n(D')=n-1$, and $D'$ contains at most one loop more than} $D$, we have $r(D')\leq r+1$. Thus {$w_{D'}(C')\geq {1}/{\log_2(n+r)}$}. If $C'$ does not contain the vertex $z$, then it is also a {directed} cycle in $D$ with the same weight. Otherwise, let $xz$ be the arc in $C'$ with head $z$. If $xy\notin A(D)$, then $C'$ is also a {directed} cycle in $D$ with the same weight. If $xy\in A(D)$, let $C$ be the {directed} cycle obtained from $C'$ by {replacing the arc $xz$ with} the path $xyz$, then $C$ is a {directed} cycle in $D$ of weight {$w_D(C)=w_{D'}(C')\geq {1}/{\log_2(n+r)}$.} \begin{subcase} For every $v\in V(D)$, $vv\in A(D)$. \end{subcase} In this case, $D$ has $r=n$ loops. And we need only prove that there exists a {directed} cycle in $D$ with weight at least {${1}/{\log_2(n+n)}={1}/{(1+\log_2n})$.} If there exists a loop with weight at least {${1}/({1+\log_2 n})$}, then we complete the proof. So we assume that {every loop} of $D$ has weight less than {${1}/({1+\log_2 n})$}. Let $D'$ be the digraph obtained from $D$ by deleting all the loops. Then $D'$ has $n$ vertices and no loops, and for each vertex $v$ in $V(D')$, {we have} $$ d_{D'}^{w+}(v)\geq 1-\frac{1}{1+\log_2 n}=\frac{\log_2 n}{1+\log_2 n}. $$ {It is easy to know that $D'$ is strongly connected. Note that for every vertex $v\in V(D')$, $vv\notin A(D')$. Using the conclusion of Case 2.1, we can obtained that} there exists a {directed} cycle $C$ in $D'$ such that $$ w_{D'}(C)\geq\frac{1}{\log_2 n}\frac{\log_2 n}{1+\log_2 n}=\frac{1}{1+\log_2 n}, $$ and $C$ is also a {directed} cycle in $D$ with the same weight. The proof is complete.\hfill$\Box$

https://arxiv.org/abs/2001.01304

Approximation of PDE eigenvalue problems involving parameter dependent matrices

We discuss the solution of eigenvalue problems associated with partial differential equations that can be written in the generalized form $\m{A}x=\lambda\m{B}x$, where the matrices $\m{A}$ and/or $\m{B}$ may depend on a scalar parameter. Parameter dependent matrices occur frequently when stabilized formulations are used for the numerical approximation of partial differential equations. With the help of classical numerical examples we show that the presence of one (or both) parameters can produce unexpected results.

\section{Introduction} \label{se:intro} Several schemes for the approximation of eigenvalue problems arising from partial differential equations lead to the algebraic form: find $\lambda\in\mathbb{R}$ and $x\in\mathbb{R}^n$ with $x\ne 0$ such that \begin{equation} \label{eq:eig} \m{A}x=\lambda\m{B}x, \end{equation} where $\m{A}$ and $\m{B}$ are matrices in $\mathbb{R}^{n\times n}$. We consider the case when the matrices $\m{A}$ and $\m{B}$ are symmetric and positive semidefinite and may depend on a parameter. This is a typical situation found in applications where elliptic partial differential equations are approximated by schemes that require suitable parameters to be tuned (for consistency and/or stability reasons). In this paper we discuss in particular applications arising from the use of the Virtual Element Method (VEM), see~\cite{MRR,BMRR,GV,MRV,MV,GMV,CGMMV}, where suitable parameters have to be chosen for the correct approximation. Similar situations are present, for instance, when a parameter-dependent stabilization is used for the approximation of discontinuous Galerkin formulations and when a penalty penalty term is added to the discretization of the eigenvalue problem associated with Maxwell's equations~\cite{CoDaMax,CoDaDurham,CoDareg,bfg,2006,WarburtonEmbree,2010,BoGue,BaCo} In general, it may be not immediate to describe how the matrices $\m{A}$ and $\m{B}$ depend on the given parameters. For simplicity, we consider the case when the dependence is linear: under suitable assumptions it is easy to discuss how the computed spectrum varies with respect to the parameters. The description of the spectrum in the linear case is not surprising and is well known to a broad scientific community~\cite{ElsnerSun,StewartSun,MR1066108,LiStewart,greenbaum2019firstorder}. Nevertheless, the main focus of perturbation theory for eigenvalues and eigenvectors is usually centered on the asymptotic behavior when the parameters tend to zero. In our case, the asymptotic parameter is usually the mesh size $h$ and we are interested in the convergence when $h$ goes to zero, that is when the size of the involved matrices tends to infinity. The presence of additional parameters makes the convergence more difficult to describe and can produce unexpected results in the pre-asymptotic regime. For this reason, we start by recalling how the spectrum of problem~\eqref{eq:eig} is influhenced by the parameter, without considering $h$, and we translate those results to an example of interest in Section~\ref{se:subVEM} where the discretization parameter $h$ is considered as well. We assume that the matrices $\m{A}$ and $\m{B}$ satisfy the following condition for $\m{C}=\m{A},\m{B}$. \begin{ass} \label{ass:C} The matrix $\m{C}$ can be split into the sum \begin{equation} \label{eq:C} \m{C}=\m{C}_1+\gamma\m{C}_2, \end{equation} where $\gamma$ is a non negative real number and $\m{C}_1$ and $\m{C}_2$ are symmetric. The matrices $\m{C}_1$ and $\m{C}_2$ satisfy the following properties: \begin{enumerate}[a)] \item $\m{C}_1$ is positive semidefinite with kernel $K_{\m{C}_1}$; \item $\m{C}_2$ is positive semidefinite and positive definite on $K_{\m{C}_1}$; \item $\m{C}_2$ vanishes on $K_{\m{C}_1}^\perp$, the orthogonal complement of $K_{\m{C}_1}$ in $\mathbb{R}^n$. \end{enumerate} \end{ass} In~\cref{se:param} we describe the spectrum of~\eqref{eq:eig} as a function of the parameters, in various situations that mimic the behavior of matrices $\m{A}$ and $\m{B}$ originating from several discretization schemes. \Cref{se:VEM}, which is the core of this paper, discusses the influence of the parameters on the VEM approximation of eigenvalue problems. Several numerical examples complete the papers, showing that the parameters have to be carefully tuned and that wrong choices can produce useless results. \section{Parametric algebraic eigenvalue problem} \label{se:param} Given two symmetric and positive semidefinite matrices $\m{A}$ and $\m{B}$ that can be written as \begin{equation} \label{eq:A} \m{A}=\Au+\alpha\Ad \end{equation} and \begin{equation} \label{eq:B} \m{B}=\Bu+\beta\Bd, \end{equation} with nonnegative parameters $\alpha$ and $\beta$, we consider the eigensolutions to the generalized problem~\eqref{eq:eig}. We assume that the splitting of the matrices $\m{A}$ and $\m{B}$ is obtained with symmetric matrices and satisfies~\cref{ass:C} for $\m{C}_1=\Au,\Bu$ and $\m{C}_2=\Ad,\Bd$. Moreover we denote by $\NA$ and $\NB$ the dimension of $\KA$ and $\KB$, respectively. \begin{remark} \label{re:geneig} Problem~\eqref{eq:eig} has $n$ eigenvalues if and only if $\mathrm{rank}\m{B}=n$, see~\cite{GVL}. If $\m{B}$ is singular the spectrum can be finite, empty, or infinite (if $\m{A}$ is singular too). If $\m{A}$ is non singular, usually one can circumvent this difficulty by computing the eigenvalues of $\m{B} x=\mu\m{A}x$ and setting $\lambda=1/\mu$. The kernel of $\m{B}$ is the eigenspace associated with the vanishing eigenvalue with multiplicity $m$, and the original problem has exactly $m$ eigenvalues conventionally set to $\infty$. \end{remark} We want to study the behavior of the eigenvalues as the parameters $\alpha$ and $\beta$ vary. We consider three cases. \subsection{Case 1} \label{se:caso1} We fix $\beta>0$ so that $\m{B}$ is positive definite. This implies that the eigenvalues of~\eqref{eq:eig} are all non negative. Let us consider first $\alpha=0$ so that~\eqref{eq:eig} reduces to \begin{equation} \label{eq:eigA1} \Au x=\lambda \m{B}x. \end{equation} Since $\Au$ is positive semidefinite, $\lambda=0$ is an eigenvalue of~\eqref{eq:eigA1} with multiplicity equal to $\NA=\dim(\KA)$ and $\KA$ is the associated eigenspace. In addition, we have $m_A=n-\NA$ positive eigenvalues $\{\mu_1\le\dots\le\mu_{m_A}\}$ counted with their multiplicity (since we are dealing with a symmetric problem, we do not distinguish between geometric and algebraic multiplicity). We denote by $v_j\in\KA^\perp$ the eigenvector associated with $\mu_j$, that is \[ \Au v_j=\mu_j \m{B}v_j. \] Thanks to property c) of~\cref{ass:C} when $\m{C}=\m{A}$, we observe that \[ \m{A}v_j=\Au v_j+\alpha\Ad v_j=\Au v_j=\mu_j\m{B} v_j. \] Therefore $(\mu_j,v_j)$, for $j=1,\dots,m_A$, are eigensolutions of the original system~\eqref{eq:eig}. On the other hand, the eigensolutions of \[ \Ad w=\nu\m{B}w \] are characterized by the fact that $\NA$ eigenvalues $\nu_i$ ($i=1,\dots,\NA$) are strictly positive with corresponding eigenvectors $w_i$ belonging to $\KA$, while the remaining $m_A$ eigenvalues vanish and have $\KA^\perp$ as eigenspace. Thus, property a) of~\cref{ass:C}, for $\m{C}=\m{A}$, yields \[ \m{A}w_i=\Au w_i+\alpha\Ad w_i=\alpha\Ad w_i=\alpha\nu_i \m{B}w_i, \] which means that $(\alpha\nu_i,w_i)$, for $i=1,\dots,\NA$, are eigensolutions of~\eqref{eq:eig}. Summarizing the eigenvalues of~\eqref{eq:eig} are: \begin{equation} \label{eq:eigA} \lambda_k= \left\{ \begin{array}{ll} \alpha\nu_k & \text{if }1\le k\le\NA\\ \mu_{k-\NA} & \text{if } \NA+1\le k\le n. \end{array} \right. \end{equation} The left panel in~\cref{fig:case1} shows the eigenvalues of a simple example where $\m{A}\in\mathbb{R}^{6\times6}$ is obtained by the combination of diagonal matrices with entries \begin{equation} \label{eq:matrici} \diag(\Au)=[3,4,5,6,0,0],\quad\diag(\Ad)=[0,0,0,0,1,2]. \end{equation} and $\m{B}=\mathbb{I}_6$ is the identity matrix. Along the vertical lines we see the eigenvalues corresponding to a fixed value of $\alpha$. The eigenvalues $3,4,5,6$ are associated with eigenvectors in $\KA^\perp$ and do not depend on $\alpha$. The solid lines starting at the origin display the eigenvalues $1,2$ multiplied by $\alpha$. \begin{figure} \begin{center} \includegraphics[width=.48\textwidth]{matlab/caso1-eps-converted-to.pdf} \includegraphics[width=.48\textwidth]{matlab/caso2-eps-converted-to.pdf} \caption{Dependence of the eigenvalues on the parameters $\alpha$ (Case~1) and $\beta$ (Case~2), respectively} \label{fig:case1} \end{center} \end{figure} \begin{remark} \label{re:intersection} We observe that if $\Ad$ is not positive definite on $\KA$, its kernel has a nonempty intersection with $\KA$. Let $n_{12}$ be the dimension of this intersection, then problem~\eqref{eq:eig} admits $n_{12}$ vanishing eigenvalues which appear in the first case of~\eqref{eq:eigA}. \end{remark} \subsection{Case 2} \label{se:caso2} Let us now fix $\alpha>0$, so that $\m{A}$ is positive definite. We have that all the eigenvalues are positive. We observe that when $\beta=0$, the matrix $\m{B}=\Bu$ may be singular, therefore it is convenient to consider the following problem: \begin{equation} \label{eq:eig2} \m{B}x=\chi\m{A}x, \end{equation} where $\chi=\frac 1\lambda$. If $\chi=0$, we conventionally set $\lambda=\infty$. Problem~\eqref{eq:eig2} reproduces the same situation we had in Case~1, with the matrices $\m{A}$ and $\m{B}$ switched. Repeating the same arguments as before, we obtain that problem~\eqref{eq:eig2} has two families of eigenvalues \[ \chi_k= \left\{ \begin{array}{ll} \beta\xi_k & \text{if }1\le k\le\NB\\ \zeta_{k-\NB} & \text{if } \NB+1\le k\le n, \end{array} \right. \] where \[ \aligned &\Bu r_j=\zeta_j \m{A} r_j,\ j=1,\dots,n-\NB &&\text{ with }r_j\in\KB^\perp\\ &\Bd s_i=\xi_i\m{A}s_i,\ i=1,\dots,\NB &&\text{ with }s_i\in\KB. \endaligned \] Going back to the original problem~\eqref{eq:eig}, we can conclude that the eigensolutions of~\eqref{eq:eig} are the following ones: \begin{equation} \label{eq:eigB} \aligned &\left(\frac1{\beta\xi_k},s_k\right) &&\text{ for } k=1,\dots,\NB\\ &\left(\frac1{\zeta_{k-\NB}},r_{k-\NB}\right)&&\text{ for }k=\NB+1,\dots,n. \endaligned \end{equation} In the right panel of~\cref{fig:case1}, we report the eigenvalues of a simple example where $\m{A}=\mathbb{I}_6$ and $\m{B}$ is obtained by combining $\Bu=\Au$ and $\Bd=\Ad$ defined in~\eqref{eq:matrici}. We see that the eigenvalues $\frac13,\frac14,\frac15,\frac16$ are independent of $\beta$ and that the remaining two eigenvalues lie along the hyperbolas $\frac1\beta$ and $\frac1{2\beta}$, plotted with solid line. \subsection{Case 3} \label{se:caso3} We consider now the case when $\alpha$ and $\beta$ can vary independently from each other. We have different situations corresponding to the relation between $\KA$ and $\KB$. To ease the reading, let us introduce the following notation: \begin{subequations} \begin{alignat}{1} &\Au v=\mu\Bu v \label{eq:notation11}\\ &\Au w=\nu \Bd w \label{eq:notation12}\\ &\Ad y=\chi\Bu y \label{eq:notation21}\\ &\Ad z=\eta\Bd z. \label{eq:notation22} \end{alignat} \end{subequations} In this case the space $\mathbb{R}^n$ can be decomposed into four mutually orthogonal subspaces \[ \mathbb{R}^n=(\KA\cap\KB)\oplus(\KA\cap\KB^\perp)\oplus(\KA^\perp\cap\KB)\oplus (\KA^\perp\cap\KB^\perp). \] Let us denote by $n_{\Au\cap\Bu}$ the dimension of $\KA\cap\KB$. If $\KA\cap\KB\ne\emptyset$, for $x\in\KA\cap\KB$ the eigenproblem to be solved is $\alpha\Ad x=\lambda\beta\Bd x$, hence the eigenvalues are given by $\frac{\alpha}{\beta}\eta_i$ $i=1,\dots,n_{\Au\cap\Bu}$, see~\eqref{eq:notation22}. Next, if $x\in\KA\cap\KB^\perp$ we have to solve $\alpha\Ad x=\lambda\Bu x$, which admits $(\alpha\chi_i,y_i)$ $i=1,\dots,\NA-n_{\Au\cap\Bu}$ as eigensolutions where $(\chi_i,y_i)$ are defined in~\eqref{eq:notation21}. Similarly, if $x\in\KA^\perp\cap\KB$, we find that the eigensolutions are $\left(\frac1{\beta}\nu_i,w,_i\right)$ $i=1,\dots,\NB-n_{\Au\cap\Bu}$ with $(\nu_i,w_i)$ given by~\eqref{eq:notation12}. In the last case, $x\in\KA^\perp\cap\KB^\perp$, the matrices $\m{A}$ and $\m{B}$ are non singular and thanks to property c) in ~\cref{ass:C}, for $\m{C}=\m{A}$ and $\m{C}=\m{B}$, we obtain that the eigenvalues are positive and independent of $\alpha$ and $\beta$ and correspond to those of~\eqref{eq:notation11}. In conclusion, we have \[ \lambda_k=\left\{ \begin{array}{ll} \displaystyle \frac{\alpha}{\beta}\eta_k &\quad\text{ if }1\le k\le n_{\Au\cap\Bu}\\ \alpha\chi_{k-n_{\Au\cap\Bu}} &\quad\text{ if }n_{\Au\cap\Bu}+1\le k\le\NA\\ \displaystyle\frac1{\beta}\nu_{k-\NA} &\quad\text{ if }\NA+1\le k\le\NA+\NB-n_{\Au\cap\Bu}\\ \mu_{k-\NA+\NB-n_{\Au\cap\Bu}} &\quad\text{ if }\NA+\NB-n_{\Au\cap\Bu}+1\le k\le n. \end{array} \right. \] \begin{figure}[ht] \begin{center} \includegraphics[width=.98\textwidth]{matlab/caso7-eps-converted-to.pdf} \caption{Eigenvalues when $\KA\cap\KB\ne\emptyset$ as a function of $\alpha$ and $\beta$} \label{fig:ab7} \end{center} \end{figure} We report in~\cref{fig:ab7} the eigenvalues illustrating this last case when we have diagonal matrices given by \[ \aligned &\diag(\Au)=[3,0,0,4,5,6] &&\quad\diag(\Ad)=[0,1,2,0,0,0]\\ &\diag(\Bu)=[7,8,0,0,9,10] &&\quad\diag(\Bd)=[0,0,0.8,1,0,0]. \endaligned \] The surface contains the eigenvalues depending on both $\alpha$ and $\beta$, the hyperbolas those depending only on $\beta$ and the straight lines those depending only on $\alpha$. If we cut the three dimensional picture with a plane at $\beta>0$ fixed we recognize the behavior analyzed in~\cref{se:caso1} and shown in~\cref{fig:case1} left. Analogously, taking a plane with $\alpha>0$ fixed, we recover Case 2 (see~\cref{se:caso2}). If $\KA\cap\KB=\emptyset$, we set $n_{\Au\cap\Bu}=0$, hence the eigenvalues are \[ \lambda_k=\left\{ \begin{array}{ll} \alpha\chi_k & \text{ if }1\le k\le \NA\\ \displaystyle\frac1\beta{\nu_{k-\NA}} & \text{ if } \NA+1\le k\le\NA+\NB\\ \mu_{k-\NA-\NB} & \text{ if } \NA+\NB+1\le k\le n \end{array}. \right. \] \begin{figure} \begin{center} \includegraphics[width=.98\textwidth]{matlab/caso4-eps-converted-to.pdf} \caption{Eigenvalues when $\KA\cap\KB=\emptyset$ as a function of $\alpha$ and $\beta$} \label{fig:ab4} \end{center} \end{figure} In order to illustrate the case $\KA\cap\KB=\emptyset$, we report in~\cref{fig:ab4} the eigenvalues computed using the following diagonal matrices with entries \[ \aligned &\diag(\Au)=[0,0,3,4,5,6],&&\quad\diag(\Ad)=[1,2,0,0,0,0],\\ &\diag(\Bu)=[7,8,9,10,0,0],&&\quad\diag(\Bd)=[0,0,0,0,0.8,1]. \endaligned \] For a fixed $\alpha$, we can see in solid line the hyperbolas $\frac{\nu_j}\beta$, $j=1,2$ while when $\beta$ is fixed we can see the straight lines $\alpha\chi_j$, $j=1,2$. The remaining two eigenvalues are independent of $\alpha$ and $\beta$. \section{Virtual element method for eigenvalue problems} \label{se:VEM} In this section we recall how algebraic eigenvalue problems similar to the ones discussed in the previous section can be obtained withing the framework of the Virtual Element Method (VEM) for the discretization of elliptic eigenvalue problems, see~\cite{GV,GMV}. We consider the model problem of the Laplacian operator. Given a connected open domain with Lipschitz continuous boundary $\Omega\subseteq\mathbb{R}^d$, with $d=2,3$, we look for eigenvalues $\lambda\in\mathbb{R}$ and eigenfunctions $u\ne0$ such that \[ \left\{ \begin{array}{ll} -\Delta u=\lambda u &\quad\text{ in }\Omega\\ u=0&\quad\text{ on }\partial\Omega. \end{array} \right. \] In view of the application of VEM, we consider the weak form: find $\lambda\in\mathbb{R}$ and $u\in\Huo$ with $u\ne0$ such that \begin{equation} \label{eq:Laplace} a(u,v)=\lambda b(u,v)\quad\forall v\in\Huo, \end{equation} where \[ a(u,v)=(\nabla u,\nabla v),\quad b(u,v)=(u,v), \] and $(\cdot,\cdot)$ is the scalar product in $L^2(\Omega)$. It is well-known that problem~\eqref{eq:Laplace} admits an infinite sequence of positive eigenvalues \[ 0<\lambda_1\le \dots\le\lambda_i\le\cdots \] repeated according to their multiplicity, each one associated with an eiegenfunction $u_i$ with the following properties \begin{equation} \label{eq:eigf} \aligned & a(u_i,u_j)=b(u_i,u_j)=0\quad \text{if }i\ne j\\ & b(u_i,u_i)=1,\quad a(u_i,u_i)=\lambda_i. \endaligned \end{equation} Let us briefly recall the definition of the virtual element spaces and of the discrete bilinear forms which we are going to use in this section, see~\cite{BBCMMR,AABMR}. We present only the two dimensional spaces, the three dimensional ones are obtained using the 2D virtual elements on the faces. We decompose $\Omega$ into polygons $P$, with diameter $h_P$ and area $|P|$. Similarly, if $e$ is an edge of an element $P$, we denote by $h_e=|e|$ its length. Depending on the context $\partial P$ refers to either the boundary of $P$ or the set of the edges of $P$. The notation $\T_h$ and $\E_h$ stands for the set of the elements and the edges, respectively. As usual, $h=\max_{P\in\T_h} h_P$. We assume the following mesh regularity condition (see~\cite{BBCMMR}): there exists a positive constant $\gamma$, independent of $h$, such that each element $P\in\T_h$ is star-shaped with respect to a ball of radius greater than $\gamma h_P$; moreover, for every element $P$ and for every edge $e\subset\partial P$, it holds $h_e \ge \gamma h_P$. For $k\ge1$ and $P\in\T_h$ we define \[ \tilde{V}_h^k(P)=\{v\in H^1(P): v|_{\partial P}\in C^0(\partial P), v|_e\in\P_k(e)\ \forall e\subset\partial P, \Delta v\in\P_k(P)\}. \] We consider the following linear forms on the space $\tilde{V}_h^k(P)$ \begin{enumerate} \item[D1]: the values $v(V_i)$ at the vertices $V_i$ of $P$, \item[D2]: the scaled edge moments up to order $k-2$ \[ \dfrac{1}{|e|}\int_e vm\,\text{d}s\quad\forall m\in\mathcal{M}_{k-2}(e),\ \forall e\subset\partial P, \] \item[D3]: the scaled element moments up to order $k-2$ \[ \dfrac{1}{|P|}\int_P vm\,\text{d}x\quad\forall m\in\mathcal{M}_{k-2}(P),\ \] \end{enumerate} where $\mathcal{M}_{k-2}(\omega)$ is the set of scaled monomials on $\omega$, namely \[ \mathcal{M}_{k-2}(\omega)=\Big\{\Big(\dfrac{\mathbf{x}-\mathbf{x}_\omega} {h_\omega}\Big)^s, |s|\le k-2\Big\}, \] with $\mathbf{x}_\omega$ the barycenter of $\omega$, and with the convention that $\mathcal{M}_{-1}(\omega)=\emptyset$. From the values of the linear operators D1--D3, on each element $P$ we can compute a projection operator $\Pinabla: \tilde{V}_h^k(P)\rightarrow \P_k(P)$ defined as the unique solution of the following problem: \begin{equation} \label{eq:pinabla} \aligned & a^P(\Pinabla v-v,p)=0\quad\forall p\in\P_k(P)\\ & \int_{\partial P}(\Pinabla v-v)\text{d}s=0, \endaligned \end{equation} where $a^P(u,v)=(\nabla u,\nabla v)_P$ and $(\cdot,\cdot)_P$ denotes the $L^2(P)$-scalar product. The local virtual space is defined as \begin{equation} \label{eq:VemP} \VemP=\left\{v\in\tilde{V}_h^k(P):\int_P (v-\Pinabla v) p\text{d}x=0\ \forall p\in(\P_k\setminus\P_{k-2})(P)\right\}, \end{equation} where $(\P_k\setminus\P_{k-2})(P)$ contains the polynomials in $\P_k(P)$ $L^2$-orthogonal to $\P_{k-2}(P)$. We recall that by construction $\P_k(P)\subset\VemP$, so that the optimal rate of convergence is ensured. Moreover, the linear operators D1--D3 provide a unisolvent set of degrees of freedom (DoFs) for $\VemP$, which allows us to define and compute $\Pinabla$ on $\VemP$. In addition, the $L^2$-projection operator $\Pio:\VemP\to\P_k(P)$ is also computable using the DoFs. The global virtual space is \begin{equation} \label{eq:Vem} \V=\{v\in\Huo: v|_P\in\VemP\ \forall P\in\T_h\}. \end{equation} In order to discretize problem~\eqref{eq:Laplace}, we introduce the discrete counterparts $a_h$ and $b_h$ of the bilinear forms $a$ and $b$, respectively. Both discrete forms are obtained as sum of the following local contributions: for all $u_h,v_h\in\V$ \begin{equation} \label{eq:abP} \aligned &a_h^P(u_h,v_h)=a^P(\Pinabla u_h,\Pinabla v_h) +S_a^P((I-\Pinabla)u_h,(I-\Pinabla)v_h)\\ &b_h^P(u_h,v_h)=b^P(\Pio u_h,\Pio v_h)+S_b^P((I-\Pio)u_h,(I-\Pio)v_h), \endaligned \end{equation} where $b^P(u,v)=(u,v)_P$, and $S_a^P$ and $S_b^P$ are symmetric positive definite bilinear forms on $\VemP\times\VemP$ such that \begin{equation} \label{eq:stab} \aligned &c_0 a^P(v,v)\le S_a^P(v,v)\le c_1a^P(v,v) &&\quad\forall v\in\VemP \text{ with }\Pinabla v=0\\ &c_2 b^P(v,v)\le S_b^P(v,v)\le c_3b^P(v,v) &&\quad\forall v\in\VemP \text{ with }\Pio v=0, \endaligned \end{equation} for some positive constants $c_i$ ($i=0,\dots,3$) independent of $h$. We define $a_h(u_h,v_h)=\sum_{P\in\T_h}a_h^P(u_h,v_h)$ and $b_h(u_h,v_h)=\sum_{P\in\T_h}b_h^P(u_h,v_h)$. The virtual element counterpart of~\eqref{eq:Laplace} reads: find $\lambda_h$ and $u_h\in\V$ with $u_h\ne0$ such that \begin{equation} \label{eq:LaplV} a_h(u_h,v_h)=\lambda_h b_h(u_h,v_h)\quad\forall v_h\in\V. \end{equation} Thanks to~\eqref{eq:stab}, the discrete problem~\eqref{eq:LaplV} admits $N_h=\dim{\V}$ positive eigenvalues \[ 0<\lambda_{1h}\le\dots\lambda_{N_hh} \] and the corresponding eigenfunctions $u_{ih}$, for $i=1,\dots,N_h$, enjoy the discrete counterpart of properties in~\eqref{eq:eigf}. The following convergence result has been proved in~\cite{GV}. \begin{thm} \label{th:conv} Let $\lambda$ be an eigenvalue of~\eqref{eq:Laplace} of multiplicity $m$ and $\mathcal{E}_\lambda$ the corresponding eigenspace. Then there are exactly $m$ discrete eigenvalues of~\eqref{eq:LaplV} $\lambda_{j(i)h}$ ($i=1,\dots,m$) tending to $\lambda$. Moreover, assuming that $u\in H^{1+r}(\Omega)$, for all $u\in\mathcal{E}_\lambda$, the following inequalities hold true: \[ \aligned &|\lambda-\lambda_{j(i)h}|\le Ch^{2t}\\ &\hat\delta(\mathcal{E}_\lambda,\oplus_i\mathcal{E}_{j(i)h})\le Ch^t, \endaligned \] where $t=\min(k,r)$, $\hat\delta(\mathcal{E},\mathcal{F})$ represents the gap between the spaces $\mathcal{E}$ and $\mathcal{F}$, and $\mathcal{E}_{\ell h}$ is the eigenspace spanned by $u_{\ell h}$. \end{thm} \begin{remark} \label{re:nonstab} It is also possible to consider on the right hand side of~\eqref{eq:LaplV} the bilinear form for $\tb(u_h,v_h)=\sum_{P\in\T_h}b^P(\Pio u_h,\Pio v_h)$. This leads to the following discrete eigenvalue problem: find $(\tl,\tu)\in\mathbb{R}\times\V$ with $\tu\ne0$ such that \begin{equation} \label{eq:nonstab} a_h(\tu,v_h)=\tl\tb(\tu,v_h)\quad\forall v_h\in\V. \end{equation} The analogue of~\cref{th:conv} holds true for this partially non stabilized discretization as well. \end{remark} \subsection{Computational aspects and numerical results} \label{se:subVEM} In order to compute the solution of problems~\eqref{eq:LaplV} and~\eqref{eq:nonstab}, we need to describe how to obtain the matrices associated to our bilinear forms. By construction the matrix $\Au$ (respectively, $\Bu$) associated with $\sum_P a^P(\Pinabla\cdot,\Pinabla\cdot)$ (respectively, $\sum_P b^P(\Pio\cdot,\Pio\cdot)$) has kernel corresponding to the elements $v_h\in\V$ such that $\Pinabla v_h$ is constant (respectively, $\Pio v_h=0$) for all $P\in\T_h$. We observe that the local contributions of the bilinear forms displayed in~\eqref{eq:abP} mimic the following exact relations \begin{equation} \label{eq:exactab} \aligned &a^P(u_h,v_h)= a^P(\Pinabla u_h,\Pinabla v_h)+a^P((I-\Pinabla)u_h,(I-\Pinabla)v_h)\\ &b^P(u_h,v_h)= b^P(\Pio u_h,\Pio v_h)+b^P((I-\Pio)u_h,(I-\Pio)v_h). \endaligned \end{equation} Let us denote by $\Aul$, $\Adl$, $\Bul$ and $\Bdl$ the matrices whose entries are given by \begin{equation} \label{eq:matr} \aligned &(\Aul)_{ij}=a^P(\Pinabla \phi_i,\Pinabla \phi_j),&\quad &(\Adl)_{ij}=a^P((I-\Pinabla)\phi_i,(I-\Pinabla)\phi_j)\\ &(\Bul)_{ij}=b^P(\Pio \phi_i,\Pio \phi_j),&\quad &(\Bdl)_{ij}=b^P((I-\Pio)\phi_i,(I-\Pio)\phi_j) \endaligned \end{equation} with $\phi_i$ basis functions for $\VemP$. Even if the global matrices $\m{A}$ and $\m{B}$ do not satisfy the properties stated in~\cref{ass:C}, it turns out that~\cref{ass:C} is fulfilled by $\m{C}=\Bul+\beta\Bdl$; moreover, $\m{C}=\Aul+\alpha\Adl$ is characterized by the situation described in~\cref{re:intersection}. We start with the pair $\Aul$ and $\Adl$. The kernel $K_{\Aul}$, with abuse of notation, is characterized by \[ K_{\Aul}=\{v\in\VemP: a^P(\Pinabla v,\Pinabla w)=0\ \forall w\in\VemP\}, \] that is, $K_{\Aul}$ is made of $v$ with constant $\Pinabla v$ on $P$. Moreover, the orthogonal complement of $K_{\Aul}$, denoted by $K_{\Aul}^\perp$ contains the elements $v\in\VemP$ such that $a^P(v,w)=0$ for all $w\in K_{\Aul}$. We now show that $\Adl( K_{\Aul}^\perp)=0$, that is, for all $v\in K_{\Aul}^\perp$, $a^P((I-\Pinabla)v,(I-\Pinabla)w)=0$ for all $w\in\VemP$. We recall that, if $v\in K_{\Aul}^\perp$, then $a^P(v,w)=0$ for all $w\in K_{\Aul}$. This implies that for $v\in K_{\Aul}^\perp$ and $w\in K_{\Aul}$, it holds true that $a^P(v,w)=a^P((I-\Pinabla)v,(I-\Pinabla)w)=0$. Now we can write for all $w\in\VemP$ \[ \aligned &a^P((I-\Pinabla)v,(I-\Pinabla)w)\\ &\quad=a^P((I-\Pinabla)v,(I-\Pinabla)(I-\Pinabla)w)+ a^P((I-\Pinabla)v,(I-\Pinabla)\Pinabla w)=0. \endaligned \] Indeed, $\Pinabla(I-\Pinabla)w=0$ implies that $(I-\Pinabla)w\in K_{\Aul}$, and thus the first term vanishes, while for the second term it is enough to observe that $\Pinabla(\Pinabla w)=\Pinabla w$. Thus property c) of~\cref{ass:C} is verified for $\m{C}=\m{A}$. Concerning property b) of~\cref{ass:C}, we have by construction, that $a^P((I-\Pinabla)v,(I-\Pinabla)v)\ge0$ for all $v\in\VemP$, see~\eqref{eq:exactab}. On the other hand, if $v$ is constant on $P$, then $\Pinabla v=v$ is constant, therefore $v\in K_{\Aul}$ and $(I-\Pinabla)v=0$ so that $v$ belongs also to the kernel of $\Adl$. Hence the pair $\Aul$ and $\Adl$ does not satisfy property b), but it is in the situation described in~\cref{re:intersection}. Let us now consider the pair $\Bul$ and $\Bdl$. We observe that the kernel of $\Bul$ is characterized by $\Pio v=0$. The analysis performed for the pair $\Aul$ and $\Adl$ can be repeated and gives that in this case~\cref{ass:C} is verified for $\m{C}=\m{B}$. As a consequence of the assembling of the local matrices, the global matrices $\Au$ and $\Ad$ ($\Bu$ and $\Bd$, respectively) do not satisfy anymore the properties listed in~\cref{ass:C}. In particular, for $k=1$ we shall see that the matrices $\Au$ and $\Bu$ are not singular. Nevertheless, we are going to show that the numerical results look pretty much similar to the ones reported in~\cref{se:param}. Moreover, in practice the matrices $\Adl$ and $\Bdl$ are not available and they are replaced by using the local bilinear forms $S_a^P$ and $S_b^P$ given in~\eqref{eq:abP} as follows. Let us denote by $\mathbf{u}_h,\mathbf{v}_h\in\mathbb{R}^{N_P}$ the vectors containing the values of the $N_P$ local DoFs associated to $u_h,v_h\in\VemP$. Then, we define the local stabilized forms as \[ S_a^P(u_h,v_h)=\sigma_P \mathbf{u}_h^\top\mathbf{v}_h,\quad S_b^P(u_h,v_h)=\tau_P h_P^2\mathbf{u}_h^\top\mathbf{v}_h \] where the stability parameters $\sigma_P$ and $\tau_P$ are positive constants which might depend on $P$ but are independent of $h$. We point out that this choice implies the stability requirements in~\eqref{eq:stab}. In the applications, the parameter $\sigma_P$ is usually chosen depending on the mean value of the eigenvalues of the matrix stemming from the term $a_P(\Pinabla\cdot,\Pinabla\cdot)$, and $\tau_P$ as the mean value of the eigenvalues of the matrix resulting from $\frac1{h^2_P}(\Pio\cdot,\Pio\cdot)_P$. The choice of the stabilized form $S_a^P$ is discussed in some papers concerning the source problem, see, e.g., \cite{BLR} and the references therein. One can find an analysis of the stabilization parameters $\sigma_P$ in~\cite{CMRS}. If $\sigma_P$ and $\tau_P$ vary in a small range, it is reasonable to take $\sigma_P=\alpha$ and $\tau_P=\beta$ for all $P$ and this is the situation which we discuss further. Therefore, the structure of the matrices is $\m{A}=\Au+\alpha\Ad$ and $\m{B}=\Bu+\beta\Bd$ where $\Ad$ and $\Bd$ are the matrices with local contribution given by $\mathbf{u}_h^\top\mathbf{v}_h$ and $h_P^2\mathbf{u}_h^\top\mathbf{v}_h$, respectively. We study the behavior of the eigenvalues as $\alpha$ and $\beta$ vary in given ranges. In the following tests $\Omega$ is the unit square partitioned using a sequence of Voronoi meshes with a given number of elements. In~\cref{fig:Voronoi} we report the coarsest mesh with 50 elements ($h=0.2350$, 151 edges, 102 vertices). We recall that the exact eigenvalues are given by $(i^2+j^2)\pi^2$ for $i,j\in\mathbb{N}\setminus\{0\}$ with eigenfunctions $\sin(i\pi x)\sin(j\pi y)$. The following numerical results have been obtained using \textsc{Matlab} and, in particular, the routine \texttt{eig} for the computation of the eigenvalues. In the following figures, we shall always report the computed eigenvalues divided by $\pi^2$. \begin{center} \begin{figure}[ht] \includegraphics[width=0.6\textwidth]{figures/mesh50square-eps-converted-to.pdf} \caption{Voronoi mesh with 50 polygons.} \label{fig:Voronoi} \end{figure} \end{center} \Cref{tb:kerAu} and~\cref{tb:kerBu} display the dimension of the kernel of the matrices $\Au$ and $\Bu$ for $k=1,2,3$, and for different numbers $N$ of the elements in the mesh. \begin{table} \caption{Dimension of $\KA$ with respect to $k$ and the number of elements} \centering \begin{tabular}{cc cc cc cc cc cc }\toprule $k$ && $N=50$ && $N=100$ &&$N=200$ &&$N=400$ &&$N=800$\\ \midrule 1 && 0 && 0 && 0 && 0 && 0\\ 2 && 3 && 30 && 99 && 258 && 565\\ 3 && 27 && 94 && 246 && 588 && 1312\\ \bottomrule \end{tabular} \label{tb:kerAu} \end{table} \begin{table} \caption{Dimension of $\KB$ with respect to $k$ and the number of elements} \centering \begin{tabular}{cc cc cc cc cc cc }\toprule $k$ && $N=50$ && $N=100$ &&$N=200$ &&$N=400$ &&$N=800$\\ \midrule 1 && 0 && 0 && 0 && 0 && 0\\ 2 && 0 && 0 && 0 && 0 && 0\\ 3 && 0 && 1 && 43 && 182 && 504\\ \bottomrule \end{tabular} \label{tb:kerBu} \end{table} In particular we see that for $k=1$ the matrix $\Au$ is nonsingular. We have computed the lowest eigenvalue of $\Au x=\lambda \Bu x$, which gives an estimate of the \emph{inf-sup} constant of the discrete problem~\eqref{eq:LaplV}. The results, presented in~\cref{tb:infsup}, show that the first eigenvalue is decreasing, and this behavior corresponds to the fact that the bilinear form $\sum_P a^P(\Pinabla\cdot,\Pinabla\cdot)$ is not stable. \begin{table} \caption{First eigenvalues of $\Au x=\lambda \Bu x$ for different meshes} \centering \begin{tabular}{c c c c c}\toprule $N=50$ & $N=100$ &$N=200$ &$N=400$ &$N=800$\\ \midrule 1.92654e+00 & 1.74193e+00 & 1.06691e+00 & 6.81927e-01 & 5.54346e-01\\ \bottomrule \end{tabular} \label{tb:infsup} \end{table} We now discuss some tests, where we present the behavior of the eigenvalues as the parameters $\alpha$ and $\beta$ vary, for the mesh with $N=200$ and different degree $k$ of the polynomials in the space $\V$. The rows of~\cref{fig:girato} contain the results for fixed $k$ and the values $\beta=0,1,5$, while, in the columns, $\beta$ is fixed and $k$ varies. In each picture, we plot in red the exact eigenvalues and with different colors those corresponding to $\alpha=10^r$ with $r=-3,\dots,1$. These plots clearly confirm that the choice of the parameters for optimal performance is not so immediate. Consider, in particular, that we are solving the Laplace eigenvalue problem (isotropic diffusion) on a domain as simple as a square. For an arbitrary elliptic problem and more general domains the situation could be much more complicated. For $\beta=0$, the first 30 eigenvalues are well approximated with higher degree of polynomials whenever $\alpha\ge0.1$. The value $\alpha=0.1$ seems to be the best choice in the case $k=1$. Increasing $\beta$ does not produce much improvement. All the pictures seem to indicate that higher values of $\alpha$ might give better results. In particular, for $k=2,3$ the first 30 eigenvalues are approximated with a reasonable accuracy for $\alpha=10$ and $\beta=1$. Increasing $\beta$ and keeping $\alpha=10$, we see that a smaller number of eigenvalues are captured. \begin{figure}[h] \begin{center} \includegraphics[width=\textwidth]{figures/figure_F/legend.png} \subfigure[\tiny{$k=1$, $\beta=0$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok1_beta0_autoval30-eps-converted-to.pdf}} \subfigure[\tiny{$k=1$, $\beta=1$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok1_beta1_autoval30-eps-converted-to.pdf}} \subfigure[\tiny{$k=1$, $\beta=5$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok1_beta5_autoval30-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\beta=0$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok2_beta0_autoval30-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\beta=1$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok2_beta1_autoval30-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\beta=5$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok2_beta5_autoval30-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\beta=0$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok3_beta0_autoval30-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\beta=1$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok3_beta1_autoval30-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\beta=5$}] {\includegraphics[width=.32\textwidth]{figures/figure_F/giratok3_beta5_autoval30-eps-converted-to.pdf}} \end{center} \caption{First 30 eigenvalues for different values of $k$, $\alpha$ and $\beta$} \label{fig:girato} \end{figure} \Cref{fig:alfa} shows the behavior of the eigenvalues as $\alpha$ varies from $0$ to $10$. At a first glance the pictures remind of~\cref{fig:case1} (left) even if, as it has been explained before, the situation is not exactly matching what we discussed in~\cref{se:param}. Each subplot reports all computed eigenvalues between $0$ and $40$; the dotted horizontal lines represent the exact solutions. The first $30$ computed eigenvalues are connected together with lines of different colors in an automated way. An ``ideal'' good approximation would correspond to a series of colored lines matching the dotted lines of the exact eigenvalues. It is interesting to look at the differences between various degrees ($k$ from $1$ to $3$ moving from the top to the bottom) and values of $\beta$ (equal to $0$, $1$, and $5$ from left to right). \begin{figure}[h] \begin{center} \subfigure[\tiny{$k=1$, $\beta=0$, $\alpha\in[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k1_b0_a10-eps-converted-to.pdf}} \subfigure[\tiny{$k=1$, $\beta=1$, $\alpha\in[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k1_b1_a10-eps-converted-to.pdf}} \subfigure[\tiny{$k=1$, $\beta=5$, $\alpha\in[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k1_b5_a10-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\beta=0$, $\alpha=[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k2_b0_a10-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\beta=1$, $\alpha=[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k2_b1_a10-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\beta=5$, $\alpha=[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k2_b5_a10-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\beta=0$, $\alpha=[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k3_b0_a10-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\beta=1$, $\alpha=[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k3_b1_a10-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\beta=5$, $\alpha=[0,10]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/rette_k3_b5_a10-eps-converted-to.pdf}} \end{center} \caption{Eigenvalues versus $\alpha$ for different values of $k$ and $\beta$} \label{fig:alfa} \end{figure} \begin{figure} \includegraphics[width=.70\textwidth]{figures/autofun/rette_k3_b1_a10_spuri-eps-converted-to.pdf} \caption{Same plot as in~\cref{fig:alfa}(h) with four marked (spurious) eigenvalues} \label{fig:marked} \end{figure} \begin{figure} \begin{center} \includegraphics[width=.45\textwidth]{figures/autofun/alfa3_sp2-eps-converted-to.pdf} \includegraphics[width=.45\textwidth]{figures/autofun/alfa11_sp7-eps-converted-to.pdf} \includegraphics[width=.45\textwidth]{figures/autofun/alfa21_sp14-eps-converted-to.pdf} \includegraphics[width=.45\textwidth]{figures/autofun/alfa31_sp20-eps-converted-to.pdf} \end{center} \caption{Eigenfunctions corresponding to the eigenvalues marked in~\cref{fig:marked}} \label{fig:autof} \end{figure} More reliable results seem to be obtained for large $k$ and small $\beta$. Actually, the limit case of $\beta=0$ appears to be the safest choice. This is in agreement with the claim of~\cite{BMRR} where the authors remark that ``even the value $\sigma_E=0$ yields very accurate results, in spite of the fact that for such a value of the parameter the stability estimate and hence most of the proofs of the theoretical results do not hold'' (note that $\sigma_E=0$ in~\cite{BMRR} has the same meaning as $\beta$ in our paper). It is interesting to observe that the analysis of~\cite{GV}, summarized in~\cref{th:conv}, covers the case $\beta=0$ as well. On the other hand $\beta=0$ may produce a singular matrix $\m{B}$ and this could be not convenient from the computational point of view. In order to better understand the behavior of the eigenvalues reported in~\cref{fig:alfa}(h), we highlight in~\cref{fig:marked} four eigenvalues that are apparently aligned along an oblique line. The corresponding eigenfunctions are reported in~\cref{fig:autof}. The four eigenfunctions look similar, so that the analogy with~\cref{fig:case1} (left) is even more evident. We conclude this discussion with an example where, for a given value of $\alpha$, a good eigenvalue (i.e., an eigenvalue corresponding to a correct approximation) is crossing a spurious one (i.e., an eigenvalue belonging to an oblique line). In this case it may happen that the two eigenfunctions mix together, thus yielding to an even more complicated situation. This behavior is reported in~\cref{fig:autof-marked}, where a region of the plot shown in~\cref{fig:alfa}(h) is blown-up close to an intersection point: actually three eigenvalues (a spurious one and two corresponding to good ones) are clustered at the marked intersection points. \begin{figure} \begin{center} \includegraphics[width=.45\textwidth]{figures/autofun/spuriomischiato1-eps-converted-to.pdf} \includegraphics[width=.45\textwidth]{figures/autofun/alfa13_sp7-eps-converted-to.pdf} \medskip \includegraphics[width=.45\textwidth]{figures/autofun/spuriomischiato2-eps-converted-to.pdf} \includegraphics[width=.45\textwidth]{figures/autofun/alfa41_sp25-eps-converted-to.pdf} \end{center} \caption{Intersections of good and spurious eigenvalues} \label{fig:autof-marked} \end{figure} \Cref{fig:beta} shows the computed eigenvalues smaller that $40$ when $\beta$ varies from $0$ to $5$ and for a fixed value of $\alpha$. As in~\cref{fig:girato} and in analogy with~\cref{fig:alfa}, the rows correspond to the degree $k$ of polynomials, while the columns refer to different values of $\alpha$. The dotted horizontal lines represent the exact eigenvalues. The lines with different colors in each picture follow the $n$-th eigenvalue for $n=1,\dots,30$. It turns out that all lines are originating from curves that look like hyperbolas when $\beta$ is large. Following each of these hyperbolas from $\beta=+\infty$ backwards, it happens that when the hyperbola meets a correct approximation of an eigenvalue of the continuous problem, it deviates from its trajectory and becomes a (almost horizontal) straight line. In the case $k=1$, we see that the higher eigenvalues are computed with decreasing accuracy as $\beta$ approaches $0$. \begin{figure}[h] \begin{center} \subfigure[\tiny{$k=1$, $\alpha=0.1$, $\beta\in[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k1_a01_b5-eps-converted-to.pdf}} \subfigure[\tiny{$k=1$, $\alpha=1$, $\beta\in[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k1_a1_b5-eps-converted-to.pdf}} \subfigure[\tiny{$k=1$, $\alpha=10$, $\beta\in[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k1_a10_b5-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\alpha=0.1$, $\beta=[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k2_a01_b5-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\alpha=1$, $\beta=[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k2_a1_b5-eps-converted-to.pdf}} \subfigure[\tiny{$k=2$, $\alpha=10$, $\beta=[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k2_a10_b5-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\alpha=0.1$, $\beta=[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k3_a01_b5-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\alpha=1$, $\beta=[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k3_a1_b5-eps-converted-to.pdf}} \subfigure[\tiny{$k=3$, $\alpha=10$, $\beta=[0,5]$}] {\includegraphics[width=.32\textwidth]{figures/figure_Lnew/iperb_k3_a10_b5-eps-converted-to.pdf}} \end{center} \caption{Eigenvalues versus $\beta$ for different values of $k$ and $\alpha$} \label{fig:beta} \end{figure} We recognize in these pictures the situation presented in~\cref{se:caso2}, corresponding to the behavior of the eigenvalues when the parameter $\beta$ in matrix $\m{B}$ varies. In this test, the kernel of matrix $\Bu$ is not empty only for $k=3$. Nevertheless, we can see that when $\beta$ approaches $0$, there are several eigenvalues going to $\infty$. On the other side, for greater values of $\beta$ we obtain several spurious eigenvalues. The range of $\beta$, which gives eigenvalues close to the exact ones, clearly depends on $k$ and $\alpha$. \Cref{fig:6400} displays, in separate pictures, the first four eigenvalues, with $k=1$, $\alpha=10$, different values of $h$, and $0\le\beta\le400$. Taking into account that the routine \texttt{eig} sorts the eigenvalues in ascending order, the four pictures display, in lexicographical order, the first, second, third and fourth computed eigenvalues. In each subplot, each line refers to a particular mesh. We can see that the eigenvalues computed with the finest mesh seem to be insensitive with respect to the value of $\beta$. On the opposite side the coarsest mesh gives approximations of the correct values only when $\beta$ is very small and, furthermore, the accuracy is rather low. For each eigenvalue and each fixed mesh we recognize a critical value of the parameter such that greater values of $\beta$ produce spurious eigenvalues. The behavior of these eigenvalues clearly reproduces that of the eigenvalues in~\cref{fig:case1} (right) referring to Case 2. The results are plotted with a different perspective depending on the fact that the results now depend also on the computational mesh. The right bottom plot of~\cref{fig:6400} highlights a phenomenon which already appears in~\cref{fig:beta}(i). Indeed, we see that the red line corresponding to the fourth computed eigenvalue for $N=400$ lies along an hyperbola until $\beta=65$ where it reaches the value $5$ associated with second and third exact eigenvalues. Between $\beta=65$ and $\beta=55$ the red line remains close to $5$, then decreasing $\beta$ it follows a different hyperbola until it reaches the expected value for $\beta=35$. \begin{figure}[h] \begin{center} \includegraphics[width=\textwidth]{figures/6400_4-eps-converted-to.pdf} \end{center} \caption{First four eigenvalues} \label{fig:6400} \end{figure} \section*{Conclusions} In this paper we have discussed how numerically computed eigenvalues can depend on discretization parameters. \Cref{se:param} shows the dependence on $\alpha$ and $\beta$ of the eigenvalues of~\eqref{eq:eig} when $\m{A}$ and $\m{B}$ have the forms~\eqref{eq:A} and~\eqref{eq:B}, respectively. In~\cref{se:VEM} we have studied the behavior of the eigenvalues of the Laplace operator computed with the Virtual Element Method. The presence of two parameters resembles the abstract setting of~\cref{se:param}; even if assumptions satisfied by the VEM matrices are more complicated than the ones previously discussed, the numerical results are pretty much in agreement. The present work opens the question of a viable choice of the parameters for eigenvalue computations when the discretization scheme depends on a suitable tuning of them (such as in the case of VEM). \section*{Acknowledgments} % The authors are members of INdAM Research group GNCS and their research is supported by PRIN/MIUR. The research of the first and third authors is partially supported by IMATI/CNR.

https://arxiv.org/abs/0908.2456

Descent polynomials for permutations with bounded drop size

Motivated by juggling sequences and bubble sort, we examine permutations on the set {1,2,...,n} with d descents and maximum drop size k. We give explicit formulas for enumerating such permutations for given integers k and d. We also derive the related generating functions and prove unimodality and symmetry of the coefficients.

\section{Introduction} There have been extensive studies of various statistics on $\mathcal{S}_n$, the set of all permutations of $\{ 1,2, \dots, n\}$. For a permutation $\pi$ in $\mathcal{S}_n$, we say that $\pi$ has a \emph{drop} at $i$ if $\pi_i < i$ and that the \emph{drop size} is $i-\pi_i$. We say that $\pi$ has a \emph{descent} at $i$ if $\pi_i > \pi_{i+1}$. One of the earliest results \cite{mac} in permutation statistics states that the number of permutations in $\mathcal{S}_n$ with $k$ drops equals the number of permutations with $k$ descents. A concept closely related to drops is that of \emph{excedances}, which is just a drop of the inverse permutation. In this paper we focus on drops instead of excedances because of their connection with our motivating applications concerning bubble sort and juggling sequences. Other statistics on a permutation $\pi$ include such things as the number of \emph{inversions}, that is, $|\{(i,j) : i < j,\; \pi_i > \pi_j\}|$, and the \emph{major index} of $\pi$ (i.e., the sum of $i$ for which a descent occurs). The enumeration of and generating functions for these statistics can be traced back to the work of Rodrigues in 1839 \cite{rodrigues} but was mainly influenced by McMahon's treatise in 1915 \cite{mac}. There is an extensive literature studying the distribution of the above statistics and their $q$-analogs (see for example Foata and Han \cite{foata_han} or the papers of Shareshian and Wachs \cite{wachs,wachs2} for more recent developments). As noted above, the drop statistic that we study is closely related to the excedances statistic. The distribution of the bivariate statistics $(\mbox{descents},\mbox{excedances})$ can be found in Foata and Han~\cite[equations (1.15) and (1.16)]{foata_han_fix}. This joint work originated from its connection with a paper \cite{jug} on sequences that can be translated into juggling patterns. The set of juggling sequences of period $n$ containing a specific state, called the ground state, corresponds to the set $\mathcal{B}_{n,k}$ of permutations in $\mathcal{S}_n$ with drops of size at most $k$. As it turns out, $\mathcal{B}_{n,k}$ can also be associated with the set of permutations that can be sorted by $k$ operations of bubble sort. These connections will be further described in the next section. We note that the {\it maxdrop} statistic has not been treated in the literature as extensively as many other statistics in permutations. As far as we know, this is the first time that the distribution of descents with respect to maxdrop has been determined. First we give some definitions concerning the statistics and polynomials that we examine. Given a permutation $\pi$ in $\mathcal{S}_n$, let $\Des(\pi)$ denote the descent set, $\{1\leq i<n: \pi_i > \pi_{i+1}\}$, of $\pi$ and let $\des(\pi)=|\Des(\pi)|$ be the number of descents. We use $\maxdrop(\pi)$ to denote the value of the maximum drop (or maxdrop) of $\pi$, \begin{equation*} \maxdrop (\pi) = \max\{\,i - \pi(i) : 1\leq i\leq n\,\}. \end{equation*} Let $\mathcal{B}_{n,k}=\{\pi \in \mathcal{S}_n : \maxdrop(\pi) \leq k\}$. It is known, and also easy to show, that $|\mathcal{B}_{n,k}| = k!(k+1)^{n-k}$; e.g., see \cite[Thm. 1]{jug} or \cite[p. 108]{knuth}. Let \begin{equation*} b_{n,k}(r) = |\{\pi \in \mathcal{B}_{n,k} : \des(\pi) = r \}|, \end{equation*} and define the ($k$-maxdrop-restricted) descent polynomial \begin{equation*} B_{n,k} (x) = \sum_{r \geq 0} b_{n,k}(r) x^r = \sum_{\pi \in \mathcal{B}_{n,k}}x^{\des(\pi)}. \end{equation*} Examining the case of $k=2$, we discovered that the coefficients $b_{n,2}(r)$ of $B_{n,2}(x)$ appear to be given by every \emph{third} coefficient of the simple polynomial \begin{equation*} (1+x^2)(1+x+x^2)^{n-1}. \end{equation*} Looking at the next two cases, $k=3$ and $k=4$, yielded more mysterious polynomials: $b_{n,3}(r)$ appeared to be every fourth coefficient of \begin{equation*} (1+x^2+2x^3+x^4+x^6)(1+x+x^2+x^3)^{n-2} \end{equation*} and $b_{n,4}(r)$ every fifth coefficient of \begin{equation*} (1+x^2+2x^3+4x^4+4x^5+4x^7+4x^8+2x^9+x^{10}+x^{12})(1+x+x^2+x^3+x^4)^{n-3}. \end{equation*} After a fierce battle with these polynomials, we were able to show that $b_{n,k}(r)$ is the coefficient of $ u^{r(k+1)}$ in the polynomial \begin{equation}\label{closed_1} P_k(u) \left(1+u+\dots + u^k \right)^{n-k} \end{equation} where \begin{equation}\label{closed_2} P_k(u) = \sum_{j=0}^k A_{k-j}(u^{k+1})(u^{k+1}-1)^{j}\sum_{i=j}^k\binom{i}{j}u^{-i}, \end{equation} and $A_k$ denotes the $k$th Eulerian polynomial (defined in the next section). Further to this, we give an expression for the generating function $\mathbf B_k(z,y) = \sum_{n \geq 0}B_{n,k}(y)z^n$, namely \begin{equation*} \mathbf B_k(z,y) = \dfrac{\displaystyle{ 1+\sum_{t=1}^k \left( A_t(y) - \sum_{i=1}^t \binom{k+1}{i} (y-1)^{i-1} A_{t-i}(y) \right)z^t }}{\displaystyle{ 1 - \sum_{i=1}^{k+1}\binom{k+1}{i}z^i (y-1)^{i-1} }}. \end{equation*} We also give some alternative formulations for $P_k$ which lead to some identities involving Eulerian numbers as well as proving the symmetry and unimodality of the polynomials $B_{n,k}(x)$. Many questions remain. For example, is there a more natural bijective proof for the formulas that we have derived for $B_{n,k}$ and $\mathbf B_k$? Why do permutations that are $k$-bubble sortable define the aforementioned juggling sequences? \section{Descent polynomials, bubble sort and juggling sequences} We first state some standard notation. The polynomial \begin{equation*} A_n(x) = \sum_{\pi\in\mathcal{S}_n} x^{\des(\pi)} \end{equation*} is called the $n$th \emph{Eulerian polynomial}. For instance, $A_0(x)=A_1(x)=1$ and $A_2(x)=1+x$. Note that $B_{n,k}(x) = A_n(x)$ for $k \geq n-1$, since $\maxdrop(\pi) \leq n-1$ for all $\pi \in \mathcal{S}_n$. The coefficient of $x^k$ in $A_n(x)$ is denoted $\euler{n}{k}$ and is called an \emph{Eulerian number}. It is well known that (\cite{concrete}) \begin{equation}\label{euler_gf} \frac{1-w}{e^{(w-1)z }- w } = \sum_{k,n \geq 0} \euler{n}{k} w^k \frac{z^n}{n!}. \end{equation} The Eulerian numbers are also known to be given explicitly as (\cite{euler, concrete}) \begin{equation*} \euler{n}{k} = \sum_{i=0}^n \binom{n+1}{i}(k+1-i)^n (-1)^i . \end{equation*} We define the operator $\bsort$ which acts recursively on permutations via \begin{equation*} \bsort(LnR)=\bsort(L)Rn. \end{equation*} In other words, to apply $\bsort$ to a permutation $\pi$ in $\mathcal{S}_n$, we split $\pi$ into (possibly empty) blocks $L$ and $R$ to the left and right, respectively, of the largest element of $\pi$ (which initially is $n$), interchange $n$ and $R$, and then recursively apply this procedure to $L$. We will use the convention that $\bsort(\emptyset) = \emptyset$; here $\emptyset$ denotes the empty permutation. This operator corresponds to one \emph{pass} of the classical bubble sort operation. Several interesting results on the analysis of bubble sort can be found in Knuth~\cite[pp. 106--110]{knuth}. We define the \emph{bubble sort complexity} of $\pi$ as \begin{equation*} \bsc(\pi) = \min \{ k: \bsort^k(\pi)=\mbox{id}\}, \end{equation*} the number of times $\bsort$ must be applied to $\pi$ to give the identity permutation. The following lemma is easy to prove using induction. \begin{lemma}\label{bubblesort} $\mathrm{(i)}$ For all permutations $\pi$ we have $\maxdrop(\pi) = \bsc(\pi)$.\\ $\mathrm{(ii)}$ The bubble sort operator maps $\mathcal{B}_{n,k}$ to $\mathcal{B}_{n,k-1}$. \end{lemma} The analysis of algorithms similar to bubble sort has been instrumental in generating interesting research. For example, the analysis of stack sort in Knuth\cite[pp. 242--243]{knu} gave rise to the area of pattern avoiding permutations. The stack sort operator $\ssort$ is defined by $\ssort(LnR)=\ssort(L)\ssort(R)n$. We see below that stack sort is at least as efficient as bubble sort. \begin{lemma} \label{stack} For all $\pi \in \mathcal{S}_n$, if $\bsort^k(\pi) = \mbox{id}$ then $\ssort^k(\pi)=\mbox{id}$. \end{lemma} \begin{proof} The proof of Lemma \ref{stack} follows from the following claim: \smallskip {\it If $A=a_1a_2 \dots a_n=LmR$ is a sequence of distinct positive integers and $m=\max_i a_i$, then either $\maxdrop(A) = 1-a_1$ or $\maxdrop(\ssort(A)) \leq \maxdrop(A) -1$.}\smallskip The Claim is certainly true for $n=1$. Suppose the claim is true for $n' < n$. If $\maxdrop(A)=1-a_1$, we are done. We may assume that $\maxdrop(A) > 1-a_1$. This implies that the maxdrop of $A$ does not occur at the entry where $m$ is located. For the $i$th entry in $\ssort(L)$, the maxdrop of $\ssort(L)$ at $i$ is reduced by one by induction. For the $j$th entry in $R$, the maxdrop of the corresponding entry in $\ssort(A)$ is reduced by $1$. Thus, the claim is proved by induction. \end{proof} The class of permutations $\mathcal{B}_{n,k}$ appears in a recent paper \cite{jug} on enumerating juggling patterns that are usually called \emph{siteswaps} by (mathematically inclined) jugglers. Suppose a juggler throws a ball at time $i$ so that the ball will be in the air for a time $t_i$ before landing at time $t_i +i$. Instead of an infinite sequence, we will consider periodic patterns, denoted by $T=(t_1, t_2, \dots, t_n)$. A \emph{juggling sequence} is just one in which two balls never land at the same time. It is not hard to show \cite{jug0} that a necessary and sufficient condition for a sequence to be a juggling sequence is that all the values $t_i+i \pmod n$ are distinct. In particular, it follows that that the average of $t_i$ is just the numbers of balls being juggled. Here is an example: If $T=(3,5,0,2,0)$ then at time 1 a ball is thrown that will land at time $1+3=4$. At time 2 a ball is thrown that will land at time $2+5=7$. At time 3 a ball is thrown that will land at time $3+0=3$. Alternatively one can say that no ball is thrown at time 3. This is represented in the following diagram. \ \\[0.6em] \centerline{\scalebox{0.75}{\includegraphics{example1}}} Repeating this for all intervals of length 5 gives \ \\[0.9em] \centerline{\scalebox{0.75}{\includegraphics{example2}}} For a given juggling sequence, it is often possible to further decompose into shorter juggling sequences, called \emph{primitive juggling sequences}, which themselves cannot be further decomposed. These primitive juggling sequences act as basic building blocks for juggling sequences \cite{jug}. However, in the other direction, it is not always possible to combine primitive juggling sequences into a longer juggling sequence. Nevertheless, if primitive juggling sequences share a common \emph{state} (which one can think of as a \emph{landing schedule}), we then can combine them to form a longer and more complicated juggling sequences. In \cite{jug}, primitive juggling sequences associated with a specified state are enumerated. Here we mention the related fact concerning $\mathcal{B}_{n,k}$:\smallskip {\it There is a bijection mapping permutations in $\mathcal{B}_{n,k}$ to primitive juggling sequences of period $n$ with $k$ balls that all share a certain state, called the ground state.}\smallskip The bijection maps $\pi$ to $\phi(\pi)= (t_1, \dots, t_n) $ with $t_i = k-i+\pi_i$. As a consequence of the above fact and Lemma \ref{bubblesort}, we can use bubble sort to transform a juggling sequence using $k$ balls to a juggling sequence using $k-1$ balls. To make this more precise, let $T=(t_1,\dots, t_n)$ be a juggling sequence that corresponds to $\pi \in \B_{n,k}$, and suppose that $T'=(s_1,\dots , s_n)$ is the juggling sequence that corresponds to $\bsort(\pi)$. Assume that the ball $B$ thrown at time $j$ is the one that lands latest out of all the $n$ throws. In other words, $t_j+j$ is the largest element in $\{t_i+i\}_{i=1}^n$. Now, write $T=L t_j R $ where $L=(t_1,\dots,t_{j-1})$ and $R=(t_{j+1},\dots,t_n)$. Then we have \begin{equation*} T' = f_k(T) = f_k(L)R\hspace{0.7pt}s, \end{equation*} where $s = t_j + j - (n+1)$. In other words, we have removed the ball $B$ thrown at time $j$ and thus throw all balls after time $j$ one time unit sooner. Then at time $n$ we throw the ball B so that it lands one time unit sooner than it would have originally landed. Then we repeat this procedure to all the balls thrown before time $j$. \section{The polynomials $B_{n,k}(y)$} In this section we will characterise the polynomials $B_{n,k}(y)$. This is done by first finding a recurrence for the polynomials and then solving the recurrence by exploiting some aspects of their associated characteristic polynomials. The latter step is quite involved and so we present the special case dealing with $B_{n,4}(y)$ first. \subsection{Deriving the recurrence for $B_{n,k}$} We will derive the following recurrence for $B_{n,k}(y)$. \begin{theorem}\label{b_rec_thm} For $n \geq 0$, \begin{equation}\label{rec} B_{n+k+1,k} (y) = \sum_{i=1}^{k+1} \binom{k+1}{i} (y-1)^{i-1} B_{n+k+1-i,k} (y) \end{equation} with the initial conditions \begin{equation*} B_{i,k}(y) = A_i(y),\quad 0 \leq i \leq k. \end{equation*} \end{theorem} We use the notation $[a,b] =\{i\in\mathbb{Z} :a\leq i\leq b\}$ and $[b]=[1,b]$. Let $A=\{a_1,\dots,a_n\}$ with $a_1<\dots<a_n$ be any finite subset of $\mathbb{N}$. The \emph{standardization} of a permutation $\pi$ on $A$ is the permutation $\st(\pi)$ on $[n]$ obtained from $\pi$ by replacing the integer $a_i$ with the integer $i$. Thus $\pi$ and $\st(\pi)$ are order isomorphic. For example, $\st(19452) = 15342$. If the set $A$ is fixed, the inverse of the standardization map is well defined, and we denote it by $\st^{-1}_A(\sigma)$; for instance, with $A=\{1,2,4,5,9\}$, we have $\st^{-1}_A(15342)=19452$. Note that $\st$ and $\st^{-1}_A$ each preserve the descent set. For any set $S \subseteq [n-1]$ we define $\mathcal{A}_{n,k}(S) = \{ \pi\in\B_{n,k} : \Des(\pi) \supseteq S \}$ and \begin{equation*} t_n(S) = \max\{ i\in\mathbb{N} : [n-i,n-1] \subseteq S\}. \end{equation*} Note that $\tl_n(S)=0$ in the case that $n-1$ is not a member of $S$. Now, for any permutation $\pi=\pi_1\dots\pi_n$ in $\mathcal{A}_{n,k}(S)$ define \begin{equation*} f(\pi) = (\sigma, X),\;\text{ where } \sigma=\st(\pi_1\dots\pi_{n-i-1}), X=\{\pi_{n-i},\dots ,\pi_n\} \text{ and } i=\tl_n(S). \end{equation*} \begin{example} Let $S=\{3,7,8\}$, and choose the permutation $\pi = 138425976$ in $\mathcal{A}_{9,3}(S)$. Notice that $\Des(\pi)=\{3,4,7,8\} \supset S$. Now $\tl_9(S)=2$. This gives $f(\pi) = (\sigma, X)$ where $\sigma = \st(138425) = 136425$ and $X=\{\pi_7, \pi_8,\pi_9\}=\{6,7,9\}$. Hence $f(138425976)=(136425, \{6,7,9\})$. \end{example} \begin{lemma} For any $\pi$ in $\mathcal{A}_{n,k}(S)$, the image $f(\pi)$ is in the Cartesian product \begin{equation*} \mathcal{A}_{n-i-1,k}(S\cap [n-\tl_n(S)-2])\times\binom{[n-k,n]}{\tl_n(S)+1}, \end{equation*} where $\binom{X}{m}$ denotes that set of all $m$-element subsets of the set $X$. \end{lemma} \begin{proof} Given $\pi\in\mathcal{A}_{n,k}(S)$, let $f(\pi)=(\sigma,X)$. Suppose $i=\tl_n(S)$. Then there are descents at positions $n-i, \dots , n-1$ (this is an empty sequence in case $i=0$). Thus \begin{equation*} n\geq\pi_{n-i}>\pi_{n-i+1}>\dots > \pi_{n-1}>\pi_n\geq n-k, \end{equation*} where the last inequality follows from the assumption that $\maxdrop(\pi)\leq k$. Hence $X$ is an $(i+1)$-element subset of $[n-k,n]$, as claimed. Clearly $\sigma\in\mathcal{S}_{n-i-1}$. Next we shall show that $\sigma$ is in $\mathcal{A}_{n-i-1,k}$. Notice that the entries of $(\pi_1,\dots, \pi_{n-i-1})$ that do not change under standardization are those $\pi_{\ell}$ which are $<\pi_n$. Since these values remain unchanged, the values $\ell-\pi_{\ell}$ are also unchanged and are thus $\leq k$. Let $(\pi_{a(1)},\dots, \pi_{a(m)})$ be the subsequence of values which are $>\pi_n$. The smallest value that any of these may take after standardization is $\pi_n \geq n-k$. So $\sigma_{a(j)} \geq \pi_n \geq n-k$ for all $j \in [1,m]$. Thus $a(j)-\sigma_{a(j)} \leq a(j) - (n-k) = k-(n-a(j)) \leq k$ for all $j \in [1,m]$. Therefore $\ell - \sigma_{\ell} \leq k$ for all $\ell \in [1,n-i-1]$ and so $\sigma \in \mathcal{A}_{n-i-1,k}$. The descent set is preserved under standardization, and consequently $\sigma$ is in $\mathcal{A}_{n-i-1,k}(S\cap [n-i-2])$, as claimed. \end{proof} We now define a function $g$ which will be shown to be the inverse of $f$. Let $\pi$ be a permutation in $\mathcal{A}_{m,k}(T)$, where $T$ is a subset of $[m-1]$. We will add $i+1$ elements to $\pi$ to yield a new permutation $\sigma$ in $\mathcal{A}_{m+i,k}(T \cup [m+1,m+i])$. Choose any $(i+1)$-element subset $X$ of the interval $[m+i+1-k,m+i+1]$, and let us write $X=\{x_1,\dots , x_{i+1}\}$, where $x_1\leq\dots\leq x_{i+1}$. Define \begin{equation*} g(\pi,X)=\st^{-1}_V (\pi_1\dots\pi_m) \, x_{i+1}x_i\dots x_1,\, \text{ where }V=[m+i+1]\setminus X. \end{equation*} \begin{example} Let $T=\{1\}$, and choose the permutation $\pi=3142$ in $\mathcal{A}_{4,3} (T)$. Notice that $\Des(\pi)=\{1,3\}\supseteq T$. Choose $i=2$ and select a subset $X$ from $[4+2+1-3,4+2+1]=\{4,5,6,7\}$ of size $i+1=3$. Let us select $X=\{4,6,7\}$. Now we have $g(\pi,X) =\st^{-1}_V(3142)\,764=3152764$, where $V$ is the set $[4+2+1]\setminus \{4,6,7\} = \{1,2,3,5\}$. \end{example} \begin{lemma} If $(\pi,X)$ is in the Cartesian product \begin{equation*} \mathcal{A}_{m,k}(T)\times\binom{[m+i+1-k,m+i+1]}{i+1} \end{equation*} for some $i>0$ then $g(\pi,X)$ is in \begin{equation*} \mathcal{A}_{m+i+1,k}(T\cup [m+1, m+i]). \end{equation*} \end{lemma} \begin{proof} Let $\sigma=g(\pi,X)$. For the first $m$ elements of $\sigma$, since $\sigma_j \geq \pi_j$ for all $1\leq j \leq m$, we have $j-\sigma_j \leq j-\pi_j$ which gives \begin{equation*} \max\{j-\sigma_j: j\in[m]\}\leq\max\{j-\pi_j : j\in[m]\}\leq k. \end{equation*} The final $i+1$ elements of $\sigma$ are decreasing so the $\maxdrop$ of these elements will be the $\maxdrop$ of the final element, \begin{equation*} m+i+1-\sigma_{m+i+1} = m+i+1 - x_1 \leq m+i+1-(m+i+1-k) = k. \end{equation*} Thus $\maxdrop(\sigma) \leq k$ and so $\sigma\in\B_{m+i+1,k}$. The descents of $\sigma$ will be in the set $T \cup [m+1,m+i]$ since descents are preserved under standardization and the final $i+1$ elements of $\sigma$ are listed in decreasing order. Hence $\sigma\in A_{m+i+1,k}(T\cup [m+1, m+i])$, as claimed. \end{proof} \begin{lemma} The function $f$ is a bijection, and $g$ is its inverse. \end{lemma} \begin{proof} Given any $(\sigma,X) \in \binom{[m+j+1-k,m+j+1]}{j+1} \times\mathcal{A}_{m,k}(T)$ where $T \subseteq [m-1]$, let $\pi=g(\sigma,X)$. We have \begin{equation*} \pi=\st^{-1}_{[m+j+1]\setminus X}(\sigma_1\dots\sigma_m) \,x_{j+1}x_j\dots x_1 \in \mathcal{A}_{m+j+1,k}(T\cup[m+1,m+j]), \end{equation*} where $X=\{x_1,\dots, x_{j+1}\}$ and $x_1\leq\dots\leq x_{j+1}$. Let $S=T\cup [m+1,m+j]$. Clearly $\Des(g(\sigma,X)) \supseteq S$ and $i=\tl_{m+j+1}(S)=j$. So $f(g(\sigma,X)) = (\tau,Y)$ where \begin{equation*} Y=\{\pi_{m+1},\dots , \pi_{m+j+1}\} = \{x_1,\dots, x_{j+1}\}= X \end{equation*} and \begin{equation*} \tau = \st(\st^{-1}_{[1,m+j+1]\setminus X}(\sigma_1\dots\sigma_m)) = \sigma. \end{equation*} Hence $f(g(\sigma,X))=(\sigma,X)$. Given $\pi\in\mathcal{A}_{n,k}(S)$, let $f(\pi)=(\sigma,X)$ with $X=\{x_1,\dots , x_{i+1}\}$ and $\sigma = \st(\pi_1\dots\pi_{n-(i+1)})$. We have \begin{align*} g(\sigma,X) &=\st^{-1}_{[1,n]\setminus X} (\sigma_1\dots\sigma_{n-i-1}) x_{i+1}\dots x_1\\ &= \st^{-1}_{[1,n]\setminus X}(\st(\pi_1\dots\pi_{n-i-1})) \pi_{n-i}\dots \pi_n \\ &= \pi_1\dots \pi_{n-i-1}\pi_{n-i}\dots \pi_n \\ &= \pi. \end{align*} Hence $g(f(\pi))=\pi$. \end{proof} \begin{corollary}\label{rec_corol} Let $a_{n,k}(S) = |\mathcal{A}_{n,k}(S)|$ and $i=\tl_n(S)$. Then \begin{equation*} a_{n,k}(S) = \binom{k+1}{i+1} a_{n-(i+1),k} (S\cap [1,n-(i+1)]). \end{equation*} \end{corollary} \begin{proposition}\label{prop_rec} For all $n\geq0$, $$\B_{n,k}(y+1) = \sum_{i=1}^{k+1} {k+1 \choose i} y^{i-1} \B_{n-i,k} (y+1). $$ \end{proposition} \begin{proof} Notice that \begin{align*} \B_{n,k}(y+1) &= \sum_{\pi \in \B_{n,k}} (y+1)^{\des(\pi)} \\ &= \sum_{\pi \in \B_{n,k}} \sum_{i=0}^{\des(\pi)} \binom{\des(\pi)}{i} y^i \\ &= \sum_{\pi \in \B_{n,k}} \sum_{S\subseteq \Des(\pi)} y^{|S|}\\ &= \sum_{S\subseteq [n-1]} y^{|S|} \sum_{\pi \in \mathcal{A}_{n,k}(S)} 1 = \sum_{S\subseteq [n-1]} y^{|S|} a_{n,k}(S). \end{align*} From Corollary \ref{rec_corol}, multiply both sides by $y^{|S|}$ and sum over all $S \subseteq [n-1]$. We have \begin{align*} \B_{n,k}(y+1) &= \sum_{S \subseteq [n-1]} y^{|S|} \binom{k+1}{\tl_n(S)+1} a_{n-(\tl_n(S)+1),k} (S \cap [n-(\tl_n(S)+2)]) \\ &= \sum_{i \geq 0 } \sum_{S \subseteq [n-1]\atop \tl_n(S)=i} y^i y^{|S|-i} \binom{k+1}{i+1} a_{n-(i+1),k} (S \cap [n-(i+2)]) \\ &= \sum_{i \geq 0 } \binom{k+1}{i+1} y^i \sum_{S \subseteq [n-1]\atop \tl_n(S)=i} a_{n-(i+1),k} (S \cap [n-(i+2)]) y^{|S|-i}\\ &= \sum_{i \geq 0 } \binom{k+1}{i+1} y^i \sum_{S \subseteq [n-(i+1)]} a_{n-(i+1),k} (S) y^{|S|}\\ &= \sum_{i \geq 0 } \binom{k+1}{i+1} y^i \B_{n-(i+1),k} (y+1) \\ &= \sum_{i \geq 1 } \binom{k+1}{i} y^{i-1} \B_{n-i,k} (y+1). \end{align*} \end{proof} \begin{proof}[Proof of Theorem \ref{b_rec_thm}] Replacing $n$ and $y$ by $n+k+1$ and $y-1$, respectively, in Proposition \ref{prop_rec} yields the recurrence (\ref{rec}): \begin{equation*} B_{n+k+1,k}(y) = \sum_{i=1}^{k+1} \binom{k+1}{i} (y-1)^{i-1} B_{n+k+1-i,k} (y) \end{equation*} for $n \geq 0$, with the initial conditions $B_{i,k}(y) = A_i(y),\,0 \leq i \leq k$. \end{proof} Consequently, by multiplying the above recurrence by $z^n$ and summing over all $n\geq 0$, we have the generating function $\mathbf B_k(z,y)$: \begin{equation}\label{BB_gen} \mathbf B_k(z,y) = \dfrac{\displaystyle{ 1+\sum_{t=1}^k \left( A_t(y) - \sum_{i=1}^t \binom{k+1}{i} (y-1)^{i-1} A_{t-i}(y) \right)z^t }}{\displaystyle{ 1 - \sum_{i=1}^{k+1}\binom{k+1}{i}z^i (y-1)^{i-1} }}. \end{equation} \subsection{Solving the recurrence for $B_{n,4}$.} Before we proceed to solve the recurrence for $B_{n,k}$, we first examine the special case of $k=4$ which is quite illuminating. We note that the characteristic polynomial for the recurrence for $B_{n,4}$ is \begin{align*} h(z) &= z^5 - 5z^4+10(1-y)z^3 - 10(1-y)^2z^2 + 5(1-y)^3z -(1-y)^4\\ &= \frac{(z-1+y)^5 -yz^5}{1-y}. \end{align*} Substituting $y = t^5$ in the expression above, we see that the roots of $h(z)$ are just \begin{equation*} \rho_j(t) = \frac{1-t^5}{1-\omega^j t},\quad 0\leq j \leq 4, \end{equation*} where $\omega = \exp(\frac{2 \pi \mathfrak{i}}{5})$ is a primitive $5th$ root of unity. Hence, the general term for $B_{n,4}(t)$ can written as \begin{equation*} B_{n,4}(t) = \sum_{i=0}^4 \alpha_i(t) \rho_i^n(t) \end{equation*} where the $\alpha_i(t)$ are appropriately chosen coefficients (polynomials in $t$). To determine the $\alpha_i(t)$ we need to solve the following system of linear equations: \begin{equation*} \sum_{i=0}^4 \alpha_i(t) \rho_i^j(t) = B_{j,4}(t) = A_j(t^5), \, 0 \leq j \leq 4. \end{equation*} Thus, $\alpha_i(t)$ can be expressed as the ratio $N_{4,i+1}(t)/D_4(t)$ of two determinants. The denominator $D_4(t)$ is just a standard Vandermonde determinant whose $(i+1,j+1)$ entry is $\rho_i^j(t)$. The numerator $N_{4,i+1}(t)$ is formed from $D_4(t)$ by replacing the elements $\rho_i^j(t)$ in the $(i+1)$st row by $A_j(t^5)$. A quick computation (using the symbolic computation package Maple) gives: \begin{align*} D_4(t) &= 25 \sqrt 5 \,(1-t^5)^6 t^{10}; \\ N_{4,1}(t) &=5 \sqrt 5 \,(t^{12}+t^{10}+2t^9+4t^8+4t^7+4t^5+4t^4+2t^3+t^2+1) (1-t^5)^3 (1-t)^3t^{10} \end{align*} and, in general, $N_{4,i+1}(t) = N_{4,1}(\omega^i t).$ Substituting the value $\alpha_0 (t) = N_{4,1}(t)/D_4(t)$ into the first term in the expansion of $B_{n,4}$, we get \begin{multline*} \alpha_0 (t)(1+t+t^2+t^3+t^4)^n \\ =\tfrac{1}{5}(t^{12}+t^{10}+2t^9+4t^8+4t^7+4t^5+4t^4+2t^3+t^2+1) (1+t+t^2+t^3+t^4)^{n-3}. \end{multline*} Now, since the other four terms $\alpha_i(t)(1+t+t^2+t^3+t^4)^n$ arise by replacing $t$ by $\omega ^i t$ then in the sum of all five terms, the only powers of $t$ that survive are those which have powers which are multiples of $5$. Thus, we can conclude that if we write \begin{equation*} (t^{12}+t^{10}+2t^9+4t^8+4t^7+4t^5+4t^4+2t^3+t^2+1)(1+t+t^2+t^3+t^4)^{n-3} = \sum_r \beta(r) t^r \end{equation*} then $b_{n,4}(d) = \beta(5d)$. In other words, the number of permutations $\pi \in \mathcal{B}_{n,4}$ with $d$ descents is given by the coefficient of $t^{5d}$ in the expansion of the above polynomial. Incidentally, the corresponding results for the earlier $\mathcal{B}_{n,i}$ are as follows: $b_{n,1}(d) = \beta(2d)$ in the expansion of \begin{equation*} (1+t)^n = \sum_r \beta(r) t^r , \end{equation*} so $b_{n,1}(d)=\binom{n}{2d}$; $b_{n,2}(d) = \beta(3d)$ in the expansion of \begin{equation*} (1+t^2)(1+t)^{n-1} = \sum_r \beta(r) t^r; \end{equation*} and $b_{n,3}(d) = \beta(4d)$ in the expansion of \begin{equation*} (1+t^2+2t^3+t^4+t^6)(1+t)^{n-2} = \sum_r \beta(r) t^r. \end{equation*} The preceding arguments have now set the stage for dealing with the general case of $B_{n,k}$. Of course, the arguments will be somewhat more involved but it is hoped that treating the above special case will be a useful guide for the reader. \subsection{Solving the recurrence for $B_{n,k}$} \begin{theorem} We have $B_{n,k}(y) = \sum_d \beta_k\big((k+1)d\big) y^{(k+1)d}$, where \begin{equation*} \sum_j \beta_k(j) u^j = P_k(u) \left( \frac{1-u^{k+1}}{1-u} \right)^{n-k} \end{equation*} and \begin{equation*} P_k(u) = \sum_{j=0}^k A_{k-j}(u^{k+1}) (u^{k+1}-1)^{j} \sum_{i=j}^k \binom{i}{j} u^{-i}. \end{equation*} \end{theorem} \begin{proof} To solve (\ref{rec}) for $B_{n,k}$, we first need to compute the roots of the corresponding characteristic polynomial \begin{equation*} z^{n+k+1} - \sum_{i=1}^{k+1} \binom{k+1}{i} (y-1)^{i-1} z^{n+k+1-i} \end{equation*} which can be rewritten as \begin{equation*} \big(z - (1-y)\big)^{k+1} - yz^{k+1}. \end{equation*} Substituting $y = u^{k+1}$, this becomes \begin{equation*} \big(z-(1-u^{k+1})\big)^{k+1} - u^{k+1}z^{k+1}. \end{equation*} The $k+1$ roots of this polynomial are easily seen to be the expressions \begin{equation*} \rho_i = \rho_i(u) = \frac{1-u^{k+1}}{1-\omega^i u},\quad 0 \leq i \leq k, \end{equation*} where $\omega = \exp(\frac{2 \pi \mathfrak{i}}{k+1})$ is a primitive $(k+1)$st root of unity. Thus, we can express $B_{n,k}$ in the form \begin{equation*} B_{n,k}(u^{k+1}) = \sum_{i=0}^k \alpha_{k,i}(u) \rho_i^n \end{equation*} for an appropriate choice of coefficients $\alpha_{k,i}$ (which depend on the initial conditions). In fact, writing down the expressions for the first $k+1$ $B_{n,k}$'s, we have: \begin{equation*} B_{j,k}(u^{k+1}) = \sum_{i=0}^k \alpha_{k,i}(u) \rho_i^j = A_j(u^{k+1}),\quad \mbox{for } 0 \leq j \leq k. \end{equation*} We can solve this as a system of $k+1$ linear equations in the $k+1$ unknown coefficients $\alpha_{k,i}(u)$ by representing the solution in the usual way as a ratio of two determinants. In particular, the expression for $\alpha_{k,0}$ is given by \begin{equation}\label{alpha_0} \alpha_{k,0}(u) = \frac{\det R_k(u)}{\det S_k(u)} \end{equation} where $S_k(u)$ and $R_k(u)$ are ($k+1$) by ($k+1$) matrices defined by \begin{equation*} S_k(u) = (S_k(i+1,j+1))\;\text{ with }\,S_k(i+1,j+1) = \rho_i^j \end{equation*} and \begin{equation*} R_k(u) =\big(R_k(i+1,j+1)\big) \;\text{ with }\, R_k(i+1,j+1) = \begin{cases} A_{j}(u^{k+1})& \text{ if } i = 0, \\ \rho_i^j & \text{ if } i > 0. \end{cases} \end{equation*} Now the bottom determinant is a standard Vandermonde determinant which has the value \begin{equation*} \det S_k(u) = \prod_{0 \leq i < j \leq k} (\rho_j - \rho_i). \end{equation*} The top determinant is almost a Vandermonde determinant (except for the first row). Its value has the form \begin{equation*} \det R_k(u) = \mathsf{Top}_k(u) \cdot \prod_{0 < i < j \leq k} (\rho_j - \rho_i) \end{equation*} where $\mathsf{Top}_k(u)$ is a polynomial in $u$ which we will soon determine. Hence, in the ratio (\ref{alpha_0}), the terms which do not involve $\rho_0$ cancel, leaving the reduced form \begin{equation*} \alpha_{k,0}(u) = \frac{\mathsf{Top}_k(u)}{\prod_{j > 0} (\rho_j - \rho_0)}. \end{equation*} However, we have \begin{align} \prod_{j > 0} (\rho_0 - \rho_j) &= \prod_{j>0} \left( \frac{1-u^{k+1}}{1-u} - \frac{1-u^{k+1}}{1-\omega^j u} \right)\nonumber \\ &= \left( \prod_{j>0} \frac{(1-\omega^j) u}{(1-u)(1-\omega^j u)}\right) (1-u^{k+1})^k \nonumber \\ &= \frac{(1-u^{k+1})^k u^k}{(1-u)^k}\cdot \frac{\prod_{j>0} (1-\omega^j) }{\prod_{j>0} (1-\omega^j u)} \nonumber\\ &= \frac{(1-u^{k+1})^k u^k}{(1-u)^k}\cdot \frac{k+1}{\frac{1-u^{k+1}}{1-u}} = \frac{(1-u^{k+1})^{k-1} u^k}{(1-u)^{k-1}} \cdot (k+1). \label{bot2} \end{align} On the other hand, for the top we have by standard properties of Vandermonde determinants: \begin{equation}\label{top} \mathsf{Top}_k(u) = \sum_{j \geq 0} (-1)^{k-j} A_{k-j}(u^{k+1}) \mathsf{S}_{k,j} (\rho_1,\dots, \rho_k), \end{equation} where $\mathsf{S}_{k,j}(x_1, \dots, x_k)$ is the elementary symmetric function of degree $j$ in the $k$ variables $x_1, \dots, x_k$ (and we recall that $A_t(y) = \sum_{j \geq 0} \euler{t}{j} y^t$). Now consider the generating function \begin{align*} X_k(z) &= (z-\rho_0) (z-\rho_1) \dots (z-\rho_k) \\ &= \sum_{t \geq 0}(-1)^t\mathsf{S}_{k+1,t}(\rho_0,\dots,\rho_k)z^{k+1-t} \\ &= \prod_{i \geq 0}\left(z - \frac{1-u^{k+1}}{1-\omega^i u} \right) \\ &= \frac{\prod_{i \geq 0} (z-(1-u^{k+1}) - \omega^i uz)} {\prod_{i \geq 0} (1-\omega^i u)} \\ &= \frac{\left(z-(1-u^{k+1})\right)^{k+1}-u^{k+1} z^{k+1}}{1 - u^{k+1}}. \end{align*} What we are interested in is \begin{align*} Y_k(z) = \frac{X_k(z)}{z-\rho_0} &= \sum_{i \geq 0} (-1)^i\mathsf{S}_{k,i}(\rho_1,\dots,\rho_k) z^{k-i}\\ &= \frac{\left( z-(1-u^{k+1})\right)^{k+1} - u^{k+1}z^{k+1}}{(1-u^{k+1})\big(z - \frac{1-u^{k+1}}{1-u}\big)}\\ &= \frac{1-u}{1-u^{k+1}} \sum_{i=0}^k (z-1+u^{k+1})^i (uz)^{k-i} \\ &= \frac{1-u}{1-u^{k+1}} \sum_{i=0}^k \sum_{j=0}^i(-1)^j\binom{i}{j}(1 - u^{k+1})^j u^{k-i}z^{k-j}\\ &= \frac{1-u}{1-u^{k+1}} \sum_{j=0}^k(1 - u^{k+1})^j z^{k-j} \sum_{i = j}^k (-1)^j \binom{i}{j} u^{k-i}. \end{align*} Thus, by identifying the coefficient of $z^{k-j}$, we have \begin{equation*} \mathsf{S}_{k,j}(\rho_1, \dots, \rho_k) = \frac{(1-u)(1 - u^{k+1})^{j}}{1-u^{k+1}}\sum_{i=j}^k\binom{i}{j}u^{k-i}. \end{equation*} Therefore, we have \begin{align*} \mathsf{Top}_k(u) &= \sum_{j \geq 0}(-1)^{k-j}A_{k-j}(u^{k+1}) \mathsf{S}_{k,j}(\rho_1,\dots, \rho_k) \\ &= \sum_{j \geq 0} (-1)^{k-j}A_{k-j}(u^{k+1}) \frac{(1-u)(1 - u^{k+1})^{j}}{1-u^{k+1}}\sum_{i=j}^k\binom{i}{j}u^{k-i}. \end{align*} As a consequence, we find by (\ref{top}) and (\ref{bot2}) that \begin{align}\label{ex} \lefteqn{\alpha_{k,0}(u) \rho_0^n} \nonumber\\ &= \frac{\mathsf{Top}_k(u)}{(-1)^k \prod_{j > 0} (\rho_0 - \rho_j)} \left(\frac{1-u^{k+1}}{1-u}\right)^{\!n}\\ &= \frac{1}{(k+1)u^k}\left( \frac{1-u^{k+1}}{1-u} \right)^{\!n-k} \sum_{j=0}^k(-1)^{j}A_{k-j}(u^{k+1}) (1-u^{k+1})^{j} \sum_{i=j}^k \binom{i}{j} u^{k-i}\nonumber\\ &= \frac{1}{(k+1)} \left( \frac{1-u^{k+1}}{1-u} \right)^{\!n-k} \sum_{j=0}^kA_{k-j}(u^{k+1}) (u^{k+1}-1)^{j} \sum_{i=j}^k \binom{i}{j} u^{-i}.\nonumber \end{align} (We should keep in mind that this expression actually is a polynomial in $u$.) To determine the other coefficients $\alpha_{k,t}(u)$, we make the following observations. First, for $1 \leq t \leq k$, we can cyclically permute the rows of the (Vandermonde) matrix $S_k(u)$ to form the new matrix $S_k^{(t)}(u) = (S_k^{(t)}(i+1,j+1)) = (\rho_i^{j+t})$. We can then form the corresponding matrix $R_k^{(t)}(u)$ by replacing the top row of $S_k^{(t)}(u)$ by $A_0(u^{k+1}), \dots, A_k(u^{k+1})$. In this way, we can express the coefficient $\alpha_{k,t}(u)$ as: \begin{equation}\label{alpha_t} \alpha_{k,t}(u) = \frac{\det R_k^{(t)}(u)}{\det S_k^{(t)}(u)}. \end{equation} However, observe that the resulting computations for determining $\alpha_{k,t}(u)$ are exactly the same as those for $\alpha_{k,0}(u)$ where we replace $u$ by $\omega^t u$. This is because $A_j(u^{k+1}) = A_j(\omega^t u)^{k+1}$. Consequently, \begin{equation*} \alpha_{k,t}(u) = \alpha_{k,0}(\omega^t u). \end{equation*} Therefore, for each $n$, the expression $\alpha_{k,t}(u) \rho_t^n$ as a polynomial in $u$ can be obtained from $ \alpha_{k,0}(u) \rho_0^n $ by replacing $u$ by $\omega^t u$. However, this implies that in the sum $ \sum_{t=0}^k \alpha_{k,t}(u) \rho_t^n$, the only terms that survive are those powers $u^m$ of $u$ which are multiples of $k+1$, since for $m \not\equiv 0 \pmod {k+1}$, we have $\sum_{i=0}^k \omega^{mi} = 0$. One the other hand, for $u^m$ with $m \equiv 0 \pmod {k+1}$, we must multiply the coefficients by $k+1$ since in this case, all the powers $\omega^t$ are $1$. Consequently, if we write \begin{equation}\label{final} (k+1) \alpha_{k,0}(u) \rho_0^n = \sum_j \beta_k(j) u^j \end{equation} then we have \begin{equation*} B_{n,k}(y) = \sum_d \beta_k\big(\,(k+1)d\,\big)\, y^{(k+1)d}. \end{equation*} In other words, the number $b_{n,k} (r)$ of permutations in $\mathcal{B}_{n,k}$ having exactly $r$ descents is equal to the $k+1$ times the coefficient of $u^{(k+1)r}$ in (\ref{ex}). \end{proof} If we express (\ref{ex}) (times $k+1$) in the form \begin{equation} P_k(u) \left(1+u+\dots+u^k\right)^{n-k} \end{equation} then the first few values of $P_k(u)$ are shown below. \begin{equation*} \begin{array}{c|l} k & P_k(u) \\ \hline 1 & 1\\ 2 & 1+u\\ 3 & 1+u+2u^2+u^3+u^4\\ 4 & 1+u+2u^2+4u^3+4u^4+4u^5+4u^6+2u^7+u^8+u^{9}\\ 5 & 1+u+2u^2+4u^3+8u^4+11u^5+11u^6+14u^7+16u^{8}+\\ & +14u^9+11u^{10}+11u^{11}+8u^{12}+4u^{13}+2u^{14}+u^{15}+u^{16} \end{array} \end{equation*} There is clearly a lot of structure in the polynomials $P_k(u)$ which will be discussed in the next section. \section{The structure of $P_k(u)$} Let us first write down the expression for $P_k(u)$ which came from (\ref{ex}): \begin{equation}\label{P} P_k(u) = \sum_{j=0}^k (-1)^{j}A_{k-j}(u^{k+1}) (1-u^{k+1})^{j} \sum_{i=j}^k \binom{i}{j} u^{-i}. \end{equation} If we write $P_k(u) = \sum_{i=0}^{k^2} \alpha_i u^i$, we will define the \emph{stretch} of $P_k(u)$ to be \begin{equation*} PP_k(u) = \alpha_0 + \sum_{i=0}^k \sum_{j=0}^{k-2} \alpha_{1+i+(k+1)} u^{2+i+(k+1)j+j}. \end{equation*} What this does to $P_k(u)$ is to insert $0$ coefficients at every $(k+1)$st term, starting after $\alpha_0$. Thus, the stretched polynomials corresponding to the values of $P_k(u)$ given in the array above are: \begin{equation*} \begin{array}{c|l} k & PP_k(u) \\ \hline 1 & 1\\ 2 & 1+u^2\\ 3 & 1+u^2+2u^3+u^4+u^6\\ 4 & 1+u^2+2u^3+4u^4+4u^5+4u^7+4u^8+2u^9+u^{10}+u^{12}\\ 5 & 1+u^2+2u^3+4u^4+8u^5+11u^6+11u^8+14u^9+16u^{10}+\\ & +14u^{11}+11u^{12}+11u^{14}+8u^{15}+4u^{16}+2u^{17}+u^{18}+u^{20} \end{array} \end{equation*} Note that if $P_k(u)$ has degree $k^2$ then $PP_k(u)$ has degree $k^2 + k$. \begin{theorem} \label{th2} For all $k \geq 1$, \begin{equation*} P_{k+1}(u) = PP_k(u)\cdot (1 + u + u^2 + \dots + u^{k+1}). \end{equation*} \end{theorem} \begin{proof} From (\ref{P}) and the definition of $PP_k(u)$, we can write \begin{equation*} PP_k(u) = \sum_{t=0}^k (-1)^t (1-u^{k+2})^t A_{k-t}(u^{k+2}) \sum_{s=t}^k \binom{s}{t} u^{-s}. \end{equation*} We want to show that \begin{equation*} P_{k+1}(u) = PP_k(u) \cdot \frac{1-u^{k+2}}{1-u}. \end{equation*} Thus, it is enough to prove that $A = B$ where \begin{align*} A &= PP_{k+1}(u) (1-u^{k+2}) \\ &= \sum_{t=0}^k (-1)^t (1-u^{k+2})^{t+1} A_{k-t}(u^{k+2})\sum_{s=t}^k\binom{s}{t} u^{-s} \shortintertext{and} B &= P_k(u) (1-u) \\ &= \sum_{t=0}^{k+1} (-1)^t (1-u^{k+2})^{t} A_{k+1-t}(u^{k+2}) \sum_{s=t}^{k+1} \binom{s}{t} u^{-s} (1-u). \end{align*} Observe that \begin{align*} B &= \sum_{t=0}^{k+1} (-1)^t (1-u^{k+2})^{t} A_{k+1-t}(u^{k+2} \left( \sum_{s=t}^{k+1} \binom{s}{t} u^{-s} (1-u)\right)\\ &= \sum_{t=1}^{k+1} (-1)^t (1-u^{k+2})^{t} A_{k+1-t}(u^{k+2}) \left( \sum_{s=t}^{k+1} \binom{s}{t} u^{-s} (1-u)\right)\\ & \quad + A_{k+1}(u^{k+2}) (1-u^{-k-2}) (-u)\\ &= \sum_{t=0}^{k} (-1)^t (1-u^{k+2})^{t+1} A_{k-t}(u^{k+2}) \left( \sum_{s=t+1}^{k}\binom{s}{t+1} u^{-s} (u-1)\right) \\ & \quad - u (1-u^{-k-2}) A_{k+1}\\ &= \sum_{t=0}^{k} (-1)^t (1-u^{k+2})^{t+1} A_{k-t}(u^{k+2}) \left( \sum_{s=t+1}^{k} \binom{s}{t} u^{-s} +u^{-t} - \binom{k+1}{t+1} u^{-k-1} \right) \\ &\quad - u (1-u^{-k-2}) A_{k+1}\\ &= \sum_{t=0}^{k} (-1)^t (1-u^{k+2})^{t+1} A_{k-t}(u^{k+2}) \left( \sum_{s=t}^{k} \binom{s}{t} u^{-s}-\binom{k+1}{t+1} u^{-k-1} \right) \\ & \quad - u (1-u^{-k-2}) A_{k+1}. \end{align*} Hence, to prove that $A = B$, we only need to establish \begin{multline*} \sum_{t=0}^{k} (-1)^t (1-u^{k+2})^{t+1} A_{k-t}(u^{k+2}) \left( -\binom{k+1}{t+1} u^{-k-1} \right) - u (1-u^{-k-2}) A_{k+1}\\ = u(1-u^{-k-2}) A_{k+1}(u^{k+2}). \end{multline*} However, this would follow from \begin{equation}\label{euler-id} \sum_{t=0}^{k'} (x-1)^t A_{k'-t}(x) \binom{k'}{t} = x A_{k'}(x) \end{equation} by taking $k' = k+1$ and $x = u^{k+2}$. So it remains to prove (\ref{euler-id}). To do this we will use the standard generating function for $A_n(w)$ from (\ref{euler_gf}): \begin{equation*} \sum_{n,m \geq 0} \euler{n}{m} w^m \frac{z^n}{n!} = \sum_{n \geq 0} A_n(w) \frac{z^n}{n!} = \frac{1-w}{e^{(w-1)z} - w}. \end{equation*} Consider \begin{align*} F(x,z) &= \sum_{k > 0} x A_k(x) \frac{z^k}{k!} \shortintertext{and} G(x,z) &= \sum_{k > 0} \sum_{t=0}^k (x-1)^t A_{k-t} (x) \binom{k}{t} \frac{z^k}{k!}. \end{align*} It will suffice to show that $F = G$. Now \begin{align*} G(x,z) &= \sum_{k \geq 0} \sum_{t \geq 0} (x-1)^t A_{k-t}(x) \binom{k}{t} \frac{z^k}{k!} - 1\\ &= \sum_{t \geq 0} \sum_{k \geq t} (x-1)^t A_{k-t}(x) \binom{k}{t} \frac{z^k}{k!} - 1\\ &= \sum_{t \geq 0} (x-1)^t \sum_{k' \geq 0} A_{k'}(x) \binom{k' + t}{t} \frac{z^{k' + t}}{(k' + t)!} - 1\\ &= \sum_{t \geq 0} (x-1)^t z^t \sum_{k' \geq 0}A_{k}(x) \frac{z^{k}}{k!} - 1\\ &= e^{(x-1)z} \cdot \frac{1-x}{e^{(x-1)z} - x} - 1 = \frac{x (1 - e^{(1-x)z})}{e^{(x-1)z} -x}. \end{align*} On the other hand, \begin{align*} F(x,z) &= \sum_{k > 0} x A_k(x) \frac{z^k}{k!}\\ &= x \left( \frac{1-x}{e^{(x-1)z}-x} - 1 \right) = \frac{x (1 - e^{(1-x)z})}{e^{(x-1)z} -x} \end{align*} as desired. This completes the proof of Theorem \ref{th2}. \end{proof} \begin{theorem} The coefficients of $P_k(u)$ are symmetric and unimodal. \end{theorem} \begin{proof} It follows from Theorem \ref{th2} that we can construct the coefficient sequence for $P_{k+1}(u)$ from that of $P_k(u)$ by the following rule (where we assume that all coefficients of $u^t$ in $P_k(u)$ are $0$ if $t < 0$ or $t > k^2$). Namely, suppose we write $P_k(u) = \sum_{i=0}^{k^2} \alpha_i u^i$ so that we have the coefficient sequence $A_k = (\alpha_0, \alpha_1, \dots, \alpha_{k^2})$. Now form the new sequence $B_k = (\beta_0, \beta_1, \dots \beta_{k^2 + k})$ by the rule \begin{equation*} \beta_i = \sum_{j=i-k}^i \alpha_j ,\quad 0 \leq i \leq k^2 + k. \end{equation*} Finally, starting with $\beta_0$, insert \emph{duplicate} values for the coefficients \begin{equation*} \beta_0,\beta_{k+1},\beta_{2(k+1)},\dots,\beta_{t(k+1)}, \dots,\beta_{(k-1)(k+1)}\text{ and }\beta_{k(k+1)}. \end{equation*} Thus, this will generate the sequence \begin{equation*} (\beta_0, \beta_0, \beta_1, \beta_2, \dots, \beta_k, \beta_{k+1}, \beta_{k+1}, \beta_{k+2}, \dots, \beta_{k^2+k-1}, \beta_{k^2+k}, \beta_{k^2+k}). \end{equation*} This new sequence will in fact just be the coefficient sequence $A_{k+1}$ for $P_{k+1}(u)$. For example, starting with $P_1(u) = 1 + u$, we have $A_1 = (1,1)$ and so $B_1 = (1, 2, 1)$. Now, inserting the duplicate values for $\beta_0 = 1$ and $\beta_2 = 1$, we get the coefficient sequence $A_2 = (1, {\bf 1}, 2, 1, {\bf 1})$ for $P_2(u) = 1 + u +2 u^2 + u^3 + u^4$. Repeating this process for $A_2$, we sum blocks of length $3$ to get $B_2 = (1, 2, 4, 4, 4, 2, 1)$. Inserting duplicates for entries at positions $0, 3$ and $6$ gives us the new coefficient sequence $A_3 = (1, {\bf 1}, 2, 4, 4, {\bf 4}, 4, 2, 1, {\bf 1})$ of $P_3 = 1 + u + 2u^2 + 4u^3 +4u^4 +4u^5 +4u^6 +2u^7 +u^8 +u^9$, etc. It is also clear from this procedure that if $A_k$ is symmetric and unimodal, then so is $B_k$, and consequently, so is $A_{k+1}$. This is what we claimed. \end{proof} \subsection{An Eulerian identity} Note that since $P_k(u)$ is symmetric and has degree $u^{k^2}$, we have $P_k(u) = u^{k^2} P_k({\textstyle \frac{1}{u}})$. Replacing $P_k(u)$ by its expression in (\ref{P}), we obtain (with some calculation) the interesting identity \begin{equation*} \sum_{j = 0}^{a+b} (-1)^j \binom{a}{j} (1-x)^j A_{a+b-j}(x) = x \sum_{j = 0}^{a+b} \binom{b}{j}(1-x)^j A_{a+b-j}(x) +\binom{b}{a+b}(1-x)^{a+b+1} \end{equation*} for \emph{all} integers $a$ and $b$ provided that $a + b \geq 0$.

https://arxiv.org/abs/1011.3027

Introduction to the non-asymptotic analysis of random matrices

This is a tutorial on some basic non-asymptotic methods and concepts in random matrix theory. The reader will learn several tools for the analysis of the extreme singular values of random matrices with independent rows or columns. Many of these methods sprung off from the development of geometric functional analysis since the 1970's. They have applications in several fields, most notably in theoretical computer science, statistics and signal processing. A few basic applications are covered in this text, particularly for the problem of estimating covariance matrices in statistics and for validating probabilistic constructions of measurement matrices in compressed sensing. These notes are written particularly for graduate students and beginning researchers in different areas, including functional analysts, probabilists, theoretical statisticians, electrical engineers, and theoretical computer scientists.

\section{Introduction} \label{s: introduction} \paragraph{Asymptotic and non-asymptotic regimes} Random matrix theory studies properties of $N \times n$ matrices $A$ chosen from some distribution on the set of all matrices. As dimensions $N$ and $n$ grow to infinity, one observes that the spectrum of $A$ tends to stabilize. This is manifested in several {\em limit laws}, which may be regarded as random matrix versions of the central limit theorem. Among them is Wigner's semicircle law for the eigenvalues of symmetric Gaussian matrices, the circular law for Gaussian matrices, the Marchenko-Pastur law for Wishart matrices $W = A^*A$ where $A$ is a Gaussian matrix, the Bai-Yin and Tracy-Widom laws for the extreme eigenvalues of Wishart matrices $W$. The books \cite{Mehta, AGZ, Deift-Gioev, Bai-Silverstein} offer thorough introduction to the classical problems of random matrix theory and its fascinating connections. The asymptotic regime where the dimensions $N,n \to \infty$ is well suited for the purposes of statistical physics, e.g. when random matrices serve as finite-dimensional models of infinite-dimensional operators. But in some other areas including statistics, geometric functional analysis, and compressed sensing, the limiting regime may not be very useful \cite{RV ICM}. Suppose, for example, that we ask about the largest singular value $s_{\max}(A)$ (i.e. the largest eigenvalue of $(A^*A)^{1/2}$); to be specific assume that $A$ is an $n \times n$ matrix whose entries are independent standard normal random variables. The asymptotic random matrix theory answers this question as follows: the Bai-Yin law (see Theorem~\ref{Bai-Yin}) states that $$ s_{\max}(A) / 2\sqrt{n} \to 1 \quad \text{almost surely} $$ as the dimension $n \to \infty$. Moreover, the limiting distribution of $s_{\max}(A)$ is known to be the Tracy-Widom law (see \cite{Soshnikov, FeSo}). In contrast to this, a non-asymptotic answer to the same question is the following: in {\em every} dimension $n$, one has $$ s_{\max}(A) \le C\sqrt{n} \quad \text{with probability at least } 1 - e^{-n}, $$ here $C$ is an absolute constant (see Theorems~\ref{Gaussian} and \ref{sub-gaussian rows}). The latter answer is less precise (because of an absolute constant $C$) but more quantitative because for fixed dimensions $n$ it gives an exponential probability of success.\footnote{For this specific model (Gaussian matrices),Theorems~\ref{Gaussian} and \ref{Gaussian deviation} even give a sharp absolute constant $C\approx 2$ here. But the result mentioned here is much more general as we will see later; it only requires independence of rows or columns of $A$.} This is the kind of answer we will seek in this text -- guarantees up to absolute constants in all dimensions, and with large probability. \paragraph{Tall matrices are approximate isometries} The following heuristic will be our guideline: {\em tall random matrices should act as approximate isometries}. So, an $N \times n$ random matrix $A$ with $N \gg n$ should act almost like an isometric embedding of $\ell_2^n$ into $\ell_2^N$: $$ (1-\delta) K \|x\|_2 \le \|Ax\|_2 \le (1+\delta) K \|x\|_2 \quad \text{for all } x \in \mathbb{R}^n $$ where $K$ is an appropriate normalization factor and $\delta \ll 1$. Equivalently, this says that all the singular values of $A$ are close to each other: $$ (1-\delta)K \le s_{\min}(A) \le s_{\max}(A) \le (1+\delta)K, $$ where $s_{\min}(A)$ and $s_{\max}(A)$ denote the smallest and the largest singular values of $A$. Yet equivalently, this means that tall matrices are well conditioned: the {\em condition number} \index{Condition number} of $A$ is $\kappa(A) = s_{\max}(A)/s_{\min}(A) \le (1+\delta)/(1-\delta) \approx 1$. In the asymptotic regime and for random matrices with independent entries, our heuristic is justified by Bai-Yin's law, which is Theorem~\ref{Bai-Yin} below. Loosely speaking, it states that as the dimensions $N,n$ increase to infinity while the aspect ratio $N/n$ is fixed, we have \begin{equation} \label{Bai-Yin heuristic} \sqrt{N} - \sqrt{n} \approx s_{\min}(A) \le s_{\max}(A) \approx \sqrt{N} + \sqrt{n}. \end{equation} In these notes, we study $N \times n$ random matrices $A$ with independent rows or independent columns, but not necessarily independent entries. We develop non-asymptotic versions of \eqref{Bai-Yin heuristic} for such matrices, which should hold for all dimensions $N$ and $n$. The desired results should have the form \begin{equation} \label{heuristic} \sqrt{N} - C \sqrt{n} \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{N} + C \sqrt{n} \end{equation} with large probability, e.g. $1-e^{-N}$, where $C$ is an absolute constant.\footnote{More accurately, we should expect $C=O(1)$ to depend on easily computable quantities of the distribution, such as its moments. This will be clear from the context.} For tall matrices, where $N \gg n$, both sides of this inequality would be close to each other, which would guarantee that $A$ is an approximate isometry. \paragraph{Models and methods} We shall study quite general models of random matrices -- those with independent rows or independent columns that are sampled from high-dimensional distributions. We will place either strong moment assumptions on the distribution (sub-gaussian growth of moments), or no moment assumptions at all (except finite variance). This leads us to four types of main results: \begin{enumerate} \setlength{\itemsep}{-3pt} \item Matrices with independent sub-gaussian rows: Theorem~\ref{sub-gaussian rows} \item Matrices with independent heavy-tailed rows: Theorem~\ref{heavy-tailed rows} \item Matrices with independent sub-gaussian columns: Theorem~\ref{sub-gaussian columns} \item Matrices with independent heavy-tailed columns: Theorem~\ref{heavy-tailed columns} \end{enumerate} These four models cover many natural classes of random matrices that occur in applications, including random matrices with independent entries (Gaussian and Bernoulli in particular) and random sub-matrices of orthogonal matrices (random Fourier matrices in particular). The analysis of these four models is based on a variety of tools of probability theory and geometric functional analysis, most of which have not been covered in the texts on the ``classical'' random matrix theory. The reader will learn basics on sub-gaussian and sub-exponential random variables, isotropic random vectors, large deviation inequalities for sums of independent random variables, extensions of these inequalities to random matrices, and several basic methods of high dimensional probability such as symmetrization, decoupling, and covering ($\varepsilon$-net) arguments. \paragraph{Applications} In these notes we shall emphasize two applications, one in statistics and one in compressed sensing. Our analysis of random matrices with independent rows immediately applies to a basic problem in statistics -- {\em estimating covariance matrices} of high-dimensional distributions. If a random matrix $A$ has i.i.d. rows $A_i$, then $A^*A = \sum_i A_i \otimes A_i$ is the {\em sample covariance matrix}. If $A$ has independent columns $A_j$, then $A^*A = (\< A_j, A_k\> )_{j,k}$ is the {\em Gram matrix}. Thus our analysis of the row-independent and column-independent models can be interpreted as a study of sample covariance matrices and Gram matrices of high dimensional distributions. We will see in Section~\ref{s: covariance} that for a general distribution in $\mathbb{R}^n$, its covariance matrix can be estimated from a sample of size $N = O(n \log n)$ drawn from the distribution. Moreover, for sub-gaussian distributions we have an even better bound $N = O(n)$. For low-dimensional distributions, much fewer samples are needed -- if a distribution lies close to a subspace of dimension $r$ in $\mathbb{R}^n$, then a sample of size $N = O(r \log n)$ is sufficient for covariance estimation. In compressed sensing, the best known measurement matrices are random. A sufficient condition for a matrix to succeed for the purposes of compressed sensing is given by the {\em restricted isometry property}. Loosely speaking, this property demands that all sub-matrices of given size be well-conditioned. This fits well in the circle of problems of the non-asymptotic random matrix theory. Indeed, we will see in Section~\ref{s: restricted isometries} that all basic models of random matrices are nice restricted isometries. These include Gaussian and Bernoulli matrices, more generally all matrices with sub-gaussian independent entries, and even more generally all matrices with sub-gaussian independent rows or columns. Also, the class of restricted isometries includes random Fourier matrices, more generally random sub-matrices of bounded orthogonal matrices, and even more generally matrices whose rows are independent samples from an isotropic distribution with uniformly bounded coordinates. \paragraph{Related sources} This text is a tutorial rather than a survey, so we focus on explaining methods rather than results. This forces us to make some concessions in our choice of the subjects. {\em Concentration of measure} and its applications to random matrix theory are only briefly mentioned. For an introduction into concentration of measure suitable for a beginner, see \cite{Ball} and \cite[Chapter~14]{Matousek}; for a thorough exposition see \cite{MS, Ledoux}; for connections with random matrices see \cite{DS, Ledoux extremal}. The monograph \cite{Ledoux-Talagrand} also offers an introduction into concentration of measure and related probabilistic methods in analysis and geometry, some of which we shall use in these notes. We completely avoid the important (but more difficult) model of {\em symmetric random matrices} with independent entries on and above the diagonal. Starting from the work of F\"uredi and Komlos \cite{FuKo}, the largest singular value (the spectral norm) of symmetric random matrices has been a subject of study in many works; see e.g. \cite{Meckes, Vu, PeSo} and the references therein. We also did not even attempt to discuss sharp small {\em deviation inequalities} (of Tracy-Widom type) for the extreme eigenvalues. Both these topics and much more are discussed in the surveys \cite{DS, Ledoux extremal, RV ICM}, which serve as bridges between asymptotic and non-asymptotic problems in random matrix theory. Because of the absolute constant $C$ in \eqref{heuristic}, our analysis of the smallest singular value (the {\em ``hard edge''}) will only be useful for sufficiently tall matrices, where $N \ge C^2 n$. For square and almost square matrices, the hard edge problem will be only briefly mentioned in Section~\ref{s: entries}. The surveys \cite{Tao-Vu survey, RV ICM} discuss this problem at length, and they offer a glimpse of connections to other problems of random matrix theory and additive combinatorics. Many of the results and methods presented in these notes are known in one form or another. Some of them are published while some others belong to the folklore of probability in Banach spaces, geometric functional analysis, and related areas. When available, historic references are given in Section~\ref{s: notes}. \paragraph{Acknowledgements} The author is grateful to the colleagues who made a number of improving suggestions for the earlier versions of the manuscript, in particular to Richard Chen, Subhroshekhar Ghosh, Alexander Litvak, Deanna Needell, Holger Rauhut, S V N Vishwanathan and the anonymous referees. Special thanks are due to Ulas Ayaz and Felix Krahmer who thoroughly read the entire text, and whose numerous comments led to significant improvements of this tutorial. \section{Preliminaries} \label{s: preliminaries} \subsection{Matrices and their singular values} The main object of our study will be an $N \times n$ matrix $A$ with real or complex entries. We shall state all results in the real case; the reader will be able to adjust them to the complex case as well. Usually but not always one should think of tall matrices $A$, those for which $N \ge n > 1$. By passing to the adjoint matrix $A^*$, many results can be carried over to ``flat'' matrices, those for which $N \le n$. It is often convenient to study $A$ through the $n \times n$ symmetric positive-semidefinite matrix the matrix $A^*A$. The eigenvalues of $|A| := \sqrt{A^*A}$ are therefore non-negative real numbers. Arranged in a non-decreasing order, they are called the {\em singular values}\footnote{In the literature, singular values are also called {\em s-numbers}.} \index{Singular values} of $A$ and denoted $s_1(A) \ge \cdots \ge s_n(A) \ge 0$. Many applications require estimates on the extreme singular values $$ s_{\max}(A) := s_1(A), \quad s_{\min}(A) := s_n(A). $$ The smallest singular value is only of interest for tall matrices, since for $N < n$ one automatically has $s_{\min}(A) = 0$. Equivalently, $s_{\max}(A)$ and $s_{\min}(A)$ are respectively the smallest number $M$ and the largest number $m$ such that \begin{equation} \label{mM} m \|x\|_2 \le \|Ax\|_2 \le M \|x\|_2 \quad \text{for all } x \in \mathbb{R}^n. \end{equation} In order to interpret this definition geometrically, we look at $A$ as a linear operator from $\mathbb{R}^n$ into $\mathbb{R}^N$. The Euclidean distance between any two points in $\mathbb{R}^n$ can increase by at most the factor $s_{\max}(A)$ and decrease by at most the factor $s_{\max}(A)$ under the action of $A$. Therefore, the extreme singular values control the distortion of the Euclidean geometry under the action of $A$. If $s_{\max}(A) \approx s_{\min}(A) \approx 1$ then $A$ acts as an {\em approximate isometry},\index{Approximate isometries} or more accurately an approximate isometric embedding of $\ell_2^n$ into $\ell_2^N$. The extreme singular values can also be described in terms of the {\em spectral norm of $A$},\index{Spectral norm} which is by definition \begin{equation} \label{spectral norm} \|A\| = \|A\|_{\ell_2^n \to \ell_2^N} = \sup_{x \in \mathbb{R}^n \setminus \{0\}} \frac{\|Ax\|_2}{\|x\|_2} = \sup_{x \in S^{n-1}} \|Ax\|_2. \end{equation} \eqref{mM} gives a link between the extreme singular values and the spectral norm: $$ s_{\max}(A) = \|A\|, \quad s_{\min}(A) = 1/\|A^\dagger\| $$ where $A^\dagger$ denotes the pseudoinverse of $A$; if $A$ is invertible then $A^\dagger = A^{-1}$. \subsection{Nets} Nets are convenient means to discretize compact sets. In our study we will mostly need to discretize the unit Euclidean sphere $S^{n-1}$ in the definition of the spectral norm \eqref{spectral norm}. Let us first recall a general definition of an $\varepsilon$-net. \begin{definition}[Nets, covering numbers] \index{Net} \index{Covering numbers} Let $(X,d)$ be a metric space and let $\varepsilon>0$. A subset $\mathcal{N}_\varepsilon$ of $X$ is called an {\em $\varepsilon$-net} of $X$ if every point $x \in X$ can be approximated to within $\varepsilon$ by some point $y \in \mathcal{N}_\varepsilon$, i.e. so that $d(x,y) \le \varepsilon$. The minimal cardinality of an $\varepsilon$-net of $X$, if finite, is denoted $\mathcal{N}(X,\varepsilon)$ and is called the {\em covering number}\footnote{Equivalently, $\mathcal{N}(X,\varepsilon)$ is the minimal number of balls with radii $\varepsilon$ and with centers in $X$ needed to cover $X$.} of $X$ (at scale $\varepsilon$). \end{definition} From a characterization of compactness we remember that $X$ is compact if and only if $\mathcal{N}(X,\varepsilon) < \infty$ for each $\varepsilon > 0$. A quantitative estimate on $\mathcal{N}(X, \varepsilon)$ would give us a {\em quantitative version of compactness} of $X$.\footnote{In statistical learning theory and geometric functional analysis, $\log \mathcal{N}(X,\varepsilon)$ is called {\em the metric entropy of $X$}. In some sense it measures the ``complexity'' of metric space $X$.} Let us therefore take a simple example of a metric space, the unit Euclidean sphere $S^{n-1}$ equipped with the Euclidean metric\footnote{A similar result holds for the geodesic metric on the sphere, since for small $\varepsilon$ these two distances are equivalent.} $d(x,y) = \|x-y\|_2$, and estimate its covering numbers. \begin{lemma}[Covering numbers of the sphere] \label{net cardinality} The unit Euclidean sphere $S^{n-1}$ equipped with the Euclidean metric satisfies for every $\varepsilon>0$ that $$ \mathcal{N}(S^{n-1},\varepsilon) \le \Big( 1 + \frac{2}{\varepsilon} \Big)^n. $$ \end{lemma} \begin{proof} This is a simple {\em volume argument}. Let us fix $\varepsilon>0$ and choose $\mathcal{N}_\varepsilon$ to be a maximal $\varepsilon$-separated subset of $S^{n-1}$. In other words, $\mathcal{N}_\varepsilon$ is such that $d(x,y) \ge \varepsilon$ for all $x, y \in \mathcal{N}_\varepsilon$, $x \ne y$, and no subset of $S^{n-1}$ containing $\mathcal{N}_\varepsilon$ has this property.\footnote{One can in fact construct $\mathcal{N}_\varepsilon$ inductively by first selecting an arbitrary point on the sphere, and at each next step selecting a point that is at distance at least $\varepsilon$ from those already selected. By compactness, this algorithm will terminate after finitely many steps and it will yield a set $\mathcal{N}_\varepsilon$ as we required.} The maximality property implies that $\mathcal{N}_\varepsilon$ is an $\varepsilon$-net of $S^{n-1}$. Indeed, otherwise there would exist $x \in S^{n-1}$ that is at least $\varepsilon$-far from all points in $\mathcal{N}_\varepsilon$. So $\mathcal{N}_\varepsilon \cup \{x\}$ would still be an $\varepsilon$-separated set, contradicting the minimality property. Moreover, the separation property implies via the triangle inequality that the balls of radii $\varepsilon/2$ centered at the points in $\mathcal{N}_\varepsilon$ are disjoint. On the other hand, all such balls lie in $(1+\varepsilon/2) B_2^n$ where $B_2^n$ denotes the unit Euclidean ball centered at the origin. Comparing the volume gives $\vol \big( \frac{\varepsilon}{2} B_2^n \big) \cdot |\mathcal{N}_\varepsilon| \le \vol \big( (1 + \frac{\varepsilon}{2}) B_2^n \big)$. Since $\vol \big( r B_2^n \big) = r^n \vol(B_2^n)$ for all $r \ge 0$, we conclude that $|\mathcal{N}_\varepsilon| \le (1+\frac{\varepsilon}{2})^n / (\frac{\varepsilon}{2})^n = (1+\frac{2}{\varepsilon})^n$ as required. \end{proof} Nets allow us to reduce the complexity of computations with linear operators. One such example is the computation of the spectral norm. To evaluate the spectral norm by definition \eqref{spectral norm} one needs to take the supremum over the whole sphere $S^{n-1}$. However, one can essentially replace the sphere by its $\varepsilon$-net: \begin{lemma}[Computing the spectral norm on a net] \index{Spectral norm!computing on a net} \label{norm on net general} Let $A$ be an $N \times n$ matrix, and let $\mathcal{N}_\varepsilon$ be an $\varepsilon$-net of $S^{n-1}$ for some $\varepsilon \in [0,1)$. Then $$ \max_{x \in \mathcal{N}_\varepsilon} \|Ax\|_2 \le \|A\| \le (1 - \varepsilon)^{-1} \max_{x \in \mathcal{N}_\varepsilon} \|Ax\|_2 $$ \end{lemma} \begin{proof} The lower bound in the conclusion follows from the definition. To prove the upper bound let us fix $x \in S^{n-1}$ for which $\|A\| = \|Ax\|_2$, and choose $y \in \mathcal{N}_\varepsilon$ which approximates $x$ as $\|x-y\|_2 \le \varepsilon$. By the triangle inequality we have $\|Ax-Ay\|_2 \le \|A\| \|x-y\|_2 \le \varepsilon \|A\|$. It follows that $$ \|Ay\|_2 \ge \|Ax\|_2 - \|Ax-Ay\|_2 \ge \|A\| - \varepsilon\|A\| = (1-\varepsilon)\|A\|. $$ Taking maximum over all $y \in \mathcal{N}_\varepsilon$ in this inequality, we complete the proof. \end{proof} A similar result holds for symmetric $n \times n$ matrices $A$, whose spectral norm can be computed via the associated quadratic form: $\|A\| = \sup_{x \in S^{n-1}} |\< Ax,x\> |$. Again, one can essentially replace the sphere by its $\varepsilon$-net: \begin{lemma}[Computing the spectral norm on a net] \index{Spectral norm!computing on a net} \label{norm on net} Let $A$ be a symmetric $n \times n$ matrix, and let $\mathcal{N}_\varepsilon$ be an $\varepsilon$-net of $S^{n-1}$ for some $\varepsilon \in [0,1)$. Then $$ \|A\| = \sup_{x \in S^{n-1}} |\< Ax, x\> | \le (1 - 2\varepsilon)^{-1} \sup_{x \in \mathcal{N}_\varepsilon} |\< Ax, x\> |. $$ \end{lemma} \begin{proof} Let us choose $x \in S^{n-1}$ for which $\|A\| = |\< Ax, x\> |$, and choose $y \in \mathcal{N}_\varepsilon$ which approximates $x$ as $\|x - y\|_2 \le \varepsilon$. By the triangle inequality we have \begin{align*} |\< Ax, x\> - \< Ay, y\> | &= |\< Ax, x-y\> + \< A(x-y), y\> |\\ &\le \|A\| \|x\|_2 \|x-y\|_2 + \|A\| \|x-y\|_2 \|y\|_2 \le 2 \varepsilon \|A\|. \end{align*} It follows that $|\< Ay, y \> | \ge |\< Ax, x\> | - 2 \varepsilon \|A\| = (1-2\varepsilon) \|A\|$. Taking the maximum over all $y \in \mathcal{N}_\varepsilon$ in this inequality completes the proof. \end{proof} \subsection{Sub-gaussian random variables} \index{Sub-gaussian!random variables} \label{s: sub-gaussian} In this section we introduce the class of sub-gaussian random variables,\footnote{It would be more rigorous to say that we study {\em sub-gaussian probability distributions}. The same concerns some other properties of random variables and random vectors we study later in this text. However, it is convenient for us to focus on random variables and vectors because we will form random matrices out of them.} those whose distributions are dominated by the distribution of a centered gaussian random variable. This is a convenient and quite wide class, which contains in particular the standard normal and all bounded random variables. Let us briefly recall some of the well known properties of the {\em standard normal random variable} $X$. The distribution of $X$ has density $\frac{1}{\sqrt{2 \pi}} e^{-x^2/2}$ and is denoted $N(0,1)$. Estimating the integral of this density between $t$ and $\infty$ one checks that the tail of a standard normal random variable $X$ decays super-exponentially: \begin{equation} \label{normal tail} \mathbb{P} \{ |X| > t \} = \frac{2}{\sqrt{2 \pi}} \int_t^\infty e^{-x^2/2} \, dx \le 2 e^{-t^2/2}, \quad t \ge 1, \end{equation} see e.g. \cite[Theorem 1.4]{Durrett} for a more precise two-sided inequality. The absolute moments of $X$ can be computed as \begin{equation} \label{normal moments} (\mathbb{E} |X|^p)^{1/p} = \sqrt{2} \Big[ \frac{\Gamma((1+p)/2)}{\Gamma(1/2)} \Big]^{1/p} = O(\sqrt{p}), \quad p \ge 1. \end{equation} The moment generating function of $X$ equals \begin{equation} \label{normal mgf} \mathbb{E} \exp(tX) = e^{t^2/2}, \quad t \in \mathbb{R}. \end{equation} Now let $X$ be a general random variable. We observe that these three properties are equivalent -- a super-exponential tail decay like in \eqref{normal tail}, the moment growth \eqref{normal moments}, and the growth of the moment generating function like in \eqref{normal mgf}. We will then focus on the class of random variables that satisfy these properties, which we shall call sub-gaussian random variables. \begin{lemma}[Equivalence of sub-gaussian properties] \label{sub-gaussian properties} Let $X$ be a random variable. Then the following properties are equivalent with parameters $K_i > 0$ differing from each other by at most an absolute constant factor.\footnote{The precise meaning of this equivalence is the following. There exists an absolute constant $C$ such that property $i$ implies property $j$ with parameter $K_j \le C K_i$ for any two properties $i,j=1,2,3$.} \begin{enumerate} \item Tails: $\mathbb{P} \{ |X| > t \} \le \exp(1-t^2/K_1^2)$ for all $t \ge 0$; \item Moments: $(\mathbb{E} |X|^p)^{1/p} \le K_2 \sqrt{p}$ for all $p \ge 1$; \item Super-exponential moment: $\mathbb{E} \exp(X^2/K_3^2) \le e$. \end{enumerate} Moreover, if $\mathbb{E} X = 0$ then properties 1--3 are also equivalent to the following one: \begin{enumerate} \setcounter{enumi}{3} \item Moment generating function: $\mathbb{E} \exp(tX) \le \exp(t^2 K_4^2)$ for all $t \in \mathbb{R}$. \end{enumerate} \end{lemma} \begin{proof} {\bf 1. $\Rightarrow$ 2.} Assume property 1 holds. By homogeneity, rescaling $X$ to $X/K_1$ we can assume that $K_1=1$. Recall that for every non-negative random variable $Z$, integration by parts yields the identity $\mathbb{E} Z = \int_0^\infty \mathbb{P} \{ Z \ge u \} \, du$. We apply this for $Z = |X|^p$. After change of variables $u = t^p$, we obtain using property 1 that $$ \mathbb{E} |X|^p = \int_0^\infty \mathbb{P} \{ |X| \ge t \} \, p t^{p-1} \, dt \le \int_0^\infty e^{1-t^2} p t^{p-1} \, dt = \big( \frac{ep}{2} \big) \Gamma \big( \frac{p}{2} \big) \le \big( \frac{ep}{2} \big) \big( \frac{p}{2} \big)^{p/2}. $$ Taking the $p$-th root yields property 2 with a suitable absolute constant $K_2$. {\bf 2. $\Rightarrow$ 3.} Assume property 2 holds. As before, by homogeneity we may assume that $K_2 = 1$. Let $c>0$ be a sufficiently small absolute constant. Writing the Taylor series of the exponential function, we obtain $$ \mathbb{E} \exp(c X^2) = 1 + \sum_{p=1}^\infty \frac{c^p \mathbb{E}(X^{2p})}{p!} \le 1 + \sum_{p=1}^\infty \frac{c^p (2p)^p}{p!} \le 1 + \sum_{p=1}^\infty (2c/e)^p. $$ The first inequality follows from property 2; in the second one we use $p! \ge (p/e)^p$. For small $c$ this gives $\mathbb{E} \exp(c X^2) \le e$, which is property 3 with $K_3 = c^{-1/2}$. {\bf 3. $\Rightarrow$ 1.} Assume property 3 holds. As before we may assume that $K_3=1$. Exponentiating and using Markov's inequality\footnote{This simple argument is sometimes called exponential Markov's inequality.} and then property 3, we have $$ \mathbb{P} \{ |X| > t \} = \mathbb{P} \{ e^{X^2} \ge e^{t^2} \} \le e^{-t^2} \mathbb{E} e^{X^2} \le e^{1-t^2}. $$ This proves property 1 with $K_1=1$. {\bf 2. $\Rightarrow$ 4.} Let us now assume that $\mathbb{E} X = 0$ and property 2 holds; as usual we can assume that $K_2 = 1$. We will prove that property 4 holds with an appropriately large absolute constant $C = K_4$. This will follow by estimating Taylor series for the exponential function \begin{equation} \label{mgf above} \mathbb{E} \exp(tX) = 1 + t \mathbb{E} X + \sum_{p=2}^\infty \frac{t^p \mathbb{E} X^p}{p!} \le 1 + \sum_{p=2}^\infty \frac{t^p p^{p/2}}{p!} \le 1 + \sum_{p=2}^\infty \Big( \frac{e|t|}{\sqrt{p}} \Big)^p. \end{equation} The first inequality here follows from $\mathbb{E} X = 0$ and property 2; the second one holds since $p! \ge (p/e)^p$. We compare this with Taylor's series for \begin{equation} \label{exp t squared} \exp(C^2 t^2) = 1 + \sum_{k=1}^\infty \frac{(C|t|)^{2k}}{k!} \ge 1 + \sum_{k=1}^\infty \Big( \frac{C|t|}{\sqrt{k}} \Big)^{2k} = 1 + \sum_{p \in 2\mathbb{N}} \Big( \frac{C|t|}{\sqrt{p/2}} \Big)^p. \end{equation} The first inequality here holds because $p! \le p^p$; the second one is obtained by substitution $p=2k$. One can show that the series in \eqref{mgf above} is bounded by the series in \eqref{exp t squared} with large absolute constant $C$. We conclude that $\mathbb{E} \exp(tX) \le \exp(C^2 t^2)$, which proves property~4. {\bf 4. $\Rightarrow$ 1.} Assume property 4 holds; we can also assume that $K_4=1$. Let $\lambda > 0$ be a parameter to be chosen later. By exponential Markov inequality, and using the bound on the moment generating function given in property 4, we obtain $$ \mathbb{P} \{ X \ge t \} = \mathbb{P} \{ e^{\lambda X} \ge e^{\lambda t} \} \le e^{-\lambda t} \mathbb{E} e^{\lambda X} \le e^{-\lambda t + \lambda^2}. $$ Optimizing in $\lambda$ and thus choosing $\lambda = t/2$ we conclude that $\mathbb{P} \{ X \ge t \} \le e^{-t^2/4}$. Repeating this argument for $-X$, we also obtain $\mathbb{P} \{ X \le -t \} \le e^{-t^2/4}$. Combining these two bounds we conclude that $\mathbb{P} \{ |X| \ge t \} \le 2 e^{-t^2/4} \le e^{1-t^2/4}$. Thus property 1 holds with $K_1=2$. The lemma is proved. \end{proof} \begin{remark} \begin{enumerate} \item The constants $1$ and $e$ in properties 1 and 3 respectively are chosen for convenience. Thus the value $1$ can be replaced by any positive number and the value $e$ can be replaced by any number greater than $1$. \item The assumption $\mathbb{E} X = 0$ is only needed to prove the necessity of property 4; the sufficiency holds without this assumption. \end{enumerate} \end{remark} \begin{definition}[Sub-gaussian random variables] \index{Sub-gaussian!random variables} A random variable $X$ that satisfies one of the equivalent properties 1 -- 3 in Lemma~\ref{sub-gaussian properties} is called a {\em sub-gaussian random variable}. The {\em sub-gaussian norm} \index{Sub-gaussian!norm} of $X$, denoted $\|X\|_{\psi_2}$, is defined to be the smallest $K_2$ in property 2. In other words,\footnote{The sub-gaussian norm is also called ${\psi_2}$ norm in the literature.} $$ \|X\|_{\psi_2} = \sup_{p \ge 1} p^{-1/2} (\mathbb{E} |X|^p)^{1/p}. $$ \end{definition} The class of sub-gaussian random variables on a given probability space is thus a normed space. By Lemma~\ref{sub-gaussian properties}, every sub-gaussian random variable $X$ satisfies: \begin{gather} \mathbb{P} \{ |X| > t \} \le \exp(1-ct^2/\|X\|_{\psi_2}^2) \quad \text{for all } t \ge 0; \label{sub-gaussian tail}\\ (\mathbb{E} |X|^p)^{1/p} \le \|X\|_{\psi_2} \sqrt{p} \quad \text{for all } p \ge 1; \label{sub-gaussian moments}\\ \mathbb{E} \exp(cX^2/\|X\|_{\psi_2}^2) \le e; \nonumber\\ \text{if $\mathbb{E} X =0$ then } \mathbb{E} \exp(tX) \le \exp(C t^2 \|X\|_{\psi_2}^2) \quad \text{for all } t \in \mathbb{R}, \label{sub-gaussian mgf} \end{gather} where $C, c > 0$ are absolute constants. Moreover, up to absolute constant factors, $\|X\|_{\psi_2}$ is the smallest possible number in each of these inequalities. \begin{example} Classical examples of sub-gaussian random variables are Gaussian, Bernoulli and all bounded random variables. \begin{enumerate} \item {\bf (Gaussian):} A standard normal random variable $X$ is sub-gaussian with $\|X\|_{\psi_2} \le C$ where $C$ is an absolute constant. This follows from \eqref{normal moments}. More generally, if $X$ is a centered normal random variable with variance $\sigma^2$, then $X$ is sub-gaussian with $\|X\|_{\psi_2} \le C \sigma$. \item {\bf (Bernoulli):} \index{Bernoulli!random variables} Consider a random variable $X$ with distribution $\mathbb{P}\{X=-1\} = \mathbb{P}\{X=1\} = 1/2$. We call $X$ a {\em symmetric Bernoulli random variable}. Since $|X|=1$, it follows that $X$ is a sub-gaussian random variable with $\|X\|_{\psi_2} = 1$. \item {\bf (Bounded):} More generally, consider any bounded random variable $X$, thus $|X| \le M$ almost surely for some $M$. Then $X$ is a sub-gaussian random variable with $\|X\|_{\psi_2} \le M$. We can write this more compactly as $\|X\|_{\psi_2} \le \|X\|_\infty$. \end{enumerate} \end{example} A remarkable property of the normal distribution is {\em rotation invariance}. Given a finite number of independent centered normal random variables $X_i$, their sum $\sum_i X_i$ is also a centered normal random variable, obviously with $\Var(\sum_i X_i) = \sum_i \Var(X_i)$. Rotation invariance passes onto sub-gaussian random variables, although approximately: \begin{lemma}[Rotation invariance] \index{Rotation invariance} \label{rotation invariance} Consider a finite number of independent centered sub-gaussian random variables $X_i$. Then $\sum_i X_i$ is also a centered sub-gaussian random variable. Moreover, $$ \big\| \sum_i X_i \big\|_{\psi_2}^2 \le C \sum_i \|X_i\|_{\psi_2}^2 $$ where $C$ is an absolute constant. \end{lemma} \begin{proof} The argument is based on estimating the moment generating function. Using independence and \eqref{sub-gaussian mgf} we have for every $t \in \mathbb{R}$: \begin{align*} \mathbb{E} \exp \big( t \sum_i X_i \big) &= \mathbb{E} \prod_i \exp(t X_i) = \prod_i \mathbb{E} \exp(t X_i) \le \prod_i \exp(C t^2 \|X_i\|_{\psi_2}^2) \\ &= \exp(t^2 K^2) \quad \text{where } K^2 = C \sum_i \|X_i\|_{\psi_2}^2. \end{align*} Using the equivalence of properties 2 and 4 in Lemma~\ref{sub-gaussian properties} we conclude that $\|\sum_i X_i\|_{\psi_2} \le C_1 K$ where $C_1$ is an absolute constant. The proof is complete. \end{proof} The rotation invariance immediately yields a {\em large deviation inequality} for sums of independent sub-gaussian random variables: \begin{proposition}[Hoeffding-type inequality] \index{Hoeffding-type inequality} \label{sub-gaussian large deviations} Let $X_1,\ldots,X_N$ be independent centered sub-gaussian random variables, and let $K = \max_i \|X_i\|_{\psi_2}$. Then for every $a = (a_1,\ldots,a_N) \in \mathbb{R}^N$ and every $t \ge 0$, we have $$ \mathbb{P} \Big\{ \Big| \sum_{i=1}^N a_i X_i \Big| \ge t \Big\} \le e \cdot \exp \Big( -\frac{ct^2}{K^2\|a\|_2^2} \Big) $$ where $c>0$ is an absolute constant. \end{proposition} \begin{proof} The rotation invariance (Lemma~\ref{rotation invariance}) implies the bound $\|\sum_i a_i X_i\|_{\psi_2}^2 \le C \sum_i a_i^2 \|X_i\|_{\psi_2}^2 \le C K^2 \|a\|_2^2$. Property \eqref{sub-gaussian tail} yields the required tail decay. \end{proof} \begin{remark} One can interpret these results (Lemma~\ref{rotation invariance} and Proposition~\ref{sub-gaussian large deviations}) as one-sided {\em non-asymptotic manifestations of the central limit theorem}. For example, consider the normalized sum of independent symmetric Bernoulli random variables $S_N = \frac{1}{\sqrt{N}} \sum_{i=1}^N \varepsilon_i$. Proposition~\ref{sub-gaussian large deviations} yields the tail bounds $\mathbb{P} \{ |S_N| > t \} \le e \cdot e^{-ct^2}$ for any number of terms $N$. Up to the absolute constants $e$ and $c$, these tails coincide with those of the standard normal random variable \eqref{normal tail}. \end{remark} Using moment growth \eqref{sub-gaussian moments} instead of the tail decay \eqref{sub-gaussian tail}, we immediately obtain from Lemma~\ref{rotation invariance} a general form of the well known Khintchine inequality: \begin{corollary}[Khintchine inequality] \index{Khinchine inequality} \label{Khintchine} Let $X_i$ be a finite number of independent sub-gaussian random variables with zero mean, unit variance, and $\|X_i\|_{{\psi_2}} \le K$. Then, for every sequence of coefficients $a_i$ and every exponent $p \ge 2$ we have $$ \big( \sum_i a_i^2 \big)^{1/2} \le \big( \mathbb{E} \big| \sum_i a_i X_i \big|^p \big)^{1/p} \le C K \sqrt{p} \, \big( \sum_i a_i^2 \big)^{1/2} $$ where $C$ is an absolute constant. \end{corollary} \begin{proof} The lower bound follows by independence and H\"older's inequality: indeed, $\big( \mathbb{E} \big| \sum_i a_i X_i \big|^p \big)^{1/p} \ge \big( \mathbb{E} \big| \sum_i a_i X_i \big|^2 \big)^{1/2} = \big( \sum_i a_i^2 \big)^{1/2}$. For the upper bound, we argue as in Proposition~\ref{sub-gaussian large deviations}, but use property \eqref{sub-gaussian moments}. \end{proof} \subsection{Sub-exponential random variables} \index{Sub-exponential!random variables} \label{s: sub-exponential} Although the class of sub-gaussian random variables is natural and quite wide, it leaves out some useful random variables which have tails heavier than gaussian. One such example is a standard exponential random variable -- a non-negative random variable with exponential tail decay \begin{equation} \label{exponential} \mathbb{P}\{X \ge t\} = e^{-t}, \quad t \ge 0. \end{equation} To cover such examples, we consider a class of {\em sub-exponential random variables}, those with at least an exponential tail decay. With appropriate modifications, the basic properties of sub-gaussian random variables hold for sub-exponentials. In particular, a version of Lemma~\ref{sub-gaussian properties} holds with a similar proof for sub-exponential properties, except for property 4 of the moment generating function. Thus for a random variable $X$ the following properties are equivalent with parameters $K_i > 0$ differing from each other by at most an absolute constant factor: \begin{gather} \mathbb{P} \{ |X| > t \} \le \exp(1-t/K_1) \quad \text{for all } t \ge 0; \label{sub-exponential tail} \\ (\mathbb{E} |X|^p)^{1/p} \le K_2 p \quad \text{for all } p \ge 1; \label{sub-exponential moments} \\ \mathbb{E} \exp(X/K_3) \le e. \label{sub-exponential integrability} \end{gather} \begin{definition}[Sub-exponential random variables] \index{Sub-exponential!random variables} A random variable $X$ that satisfies one of the equivalent properties \eqref{sub-exponential tail} -- \eqref{sub-exponential integrability} is called a {\em sub-exponential random variable}. The {\em sub-exponential norm} \index{Sub-exponential!norm} of $X$, denoted $\|X\|_{\psi_1}$, is defined to be the smallest parameter $K_2$. In other words, $$ \|X\|_{\psi_1} = \sup_{p \ge 1} p^{-1} (\mathbb{E} |X|^p)^{1/p}. $$ \end{definition} \begin{lemma}[Sub-exponential is sub-gaussian squared] \label{sub-exponential squared} A random variable $X$ is sub-gaussian if and only if $X^2$ is sub-exponential. Moreover, $$ \|X\|_{\psi_2}^2 \le \|X^2\|_{\psi_1} \le 2 \|X\|_{\psi_2}^2. $$ \end{lemma} \begin{proof} This follows easily from the definition. \end{proof} The moment generating function of a sub-exponential random variable has a similar upper bound as in the sub-gaussian case (property 4 in Lemma~\ref{sub-gaussian properties}). The only real difference is that the bound only holds in a neighborhood of zero rather than on the whole real line. This is inevitable, as the moment generating function of an exponential random variable \eqref{exponential} does not exist for $t \ge 1$. \begin{lemma}[Mgf of sub-exponential random variables] \label{sub-exponential mgf} Let $X$ be a centered sub-exponential random variable. Then, for $t$ such that $|t| \le c/\|X\|_{\psi_1}$, one has $$ \mathbb{E} \exp(t X) \le \exp(C t^2 \|X\|_{\psi_1}^2) $$ where $C, c > 0$ are absolute constants. \end{lemma} \begin{proof} The argument is similar to the sub-gaussian case. We can assume that $\|X\|_{\psi_1}=1$ by replacing $X$ with $X/\|X\|_{\psi_1}$ and $t$ with $t \|X\|_{\psi_1}$. Repeating the proof of the implication 2 $\Rightarrow$ 4 of Lemma~\ref{sub-gaussian properties} and using $\mathbb{E}|X|^p \le p^p$ this time, we obtain that $\mathbb{E} \exp(tX) \le 1 + \sum_{p=2}^\infty (e|t|)^p$. If $|t| \le 1/2e$ then the right hand side is bounded by $1 + 2e^2 t^2 \le \exp(2e^2 t^2)$. This completes the proof. \end{proof} Sub-exponential random variables satisfy a {\em large deviation inequality} similar to the one for sub-gaussians (Proposition~\ref{sub-gaussian large deviations}). The only significant difference is that {\em two tails} have to appear here -- a gaussian tail responsible for the central limit theorem, and an exponential tail coming from the tails of each term. \begin{proposition}[Bernstein-type inequality] \index{Bernstein-type inequality} \label{sub-exponential large deviations} Let~$X_1,\ldots,X_N$ be independent centered sub-exponential random variables, and $K = \max_i \|X_i\|_{\psi_1}$. Then for every $a = (a_1,\ldots,a_N) \in \mathbb{R}^N$ and every $t \ge 0$, we have $$ \mathbb{P} \Big\{ \Big| \sum_{i=1}^N a_i X_i \Big| \ge t \Big\} \le 2 \exp \Big[ -c \min \Big( \frac{t^2}{K^2\|a\|_2^2}, \; \frac{t}{K\|a\|_\infty} \Big) \Big] $$ where $c>0$ is an absolute constant. \end{proposition} \begin{proof} Without loss of generality, we assume that $K=1$ by replacing $X_i$ with $X_i/K$ and $t$ with $t/K$. We use the exponential Markov inequality for the sum $S = \sum_i a_i X_i$ and with a parameter $\lambda>0$: $$ \mathbb{P} \{ S \ge t \} = \mathbb{P} \{ e^{\lambda S} \ge e^{\lambda t} \} \le e^{-\lambda t} \mathbb{E} e^{\lambda S} = e^{-\lambda t} \prod_i \mathbb{E} \exp (\lambda a_i X_i). $$ If $|\lambda| \le c/\|a\|_\infty$ then $|\lambda a_i| \le c$ for all $i$, so Lemma~\ref{sub-exponential mgf} yields $$ \mathbb{P} \{ S \ge t \} \le e^{-\lambda t} \; \prod_i \exp (C \lambda^2 a_i^2) = \exp(-\lambda t + C \lambda^2 \|a\|_2^2). $$ Choosing $\lambda = \min( t/2C\|a\|_2^2, \, c/\|a\|_\infty )$, we obtain that $$ \mathbb{P} \{ S \ge t \} \le \exp \Big[ - \min \Big( \frac{t^2}{4C\|a\|_2^2}, \; \frac{ct}{2\|a\|_\infty} \Big) \Big]. $$ Repeating this argument for $-X_i$ instead of $X_i$, we obtain the same bound for $\mathbb{P} \{ -S \ge t \}$. A combination of these two bounds completes the proof. \end{proof} \begin{corollary} \label{average sub-exponentials} Let $X_1,\ldots,X_N$ be independent centered sub-exponential random variables, and let $K = \max_i \|X_i\|_{\psi_1}$. Then, for every $\varepsilon \ge 0$, we have $$ \mathbb{P} \Big\{ \Big| \sum_{i=1}^N X_i \Big| \ge \varepsilon N \Big\} \le 2 \exp \Big[ -c \min \Big( \frac{\varepsilon^2}{K^2}, \; \frac{\varepsilon}{K} \Big) N \Big] $$ where $c>0$ is an absolute constant. \end{corollary} \begin{proof} This follows from Proposition~\ref{sub-exponential large deviations} for $a_i=1$ and $t=\varepsilon N$. \end{proof} \begin{remark}[Centering] \label{centering} The definitions of sub-gaussian and sub-exponential random variables $X$ do not require them to be centered. In any case, one can always center $X$ using the simple fact that if $X$ is sub-gaussian (or sub-exponential), then so is $X - \mathbb{E} X$. Moreover, $$ \|X - \mathbb{E} X\|_{\psi_2} \le 2 \|X\|_{\psi_2}, \quad \|X - \mathbb{E} X\|_{\psi_1} \le 2 \|X\|_{\psi_1}. $$ This follows by triangle inequality $\|X - \mathbb{E} X\|_{\psi_2} \le \|X\|_{\psi_2} + \|\mathbb{E} X\|_{\psi_2}$ along with $\|\mathbb{E} X\|_{\psi_2} = |\mathbb{E} X| \le \mathbb{E}|X| \le \|X\|_{\psi_2}$, and similarly for the sub-exponential norm. \end{remark} \subsection{Isotropic random vectors} \index{Isotropic random vectors} \label{s: isotropic} Now we carry our work over to higher dimensions. We will thus be working with random vectors $X$ in $\mathbb{R}^n$, or equivalently probability distributions in $\mathbb{R}^n$. While the concept of the mean $\mu = \mathbb{E} Z$ of a random variable $Z$ remains the same in higher dimensions, the second moment $\mathbb{E} Z^2$ is replaced by the $n \times n$ {\em second moment matrix} \index{Second moment matrix} of a random vector $X$, defined as $$ \Sigma = \Sigma(X) = \mathbb{E} X \otimes X = \mathbb{E} X X^T $$ where $\otimes$ denotes the outer product of vectors in $\mathbb{R}^n$. Similarly, the concept of variance $\Var(Z) = \mathbb{E}(Z - \mu)^2 = \mathbb{E} Z^2 - \mu^2$ of a random variable is replaced in higher dimensions with the {\em covariance matrix} \index{Covariance matrix} of a random vector $X$, defined as $$ \Cov(X) = \mathbb{E} (X-\mu) \otimes (X-\mu) = \mathbb{E} X \otimes X - \mu \otimes \mu $$ where $\mu = \mathbb{E} X$. By translation, many questions can be reduced to the case of centered random vectors, for which $\mu = 0$ and $\Cov(X) = \Sigma(X)$. We will also need a higher-dimensional version of unit variance: \begin{definition}[Isotropic random vectors] \index{Isotropic random vectors} A random vector $X$ in $\mathbb{R}^n$ is called {\em isotropic} if $\Sigma(X) = I$. Equivalently, $X$ is isotropic if \begin{equation} \label{isotropy} \mathbb{E} \< X, x\> ^2 = \|x\|_2^2 \quad \text{for all } x \in \mathbb{R}^n. \end{equation} \end{definition} Suppose $\Sigma(X)$ is an invertible matrix, which means that the distribution of $X$ is not essentially supported on any proper subspace of $\mathbb{R}^n$. Then $\Sigma(X)^{-1/2} X$ is an isotropic random vector in $\mathbb{R}^n$. Thus every non-degenerate random vector can be made isotropic by an appropriate linear transformation.\footnote{This transformation (usually preceded by centering) is a higher-dimensional version of {\em standardizing} of random variables, which enforces zero mean and unit variance.} This allows us to mostly focus on studying isotropic random vectors in the future. \begin{lemma} \label{norm isotropic} Let $X,Y$ be independent isotropic random vectors in $\mathbb{R}^n$. Then $\mathbb{E} \|X\|_2^2 = n$ and $\mathbb{E} \< X,Y\> ^2 = n$. \end{lemma} \begin{proof} The first part follows from $\mathbb{E} \|X\|_2^2 = \mathbb{E} \tr(X \otimes X) = \tr(\mathbb{E} X \otimes X) = \tr(I) = n$. The second part follows by conditioning on $Y$, using isotropy of $X$ and using the first part for $Y$: this way we obtain $\mathbb{E} \< X,Y\> ^2 = \mathbb{E} \|Y\|_2^2 = n$. \end{proof} \begin{example} \label{random vectors} \begin{enumerate} \item {\bf (Gaussian):} The (standard) {\em Gaussian random vector} \index{Gaussian!random vectors} $X$ in $\mathbb{R}^n$ chosen according to the standard normal distribution $N(0, I)$ is isotropic. The coordinates of $X$ are independent standard normal random variables. \item {\bf (Bernoulli):} \index{Bernoulli!random vectors} A similar example of a discrete isotropic distribution is given by a {\em Bernoulli random vector} $X$ in $\mathbb{R}^n$ whose coordinates are independent symmetric Bernoulli random variables. \item {\bf (Product distributions):} More generally, consider a random vector $X$ in $\mathbb{R}^n$ whose coordinates are independent random variables with zero mean and unit variance. Then clearly $X$ is an isotropic vector in $\mathbb{R}^n$. \item {\bf (Coordinate):} \index{Coordinate random vectors} Consider a {\em coordinate random vector} $X$, which is uniformly distributed in the set $\{ \sqrt{n} \, e_i \}_{i=1}^n$ where $\{e_i \}_{i=1}^n$ is the canonical basis of $\mathbb{R}^n$. Clearly $X$ is an isotropic random vector in $\mathbb{R}^n$.\footnote{The examples of Gaussian and coordinate random vectors are somewhat opposite -- one is very continuous and the other is very discrete. They may be used as test cases in our study of random matrices.} \item {\bf (Frame):} \index{Frames} This is a more general version of the coordinate random vector. A {\em frame} is a set of vectors $\{u_i\}_{i=1}^M$ in $\mathbb{R}^n$ which obeys an approximate Parseval's identity, i.e. there exist numbers $A,B>0$ called {\em frame bounds} such that $$ A\|x\|_2^2 \le \sum_{i=1}^M \< u_i, x \> ^2 \le B\|x\|_2^2 \quad \text{for all } x \in \mathbb{R}^n. $$ If $A=B$ the set is called a {\em tight frame}. Thus, tight frames are generalizations of orthogonal bases without linear independence. Given a tight frame $\{u_i\}_{i=1}^M$ with bounds $A=B=M$, the random vector $X$ uniformly distributed in the set $\{u_i \}_{i=1}^M$ is clearly isotropic in $\mathbb{R}^n$.\footnote{There is clearly a reverse implication, too, which shows that the class of tight frames can be identified with the class of discrete isotropic random vectors.} \item{\bf (Spherical):} \index{Spherical random vector} Consider a random vector $X$ uniformly distributed on the unit Euclidean sphere in $\mathbb{R}^n$ with center at the origin and radius $\sqrt{n}$. Then $X$ is isotropic. Indeed, by rotation invariance $\mathbb{E} \< X,x\> ^2$ is proportional to $\|x\|_2^2$; the correct normalization $\sqrt{n}$ is derived from Lemma~\ref{norm isotropic}. \item {\bf (Uniform on a convex set):} In convex geometry, a convex set $K$ in $\mathbb{R}^n$ is called isotropic if a random vector $X$ chosen uniformly from $K$ according to the volume is isotropic. As we noted, every full dimensional convex set can be made into an isotropic one by an affine transformation. Isotropic convex sets look ``well conditioned'', which is advantageous in geometric algorithms (e.g. volume computations). \end{enumerate} \end{example} We generalize the concepts of sub-gaussian random variables to higher dimensions using one-dimensional marginals. \begin{definition}[Sub-gaussian random vectors] \index{Sub-gaussian!random vectors} \label{d: sub-gaussian} We say that a random vector $X$ in $\mathbb{R}^n$ is {\em sub-gaussian} if the one-dimensional marginals $\< X, x\> $ are sub-gaussian random variables for all $x \in \mathbb{R}^n$. The {\em sub-gaussian norm} \index{Sub-gaussian!norm} of $X$ is defined as $$ \|X\|_{\psi_2} = \sup_{x \in S^{n-1}} \|\< X, x\> \|_{\psi_2}. $$ \end{definition} \begin{remark}[Properties of high-dimensional distributions] The definitions of isotropic and sub-gaussian distributions suggest that more generally, natural properties of high-dimensional distributions may be defined via one-dimensional marginals. This is a natural way to generalize properties of random variables to random vectors. For example, we shall call a random vector sub-exponential if all of its one-dimensional marginals are sub-exponential random variables, etc. \end{remark} One simple way to create sub-gaussian distributions in $\mathbb{R}^n$ is by taking a product of $n$ sub-gaussian distributions on the line: \begin{lemma}[Product of sub-gaussian distributions] \label{sub-gaussian products} Let $X_1,\ldots,X_n$ be independent centered sub-gaussian random variables. Then $X = (X_1,\ldots,X_n)$ is a centered sub-gaussian random vector in $\mathbb{R}^n$, and $$ \|X\|_{\psi_2} \le C \max_{i \le n} \|X_i\|_{\psi_2} $$ where $C$ is an absolute constant. \end{lemma} \begin{proof} This is a direct consequence of the rotation invariance principle, Lemma~\ref{rotation invariance}. Indeed, for every $x = (x_1,\ldots,x_n) \in S^{n-1}$ we have $$ \|\< X, x\> \|_{\psi_2} = \Big\| \sum_{i=1}^n x_i X_i \Big\|_{\psi_2} \le C \sum_{i=1}^n x_i^2 \|X_i\|_{\psi_2}^2 \le C \max_{i \le n} \|X_i\|_{\psi_2} $$ where we used that $\sum_{i=1}^n x_i^2 = 1$. This completes the proof. \end{proof} \begin{example} \label{random vectors sub-gaussian} Let us analyze the basic examples of random vectors introduced earlier in Example~\ref{random vectors}. \begin{enumerate} \item {\bf (Gaussian, Bernoulli):} \index{Gaussian!random vectors} \index{Bernoulli!random vectors} Gaussian and Bernoulli random vectors are sub-gaussian; their sub-gaussian norms are bounded by an absolute constant. These are particular cases of Lemma~\ref{sub-gaussian products}. \item {\bf (Spherical):} \index{Spherical random vector} A spherical random vector is also sub-gaussian; its sub-gaussian norm is bounded by an absolute constant. Unfortunately, this does not follow from Lemma~\ref{sub-gaussian products} because the coordinates of the spherical vector are not independent. Instead, by rotation invariance, the claim clearly follows from the following geometric fact. For every $\varepsilon \ge 0$, the spherical cap $\{ x \in S^{n-1}:\; x_1 > \varepsilon\}$ makes up at most $\exp(-\varepsilon^2 n/2)$ proportion of the total area on the sphere.\footnote{This fact about spherical caps may seem counter-intuitive. For example, for $\varepsilon = 0.1$ the cap looks similar to a hemisphere, but the proportion of its area goes to zero very fast as dimension $n$ increases. This is a starting point of the study of the {\em concentration of measure phenomenon}, see \cite{Ledoux}.} This can be proved directly by integration, and also by elementary geometric considerations \cite[Lemma~2.2]{Ball}. \item {\bf (Coordinate):} \index{Coordinate random vectors} Although the coordinate random vector $X$ is formally sub-gaussian as its support is finite, its sub-gaussian norm is too big: $\|X\|_{\psi_2} = \sqrt{n} \gg 1$. So we would not think of $X$ as a sub-gaussian random vector. \item {\bf (Uniform on a convex set):} For many isotropic convex sets $K$ (called $\psi_2$ bodies), a random vector $X$ uniformly distributed in $K$ is sub-gaussian with $\|X\|_{\psi_2} = O(1)$. For example, the cube $[-1,1]^n$ is a $\psi_2$ body by Lemma~\ref{sub-gaussian products}, while the appropriately normalized cross-polytope $\{ x \in \mathbb{R}^n:\; \|x\|_1 \le M \}$ is not. Nevertheless, Borell's lemma (which is a consequence of Brunn-Minkowski inequality) implies a weaker property, that $X$ is always {\em sub-exponential}, and $\|X\|_{\psi_1} = \sup_{x \in S^{n-1}} \|\< X, x\> \|_{\psi_1}$ is bounded by absolute constant. See \cite[Section~2.2.b$_3$]{GiMi} for a proof and discussion of these ideas. \end{enumerate} \end{example} \subsection{Sums of independent random matrices} \label{s: sums matrices} In this section, we mention without proof some results of classical probability theory in which scalars can be replaced by matrices. Such results are useful in particular for problems on random matrices, since we can view a random matrix as a generalization of a random variable. One such remarkable generalization is valid for Khintchine inequality, Corollary~\ref{Khintchine}. The scalars $a_i$ can be replaced by matrices, and the absolute value by the {\em Schatten norm}. \index{Schatten norm} Recall that for $1 \le p \le \infty$, the $p$-Schatten norm of an $n \times n$ matrix $A$ is defined as the $\ell_p$ norm of the sequence of its singular values: $$ \|A\|_{C_p^n} = \| (s_i(A))_{i=1}^n\|_p = \big( \sum_{i=1}^n s_i(A)^p \big)^{1/p}. $$ For $p=\infty$, the Schatten norm equals the spectral norm $\|A\| = \max_{i \le n} s_i(A)$. Using this one can quickly check that already for $p = \log n$ the Schatten and spectral norms are equivalent: $\|A\|_{C_p^n} \le \|A\| \le e \|A\|_{C_p^n}$. \begin{theorem}[Non-commutative Khintchine inequality, see \cite{Pisier operator} Section 9.8] \index{Khinchine inequality!non-commutative} \label{non-commutative Khintchine} \hfill Let $A_1, \ldots, A_N$ be self-adjoint $n \times n$ matrices and $\varepsilon_1, \ldots, \varepsilon_N$ be independent symmetric Bernoulli random variables. Then, for every $2 \le p < \infty$, we have $$ \Big\| \Big( \sum_{i=1}^N A_i^2 \Big)^{1/2} \Big\|_{C_p^n} \le \Big( \mathbb{E} \Big\| \sum_{i=1}^N \varepsilon_i A_i \Big\|_{C_p^n}^p \Big)^{1/p} \le C \sqrt{p} \, \Big\| \Big( \sum_{i=1}^N A_i^2 \Big)^{1/2} \Big\|_{C_p^n} $$ where $C$ is an absolute constant. \end{theorem} \begin{remark} \begin{enumerate} \item The scalar case of this result, for $n=1$, recovers the classical Khintchine inequality, Corollary~\ref{Khintchine}, for $X_i = \varepsilon_i$. \item By the equivalence of Schatten and spectral norms for $p=\log n$, a version of non-commutative Khintchine inequality holds for the spectral norm: \begin{equation} \label{Khintchine operator norm} \mathbb{E} \Big\| \sum_{i=1}^N \varepsilon_i A_i \Big\| \le C_1 \sqrt{\log n} \, \Big\| \Big( \sum_{i=1}^N A_i^2 \Big)^{1/2} \Big\| \end{equation} where $C_1$ is an absolute constant. The logarithmic factor is unfortunately essential; it role will be clear when we discuss applications of this result to random matrices in the next sections. \end{enumerate} \end{remark} \begin{corollary}[Rudelson's inequality \cite{Rudelson isotropic}] \index{Rudelson's inequality} \label{Rudelson} Let $x_1, \ldots, x_N$ be vectors in $\mathbb{R}^n$ and $\varepsilon_1, \ldots, \varepsilon_N$ be independent symmetric Bernoulli random variables. Then $$ \mathbb{E} \Big\| \sum_{i=1}^N \varepsilon_i x_i \otimes x_i \Big\| \le C \sqrt{\log \min(N,n)} \cdot \max_{i \le N} \|x_i\|_2 \cdot \Big\| \sum_{i=1}^N x_i \otimes x_i \Big\|^{1/2} $$ where $C$ is an absolute constant. \end{corollary} \begin{proof} One can assume that $n \le N$ by replacing $\mathbb{R}^n$ with the linear span of $\{x_1,\ldots,x_N\}$ if necessary. The claim then follows from \eqref{Khintchine operator norm}, since $$ \Big\| \Big( \sum_{i=1}^N (x_i \otimes x_i)^2 \Big)^{1/2} \Big\| = \Big\| \sum_{i=1}^N \|x_i\|_2^2 \; x_i \otimes x_i \Big\|^{1/2} \le \max_{i \le N} \|x_i\|_2 \Big\| \sum_{i=1}^N x_i \otimes x_i \Big\|^{1/2}. \qedhere $$ \end{proof} Ahlswede and Winter \cite{AW} pioneered a different approach to matrix-valued inequalities in probability theory, which was based on trace inequalities like Golden-Thompson inequality. A development of this idea leads to remarkably sharp results. We quote one such inequality from \cite{Tropp}: \begin{theorem}[Non-commutative Bernstein-type inequality \cite{Tropp}] \index{Bernstein-type inequality!non-commutative} \label{matrix Bernstein} Consider a finite sequence $X_i$ of independent centered self-adjoint random $n \times n$ matrices. Assume we have for some numbers $K$ and $\sigma$ that $$ \|X_i\| \le K \text{ almost surely}, \quad \big\| \sum_i \mathbb{E} X_i^2 \big\| \le \sigma^2. $$ Then, for every $t \ge 0$ we have \begin{equation} \label{eq matrix Bernstein} \mathbb{P} \Big\{ \big\| \sum_i X_i \big\| \ge t \Big\} \le 2 n \cdot \exp \Big( \frac{-t^2/2}{\sigma^2 + Kt/3} \Big). \end{equation} \end{theorem} \begin{remark} \label{mixed tail} This is a direct matrix generalization of a classical Bernstein's inequality for bounded random variables. To compare it with our version of Bernstein's inequality for sub-exponentials, Proposition~\ref{sub-exponential large deviations}, note that the probability bound in \eqref{eq matrix Bernstein} is equivalent to $2n \cdot \exp \big[ -c \min \big( \frac{t^2}{\sigma^2}, \frac{t}{K} \big) \big]$ where $c>0$ is an absolute constant. In both results we see a mixture of gaussian and exponential tails. \end{remark} \section{Random matrices with independent entries} \label{s: entries} We are ready to study the extreme singular values of random matrices. In this section, we consider the classical model of random matrices whose entries are independent and centered random variables. Later we will study the more difficult models where only the rows or the columns are independent. The reader may keep in mind some classical examples of $N \times n$ random matrices with independent entries. The most classical example is the {\em Gaussian random matrix} $A$ \index{Gaussian!random matrices} whose entries are independent standard normal random variables. In this case, the $n \times n$ symmetric matrix $A^*A$ is called Wishart matrix; it is a higher-dimensional version of chi-square distributed random variables. The simplest example of discrete random matrices is the {\em Bernoulli random matrix} $A$ \index{Bernoulli!random matrices} whose entries are independent symmetric Bernoulli random variables. In other words, Bernoulli random matrices are distributed uniformly in the set of all $N \times n$ matrices with $\pm 1$ entries. \subsection{Limit laws and Gaussian matrices} Consider an $N \times n$ random matrix $A$ whose entries are independent centered identically distributed random variables. By now, the {\em limiting behavior} of the extreme singular values of $A$, as the dimensions $N, n \to \infty$, is well understood: \begin{theorem}[Bai-Yin's law, see \cite{Bai-Yin}] \index{Bai-Yin's law} \label{Bai-Yin} Let $A = A_{N,n}$ be an $N \times n$ random matrix whose entries are independent copies of a random variable with zero mean, unit variance, and finite fourth moment. Suppose that the dimensions $N$ and $n$ grow to infinity while the aspect ratio $n/N$ converges to a constant in $[0,1]$. Then $$ s_{\min}(A) = \sqrt{N} - \sqrt{n} + o(\sqrt{n}), \quad s_{\max}(A) = \sqrt{N} + \sqrt{n} + o(\sqrt{n}) \quad \text{almost surely}. $$ \end{theorem} As we pointed out in the introduction, our program is to find non-asymptotic versions of Bai-Yin's law. There is precisely one model of random matrices, namely Gaussian, where an {\em exact} non-asymptotic result is known: \begin{theorem}[Gordon's theorem for Gaussian matrices] \index{Gordon's theorem} \label{Gaussian} Let $A$ be an $N \times n$ matrix whose entries are independent standard normal random variables. Then $$ \sqrt{N} - \sqrt{n} \le \mathbb{E} s_{\min}(A) \le \mathbb{E} s_{\max}(A) \le \sqrt{N} + \sqrt{n}. $$ \end{theorem} The proof of the upper bound, which we borrowed from \cite{DS}, is based on Slepian's comparison inequality for Gaussian processes.\footnote{Recall that a Gaussian process $(X_t)_{t \in T}$ is a collection of centered normal random variables $X_t$ on the same probability space, indexed by points $t$ in an abstract set $T$.} \begin{lemma}[Slepian's inequality, see \cite{Ledoux-Talagrand} Section 3.3] \index{Slepian's inequality} \label{Slepian} Consider two Gaussian processes $(X_t)_{t \in T}$ and $(Y_t)_{t \in T}$ whose increments satisfy the inequality $\mathbb{E} |X_s - X_t|^2 \le \mathbb{E} |Y_s - Y_t|^2$ for all $s,t \in T$. Then $\mathbb{E} \sup_{t \in T} X_t \le \mathbb{E} \sup_{t \in T} Y_t$. \end{lemma} \begin{proof}[Proof of Theorem~\ref{Gaussian}] We recognize $s_{\max}(A) = \max_{u \in S^{n-1}, \; v \in S^{N-1}} \< Au, v\> $ to be the supremum of the Gaussian process $X_{u,v} = \< Au, v\> $ indexed by the pairs of vectors $(u,v) \in S^{n-1} \times S^{N-1}$. We shall compare this process to the following one whose supremum is easier to estimate: $Y_{u,v} = \< g, u\> + \< h, v\> $ where $g \in \mathbb{R}^n$ and $h \in \mathbb{R}^N$ are independent standard Gaussian random vectors. The rotation invariance of Gaussian measures makes it easy to compare the increments of these processes. For every $(u,v), (u',v') \in S^{n-1} \times S^{N-1}$, one can check that $$ \mathbb{E} |X_{u,v} - X_{u',v'}|^2 = \sum_{i=1}^n \sum_{j=1}^N |u_i v_j - u'_i v'_j|^2 \le \|u - u'\|_2^2 + \|v - v'\|_2^2 = \mathbb{E} |Y_{u,v} - Y_{u',v'}|^2. $$ Therefore Lemma~\ref{Slepian} applies, and it yields the required bound $$ \mathbb{E} s_{\max}(A) = \mathbb{E} \max_{(u,v)} X_{u,v} \le \mathbb{E} \max_{(u,v)}Y_{u,v} = \mathbb{E} \|g\|_2 + \mathbb{E} \|h\|_2 \le \sqrt{N} + \sqrt{n}. $$ Similar ideas are used to estimate $\mathbb{E} s_{\min}(A) = \mathbb{E} \max_{v \in S^{N-1}} \min_{u \in S^{n-1}} \< Au, v\> $, see \cite{DS}. One uses in this case Gordon's generalization of Slepian's inequality for minimax of Gaussian processes \cite{Gordon 84, Gordon 85, Gordon 92}, see \cite[Section 3.3]{Ledoux-Talagrand}. \end{proof} While Theorem~\ref{Gaussian} is about the expectation of singular values, it also yields a large deviation inequality for them. It can be deduced formally by using the {\em concentration of measure} in the Gauss space. \begin{proposition}[Concentration in Gauss space, see \cite{Ledoux}] \index{Concentration of meaure} \label{Gaussian concentration} Let $f$ be a real valued Lipschitz function on $\mathbb{R}^n$ with Lipschitz constant $K$, i.e. $|f(x)-f(y)| \le K \|x-y\|_2$ for all $x,y \in \mathbb{R}^n$ (such functions are also called $K$-Lipschitz). Let $X$ be the standard normal random vector in $\mathbb{R}^n$. Then for every $t \ge 0$ one has $$ \mathbb{P} \{ f(X) - \mathbb{E} f(X) > t \} \le \exp(-t^2/2K^2). $$ \end{proposition} \begin{corollary}[Gaussian matrices, deviation; see \cite{DS}] \index{Gaussian!random matrices} \label{Gaussian deviation} Let $A$ be an $N \times n$ matrix whose entries are independent standard normal random variables. Then for every $t \ge 0$, with probability at least $1 - 2 \exp(-t^2/2)$ one has $$ \sqrt{N} - \sqrt{n} - t \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{N} + \sqrt{n} + t. $$ \end{corollary} \begin{proof} Note that $s_{\min}(A)$, $s_{\max}(A)$ are $1$-Lipschitz functions of matrices $A$ considered as vectors in $\mathbb{R}^{Nn}$. The conclusion now follows from the estimates on the expectation (Theorem~\ref{Gaussian}) and Gaussian concentration (Proposition~\ref{Gaussian concentration}). \end{proof} Later in these notes, we find it more convenient to work with the $n \times n$ positive-definite symmetric matrix $A^*A$ rather than with the original $N \times n$ matrix $A$. Observe that the normalized matrix $\bar{A} = \frac{1}{\sqrt{N}} A$ is an approximate isometry (which is our goal) if and only if $\bar{A}^*\bar{A}$ is an approximate identity: \begin{lemma}[Approximate isometries] \index{Approximate isometries} \label{approximate isometries} Consider a matrix $B$ that satisfies \begin{equation} \label{B*B} \|B^*B - I\| \le \max(\delta,\delta^2) \end{equation} for some $\delta > 0$. Then \begin{equation} \label{smin smax B} 1-\delta \le s_{\min}(B) \le s_{\max}(B) \le 1+\delta. \end{equation} Conversely, if $B$ satisfies \eqref{smin smax B} for some $\delta > 0$ then $\|B^*B - I\| \le 3 \max(\delta,\delta^2)$. \end{lemma} \begin{proof} Inequality \eqref{B*B} holds if and only if $\big| \|Bx\|_2^2 - 1 \big| \le \max(\delta,\delta^2)$ for all $x \in S^{n-1}$. Similarly, \eqref{smin smax B} holds if and only if $\big| \|Bx\|_2 - 1 \big| \le \delta$ for all $x \in S^{n-1}$. The conclusion then follows from the elementary inequality $$ \max(|z-1|, |z-1|^2) \le |z^2-1| \le 3 \max(|z-1|, |z-1|^2) \quad \text{for all } z \ge 0. \qedhere $$ \end{proof} Lemma~\ref{approximate isometries} reduces our task of proving inequalities \eqref{heuristic} to showing an equivalent (but often more convenient) bound $$ \big\| \frac{1}{N} A^*A-I \big\| \le \max(\delta, \delta^2) \quad \text{where } \delta = O(\sqrt{n/N}). $$ \subsection{General random matrices with independent entries} Now we pass to a more general model of random matrices whose entries are independent centered random variables with some general distribution (not necessarily normal). The largest singular value (the spectral norm) can be estimated by Latala's theorem for general random matrices with non-identically distributed entries: \begin{theorem}[Latala's theorem \cite{Latala}] \index{Latala's theorem} \label{Latala} Let $A$ be a random matrix whose entries $a_{ij}$ are independent centered random variables with finite fourth moment. Then $$ \mathbb{E} s_{\max}(A) \le C \Big[ \max_i \big( \sum_j \mathbb{E} a_{ij}^2 \big)^{1/2} + \max_j \big( \sum_i \mathbb{E} a_{ij}^2 \big)^{1/2} + \big( \sum_{i,j} \mathbb{E} a_{ij}^4 \big)^{1/4} \Big]. $$ \end{theorem} If the variance and the fourth moments of the entries are uniformly bounded, then Latala's result yields $s_{\max}(A) = O(\sqrt{N} + \sqrt{n})$. This is slightly weaker than our goal \eqref{heuristic}, which is $s_{\max}(A) = \sqrt{N} + O(\sqrt{n})$ but still satisfactory for most applications. Results of the latter type will appear later in the more general model of random matrices with independent rows or columns. Similarly, our goal \eqref{heuristic} for the smallest singular value is $s_{\min}(A) \ge \sqrt{N} - O(\sqrt{n})$. Since the singular values are non-negative anyway, such inequality would only be useful for sufficiently tall matrices, $N \gg n$. For almost square and square matrices, estimating the smallest singular value (known also as the {\em hard edge} of spectrum) is considerably more difficult. The progress on estimating the hard edge is summarized in \cite{RV ICM}. If $A$ has independent entries, then indeed $s_{\min}(A) \ge c (\sqrt{N} - \sqrt{n})$, and the following is an optimal probability bound: \begin{theorem}[Independent entries, hard edge \cite{RV rectangular}] \index{Hard edge of spectrum} \label{RV rectangular} Let $A$ be an $n \times n$ random matrix whose entries are independent identically distributed subgaussian random variables with zero mean and unit variance. Then for $\varepsilon \ge 0$, $$ \mathbb{P} \big( s_{\min}(A) \le \varepsilon (\sqrt{N} - \sqrt{n-1}) \big) \le (C\varepsilon)^{N-n+1} + c^N $$ where $C > 0$ and $c \in (0,1)$ depend only on the subgaussian norm of the entries. \end{theorem} This result gives an optimal bound for square matrices as well ($N=n$). \section{Random matrices with independent rows} \label{s: rows} In this section, we focus on a more general model of random matrices, where we only assume independence of the rows rather than all entries. Such matrices are naturally {\em generated by high-dimensional distributions}. Indeed, given an arbitrary probability distribution in $\mathbb{R}^n$, one takes a sample of $N$ independent points and arranges them as the rows of an $N \times n$ matrix $A$. By studying spectral properties of $A$ one should be able to learn something useful about the underlying distribution. For example, as we will see in Section~\ref{s: covariance}, the extreme singular values of $A$ would tell us whether the covariance matrix of the distribution can be estimated from a sample of size $N$. The picture will vary slightly depending on whether the rows of $A$ are sub-gaussian or have arbitrary distribution. For heavy-tailed distributions, an extra logarithmic factor has to appear in our desired inequality \eqref{heuristic}. The analysis of sub-gaussian and heavy-tailed matrices will be completely different. There is an abundance of examples where the results of this section may be useful. They include all matrices with independent entries, whether sub-gaussian such as Gaussian and Bernoulli, or completely general distributions with mean zero and unit variance. In the latter case one is able to surpass the fourth moment assumption which is necessary in Bai-Yin's law, Theorem~\ref{Bai-Yin}. Other examples of interest come from non-product distributions, some of which we saw in Example~\ref{random vectors}. Sampling from discrete objects (matrices and frames) fits well in this framework, too. Given a deterministic matrix $B$, one puts a uniform distribution on the set of the rows of $B$ and creates a random matrix $A$ as before -- by sampling some $N$ random rows from $B$. Applications to sampling will be discussed in Section~\ref{s: sub-matrices}. \subsection{Sub-gaussian rows} \index{Sub-gaussian!random matrices with independent rows} \label{s: sub-gaussian rows} The following result goes in the direction of our goal \eqref{heuristic} for random matrices with independent sub-gaussian rows. \begin{theorem}[Sub-gaussian rows] \label{sub-gaussian rows} Let $A$ be an $N \times n$ matrix whose rows $A_i$ are independent sub-gaussian isotropic random vectors in $\mathbb{R}^n$. Then for every $t \ge 0$, with probability at least $1 - 2\exp(-ct^2)$ one has \begin{equation} \label{smin smax rectangular} \sqrt{N} - C \sqrt{n} - t \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{N} + C \sqrt{n} + t. \end{equation} Here $C = C_K$, $c = c_K > 0$ depend only on the subgaussian norm $K = \max_i \|A_i\|_{\psi_2}$ of the rows. \end{theorem} This result is a general version of Corollary~\ref{Gaussian deviation} (up to absolute constants); instead of independent Gaussian entries we allow independent sub-gaussian rows. This of course covers all matrices with independent sub-gaussian entries such as Gaussian and Bernoulli. It also applies to some natural matrices whose entries are not independent. One such example is a matrix whose rows are independent spherical random vectors (Example~\ref{random vectors sub-gaussian}). \begin{proof} The proof is a basic version of a {\em covering argument}, \index{Covering argument} and it has three steps. We need to control $\|Ax\|_2$ for all vectors $x$ on the unit sphere $S^{n-1}$. To this end, we discretize the sphere using a net $\mathcal{N}$ (the approximation step), establish a tight control of $\|Ax\|_2$ for every fixed vector $x \in \mathcal{N}$ with high probability (the concentration step), and finish off by taking a union bound over all $x$ in the net. The concentration step will be based on the deviation inequality for sub-exponential random variables, Corollary~\ref{average sub-exponentials}. {\bf Step 1: Approximation.} Recalling Lemma~\ref{approximate isometries} for the matrix $B=A/\sqrt{N}$ we see that the conclusion of the theorem is equivalent to \begin{equation} \label{A*A rows} \big\| \frac{1}{N}A^*A-I \big\| \le \max(\delta, \delta^2) =:\varepsilon \quad \text{where} \quad \delta = C \sqrt{\frac{n}{N}} + \frac{t}{\sqrt{N}}. \end{equation} Using Lemma~\ref{norm on net}, we can evaluate the operator norm in \eqref{A*A rows} on a $\frac{1}{4}$-net $\mathcal{N}$ of the unit sphere $S^{n-1}$: $$ \big\| \frac{1}{N}A^*A-I \big\| \le 2 \max_{x \in \mathcal{N}} \big| \big\langle (\frac{1}{N}A^*A-I)x, x \big\rangle \big| = 2 \max_{x \in \mathcal{N}} \big| \frac{1}{N} \|Ax\|_2^2 - 1 \big|. $$ So to complete the proof it suffices to show that, with the required probability, $$ \max_{x \in \mathcal{N}} \big| \frac{1}{N} \|Ax\|_2^2 - 1 \big| \le \frac{\varepsilon}{2}. $$ By Lemma~\ref{net cardinality}, we can choose the net $\mathcal{N}$ so that it has cardinality $|\mathcal{N}| \le 9^n$. {\bf Step 2: Concentration.} Let us fix any vector $x \in S^{n-1}$. We can express $\|Ax\|_2^2$ as a sum of independent random variables \begin{equation} \label{Ax as sum} \|Ax\|_2^2 = \sum_{i=1}^N \< A_i, x\> ^2 =: \sum_{i=1}^N Z_i^2 \end{equation} where $A_i$ denote the rows of the matrix $A$. By assumption, $Z_i = \< A_i, x\> $ are independent sub-gaussian random variables with $\mathbb{E} Z_i^2 = 1$ and $\|Z_i\|_{\psi_2} \le K$. Therefore, by Remark~\ref{centering} and Lemma~\ref{sub-exponential squared}, $Z_i^2 - 1$ are independent centered sub-exponential random variables with $\|Z_i^2-1\|_{\psi_1} \le 2\|Z_i^2\|_{\psi_1} \le 4 \|Z_i\|_{\psi_2}^2 \le 4 K^2$. We can therefore use an exponential deviation inequality, Corollary~\ref{average sub-exponentials}, to control the sum \eqref{Ax as sum}. Since $K \ge \|Z_i\|_{\psi_2} \ge \frac{1}{\sqrt{2}} (\mathbb{E}|Z_i|^2)^{1/2} = \frac{1}{\sqrt{2}}$, this gives \begin{align*} \mathbb{P} \Big\{ \big| \frac{1}{N} \|Ax\|_2^2 - 1 \big| \ge \frac{\varepsilon}{2} \Big\} &= \mathbb{P} \Big\{ \big| \frac{1}{N}\sum_{i=1}^N Z_i^2 - 1 \big| \ge \frac{\varepsilon}{2} \Big\} \le 2 \exp \Big[ - \frac{c_1}{K^4} \min(\varepsilon^2, \varepsilon) N \Big] \\ &= 2 \exp \Big[ - \frac{c_1}{K^4} \delta^2 N \Big] \le 2 \exp \Big[ - \frac{c_1}{K^4} (C^2 n + t^2) \Big] \end{align*} where the last inequality follows by the definition of $\delta$ and using the inequality $(a+b)^2 \ge a^2 + b^2$ for $a,b \ge 0$. {\bf Step 3: Union bound.} Taking the union bound over all vectors $x$ in the net $\mathcal{N}$ of cardinality $|\mathcal{N}| \le 9^n$, we obtain $$ \mathbb{P} \Big\{ \max_{x \in \mathcal{N}} \big| \frac{1}{N} \|Ax\|_2^2 - 1 \big| \ge \frac{\varepsilon}{2} \Big\} \le 9^n \cdot 2 \exp \Big[ - \frac{c_1}{K^4} (C^2 n + t^2) \Big] \le 2 \exp \Big( - \frac{c_1 t^2}{K^4} \Big) $$ where the second inequality follows for $C = C_K$ sufficiently large, e.g. $C = K^2 \sqrt{\ln 9/c_1}$. As we noted in Step~1, this completes the proof of the theorem. \end{proof} \begin{remark}[Non-isotropic distributions] \label{r: non-isotropic} \begin{enumerate} \item A version of Theorem~\ref{sub-gaussian rows} holds for general, non-isotropic sub-gaussian distributions. Assume that $A$ is an $N \times n$ matrix whose rows $A_i$ are independent sub-gaussian random vectors in $\mathbb{R}^n$ with second moment matrix $\Sigma$. Then for every $t \ge 0$, the following inequality holds with probability at least $1 - 2\exp(-ct^2)$: \begin{equation} \label{A*A rows non-isotropic} \big\| \frac{1}{N}A^*A-\Sigma \big\| \le \max(\delta, \delta^2) \quad \text{where} \quad \delta = C \sqrt{\frac{n}{N}} + \frac{t}{\sqrt{N}}. \end{equation} Here as before $C = C_K$, $c = c_K > 0$ depend only on the subgaussian norm $K = \max_i \|A_i\|_{\psi_2}$ of the rows. This result is a general version of \eqref{A*A rows}. It follows by a straighforward modification of the argument of Theorem~\ref{sub-gaussian rows}. \item A more natural, multiplicative form of \eqref{A*A rows non-isotropic} is the following. Assume that $\Sigma^{-1/2} A_i$ are isotropic sub-gaussian random vectors, and let $K$ be the maximum of their sub-gaussian norms. Then for every $t \ge 0$, the following inequality holds with probability at least $1 - 2\exp(-ct^2)$: \begin{equation} \label{A*A rows non-isotropic multiplicative} \big\| \frac{1}{N}A^*A-\Sigma \big\| \le \max(\delta, \delta^2) \, \|\Sigma\| \quad \text{where} \quad \delta = C \sqrt{\frac{n}{N}} + \frac{t}{\sqrt{N}} \end{equation} Here again $C = C_K$, $c = c_K > 0$. This result follows from Theorem~\ref{sub-gaussian rows} applied to the isotropic random vectors $\Sigma^{-1/2} A_i$. \end{enumerate} \end{remark} \subsection{Heavy-tailed rows} \index{Heavy-tailed!random matrices with independent rows} \label{s: heavy-tailed rows} The class of sub-gaussian random variables in Theorem~\ref{sub-gaussian rows} may sometimes be too restrictive in applications. For example, if the rows of $A$ are independent coordinate or frame random vectors (Examples~\ref{random vectors} and \ref{random vectors sub-gaussian}), they are poorly sub-gaussian and Theorem~\ref{sub-gaussian rows} is too weak. In such cases, one would use the following result instead, which operates in remarkable generality. \begin{theorem}[Heavy-tailed rows] \label{heavy-tailed rows} Let $A$ be an $N \times n$ matrix whose rows $A_i$ are independent isotropic random vectors in $\mathbb{R}^n$. Let $m$ be a number such that $\|A_i\|_2 \le \sqrt{m}$ almost surely for all $i$. Then for every $t \ge 0$, one has \begin{equation} \label{eq heavy-tailed rows} \sqrt{N} - t \sqrt{m} \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{N} + t \sqrt{m} \end{equation} with probability at least $1 - 2 n \cdot \exp(-ct^2)$, where $c>0$ is an absolute constant. \end{theorem} Recall that $(\mathbb{E} \|A_i\|_2^2)^{1/2} = \sqrt{n}$ by Lemma~\ref{norm isotropic}. This indicates that one would typically use Theorem~\ref{heavy-tailed rows} with $m = O(n)$. In this case the result takes the form \begin{equation} \label{heavy-tailed m=n} \sqrt{N} - t \sqrt{n} \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{N} + t \sqrt{n} \end{equation} with probability at least $1 - 2n \cdot \exp(-c't^2)$. This is a form of our desired inequality \eqref{heuristic} for heavy-tailed matrices. We shall discuss this more after the proof. \begin{proof} We shall use the non-commutative Bernstein's inequality, Theorem~\ref{matrix Bernstein}. {\bf Step 1: Reduction to a sum of independent random matrices.} We first note that $m \ge n \ge 1$ since by Lemma~\ref{norm isotropic} we have $\mathbb{E} \|A_i\|_2^2 = n$. Now we start an argument parallel to Step~1 of Theorem~\ref{sub-gaussian rows}. Recalling Lemma~\ref{approximate isometries} for the matrix $B=A/\sqrt{N}$ we see that the desired inequalities \eqref{eq heavy-tailed rows} are equivalent to \begin{equation} \label{A*A heavy-tailed} \big\| \frac{1}{N}A^*A-I \big\| \le \max(\delta, \delta^2) =:\varepsilon \quad \text{where} \quad \delta = t \sqrt{\frac{m}{N}}. \end{equation} We express this random matrix as a sum of independent random matrices: $$ \frac{1}{N} A^*A - I = \frac{1}{N} \sum_{i=1}^N A_i \otimes A_i - I = \sum_{i=1}^N X_i, \quad \text{where } X_i := \frac{1}{N} (A_i \otimes A_i - I); $$ note that $X_i$ are independent centered $n \times n$ random matrices. {\bf Step 2: Estimating the mean, range and variance.} We are going to apply the non-commutative Bernstein inequality, Theorem~\ref{matrix Bernstein}, for the sum $\sum_i X_i$. Since $A_i$ are isotropic random vectors, we have $\mathbb{E} A_i \otimes A_i = I$ which implies that $\mathbb{E} X_i = 0$ as required in the non-commutative Bernstein inequality. We estimate the range of $X_i$ using that $\|A_i\|_2 \le \sqrt{m}$ and $m \ge 1$: $$ \|X_i\| \le \frac{1}{N} ( \|A_i \otimes A_i\| + 1) = \frac{1}{N} (\|A_i\|_2^2 + 1) \le \frac{1}{N} (m + 1) \le \frac{2 m}{N} =: K $$ To estimate the total variance $\|\sum_i \mathbb{E} X_i^2\|$, we first compute $$ X_i^2 = \frac{1}{N^2} \big[ (A_i \otimes A_i)^2 - 2(A_i \otimes A_i) + I \big], $$ so using that the isotropy assumption $\mathbb{E} A_i \otimes A_i = I$ we obtain \begin{equation} \label{Xi squared} \mathbb{E} X_i^2 = \frac{1}{N^2} \big[ \mathbb{E} (A_i \otimes A_i)^2 - I \big]. \end{equation} Since $(A_i \otimes A_i)^2 = \|A_i\|_2^2 \, A_i \otimes A_i$ is a positive semi-definite matrix and $\|A_i\|_2^2 \le m$ by assumption, we have $\big\| \mathbb{E} (A_i \otimes A_i)^2 \big\| \le m \cdot \| \mathbb{E} A_i \otimes A_i \| = m$. Putting this into \eqref{Xi squared} we obtain $$ \| \mathbb{E} X_i^2 \| \le \frac{1}{N^2} (m + 1) \le \frac{2 m}{N^2} $$ where we again used that $m \ge 1$. This yields\footnote{Here the seemingly crude application of triangle inequality is actually not so loose. If the rows $A_i$ are identically distributed, then so are $X_i^2$, which makes the triangle inequality above into an equality.} $$ \Big\| \sum_{i=1}^N \mathbb{E} X_i^2 \Big\| \le N \cdot \max_i \| \mathbb{E} X_i^2 \| = \frac{2m}{N} =: \sigma^2. $$ {\bf Step 3: Application of the non-commutative Bernstein's inequality.} \index{Bernstein-type inequality!non-commutative} Applying Theorem~\ref{matrix Bernstein} (see Remark~\ref{mixed tail}) and recalling the definitions of $\varepsilon$ and $\delta$ in \eqref{A*A heavy-tailed}, we we bound the probability in question as \begin{align*} \mathbb{P} &\Big\{ \Big\| \frac{1}{N} A^*A - I \Big\| \ge \varepsilon \Big\} = \mathbb{P} \Big\{ \Big\| \sum_{i=1}^N X_i \Big\| \ge \varepsilon \Big\} \le 2n \cdot \exp \Big[ -c \min \Big( \frac{\varepsilon^2}{\sigma^2}, \frac{\varepsilon}{K} \Big) \Big] \\ &\le 2n \cdot \exp \Big[ -c \min(\varepsilon^2,\varepsilon) \cdot \frac{N}{2m} \Big] = 2n \cdot \exp \Big( - \frac{c \delta^2 N}{2m} \Big) = 2n \cdot \exp(-ct^2/2). \end{align*} This completes the proof. \end{proof} Theorem~\ref{heavy-tailed rows} for heavy-tailed rows is different from Theorem~\ref{sub-gaussian rows} for sub-gaussian rows in two ways: the boundedness assumption\footnote{Going a little ahead, we would like to point out that the almost sure boundedness can be relaxed to the bound in expectation $\mathbb{E} \max_i \|A_i\|_2^2 \le m$, see Theorem~\ref{heavy-tailed rows exp si}.} $\|A_i\|_2^2 \le m$ appears, and the probability bound is weaker. We will now comment on both differences. \begin{remark}[Boundedness assumption] \label{r: boundedness} Observe that some boundendess assumption on the distribution is needed in Theorem~\ref{heavy-tailed rows}. Let us see this on the following example. Choose $\delta \ge 0$ arbitrarily small, and consider a random vector $X = \delta^{-1/2} \xi Y$ in $\mathbb{R}^n$ where $\xi$ is a $\{0,1\}$-valued random variable with $\mathbb{E} \xi = \delta$ (a ``selector'') and $Y$ is an independent isotropic random vector in $\mathbb{R}^n$ with an arbitrary distribution. Then $X$ is also an isotropic random vector. Consider an $N \times n$ random matrix $A$ whose rows $A_i$ are independent copies of $X$. However, if $\delta \ge 0$ is suitably small then $A = 0$ with high probability, hence no nontrivial lower bound on $s_{\min}(A)$ is possible. \end{remark} Inequality \eqref{heavy-tailed m=n} fits our goal \eqref{heuristic}, but not quite. The reason is that the probability bound is only non-trivial if $t \ge C \sqrt{\log n}$. Therefore, in reality Theorem~\ref{heavy-tailed rows} asserts that \begin{equation} \label{goal log} \sqrt{N} - C\sqrt{n \log n} \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{N} + C\sqrt{n \log n} \end{equation} with probability, say $0.9$. This achieves our goal \eqref{heuristic} up to a logarithmic factor. \begin{remark}[Logarithmic factor] The logarithmic factor can not be removed from \eqref{goal log} for some heavy-tailed distributions. Consider for instance the coordinate distribution introduced in Example~\ref{random vectors}. In order that $s_{\min}(A) > 0$ there must be no zero columns in $A$. Equivalently, each coordinate vector $e_1, \ldots,e_n$ \index{Coordinate random vectors} must be picked at least once in $N$ independent trials (each row of $A$ picks an independent coordinate vector). Recalling the classical coupon collector's problem, one must make at least $N \ge C n \log n$ trials to make this occur with high probability. Thus the logarithm is necessary in the left hand side of \eqref{goal log}.\footnote{This argument moreover shows the optimality of the probability bound in Theorem~\ref{heavy-tailed rows}. For example, for $t = \sqrt{N}/2\sqrt{n}$ the conclusion \eqref{heavy-tailed m=n} implies that $A$ is well conditioned (i.e. $\sqrt{N}/2 \le s_{\min}(A) \le s_{\max}(A) \le 2 \sqrt{N}$) with probability $1 - n \cdot \exp(-cN/n)$. On the other hand, by the coupon collector's problem we estimate the probability that $s_{\min}(A) > 0$ as $1 - n \cdot (1- \frac{1}{n})^N \approx 1 - n \cdot \exp(-N/n)$.} \end{remark} A version of Theorem~\ref{heavy-tailed rows} holds for general, non-isotropic distributions. It is convenient to state it in terms of the equivalent estimate \eqref{A*A heavy-tailed}: \begin{theorem}[Heavy-tailed rows, non-isotropic] \label{heavy-tailed rows non-isotropic} Let $A$ be an $N \times n$ matrix whose rows $A_i$ are independent random vectors in $\mathbb{R}^n$ with the common second moment matrix $\Sigma = \mathbb{E} A_i \otimes A_i$. Let $m$ be a number such that $\|A_i\|_2 \le \sqrt{m}$ almost surely for all $i$. Then for every $t \ge 0$, the following inequality holds with probability at least $1 - n \cdot \exp(-ct^2)$: \begin{equation} \label{A*A heavy-tailed rows non-isotropic} \big\| \frac{1}{N}A^*A-\Sigma \big\| \le \max(\|\Sigma\|^{1/2}\delta, \delta^2) \quad \text{where} \quad \delta = t \sqrt{\frac{m}{N}}. \end{equation} Here $c>0$ is an absolute constant. In particular, this inequality yields \begin{equation} \label{A heavy-tailed rows non-isotropic} \|A\| \le \|\Sigma\|^{1/2} \sqrt{N} + t \sqrt{m}. \end{equation} \end{theorem} \begin{proof} We note that $m \ge \|\Sigma\|$ because $\|\Sigma\| = \|\mathbb{E} A_i \otimes A_i\| \le \mathbb{E} \|A_i \otimes A_i\| = \mathbb{E} \|A_i\|_2^2 \le m$. Then \eqref{A*A heavy-tailed rows non-isotropic} follows by a straightforward modification of the argument of Theorem~\ref{heavy-tailed rows}. Furthermore, if \eqref{A*A heavy-tailed rows non-isotropic} holds then by triangle inequality \begin{align*} \frac{1}{N} \|A\|^2 &= \big\| \frac{1}{N} A^*A \big\| \le \|\Sigma\| + \big\| \frac{1}{N}A^*A-\Sigma \big\| \\ &\le \|\Sigma\| + \|\Sigma\|^{1/2}\delta + \delta^2 \le (\|\Sigma\|^{1/2} + \delta)^2. \end{align*} Taking square roots and multiplying both sides by $\sqrt{N}$, we obtain \eqref{A heavy-tailed rows non-isotropic}. \end{proof} \bigskip The {\em almost sure} boundedness requirement in Theorem~\ref{heavy-tailed rows} may sometimes be too restrictive in applications, and it can be relaxed to a bound {\em in expectation}: \begin{theorem}[Heavy-tailed rows; expected singular values] \label{heavy-tailed rows exp si} Let $A$ be an $N \times n$ matrix whose rows $A_i$ are independent isotropic random vectors in $\mathbb{R}^n$. Let $m := \mathbb{E} \max_{i \le N} \|A_i\|_2^2$. Then $$ \mathbb{E} \max_{j \le n} |s_j(A) - \sqrt{N}| \le C \sqrt{m \log \min(N,n)} $$ where $C$ is an absolute constant. \end{theorem} The proof of this result is similar to that of Theorem~\ref{heavy-tailed rows}, except that this time we will use Rudelson's Corollary~\ref{Rudelson} instead of matrix Bernstein's inequality. To this end, we need a link to symmetric Bernoulli random variables. This is provided by a general {\em symmetrization argument}: \begin{lemma}[Symmetrization] \index{Symmetrization} \label{symmetrization} Let $(X_i)$ be a finite sequence of independent random vectors valued in some Banach space, and $(\varepsilon_i)$ be independent symmetric Bernoulli random variables. Then \begin{equation} \label{eq symmetrization} \mathbb{E} \Big\| \sum_i (X_i - \mathbb{E} X_i) \Big\| \le 2 \mathbb{E} \Big\| \sum_i \varepsilon_i X_i \Big\|. \end{equation} \end{lemma} \begin{proof} We define random variables $\tilde{X}_i = X_i - X_i'$ where $(X_i')$ is an independent copy of the sequence $(X_i)$. Then $\tilde{X}_i$ are independent symmetric random variables, i.e. the sequence $(\tilde{X}_i)$ is distributed identically with $(-\tilde{X}_i)$ and thus also with $(\varepsilon_i \tilde{X}_i)$. Replacing $\mathbb{E} X_i$ by $\mathbb{E} X_i'$ in \eqref{eq symmetrization} and using Jensen's inequality, symmetry, and triangle inequality, we obtain the required inequality \begin{align*} \mathbb{E} \Big\| \sum_i (X_i - \mathbb{E} X_i) \Big\| &\le \mathbb{E} \Big\| \sum_i \tilde{X}_i \Big\| = \mathbb{E} \Big\| \sum_i \varepsilon_i \tilde{X}_i \Big\| \\ &\le \mathbb{E} \Big\| \sum_i \varepsilon_i X_i \Big\| + \mathbb{E} \Big\| \sum_i \varepsilon_i X_i' \Big\| = 2 \mathbb{E} \Big\| \sum_i \varepsilon_i X_i \Big\|. \qedhere \end{align*} \end{proof} We will also need a probabilistic version of Lemma~\ref{approximate isometries} on approximate isometries. The proof of that lemma was based on the elementary inequality $|z^2-1| \ge \max( |z-1|, |z-1|^2 )$ for $z \ge 0$. Here is a probabilistic version: \begin{lemma} \label{deviation from 1} Let $Z$ be a non-negative random variable. Then $\mathbb{E}|Z^2-1| \ge \max( \mathbb{E}|Z-1|, (\mathbb{E}|Z-1|)^2 )$. \end{lemma} \begin{proof} Since $|Z-1| \le |Z^2-1|$ pointwise, we have $\mathbb{E} |Z-1| \le \mathbb{E} |Z^2-1|$. Next, since $|Z-1|^2 \le |Z^2-1|$ pointwise, taking square roots and expectations we obtain $\mathbb{E}|Z-1| \le \mathbb{E}|Z^2-1|^{1/2} \le (\mathbb{E}|Z^2-1|)^{1/2}$, where the last bound follows by Jensen's inequality. Squaring both sides completes the proof. \end{proof} \begin{proof}[Proof of Theorem~\ref{heavy-tailed rows exp si}] {\bf Step 1: Application of Rudelson's inequality.} \index{Rudelson's inequality} As in the proof of Theorem~\ref{heavy-tailed rows}, we are going to control $$ E := \mathbb{E} \big\| \frac{1}{N} A^*A - I \big\| = \mathbb{E} \Big\| \frac{1}{N} \sum_{i=1}^N A_i \otimes A_i - I \Big\| \le \frac{2}{N} \, \mathbb{E} \Big\| \sum_{i=1}^N \varepsilon_i A_i \otimes A_i \Big\| $$ where we used Symmetrization Lemma~\ref{symmetrization} with independent symmetric Bernoulli random variables $\varepsilon_i$ (which are independent of $A$ as well). The expectation in the right hand side is taken both with respect to the random matrix $A$ and the signs $(\varepsilon_i)$. Taking first the expectation with respect to $(\varepsilon_i)$ (conditionally on $A$) and afterwards the expectation with respect to $A$, we obtain by Rudelson's inequality (Corollary~\ref{Rudelson}) that $$ E \le \frac{C \sqrt{l}}{N} \, \mathbb{E} \Big( \max_{i \le N} \|A_i\|_2 \cdot \Big\| \sum_{i=1}^N A_i \otimes A_i \Big\|^{1/2} \Big) $$ where $l = \log \min(N,n)$. We now apply the Cauchy-Schwarz inequality. Since by the triangle inequality $\mathbb{E} \big\| \frac{1}{N} \sum_{i=1}^N A_i \otimes A_i \big\| = \mathbb{E} \big\| \frac{1}{N} A^*A \big\| \le E + 1$, it follows that $$ E \le C \sqrt{\frac{ml}{N}} (E+1)^{1/2}. $$ This inequality is easy to solve in $E$. Indeed, considering the cases $E \le 1$ and $E > 1$ separately, we conclude that $$ E = \mathbb{E} \big\| \frac{1}{N} A^*A - I \big\| \le \max(\delta, \delta^2) \quad \text{where } \delta := C \sqrt{\frac{2ml}{N}}. $$ {\bf Step 2: Diagonalization.} Diagonalizing the matrix $A^*A$ one checks that $$ \big\| \frac{1}{N} A^*A - I \big\| = \max_{j \le n} \big| \frac{s_j(A)^2}{N} - 1 \big| = \max \Big( \big| \frac{s_{\min}(A)^2}{N} - 1 \big|, \big| \frac{s_{\max}(A)^2}{N} - 1 \big| \Big). $$ It follows that $$ \max \Big( \mathbb{E} \big| \frac{s_{\min}(A)^2}{N} - 1 \big|, \mathbb{E} \big| \frac{s_{\max}(A)^2}{N} - 1 \big| \Big) \le \max(\delta,\delta^2). $$ (we replaced the expectation of maximum by the maximum of expectations). Using Lemma~\ref{deviation from 1} separately for the two terms on the left hand side, we obtain $$ \max \Big( \mathbb{E} \big| \frac{s_{\min}(A)}{\sqrt{N}} - 1 \big|, \mathbb{E} \big| \frac{s_{\max}(A)}{\sqrt{N}} - 1 \big| \Big) \le \delta. $$ Therefore \begin{align*} \mathbb{E} \max_{j \le n} \big| \frac{s_j(A)}{\sqrt{N}}-1 \big| &= \mathbb{E} \max \Big( \big| \frac{s_{\min}(A)}{\sqrt{N}} - 1 \big|, \Big| \frac{s_{\max}(A)}{\sqrt{N}} - 1 \Big| \Big) \\ &\le \mathbb{E} \Big( \big| \frac{s_{\min}(A)}{\sqrt{N}} - 1 \big| + \big| \frac{s_{\max}(A)}{\sqrt{N}} - 1 \big| \Big) \le 2\delta. \end{align*} Multiplying both sides by $\sqrt{N}$ completes the proof. \end{proof} In a way similar to Theorem~\ref{heavy-tailed rows non-isotropic} we note that a version of Theorem~\ref{heavy-tailed rows exp si} holds for general, non-isotropic distributions. \begin{theorem}[Heavy-tailed rows, non-isotropic, expectation] \label{heavy-tailed rows exp si non-isotropic} Let $A$ be an $N \times n$ matrix whose rows $A_i$ are independent random vectors in $\mathbb{R}^n$ with the common second moment matrix $\Sigma = \mathbb{E} A_i \otimes A_i$. Let $m := \mathbb{E} \max_{i \le N} \|A_i\|_2^2$. Then $$ \mathbb{E} \big\| \frac{1}{N}A^*A-\Sigma \big\| \le \max(\|\Sigma\|^{1/2}\delta, \delta^2) \quad \text{where} \quad \delta = C \sqrt{\frac{m \log \min(N,n)}{N}}. $$ Here $C$ is an absolute constant. In particular, this inequality yields $$ \big( \mathbb{E} \|A\|^2 \big)^{1/2} \le \|\Sigma\|^{1/2} \sqrt{N} + C \sqrt{m \log \min(N,n)}. $$ \end{theorem} \begin{proof} The first part follows by a simple modification of the proof of Theorem~\ref{heavy-tailed rows exp si}. The second part follows from the first like in Theorem~\ref{heavy-tailed rows non-isotropic}. \end{proof} \begin{remark}[Non-identical second moments] \label{r: different second moments} The assumption that the rows $A_i$ have a common second moment matrix $\Sigma$ is not essential in Theorems~\ref{heavy-tailed rows non-isotropic} and \ref{heavy-tailed rows exp si non-isotropic}. The reader will be able to formulate more general versions of these results. For example, if $A_i$ have arbitrary second moment matrices $\Sigma_i = \mathbb{E} A_i \otimes A_i$ then the conclusion of Theorem~\ref{heavy-tailed rows exp si non-isotropic} holds with $\Sigma = \frac{1}{N} \sum_{i=1}^N \Sigma_i$. \end{remark} \subsection{Applications to estimating covariance matrices} \index{Covariance matrix!estimation}\label{s: covariance} One immediate application of our analysis of random matrices is in statistics, for the fundamental problem of {\em estimating covariance matrices}. Let $X$ be a random vector in $\mathbb{R}^n$; for simplicity we assume that $X$ is centered,\footnote{More generally, in this section we estimate the {\em second moment matrix} $\mathbb{E} X \otimes X$ of an arbitrary random vector $X$ (not necessarily centered).} $\mathbb{E} X = 0$. Recall that the covariance matrix of $X$ is the $n \times n$ matrix $\Sigma = \mathbb{E} X \otimes X$, see Section~\ref{s: isotropic}. The simplest way to estimate $\Sigma$ is to take some $N$ independent samples $X_i$ from the distribution and form the {\em sample covariance matrix} \index{Sample covariance matrix} $\Sigma_N = \frac{1}{N} \sum_{i=1}^N X_i \otimes X_i$. By the law of large numbers, $\Sigma_N \to \Sigma$ almost surely as $N \to \infty$. So, taking sufficiently many samples we are guaranteed to estimate the covariance matrix as well as we want. This, however, does not address the quantitative aspect: what is the minimal {\em sample size} $N$ that guarantees approximation with a given accuracy? The relation of this question to random matrix theory becomes clear when we arrange the samples $X_i =: A_i$ as rows of the $N \times n$ random matrix $A$. Then the sample covariance matrix is expressed as $\Sigma_N = \frac{1}{N}A^*A$. Note that $A$ is a matrix with independent rows but usually not independent entries (unless we sample from a product distribution). We worked out the analysis of such matrices in Section~\ref{s: rows}, separately for sub-gaussian and general distributions. As an immediate consequence of Theorem~\ref{sub-gaussian rows}, we obtain: \begin{corollary}[Covariance estimation for sub-gaussian distributions] \label{covariance sub-gaussian} \hfill Consider a sub-gaussian distribution in $\mathbb{R}^n$ with covariance matrix $\Sigma$, and let $\varepsilon \in (0,1)$, $t \ge 1$. Then with probability at least $1 - 2 \exp(- t^2 n)$ one has $$ \text{If } N \ge C(t/\varepsilon)^2 n \quad \text{then } \|\Sigma_N - \Sigma\| \le \varepsilon. $$ Here $C = C_K$ depends only on the sub-gaussian norm $K = \|X\|_{\psi_2}$ of a random vector taken from this distribution. \end{corollary} \begin{proof} It follows from \eqref{A*A rows non-isotropic} that for every $s \ge 0$, with probability at least $1 - 2\exp(-cs^2)$ we have $\|\Sigma_N-\Sigma\| \le \max(\delta, \delta^2)$ where $\delta = C \sqrt{n/N} + s/\sqrt{N}$. The conclusion follows for $s = C' t \sqrt{n}$ where $C' = C'_K$ is sufficiently large. \end{proof} Summarizing, Corollary~\ref{covariance sub-gaussian} shows that the sample size $$ N = O(n) $$ suffices to approximate the covariance matrix of a sub-gaussian distribution in $\mathbb{R}^n$ by the sample covariance matrix. \begin{remark}[Multiplicative estimates, Gaussian distributions] A weak point of Corollary~\ref{covariance sub-gaussian} is that the sub-gaussian norm $K$ may in turn depend on $\|\Sigma\|$. To overcome this drawback, instead of using \eqref{A*A rows non-isotropic} in the proof of this result one can use the multiplicative version \eqref{A*A rows non-isotropic multiplicative}. The reader is encouraged to state a general result that follows from this argument. We just give one special example for arbitrary {\em centered Gaussian distributions} in $\mathbb{R}^n$. For every $\varepsilon \in (0,1)$, $t \ge 1$, the following holds with probability at least $1 - 2 \exp(- t^2 n)$: $$ \text{If } N \ge C(t/\varepsilon)^2 n \quad \text{then } \|\Sigma_N - \Sigma\| \le \varepsilon \|\Sigma\|. $$ Here $C$ is an absolute constant. \end{remark} Finally, Theorem~\ref{heavy-tailed rows non-isotropic} yields a similar estimation result for arbitrary distributions, possibly heavy-tailed: \begin{corollary}[Covariance estimation for arbitrary distributions] \label{covariance heavy-tailed} \hfill Consider a distribution in $\mathbb{R}^n$ with covariance matrix $\Sigma$ and supported in some centered Euclidean ball whose radius we denote $\sqrt{m}$. Let $\varepsilon \in (0,1)$ and $t \ge 1$. Then the following holds with probability at least $1 - n^{-t^2}$: $$ \text{If } N \ge C(t/\varepsilon)^2 \|\Sigma\|^{-1} m \log n \quad \text{then } \|\Sigma_N - \Sigma\| \le \varepsilon \|\Sigma\|. $$ Here $C$ is an absolute constant. \end{corollary} \begin{proof} It follows from Theorem~\ref{heavy-tailed rows non-isotropic} that for every $s \ge 0$, with probability at least $1 - n \cdot \exp(-cs^2)$ we have $\|\Sigma_N-\Sigma\| \le \max(\|\Sigma\|^{1/2}\delta, \delta^2)$ where $\delta = s \sqrt{m/N}$. Therefore, if $N \ge (s/\varepsilon)^2 \|\Sigma\|^{-1} m$ then $\|\Sigma_N - \Sigma\| \le \varepsilon \|\Sigma\|$. The conclusion follows with $s = C' t \sqrt{\log n}$ where $C'$ is a sufficiently large absolute constant. \end{proof} Corollary~\ref{covariance heavy-tailed} is typically used with $m = O(\|\Sigma\| n)$. Indeed, if $X$ is a random vector chosen from the distribution in question, then its expected norm is easy to estimate: $\mathbb{E} \|X\|_2^2 = \tr(\Sigma) \le n \|\Sigma\|$. So, by Markov's inequality, most of the distribution is supported in a centered ball of radius $\sqrt{m}$ where $m = O(n \|\Sigma\|)$. If all distribution is supported there, i.e. if $\|X\| = O(\sqrt{n \|\Sigma\|})$ almost surely, then the conclusion of Corollary~\ref{covariance heavy-tailed} holds with sample size $N \ge C(t/\varepsilon)^2 n \log n$. \begin{remark}[Low-rank estimation] In certain applications, the distribution in $\mathbb{R}^n$ lies close to a low dimensional subspace. In this case, a smaller sample suffices for covariance estimation. The intrinsic dimension of the distribution can be measured with the {\em effective rank} \index{Effective rank} of the matrix $\Sigma$, defined as $$ r(\Sigma) = \frac{\tr(\Sigma)}{\|\Sigma\|}. $$ One always has $r(\Sigma) \le \rank(\Sigma) \le n$, and this bound is sharp. For example, if $X$ is an isotropic random vector in $\mathbb{R}^n$ then $\Sigma = I$ and $r(\Sigma) = n$. A more interesting example is where $X$ takes values in some $r$-dimensional subspace $E$, and the restriction of the distribution of $X$ onto $E$ is isotropic. The latter means that $\Sigma = P_E$, where $P_E$ denotes the orthogonal projection in $\mathbb{R}^n$ onto $E$. Therefore in this case $r(\Sigma) = r$. The effective rank is a stable quantity compared with the usual rank. For distributions that are approximately low-dimenional, the effective rank is still small. The effective rank $r = r(\Sigma)$ always controls the typical norm of $X$, as $\mathbb{E} \|X\|_2^2 = \tr(\Sigma) = r \|\Sigma\|$. It follows by Markov's inequality that most of the distribution is supported in a ball of radius $\sqrt{m}$ where $m = O(r \|\Sigma\|)$. Assume that all of the distribution is supported there, i.e. if $\|X\| = O(\sqrt{r \|\Sigma\|})$ almost surely. Then the conclusion of Corollary~\ref{covariance heavy-tailed} holds with sample size $N \ge C(t/\varepsilon)^2 r \log n$. \end{remark} We can summarize this discussion in the following way: the sample size $$ N = O(n \log n) $$ suffices to approximate the covariance matrix of a general distribution in $\mathbb{R}^n$ by the sample covariance matrix. Furthermore, for distributions that are approximately low-dimensional, a smaller sample size is sufficient. Namely, if the effective rank of $\Sigma$ equals $r$ then a sufficient sample size is $$ N = O(r \log n). $$ \begin{remark}[Boundedness assumption] \label{r: covariance boundedness} Without the boundedness assumption on the distribution, Corollary~\ref{covariance heavy-tailed} may fail. The reasoning is the same as in Remark~\ref{r: boundedness}: for an isotropic distribution which is highly concentrated at the origin, the sample covariance matrix will likely equal $0$. Still, one can weaken the boundedness assumption using Theorem~\ref{heavy-tailed rows exp si non-isotropic} instead of Theorem~\ref{heavy-tailed rows non-isotropic} in the proof of Corollary~\ref{covariance heavy-tailed}. The weaker requirement is that $\mathbb{E} \max_{i \le N} \|X_i\|_2^2 \le m$ where $X_i$ denote the sample points. In this case, the covariance estimation will be guaranteed in expectation rather than with high probability; we leave the details for the interested reader. A different way to enforce the boundedness assumption is to reject any sample points $X_i$ that fall outside the centered ball of radius $\sqrt{m}$. This is equivalent to sampling from the conditional distribution inside the ball. The conditional distribution satisfies the boundedness requirement, so the results discussed above provide a good covariance estimation for it. In many cases, this estimate works even for the original distribution -- namely, if only a small part of the distribution lies outside the ball of radius $\sqrt{m}$. We leave the details for the interested reader; see e.g. \cite{V marginals}. \end{remark} \subsection{Applications to random sub-matrices and sub-frames} \index{Sampling from matrices and frames} \label{s: sub-matrices} The absence of any moment hypotheses on the distribution in Section~\ref{s: heavy-tailed rows} (except finite variance) makes these results especially relevant for discrete distributions. One such situation arises when one wishes to sample entries or rows from a given matrix $B$, thereby creating a {\em random sub-matrix} $A$. It is a big program to understand what we can learn about $B$ by seeing $A$, see \cite{GMDL, DKM, RV sampling}. In other words, we ask -- what properties of $B$ pass onto $A$? Here we shall only scratch the surface of this problem: we notice that random sub-matrices of certain size preserve the property of being an {\em approximate isometry}. \begin{corollary}[Random sub-matrices] \index{Sub-matrices} \label{random sub-matrices} Consider an $M \times n$ matrix $B$ such that\footnote{The first hypothesis says $B^*B = MI$. Equivalently, $\bar{B} := \frac{1}{\sqrt{M}}B$ is an isometry, i.e. $\|\bar{B}x\|_2 = \|x\|_2$ for all $x$. Equivalently, the columns of $\bar{B}$ are orthonormal.} $s_{\min}(B) = s_{\max}(B) = \sqrt{M}$. Let $m$ be such that all rows $B_i$ of $B$ satisfy $\|B_i\|_2 \le \sqrt{m}$. Let $A$ be an $N \times n$ matrix obtained by sampling $N$ random rows from $B$ uniformly and independently. Then for every $t \ge 0$, with probability at least $1 - 2n \cdot \exp(-ct^2)$ one has $$ \sqrt{N} - t \sqrt{m} \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{N} + t \sqrt{m}. $$ Here $c>0$ is an absolute constant. \end{corollary} \begin{proof} By assumption, $I = \frac{1}{M} B^*B = \frac{1}{M} \sum_{i=1}^M B_i \otimes B_i$. Therefore, the uniform distribution on the set of the rows $\{B_1,\ldots,B_M\}$ is an isotropic distribution in $\mathbb{R}^n$. The conclusion then follows from Theorem~\ref{heavy-tailed rows}. \end{proof} Note that the conclusion of Corollary~\ref{random sub-matrices} does not depend on the dimension $M$ of the ambient matrix $B$. This happens because this result is a specific version of sampling from a discrete isotropic distribution (uniform on the rows of $B$), where size $M$ of the support of the distribution is irrelevant. The hypothesis of Corollary~\ref{random sub-matrices} implies\footnote{To recall why this is true, take trace of both sides in the identity $I = \frac{1}{M} \sum_{i=1}^M B_i \otimes B_i$.} that $\frac{1}{M} \sum_{i=1}^M \|B_i\|_2^2 = n$. Hence by Markov's inequality, most of the rows $B_i$ satisfy $\|B_i\|_2 = O(\sqrt{n})$. This indicates that Corollary~\ref{random sub-matrices} would be often used with $m = O(n)$. Also, to ensure a positive probability of success, the useful magnitude of $t$ would be $t \sim \sqrt{\log n}$. With this in mind, the extremal singular values of $A$ will be close to each other (and to $\sqrt{N}$) if $N \gg t^2 m \sim n \log n$. Summarizing, Corollary~\ref{random sub-matrices} states that a random $O(n \log n) \times n$ sub-matrix of an $M \times n$ isometry is an approximate isometry.\footnote{For the purposes of compressed sensing, we shall study the more difficult {\em uniform} problem for random sub-matrices in Section~\ref{s: restricted isometries}. There $B$ itself will be chosen as a column sub-matrix of a given $M \times M$ matrix (such as DFT), and one will need to control all such $B$ simultaneously, see Example~\ref{random measurements}.} \medskip Another application of random matrices with heavy-tailed isotropic rows is for {\em sampling from frames}. Recall that frames are generalizations of bases without linear independence, see Example~\ref{random vectors}. Consider a tight frame $\{u_i\}_{i=1}^M$ in $\mathbb{R}^n$, and for the sake of convenient normalization, assume that it has bounds $A=B=M$. We are interested in whether a small random subset of $\{u_i\}_{i=1}^M$ is still a nice frame in $\mathbb{R}^n$. Such question arises naturally because frames are used in signal processing to create {\em redundant representations} of signals. Indeed, every signal $x \in \mathbb{R}^n$ admits frame expansion $x = \frac{1}{M} \sum_{i=1}^M \< u_i, x\> u_i$. Redundancy makes frame representations more robust to errors and losses than basis representations. Indeed, we will show that if one loses all except $N = O(n \log n)$ random coefficients $\< u_i, x\> $ one is still able to reconstruct $x$ from the received coefficients $\< u_{i_k}, x\> $ as $x \approx \frac{1}{N} \sum_{k=1}^N \< u_{i_k}, x\> u_{i_k}$. This boils down to showing that a random subset of size $N = O(n \log n)$ of a tight frame in $\mathbb{R}^n$ is an approximate tight frame. \begin{corollary}[Random sub-frames, see \cite{V frames}] \index{Frames} \label{random sub-frames} Consider a tight frame $\{u_i\}_{i=1}^M$ in $\mathbb{R}^n$ with frame bounds $A=B=M$. Let number $m$ be such that all frame elements satisfy $\|u_i\|_2 \le \sqrt{m}$. Let $\{v_i\}_{i=1}^N$ be a set of vectors obtained by sampling $N$ random elements from the frame $\{u_i\}_{i=1}^M$ uniformly and independently. Let $\varepsilon \in (0,1)$ and $t \ge 1$. Then the following holds with probability at least $1 - 2n^{-t^2}$: $$ \text{If } N \ge C(t/\varepsilon)^2 m \log n \quad \text{then $\{v_i\}_{i=1}^N$ is a frame in $\mathbb{R}^n$} $$ with bounds $A = (1-\varepsilon)N$, $B = (1+\varepsilon)N$. Here $C$ is an absolute constant. In particular, if this event holds, then every $x \in \mathbb{R}^n$ admits an approximate representation using only the sampled frame elements: $$ \Big\| \frac{1}{N} \sum_{i=1}^N \< v_i, x\> v_i - x \Big\| \le \varepsilon \|x\|. $$ \end{corollary} \begin{proof} The assumption implies that $I = \frac{1}{M} \sum_{i=1}^M u_i \otimes u_i$. Therefore, the uniform distribution on the set $\{u_i\}_{i=1}^M$ is an isotropic distribution in $\mathbb{R}^n$. Applying Corollary~\ref{covariance heavy-tailed} with $\Sigma = I$ and $\Sigma_N = \frac{1}{N} \sum_{i=1}^N v_i \otimes v_i$ we conclude that $\|\Sigma_N - I\| \le \varepsilon$ with the required probability. This clearly completes the proof. \end{proof} As before, we note that $\frac{1}{M} \sum_{i=1}^M \|u_i\|_2^2 = n$, so Corollary~\ref{random sub-frames} would be often used with $m = O(n)$. This shows, liberally speaking, that a random subset of a frame in $\mathbb{R}^n$ of size $N = O(n \log n)$ is again a frame. \begin{remark}[Non-uniform sampling] The boundedness assumption $\|u_i\|_2 \le \sqrt{m}$, although needed in Corollary~\ref{random sub-frames}, can be removed by non-uniform sampling. To this end, one would sample from the set of normalized vectors $\bar{u}_i := \sqrt{n} \frac{u_i}{\|u_i\|_2}$ with probabilities proportional to $\|u_i\|_2^2$. This defines an isotropic distribution in $\mathbb{R}^n$, and clearly $\|\bar{u}_i\|_2 = \sqrt{n}$. Therefore, by Theorem~\ref{random sub-frames}, a random sample of $N = O(n \log n)$ vectors obtained this way forms an almost tight frame in $\mathbb{R}^n$. This result does not require any bound on $\|u_i\|_2$. \end{remark} \section{Random matrices with independent columns} \label{s: columns} In this section we study the extreme singular values of $N \times n$ random matrices $A$ with independent columns $A_j$. We are guided by our ideal bounds \eqref{heuristic} as before. The same phenomenon occurs in the column independent model as in the row independent model -- sufficiently tall random matrices $A$ are approximate isometries. As before, being tall will mean $N \gg n$ for sub-gaussian distributions and $N \gg n \log n$ for arbitrary distributions. The problem is equivalent to studying {\em Gram matrices} $G = A^*A = (\< A_j, A_k\> )_{j,k=1}^n$ \index{Gram matrix} of independent isotropic random vectors $A_1,\ldots, A_n$ in $\mathbb{R}^N$. Our results can be interpreted using Lemma~\ref{approximate isometries} as showing that the normalized Gram matrix $\frac{1}{N} G$ is an {\em approximate identity} for $N, n$ as above. Let us first try to prove this with a heuristic argument. By Lemma~\ref{norm isotropic} we know that the diagonal entries of $\frac{1}{N} G$ have mean $\frac{1}{N} \mathbb{E} \|A_j\|_2^2 = 1$ and off-diagonal ones have zero mean and standard deviation $\frac{1}{N} (\mathbb{E} \< A_j, A_k\> ^2)^{1/2} = \frac{1}{\sqrt{N}}$. If, hypothetically, the off-diagonal entries were independent, then we could use the results of matrices with independent entries (or even rows) developed in Section~\ref{s: rows}. The off-diagonal part of $\frac{1}{N}G$ would have norm $O(\sqrt{\frac{n}{N}})$ while the diagonal part would approximately equal $I$. Hence we would have \begin{equation} \label{Gram ideal} \big\| \frac{1}{N} G - I \big\| = O \Big( \sqrt{\frac{n}{N}} \Big), \end{equation} i.e. $\frac{1}{N} G$ is an approximate identity for $N \gg n$. Equivalently, by Lemma~\ref{approximate isometries}, \eqref{Gram ideal} would yield the ideal bounds \eqref{heuristic} on the extreme singular values of $A$. Unfortunately, the entries of the Gram matrix $G$ are obviously not independent. To overcome this obstacle we shall use the {\em decoupling} technique of probability theory \cite{dG}. We observe that there is still enough independence encoded in $G$. Consider a principal sub-matrix $(A_S)^*(A_T)$ of $G = A^*A$ with disjoint index sets $S$ and $T$. If we condition on $(A_k)_{k \in T}$ then this sub-matrix has independent rows. Using an elementary decoupling technique, we will indeed seek to replace the full Gram matrix $G$ by one such decoupled $S \times T$ matrix with independent rows, and finish off by applying results of Section~\ref{s: rows}. \medskip By transposition one can try to reduce our problem to studying the $n \times N$ matrix $A^*$. It has independent rows and the same singular values as $A$, so one can apply results of Section~\ref{s: rows}. The conclusion would be that, with high probability, $$ \sqrt{n} - C \sqrt{N} \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{n} + C \sqrt{N}. $$ Such estimate is only good for {\em flat} matrices ($N \le n$). For {\em tall} matrices ($N \ge n$) the lower bound would be trivial because of the (possibly large) constant $C$. So, from now on we can focus on tall matrices ($N \ge n$) with independent columns. \subsection{Sub-gaussian columns} \index{Sub-gaussian!random matrices with independent columns} \label{s: sub-gaussian columns} Here we prove a version of Theorem~\ref{sub-gaussian rows} for matrices with independent columns. \begin{theorem}[Sub-gaussian columns] \label{sub-gaussian columns} Let $A$ be an $N \times n$ matrix ($N \ge n$) whose columns $A_i$ are independent sub-gaussian isotropic random vectors in $\mathbb{R}^N$ with $\|A_j\|_2 = \sqrt{N}$ a. s. Then for every $t \ge 0$, the inequality holds \begin{equation} \label{smin smax columns} \sqrt{N} - C \sqrt{n} - t \le s_{\min}(A) \le s_{\max}(A) \le \sqrt{N} + C \sqrt{n} + t \end{equation} with probability at least $1 - 2\exp(-ct^2)$, where $C = C'_K$, $c = c'_K > 0$ depend only on the subgaussian norm $K = \max_j \|A_j\|_{\psi_2}$ of the columns. \end{theorem} The only significant difference between Theorem~\ref{sub-gaussian rows} for independent rows and Theorem~\ref{sub-gaussian columns} for independent columns is that the latter requires {\em normalization of columns}, $\|A_j\|_2 = \sqrt{N}$ almost surely. Recall that by isotropy of $A_j$ (see Lemma~\ref{norm isotropic}) one always has $(\mathbb{E}\|A_j\|_2^2)^{1/2} = \sqrt{N}$, but the normalization is a bit stronger requirement. We will discuss this more after the proof of Theorem~\ref{sub-gaussian columns}. \begin{remark}[Gram matrices are an approximate identity] By Lemma~\ref{approximate isometries}, the conclusion of Theorem~\ref{sub-gaussian columns} is equivalent to $$ \big\| \frac{1}{N} A^*A - I \| \le C \sqrt{\frac{n}{N}} + \frac{t}{\sqrt{N}} $$ with the same probability $1 - 2\exp(-ct^2)$. This establishes our ideal inequality \eqref{Gram ideal}. In words, the normalized Gram matrix of $n$ independent sub-gaussian isotropic random vectors in $\mathbb{R}^N$ is an approximate identity whenever $N \gg n$. \end{remark} The proof of Theorem~\ref{sub-gaussian columns} is based on the decoupling technique \cite{dG}. What we will need here is an elementary decoupling lemma for double arrays. Its statement involves the notion of a {\em random subset} \index{Random subset} of a given finite set. To be specific, we define a random set $T$ of $[n]$ with a given average size $m \in [0,n]$ as follows. Consider independent $\{0,1\}$ valued random variables $\delta_1,\ldots,\delta_n$ with $\mathbb{E}\delta_i = m/n$; these are sometimes called {\em independent selectors}. \index{Selectors} Then we define the random subset $T = \{ i \in [n]:\; \delta_i=1 \}$. Its average size equals $\mathbb{E}|T| = \mathbb{E} \sum_{i=1}^n \delta_i = m$. \begin{lemma}[Decoupling] \index{Decoupling} \label{decoupling} Consider a double array of real numbers $(a_{ij})_{i,j=1}^n$ such that $a_{ii} = 0$ for all $i$. Then $$ \sum_{i,j \in [n]} a_{ij} = 4 \mathbb{E} \sum_{i \in T,\, j \in T^c} a_{ij} $$ where $T$ is a random subset of $[n]$ with average size $n/2$. In particular, $$ 4 \min_{T \subseteq [n]} \sum_{i \in T,\, j \in T^c} a_{ij} \le \sum_{i,j \in [n]} a_{ij} \le 4 \max_{T \subseteq [n]} \sum_{i \in T,\, j \in T^c} a_{ij} $$ where the minimum and maximum are over all subsets $T$ of $[n]$. \end{lemma} \begin{proof} Expressing the random subset as $T = \{ i \in [n]:\; \delta_i=1 \}$ where $\delta_i$ are independent selectors with $\mathbb{E}\delta_i=1/2$, we see that $$ \mathbb{E} \sum_{i \in T,\, j \in T^c} a_{ij} = \mathbb{E} \sum_{i,j \in [n]} \delta_i (1-\delta_j) a_{ij} = \frac{1}{4} \sum_{i,j \in [n]} a_{ij}, $$ where we used that $\mathbb{E} \delta_i (1-\delta_j) = \frac{1}{4}$ for $i \ne j$ and the assumption $a_{ii}=0$. This proves the first part of the lemma. The second part follows trivially by estimating expectation by maximum and minimum. \end{proof} \begin{proof}[Proof of Theorem~\ref{sub-gaussian columns}] {\bf Step 1: Reductions.} Without loss of generality we can assume that the columns $A_i$ have zero mean. Indeed, multiplying each column $A_i$ by $\pm 1$ arbitrarily preserves the extreme singular values of $A$, the isotropy of $A_i$ and the sub-gaussian norms of $A_i$. Therefore, by multiplying $A_i$ by independent symmetric Bernoulli random variables we achieve that $A_i$ have zero mean. For $t = O(\sqrt{N})$ the conclusion of Theorem~\ref{sub-gaussian columns} follows from Theorem~\ref{sub-gaussian rows} by transposition. Indeed, the $n \times N$ random matrix $A^*$ has independent rows, so for $t \ge 0$ we have \begin{equation} \label{smax smax star} s_{\max}(A) = s_{\max}(A^*) \le \sqrt{n} + C_K \sqrt{N} + t \end{equation} with probability at least $1 - 2 \exp(-c_K t^2)$. Here $c_K > 0$ and we can obviously assume that $C_K \ge 1$. For $t \ge C_K\sqrt{N}$ it follows that $s_{\max}(A) \le \sqrt{N} + \sqrt{n} + 2t$, which yields the conclusion of Theorem~\ref{sub-gaussian columns} (the left hand side of \eqref{smin smax columns} being trivial). So, it suffices to prove the conclusion for $t \le C_K \sqrt{N}$. Let us fix such $t$. It would be useful to have some a priori control of $s_{\max}(A) = \|A\|$. We thus consider the desired event $$ \mathcal{E} := \big\{ s_{\max}(A) \le 3C_K \sqrt{N} \big\}. $$ Since $3C_K \sqrt{N} \ge \sqrt{n} + C_K \sqrt{N} + t$, by \eqref{smax smax star} we see that $\mathcal{E}$ is likely to occur: \begin{equation} \label{EEc} \mathbb{P}(\mathcal{E}^c) \le 2 \exp(-c_K t^2). \end{equation} {\bf Step 2: Approximation.} This step is parallel to Step~1 in the proof of Theorem~\ref{sub-gaussian rows}, except now we shall choose $\varepsilon := \delta$. This way we reduce our task to the following. Let $\mathcal{N}$ be a $\frac{1}{4}$-net of the unit sphere $S^{n-1}$ such that $|\mathcal{N}| \le 9^n$. It suffices to show that with probability at least $1 - 2\exp(-c'_K t^2)$ one has $$ \max_{x \in \mathcal{N}} \Big| \frac{1}{N} \|Ax\|_2^2 - 1 \Big| \le \frac{\delta}{2}, \quad \text{where } \delta = C \sqrt{\frac{n}{N}} + \frac{t}{\sqrt{N}}. $$ By \eqref{EEc}, it is enough to show that the probability \begin{equation} \label{desired p} p:= \mathbb{P} \Big\{ \max_{x \in \mathcal{N}} \Big| \frac{1}{N} \|Ax\|_2^2 - 1 \Big| > \frac{\delta}{2} \text{ and } \mathcal{E} \Big\} \end{equation} satisfies $p \le 2\exp(-c_K'' t^2)$, where $c_K''>0$ may depend only on $K$. {\bf Step 3: Decoupling.} As in the proof of Theorem~\ref{sub-gaussian rows}, we will obtain the required bound for a fixed $x \in \mathcal{N}$ with high probability, and then take a union bound over $x$. So let us fix any $x = (x_1,\ldots,x_n) \in S^{n-1}$. We expand \begin{equation} \label{norm expansion} \|Ax\|_2^2 = \Big\| \sum_{j=1}^n x_j A_j \Big\|_2^2 = \sum_{j=1}^n x_j^2 \|A_j\|_2^2 + \sum_{j,k \in [n], \, j \ne k} x_j x_k \< A_j, A_k \> . \end{equation} Since $\|A_j\|_2^2 = N$ by assumption and $\|x\|_2 =1$, the first sum equals $N$. Therefore, subtracting $N$ from both sides and dividing by $N$, we obtain the bound $$ \Big| \frac{1}{N} \|Ax\|_2^2 - 1 \Big| \le \Big| \frac{1}{N} \sum_{j,k \in [n], \, j \ne k} x_j x_k \< A_j, A_k \> \Big| . $$ The sum in the right hand side is $\< G_0 x, x\> $ where $G_0$ is the off-diagonal part of the Gram matrix $G = A^*A$. As we indicated in the beginning of Section~\ref{s: columns}, we are going to replace $G_0$ by its decoupled version whose rows and columns are indexed by disjoint sets. This is achieved by Decoupling Lemma~\ref{decoupling}: we obtain $$ \Big| \frac{1}{N} \|Ax\|_2^2 - 1 \Big| \le \frac{4}{N} \max_{T \subseteq [n]} |R_T(x)|, \quad \text{where } R_T(x) = \sum_{j \in T, \, k \in T^c} x_j x_k \< A_j, A_k \> . $$ We substitute this into \eqref{desired p} and take union bound over all choices of $x \in \mathcal{N}$ and $T \subseteq [n]$. As we know, $|\mathcal{N}| \le 9^n$, and there are $2^n$ subsets $T$ in $[n]$. This gives \begin{align} \label{p} p &\le \mathbb{P} \Big\{ \max_{x \in \mathcal{N}, \, T \subseteq [n]} |R_T(x)| > \frac{\delta N}{8} \text{ and } \mathcal{E} \Big\} \notag\\ &\le 9^n \cdot 2^n \cdot \max_{x \in \mathcal{N}, \, T \subseteq [n]} \mathbb{P} \Big\{ |R_T(x)| > \frac{\delta N}{8} \text{ and } \mathcal{E} \Big\}. \end{align} {\bf Step 4: Conditioning and concentration.} To estimate the probability in \eqref{p}, we fix a vector $x \in \mathcal{N}$ and a subset $T \subseteq [n]$ and we condition on a realization of random vectors $(A_k)_{k \in T^c}$. We express \begin{equation} \label{RJ equivalent} R_T(x) = \sum_{j \in T} x_j \langle A_j, z \rangle \quad \text{where } z = \sum_{k \in T^c} x_k A_k. \end{equation} Under our conditioning $z$ is a fixed vector, so $R_T(x)$ is a sum of independent random variables. Moreover, if event $\mathcal{E}$ holds then $z$ is nicely bounded: \begin{equation} \label{norm z} \|z\|_2 \le \|A\| \|x\|_2 \le 3 C_K \sqrt{N}. \end{equation} If in turn \eqref{norm z} holds then the terms $\< A_j, z\> $ in \eqref{RJ equivalent} are independent centered sub-gaussian random variables with $\|\< A_j, z\> \|_{\psi_2} \le 3 K C_K \sqrt{N}$. By Lemma~\ref{rotation invariance}, their linear combination $R_T(x)$ is also a sub-gaussian random variable with \begin{equation} \label{RT sub-gaussian} \|R_T(x)\|_{\psi_2} \le C_1 \Big( \sum_{j \in T} x_j^2 \|\< A_j, z\> \|_{\psi_2}^2 \Big)^{1/2} \le \widehat{C}_K \sqrt{N} \end{equation} where $\widehat{C}_K$ depends only on $K$. We can summarize these observations as follows. Denoting the conditional probability by $\mathbb{P}_T = \mathbb{P} \{ \; \cdot \; | (A_k)_{k \in T^c} \}$ and the expectation with respect to $(A_k)_{k \in T^c}$ by $\mathbb{E}_{T^c}$, we obtain by \eqref{norm z} and \eqref{RT sub-gaussian} that \begin{align*} \mathbb{P} &\Big\{ |R_T(x)| > \frac{\delta N}{8} \text{ and } \mathcal{E} \Big\} \le \mathbb{E}_{T^c}\mathbb{P}_T \Big\{ |R_T(x)| > \frac{\delta N}{8} \text{ and } \|z\|_2 \le 3 C_K \sqrt{N} \Big\} \\ &\le 2 \exp \Big[ -c_1 \Big( \frac{\delta N/8}{\widehat{C}_K \sqrt{N}} \Big)^2 \Big] = 2 \exp \Big( -\frac{c_2 \delta^2 N}{\widehat{C}_K^2} \Big) \le 2 \exp \Big( -\frac{c_2 C^2 n}{\widehat{C}_K^2} - \frac{c_2 t^2}{\widehat{C}_K^2} \Big). \end{align*} The second inequality follows because $R_T(x)$ is a sub-gaussian random variable \eqref{RT sub-gaussian} whose tail decay is given by \eqref{sub-gaussian tail}. Here $c_1,c_2>0$ are absolute constants. The last inequality follows from the definition of $\delta$. Substituting this into \eqref{p} and choosing $C$ sufficiently large (so that $\ln 36 \le c_2 C^2/\widehat{C}_K^2$), we conclude that $$ p \le 2 \exp \big( - c_2 t^2/\widehat{C}_K^2 \big). $$ This proves an estimate that we desired in Step 2. The proof is complete. \end{proof} \begin{remark}[Normalization assumption] Some a priori control of the norms of the columns $\|A_j\|_2$ is necessary for estimating the extreme singular values, since $$ s_{\min}(A) \le \min_{i \le n} \|A_j\|_2 \le \max_{i \le n} \|A_j\|_2 \le s_{\max}(A). $$ With this in mind, it is easy to construct an example showing that a normalization assumption $\|A_i\|_2 = \sqrt{N}$ is essential in Theorem~\ref{sub-gaussian columns}; it can not even be replaced by a boundedness assumption $\|A_i\|_2 = O(\sqrt{N})$. Indeed, consider a random vector $X = \sqrt{2} \xi Y$ in $\mathbb{R}^N$ where $\xi$ is a $\{0,1\}$-valued random variable with $\mathbb{E} \xi = 1/2$ (a ``selector'') and $X$ is an independent spherical random vector in $\mathbb{R}^n$ (see Example~\ref{random vectors sub-gaussian}). Let $A$ be a random matrix whose columns $A_j$ are independent copies of $X$. Then $A_j$ are independent centered sub-gaussian isotropic random vectors in $\mathbb{R}^n$ with $\|A_j\|_{\psi_2} = O(1)$ and $\|A_j\|_2 \le \sqrt{2N}$ a.s. So all assumptions of Theorem~\ref{sub-gaussian columns} except normalization are satisfied. On the other hand $\mathbb{P}\{X=0\}=1/2$, so matrix $A$ has a zero column with overwhelming probability $1 - 2^{-n}$. This implies that $s_{\min}(A)=0$ with this probability, so the lower estimate in \eqref{smin smax columns} is false for all nontrivial $N,n,t$. \end{remark} \subsection{Heavy-tailed columns} \index{Heavy-tailed!random matrices with independent columns} Here we prove a version of Theorem~\ref{heavy-tailed rows exp si} for independent heavy-tailed columns. We thus consider $N \times n$ random matrices $A$ with independent columns $A_j$. In addition to the normalization assumption $\|A_j\|_2 = \sqrt{N}$ already present in Theorem~\ref{sub-gaussian columns} for sub-gaussian columns, our new result must also require an a priori control of the off-diagonal part of the Gram matrix $G = A^*A = (\< A_j, A_k \> )_{j,k=1}^n$. \begin{theorem}[Heavy-tailed columns] \label{heavy-tailed columns} Let $A$ be an $N \times n$ matrix ($N \ge n$) whose columns $A_j$ are independent isotropic random vectors in $\mathbb{R}^N$ with $\|A_j\|_2 = \sqrt{N}$ a. s. Consider the incoherence parameter $$ m := \frac{1}{N} \mathbb{E} \max_{j \le n} \sum_{k \in [n], \, k \ne j} \< A_j, A_k \> ^2. $$ Then $\mathbb{E} \big\| \frac{1}{N} A^*A - I \big\| \le C_0 \sqrt{ \frac{m \log n}{N}}$. In particular, \begin{equation} \label{eq heavy-tailed columns} \mathbb{E} \max_{j \le n} |s_j(A) - \sqrt{N}| \le C \sqrt{m \log n}. \end{equation} \end{theorem} Let us briefly clarify the role of the incoherence parameter $m$, which controls the lengths of the rows of the off-diagonal part of $G$. After the proof we will see that a control of $m$ is essential in Theorem~\ref{heavy-tailed rows}. But for now, let us get a feel of the typical size of $m$. We have $\mathbb{E} \< A_j, A_k\> ^2 = N$ by Lemma~\ref{norm isotropic}, so for every row $j$ we see that $\frac{1}{N} \sum_{k \in [n], \, k \ne j} \< A_j, A_k \> ^2 = n-1$. This indicates that Theorem~\ref{heavy-tailed columns} would be often used with $m = O(n)$. In this case, Theorem~\ref{heavy-tailed rows} establishes our ideal inequality \eqref{Gram ideal} up to a logarithmic factor. In words, the normalized Gram matrix \index{Gram matrix} of $n$ independent isotropic random vectors in $\mathbb{R}^N$ is an approximate identity whenever $N \gg n \log n$. \medskip Our proof of Theorem~\ref{heavy-tailed columns} will be based on decoupling, symmetrization and an application of Theorem~\ref{heavy-tailed rows exp si non-isotropic} for a decoupled Gram matrix with independent rows. The decoupling is done similarly to Theorem~\ref{sub-gaussian columns}. However, this time we will benefit from formalizing the decoupling inequality for Gram matrices: \begin{lemma}[Matrix decoupling] \index{Decoupling} \label{matrix decoupling} Let $B$ be a $N \times n$ random matrix whose columns $B_j$ satisfy $\|B_j\|_2 = 1$. Then $$ \mathbb{E} \|B^*B-I\| \le 4 \max_{T \subseteq [n]} \mathbb{E} \|(B_T)^* B_{T^c} \|. $$ \end{lemma} \begin{proof} We first note that $\|B^*B-I\| = \sup_{x \in S^{n-1}} \big| \|Bx\|_2^2 - 1 \big|$. We fix $x = (x_1,\ldots,x_n) \in S^{n-1}$ and, expanding as in \eqref{norm expansion}, observe that $$ \|Bx\|_2^2 = \sum_{j=1}^n x_j^2 \|B_j\|_2^2 + \sum_{j,k \in [n], \, j \ne k} x_j x_k \< B_j, B_k \> . $$ The first sum equals $1$ since $\|B_j\|_2 = \|x\|_2 = 1$. So by Decoupling Lemma~\ref{decoupling}, a random subset $T$ of $[n]$ with average cardinality $n/2$ satisfies $$ \|Bx\|_2^2 - 1 = 4 \mathbb{E}_T \sum_{j \in T, k \in T^c} x_j x_k \< B_j, B_k \> . $$ Let us denote by $\mathbb{E}_T$ and $\mathbb{E}_B$ the expectations with respect to the random set $T$ and the random matrix $B$ respectively. Using Jensen's inequality we obtain \begin{align*} \mathbb{E}_B \|B^*B-I\| &= \mathbb{E}_B \sup_{x \in S^{n-1}} \big| \|Bx\|_2^2 - 1 \big| \\ &\le 4 \mathbb{E}_B \mathbb{E}_T \sup_{x \in S^{n-1}} \Big| \sum_{j \in T, k \in T^c} x_j x_k \< B_j, B_k \> \Big| = 4 \mathbb{E}_T \mathbb{E}_B \|(B_T)^* B_{T^c} \|. \end{align*} The conclusion follows by replacing the expectation by the maximum over $T$. \end{proof} \begin{proof}[Proof of Theorem~\ref{heavy-tailed columns}] {\bf Step 1: Reductions and decoupling.} It would be useful to have an a priori bound on $s_{\max}(A) = \|A\|$. We can obtain this by transposing $A$ and applying one of the results of Section~\ref{s: rows}. Indeed, the random $n \times N$ matrix $A^*$ has independent rows $A_i^*$ which by our assumption are normalized as $\|A_i^*\|_2 = \|A_i\|_2 = \sqrt{N}$. Applying Theorem~\ref{heavy-tailed rows exp si} with the roles of $n$ and $N$ switched, we obtain by the triangle inequality that \begin{equation} \label{norm A} \mathbb{E} \|A\| = \mathbb{E} \|A^*\| = \mathbb{E} s_{\max}(A^*) \le \sqrt{n} + C \sqrt{N \log n} \le C \sqrt{N \log n}. \end{equation} Observe that $n \le m$ since by Lemma~\ref{norm isotropic} we have $\frac{1}{N} \mathbb{E} \< A_j, A_k \> ^2 =1$ for $j \ne k$. We use Matrix Decoupling Lemma~\ref{matrix decoupling} for $B = \frac{1}{\sqrt{N}} A$ and obtain \begin{equation} \label{E via Sigma} E \le \frac{4}{N} \max_{T \subseteq [n]} \mathbb{E} \|(A_T)^* A_{T^c} \| = \frac{4}{N} \max_{T \subseteq [n]} \mathbb{E} \|\Gamma\| \end{equation} where $\Gamma = \Gamma(T)$ denotes the decoupled Gram matrix $$ \Gamma = (A_T)^* A_{T^c} = \big( \< A_j, A_k\> \big)_{j \in T, k \in T^c}. $$ Let us fix $T$; our problem then reduces to bounding the expected norm of $\Gamma$. {\bf Step 2: The rows of the decoupled Gram matrix.} For a subset $S \subseteq [n]$, we denote by $\mathbb{E}_{A_S}$ the conditional expectation given $A_{S^c}$, i.e. with respect to $A_S = (A_j)_{j \in S}$. Hence $\mathbb{E} = \mathbb{E}_{A_{T^c}} \mathbb{E}_{A_T}$. Let us condition on $A_{T^c}$. Treating $(A_k)_{k \in T^c}$ as fixed vectors we see that, conditionally, the random matrix $\Gamma$ has independent rows $$ \Gamma_j = \big( \< A_j, A_k\> \big)_{k \in T^c}, \quad j \in T. $$ So we are going to use Theorem~\ref{heavy-tailed rows exp si non-isotropic} to bound the norm of $\Gamma$. To do this we need estimates on (a) the norms and (b) the second moment matrices of the rows $\Gamma_j$. (a) Since for $j \in T$, $\Gamma_j$ is a random vector valued in $\mathbb{R}^{T^c}$, we estimate its second moment matrix by choosing $x \in \mathbb{R}^{T^c}$ and evaluating the scalar second moment \begin{align*} \mathbb{E}_{A_T} \< \Gamma_j, x\> ^2 &= \mathbb{E}_{A_T} \Big( \sum_{k \in T^c} \< A_j, A_k\> x_k \Big)^2 = \mathbb{E}_{A_T} \Big\langle A_j, \sum_{k \in T^c} x_k A_k \Big\rangle ^2 \\ &= \Big\| \sum_{k \in T^c} x_k A_k \Big\|^2 = \|A_{T^c}x\|_2^2 \le \|A_{T^c}\|_2^2 \|x\|_2^2. \end{align*} In the third equality we used isotropy of $A_j$. Taking maximum over all $j \in T$ and $x \in \mathbb{R}^{T^c}$, we see that the second moment matrix $\Sigma(\Gamma_j) = \mathbb{E}_{A_T} \Gamma_j \otimes \Gamma_j$ satisfies \begin{equation} \label{Sigma Gj} \max_{j \in T} \|\Sigma(\Gamma_j)\| \le \|A_{T^c}\|^2. \end{equation} (b) To evaluate the norms of $\Gamma_j$, $j \in T$, note that $\|\Gamma_j\|_2^2 = \sum_{k \in T^c} \< A_j, A_k\> ^2$. This is easy to bound, because the assumption says that the random variable $$ M := \frac{1}{N} \max_{j \in [n]} \sum_{k \in [n], \, k \ne j} \< A_j, A_k \> ^2 \quad \text{satisfies } \mathbb{E} M = m. $$ This produces the bound $\mathbb{E} \max_{j \in T} \|\Gamma_j\|_2^2 \le N \cdot \mathbb{E} M = Nm$. But at this moment we need to work conditionally on $A_{T^c}$, so for now we will be satisfied with \begin{equation} \label{rows of G} \mathbb{E}_{A_T} \max_{j \in T} \|\Gamma_j\|_2^2 \le N \cdot \mathbb{E}_{A_T} M. \end{equation} {\bf Step 3: The norm of the decoupled Gram matrix.} We bound the norm of the random $T \times T^c$ Gram matrix $\Gamma$ with (conditionally) independent rows using Theorem~\ref{heavy-tailed rows exp si non-isotropic} and Remark~\ref{r: different second moments}. Since by \eqref{Sigma Gj} we have $\big\| \frac{1}{|T|} \sum_{j \in T} \Sigma(\Gamma_j) \big\| \le \frac{1}{|T|} \sum_{j \in T} \|\Sigma(\Gamma_j)\| \le \|A_{T^c}\|^2$, we obtain using \eqref{rows of G} that \begin{align} \label{EAT Sigma} \mathbb{E}_{A_T} \|\Gamma\| &\le (\mathbb{E}_{A_T} \|\Gamma\|^2)^{1/2} \le \|A_{T^c}\| \sqrt{|T|} + C \sqrt{N \cdot \mathbb{E}_{A_T} (M) \log |T^c|} \nonumber\\ &\le \|A_{T^c}\| \sqrt{n} + C \sqrt{N \cdot \mathbb{E}_{A_T} (M) \log n}. \end{align} Let us take expectation of both sides with respect to $A_{T^c}$. The left side becomes the quantity we seek to bound, $\mathbb{E} \|\Gamma\|$. The right side will contain the term which we can estimate by \eqref{norm A}: $$ \mathbb{E}_{A_{T^c}} \|A_{T^c}\| = \mathbb{E} \|A_{T^c}\| \le \mathbb{E} \|A\| \le C \sqrt{N \log n}. $$ The other term that will appear in the expectation of \eqref{EAT Sigma} is $$ \mathbb{E}_{A_{T^c}} \sqrt{\mathbb{E}_{A_T} (M)} \le \sqrt{\mathbb{E}_{A_{T^c}} \mathbb{E}_{A_T} (M)} \le \sqrt{\mathbb{E} M} = \sqrt{m}. $$ So, taking the expectation in \eqref{EAT Sigma} and using these bounds, we obtain $$ \mathbb{E} \|\Gamma\| = \mathbb{E}_{A_{T^c}} \mathbb{E}_{A_T} \|\Gamma\| \le C \sqrt{N \log n} \sqrt{n} + C \sqrt{N m \log n} \le 2C \sqrt{N m \log n} $$ where we used that $n \le m$. Finally, using this estimate in \eqref{E via Sigma} we conclude $$ E \le 8C \sqrt{\frac{m \log n}{N}}. $$ This establishes the first part of Theorem~\ref{heavy-tailed columns}. The second part follow by the diagonalization argument as in Step 2 of the proof of Theorem~\ref{heavy-tailed rows exp si}. \end{proof} \begin{remark}[Incoherence] A priori control on the {\em incoherence} \index{Incoherence} is essential in Theorem~\ref{heavy-tailed columns}. Consider for instance an $N \times n$ random matrix $A$ whose columns are independent coordinate random vectors in $\mathbb{R}^N$. Clearly $s_{\max}(A) \ge \max_j \|A_i\|_2 = \sqrt{N}$. On the other hand, if the matrix is not too tall, $n \gg \sqrt{N}$, then $A$ has two identical columns with high probability, which yields $s_{\min}(A)=0$. \end{remark} \section{Restricted isometries} \index{Restricted isometries} \label{s: restricted isometries} In this section we consider an application of the non-asymptotic random matrix theory in compressed sensing. For a thorough introduction to compressed sensing, see the introductory chapter of this book and \cite{FR, CS website}. In this area, $m \times n$ matrices $A$ are considered as measurement devices, taking as input a signal $x \in \mathbb{R}^n$ and returning its measurement $y = Ax \in \mathbb{R}^m$. One would like to take measurements economically, thus keeping $m$ as small as possible, and still to be able to recover the signal $x$ from its measurement $y$. The interesting regime for compressed sensing is where we take very few measurements, $m \ll n$. Such matrices $A$ are not one-to-one, so recovery of $x$ from $y$ is not possible for all signals $x$. But in practical applications, the amount of ``information'' contained in the signal is often small. Mathematically this is expressed as {\em sparsity} of $x$. In the simplest case, one assumes that $x$ has few non-zero coordinates, say $|\supp(x)| \le k \ll n$. In this case, using any non-degenerate matrix $A$ one can check that $x$ can be recovered whenever $m > 2k$ using the optimization problem $\min \{ |\supp(x)|: \; Ax=y \}$. This optimization problem is highly non-convex and generally NP-complete. So instead one considers a convex relaxation of this problem, $\min \{ \|x\|_1: \; Ax=y \}$. A basic result in compressed sensing, due to Cand\`es and Tao \cite{Candes-Tao, Candes}, is that for sparse signals $|\supp(x)| \le k$, the convex problem recovers the signal $x$ from its measurement $y$ exactly, provided that the measurement matrix $A$ is quantitatively non-degenerate. Precisely, the non-degeneracy of $A$ means that it satisfies the following {\em restricted isometry property} with $\delta_{2k}(A) \le 0.1$. \begin{definition-notag}[Restricted isometries] An $m \times n$ matrix $A$ satisfies the {\em restricted isometry property} of order $k \ge 1$ if there exists $\delta_k \ge 0$ such that the inequality \begin{equation} \label{eq RIP} (1-\delta_k) \|x\|_2^2 \le \|Ax\|_2^2 \le (1+\delta_k) \|x\|_2^2 \end{equation} holds for all $x \in \mathbb{R}^n$ with $|\supp(x)| \le k$. The smallest number $\delta_k = \delta_k(A)$ is called the {\em restricted isometry constant} of $A$. \end{definition-notag} In words, $A$ has a restricted isometry property if $A$ acts as an approximate isometry on all sparse vectors. Clearly, \begin{equation} \label{equiv RIP} \delta_k(A) = \max_{|T| \le k} \|A_T^* A_T - I_{\mathbb{R}^T}\| = \max_{|T| = \lfloor k \rfloor} \|A_T^* A_T - I_{\mathbb{R}^T}\| \end{equation} where the maximum is over all subsets $T \subseteq [n]$ with $|T| \le k$ or $|T| = \lfloor k \rfloor$. The concept of restricted isometry can also be expressed via extreme singular values, which brings us to the topic we studied in the previous sections. $A$ is a restricted isometry if and only if all $m \times k$ sub-matrices $A_T$ of $A$ (obtained by selecting arbitrary $k$ columns from $A$) are approximate isometries. Indeed, for every $\delta \ge 0$, Lemma~\ref{approximate isometries} shows that the following two inequalities are equivalent up to an absolute constant: \begin{gather} \delta_k(A) \le \max(\delta,\delta^2); \label{dk dd} \\ 1-\delta \le s_{\min}(A_T) \le s_{\max}(A_T) \le 1+\delta \label{s restricted} \quad \text{for all } |T| \le k. \end{gather} More precisely, \eqref{dk dd} implies \eqref{s restricted} and \eqref{s restricted} implies $\delta_k(A) \le 3\max(\delta,\delta^2)$. \medskip Our goal is thus to find matrices that are good restricted isometries. What good means is clear from the goals of compressed sensing described above. First, we need to keep the restricted isometry constant $\delta_k(A)$ below some small absolute constant, say $0.1$. Most importantly, we would like the number of measurements $m$ to be small, ideally proportional to the sparsity $k \ll n$. This is where non-asymptotic random matrix theory enters. We shall indeed show that, with high probability, $m \times n$ random matrices $A$ are good restricted isometries of order $k$ with $m = O^*(k)$. Here the $O^*$ notation hides some logarithmic factors of $n$. Specifically, in Theorem~\ref{sub-gaussian RIP} we will show that $$ m = O(k \log(n/k)) $$ for sub-gaussian random matrices $A$ (with independent rows or columns). This is due to the strong concentration properties of such matrices. A general observation of this kind is Proposition~\ref{concentration RIP}. It says that if for a given $x$, a random matrix $A$ (taken from any distribution) satisfies inequality \eqref{eq RIP} with high probability, then $A$ is a good restricted isometry. In Theorem~\ref{heavy-tailed RIP} we will extend these results to random matrices without concentration properties. Using a uniform extension of Rudelson's inequality, Corollary~\ref{Rudelson}, we shall show that \begin{equation} \label{m heavy-tailed} m = O(k \log^4 n) \end{equation} for heavy-tailed random matrices $A$ (with independent rows). This includes the important example of random Fourier matrices. \subsection{Sub-gaussian restricted isometries} \index{Sub-gaussian!restricted isometries} In this section we show that $m \times n$ sub-gaussian random matrices $A$ are good restricted isometries. We have in mind either of the following two models, which we analyzed in Sections~\ref{s: sub-gaussian rows} and \ref{s: sub-gaussian columns} respectively: \begin{description} \item[Row-independent model:] the rows of $A$ are independent sub-gaussian isotropic random vectors in $\mathbb{R}^n$; \item[Column-independent model:] the columns $A_i$ of $A$ are independent sub-gaussian isotropic random vectors in $\mathbb{R}^m$ with $\|A_i\|_2 = \sqrt{m}$ a.s. \end{description} Recall that these models cover many natural examples, including Gaussian and Bernoulli matrices (whose entries are independent standard normal or symmetric Bernoulli random variables), general sub-gaussian random matrices (whose entries are independent sub-gaussian random variables with mean zero and unit variance), ``column spherical'' matrices whose columns are independent vectors uniformly distributed on the centered Euclidean sphere in $\mathbb{R}^m$ with radius $\sqrt{m}$, ``row spherical'' matrices whose rows are independent vectors uniformly distributed on the centered Euclidean sphere in $\mathbb{R}^d$ with radius $\sqrt{d}$, etc. \begin{theorem}[Sub-gaussian restricted isometries] \label{sub-gaussian RIP} Let $A$ be an $m \times n$ sub-gaussian random matrix with independent rows or columns, which follows either of the two models above. Then the normalized matrix $\bar{A} = \frac{1}{\sqrt{m}} A$ satisfies the following for every sparsity level $1 \le k \le n$ and every number $\delta \in (0,1)$: $$ \text{if } m \ge C \delta^{-2} k \log (en/k) \quad \text{then } \delta_k(\bar{A}) \le \delta $$ with probability at least $1 - 2\exp(-c \delta^2 m)$. Here $C = C_K$, $c = c_K > 0$ depend only on the subgaussian norm $K = \max_i \|A_i\|_{\psi_2}$ of the rows or columns of $A$. \end{theorem} \begin{proof} Let us check that the conclusion follows from Theorem~\ref{sub-gaussian rows} for the row-independent model, and from Theorem~\ref{sub-gaussian columns} for the column-independent model. We shall control the restricted isometry constant using its equivalent description \eqref{equiv RIP}. We can clearly assume that $k$ is a positive integer. Let us fix a subset $T \subseteq [n]$, $|T| = k$ and consider the $m \times k$ random matrix $A_T$. If $A$ folows the row-independent model, then the rows of $A_T$ are orthogonal projections of the rows of $A$ onto $\mathbb{R}^T$, so they are still independent sub-gaussian isotropic random vectors in $\mathbb{R}^T$. If alternatively, $A$ follows the column-independent model, then trivially the columns of $A_T$ satisfy the same assumptions as the columns of $A$. In either case, Theorem~\ref{sub-gaussian rows} or Theorem~\ref{sub-gaussian columns} applies to $A_T$. Hence for every $s \ge 0$, with probability at least $1-2\exp(-cs^2)$ one has \begin{equation} \label{AT smin smax} \sqrt{m} - C_0\sqrt{k} - s \le s_{\min}(A_T) \le s_{\max}(A_T) \le \sqrt{m} + C_0\sqrt{k} + s. \end{equation} Using Lemma~\ref{approximate isometries} for $\bar{A}_T = \frac{1}{\sqrt{m}}{A_T}$, we see that \eqref{AT smin smax} implies that $$ \| \bar{A}_T^* \bar{A}_T - I_{\mathbb{R}^T} \| \le 3 \max(\delta_0,\delta_0^2) \quad \text{where } \delta_0 = C_0 \sqrt{\frac{k}{m}} + \frac{s}{\sqrt{m}}. $$ Now we take a union bound over all subsets $T \subset [n]$, $|T| = k$. Since there are $\binom{n}{k} \le (en/k)^k$ ways to choose $T$, we conclude that $$ \max_{|T| = k} \| \bar{A}_T^* \bar{A}_T - I_{\mathbb{R}^T} \| \le 3 \max(\delta_0,\delta_0^2) $$ with probability at least $1 - \binom{n}{k} \cdot 2\exp(-cs^2) \ge 1 - 2 \exp \big( k \log (en/k) - cs^2)$. Then, once we choose $\varepsilon > 0$ arbitrarily and let $s = C_1 \sqrt{k \log(en/k)} + \varepsilon \sqrt{m}$, we conclude with probability at least $1 - 2\exp(-c \varepsilon^2 m)$ that $$ \delta_k(\bar{A}) \le 3 \max(\delta_0,\delta_0^2) \quad \text{where } \delta_0 = C_0 \sqrt{\frac{k}{m}} + C_1 \sqrt{\frac{k \log(en/k)}{m}} + \varepsilon. $$ Finally, we apply this statement for $\varepsilon := \delta/6$. By choosing constant $C$ in the statement of the theorem sufficiently large, we make $m$ large enough so that $\delta_0 \le \delta/3$, which yields $3 \max(\delta_0,\delta_0^2) \le \delta$. The proof is complete. \end{proof} The main reason Theorem~\ref{sub-gaussian RIP} holds is that the random matrix $A$ has a strong concentration property, i.e. that $\|\bar{A}x\|_2 \approx \|x\|_2$ with high probability for every fixed sparse vector $x$. This concentration property alone implies the restricted isometry property, regardless of the specific random matrix model: \begin{proposition}[Concentration implies restricted isometry, see \cite{BDDW}] \label{concentration RIP} Let $A$ be an $m \times n$ random matrix, and let $k \ge 1$, $\delta \ge 0$, $\varepsilon > 0$. Assume that for every fixed $x \in \mathbb{R}^n$, $|\supp(x)| \le k$, the inequality $$ (1-\delta) \|x\|_2^2 \le \|Ax\|_2^2 \le (1+\delta) \|x\|_2^2 $$ holds with probability at least $1 - \exp(-\varepsilon m)$. Then we have the following: $$ \text{if } m \ge C \varepsilon^{-1} k \log (en/k) \quad \text{then } \delta_k(\bar{A}) \le 2\delta $$ with probability at least $1 - \exp(-\varepsilon m / 2)$. Here $C$ is an absolute constant. \end{proposition} In words, the restricted isometry property can be checked on each individual vector $x$ with high probability. \begin{proof} We shall use the expression \eqref{equiv RIP} to estimate the restricted isometry constant. We can clearly assume that $k$ is an integer, and focus on the sets $T \subseteq [n]$, $|T| = k$. By Lemma~\ref{net cardinality}, we can find a net $\mathcal{N}_T$ of the unit sphere $S^{n-1} \cap \mathbb{R}^T$ with cardinality $|\mathcal{N}_T| \le 9^k$. By Lemma~\ref{norm on net}, we estimate the operator norm as $$ \big\| A_T^* A_T - I_{\mathbb{R}^T} \big\| \le 2 \max_{x \in \mathcal{N}_T} \big| \big\langle (A_T^* A_T - I_{\mathbb{R}^T})x, x \big\rangle \big| = 2 \max_{x \in \mathcal{N}_T} \big| \|Ax\|_2^2 - 1 \big|. $$ Taking maximum over all subsets $T \subseteq [n]$, $|T| = k$, we conclude that $$ \delta_k(A) \le 2 \max_{|T| = k} \max_{x \in \mathcal{N}_T} \big| \|Ax\|_2^2 - 1 \big|. $$ On the other hand, by assumption we have for every $x \in \mathcal{N}_T$ that $$ \mathbb{P} \big\{ \big| \|Ax\|_2^2 - 1 \big| > \delta \big\} \le \exp(-\varepsilon m). $$ Therefore, taking a union bound over $\binom{n}{k} \le (en/k)^k$ choices of the set $T$ and over $9^k$ elements $x \in \mathcal{N}_T$, we obtain that \begin{align*} \mathbb{P} \{ \delta_k(A) > 2\delta \} &\le \binom{n}{k} 9^k \exp(-\varepsilon m) \le \exp \big( k \ln(en/k) + k \ln 9 - \varepsilon m \big) \\ &\le \exp(- \varepsilon m / 2) \end{align*} where the last line follows by the assumption on $m$. The proof is complete. \end{proof} \subsection{Heavy-tailed restricted isometries} \index{Heavy-tailed!restricted isometries} \label{s: heavy-tailed RIP} In this section we show that $m \times n$ random matrices $A$ with independent heavy-tailed rows (and uniformly bounded coefficients) are good restricted isometries. This result will be established in Theorem~\ref{heavy-tailed RIP}. As before, we will prove this by controlling the extreme singular values of all $m \times k$ sub-matrices $A_T$. For each individual subset $T$, this can be achieved using Theorem~\ref{heavy-tailed rows}: one has \begin{equation} \label{AT individual control} \sqrt{m} - t \sqrt{k} \le s_{\min}(A_T) \le s_{\max}(A_T) \le \sqrt{m} + t \sqrt{k} \end{equation} with probability at least $1 - 2 k \cdot \exp(-ct^2)$. Although this optimal probability estimate has optimal order, it is too weak to allow for a union bound over all $\binom{n}{k} = (O(1) n/k)^k$ choices of the subset $T$. Indeed, in order that $1 - \binom{n}{k} 2 k \cdot \exp(-ct^2) > 0$ one would need to take $t > \sqrt{k \log(n/k)}$. So in order to achieve a nontrivial lower bound in \eqref{AT individual control}, one would be forced to take $m \ge k^2$. This is too many measurements; recall that our hope is $m = O^*(k)$. This observation suggests that instead of controlling each sub-matrix $A_T$ separately, we should learn how to control all $A_T$ at once. This is indeed possible with the following uniform version of Theorem~\ref{heavy-tailed rows exp si}: \begin{theorem}[Heavy-tailed rows; uniform] \label{heavy-tailed rows uniform} Let $A=(a_{ij})$ be an $N \times d$ matrix ($1 < N \le d$) whose rows $A_i$ are independent isotropic random vectors in $\mathbb{R}^d$. Let $K$ be a number such that all entries $|a_{ij}| \le K$ almost surely. Then for every $1 < n \le d$, we have $$ \mathbb{E} \max_{|T| \le n} \max_{j \le |T|} |s_j(A_T) - \sqrt{N}| \le C l \sqrt{n} $$ where $l = \log(n) \sqrt{\log d} \sqrt{\log N}$ and where $C=C_K$ may depend on $K$ only. The maximum is, as usual, over all subsets $T \subseteq [d]$, $|T| \le n$. \end{theorem} The non-uniform prototype of this result, Theorem~\ref{heavy-tailed rows exp si}, was based on Rudelson's inequality, Corollary~\ref{Rudelson}. In a very similar way, Theorem~\ref{heavy-tailed rows uniform} is based on the following uniform version of Rudelon's inequality. \begin{proposition}[Uniform Rudelson's inequality \cite{RV Fourier}] \index{Rudelson's inequality} \label{RV Fourier} Let $x_1, \ldots, x_N$ be vectors in $\mathbb{R}^d$, $1 < N \le d$, and let $K$ be a number such that all $\|x_i\|_\infty \le K$. Let $\varepsilon_1, \ldots, \varepsilon_N$ be independent symmetric Bernoulli random variables. Then for every $1 < n \le d$ one has $$ \mathbb{E} \max_{|T| \le n} \Big\| \sum_{i=1}^N \varepsilon_i (x_i)_T \otimes (x_i)_T \Big\| \le C l \sqrt{n} \cdot \max_{|T| \le n} \Big\| \sum_{i=1}^N (x_i)_T \otimes (x_i)_T \Big\|^{1/2} $$ where $l = \log(n) \sqrt{\log d} \sqrt{\log N}$ and where $C = C_K$ may depend on $K$ only. \end{proposition} The non-uniform Rudelson's inequality (Corollary~\ref{Rudelson}) was a consequence of a non-commutative Khintchine inequality. Unfortunately, there does not seem to exist a way to deduce Proposition~\ref{RV Fourier} from any known result. Instead, this proposition is proved using Dudley's integral inequality for Gaussian processes and estimates of covering numbers going back to Carl, see \cite{RV Fourier}. It is known however that such usage of Dudley's inequality is not optimal (see e.g. \cite{Ta book}). As a result, the logarithmic factors in Proposition~\ref{RV Fourier} are probably not optimal. In contrast to these difficulties with Rudelson's inequality, proving uniform versions of the other two ingredients of Theorem~\ref{heavy-tailed rows exp si} -- the deviation Lemma~\ref{deviation from 1} and Symmetrization Lemma~\ref{symmetrization} -- is straightforward. \begin{lemma} \label{deviation from 1 uniform} Let $(Z_t)_{t \in \mathcal{T}}$ be a stochastic process\footnote{A stochastic process $(Z_t)$ is simply a collection of random variables on a common probability space indexed by elements $t$ of some abstract set $\mathcal{T}$. In our particular application, $\mathcal{T}$ will consist of all subsets $T \subseteq [d]$, $|T| \le n$.} such that all $Z_t \ge 0$. Then $ \mathbb{E} \sup_{t \in \mathcal{T}} |Z_t^2-1| \ge \max( \mathbb{E} \sup_{t \in \mathcal{T}} |Z_t-1|, (\mathbb{E} \sup_{t \in \mathcal{T}} |Z_t-1|)^2 ). $ \end{lemma} \begin{proof} The argument is entirely parallel to that of Lemma~\ref{deviation from 1}. \end{proof} \begin{lemma}[Symmetrization for stochastic processes] \index{Symmetrization} \label{symmetrization uniform} Let $X_{it}$, $1 \le i \le N$, $t \in \mathcal{T}$, be random vectors valued in some Banach space $B$, where $\mathcal{T}$ is a finite index set. Assume that the random vectors $X_i = (X_{ti})_{t \in \mathcal{T}}$ (valued in the product space $B^\mathcal{T}$) are independent. Let $\varepsilon_1,\ldots, \varepsilon_N$ be independent symmetric Bernoulli random variables. Then $$ \mathbb{E} \sup_{t \in \mathcal{T}} \Big\| \sum_{i=1}^N (X_{it} - \mathbb{E} X_{it}) \Big\| \le 2 \mathbb{E} \sup_{t \in \mathcal{T}} \Big\| \sum_{i=1}^N \varepsilon_i X_{it} \Big\|. $$ \end{lemma} \begin{proof} The conclusion follows from Lemma~\ref{symmetrization} applied to random vectors $X_i$ valued in the product Banach space $B^\mathcal{T}$ equipped with the norm $||| (Z_t)_{t \in \mathcal{T}} ||| = \sup_{t \in \mathcal{T}} \|Z_t\|$. The reader should also be able to prove the result directly, following the proof of Lemma~\ref{symmetrization}. \end{proof} \begin{proof}[Proof of Theorem~\ref{heavy-tailed rows uniform}] Since the random vectors $A_i$ are isotropic in $\mathbb{R}^d$, for every fixed subset $T \subseteq [d]$ the random vectors $(A_i)_T$ are also isotropic in $\mathbb{R}^T$, so $\mathbb{E} (A_i)_T \otimes (A_i)_T = I_{\mathbb{R}^T}$. As in the proof of Theorem~\ref{heavy-tailed rows exp si}, we are going to control \begin{align*} E := \mathbb{E} \max_{|T| \le n} \big\| \frac{1}{N} A_T^* A_T - I_{\mathbb{R}^T} \big\| &= \mathbb{E} \max_{|T| \le n} \Big\| \frac{1}{N} \sum_{i=1}^N (A_i)_T \otimes (A_i)_T - I_{\mathbb{R}^T} \Big\| \\ &\le \frac{2}{N} \, \mathbb{E} \max_{|T| \le n} \Big\| \sum_{i=1}^N \varepsilon_i (A_i)_T \otimes (A_i)_T \Big\| \end{align*} where we used Symmetrization Lemma~\ref{symmetrization uniform} with independent symmetric Bernoulli random variables $\varepsilon_1,\ldots, \varepsilon_N$. The expectation in the right hand side is taken both with respect to the random matrix $A$ and the signs $(\varepsilon_i)$. First taking the expectation with respect to $(\varepsilon_i)$ (conditionally on $A$) and afterwards the expectation with respect to $A$, we obtain by Proposition~\ref{RV Fourier} that $$ E \le \frac{C_K l \sqrt{n}}{N} \, \mathbb{E} \max_{|T| \le n} \Big\| \sum_{i=1}^N (A_i)_T \otimes (A_i)_T \Big\|^{1/2} = \frac{C_K l \sqrt{n}}{\sqrt{N}} \, \mathbb{E} \max_{|T| \le n} \big\| \frac{1}{N} A_T^* A_T \big\|^{1/2} $$ By the triangle inequality, $\mathbb{E} \max_{|T| \le n} \big\| \frac{1}{N} A_T^* A_T \big\| \le E + 1$. Hence we obtain $$ E \le C_K l \sqrt{\frac{n}{N}} (E+1)^{1/2} $$ by H\"older's inequality. Solving this inequality in $E$ we conclude that \begin{equation} \label{A*A heavy-tailed exp} E = \mathbb{E} \max_{|T| \le n} \big\| \frac{1}{N} A_T^* A_T - I_{\mathbb{R}^T} \big\| \le \max(\delta, \delta^2) \quad \text{where } \delta = C_K l \sqrt{\frac{2n}{N}}. \end{equation} The proof is completed by a diagonalization argument similar to Step 2 in the proof of Theorem~\ref{heavy-tailed rows exp si}. One uses there a uniform version of deviation inequality given in Lemma~\ref{deviation from 1 uniform} for stochastic processes indexed by the sets $|T| \le n$. We leave the details to the reader. \end{proof} \begin{theorem}[Heavy-tailed restricted isometries] \label{heavy-tailed RIP} Let $A=(a_{ij})$ be an $m \times n$ matrix whose rows $A_i$ are independent isotropic random vectors in $\mathbb{R}^n$. Let $K$ be a number such that all entries $|a_{ij}| \le K$ almost surely. Then the normalized matrix $\bar{A} = \frac{1}{\sqrt{m}} A$ satisfies the following for $m \le n$, for every sparsity level $1 < k \le n$ and every number $\delta \in (0,1)$: \begin{equation} \label{eq heavy-tailed RIP} \text{if } m \ge C \delta^{-2} k \log n \log^2(k) \log(\delta^{-2} k \log n \log^2 k) \quad \text{then } \mathbb{E} \delta_k(\bar{A}) \le \delta. \end{equation} Here $C = C_K > 0$ may depend only on $K$. \end{theorem} \begin{proof} The result follows from Theorem~\ref{heavy-tailed rows uniform}, more precisely from its equivalent statement \eqref{A*A heavy-tailed exp}. In our notation, it says that $$ \mathbb{E} \delta_k(\bar{A}) \le \max(\delta,\delta^2) \quad \text{where } \delta = C_K l \sqrt{\frac{k}{m}} = C_K \sqrt{\frac{k \log m}{m}} \log(k) \sqrt{\log n}. $$ The conclusion of the theorem easily follows. \end{proof} In the interesting sparsity range $k \ge \log n$ and $k \ge \delta^{-2}$, the condition in Theorem~\ref{heavy-tailed RIP} clearly reduces to $$ m \ge C \delta^{-2} k \log (n) \log^{3} k. $$ \begin{remark}[Boundedness requirement] The {\em boundedness assumption} on the entries of $A$ is essential in Theorem~\ref{heavy-tailed RIP}. Indeed, if the rows of $A$ are independent coordinate vectors in $\mathbb{R}^n$, then $A$ necessarily has a zero column (in fact $n-m$ of them). This clearly contradicts the restricted isometry property. \end{remark} \begin{example} \label{random measurements} \begin{enumerate} \item {\bf (Random Fourier measurements):} \index{Fourier measurements} An important example for Theorem~\ref{heavy-tailed rows} is where $A$ realizes random Fourier measurements. Consider the $n \times n$ Discrete Fourier Transform (DFT) matrix $W$ with entries $$ W_{\omega,t} = \exp \Big( -\frac{2 \pi i \omega t}{n} \Big), \quad \omega, t \in \{0,\ldots,n-1\}. $$ Consider a random vector $X$ in $\mathbb{C}^n$ which picks a random row of $W$ (with uniform distribution). It follows from Parseval's inequality that $X$ is isotropic.\footnote{For convenience we have developed the theory over $\mathbb{R}$, while this example is over $\mathbb{C}$. As we noted earlier, all our definitions and results can be carried over to the complex numbers. So in this example we use the obvious complex versions of the notion of isotropy and of Theorem~\ref{heavy-tailed RIP}.} Therefore the $m \times n$ random matrix $A$ whose rows are independent copies of $X$ satisfies the assumptions of Theorem~\ref{heavy-tailed rows} with $K=1$. Algebraically, we can view $A$ as a {\em random row sub-matrix of the DFT matrix}. In compressed sensing, such matrix $A$ has a remarkable meaning -- it realizes $m$ {\em random Fourier measurements} of a signal $x \in \mathbb{R}^n$. Indeed, $y=Ax$ is the DFT of $x$ evaluated at $m$ random points; in words, $y$ consists of $m$ random frequencies of $x$. Recall that in compressed sensing, we would like to guarantee that with high probability every sparse signal $x \in \mathbb{R}^n$ (say, $|\supp(x)| \le k$) can be effectively recovered from its $m$ random frequencies $y=Ax$. Theorem~\ref{heavy-tailed RIP} together with Cand\`es-Tao's result (recalled in the beginning of Section~\ref{s: restricted isometries}) imply that an exact recovery is given by the convex optimization problem $\min\{ \|x\|_1 : Ax=y\}$ provided that we observe {\em slightly more frequencies than the sparsity of a signal}: $m \gtrsim \ge C \delta^{-2} k \log (n) \log^{3} k$. \item {\bf (Random sub-matrices of orthogonal matrices):} \index{Sub-matrices} In a similar way, Theorem~\ref{heavy-tailed RIP} applies to a random row sub-matrix $A$ of an {\em arbitrary bounded orthogonal matrix} $W$. Precisely, $A$ may consist of $m$ randomly chosen rows, uniformly and without replacement,\footnote{Since in the interesting regime very few rows are selected, $m \ll n$, sampling with or without replacement are formally equivalent. For example, see \cite{RV Fourier} which deals with the model of sampling without replacement.} from an arbitrary $n \times n$ matrix $W = (w_{ij})$ such that $W^*W = n I$ and with uniformly bounded coefficients, $\max_{ij} |w_{ij}| = O(1)$. The examples of such $W$ include the class of {\em Hadamard matrices} \index{Hadamard matrices} -- orthogonal matrices in which all entries equal $\pm 1$. \end{enumerate} \end{example} \section{Notes} \label{s: notes} \paragraph{For Section~\ref{s: introduction}} We work with two kinds of moment assumptions for random matrices: sub-gaussian and heavy-tailed. These are the two extremes. By the central limit theorem, the sub-gaussian tail decay is the strongest condition one can demand from an isotropic distribution. In contrast, our heavy-tailed model is completely general -- no moment assumptions (except the variance) are required. It would be interesting to analyze random matrices with independent rows or columns in the intermediate regime, {\em between sub-gaussian and heavy-tailed} moment assumptions. We hope that for distributions with an appropriate finite moment (say, $(2+\varepsilon)$th or $4$th), the results should be the same as for sub-gaussian distributions, i.e. no $\log n$ factors should occur. In particular, tall random matrices ($N \gg n)$ should still be approximate isometries. This indeed holds for sub-exponential distributions \cite{ALPT}; see \cite{V covariance} for an attempt to go down to finite moment assumptions. \paragraph{For Section~\ref{s: preliminaries}} The material presented here is well known. The volume argument presented in Lemma~\ref{net cardinality} is quite flexible. It easily generalizes to covering numbers of more general metric spaces, including convex bodies in Banach spaces. See \cite[Lemma 4.16]{Pisier volume} and other parts of \cite{Pisier volume} for various methods to control covering numbers. \paragraph{For Section~\ref{s: sub-gaussian}} The concept of sub-gaussian random variables is due to Kahane~\cite{Kahane}. His definition was based on the moment generating function (Property 4 in Lemma~\ref{sub-gaussian properties}), which automatically required sub-gaussian random variables to be centered. We found it more convenient to use the equivalent Property 3 instead. The characterization of sub-gaussian random variables in terms of tail decay and moment growth in Lemma~\ref{sub-gaussian properties} also goes back to \cite{Kahane}. The rotation invariance of sub-gaussian random variables (Lemma~\ref{rotation invariance}) is an old observation \cite{BuKo}. Its consequence, Proposition~\ref{sub-gaussian large deviations}, is a general form of {\em Hoeffding's inequality}, which is usually stated for bounded random variables. For more on large deviation inequalities, see also notes for Section~\ref{s: sub-exponential}. Khintchine inequality is usually stated for the particular case of symmetric Bernoulli random variables. It can be extended for $0<p<2$ using a simple extrapolation argument based on H\"older's inequality, see \cite[Lemma~4.1]{Ledoux-Talagrand}. \paragraph{For Section~\ref{s: sub-exponential}} Sub-gaussian and sub-exponential random variables can be studied together in a general framework. For a given exponent $0 < \alpha < \infty$, one defines general $\psi_\alpha$ random variables, those with moment growth $(\mathbb{E} |X|^p)^{1/p} = O(p^{1/\alpha})$. Sub-gaussian random variables correspond to $\alpha = 2$ and sub-exponentials to $\alpha = 1$. The reader is encouraged to extend the results of Sections~\ref{s: sub-gaussian} and \ref{s: sub-exponential} to this general class. Proposition~\ref{sub-exponential large deviations} is a form of {\em Bernstein's inequality}, which is usually stated for bounded random variables in the literature. These forms of Hoeffding's and Bernstein's inequalities (Propositions~\ref{sub-gaussian large deviations} and \ref{sub-exponential large deviations}) are partial cases of a large deviation inequality for general $\psi_\alpha$ norms, which can be found in \cite[Corollary~2.10]{Talagrand canonical} with a similar proof. For a thorough introduction to large deviation inequalities for sums of independent random variables (and more), see the books \cite{Petrov, Ledoux-Talagrand, Dembo-Zeitouni} and the tutorial \cite{BBL}. \paragraph{For Section~\ref{s: isotropic}} Sub-gaussian distributions in $\mathbb{R}^n$ are well studied in geometric functional analysis; see \cite{MePaTJ reconstruction} for a link with compressed sensing. General $\psi_\alpha$ distributions in $\mathbb{R}^n$ are discussed e.g. in \cite{GiMi concentration}. Isotropic distributions on convex bodies, and more generally isotropic log-concave distributions, are central to asymptotic convex geometry (see \cite{Giannopoulos isotropic, Paouris}) and computational geometry \cite{Vempala}. A completely different way in which isotropic distributions appear in convex geometry is from {\em John's decompositions} for contact points of convex bodies, see \cite{Ball, Rudelson contact, Vershynin John}. Such distributions are finitely supported and therefore are usually heavy-tailed. For an introduction to the concept of {\em frames} (Example~\ref{random vectors}), see \cite{KC, Christensen}. \paragraph{For Section~\ref{s: sums matrices}} The non-commutative Khintchine inequality, Theorem~\ref{non-commutative Khintchine}, was first proved by Lust-Piquard \cite{Lust-Piquard} with an unspecified constant $B_p$ in place of $C \sqrt{p}$. The optimal value of $B_p$ was computed by Buchholz \cite{Buc01, Buc05}; see \cite[Section~6.5]{Rauhut structured} for an thorough introduction to Buchholz's argument. For the complementary range $1 \le p \le 2$, a corresponding version of non-commutative Khintchine inequality was obtained by Lust-Piquard and Pisier \cite{Lu-Pi}. By a duality argument implicitly contained in \cite{Lu-Pi} and independently observed by Marius Junge, this latter inequality also implies the optimal order $B_p = O(\sqrt{p})$, see \cite{Rudelson isotropic} and \cite[Section~9.8]{Pisier operator}. Rudelson's Corollary~\ref{Rudelson} was initially proved using a majorizing measure technique; our proof follows Pisier's argument from \cite{Rudelson isotropic} based on the non-commutative Khintchine inequality. \paragraph{For Section~\ref{s: entries}} The ``Bai-Yin law'' (Theorem~\ref{Bai-Yin}) was established for $s_{\max}(A)$ by Geman \cite{Geman} and Yin, Bai and Krishnaiah \cite{YBK}. The part for $s_{\min}(A)$ is due to Silverstein \cite{Silverstein} for Gaussian random matrices. Bai and Yin \cite{Bai-Yin} gave a unified treatment of both extreme singular values for general distributions. The fourth moment assumption in Bai-Yin's law is known to be necessary \cite{Bai-Silverstein-Yin}. Theorem~\ref{Gaussian} and its argument is due to Gordon \cite{Gordon 84, Gordon 85, Gordon 92}. Our exposition of this result and of Corollary~\ref{Gaussian deviation} follows \cite{DS}. Proposition~\ref{Gaussian concentration} is just a tip of an iceberg called {\em concentration of measure phenomenon}. \index{Concentration of measure} We do not discuss it here because there are many excellent sources, some of which were mentioned in Section~\ref{s: introduction}. Instead we give just one example related to Corollary~\ref{Gaussian deviation}. For a general random matrix $A$ with independent centered entries bounded by $1$, one can use Talagrand's concentration inequality for convex Lipschitz functions on the cube \cite{Tal1, Tal2}. Since $s_{\max}(A)= \|A\|$ is a convex function of $A$, Talagrand's concentration inequality implies $\mathbb{P} \big\{ |s_{\max}(A) - \Median(s_{\max}(A))| \ge t \big\} \le 2 e^{-ct^2}$. Although the precise value of the median may be unknown, integration of this inequality shows that $|\mathbb{E} s_{\max}(A)-\Median(s_{\max}(A))| \le C$. For the recent developments related to the {\em hard edge} problem for almost square and square matrices (including Theorem~\ref{RV rectangular}) see the survey \cite{RV ICM}. \paragraph{For Section~\ref{s: rows}} Theorem~\ref{sub-gaussian rows} on random matrices with sub-gaussian rows, as well as its proof by a covering argument, is a folklore in geometric functional analysis. The use of covering arguments in a similar context goes back to Milman's proof of Dvoretzky's theorem \cite{Milman Dvoretzky}; see e.g. \cite{Ball} and \cite[Chapter 4]{Pisier volume} for an introduction. In the more narrow context of extreme singular values of random matrices, this type of argument appears recently e.g. in \cite{ALPT}. The breakthrough work on heavy-tailed isotropic distributions is due to Rudelson \cite{Rudelson isotropic}. He used Corollary~\ref{Rudelson} in the way we described in the proof of Theorem~\ref{heavy-tailed rows exp si} to show that $\frac{1}{N} A^*A$ is an approximate isometry. Probably Theorem~\ref{heavy-tailed rows} can also be deduced by a modification of this argument; however it is simpler to use the non-commutative Bernstein's inequality. The symmetrization technique is well known. For a slightly more general two-sided inequality than Lemma~\ref{symmetrization}, see \cite[Lemma~6.3]{Ledoux-Talagrand}. The problem of estimating covariance matrices described in Section~\ref{s: covariance} is a basic problem in statistics, see e.g. \cite{Johnstone}. However, most work in the statistical literature is focused on the normal distribution or general product distributions (up to linear transformations), which corresponds to studying random matrices with independent entries. For non-product distributions, an interesting example is for uniform distributions on convex sets \cite{KLS}. As we mentioned in Example~\ref{random vectors sub-gaussian}, such distributions are sub-exponential but not necessarily sub-gaussian, so Corollary~\ref{covariance sub-gaussian} does not apply. Still, the sample size $N = O(n)$ suffices to estimate the covariance matrix in this case \cite{ALPT}. It is conjectured that the same should hold for general distributions with finite (e. g. $4$th) moment assumption \cite{V covariance}. Corollary~\ref{random sub-matrices} on random sub-matrices is a variant of the Rudelson's result from \cite{Rudelson sub-matrices}. The study of random sub-matrices was continued in \cite{RV sampling}. Random sub-frames were studied in \cite{V frames} where a variant of Corollary~\ref{random sub-frames} was proved. \paragraph{For Section~\ref{s: columns}} Theorem~\ref{sub-gaussian columns} for sub-gaussian columns seems to be new. However, historically the efforts of geometric functional analysts were immediately focused on the more difficult case of sub-exponential tail decay (given by uniform distributions on convex bodies). An indication to prove results like Theorem~\ref{sub-gaussian columns} by decoupling and covering is present in \cite{Bourgain} and is followed in \cite{GiMi concentration, ALPT}. The normalization condition $\|A_j\|_2 = \sqrt{N}$ in Theorem~\ref{sub-gaussian columns} can not be dropped but can be relaxed. Namely, consider the random variable $\delta := \max_{i \le n} \big| \frac{\|A_j\|_2^2}{N} - 1 \big|$. Then the conclusion of Theorem~\ref{sub-gaussian columns} holds with \eqref{smin smax columns} replaced by $$ (1-\delta)\sqrt{N} - C \sqrt{n} - t \le s_{\min}(A) \le s_{\max}(A) \le (1+\delta)\sqrt{N} + C \sqrt{n} + t. $$ Theorem~\ref{heavy-tailed columns} for heavy-tailed columns also seems to be new. The incoherence parameter $m$ is meant to prevent collisions of the columns of $A$ in a quantitative way. It is not clear whether the {\em logarithmic factor} is needed in the conclusion of Theorem~\ref{heavy-tailed columns}, or whether the incoherence parameter alone takes care of the logarithmic factors whenever they appear. The same question can be raised for all other results for heavy-tailed matrices in Section~\ref{s: heavy-tailed rows} and their applications -- can we replace the logarithmic factors by more sensitive quantities (e.g. the logarithm of the incoherence parameter)? \paragraph{For Section~\ref{s: restricted isometries}} For a mathematical introduction to compressed sensing, see the introductory chapter of this book and \cite{FR, CS website}. A version of Theorem~\ref{sub-gaussian RIP} was proved in \cite{MePaTJ} for the row-independent model; an extension from sub-gaussian to sub-exponential distributions is given in \cite{ALPT RIP}. A general framework of stochastic processes with sub-exponential tails is discussed in \cite{Mendelson}. For the column-independent model, Theorem~\ref{sub-gaussian RIP} seems to be new. Proposition~\ref{concentration RIP} that formalizes a simple approach to restricted isometry property based on concentration is taken from \cite{BDDW}. Like Theorem~\ref{sub-gaussian RIP}, it can also be used to show that Gaussian and Bernoulli random matrices are restricted isometries. Indeed, it is not difficult to check that these matrices satisfy a concentration inequality as required in Proposition~\ref{concentration RIP} \cite{Achlioptas}. Section~\ref{s: heavy-tailed RIP} on heavy-tailed restricted isometries is an exposition of the results from \cite{RV Fourier}. Using concentration of measure techniques, one can prove a version of Theorem~\ref{heavy-tailed RIP} with high probability $1 - n^{-c \log^3 k}$ rather than in expectation \cite{Rauhut structured}. Earlier, Candes and Tao \cite{Candes-Tao Fourier} proved a similar result for random Fourier matrices, although with a slightly higher exponent in the logarithm for the number of measurements in \eqref{m heavy-tailed}, $m = O(k \log^6 n)$. The survey \cite{Rauhut structured} offers a thorough exposition of the material presented in Section~\ref{s: heavy-tailed RIP} and more.

https://arxiv.org/abs/2110.06244

Congruence properties of combinatorial sequences via Walnut and the Rowland-Yassawi-Zeilberger automaton

Certain famous combinatorial sequences, such as the Catalan numbers and the Motzkin numbers, when taken modulo a prime power, can be computed by finite automata. Many theorems about such sequences can therefore be proved using Walnut, which is an implementation of a decision procedure for proving various properties of automatic sequences. In this paper we explore some results (old and new) that can be proved using this method.

\section{Introduction} We study the properties of two famous combinatorial sequences. For $n \geq 0$, let $$C_n = \frac{1}{n+1}\binom{2n}{n}$$ denote the \emph{$n$-th Catalan number} and let $$M_n = \sum_{k=0}^{\lfloor n/2 \rfloor}\binom{n}{2k}C_k$$ denote the \emph{$n$-th Motzkin number}. For more about the Catalan numbers, see \cite{Stanley}. Many authors have studied congruence properties of these and other sequences modulo primes $p$ or prime powers $p^\alpha$. Notably, Alter and Kubota \cite{AK73} studied the Catalan numbers modulo $p$, and Deutsch and Sagan \cite{DS06} studied many sequences, including the Catalan numbers, Motzkin numbers, Central Delannoy numbers, Ap\'ery numbers, etc., modulo certain prime powers. Eu, Liu, and Yeh \cite{ELY08} studied the Catalan and Motzkin numbers modulo $4$ and $8$, and Krattenthaler and M\"uller \cite{KM18} studied the Motzkin numbers and related sequences modulo powers of $2$. Rowland and Yassawi \cite{RY15} and Rowland and Zeilberger \cite{RZ14} gave different methods to compute finite automata that compute the sequences $(C_n \bmod p^\alpha)_{n \geq 0}$ and $(M_n \bmod p^\alpha)_{n \geq 0}$ (and many other similar sequences), where $p^\alpha$ is a prime power. Rowland and Zeilberger provide a number of these automata for different $p^\alpha$ at the website \begin{center} \url{https://sites.math.rutgers.edu/~zeilberg/mamarim/mamarimhtml/meta.html} \end{center} along with the Maple code used to compute them. We use some of these automata, along with the program Walnut \cite{Walnut}, available at the website \begin{center} \url{https://cs.uwaterloo.ca/~shallit/walnut.html} \end{center} to study properties of these sequences. We use the Rowland--Zeilberger algorithm (and Walnut) as a black-box, so we do not discuss the theory behind it. We just mention that this algorithm applies to any sequence of numbers that can be defined as the \emph{constant term} of $[P(x)]^nQ(x)$, where $P$ and $Q$ are \emph{Laurent polynomials}. In particular, the $n$-th Catalan number is the constant term of $(1/x+2+x)^n(1-x)$ and the $n$-th Motzkin number is the constant term of $(1/x+1+x)^n(1-x^2)$. Note that the automata produced by this method read their input in least-significant-digit-first format. All of the automata in this paper therefore also follow this convention. We use the notation $(n)_k$ to denote the base-$k$ representation of $n$ in the lsd-first format. Burns has posted several manuscripts to the arXiv \cite{Bur_A,Bur_B,Bur_C,Bur_D,Bur_E,Bur_F} in which he investigates various properties of the Catalan and Motzkin numbers modulo primes $p$ by analyzing structural properties of automata computed using the Rowland--Yassawi algorithm. This paper takes a similar approach, but we use Walnut to simplify/automate much of the analysis. \section{Motzkin numbers} Deutsch and Sagan \cite{DS06} gave a characterization of ${\bf m}_2 = (M_n \bmod 2)_{n \geq 0}$ that involves the Thue--Morse sequence $${\bf t} = (t_n)_{n \geq 0} = (0,1,1,0,1,0,0,1,\ldots).$$ Let $${\bf c} = (c_n)_{n \geq 0} = (1,3,4,5,7,\ldots)$$ denote the starting positions of the ``runs'' in ${\bf t}$, excluding the first run (which, of course, starts at position $0$). \begin{theorem}[Deutsch and Sagan] The Motzkin number $M_n$ is even if and only if either $n \in 4{\bf c}-2$ or $n \in 4{\bf c}-1$. \end{theorem} \begin{proof} We can prove this result using Walnut. The Rowland--Zeilberger algorithm produces the algorithm in Figure~\ref{MOT2}, which, when fed with $(n)_2$,\ outputs $M_n \bmod 2$. \begin{figure}[htb] \centering \includegraphics[scale=0.75]{MOT2.pdf} \caption{Automaton for $M_n \bmod 2$}\label{MOT2} \end{figure} Next we use Walnut to construct an automaton for the sequence ${\bf c}$. The commands \begin{verbatim} def tm_blocks "?lsd_2 n>=1 & (At t<n => T_lsd[i+t]=T_lsd[i]) & T_lsd[i+n]!=T_lsd[i] & (i=0|T_lsd[i-1]!=T_lsd[i])": def tm_block_start "?lsd_2 i>=1 & ($tm_blocks(i,1)|$tm_blocks(i,2))": \end{verbatim} produce the automaton \texttt{tm\_block\_start} given in Figure~\ref{tm_block_start}, which computes ${\bf c}$. We see that the elements of ${\bf c}$ are $$ \{ m4^k : m \text{ is odd and } k \geq 0 \}.$$ \begin{figure}[htb] \centering \includegraphics[scale=0.75]{tm_block_start.pdf} \caption{Automaton for starting positions of ``runs'' in {\bf t}}\label{tm_block_start} \end{figure} To complete the proof of the theorem, it suffices to execute the Walnut command \begin{verbatim} eval even_mot "?lsd_2 An (MOT2[n]=@0 <=> Ei $tm_block_start(i) & (n+2=4*i | n+1=4*i))": \end{verbatim} which produces the output ``TRUE''. \end{proof} Deutsch and Sagan also characterized ${\bf m}_3 = (M_n \bmod 3)_{n \geq 0}$: \begin{theorem}{(Deutsch and Sagan)} The Motzkin number $M_n$ satisfies \[ M_n \equiv_3 \begin{cases} 1, & \text{ if either } (n)_3 = 0w, w \in \{0,1\}^* \text{ or } (n+2)_3 = 0w, w \in \{0,1\}^*, \\ 2, & \text{ if } (n+1)_3 = 0w, w \in \{0,1\}^*, \\ 0, & \text{ otherwise.} \end{cases} \] \end{theorem} This can also be obtained directly from the automaton for ${\bf m}_3$. If we examine ${\bf m}_5 = (M_n \bmod 5)_{n \geq 0}$, however, we discover that its behaviour is very different from that of ${\bf m}_3$. Deutsch and Sagan determined the positions of the $0$'s in ${\bf m}_5$. \begin{theorem}{(Deutsch and Sagan)} The Motzkin number $M_n$ is divisible by $5$ if and only if $n$ is of the form \[ (5i+1)5^{2j}-2, \quad (5i+2)5^{2j-1}-1 \quad (5i+3)5^{2j-1}-2 \quad (5i+4)5^{2j}-1 . \] \end{theorem} \begin{proof} The Rowland--Zeilberger algorithm gives the automaton in Figure~\ref{MOT5}, which fully characterizes ${\bf m}_5$. \begin{figure}[htb] \centering \includegraphics[scale=0.60]{MOT5.pdf} \caption{Automaton for $M_n \bmod 5$}\label{MOT5} \end{figure} The Walnut command \begin{verbatim} eval mot5mod0 "?lsd_5 MOT5[n]=@0": \end{verbatim} produces the automaton in Figure~\ref{mot5mod0}, from which one easily derives the result. \begin{figure}[htb] \centering \includegraphics[scale=0.75]{mot5mod0.pdf} \caption{Automaton for the positions of the $0$'s in ${\bf m}_5$}\label{mot5mod0} \end{figure} \end{proof} One notices that ${\bf m}_3$ contains arbitrarily large runs of $0$'s, whereas ${\bf m}_5$ does not have this property. We can use Walnut to determine the types of repetitions that are present in ${\bf m}_5$, but we first need to introduce some definitions. Let $w = w_1w_2\cdots w_n$ be a word of length $n$ and \emph{period} $p$; i.e., $w_i = w_{i+p}$ for $i = 0, \ldots, n-p$. If $p$ is the smallest period of $w$, we say that the \emph{exponent} of $w$ is $n/p$. We also say that $w$ is an \emph{$(n/p)$-power} of \emph{order $p$}. Words of exponent $2$ (resp.~$3$) are called \emph{squares} (resp.~\emph{cubes}). If ${\bf x}$ is an infinite sequence, we define the \emph{critical exponent} of ${\bf x}$ as $$ \sup \{ e \in \mathbb{Q}: \text{ there is a factor of } {\bf x} \text{ with exponent } e \}. $$ \begin{theorem} The sequence ${\bf m}_5$ has critical exponent $3$. Furthermore, the only cubes in ${\bf m}_5$ are $111$, $222$, $333$, and $444$. \end{theorem} \begin{proof} We execute the Walnut commands \begin{verbatim} eval tmp "?lsd_5 Ei,n (n>=1) & At (t<=2*n) => MOT5[i+t]=MOT5[i+t+n]": eval tmp "?lsd_5 Ei (n>=1) & At (t<2*n) => MOT5[i+t]=MOT5[i+t+n]": \end{verbatim} and note that the first outputs ``FALSE'', indicating that ${\bf m}_5$ has no factors of exponent larger than $3$, and the second produces an automaton that only accepts $n=1$, indicating that the only cubes in ${\bf m}_5$ have order $1$. By inspecting a prefix of ${\bf m}_5$, one sees that $111$, $222$, $333$, and $444$ all occur. \end{proof} We can also prove that every pattern of residues that appears in ${\bf m}_5$ appears infinitely often, and furthermore, we can give a bound on when the next occurrence of a pattern will appear in ${\bf m}_5$. We say that a sequence ${\bf x}$ is \emph{uniformly recurrent} if for every factor $w$ of ${\bf x}$, there is a constant $c$ such that every occurrence of $w$ in ${\bf x}$ is followed by another occurrence of $w$ at distance at most $c$. Note that ${\bf m}_3$ is \emph{not} uniformly recurrent. This is due to the presence of arbitrarily large runs of $0$'s in ${\bf m}_3$. On the other hand, the sequence ${\bf m}_5$ exhibits rather different behaviour. \begin{theorem} The sequence ${\bf m}_5$ is uniformly recurrent. Furthermore, if $w$ has length $n$ and occurs at position $i$ in ${\bf m}_5$, then there is another occurrence of $w$ at some position $j$, where $i < j \leq i+200n$. The bound $200n$ cannot be replaced by $200n-1$. \end{theorem} \begin{proof} This is proved with the Walnut commands \begin{verbatim} def mot5faceq "?lsd_5 At (t<n) => (MOT5[i+t]=MOT5[j+t])": eval tmp "?lsd_5 An (n>=1) => Ai Ej (j>i) & (j<i+200*n+1) & $mot5faceq(i,j,n)": eval tmp "?lsd_5 An (n>=1) => Ai Ej (j>i) & (j<i+200*n) & $mot5faceq(i,j,n)": \end{verbatim} noting that the first \texttt{eval} command returns ``TRUE'' and the second returns ``FALSE''. \end{proof} Burns \cite{Bur_E} studied ${\bf m}_p$ for $p$ between $7$ and $29$ using automata computed using the Rowland--Yassawi algorithm. Among other things, his work suggests that depending on the value of $p$, the sequence ${\bf m}_p$ either behaves like ${\bf m}_3$, where $0$ has density $1$ (i.e., $p = 7, 17, 19$), or ${\bf m}_p$ behaves like ${\bf m}_5$, where $0$ has density $<1$ (i.e., $p = 11, 13, 23, 29$). Many of Burns' results could also be obtained using Walnut. \begin{problem}\label{mot_rec}Characterize the primes $p$ for which ${\bf m}_p$ is uniformly recurrent. \end{problem} Indeed, based on Burns' results and the discussion in the next section, we guess that the answer to this problem is given by the sequence $$2, 5, 11, 13, 23, 29, 31, 37, 53, \ldots$$ of primes that do not divide any central trinomial number. This is sequence \seqnum{A113305} of \cite{OEIS}. \section{Central trinomial coefficients} The Motzkin numbers are closely related to the \emph{central trinomial coefficients} $T_n$. The usual definition of $T_n$ is as the coefficient of $x^n$ in $(1+x+x^2)^n$, but the definition $$T_n = \sum_{k \geq 0}\binom{n}{2k}\binom{2k}{k}$$ better illustrates the connection between these numbers and the Motzkin numbers. The number $T_n$ is also the constant term of $(1/x+1+x)^n$, which is the form needed for the Rowland--Zeilberger algorithm. Deutsch and Sagan studied the divisibility of $T_n$ modulo primes and Noe \cite{Noe06} did the same for generalized central trinomial numbers. \begin{theorem}[Deutsch and Sagan] The central trinomial coefficient $T_n$ satisfies \[ T_n \equiv_3 \begin{cases} 1, & \text{ if } (n)_3 \text{ does not contain a } 2;\\ 0, & \text{ otherwise.} \end{cases} \] \end{theorem} Deutsch and Sagan proved this by an application of Lucas' Theorem; it is also immediate from the automaton produced by the Rowland--Zeilberger algorithm. As with the Motzkin numbers, the behaviour of $T_n$ modulo $5$ is rather different from that modulo $3$. We collect some properties below (compare with those of ${\bf m}_5$ from the previous section). \begin{theorem} Let ${\bf t}_5 = (T_n \bmod 5)_{n \geq 0}$. Then \begin{enumerate} \item ${\bf t}_5$ does not contain $0$ (i.e., $T_n$ is never divisible by $5$); \item ${\bf t}_5$ has critical exponent $3$; furthermore, the only cubes in ${\bf t}_5$ are $111$, $222$, $333$, and $444$; \item ${\bf t}_5$ is uniformly recurrent; Furthermore, if $w$ has length $n$ and occurs at position $i$ in ${\bf t}_5$, then there is another occurrence of $w$ at some position $j$, where $i < j \leq i+200n-192$. The constant $192$ cannot be replaced with $193$. \item If $w$ has length $n$ and appears in ${\bf t}_5$, then $w$ appears in the prefix of ${\bf t}_5$ of length $121n$. The quantity $121n$ cannot be replaced with $121n-1$. \end{enumerate} \end{theorem} \begin{proof} Properties 1)--3) can all be obtained by similar Walnut commands to those used in the previous section for the Motzkin numbers. We just need the automaton for ${\bf t}_5$. The Rowland--Zeilberger algorithm gives the pleasantly symmetric automaton in Figure~\ref{TRI5}. \begin{figure}[htb] \centering \includegraphics[scale=0.60]{TRI5.pdf} \caption{Automaton for $T_n \bmod 5$}\label{TRI5} \end{figure} For Property 4), we use the Walnut commands \begin{verbatim} def pr_tri5 "?lsd_5 Aj Ei i+n<=s & At t<n => TRI5[i+t]=TRI5[j+t]": eval tmp "?lsd_5 An $pr_tri5(n,121*n)": eval tmp "?lsd_5 An $pr_tri5(n,121*n-1)": \end{verbatim} and note that the last two commands return "TRUE" and "FALSE", respectively. \end{proof} We should note that in the special case of the central trinomial coefficients, it is not necessary to resort to either the Rowland--Zeilberger or Rowland--Yassawi algorithms to compute the automaton for $T_n \bmod p$. Using the following result of Deutsch and Sagan, one can directly define the automaton for $T_n \bmod p$. \begin{theorem}{(Deutsch and Sagan)}\label{tri_lucas} Let $(n)_p = n_0n_1\cdots n_r$. Then $$T_n \equiv_p \prod_{i=0}^r T_{n_i}.$$ \end{theorem} An immediate consequence is that $T_n$ is divisible by $p$ if and only if one of the $T_{n_i}$ is divisible by $p$. This criterion allows one to determine the primes that do not divide any central trinomial coefficient; i.e., those in \seqnum{A113305} of \cite{OEIS}, which we conjectured in the previous section to be the ones that answer the question of Problem~\ref{mot_rec}. We can also give the following sufficient condition for ${\bf t}_p = (T_n \bmod p)_{n \geq 0}$ to be uniformly recurrent. For $i =0,\ldots,p-1$, let $\tau_i = T_i \bmod p$. \begin{theorem} Let $p$ be prime and let $\Sigma = \{\tau_i : i = 0,\ldots,p-1\}$. If $\Sigma$ does not contain $0$ but does contain a primitive root modulo $p$, then ${\bf t}_p$ is uniformly recurrent. \end{theorem} \begin{proof} Clearly the order of the product in Theorem~\ref{tri_lucas} does not matter; it follows then that Theorem~\ref{tri_lucas} holds for the most-significant-digit-first representation of $n$, as well as the least-significant-digit-first representation. If we consider Theorem~\ref{tri_lucas} with $n$ written in msd-first notation, we see that ${\bf t}_p$ is generated by iterating the morphism $f : \Sigma^* \to \Sigma^*$ defined by $$f(\tau_i) = (\tau_i\tau_0 \bmod p)(\tau_i\tau_1 \bmod p)\cdots (\tau_i\tau_{p-1} \bmod p)$$ for $i = 0,\ldots,p-1$; i.e., ${\bf t}_p = f^\omega(1)$. Recall that if there exists $t$ such that for every $a,b \in \Sigma$ the word $f^t(a)$ contains $b$, we say that $f$ is a \emph{primitive morphism}. Now $\tau_0=1$, so for $i=0,\ldots,p-1$, we can write $f(\tau_i)=\tau_i x_i$ for some word $x_i$. It follows that $f^p(\tau_i) = \tau_i f(x_i) f^2(x_i)\cdots f^{p-1}(x_i)$. Furthermore, if $0 \notin \Sigma$ and $x_0$ contains a primitive root modulo $p$, then for every $i$, each non-zero residue modulo $p$ appears in one of $\tau_i, f(x_i), f^2(x_i), \ldots, f^{p-1}(x_i)$. This proves that the morphism $f$ is primitive. A standard result from the theory of morphic sequences states that any fixed point of a primitive morphism is uniformly recurrent \cite[Theorem~10.9.5]{AS03}. \end{proof} \begin{example} For $p=5$, we have $(T_0,T_1,T_2,T_3,T_4) = (1,1,3,7,19)$, so $(\tau_0,\tau_1,\tau_2,\tau_3,\tau_4) = (1,1,3,2,4)$ contains the primitive root $2$. The word $${\bf t}_5 = 113241132433412221434423111324\cdots$$ is uniformly recurrent and is equal to $f^\omega(1)$, where $f$ is the morphism defined by \begin{align*} 1 &\to 11324\\ 2 &\to 22143\\ 3 &\to 33412\\ 4 &\to 44231. \end{align*} \end{example} A computer calculation shows that for each prime $p$ appearing in the list of initial values $2,5,11,13,\ldots,479$ of \seqnum{A113305}, the first $p$ terms of ${\bf t}_p$ always contain a primitive root modulo $p$. Hence, each of these ${\bf t}_p$'s are uniformly recurrent. \section{Catalan numbers} Alter and Kubota \cite{AK73} studied the sequences ${\bf c}_p = (C_n \bmod p)_{n \geq 0}$, where $p$ is prime. They proved that the runs of $0$'s in ${\bf c}_p$ have lengths \begin{equation}\label{cp_0runs} \frac{p^{m+1+\delta_{3p}}-3}{2}, \end{equation} where $\delta_{3p}$ is $1$ when $p=3$ and $0$ otherwise. This implies, of course, that for every prime $p$, the sequence ${\bf c}_p$ is not uniformly recurrent. Alter and Kubota also proved that the blocks of non-zero values in ${\bf c}_p$ have length $$\frac{p+3(1+2\delta_{3p})}{2}.$$ For $p=3$, Deutsch and Sagan~\cite[Theorem~5.2]{DS06} gave a complete characterization of ${\bf c}_3$. We can obtain a similar characterization using Walnut. \begin{theorem}[Deutsch and Sagan]\label{c3_0runs} The runs of $0$'s in ${\bf c}_3$ begin at positions $n$, where either $$(n)_3 \in 211^* \text{ or } (n)_3 \in 211^*0\{0,1\}^*,$$ and have length $(3^{i+2}-3)/2$, where $i$ is the length of the leftmost block of $1$'s in $(n)_3$. The blocks of non-zero values in ${\bf c}_3$ are given by the following: \begin{itemize} \item The block $11222$ occurs at position $0$. \item The block $111222$ occurs at all positions $n$ where $(n)_3 \in 222^*0w$ for some $w \in \{0,1\}^*$ that contains an odd number of $1$'s. \item The block $222111$ occurs at all positions $n$ where $(n)_3 \in 222^*0w$ for some $w \in \{0,1\}^*$ that contains an even number of $1$'s. \end{itemize} \end{theorem} \begin{proof} We use the automaton for ${\bf c}_3$ given in Figure~\ref{CAT3}. \begin{figure}[htb] \centering \includegraphics[scale=0.75]{CAT3.pdf} \caption{Automaton for $C_n \bmod 3$}\label{CAT3} \end{figure} The Walnut command \begin{verbatim} eval cat3max0 "?lsd_3 n>=1 & (At t<n => CAT3[i+t]=@0) & CAT3[i+n]!=@0 & (i=0|CAT3[i-1]!=@0)": \end{verbatim} produces the automaton in Figure~\ref{cat3max0}. \begin{figure}[htb] \centering \includegraphics[scale=0.75]{cat3max0.pdf} \caption{Automaton for runs of $0$'s in ${\bf c}_3$}\label{cat3max0} \end{figure} Examining the transition labels of the first component of the input gives the claimed representation for the starting positions of the runs of $0$'s and examining the transition labels of the second component gives the claimed length. For the blocks of non-zero values, we execute the Walnut commands \begin{verbatim} eval cat3max12 "?lsd_3 n>=1 & (At t<n => CAT3[i+t]!=@0) & CAT3[i+n]=@0 & (i=0|CAT3[i-1]=@0)": eval cat3_111222 "?lsd_3 $cat3max12(i,6) & CAT3[i]=@1 & CAT3[i+1]=@1 & CAT3[i+2]=@1 & CAT3[i+3]=@2 & CAT3[i+4]=@2 & CAT3[i+5]=@2": eval cat3_222111 "?lsd_3 $cat3max12(i,6) & CAT3[i]=@2 & CAT3[i+1]=@2 & CAT3[i+2]=@2 & CAT3[i+3]=@1 & CAT3[i+4]=@1 & CAT3[i+5]=@1": eval cat3all12 "?lsd_3 Ai,n $cat3max12(i,n) => (i=0 | $cat3_111222(i) | $cat3_222111(i))": \end{verbatim} to obtain the automata in Figures~\ref{cat3_111222} and \ref{cat3_222111}. \begin{figure}[htb] \centering \includegraphics[scale=0.75]{cat3_111222.pdf} \caption{Automaton for blocks $111222$ in ${\bf c}_3$}\label{cat3_111222} \end{figure} \begin{figure}[htb] \centering \includegraphics[scale=0.75]{cat3_222111.pdf} \caption{Automaton for blocks $222111$ in ${\bf c}_3$}\label{cat3_222111} \end{figure} \end{proof} Note that the length of the runs given in Theorem~\ref{c3_0runs} is exactly what is given by the result of Alter and Kubota stated above in Eq.~\eqref{cp_0runs}. We can also perform the same calculation for $p=5$ to obtain \begin{theorem} The runs of $0$'s in ${\bf c}_5$ begin at positions $n$, where either $$(n)_5 \in 32^* \text{ or } (n)_5 \in 32^*\{0,1\}\{0,1,2\}^*,$$ and have length $(5^{i+2}-3)/2$, where $i$ is the length of the leftmost block of $2$'s in $(n)_5$. \end{theorem} \begin{proof} We use the automaton for ${\bf c}_5$ given in Figure~\ref{CAT5}. \begin{figure}[htb] \centering \includegraphics[scale=0.65]{CAT5.pdf} \caption{Automaton for $C_n \bmod 5$}\label{CAT5} \end{figure} The Walnut command \begin{verbatim} eval cat5max0 "?lsd_5 n>=1 & (At t<n => CAT5[i+t]=@0) & CAT5[i+n]!=@0 & (i=0|CAT5[i-1]!=@0)": \end{verbatim} produces the automaton in Figure~\ref{cat5max0}. \begin{figure}[htb] \centering \includegraphics[scale=0.75]{cat5max0.pdf} \caption{Automaton for runs of $0$'s in ${\bf c}_5$}\label{cat5max0} \end{figure} Examining the transition labels, as in the proof of Theorem~\ref{c3_0runs}, gives the result. \end{proof} Again, note that the lengths of the runs match what is given by Eq.~\eqref{cp_0runs}. \begin{theorem} The sequence $c_5$ begins with the non-zero block $112$. The other non-zero blocks in $c_5$ are $1331$, $2112$, $3443$, and $4224$. \end{theorem} \begin{proof} The Walnut command \begin{verbatim} eval cat5max1234 "?lsd_5 n>=1 & (At t<n => CAT5[i+t]!=@0) & CAT5[i+n]=@0 & (i=0|CAT5[i-1]=@0)": \end{verbatim} produces the automaton in Figure~\ref{cat5max1234}. \begin{figure}[htb] \centering \includegraphics[scale=0.75]{cat5max1234.pdf} \caption{Automaton for non-zero blocks in ${\bf c}_5$}\label{cat5max1234} \end{figure} We see that the initial non-zero block has length $3$ and all others have length $4$. We omit the Walnut command to verify the values of these length $4$ blocks, but it is easy to formulate. \end{proof} \section{Conclusion} We have shown how to use Walnut to obtain automated proofs of certain results in the literature concerning the Catalan and Motzkin numbers modulo $p$, as well as the central trinomial coefficients modulo $p$. We were also able to use Walnut to examine other properties of these sequences that have not previously been explored, such as the presence (or absence) of certain repetitive patterns and the property of being uniformly recurrent. We hope these results encourage other researchers to continue to further explore these properties for other sequences.

https://arxiv.org/abs/1111.0433

A closed-form approximation for the median of the beta distribution

A simple closed-form approximation for the median of the beta distribution Beta(a, b) is introduced: (a-1/3)/(a+b-2/3) for (a,b) both larger than 1 has a relative error of less than 4%, rapidly decreasing to zero as both shape parameters increase.

\section{Introduction} \label{s1} Consider the the beta distribution $\mathrm{Beta}(a,b)$, with the density function, $$ \frac{\Gamma(a+b)}{\Gamma(a)\Gamma(b)} \theta^{a-1}(1-\theta)^{b-1}. $$ The mean of $\mathrm{Beta}(a, b)$ is readily obtained by the formula $a/(a+b)$, but there is no general closed formula for the median. The median function, here denoted by $m(a,b)$, is the function that satisfies, $$ \frac{\Gamma(a+b)}{\Gamma(a)\Gamma(b)} \int_0^{m(a,b)}\theta^{a-1}(1-\theta)^{b-1} \mathrm{d}\theta = \frac{1}{2}. $$ The relationship $m(a,b)=1-m(b, a)$ holds. Only for the special cases $a=1$ or $b=1$ we may obtain an exact formula: $m(a, 1)=2^{-1/a}$ and $m(1, b)=1-2^{-1/b}$. Moreover, when $a=b$, the median is exactly $1/2$. There has been much literature about the incomplete beta function and its inverse (see e.g. \citet{Dutka:1981} for a review). The focus in literature has been on finding accurate numerical results, but a simple and practical approximation that is easy to compute has not been found. \begin{figure}[t] \begin{center} \scalebox{0.80}{\includegraphics{fig-betaerr}} \caption{\label{fig-betaerrors} Relative errors of the approximation $(a-1/3)/(a+b-2/3)$ of the median of the $\mathrm{Beta}(a, b)$ distribution, compared with the numerically computed value for several fixed $p=a/(a+b)<1/2$. The horizontal axis shows the shape parameter $a$ on logarithmic scale. From left to right, $p=0.499$, 0.49, 0.45, 0.35, 0.25, and 0.001. } \end{center} \end{figure} \section{A new closed-form approximation for the median} Trivial bounds for the median can be derived \citep{Payton:1989}, which are a consequence of the more general mode-median-mean inequality \citep{Groeneveld:Meeden:1977}. In the case of the beta distribution with $1<a<b$, the median is bounded by the mode $(a-1)/(a+b-2)$ and the mean $a/(a+b)$: $$ \frac{a-1}{a+b-2} \le m(a,b) \le \frac{a}{a+b}. $$ For $a\le1$ the formula for the mode does not hold as there is no mode. If $1<b<a$, the order of the inequality is reversed. Equality holds if and only if $a=b$; in this case the mean, median, and mode are all equal to $1/2$. This inequality shows that if the mean is kept fixed at some $p$, and one of the shape parameters is increased, say $a$, then the median is sandwiched between $p(a-1)/(a-2p)$ and $p$, hence the median tends to $p$. From the formulas for the mode and mean, it can be conjectured that the median $m(a,b)$ could be approximated by $m(a,b;d)=(a-d)/(a+b-2d)$ for some $d\in(0,1)$, as this form would satisfy the above inequality while agreeing with the symmetry requirement, that is, $m(a,b;d)=1-m(b,a;d)$. \begin{figure}[t] \begin{center} \scalebox{0.80}{\includegraphics{fig-betaerrp}} \caption{\label{fig-betaerrp} Relative errors of the approximation $(a-1/3)/(a+b-2/3)$ of the median of the $\mathrm{Beta}(a, b)$ distribution over the whole range of possible distribution means $p=a/(a+b)$. The smaller of the shape parameters is fixed, i.e. for $p\le 0.5$, the median is computed for $\mathrm{Beta}(a, a(1-p)/p)$ and for $p>0.5$, the median is computed for $\mathrm{Beta}(bp/(1-p), b)$. } \end{center} \end{figure} Since a $\mathrm{Beta}(a,b)$ variate can be expressed as the ratio $\gamma_1/(\gamma_1+\gamma_2)$ where $\gamma_1\sim\mathrm{Gamma}(a)$ and $\gamma_2\sim\mathrm{Gamma}(b)$ (both with unit scale), it is useful to have a look at the median of the gamma distribution. \citet{Berg:Pedersen:2006} studied the median function of the unit-scale gamma distribution median function, denoted here by $M(a)$, for any shape parameter $a>0$, and obtained $M(a) = a - 1/3 + o(1)$, rapidly approaching $a-1/3$ as $a$ increases. It can therefore be conjectured that the distribution median may be approximated by, \begin{equation}\label{eq-beta} m(a, b) \approx m(a, b; 1/3) = \frac{a-1/3}{(a-1/3)+(b-1/3)} = \frac{a-1/3}{a+b-2/3}. \end{equation} Figure (\ref{fig-betaerrors}) shows that this approximation indeed appears to approach the numerically computed median asymptotically for all distribution means $p=a/(a+b)$ as the (smaller) shape parameter $a\to\infty$. For $a\ge1$, the relative error is less than 4\%, and for $a\ge2$ this is already less than 1\%. \begin{figure}[t] \begin{center} \scalebox{0.80}{\includegraphics{fig-betadisterr}} \caption{\label{fig-betadisterr} Logarithm of the scaled absolute error (distance) $\log(|m(a,b;d)-m(a,b)|/p)$, computed for a fixed distribution mean $p=0.01$ and various $d$. The approximate median of the $\mathrm{Beta}(a,b)$ distribution is defined as $m(a,b;d)=(a-d)/(a+b-2d)$. Due to scaling of the error, the graph and its scale will not essentially change even if the error is computed for other values of $p<0.5$. The approximation $m(a,b;1/3)$ performs the most consistently, attaining the lowest absolute error eventually as the precision of the distribution increases. } \end{center} \end{figure} Figure (\ref{fig-betaerrp}) shows the relative error over all possible distribution means $p=a/(a+b)$, as the smallest of the two shape parameters varies from $1$ to $4$. This illustrates how the relative error tends uniformly to zero over all $p$ as the shape parameters increase. The figure also shows that the formula consistently either underestimates or overestimates the median depending on whether $p<0.5$ or $p>0.5$. However, the function $m(a,b;d)$ approximates the median fairly accurately if some other $d$ close to $1/3$ (say $d=0.3$) is chosen. Figure (\ref{fig-betadisterr}) displays curves of the logarithm of the absolute difference from the numerically computed median for a fixed $p=0.01$, as the shape parameter $a$ increases. The absolute difference has been scaled by $p$ before taking the logarithm: due to this scaling, the error stays approximately constant as $p$ decreases so the picture and its scale will not essentially change even if the error is computed for other values of $p<0.5$. The figure shows that although some approximations such as $d=0.3$ has a lower absolute error for some $a$, the error of $m(a, b; 1/3)$ tends to be lower in the long run, and moreover performs more consistently by decreasing at the same rate on the logarithmic scale. In practical applications, $d=0.333$ should be a sufficiently good approximation of $d=1/3$. \begin{figure}[t] \begin{center} \scalebox{0.7}{\includegraphics{fig-betatail}} \caption{\label{fig-betatail} Tail probabilities $\Pr(\theta<m)$ of the $\mathrm{Beta}(a, b)$ distribution when $m=(a-1/3)/(a+b-2/3)$. As the smaller of the two shape parameters increases, the tail probability tends rapidly and uniformly to $0.5$. } \end{center} \end{figure} Another measure of the accuracy is the tail probability $\Pr(\theta \le m(a,b;1/3))$ of a $\mathrm{Beta}(a, b)$ variate $\theta$: good approximators of the median should yield probabilities close to $1/2$. Figure (\ref{fig-betatail}) shows that as long as the smallest of the shape parameters is at least 1, the tail probability is bound between $0.4865$ and $0.5135$. As the shape parameters increase, the probability tends rapidly and uniformly to $0.5$. Finally, let us have a look at a well-known paper that provides further support for the uniqueness of $m(a,b;1/3)$. \citet{Peizer:Pratt:1968} and \citet{Pratt:1968} provide approximations for the probability function $\Pr(\theta\le x)$ of a $\mathrm{Beta}(a,b)$ variate $\theta$. Although they do not provide a formula for the inverse, it is the probability function at the approximate median. According to \citet{Peizer:Pratt:1968}, $\Pr(\theta\le x)$ is well approximated by $\Phi(z(a,b; x))$ where $\Phi$ is the standard normal probability function, and $z$ is a function of the shape parameters and the quantile $x$. Consider $m=m(a,b;d)$: $z(a,b;m)$ should be close to zero and at least tend to zero fast as $a$ and $b$ increase. Now assume that $p$ is fixed, $a$ varies and $b=a(1-p)/p$. The function $z(a, b; m)$ equals, rewritten with the notation in this paper, \begin{equation}\label{eq-peizer-beta} \sqrt{p}\frac{1-2m}{(a-p)^{1/2}}\left( 1/3-d - \frac{0.02p}{a}\left[ \frac{1}{2} + \frac{1-dp/a}{p(1-p)} \right] \right) \left(\frac{1+f(a,p;d)}{m(1-m)}\right)^{1/2}, \end{equation} where the function $f(a,p;d)$ tends to zero as $a$ increases, being exactly zero only when $d=1/2$ or $m=1/2$. It is evident that for the fastest convergence rate to zero, one should choose $d=1/3$. This is of the order $O(a^{-3/2})$; if $d\ne 1/3$, for example if we choose the mean $p$ as the approximation of the median ($d=0$), the rate is at most $O(a^{-1/2})$.

https://arxiv.org/abs/0708.2643

On fixed points of permutations

The number of fixed points of a random permutation of 1,2,...,n has a limiting Poisson distribution. We seek a generalization, looking at other actions of the symmetric group. Restricting attention to primitive actions, a complete classification of the limiting distributions is given. For most examples, they are trivial -- almost every permutation has no fixed points. For the usual action of the symmetric group on k-sets of 1,2,...,n, the limit is a polynomial in independent Poisson variables. This exhausts all cases. We obtain asymptotic estimates in some examples, and give a survey of related results.

\section{Introduction} \label{intro} One of the oldest theorems in probability theory is the Montmort (1708) limit theorem for the number of fixed points of a random permutation of $\{1, 2, \ldots, n\}$. Let $S_n$ be the symmetric group. For an element $w \in S_n$, let $A(w) = \{i : w(i) = i\}$. Montmort \cite{monmort} proved that \refstepcounter{thm} \begin{equation} \label{eqn0} {|\{w : A(w) = j\}| \over n!} \rightarrow {1 \over e} \ {1 \over j!} \end{equation} for $j$ fixed as $n$ tends to infinity. The limit theorem (\ref{eqn0}) has had many refinements and variations. See Tak\'{a}cs \cite{Ta} for its history, Chapter 4 of Barbour, Holst, Janson \cite{BHJ} or Chatterjee, Diaconis, Meckes \cite{CDM} for modern versions. The limiting distribution $P_{\lambda}(j) = e^{-\lambda} \lambda^j / j!$ (in (\ref{eqn0}) $\lambda = 1$) is the Poisson distribution of ``the law of small numbers''. Its occurrence in many other parts of probability (see e.g. Aldous \cite{Al}) suggests that we seek generalizations of (\ref{eqn0}), searching for new limit laws. In the present paper we look at other finite sets on which $S_n$ acts. It seems natural to restrict to transitive action -- otherwise, things break up into orbits in a transparent way. It is also natural to restrict to primitive actions. Here $S_n$ acts {\it primitively} on the finite set $\Omega$ if we cannot partition $\Omega$ into disjoint blocks $\Delta_1, \Delta_2, \ldots, \Delta_h$ where $S_n$ permutes the blocks (if $\Delta_i^w \cap \Delta_j \not = \phi$ then $\Delta_i^w = \Delta_j)$. The familiar wreath products which permute within blocks and between blocks are an example of an imprimitive action. The primitive actions of $S_n$ have been classified in the O'Nan-Scott theorem. We describe this carefully in Section \ref{Onan}. For the study of fixed points most of the cases can be handled by a marvelous theorem of Luczak-Pyber \cite{LP}. This shows that, except for the action of $S_n$ on $k$-sets of an $n$ set, almost all permutations have no fixed points (we say $w$ is a derangement). This result is explained in Section \ref{lucpyb}. For $S_n$ acting on $k$-sets, one can assume that $k < n/2$, and there is a nontrivial limit if and only if $k$ stays fixed as $n$ tends to infinity. In these cases, the limit is shown to be an explicit polynomial in independent Poisson random variables. This is the main content of Section \ref{ksets}. Section \ref{matchings} works out precise asymptotics for the distribution of fixed points in the action of $S_n$ on matchings. Section \ref{impriv} considers more general imprimitive subgroups. Section \ref{prim} proves that the proportion of elements of $S_n$ which belong to a primitive subgroup not containing $A_n$ is at most $O(n^{-2/3+\alpha})$ for any $\alpha>0$; this improves on the bound of Luczak and Pyber \cite{LP}. Finally, Section \ref{survey} surveys related results (including analogs of our main results for finite classical groups) and applications of the distribution of fixed points and derangements. If a finite group $G$ acts on $\Omega$ with $F(w)$ the number of fixed points of $w$, the ``lemma that is not Burnside's'' implies that $$ \begin{aligned} E(F(w)) &= \# \ \mbox{orbits of} \ G \ \mbox{on} \ \Omega \\ E(F^2 (w)) &= \# \ \mbox{orbits of} \ G \ \mbox{on} \ \Omega \times \Omega = \ \mbox{rank} := r. \end{aligned} $$ If $G$ is transitive on $\Omega$ with isotropy group $H$, then the rank is also the number of orbits of $H$ on $\Omega$ and so equal to the number of $H-H$ double cosets in $G$. Thus for transitive actions \refstepcounter{thm} \begin{equation} E(F(w)) = 1, \quad {\rm Var} (F(w)) = \ \mbox{rank}-1 \end{equation} In most of our examples $P(F(w) = 0) \rightarrow 1$ but because of (1.2), this cannot be seen by moment methods. The standard second moment method (Durrett \cite{Du}, page 16) says that a non-negative integer random variable satisfies $P(X > E(X)/2) \geq {1 \over 4} (EX)^2 / E(X^2)$. Specializing to our case, $P(F(w) > 0) \geq 1/(4 r)$; thus $P(F(w) = 0) \leq 1 - 1/(4 r)$. This shows that the convergence to $1$ cannot be too rapid. There is also a quite easy lower bound for $P(F(w)=0)$ \cite{GW}. Even the simplest instance of this lower bound was only observed in 1992 in \cite{CaCo}. We reproduce the simple proof from \cite{GW}. Let $n=|\Omega|$ and let $G_0$ be the set of elements of $G$ with no fixed points. Note that $F(w) \le n$, whence $$ \sum_G (F(w)-1)(F(w)-n) \le \sum_{G_0} (F(w)-1)(F(w)-n) = n|G_0|. $$ On the other hand, the left hand side is equal to $|G|(r -1)$. Thus, $P(F(w)=0) \ge (r-1)/n$. We record these bounds. \begin{theorem} \label{basic} Let $G$ be a finite transitive permutation group of degree $n$ and rank $r$. Then $$ \frac{r-1}{n} \le P(F(w)=0) \le 1 - \frac{1}{4r}. $$ \end{theorem} Frobenius groups of order $n(n-1)$ with $n$ a prime power are only the possibilities when the lower bound is achieved. The inequality above shows that $P(F(w)=0)$ tends to $1$ implies that the rank tends to infinity. Indeed, for primitive actions of symmetric and alternating groups, this is also a sufficient condition -- see Theorem \ref{theC}. \section{O'Nan-Scott Theorem} \label{Onan} Let $G$ act transitively on the finite set $\Omega$. By standard theory we may represent $\Omega = G/ G_{\alpha}$, with any fixed $\alpha \in \Omega$. Here $G_{\alpha} = \{w : \alpha^w = \alpha\}$ with the action being left multiplication on the cosets. Further (Passman \cite[3.4]{P}) the action of $G$ on $\Omega$ is primitive if and only if the isotropy group $G_{\alpha}$ is maximal. Thus, classifying primitive actions of $G$ is the same problem as classifying maximal subgroups $H$ of $G$. The O'Nan-Scott theorem classifies maximal subgroups of $A_n$ and $S_n$ up to determining the almost simple primitive groups of degree $n$. \begin{theorem} {\rm [O'Nan-Scott]} Let $H$ be a maximal subgroup of $G=A_n$ or $S_n$. Then, one of the following three cases holds: \begin{description} \item [I] $H$ acts primitively as a subgroup of $S_n$ (primitive case), \item [II] $H = (S_a \wr S_b) \cap G$ (wreath product), $n = a \cdot b, | \Omega | = \frac{n!}{(a!)^b \cdot b!}$ (imprimitive case), or \item [III] $H = (S_k \times S_{n-k}) \cap G, | \Omega | = {n \choose k}$ with $1 \le k < n/2$ (intransitive case). \end{description} Further, in case I, one of the following holds: \begin{description} \item [Ia] $H$ is almost simple, \item [Ib] $H$ is diagonal, \item [Ic] $H$ preserves product structure, or \item [Id] $H$ is affine. \end{description} \end{theorem} {\it Remarks and examples:} \begin{enumerate} \item Note that in cases I, II, III, the modifiers `primitive', `imprimitive', `intransitive' apply to $H$. Since $H$ is maximal in $G$, $\Omega \cong G/H$ is a primitive $G$-set. We present an example and suitable additional definitions for each case. \item In case III, $\Omega$ is the $k$-sets of $\{1, 2, \ldots, n\}$ with the obvious action of $S_n$. This case is discussed extensively in Section \ref{ksets} below. \item In case II, take $n$ even with $a = 2, b = n/2$. We may identify $\Omega$ with the set of perfect matchings on $n$ points -- partitions of $n$ into $n/2$ two-element subsets where order within a subset or among subsets does not matter. For example if $n = 6, \{1,2\} \{3,4\} \{5,6\}$ is a perfect matching. For this case, $|\Omega| = \frac{n!}{2^{n/2} (n/2)!} = (2 n-1)(2n-3) \cdots (1)$. Careful asymptotics for this case are developed in Section \ref{matchings}. More general imprimitive subgroups are considered in Section \ref{impriv}. \item While every maximal subgroup of $A_n$ or $S_n$ falls into one of the categories of the O'Nan-Scott theorem, not every group is maximal. A complete list of the exceptional examples is in Liebeck, Praeger and Saxl \cite{LPS1}. \item In case Ia, $H$ is {\it almost simple} if for some non-abelian simple group $G$, $G \leq H \leq \Aut(G)$. For example, fix $1 < k < m$. Let $n = {m \choose k}$. Let $S_n$ be all $n!$ permutations of the $k$ sets of $\{1, 2, \ldots, m\}$. Take $S_m \leq S_n$ acting on the $k$-sets in the usual way. For $m \geq 5$, $S_m$ is almost simple and primitive. Here $\Omega = S_n / S_m$ does not have a simple combinatorial description, but this example is crucial and the $k=2$ case will be analyzed in Section \ref{prim}. Let $\tau \in S_m$ be a transposition. Then $\tau$ moves precisely $ 2 \binom{m-2}{k-1}$ elements of $\Omega$. Thus, $S_m$ embeds in $A_n$ if and only if $\binom{m-2}{k-1}$ is even. Indeed for most primitive embeddings of $S_m$ into $S_n$, the image is contained in $A_n$ \cite{wisc}. It is not difficult to see that the image of $S_m$ is maximal in either $A_n$ or $S_n$. This follows from the general result in \cite{LPS1}. It also follows from the classification of primitive groups containing a non-trivial element fixing at least $n/2$ points \cite{GM}. Similar examples can be constructed by looking at the action of $P\Gamma L_d (q)$ on $k$-spaces (recall the $P\Gamma L_d(q)$ is the projective group of all semilinear transformations of a $d$ dimensional vector space over $\F_q$). All of these are covered by case $Ia$. \item In case Ib, {\it $H$ is diagonal} if $H = G^k \cdot (\Out(G) \times S_k)$ for $G$ a non-abelian simple group, $k \geq 2$ (the dot denotes semidirect product). Let $\Omega = G^k / D$ with $D = \{(g, g, \ldots g)\}_{g \in G}$ the diagonal subgroup. Clearly $G^k$ acts on $\Omega$. Let $\Out(G)$ (the outer automorphisms) act coordinate-wise and let $S_k$ act by permuting coordinates. These transformations determine a permutation group $H$ on the set $\Omega$. The group $H$ has normal subgroup $G^k$ with quotient isomorphic to $\Out(G) \times S_k$. The extension usually splits (but it doesn't always split). Here is an specific example. Take $G = A_m$ for $m \geq 8$ and $k = 2$. Then $\Out(A_m) = C_2$ and so $H = \langle A_m \times A_m, \tau, (s,s) \rangle$ where $s$ is a transposition (or any element in $S_m$ outside of $A_m$) and $\tau$ is the involution changing coordinates. More precisely, each coset of $D$ has a unique representative of the form $(1, x)$. We have $(g_1, g_2) (1, x)D = (g_1, g_2 x)D = (1,g_2 x g_1^{-1})D$. The action of $\tau \in C_2$ takes $(1, x) \rightarrow (1, x^{-1})$ and the action of $(s,s) \in \Out(A_m)$ takes $(1, x)$ to $(1, sxs^{-1})$. The maximality of $H$ is somewhat subtle. We first show that if $m \geq 8$, then $H$ is contained in $\Alt(\Omega)$. Clearly $A_m \times A_m$ is contained in $\Alt(\Omega)$. Observe that $(s,s)$ is contained in $\Alt(\Omega)$. Indeed, taking $s$ to be a transposition, the number of fixed points of $(s,s)$ is the size of its centralizer in $A_m$ which is $|S_{m-2}|$, and so $\frac{m!}{2}-(m-2)!$ points are moved and this is divisible by $4$ since $m \geq 8$. To see that $\tau$ is contained in $Alt(\Omega)$ for $m \geq 8$, note that the number of fixed points of $\tau$ is the number of involutions (including the identity) in $A_m$, so it is sufficient to show that $\frac{m!}{2}$ minus this number is a multiple of 4. This follows from the next proposition, which is of independent combinatorial interest. \begin{prop} \label{countinv} Suppose that $m \geq 8$. Then the number of involutions in $A_m$ and the number of involutions in $S_m$ are multiples of $4$. \end{prop} \begin{proof} Let $a(m)$ be the number of involutions in $A_m$ (including the identity). Let $b(m)$ be the number of involutions in $S_m-A_m$. It suffices to show that $a(m)=b(m) = 0 \mod 4$. For $n = 8,9$ we compute directly. For $n > 9$, we observe that \[ a(n) = a(n-1) + (n-1)b(n-2) \] and \[ b(n) = b(n-1) + (n-1)a(n-2) \] (because an involution either fixes 1 giving the first term or swaps 1 with $j > 1$, giving rise to the second term). The result follows by induction. \end{proof} Having verified that $H$ is contained in $\Alt(\Omega)$ for $m \geq 8$, maximality now follows from Liebeck-Praeger-Saxl \cite{LPS1}. \item In case Ic, {\it $H$ preserves a product structure}. Let $\Gamma=\{1,...,m\}$, $\Delta=\{1,...,t\}$, and let $\Omega$ be the $t$-fold Cartesian product of $\Gamma$. If $C$ is a permutation group on $\Gamma$ and $D$ is a permutation group on $\Delta$, we may define a group $H = C \wr D$ by having $C$ act on the coordinates, and having $D$ permute the coordinates. Primitivity of $H$ is equivalent to $C$ acting primitively on $\Gamma$ with some non identity element having a fixed point and $D$ acting transitively on $\Delta$ (see, e.g. Cameron \cite{Ca1}, Th. 4.5). There are many examples of case Ic but $|\Omega| = m^t$ is rather restricted and $H$ has a simple form. One specific example is as follows: $G=S_{m^t}$, $H=S_m \wr S_t$ and $\Omega$ is the t-fold Cartesian product $\{1,\cdots,m\}^t$. The case $t=2$ will be analyzed in detail in Section \ref{prim}. It is easy to determine when $H$ embeds in $A_{m^t}$. We just note that if $t=2$, then this is case if and only if $4|m$. \item In case Id $H$ is affine. Thus $\Omega = V$, a vector space of dimension $k$ over a field of $q$ elements (so $n= |\Omega | = q^k$) and $H$ is the semidirect product $V \cdot GL(V)$. Since we are interested only in maximal subgroups, $q$ must be prime. Note that if $q$ is odd, then $H$ contains an $n-1$ cycle and so is not contained in $A_n$. If $q=2$, then for $k > 2$, $H$ is perfect and so is contained in $A_n$. The maximality of $H$ in $A_n$ or $S_n$ follows by Mortimer \cite{mortimer} for $k >1$ and \cite{gurkim} if $k=1$. \item The proof of the O'Nan-Scott theorem is not extremely difficult. O'Nan and Scott each presented proofs at the Santa Cruz Conference in 1979. There is a more delicate version which describes all primitive permutation groups. This was proved in Aschbacher-Scott \cite{AS} giving quite detailed information. A short proof of the Aschbacher-O'Nan Scott Theorem is in \cite{msri}. See also Liebeck, Praeger and Saxl \cite{LPS2}). A textbook presentation is in Dixon and Mortimer \cite{DxM}. We find the lively lecture notes of Cameron (\cite{Ca1}, Chapter 4) very helpful. The theorem has a life of its own, away from permutation groups, in the language of the generalized Fitting subgroup $F^*$. See Kurtzweil and Stellmacher \cite{KS}. See also the lively lecture notes of Cameron (\cite{Ca1}, Chapter 4). The notion of generalized Fitting subgroup is quite useful in both the proof and statement of the theorem. See Kurtzweil and Stellmacher \cite{KS}. \item It turns out that many of the details above are not needed for our main results. Only case III $(H = S_k \times S_{n-k})$ allows non-trivial limit theorems. This is the subject of the next section. The other cases are of interest when we try to get explicit bounds (Sections \ref{matchings}, \ref{impriv}, \ref{prim}). \end{enumerate} \section{Two Theorems of Luczak-Pyber} \label{lucpyb} The following two results are due to Luczak and Pyber. \begin{theorem} \label{theA} (\cite{LP}) Let $S_n$ act on $\{1, 2, \ldots, n\}$ as usual and let $i(n, k)$ be the number of $w \in S_n$ that leave some $k$-element set invariant. Then, ${i (n, k) \over n!} \leq a k^{-.01}$ for an absolute constant $a$. \end{theorem} \begin{theorem} \label{theB} (\cite{LP}) Let $t_n$ denote the number of elements of the symmetric group $S_n$ which belong to transitive subgroups different from $S_n$ or $A_n$. Then $$ \lim_{n \rightarrow \infty} t_n / n! = 0. $$ \end{theorem} Theorem \ref{theA} is at the heart of the proof of Theorem \ref{theB}. We use them both to show that a primitive action of $S_n$ is a derangement with probability approaching one, unless $S_n$ acts on $k$-sets with fixed $k$. Note that we assume that $k \leq n/2$ since the action on $k$-sets is isomorphic to the action on $n-k$ sets. \begin{theorem} \label{theC} Let $G_i$ be a finite symmetric or alternating group of degree $n_i$ acting primitively on a finite set $\Omega_i$ of cardinality at least $3$. Assume that $n_i \rightarrow \infty$. Let $d_i$ be the proportion of $w \in G_i$ with no fixed points. Then the following are equivalent: \begin{enumerate} \item $ \lim_{i \rightarrow \infty} d_i = 1, $ \item there is no fixed $k$ with $\Omega_i= \{k-\mbox{sets of} \ \{1, 2, \ldots, n_i\}\}$ for infinitely many $i$, and \item the rank of $G_i$ acting on $\Omega_i$ tends to $\infty$. \end{enumerate} \end{theorem} \begin{proof} Let $H_i$ be an isotropy group for $G_i$ acting on $\Omega_i$. If $H_i$ falls into category I or II of the O'Nan-Scott theorem, $H_i$ is transitive. Writing out Theorem \ref{theB} above more fully, Luczak-Pyber prove that $$ {\left| \bigcup_H H \right| \over n!} \rightarrow 0 $$ where the union is over {\it all} transitive subgroups of $S_n$ not equal to $S_n$ or $A_n$. Thus a randomly chosen $w \in S_n$ is not in $x H_n x^{-1}$ for any $x$ if $H_n$ falls into category I or II. Having ruled out categories I and II, we turn to category III ($k$-sets of an $n$ set). Here, Theorem \ref{theA} shows the chance of a derangement tends to one as $a/k^{.01}$, for an absolute constant $a$. The previous paragraphs show that (2) implies (1). If the rank does not go to $\infty$, then $d_i$ cannot approach 1 by Theorem \ref{basic}. Thus (1) implies (3), and also (2) since the rank of the action on k-sets is $k+1$. Clearly (3) implies (2), completing the proof. \end{proof} \section{$k$-Sets of an $n$-Set} \label{ksets} In this section the limiting distribution of the number of fixed points of a random permutation acting on $k$-sets of an $n$-set is determined. \begin{theorem} \label{klim} Fix $k$ and let $S_n$ act on $\Omega_{n, k}$ -- the $k$ sets of $\{1, 2, \ldots, n\}$. Let $A_i (w)$ be the number of $i$-cycles of $w \in S_n$ in its usual action on $\{1, 2, \ldots, n\}$. Let $F_k(w)$ be the number of fixed points of $w$ acting on $\Omega_{n, k}$. Then \refstepcounter{thm} \begin{equation} F_k(w) = \sum_{|\lambda|=k} \prod_{i=1}^k {A_i (w) \choose \alpha_i (\lambda)}. \end{equation} Here the sum is over partitions $\lambda$ of $k$ and $\alpha_i (\lambda)$ is the number of parts of $\lambda$ equal to $i$. (2) For all $n \geq 2, E(F_k) = 1, {\rm Var} (F_k) = k$. (3) As $n$ tends to infinity, $A_i(w)$ converge to independent Poisson $(1/i)$ random variables. \end{theorem} \begin{proof} If $w \in S_n$ is to fix a $k$ set, the cycles of $w$ must be grouped to partition $k$. The expression for $F_k$ just counts the distinct ways to do this. See the examples below. This proves (1). The rank of $S_n$ acting on $k$ sets is $k+1$, proving (2). The joint limiting distribution of the $A_i$ is a classical result due to Goncharov \cite{Go}. In fact, letting $X_1,X_2,\cdots,X_k$ be independent Poisson with parameters $1,\frac{1}{2},\cdots,\frac{1}{k}$, one has from \cite{DS} that for all $n \geq \sum_{i=1}^k ib_i$, \[ E \left( \prod_{i=1}^k A_i(w)^{b_i} \right) = \prod_{i=1}^k E(X_i^{b_i}).\] For total variation bounds see \cite{AT}. This proves (3). \end{proof} {\it Examples}. Throughout, let $X_1, X_2, \ldots, X_k$ be independent Poisson random variables with parameters $1, 1/2, 1/3, \ldots, 1/k$ respectively. {$k=1$:} This is the usual action of $S_n$ on $\{1, 2, \ldots, n\}$ and Theorem \ref{klim} yields (1) of the introduction: In particular, for derangements $$ P(F_1 (w) = 0) \rightarrow 1/e \ \dot = \ .36788 . $$ {$k=2$:} Here $F_2(w) = {A_1 (w) \choose 2} + A_2 (w)$ and Theorem \ref{klim} says that $P(F_2(w) = j) \rightarrow P \left( {X_1 \choose 2} + X_2 = j \right)$ with $X_1$ Poisson$(1)$, $X_2$ Poisson$({1 \over 2})$. In particular $$ P(F_2(w) = 0) \rightarrow {2 \over e^{3/2}} \ \dot = \ .44626 . $$ {$k=3$:} Here $F_3(w) = { A_1 (w) \choose 3} + A_1 (w) A_2 (w) + A_3 (w)$ and $$ P(F_3 (w) = j) \rightarrow P\left( {X_1 \choose 3} + X_1 X_2 + X_3 = j \right) . $$ In particular $$ P(F_3(w) = 0) \rightarrow {1 \over e^{4/3}} (1 + {3 \over 2} e^{-1/2}) \ \dot = \ .50342. $$ We make the following conjecture, which has also been independently stated as a problem by Cameron \cite{Ca2}. \\ \noindent {\it Conjecture:} $\lim_{n \rightarrow \infty} P(F_k (w) = 0)$ is increasing in $k$. \\ Using Theorem \ref{klim}, one can prove the following result which improves, in this context, the upper bound given in Theorem \ref{basic}. \begin{prop} \[ lim_{n \rightarrow \infty} P(F_k (w) = 0) \leq 1 - \frac{\log(k)}{k} + O \left( \frac{1}{k} \right). \] \end{prop} \begin{proof} Clearly \begin{eqnarray*} P(F_k(w) > 0) & \geq & P \left(\bigcup_{j=1}^{\lfloor \frac{k-1}{2} \rfloor} (A_{k-j} > 0 \ \mbox{and} \ A_j > 0) \right)\\ & = & 1 - P \left( \bigcap_{j=1}^{\lfloor \frac{k-1}{2} \rfloor} \overline{ (A_{k-j} > 0 \ \mbox{and} \ A_j > 0) } \right). \end{eqnarray*} By Theorem \ref{klim}, this converges to \[ 1 - \prod_{j=1}^{\lfloor \frac{k-1}{2} \rfloor} \left[ 1-(1- e^{-{1/j}}) (1 - e^{-{1/(k-j)}}) \right]. \] Let $a_j=(1-e^{-1/j})$. Write the general term in the product as $e^{\log(1-a_ja_{k-j})}$. Expand the log to $-a_ja_{k-j} + O \left( (a_ja_{k-j})^2 \right)$. Writing $a_j= \left( \frac{1}{j}+O(\frac{1}{j^2}) \right)$ and multiplying out, we must sum \[ \sum_{j=1}^{\lfloor \frac{k-1}{2} \rfloor} \frac{1}{j} \frac{1}{k-j} , \sum_{j=1}^{\lfloor \frac{k-1}{2} \rfloor} \frac{1}{j^2} \frac{1}{k-j} , \sum_{j=1}^{\lfloor \frac{k-1}{2} \rfloor} \frac{1}{(k-j)^2} \frac{1}{k} , \sum_{j=1}^{\lfloor \frac{k-1}{2} \rfloor} \frac{1}{(k-j)^2} \frac{1}{j^2}.\] Writing $\frac{1}{j} \frac{1}{k-j} = \frac{1}{k} \left( \frac{1}{j} + \frac{1}{k-j} \right),$ the first sum is $\frac{\log(k)}{k}+ O \left( \frac{1}{k} \right)$. The second sum is $O \left( \frac{1}{k} \right)$, the third sum is $O \left( \frac{\log(k)}{k^2} \right)$ and the fourth is $O \left( \frac{1}{k^2} \right)$. Thus $-a_j a_{k-j}$ summed over $1 \leq j \leq (k-1)/2$ is $- \frac{\log(k)}{k} + O \left( \frac{1}{k} \right)$. The sum of $(a_ja_{k-j})^2$ is of lower order by similar arguments. In all, the lower bound on $lim_{n \rightarrow \infty} P(F_k(w)>0)$ is \[ 1-e^{-\frac{\log(k)}{k} + O \left( \frac{1}{k} \right)} = \frac{\log(k)}{k} + O \left( \frac{1}{k} \right).\] \end{proof} To close this section, we give a combinatorial interpretation for the moments of the numbers $F_k(w)$ of Theorem \ref{klim} above. This involves the ``top k to random'' shuffle, which removes k cards from the top of the deck, and randomly interleaves them with the other n-k cards (choosing one of the ${n \choose k}$ possible interleavings uniformly at random). \begin{prop} \label{shufeig} \begin{enumerate} \item The eigenvalues of the top k to random shuffle are the numbers $\left\{ \frac{F_k(w)}{{n \choose k}} \right \}$, where $w$ ranges over $S_n$. \item For all values of $n,k,r$, the rth moment of the distribution of fixed k-sets is equal to ${n \choose k}^r$ multiplied by the chance that the top k to random shuffle is at the identity after r steps. \end{enumerate} \end{prop} \begin{proof} Note that the top k to random shuffle is the inverse of the move k to front shuffle, which picks k cards at random and moves them to the front of the deck, preserving their relative order. Hence their transition matrices are transposes, so have the same eigenvalues. The move k to front shuffle is a special case of the theory of random walk on chambers of hyperplane arrangements developed in \cite{BHR}. The arrangement is the braid arrangement and one assigns weight $\frac{1}{{n \choose k}}$ to each of the block ordered partitions where the first block has size $k$ and the second block has size $n-k$. The result now follows from Corollary 2.2 of \cite{BHR}, which determined the eigenvalues of such hyperplane walks. For the second assertion, let $M$ be the transition matrix for the top k to random shuffle. Clearly $Tr(M^r)$ (the trace of $M^r$) is equal to $n!$ multiplied by the chance that the top k to random shuffle is at the identity after r steps. The first part gives that \[ Tr(M^r) = \sum_{\ w \in S_n} \left( \frac{F_k(w)}{{n \choose k}} \right)^r, \] which implies the result. \end{proof} As an example of part 2 of Proposition \ref{shufeig}, the chance of being at the identity after 1 step is $\frac{1}{{n \choose k}}$ and the chance of being at the identity after 2 steps is $\frac{k+1}{{n \choose k}^2}$, giving another proof that $E(F_k(w))=1$ and $E(F_k^2(w))=k+1$. {\it Remarks} \begin{enumerate} \item As in the proof of Proposition \ref{klim}, the moments of $F_k(w)$ can be expressed exactly in terms of the moments of Poisson random variables, provided that $n$ is sufficiently large. \item There is a random walk on the irreducible representations of $S_n$ which has the same eigenvalues as the top k to random walk, but with different multiplicities. Unlike the top k to random walk, this walk is reversible with respect to its stationary distribution, so that spectral techniques (and hence information about the distribution of fixed points) can be used to analyze its convergence rate. For details, applications, and a generalization to other actions, see \cite{F1}, \cite{F2}. \end{enumerate} \section{Fixed Points on Matchings} \label{matchings} Let $M_{2n}$ be the set of perfect matchings on $2n$ points. Thus, if $2n = 4, M_{2n} = \{(1,2)(3,4), (1,3)(2,4), (1,4)(2,3)\}$. It is well known that $$ |M_{2n}| = (2n - 1)!! = (2n - 1)(2n - 3) \cdots (3) (1). $$ The literature on perfect matchings is enormous. See Lov\'{a}sz and Plummer \cite{LoPl} for a book length treatment. Connections with phylogenetic trees and further references are in \cite{DH,DH2}. As explained above, the symmetric group $S_{2n}$ acts primitively on $M_{2n}$. Results of Luczak-Pyber \cite{LP} imply that, in this action, almost every permutation is a derangement. In this section we give sharp asymptotic rates for this last result. We show that the proportion of derangements in $S_{2n}$ is \refstepcounter{thm} \begin{equation} \label{eqn5.1} 1 - {A (1) \over \sqrt{\pi n}} + o \Bigg( {1 \over \sqrt{n}} \Bigg) , \quad A(1) = \prod_{i=1}^\infty {\rm cosh} (1 / (2i-1)). \end{equation} Similar asymptotics are given for the proportion of permutations with $j > 0$ fixed points. This is zero if $j$ is even. For odd $j$, it is \refstepcounter{thm} \begin{equation} \label{eqn5.2} {C(j) B(1) \over \sqrt{\pi n}} + o \Bigg( {1 \over \sqrt{n}} \Bigg), \quad B(1) = \prod_{i=1}^\infty \Bigg( 1 + {1 \over 2i} \Bigg) e^{-{1 / 2i}} \end{equation} and $C(j)$ explicit rational numbers. In particular \refstepcounter{thm} \begin{equation} \label{eqn5.3} C(1) = {3 \over 2}, \ C(3) = {1 \over 4}, \ C(5) = {27 \over 400}, \ C(7) = {127 \over 2352}. \end{equation} The argument proceeds by finding explicit closed forms for generating functions followed by standard asymptotics. It is well known that the rank of this action is $p(n)$, the number of partitions of $n$. Thus (\ref{eqn5.1}) is a big improvement over the upper bound given in Theorem \ref{basic}. For $w \in S_{2n}$, let $a_i (w)$ be the number of $i$-cycles in the cycle decomposition. Let $F(w)$ be the number of fixed points of $w$ acting on $M_{2n}$. The following proposition determines $F(w)$ in terms of $a_i(w), \ 1 \leq i \leq 2n$. \begin{prop} \label{numfix} The number of fixed points, $F(w)$ of $w \in S_{2n}$ on $M_{2n}$ is $$ F(w) = \prod_{i=1}^{2n} F_i (a_i (w)) $$ with $$ F_{2i-1} (a) = \begin{cases} 1 &{\rm if} \ a = 0 \\ 0 &{\rm if} \ a \ \mbox{is odd} \\ (a-1)!!(2i-1)^{a/2} &{\rm if} \ a > 0 \ \mbox{is even} \end{cases} $$ $$ F_{2i} (a) = 1+ \sum_{k=1}^{\lfloor a/2 \rfloor} (2k-1)!! {a \choose 2k} (2i)^k $$ In particular, \refstepcounter{thm} \begin{equation} F(w) \not = 0 \ \mbox{if and only if} \ a_{2i-1} (w) \ \mbox{is even for all} \ i \end{equation} \refstepcounter{thm} \begin{equation} F(w) \ \mbox{does not take on non-zero even values}. \end{equation} \end{prop} \begin{proof} Consider first the cycles of $w$ of length $2 i-1$. If $a_{2i-1}$ is even, the cycles may be matched in pairs, then each pair of length $2i-1$ can be broken into matched two element subsets by first pairing the lowest element among the numbers with any of the $2i-1$ numbers in the second cycle. The rest is determined by cyclic action. For example, if the two three-cycles $(123) (456)$ appear, the matched pairs $(14)(25)(36)$ are fixed, so are $(15)(26)(34)$ or $(16)(24)(35)$. Thus $F_3 (2) = 3$. If $a_{2i-1}$ is odd, some element cannot be matched and $F_{2i-1}(a_{2i-1}) = 0$. Consider next the cycles of $w$ of length $2i$. Now, there are two ways to create parts of a fixed perfect matching. First, some of these cycles can be paired and, for each pair, the previous construction can be used. Second, for each unpaired cycle, elements $i$ apart can be paired. For example, from $(1234)$ the pairing $(13)(24)$ may be formed. The sum in $F_{2i} (a)$ simply enumerates by partial matchings. To see that $F(w)$ cannot take on non-zero even values, observe that $F_{2i-1} (a)$ and $F_{2i}(a)$ only take on odd values if they are non-zero. \end{proof} Let $P_{2n} (j) = \frac{|\{w \in S_{2n} : F(w) = j\}|}{2n!}$. For $j \geq 1$, let $g_j (t) = \sum_{n=0}^\infty t^{2n} P_{2n} (j)$ and let $\bar g_0(t) = \sum_{n=0}^\infty t^{2n} (1 - P_{2n} (0))$. \begin{prop} \label{cycindex} $$ \bar g_0(t) = {\displaystyle \prod_{i=1}^\infty \cosh (t^{2i-1} / (2i-1)) \over \sqrt {1 - t^2}} \qquad \mbox{for} \ 0 < t < 1. $$ \end{prop} \begin{proof} From Proposition \ref{numfix}, $w \in S_{2n}$ has $F(w) \not = 0$ if and only if $a_{2i-1} (w)$ is even for all $i$. From Shepp-Lloyd \cite{SL}, if $N$ is chosen in $\{0, 1, 2, \ldots\}$ with $$ P (N = n) = (1 - t) t^n $$ and then $w$ is chosen uniformly in $S_N$, the $a_i(w)$ are independent Poisson random variables with parameter $t^i / i$ respectively. If $X$ is a Poisson $(\lambda)$ random variable, $P$ ($X$ is even) = $\Bigg( {1 \over 2} + {e^{-2 \lambda} \over 2} \Bigg)$. It follows that \begin{eqnarray*} \sum_{n=0}^\infty (1 - t) t^{2n} (1 - P_{2n} (0)) & = & \prod_{i=1}^\infty \Bigg( {1 + e^{-2{t}^{2i-1} / (2i-1)} \over 2} \Bigg)\\ & = & \prod_{i=1}^\infty e^{{-t}^{2i-1} / (2i-1)} \prod_{i=1}^\infty {\rm cosh} \Bigg( {t^{2i-1} \over (2i-1)} \Bigg) \\ & = & \sqrt{{1 - t \over 1 + t}} \prod_{i=1}^\infty {\rm cosh} (t^{2i-1} / (2i-1)). \end{eqnarray*} \end{proof} \begin{cor} \label{asymcor} As $n$ tends to infinity, $$ 1 - P_{2n} (0) \sim {\displaystyle \prod_{i=1}^\infty {\rm cosh} (1/(2i-1)) \over \sqrt {\pi n}}. $$ \end{cor} \begin{proof} By Proposition \ref{cycindex}, $1-P_{2n}(0)$ is the coefficient of $t^n$ in \[ \frac{\prod_{i=1}^\infty \cosh (t^{(2i-1)/2} / (2i-1))}{ \sqrt {1 - t}}.\] It is straightforward to check that the numerator is analytic near $t=1$, so the result follows from Darboux's theorem (\cite{O}, Theorem 11.7). \end{proof} Proposition \ref{numfix} implies that the event $F(w) = j$ is contained in the event $\{a_{2i-1} (w)$ is even for all $i$ and $a_{2i}(w) \in \{0, 1\} \ \mbox{for} \ 2i > j\}$. This is evidently complicated for large $j$. We prove \begin{prop} \label{genfunc} For positive odd $j$, $$g_j (t) = {P_j(t) \displaystyle \prod_{i=1}^{\infty} \Big(1 + {t^{2i} \over 2i} \Big) e^{-t^{2i} / 2i} \over \sqrt {1 - t^2}} $$ for $P_j(t)$ an explicit rational function in $t$ with positive rational coefficients. In particular, $$ P_1 (t) = \Bigg(1 + {t^2 \over 2} \Bigg), P_3 (t) = \Bigg( \frac{t^4}{6} + \frac{t^6}{18} + \frac{t^8}{36} \Bigg) $$ $$P_5(t) = {\frac{1}{2} \Big({1 + \frac{t^2}{2}} \Big) \Big( {t^2 \over 4} \Big)^2 \over 1 + {t^{4} \over 4}} + \frac{1}{2} \left(1+ \frac{t^2}{2} \right) \left( \frac{t^5}{5} \right)^2 $$ $$ P_7 (t) = { \frac{1}{2} \Big( 1 + \frac{t^2}{2} \Big) \over 1 + {t^6 \over 6}} \Big({t^6 \over 6} \Big)^2 + \frac{1}{6} \Big( {t^2 \over 2} \Big)^3 + \frac{1}{2} \left(1+\frac{t^2}{2} \right) \left( \frac{t^7}{7} \right)^2. $$ \end{prop} \begin{proof} Consider first the case of $j=1$. From Proposition \ref{numfix}, $F(w) = 1$ if and only if $a_{2i-1} (w) = 0$ for $i \geq 2$, $a_1 (w) \in \{0, 2\}$ and $a_{2i} (w) \in \{0, 1\}$ for all $i \geq 1$. For example, if $2n = 10$ and $w = (1) (2) (3456789 10)$, the unique fixed matching is $(1\ 2) (3 \ 7) (4\ 8)(5\ 9)(6\ 10)$. From the cycle index argument used in Proposition \ref{cycindex}, \begin{eqnarray*} & & \sum_{n=0}^{\infty} (1 - t) t^{2n} P_{2n} (F(w) = 1)\\ & = & e^{-t} \Big( 1 + {t^2 \over 2} \Big) \displaystyle \prod_{i=2}^\infty e^{-t^{2i-1} / (2i-1)} \prod_{i=1}^{\infty} e^{-t^{2i}/2i} \Big( 1 + {t^{2i} \over 2i} \Big) \\ & = & (1 - t) \Big(1 + {t^2 \over 2} \Big) \displaystyle \prod_{i=1}^\infty \Big( 1 + {t^{2i} \over 2i} \Big)\\ & = & {(1-t)\Big( 1 + {t^2 \over 2} \Big) \displaystyle \prod_{i=1}^\infty \Big( 1 + {t^{2i} \over 2i} \Big) e^{-t^{2i} / 2i} \over \sqrt {1 - t^2}}. \end{eqnarray*} The arguments for the other parts are similar. In particular, $F(w)=3$ iff one of the following holds: \begin{itemize} \item $a_1(w) = 4$, all $a_{2i-1}(w) = 0 \ i \geq 2$, all $a_{2i}(w) \in \{0, 1\}$ \item $a_1(w) \in \{0, 2\}, \ a_2(w) = 2$ all $a_{2i-1}(w) = 0$ and $a_{2i} (w) \in \{0, 1\} \ i \geq 2$ \item $a_1(w) \in \{0, 2\}, \ a_3(w) = 2, a_{2i-1} (w) = 0 \ i \geq 3, \ a_{2i} \in \{0, 1\}$ \end{itemize} Similarly, $F(w)=5$ iff one of the following holds: \begin{itemize} \item $a_4(w) = 2, \ a_1 (w) \in \{0, 2\}, a_{2i-1}(w) = 0, a_{2i} (w) \in \{0, 1\}$ else \item $a_5(w)=2, a_1(w) \in \{0,2\}, a_{2i-1}(w)=0$, $a_{2i}(w) \in \{0,1\}$ else \end{itemize} Finally, $F(w)=7$ iff one of the following holds: \begin{itemize} \item $a_1(w) \in \{0, 2\}, \ a_6 (w) = 2 \ \mbox{or} \ a_2 (w) = 3, a_{2i-1} (w) = 0, a_{2i} (w) \in \{0, 1\}$ else \item $a_7(w)=2,a_1(w) \in \{0,2\}, a_{2i-1}(w)=0$, $a_{2i}(w) \in \{0,1 \}$ else \end{itemize} Further details are omitted. \end{proof} The asymptotics in (\ref{eqn5.2}) follow from Proposition \ref{genfunc}, by the same method used to prove (\ref{eqn5.1}) in Corollary \ref{asymcor}. \section{More imprimitive subgroups} \label{impriv} Section \ref{matchings} studied fixed points on matchings, or equivalently fixed points of $S_{2n}$ on the left cosets of $S_2 \wr S_n$. This section uses a quite different approach to study derangements of $S_{an}$ on the left cosets of $S_a \wr S_n$, where $a \geq 2$ is constant. It is proved that the proportion of elements of $S_{an}$ which fix at least one left coset of $S_a \wr S_n$ (or equivalently are conjugate to an element of $S_a \wr S_n$ or equivalently fix a system of $n$ blocks of size $a$) is at most the coefficient of $u^n$ in \[ \exp \left( \sum_{k \geq 1} \frac{u^k}{a!} (\frac{1}{k})(\frac{1}{k}+1) \cdots (\frac{1}{k}+a-1) \right), \] and that this coefficient is asymptotic to $C_a n^{\frac{1}{a}-1}$ as $n \rightarrow \infty$, where $C_a$ is an explicit constant depending on $a$ (defined in Theorem \ref{genfunction} below). In the special case of matchings ($a=2$), this becomes $\frac{ e^{\frac{\pi^2}{12}} } {\sqrt{\pi n}}$, which is extremely close to the true asymptotics obtained in Section \ref{matchings}. Moreover, this generating function will be crucially applied when we sharpen a result of Luczak and Pyber in Section \ref{prim}. The method of proof is straightforward. Clearly the number of permutations in $S_{an}$ conjugate to an element of $S_a \wr S_n$ is upper bounded by the sum over conjugacy classes $C$ of $S_a \wr S_n$ of the size of the $S_{an}$ conjugacy class of $C$. Unfortunately this upper bound is hard to compute, but we show it to be smaller than something which can be exactly computed as a coefficient in a generating function. This will prove the result. From Section 4.2 of \cite{JK}, there is the following useful description of conjugacy classes of $G \wr S_n$ where $G$ is a finite group. The classes correspond to matrices $M$ with natural number entries $M_{i,k}$, rows indexed by the conjugacy classes of $G$, columns indexed by the numbers $1,2,\cdots,n$, and satisfying the condition that $\sum_{i,k} k M_{i,k} = n$. More precisely, given an element $(g_1,\cdots,g_n; \pi)$ in $G \wr S_n$, for each k-cycle of $\pi$ one multiplies the $g$'s whose subscripts are the elements of the cycle in the order specified by the cycle. Taking the conjugacy class in $G$ of the resulting product contributes 1 to the matrix entry whose row corresponds to this conjugacy class in $G$ and whose column is $k$. The remainder of this section specializes to $G=S_a$. Since conjugacy classes of $S_a$ correspond to partitions $\lambda$ of $a$, the matrix entries are denoted by $M_{\lambda,k}$. We write $|\lambda|= a$ if $\lambda$ is a partition of $a$. Given a partition $\lambda$, let $n_i(\lambda)$ denote the number of parts of size $i$ of $\lambda$. \begin{prop} \label{param} Let the conjugacy class $C$ of $S_a \wr S_n$ correspond to the matrix $(M_{\lambda,k})$ where $\lambda$ is a partition of $a$. Then the proportion of elements of $S_{an}$ conjugate to an element of $C$ is at most \[ \frac{1}{\prod_{k} \prod_{|\lambda|= a} M_{\lambda,k}! [ \prod_i (ik)^{n_i(\lambda)} n_i(\lambda)!]^{M_{\lambda,k}}}.\] \end{prop} \begin{proof} Observe that the number of cycles of length $j$ of an element of $C$ is equal to \[ \sum_{k|j} \sum_{|\lambda|=a} M_{\lambda,k} n_{j/k}(\lambda).\] To see this, note that $S_a \wr S_n$ can be viewed concretely as a permutation of $an$ symbols by letting it act on an array of $n$ rows of length $a$, with $S_a$ permuting within each row and $S_n$ permuting among the rows. Hence by a well known formula for conjugacy class sizes in a symmetric group, the proportion of elements of $S_{an}$ conjugate to an element of $C$ is equal to \begin{eqnarray*} & & \frac{1}{\prod_j j^{\sum_{k|j} \sum_{|\lambda|= a} M_{\lambda,k} n_{j/k}(\lambda)} [\sum_{k|j} \sum_{|\lambda|= a} M_{\lambda,k} n_{j/k}(\lambda)] !}\\ & \leq & \frac{1}{\prod_j j^{\sum_{k|j} \sum_{|\lambda|= a} M_{\lambda,k} n_{j/k}(\lambda)} \prod_{k|j} \prod_{|\lambda|= a} M_{\lambda,k} n_{j/k}(\lambda) !}\\ & \leq & \frac{1}{\prod_j j^{\sum_{k|j} \sum_{|\lambda|= a} M_{\lambda,k} n_{j/k}(\lambda)} \prod_{k|j} \prod_{|\lambda|= a} [ M_{\lambda,k}! n_{j/k}(\lambda)!^{M_{\lambda,k}}]}\\ & = & \frac{1}{\prod_k \prod_{|\lambda|= a} M_{\lambda,k}! [\prod_i (ik)^{n_i(\lambda)} n_i(\lambda)!]^{M_{\lambda,k}}}, \end{eqnarray*} as desired. The first inequality uses the fact that $(x_1+\cdots+x_n)! \geq x_1! \cdots x_n!$. The second inequality uses that $(xy)! \geq x! y!^x$ for $x,y \geq 1$ integers, which is true since \[ (xy)! = \prod_{i=1}^x \prod_{j=0}^{y-1} (i+jx) \geq \prod_{i=1}^x \prod_{j=0}^{y-1} i(1+j) = (x!)^y (y!)^x \geq x! (y!)^x.\] The final equality used the change of variables $i=j/k$. \end{proof} To proceed further, the next lemma is useful. \begin{lemma} \label{numcycles} \[ \sum_{|\lambda|= a} \frac{1}{\prod_i (ik)^{n_i(\lambda)} n_i(\lambda)!} = \frac{(\frac{1}{k}) (\frac{1}{k}+1) \cdots (\frac{1}{k}+a-1)}{a!} .\] \end{lemma} \begin{proof} Let $c(\pi)$ denote the number of cycles of a permutation $\pi$. Since the number of permutations in $S_a$ with $n_i$ cycles of length $i$ is $\frac{a!}{\prod_i i^{n_i} n_i!}$, the left hand side is equal to \[ \frac{1}{a!} \sum_{\pi \in S_a} k^{-c(\pi)} .\] It is well known and easily proved by induction that \[ \sum_{\pi \in S_a} x^{c(\pi)} = x(x+1) \cdots (x+a-1).\] \end{proof} Theorem \ref{genfunction} applies the preceding results to obtain a useful generating function. \begin{theorem} \label{genfunction} \begin{enumerate} \item The proportion of elements in $S_{an}$ conjugate to an element of $S_a \wr S_n$ is at most the coefficient of $u^n$ in \[ \exp \left(\sum_{k \geq 1} \frac{u^k}{a!} (\frac{1}{k})(\frac{1}{k}+1) \cdots (\frac{1}{k}+a-1) \right).\] \item For $a$ fixed and $n \rightarrow \infty$, the coefficient of $u^n$ in this generating function is asymptotic to \[ \frac{e^{\sum_{r=2}^a p(a,r) \zeta(r)}}{\Gamma(1/a)} n^{\frac{1}{a}-1}\] where $p(a,r)$ is the proportion of permutations in $S_a$ with exactly r cycles, $\zeta$ is the Riemann zeta function, and $\Gamma$ is the gamma function. \end{enumerate} \end{theorem} \begin{proof} Proposition \ref{param} implies that the sought proportion is at most the coefficient of $u^n$ in \begin{eqnarray*} & & \prod_{k} \prod_{|\lambda|=a} \sum_{M_{\lambda,k} \geq 0} \frac{u^{k M_{\lambda,k}}}{ M_{\lambda,k}! [(ik)^{n_i(\lambda)} n_i(\lambda)!]^{M_{\lambda,k}}}\\ & = & \prod_k \prod_{|\lambda|=a} \exp \left(\frac{u^k}{(ik)^{n_i(\lambda)} n_i(\lambda)!} \right)\\ & = & \prod_k \exp \left(\sum_{|\lambda|=a} \frac{u^k}{(ik)^{n_i(\lambda)} n_i(\lambda)!} \right)\\ & = & \exp \left(\sum_{k \geq 1} \frac{u^k}{a!} (\frac{1}{k})(\frac{1}{k}+1) \cdots (\frac{1}{k}+a-1) \right). \end{eqnarray*} The last equality used Lemma \ref{numcycles}. For the second assertion, one uses Darboux's lemma (see \cite{O} for an exposition), which gives the asymptotics of functions of the form $(1-u)^{\alpha} g(u)$ where $g(u)$ is analytic near 1, $g(1) \neq 0$, and $\alpha \not \in \{0,1,2, \cdots \}$. More precisely it gives that the coefficient of $u^n$ in $(1-u)^{\alpha} g(u)$ is asymptotic to $\frac{g(1)}{\Gamma(-\alpha)} n^{-\alpha-1}$. By Lemma \ref{numcycles}, \begin{eqnarray*} & & \exp \left( \sum_{k \geq 1} \frac{u^k}{a!} (\frac{1}{k})(\frac{1}{k}+1) \cdots (\frac{1}{k}+a-1) \right)\\ & = & \exp \left( \sum_{k \geq 1} \frac{u^k}{ak} + \sum_{k \geq 1} u^k \sum_{r=2}^a p(a,r) k^{-r} \right)\\ & = & (1-u)^{-\frac{1}{a}} \cdot \exp \left(\sum_{r=2}^a p(a,r) \sum_{k \geq 1} \frac{u^k}{k^r} \right). \end{eqnarray*} Taking $g(u) = \exp \left(\sum_{r=2}^a p(a,r) \sum_{k \geq 1} \frac{u^k}{k^r} \right)$ proves the result. \end{proof} {\it Remark} The upper bound in Theorem \ref{genfunction} is not perfect. In fact when $n=2$, it does not approach 0 as $a \rightarrow \infty$, whereas the true answer must by Theorem \ref{theC}. However by part 2 of Theorem \ref{genfunction}, the bound is useful for $a$ fixed and $n$ growing, and it will be crucially applied in Section \ref{prim} when $a=n$ are both growing. \section{Primitive subgroups} \label{prim} A main goal of this section is to prove that the proportion of elements of $S_n$ which belong to a primitive subgroup not containing $A_n$ is at most $O(n^{-2/3+\alpha})$ for any $\alpha>0$. This improves on the bound $O(n^{-1/2+\alpha})$ in \cite{LP}, which was used in proving Theorem \ref{theB} in Section \ref{Onan}. We conjecture that this can in fact be replaced by $O(n^{-1})$ (and the examples with $n = (q^d-1)/(q-1)$ with the subgroup containing $PGL(d,q)$ or $n=p^d$ with subgroup $AGL(d,p)$ show that in general one can do no better). The minimal degree of a permutation group is defined as the least number of points moved by a nontrivial element. The first step is to classify the degree $n$ primitive permutation groups with minimal degree at most $n^{2/3}$. We note that Babai \cite{babai} gave an elegant proof (not requiring the classification of finite simple groups) that there are no primitive permutation groups of degree $n$ other than $A_n$ or $S_n$ with minimal degree at most $n^{1/2}$. \begin{theorem} \label{bobemailprim} Let $G$ be a primitive permutation group of degree $n$. Assume that there is a nontrivial $g \in G$ moving at most $n^{2/3}$ points. Then one of the following holds: \begin{enumerate} \item $G=A_n$ or $S_n$; \item $G=S_m$ or $A_m$ with $m \ge 5$ and $n = \binom{m}{2}$ (acting on subsets of size $2$) ; or \item $A_m \times A_m < G \le S_m \wr S_2$ with $m \ge 4$ and $n=m^2$ (preserving a product structure). \end{enumerate} If there is a nontrivial $g \in G$ moving fewer than $n^{1/2}$ points, then $G=A_n$ or $S_n$. \end{theorem} \begin{proof} First note that the minimal degree in (2) is $2(m-2)$ and in (3) is $2n^{1/2}$. In particular, aside from (1), we always have the minimal degree is at least $n^{1/2}$. Thus, the last statement follows from the first part. It follows by the main result of \cite{GM} that if there is a $g \in G$ moving fewer than $n/2$ points, then one the following holds: (a) $G$ is almost simple with socle (the subgroup generated by the minimal normal subgroups) $A_m$ and $n = \binom{m}{k}$ with the action on subsets of size $k < m/2$; (b) $n=m^t$, with $t > 1$, $m \ge 5$, $G$ has a unique minimal normal subgroup $N = L_1 \times \ldots \times L_t$ with $t >1$ and $G$ preserves a product structure -- i.e. if $\Omega =\{1, \ldots, n\}$, then as $G$-sets, $\Omega \cong X^t$ where $m=|X|$, $G \le S_m \wr S_t$ acts on $X^t$ by acting on each coordinate and permuting the coordinates. Note that $n/2 \geq n^{2/3}$ as long as $n \geq 8$. If $n<8$, then $G$ contains an element moving at most $3$ points, i.e. either a transposition or a $3$ cycle, and so contains $A_n$ (Theorem 3.3A in \cite{DxM}). Consider (a) above. If $k=1$, then $(1)$ holds. If $3 \le k < m/2$, then it is an easy exercise to see that the element of $S_m$ moving the fewest $k$ sets is a transposition. The number of $k$-sets moved is $2 \binom{m-2}{k-1}$. We claim that this is greater than $n^{2/3}$. Indeed, the sought inequality is equivalent to checking that $\frac{2k(m-k)}{m(m-1)} > {m \choose k}^{-1/3}$. The worst case is clearly $k=3$, which is checked by taking cubes. This settles the case $3 \le k < m/2$, and if $k=2$, we are in case (2). Now consider (b) above. Suppose that $t \geq 3$. Then if $g \in S_m \times \cdots \times S_m$ is nontrivial, it moves at least $2m^{t-1}>n^{2/3}$ many points. If $g \in S_m \wr S_t$ and is not in $S_m \times \cdots \times S_m$, then up to conjugacy we may write $g=(g_1,\cdots,g_t;\sigma)$ where say $\sigma$ has an orbit $\{ 1, \ldots, s\}$ with $s > 1$. Viewing our set as $A \times B$ with $A$ being the first $s$ coordinates, we see that $g$ fixes at most $m$ points on $A$ (since there is at most one $g$ fixed point with a given coordinate) and so on the whole space, $g$ fixes at most $m^{t-s+1} \leq m^{t-1}$ points and so moves at least $m^t - m^{t-1}$ points. Since $t \ge 3$, this is greater than $n^{2/3}$. Summarizing, we have shown that in case (b), $t \geq 3$ leads to a contradiction. So finally consider (b) with $t=2$. We claim that $L$ must be $A_m$. Enlarging the group slightly, we may assume that $G = S \wr S_2$ where $L \le S \le Aut(L)$ and $S$ is primitive of degree $m$. If $g \notin S \times S$, then arguing as in the $t=3$ case shows that $g$ moves at least $m^2-m$ points. This is greater than $m^{4/3}=n^{2/3}$ since $m \ge 5$, a contradiction. So write $g = (g_1, g_2) \in S \times S$ with say $g_1 \ne 1$. If $g_1$ moves at least $d$ points, then $g$ moves at least $dm$ points. This is greater than $n^{2/3}$ unless $d \le m^{1/3}$. By this theorem (for $m$), this implies that $L=A_m$, whence (1) holds. \end{proof} Next, we focus on Case 2 of Theorem \ref{bobemailprim}. \begin{lemma} \label{cyc} Let $S_m$ be viewed as a subgroup of $S_{{m \choose 2}}$ using its action on 2-sets of $\{1,\cdots,m\}$. For $w \in S_m$, let $A_i(w)$ denote the number of cycles of $w$ of length $i$ in its usual action on $\{1,\cdots,m\}$. The total number of orbits of $w$ on 2-sets $\{j,k\}$ of symbols which are in a common cycle of $w$ is \[ \frac{m}{2} - \sum_{i \ odd} \frac{A_i(w)}{2}.\] \end{lemma} \begin{proof} First suppose that $w$ is a single cycle of length $i \geq 2$. If $i$ is odd, then all orbits of $w$ on pairs of symbols in the $i$-cycle have length $i$, so the total number of orbits is $\frac{{i \choose 2}}{i} = \frac{i-1}{2}$. If $i$ is even, there is 1 orbit of size $\frac{i}{2}$ and all other orbits have size $i$, giving a total of $\frac{i}{2}$ orbits. Hence for general $w$, the total number of orbits on pairs of symbols in a common cycle of $w$ is \begin{eqnarray*} \sum_{i \ odd \atop i \geq 3} A_i(w) \frac{i-1}{2} + \sum_{i \ even} A_i(w) \frac{i}{2} & = & \sum_{i \ odd \atop i \geq 1} A_i(w) \frac{i-1}{2} + \sum_{i \ even} A_i(w) \frac{i}{2}\\ & = & \frac{m}{2} - \sum_{i \ odd} \frac{A_i(w)}{2}. \end{eqnarray*} \end{proof} \begin{theorem} \label{cas2} Let $S_m$ be viewed as a subgroup of $S_n$ with $n={{m \choose 2}}$ using its action on 2-sets of $\{1,\cdots,m\}$. Then the proportion of elements of $S_n$ contained in a conjugate of $S_m$ is at most $O \left( \frac{\log(n)}{n} \right)$. \end{theorem} \begin{proof} We claim that any element $w$ of $S_m$ has at least $\frac{m}{12}$ cycles when viewed as an element of $S_n$. Indeed, if $A_1(w)> \frac{m}{2}$, then $w$ fixes at least $\frac{m(m-1)}{8} \geq \frac{m}{12}$ two-sets. So we suppose that $A_1(w) \leq \frac{m}{2}$. Clearly $\sum_{i \geq 3 \ odd} A_i(w) \leq \frac{m}{3}$. Thus Lemma \ref{cyc} implies that $w$ has at least $\frac{m}{2} - \frac{m}{4} - \frac{m}{6} = \frac{m}{12}$ cycles as an element of $S_n$. The number of cycles of a random element of $S_n$ has mean and variance asymptotic to $\log(n) \sim 2 \log(m)$ (and is in fact asymptotically normal) \cite{Go}. Thus by Chebyshev's inequality, the proportion of elements in $S_m$ with at least $\frac{m}{12}$ cycles is $O \left( \frac{\log(m)}{m^2} \right) = O \left( \frac{\log(n)}{n} \right)$, as desired. \end{proof} To analyze Case 3 of Theorem \ref{bobemailprim}, the following bound, based on the generating function from Section \ref{impriv}, will be needed. \begin{prop} \label{blockbound} The proportion of elements in $S_{m^2}$ which fix a system of $m$ blocks of size m is $O(n^{-3/4+\alpha})$ for any $\alpha>0$. \end{prop} \begin{proof} By Theorem \ref{genfunction}, the proportion in question is at most the coefficient of $u^m$ in \[ \exp \left( \sum_{k \geq 1} \frac{u^k}{mk} (1+\frac{1}{k}) (1+\frac{1}{2k}) \cdots (1+\frac{1}{(m-1)k}) \right).\] If $f(u)$ and $g(u)$ are power series in $u$, we write $f(u)<<g(u)$ if the coefficient of $u^n$ in $f(u)$ is less than or equal to the corresponding coefficient in $g(u)$, for all $n$. Since $\log(1+x) \leq x$ for $0<x<1$, one has that \[ \log \left( \prod_{i=1}^{m-1} (1+\frac{1}{ik}) \right) \leq \sum_{i=1}^{m-1} \frac{1}{ki} \leq \frac{1}{k} (1+ \log(m-1)).\] Thus \begin{eqnarray*} & & \exp \left( \sum_{k \geq 1} \frac{u^k}{mk} (1+\frac{1}{k}) (1+\frac{1}{2k}) \cdots (1+\frac{1}{(m-1)k}) \right)\\ & << & \exp \left( \sum_{k \geq 1} \frac{u^k}{mk} e^{1/k} (m-1)^{1/k} \right)\\ & << & e^{ue} \exp \left( \sum_{k \geq 2} \frac{u^k}{k} \sqrt{\frac{e}{m}} \right)\\ & << & e^{ue} \exp \left( \sum_{k \geq 1} \frac{u^k}{k} \sqrt{\frac{e}{m}} \right) \\ & = & e^{ue} (1-u)^{-\sqrt{\frac{e}{m}}}. \end{eqnarray*} The coefficient of $u^i$ in $(1-u)^{-\sqrt{\frac{e}{m}}}$ is \[ \frac{1}{i!} \sqrt{\frac{e}{m}} \prod_{j=1}^{i-1} (\sqrt{\frac{e}{m}}+j-1) = \frac{1}{i} \sqrt{\frac{e}{m}} \prod_{j=1}^{i-1} (1+ \frac{1}{j} \sqrt{\frac{e}{m}}).\] Since \[ \log \left( \prod_{j=1}^{i-1} (1+\frac{1}{j} \sqrt{\frac{e}{m}}) \right) \leq \sum_{j=1}^{i-1} \frac{1}{j} \sqrt{\frac{e}{m}} \leq \sqrt{\frac{e}{m}} (1+ \log(i-1)),\] it follows that \[ \prod_{j=1}^{i-1} (1+ \frac{1}{j} \sqrt{\frac{e}{m}}) \leq [e(i-1)]^{\sqrt{\frac{e}{m}}}.\] This is at most a universal constant $A$ if $0 \leq i \leq m$. Thus the coefficient of $u^m$ in $e^{ue} (1-u)^{-\sqrt{\frac{e}{m}}}$ is at most \[ \frac{e^m}{m!} + A \sqrt{\frac{e}{m}} \sum_{i=1}^m \frac{1}{i} \frac{e^{m-i}}{(m-i)!}.\] By Stirling's formula (page 52 of \cite{Fe}), $m!> m^m e^{-m+1/(12m+1)} \sqrt{2 \pi m}$, which implies that the first term is very small for large $m$. To bound the sum, consider the terms for $i \ge m^{1-\alpha}$, where $0<\alpha<1$. These contribute at most $\frac{B m^{\alpha}}{m^{3/2}}$ for a universal constant $B$. The contribution of the other terms is negligible in comparison, by Stirling's formula. Summarizing, the contribution of the sum is $O(m^{-3/2+\alpha}) = O(n^{-3/4+\alpha/2})$, as desired. \end{proof} The following theorem gives a bound for Case 3 of Theorem \ref{bobemailprim}. \begin{theorem} \label{cas3} Let $S_m \wr S_2$ be viewed as a subgroup of $S_n$ with $n=m^2$ using its action on the Cartesian product $\{1,\cdots,m\}^2$. Then the proportion of elements of $S_n$ conjugate to an element of $S_m \wr S_2$ is $O(n^{-3/4+\alpha})$ for any $\alpha>0$. \end{theorem} \begin{proof} Consider elements of $S_m \wr S_2$ of the form $(w_1,w_2;id)$. These all fix $m$ blocks of size $m$ in the action on $\{1,\cdots,m\}^2$; the blocks consist of points with a given first coordinate. By Proposition \ref{blockbound}, the proportion of elements of $S_n$ conjugate to some $(w_1,w_2;id)$ is $O(n^{-3/4+\alpha})$ for any $\alpha>0$. Next, consider an element of $S_m \wr S_2$ of the form $\sigma=(w_1,w_2;(12))$. Then $\sigma^2=(w_1w_2,w_2w_1;id)$. Note that $w_1w_2$ and $w_2w_1$ are conjugate in $S_m$, and let $A_i$ denote their common number of $i$-cycles. Observe that if $x$ is in an $i$-cycle of $w_1w_2$, and $y$ is in an $i$-cycle of $w_2w_1$, then $(x,y) \in \Omega$ is in an orbit of $\sigma^2$ of size $i$. Hence the total number of orbits of $\sigma^2$ of size $i$ is at least $\frac{(iA_i)^2}{i} \geq iA_i$. Thus the total number of orbits of $\sigma^2$ on $\Omega$ is at least $\sum_i iA_i=m$. Hence the total number of orbits of $\sigma$ is at least $\frac{m}{2}$. Arguing as in the proof of Theorem \ref{cas2}, it follows that the proportion of elements of $S_n$ conjugate to an element of the form $\sigma$ is $O \left( \frac{\log(n)}{n} \right)$, and so is $O(n^{-3/4+\alpha})$ for any $\alpha>0$. \end{proof} Now the main result of this section can be proved. \begin{theorem} \label{mainres} The proportion of elements of $S_n$ which belong to a primitive subgroup not containing $A_n$ is at most $O(n^{-2/3+\alpha})$ for any $\alpha>0$. \end{theorem} \begin{proof} Fix $\alpha>0$. By Bovey \cite{Bo}, the proportion of elements $w$ of $S_n$ such that $\langle w \rangle$ has minimum degree greater than $n^{2/3}$ is $O(n^{-2/3+\alpha})$. Thus the proportion of $w \in S_n$ which lie in a primitive permutation group having minimal degree greater than $n^{2/3}$ is $O(n^{-2/3+\alpha})$. The only primitive permutation groups of degree $n$ with minimal degree $\leq n^{2/3}$, and not containing $A_n$ are given by Cases 2 and 3 of Theorem \ref{bobemailprim}. Theorems \ref{cas2} and \ref{cas3} imply that the proportion of $w$ lying in the union of all such subgroups is $O(n^{-2/3+\alpha})$, so the result follows. \end{proof} A trivial corollary to the theorem is that this holds for $A_n$ as well. \\ {\it Remark}\ The actions of the symmetric group studied in this section embed the group as a subgroup of various larger symmetric groups. Any such embedding can be thought of as a code in the larger symmetric group. Such codes may be used for approximating sums of various functions over the larger symmetric group via a sum over the smaller symmetric group. Our results can be interpreted as giving examples of functions where the approximation is not particularly accurate. For example, the proof of Theorem \ref{cas2} shows this to be the case when $S_m$ is viewed as a subgroup of $S_n, n = \binom{m}{2}$ using the actions on 2-sets, and the function is the number of cycles. \section{Related results and applications} \label{survey} There are numerous applications of the distribution of fixed points and derangements. Subsection \ref{motivnum} mentions some motivation from number theory. Subsection \ref{shalev} discusses some literature on the proportion of derangements and an analog of the main result of our paper for finite classical groups. Subsection \ref{fpr} discusses fixed point ratios, emphasizing the application to random generation. Subsection \ref{miscell} collects some miscellany about fixed points and derangements, including algorithmic issues and appearances in algebraic combinatorics. While this section does cover many topics, the survey is by no means comprehensive. Some splendid presentations of several other topics related to derangements are Serre \cite{Se}, Cameron's lecture notes \cite{Ca2} and Section 6.6 of \cite{Ca1}. For the connections with permutations with restricted positions and rook polynomials see \cite[2.3, 2.4]{Stanley}. \subsection{Motivation from number theory} \label{motivnum} We describe two number theoretic applications of derangements which can be regarded as motivation for their study:\\ (1) {\it Zeros of polynomials} Let $h(T)$ be a polynomial with integer coefficients which is irreducible over the integers. Let $\pi(x)$ be the number of primes $\leq x$ and let $\pi_h(x)$ be the number of primes $\leq x$ for which $h$ has no zeros mod $p$. It follows from Chebotarev's density theorem (see \cite{LS} for history and a proof sketch), that $lim_{x \rightarrow \infty} \frac{\pi_h(x)}{\pi(x)}$ is equal to the proportion of derangements in the Galois group $G$ of $h(T)$ (viewed as permutation of the roots of $h(T)$). Several detailed examples are worked out in Serre's survey \cite{Se}. In addition, there are applications such as the the number field sieve for factoring integers (Section 9 of \cite{BLP}), where it is important to understand the proportion of primes for which $h$ has no zeros mod $p$. This motivated Lenstra (1990) to pose the question of finding a good lower bound for the proportion of derangements of a transitive permutation group acting on a set of $n$ letters with $n \geq 2$. Results on this question are described in Subsection \ref{shalev}. \\ (2) {\it The value problem} Let $\mathbb{F}_q$ be a finite field of size $q$ with characteristic $p$ and let $f(T)$ be a polynomial of degree $n>1$ in $\mathbb{F}_q[T]$ which is not a polynomial in $T^p$. The arithmetic question raised by Chowla \cite{Ch} is to estimate the number $V_f$ of distinct values taken by $f(T)$ as $T$ runs over $\mathbb{F}_q$. There is an asymptotic formula for $V_f$ in terms of certain Galois groups and derangements. More precisely, let $G$ be the Galois group of $f(T)-t=0$ over $\mathbb{F}_q(t)$ and let $N$ be the Galois group of $f(T)-t=0$ over $\overline{\mathbb{F}}_q(t)$, where $\overline{\mathbb{F}}_q$ is an algebraic closure of $\mathbb{F}_q$ (we are viewing $f(T)-t$ as a polynomial with variable $T$ with coefficients in $\F_q(t)$). Both groups act transitively on the $n$ roots of $f(T)-t=0$. The geometric monodromy group $N$ is a normal subgroup of the arithmetic monodromy group $G$. The quotient group $G/N$ is a cyclic group (possibly trivial). \begin{theorem} (\cite{Co}) Let $xN$ be the coset which is the Frobenius generator of the cyclic group $G/N$. The Chebotarev density theorem for function fields yields the following asymptotic formula: \[ V_f = \left( 1 - \frac{|S_0|}{|N|} \right) q + O(\sqrt{q}) \] where $S_0$ is the set of group elements in the coset $xN$ which act as derangements on the set of roots of $f(T)-t=0$. The constant in the above error term depends only on $n$, not on $q$. \end{theorem} As an example, let $f(T)=T^r$ with $r$ prime and different from $p$ (the characteristic of the base field $\mathbb{F}_q$). The Galois closure of $\mathbb{F}_q(T)/\mathbb{F}_q(T^r)$ is $\mathbb{F}_q(\mu,T)$ where $\mu$ is a nontrivial $r$th root of $1$. Thus $N$ is cyclic of order $r$ and $G/N$ is isomorphic to the Galois group of $\mathbb{F}_q(\mu)/\mathbb{F}_q$. The permutation action is of degree $r$. If $G=N$, then every non-trivial element is a derangement and so the image of $f$ has order roughly $\frac{q}{r} + O(\sqrt{q})$. If $G \neq N$, then $G$ is a Frobenius group and every fixed point free element is contained in $N$. Indeed, since in this case $(r,q-1)=1$, we see that $T^r$ is bijective on $\mathbb{F}_q$. For further examples, see Guralnick-Wan \cite{GW} and references therein. Using work on derangements, they prove that if the degree of $f$ is relatively prime to the characteristic $p$, then either $f$ is bijective or $V_f \leq \frac{5q}{6} + O(\sqrt{q})$. \subsection{Proportion of derangements and Shalev's conjecture} \label{shalev} Let $G$ be a finite permutation group acting transitively on a set $X$ of size $n>1$. Subsection \ref{motivnum} motivated the study of $\delta(G,X)$, the proportion of derangements of $G$ acting on $X$. We describe some results on this question, focusing particularly on lower bounds and analogs of our main results for classical groups. Perhaps the earliest such result is due to Jordan \cite{Jo}, who showed that $\delta(G,X)>0$. Cameron and Cohen \cite{CaCo} proved that $\delta(G,X) \geq 1/n$ with equality if and only if $G$ is a Frobenius group of order $n(n-1)$, where $n$ is a prime power. See also \cite{Se}, who also notes a topological application of Jordan's theorem. Based on extensive computations, it was asked in \cite{Bet} whether there is a universal constant $\delta > 0$ (which they speculate may be optimally chosen as $\frac{2}{7}$) such that $\delta(G,X)> \delta$ for all finite simple groups $G$. The existence of such a $\delta > 0$ was also conjectured by Shalev. Shalev's conjecture was proved by Fulman and Guralnick in the series of papers \cite{FG1},\cite{FG2},\cite{FG3}. We do not attempt to sketch a proof of Shalev's conjecture here, but make a few remarks: \begin{enumerate} \item One can assume that the action of $G$ on $X$ is primitive, for if $f:Y \mapsto X$ is a surjection of $G$-sets, then $\delta(G,Y) \geq \delta(G,X)$. \item By Jordan's theorem \cite{Jo} that $\delta(G,X)>0$, the proof of Shalev's conjecture is an asymptotic result: we only need to show that there exists a $\delta>0$ such that for any sequence $G_i,X_i$ with $|X_i| \rightarrow \infty$, one has that $\delta(G_i,X_i)>\delta$ for all sufficiently large $i$. \item When $G$ is the alternating group, by Theorem \ref{theC}, for all primitive actions of $A_n$ except the action on $k$-sets, the proportion of derangements tends to $1$. For the case of $A_n$ on $k$-sets, one can give arguments similar to those Dixon \cite{Dx1}, who proved that the proportion of elements of $S_n$ which are derangements on $k$-sets is at least $\frac{1}{3}$. \item When $G$ is a finite Chevalley group, the key is to study the set of regular semisimple elements of $G$. Typically (there are some exceptions in the orthogonal cases) this is the set of elements of $G$ whose characteristic polynomial is square-free. Now a regular semisimple element is contained in a unique maximal torus, and there is a map from maximal tori to conjugacy classes of the Weyl group. This allows one to relate derangements in $G$ to derangements in the Weyl group. For example, one concludes that the proportion of elements of $GL(n,q)$ which are regular semisimple and fix some k-space is at most the proportion of elements in $S_n$ which fix a k-set. For large $q$, algebraic group arguments show that nearly all elements of $GL(n,q)$ are regular semisimple, and for fixed $q$, one uses generating functions to uniformly bound the proportion of regular semisimple elements away from $0$. \end{enumerate} To close this subsection, we note that the main result of this paper has an analog for finite classical groups. The following result was stated in \cite{FG1} and is proved in \cite{FG2}. \begin{theorem} Let $G_i$ be a sequence of classical groups with the natural module of dimension $d_i$. Let $X_i$ be a $G_i$-orbit of either totally singular or nondegenerate subspaces (of the natural module) of dimension $k_i \le d_i/2$. If $k_i \rightarrow \infty$, then $\lim \delta(G_i,X_i)=1$. If $k_i$ is a bounded sequence, then there exist $0 < \delta_1 < \delta_2 < 1$ so that $\delta_1 < \delta(G_i,X_i) < \delta_2$. \end{theorem} This result applies to any subgroup between the classical group and its socle. Note that in the case that $G_i= PSL$, we view all subspaces as being totally singular (note that the totally singular spaces have parabolic subgroups as stabilizers). We also remark that in characteristic $2$, we consider the orthogonal group inside the symplectic group as the stabilizer of a subspace (indeed, if we view $Sp(2m,2^e)=O(2m+1,2^e)$, then the orthogonal groups are stabilizers of nondegenerate hyperplanes). In fact, Fulman and Guralnick prove an analog of the Luczak-Pyber result for symmetric groups. This result was proved by Shalev \cite{Sh1} for $PGL(d,q)$ with $q$ fixed. \begin{theorem} Let $G_i$ be a sequence of simple classical groups with the natural module $V_i$ of dimension $d_i$ with $d_i \rightarrow \infty$. Let $H_i$ be the union of all proper irreducible subgroups (excluding orthogonal subgroups of the symplectic group in characteristic $2$). Then $\lim_{i \rightarrow \infty} |H_i|/|G_i| = 0$. \end{theorem} If the $d_i$ are fixed, then this result is false. For example, if $G_i=PSL(2,q)$ and $H$ is the normalizer of a maximal torus of $G$, then $\lim_{q \rightarrow \infty} \delta(G, G/H) = 1/2$. However, the analog of the previous theorem is proved in \cite{FG1} if the rank of the Chevalley group is fixed. In this case, we take $H_i$ to be the union of maximal subgroups which do not contain a maximal torus. The example given above shows that the rank of the permutation action going to $\infty$ does not imply that the proportion of derangements tends to $1$. The results of Fulman and Guralnick do show this is true if one considers simple Chevalley groups over fields of bounded size. \subsection{Fixed point ratios} \label{fpr} Previous sections of this paper have discussed $fix(x)$, the number of fixed points of an element $x$ of $G$ on a set $\Omega$. This subsection concerns the fixed point ratio $rfix(x)=\frac{fix(x)}{|\Omega|}$. We describe applications to random generation. For many other applications (base size, Guralnick-Thompson conjecture, etc.), see the survey \cite{Sh2}. It should also be mentioned that fixed point ratios are a special case of character ratios, which have numerous applications to areas such as random walk \cite{D} and number theory \cite{GlM}. Let $P(G)$ denote the probability that two random elements of a finite group $G$ generate $G$. One of the first results concerning $P(G)$ is due to Dixon \cite{Dx2}, who proved that $lim_{n \rightarrow \infty} P(A_n) = 1$. The corresponding result for finite simple classical groups is due to Kantor and Lubotzky \cite{KL}. The strategy adopted by Kantor and Lubotzky was to first note that for any pair $g,h \in G$, one has that $\langle g,h \rangle \neq G$ if and only if $\langle g,h \rangle$ is contained in a maximal subgroup $M$ of $G$. Since $P(g,h \in M) = (|M|/|G|)^2$, it follows that \[ 1 - P(G) \leq \sum_M \left( \frac{|M|}{|G|} \right)^2 \leq \sum_i \left( \frac{|M_i|}{|G|} \right)^2 \left( \frac{|G|}{|M_i|} \right) = \sum_i \frac{|M_i|}{|G|}.\] Here $M$ denotes a maximal subgroup and $\{\it M_i \}$ are representatives of conjugacy classes of maximal subgroups. Roughly, to show that this sum is small, one can use Aschbacher's classification of maximal subgroups \cite{As}, together with Liebeck's upper bounds on sizes of maximal subgroups \cite{Li}. Now suppose that one wants to study $P_x(G)$, the chance that a fixed element $x$ and a random element $g$ of $G$ generate $G$. Then \[ 1 - P_x(G) = P(\langle x,g \rangle \neq G) \leq \sum_{M \ maximal \atop x \in M} P(g \in M) = \sum_{M \ maximal \atop x \in M} \frac{|M|}{|G|}.\] Here the sum is over maximal subgroups $M$ containing $x$. Let $\{M_i\}$ be a set of representatives of maximal subgroups of $G$, and write $M \sim M_i$ if $M$ is conjugate to $M_i$. Then the above sum becomes \[ \sum_i \frac{|M_i|}{|G|} \sum_{M \sim M_i \atop x \in M} 1.\] To proceed further we assume that $G$ is simple. Then, letting $N_G(M_i)$ denote the normalizer of $M_i$ in $G$, one has that \[ g_1 M_i g_1^{-1} = g_2 M_i g_2^{-1} \leftrightarrow g_1^{-1}g_2 \in N_G(M_i) \leftrightarrow g_1^{-1}g_2 \in M_i \leftrightarrow g_1 M_i = g_2 M_i.\] In other words, there is a bijection between conjugates of $M_i$ and left cosets of $M_i$. Moreover, $x \in gM_ig^{-1}$ if and only if $xgM_i = gM_i$. Thus \[ \frac{|M_i|}{|G|} \sum_{M \sim M_i \atop x \in M} 1 = rfix(x,M_i). \] Here $rfix(x,M_i)$ denotes the fixed point ratio of $x$ on left cosets of $M_i$, that is the proportion of left cosets of $M_i$ fixed by $x$. Summarizing, $P_x(G)$ can be upper bounded in terms of the quantities $rfix(x,M_i)$. This fact has been usefully applied in quite a few papers (see \cite{GK}, \cite{FG4} and the references therein, for example). \subsection{Miscellany} \label{miscell} This subsection collects some miscellaneous facts about fixed points and derangements. (1) {\it Formulae for fixed points} We next state a well-known elementary proposition which gives different formulae for the number of fixed points of an element in a group action. \begin{prop} Let $G$ be a finite group acting transitively on $X$. Let $C$ be a conjugacy class of $G$ and $g$ in $C$. Let $H$ be the stabilizer of a point in $X$. \begin{enumerate} \item The number of fixed points of $g$ on $X$ is $\frac{|C \cap H|}{|C|} |X|$. \item The fixed point ratio of $g$ on $X$ is $\frac{|C \cap H|}{|C|}$. \item The number of fixed points of $g$ on $X$ is $|C_G(g)| \sum_i |C_H(g_i)|^{-1}$ where the $g_i$ are representatives for the $H$ classes of $C \cap H$. \end{enumerate} \end{prop} \begin{proof} Clearly (1) and (2) are equivalent. To prove (1), we determine the cardinality of the set $\{(u,x) \in C \times X | ux=x \}$. On the one hand, this set has size $|C| f(g)$ where $f(g)$ is the number of fixed points of $g$. On the other hand, it is $|X| |C \cap H|$, whence (a) holds. For (c), note that $|C|= \frac{|G|}{|C_G(g)|}$ and $|C \cap H| = \sum_i \frac{|H|}{|C_H(g_i)|}$ where the $g_i$ are representatives for the $H$-classes of $C \cap H$. Plugging this into (1) and using $|X|= \frac{|G|}{|H|}$ completes the proof. \end{proof} (2) {\it Algorithmic issues} It is natural to ask for an algorithm to generate a random derangement in $S_n$, for example for cryptographic purposes. Of course, one method is to simply generate random permutations until a derangements is reached. A more closed form algorithm has been suggested by Sam Payne. This begins by generating a random permutation and then, working left to right, each fixed point is transposed with a randomly chosen place. Each such transposition decreases the number of fixed points and a clever non-inductive argument shows that after one pass, the resulting derangement is uniformly distributed. We do not know if this works starting with the identity permutation instead of a random permutation. A very different, direct algorithm for generating a uniformly chosen derangement appears in \cite{De}. There is also a literature on Gray codes for running through all derangements in the symmetric group; see \cite{BV} and \cite{KoL}. \\ (3) {\it Algebraic combinatorics} The set of derangements has itself been the subject of some combinatorial study. For example, D\'{e}sarm\'{e}nien \cite{De} has shown that there is a bijection between derangements in $S_n$ and the set of permutations with first ascent occurring in an even position. This is extended and refined by D\'{e}sarm\'{e}nien and Wachs \cite{DeW}. Diaconis, McGrath, and Pitman \cite{DMP} study the set of derangements with a single descent. They show that this set admits an associative, commutative product and unique factorization into cyclic elements. B\'{o}na \cite{Bn} studies the distribution of cycles in derangements, using among other things a result of E. Canfield that the associated generating function has all real zeros. \\ (4) {\it Statistics} The fixed points of a permutation give rise to a useful metric on the permutation group: the Hamming metric. Thus $d(\pi,\sigma)$ is equal to the number of places where $\pi$ and $\sigma$ disagree. This is a bi-invariant metric on the permutation group and $$d(\pi,\sigma) = d(id,\pi^{-1} \sigma) = \mbox{number of fixed points in \ } \pi^{-1} \sigma.$$ Such metrics have many statistical applications (Chapter 6 of \cite{D}).

https://arxiv.org/abs/2102.12430

Noisy Gradient Descent Converges to Flat Minima for Nonconvex Matrix Factorization

Numerous empirical evidences have corroborated the importance of noise in nonconvex optimization problems. The theory behind such empirical observations, however, is still largely unknown. This paper studies this fundamental problem through investigating the nonconvex rectangular matrix factorization problem, which has infinitely many global minima due to rotation and scaling invariance. Hence, gradient descent (GD) can converge to any optimum, depending on the initialization. In contrast, we show that a perturbed form of GD with an arbitrary initialization converges to a global optimum that is uniquely determined by the injected noise. Our result implies that the noise imposes implicit bias towards certain optima. Numerical experiments are provided to support our theory.

\section{Main Results}\label{main} We study the algorithmic behavior of our proposed perturbed gradient descent (Perturbed GD) algorithm. We primarily focus on the rank-1 nonconvex matrix factorization. We first characterize the landscape of the original problem \eqref{mat_fct_ncvx}, and show that noise effectively smooths the original problem, yielding a smoothed problem \eqref{mat_fact_eq} with benign landscape. We then provide a non-asymptotic convergence analysis of the Perturbed GD, and demonstrate the implicit bias induced by the noise. In particular, we consider the case where noise is balanced, see more details in Theorem \ref{thm_main}. Due to space limit, we defer all the proofs to the appendix. We analyze the landscape of the original problem \eqref{mat_fct_ncvx} and the smoothed problem \eqref{mat_fact_eq}. Note that due to the bilinear form in $\cF(x,y)$, we have for $\alpha \neq 0,$ $\displaystyle\cF\left(\alpha x,\alpha^{-1} y\right)=\cF( x,y).$ The scaling invariance nature of \eqref{mat_fct_ncvx} results in undesired landscape properties, and makes the analysis of first order algorithms particularly difficult. To facilitate further discussions, below we characterize the landscape of \eqref{mat_fct_ncvx}. \begin{lemma}[\bf Landscape Analysis]\label{lem_landscape} The gradients of $\cF$ with respect to $x$ and $y$ take the form: \begin{align*} \nabla_{x}\cF(x,y)=(xy^\top-M)y,~ \nabla_{y}\cF(x,y)=(xy^\top-M)^{\top}x. \end{align*} Then $\cF$ has two types of stationary points: (i) For any $\alpha\neq 0,$ $\left(\alpha u_*, \alpha^{-1} v_*\right)$ is a global optimum; (ii) For any $x\in \RR^{d_1},y\in\RR^{d_2}$ such that $x^\top u_*=y^\top v_*=0,$ $(x,0)$ and $(0,y)$ are strict saddle. \end{lemma} The scaling-invariance leads to infinitely many global optima for \eqref{mat_fct_ncvx}, each taking the form $(\alpha u_*, \alpha^{-1} v_*)$. However, Lemma \ref{lem_condition_number} shows that different values of $\alpha$ lead to significantly different local landscape around the global minima. \begin{lemma}\label{lem_condition_number} The condition number of the Hessian matrix of $\cF$ at the global optimum $(\alpha u_*,\alpha^{-1} v_*)$ is $$\kappa\left(\nabla^2 \cF\left(\alpha u_*,\alpha^{-1} v_*\right)\right)=\max\left\{\alpha^4,\alpha^{-4}\right\}+1.$$ \end{lemma} \begin{figure*} \centering \includegraphics[width=\linewidth]{landscape1} \vspace{-0.2in} \caption{ The visualization of objective functions $\cF(x,y)=(1-xy)^2$ and $\tilde\cF(x,y)$ with $x,y\in \RR$ and $\sigma_1^2=\sigma_2^2=0.0975.$ For $\cF(x,y),$ any $(x,y)$ that satisfies $xy=1$ is a global minimum (as shown in (a)). $\tilde\cF(x,y)$ only has global minima close to $\pm(1,1)$ (as shown in (c)).}\label{fig:landscape} \vspace{-0.1in} \end{figure*} From Lemma \ref{lem_condition_number} we know that when $|\alpha|$ is extremely large or small, the global minimum $(\alpha u_*, \alpha^{-1} v_*)$ is ill-conditioned (large condition number). The condition number also relates to the sharp/flat minima in neural networks, which we will discuss in detail in Section \ref{sec_discussion}. Our previous discussion implies that any global optimum $(x_*, y_*)$ to \eqref{mat_fct_ncvx} will be ill-conditioned if the norm ratio $\norm{x_*}_2/\norm{y_*}_2$ is close to zero or infinity. To facilitate further discussions, we define the balancedness property as follows. \begin{definition}[$\gamma$-balancedness]\label{def:balance} We say that $(x,y)\in(\RR^{d_1},\RR^{d_2})$ is $\gamma-$balanced for some positive number $\gamma$ if $\norm{x}_2=\gamma\norm{y}_2.$ Informally, we say $(x,y)$ is unbalanced if $\gamma$ is close to zero or infinity, and $(x,y)$ is balanced if $\gamma$ is close to 1.\end{definition} Our next lemma shows that, in the presence of noise, the smoothed problem \eqref{mat_fact_eq} has only balanced optima, and the balancedness is completely determined by the ratio of the variance of the noise injected to the iterates. \begin{lemma}\label{lem_convolutional} We say the noise in the Perturbed GD is $\gamma-$balanced if $\EE\left[\norm{\xi_1}_2^2\right]=\gamma^2\EE\left[\norm{\xi_2}_2^2\right]$. For $\gamma = 1$, we say the noise is balanced. For any noise ratio $\gamma>0,$ if we take $\sigma^2=\EE\left[\norm{\xi_1}_2^2\right]=\gamma^2\EE\left[\norm{\xi_2}_2^2\right]\leq \min\{\gamma,1\},$ \eqref{mat_fact_eq} has two global optima $(\tilde u_*(\gamma),\tilde v_*(\gamma))=\pm\sqrt{\gamma-\sigma^2}\left(u_*,\gamma^{-1}v_*\right)$ and one saddle point $(0,0)$ with $\lambda_{\min}(\nabla^2 \tilde F(0,0))<0$. \end{lemma} We see that when the noise is $\gamma$-balanced and the noise level is sufficiently small, the global minima of the smoothed problem \eqref{mat_fact_eq} can be arbitrarily close to the $\gamma$-balanced global minima of the original rank-1 matrix factorization problem \eqref{mat_fct_ncvx}. Compared with Lemma \ref{lem_landscape}, Lemma \ref{lem_convolutional} shows that noise effectively smooths the landscape of the original problem \eqref{mat_fct_ncvx}, the unbalanced optima (with ill-conditioned Hessian) to \eqref{mat_fct_ncvx} are no longer optima after convolution. See Figure \ref{fig:landscape} for detailed illustration of the landscape for the original problem \eqref{mat_fct_ncvx} and the smoothed problem \eqref{mat_fact_eq}. Moreover, one can verify that the saddle point $(0,0)$ of problem \eqref{mat_fact_eq} enjoys the strict saddle property, i.e., the smallest eigenvalue of Hessian at $(0,0)$ is negative. By the stable manifold theory proposed in \citet{lee2019first}, Perturbed GD with an infinitesimal step size and vanishing noise avoids all the strict saddle points and converges to the global optima, asymptotically. However, this asymptotic result does not provide any finite time guarantee. We next present the non-asymptotic convergence analysis of our proposed Perturbed GD algorithm. Our analysis shows that Perturbed GD has implicit bias determined by the injected noise. Specifically, we consider balanced noise, i.e., $\EE\left[\norm{\xi_1}_2^2\right]=\EE\left[\norm{\xi_2}_2^2\right]$, and show that Perturbed GD converges to the balanced $\pm(u_*,v_*)$ in polynomial time in Theorem \ref{thm_main}. \begin{theorem}[\bf Convergence Analysis]\label{thm_main} Suppose $x_0\in \RR^{d_1},$ $y_0\in \RR^{d_2}.$ For any $\epsilon>0$ and for any $\delta\in(0,1),$ we take $\sigma^2=\EE\left[\norm{\xi_1}_2^2\right]=\EE\left[\norm{\xi_2}_2^2\right]=\rm{poly}(\epsilon,(\log(1/\delta))^{-1}),$ $$\eta={\rm{poly}}( \sigma,\epsilon, (d_1+d_2)^{-1},(\log{(d_1+d_2)})^{-1},(\log(1/\delta))^{-1}).$$ With probability at least $1-\delta,$ we have $\norm{x_t-u_*}_2\leq \epsilon$ and $\norm{y_t-v_*}_2\leq \epsilon.$ for all $t_1\leq t\leq T_1= O\left({1}/{\eta^2}\right),$ where $t_1=O\left(\eta^{-1} \sigma^{-2}\log(1/\eta )\log(1/\delta)\right).$ \end{theorem} Theorem \ref{thm_main} differs from the convergence analysis of GD in \citet{du2018algorithmic} in the following aspects: (i) Perturbed GD converges regardless of initializations, while GD requires a random initialization near $(0,0)$; (ii) Perturbed GD guarantees convergence with high probability, while GD converges with only constant probability over randomness of initializations; (iii) Perturbed GD converges to the balanced solutions ($\norm{x}_2/\norm{y}_2=1$) while GD only maintains approximate balancedness, i.e., $c_0\leq\norm{x}_2/\norm{y}_2\leq C_0$ for some absolute constants $c_0, C_0>0.$ We provide a proof sketch that contains the essential ingredients of characterizing the convergence, since the proof of Theorem \ref{thm_main} is very technical and highly involved. See more details and the proof of technical lemmas in Appendix \ref{pf_2}. \begin{proof}[Proof Sketch] The convergence of Perturbed GD consists of three phases: {\bf Phase I}: Regardless of initialization, within polynomial time the noise encourages Perturbed GD to escape from region with undesired landscape (e.g., strict saddle points) and enter the region with benign landscape. {\bf Phase II}: Perturbed GD drives the loss to zero and approaches the set of global minima $\left\{(x,y)\big|xy^\top=M\right\}.$ {\bf Phase III}: After the loss is sufficiently small, the injected noise dominates the update, which helps the Perturbed GD to balance $x$ and $y,$ and converge to a balanced optimum. Before we proceed with our proof, we first define some notations. Let $U$ and $V$ denote the linear span of $u_*$ and $v_*$, respectively, i.e., $ \cU=\{\alpha_1 u_*:\alpha_1\in\RR\},~~\cV=\{\alpha_1 v_*:\alpha_1\in\RR\}. $ The corresponding orthogonal complement of $\cU$ (or $\cV$) in $\RR^{d_1}$ (or $\RR^{d_2}$) is denoted as $\cU^\perp$ (or $\cV^\perp$). Our analysis considers the convergence of Perturbed GD in $(\cU,\cV)$ and $(\cU^\perp,\cV^\perp),$ respectively. Specifically, we take the following orthogonal decomposition of $x_t$ and $y_t$: \begin{align*} x_t&=u_*^\top x_t u_*+(x_t-u_*^\top x_t u_*),\\y_t&=v_*^\top y_t v_*+(y_t-v_*^\top y_t v_*). \end{align*} One can check that $(x_t-u_*^\top x_t u_*)\in\cU^\perp$ and $(y_t-v_*^\top y_t v_*)\in \cV^\perp,$ respectively. \noindent $\bullet$ {\bf Phase I}: Regardless of initializations, the following lemma shows that $(x_t-u_*^\top x_t u_*)$ and $(y_t-v_*^\top y_t v_*)$ vanish after polynomial time. \begin{lemma}\label{lem_b_converge} Suppose $\norm{x_t}^2_2+\norm{y_t}^2_2\leq 2/\sigma^2$ holds for all $t>0.$ For any $\delta\in(0,1)$ we take $$\eta\leq \eta_2=C_4{\sigma^8}\left(\log((d_1+d_2)/\delta)\log(1/\delta)\right)^{-1},$$ where $C_4$ is some positive constant. Then with probability at least $1-\delta,$ we have \begin{align*} \norm{x_t-u_*^\top x_t u_*}_2^2&\leq 2\eta C_2\sigma^{-2},\\ \norm{y_t-v_*^\top y_t v_*}_2^2&\leq 2\eta C_2\sigma^{-2} \end{align*} for any $\tau_1\leq t\leq T_1= O(1/\eta^2),$ where $\tau_1=O(\eta^{-1}\sigma^{-2}\log(1/\eta)\log(1/\delta))$ and $C_2=(\sigma^2+1/\sigma^2)(2/\sigma^4+6/d_1+6/d_2+6\sigma^4).$ \end{lemma} Lemma \ref{lem_b_converge} shows that with an arbitrary initialization, the projection of $x_t$ (and $y_t$) onto $\cU^\perp$ (and $\cV^\perp$) vanishes. Thus, we only need to characterize the algorithmic behavior of Perturbed GD in subspace $(\cU,\cV).$ However, as the projection of $x_t$ (and $y_t$) onto $\cU^\perp$ (and $\cV^\perp$) vanishes, Perturbed GD can possibly approach the region with undesired landscape, e.g., the small neighborhood around the saddle point. The next lemma shows that the Perturbed GD will escape such regions within polynomial time. \begin{lemma}\label{lem_escape} Suppose $\norm{x_t-u_*^\top x_t u_*}_2^2\leq 2\eta C_2\sigma^{-2}$ and $\norm{y_t-v_*^\top y_t v_*}_2^2\leq 2\eta C_2\sigma^{-2}$ hold for all $t>0.$ For any $\delta\in(0,1),$ we take $\eta\leq\eta_3=C_3\sigma^{12}\left(\log((d_1+d_2)/\delta)\log(1/\delta)\right)^{-1},$ where $C_3$ is some positive constant. Then with probability at least $1-\delta,$ we have \begin{align}\label{eq_escape} x_{t}^\top u_* v_*^\top y_{t}=x_{t}^\top M y_{t}\geq 1/4, \end{align} for all $\tau_2\leq t\leq T_1=O\left({1}/{\eta^2}\right),$ where $\tau_2=O\left(\eta^{-1}\sigma^{-2}\log(1/\eta)\log(1/\delta)\right).$ \end{lemma} Since the saddle point $(x,y)=(0,0)$ satisfies $x^\top M y = 0,$ Lemmas \ref{lem_b_converge} and \ref{lem_escape} together imply that regardless of initializations, Perturbed GD is bounded away from the saddle point after polynomial time. The algorithm then enters Phase II and approaches the set of global optima $\left\{(x,y)\big|xy^\top=M\right\}.$ \noindent $\bullet$ {\bf Phase II}: In this phase, $(x_t,y_t)$ is still away from the region $\left\{(x,y)\big|xy^\top=M\right\}.$ Thus, $\nabla\cF(x_{t},y_{t})$ dominates the update of Perturbed GD. Perturbed GD behaves similarly to gradient descent while driving the loss to zero. The next lemma formally characterizes this behavior, showing that $xy^ \top$ converges to $M.$ \begin{lemma}\label{lem_loss} Suppose $x_{t}^\top u_* v_*^\top y_{t}\geq\frac{1}{4}$ holds for all $t>0.$ For any $\epsilon>0$ and for any $\delta\in(0,1)$, we choose $\sigma\leq\sigma'_1=C_4\sqrt{\epsilon}$ and take $\eta\leq \eta_4={C}_5{\sigma^6}\left(\log((d_1+d_2)/\delta)\log(1/\delta)\right)^{-1},$ where $C_4,C_5$ are some positive constants. Then with probability at least $1-\delta,$ we have $$\norm{x_ty_t^\top-M}_{\rm{F}}\leq \epsilon,$$ for all $\tau_3\leq t\leq T_1= O\left({1}/{\eta^2}\right),$ where $\tau_3=O\left(\eta^{-1}\log\frac{1}{\sigma}\log(1/\delta)\right).$ \end{lemma} \noindent $\bullet$ {\bf Phase III}: After Phase II, Perturbed GD enters the region where $x_ty_t^\top\approx M.$ Thus, the noise will dominate the update of the Perturb GD. Recall in Lemma \ref{lem_condition_number}, we show that the Hessian of unbalanced optima has a large condition number, hence a small perturbation will significantly change the gradient of the objective. This implies that the unbalanced optima would be unstable against the noise, and Perturbed GD will escape from such optima and continue iterating towards the balanced optima. The next lemma shows that $x_t $ and $y_t $ converge to $u_*, v_*,$ respectively. \begin{lemma}\label{lem_convergence} For $\forall\epsilon>0$, suppose $\norm{x_ty_t^\top-M}_{\rm{F}}\leq \epsilon$ holds for all $t>0.$ For any $\delta\in(0,1)$, we choose $\sigma\leq\sigma'_2={C}_6\left(\log(1/\delta)\right)^{-1/3}$ and take $\eta\leq \eta_5=C_7\sigma^{10}\epsilon,$ where $C_6, C_7$ are some positive constants. Then with probability at least $1-\delta,$ we have \begin{align*} \norm{x_t -u_*}_2\leq \epsilon,~\norm{y_t-v_*}_2\leq \epsilon, \end{align*} for all $\tau_3\leq t\leq T_1= O\left({1}/{\eta^2}\right),$ where $\tau_4=O\left(\eta^{-1}\sigma^{-2}\log\eta^{-1}\log(1/\delta)\right).$ \end{lemma} With all the lemmas in place, we take $\sigma\leq \min\{\sigma'_1,\sigma'_2\},$ $\eta\leq \min\{\eta_1,\eta_2,\eta_3,\eta_4,\eta_5\}$ and $t_1=\tau_1+\tau_2+\tau_3+\tau_4$, then the claim of Theorem \ref{thm_main} follows immediately. \end{proof} \vspace{-0.12in} \section{Extension to Rank-r Matrix Factorization}\label{sec_extend} \vspace{-0.1in} We extend our results to the rank-$r$ matrix factorization, which solves the following problem: \begin{align}\label{rankr_mf} \min_{X \in \RR^{d_1 \times r},Y \in \RR^{d_2 \times r}} \cF(X,Y) = \frac{1}{2}\norm{XY^\top - M}_{\rm{F}}^2, \end{align} where $M\in \RR^{d_1\times d_2}$ is a rank-$r$ matrix. Let $M=A\Sigma B^\top$ be the SVD of $M.$ Let $U_*=A\Sigma^{\frac{1}{2}}$ and $V_*=B\Sigma^{\frac{1}{2}},$ then $(U_*, V_*)$ is a global minimum for problem \eqref{rankr_mf}. Similar to the rank-1 case, using Perturbed GD to solve problem \eqref{rankr_mf} can be viewed as solving the following smoothed problem: \begin{align}\label{rankr_eq} \min_{X,Y}\tilde\cF(X,Y)=\EE_{\xi_{1},\xi_{2}}\cF(X+\xi_1,Y+\xi_2), \end{align} where $\xi_{1}\in \RR^{d_1\times r},$ $\xi_{2}\in \RR^{d_2\times r}$ have i.i.d. elements drawn from $N(0,\sigma_1^2)$ and $N(0,\sigma_2^2),$ respectively. The next theorem shows that the noise in Perturbed GD effectively addresses the scaling invariance issue of \eqref{rankr_mf}. The smoothed problem \eqref{rankr_eq} only has balanced global minima. \begin{theorem}\label{thm_rankr} Let $\sigma_{\min}(M)$ be the smallest singular value of $M$. Suppose $$\EE\left[\norm{\xi_1}_{\rm{F}}^2\right]=\gamma^2\EE\left[\norm{\xi_2}_{\rm{F}}^2\right]=r\gamma^2\sigma^2,$$ and $\gamma\sigma^2<\sigma_{\min}(M).$ Then for $\forall(U,V)\in (\RR^{d_1},\RR^{d_2})$ such that $\nabla\tilde\cF(U,V)=0,$ we have $U^\top U=\gamma^2 V^\top V.$ Moreover, denote $(\tilde U,\tilde V)=\left(\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}},\gamma^{-1/2}B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}\right),$ then the set $\{(\tilde U,\tilde V)R \big|R \in \RR^{r\times r}, RR^\top=R^\top R=I_r \}$ contains all the global optima. All other stationary points are strict saddles, i.e., $\lambda_{\min}\left(\nabla^2 \tilde\cF(U,V)\right)<0.$ \end{theorem} Theorem \ref{thm_rankr} shows that when noise is balanced, \eqref{rankr_eq} only has balanced global optima and strict saddle points. We can invoke \citet{lee2019first} again and show Perturbed GD with an infinitesimal step size and vanishing noise converges to the balanced global optima, asymptotically. Note that compared to the rank-1 case, in addition to scaling invariance, the objective is also rotation invariant, i.e., $\tilde\cF(X,Y)= \tilde\cF(XR,YR),$ where $R\in \RR^{r\times r}$ is an orthogonal matrix. Thus, we can only recover $U_*$ and $V_*$ up to a rotation factor. To establish the non-asymptotic convergence result, the optimization error should be measured by the following metric that is rotation invariant: \begin{align*} \mathrm{dist}_\cR(D_1,D_2)=\min_{R\in \RR^{r\times r}:RR^\top=I_r}\norm{D_1-D_2R}_{\rm{F}}. \end{align*} However, it is more challenging and involved to handle the complex nature of the distance $\mathrm{dist}_\cR(\cdot, \cdot)$. As our results for the rank-1 case already provide insights on understanding the implicit bias of noise, we leave the non-asymptotic analysis of rank-r matrix factorization for future investigation. \section{Discussions}\label{sec_discussion} \vspace{-0.1in} \noindent {\bf Connections to SGD.} This paper studies the implicit bias of the noise in nonconvex optimization. Direct analysis on SGD is beyond current technical limit due to complex dependencies. Specifically, the noise in SGD comes from random sampling of the training data, and heavily depends on the iterate. This induces a complex dependency between iterate and the noise, and makes it difficult to characterize the distribution of the noise. Our Perturbed GD can be viewed as a close variant of SGD. Moreover, the noise in Perturbed GD follows Gaussian distribution and is independent of the iterates. Hence, the analysis of Perturbed GD, though still highly non-trivial, is now technically manageable. \noindent {\bf Biased Gradient Estimator.} Different from SGD, Perturbed GD implements a biased gradient estimator, i.e., $ \EE_{\xi_{1},\xi_{2}}\nabla\cF(x+\xi_1,y+\xi_2)\neq\nabla\cF(x,y).$ Such biased gradient also appears in training deep neural networks combined with computational heuristics. Specifically, \cite{luo2018towards} show that with batch normalization, the gradient estimator in SGD is also biased with respect to the original loss. The similarity between this biased gradient and our perturbed gradient is worth future investigation. \noindent {\bf Extension to Other Types of Noise.} Our work considers Gaussian noise, but can be extended to analyzing other types of noise. For example, we can show that anisotropic noise will have different smoothing effects along different directions. The implicit bias will thus depend on the covariance of noise in addition to the noise level. For another example, heavy tailed distribution of noise will affect the probability of the convergence of Perturbed GD. It may be difficult to achieve high probability convergence as we have shown for light tailed distributions. \noindent {\bf Sharp/Flat Minima and Phase Transition.} Lemma \ref{lem_condition_number} shows that the Hessian matrix of an unbalanced global optimum is ill-conditioned, and the landscape around such an optimum is sharp in some directions and flat in others. For nonconvex matrix factorization, all the global optima are connected and form a path. The landscape around the path forms a valley, which is narrow around unbalanced optima and wide around the balanced ones (See Figure \ref{fig:landscape2}). Our three-phase convergence analysis shows that the Perturbed GD first falls into the valley, and then traverses within the valley until it finds the balanced optima. \begin{figure}[t] \centering \includegraphics[width=0.95\linewidth]{landscape4} \caption{ The visualization of objective $\cF(x,y)=(1-xy)^2.$ All the global optima are connected and form a path. The landscape around the path forms a valley. Around unbalanced optima, the landscape is sharp in some directions and flat in others. Around balanced optima, the landscape only contains flat directions.} \label{fig:landscape2} \end{figure} As we have mentioned earlier, people have shown that the local optima of deep neural networks are also connected. Thus, the phase transition also provides a new perspective to explain the plateau of training curves after learning rate decay in training neural networks. Our analysis suggests that after adjusting the learning rate, the algorithm enters a new phase, where the noise slowly re-adjusts the landscape. At the beginning of this phase, the loss decreases rapidly due to the reduced noise level. By the end of this phase, the algorithm falls into a region with benign landscape that is suitable for further decreasing the step size. \noindent {\bf Related Literature.} Implicit bias of noise has also been studied in \citet{haochen2020shape,blanc2020implicit}. However, they consider perturbing labels while our work considers perturbing parameters. In a broader sense, our work is also related to \citet{li2019towards} which show that the noise scale of SGD may change the learning order of patterns. However, they do not study the implicit bias of noise towards certain optima. \section{Introduction} Nonconvex optimization has been widely adopted in various domains, including image recognition \citep{hinton2012deep, krizhevsky2012imagenet}, Bayesian graphical models \citep{jordan2004graphical, attias2000variational}, recommendation systems \citep{salakhutdinov2007restricted}, etc. Despite the fact that solving a nonconvex problem is generally difficult, empirical evidences have shown that simple first order algorithms such as stochastic gradient descent (SGD), are able to solve a majority of the aforementioned nonconvex problems efficiently. The theory behind these empirical observations, however, is still largely unexplored. In classical optimization literature, there have been fruitful results on characterizing the convergence of SGD to first-order stationary points for nonconvex problems. However, these types of results fall short of explaining the empirical evidences that SGD often converges to global minima for a wide class of nonconvex problems used in practice. More recently, understanding the role of noise in the algorithmic behavior of SGD has received significant attention. For instance, \citet{jin2017escape} show that a perturbed form of gradient descent is able to escape from strict saddle points and converge to second-order stationary points (i.e., local minima). \citet{zhou2019towards} further show that noise in the update can help SGD to escape from spurious local minima and converge to the global minima. We argue that, despite all these recent results on showing the convergence of SGD to global minima for various nonconvex problems, there is still an important question yet to be addressed. Specifically, the convergence of SGD in the presence of multiple global minima remains uncleared for many important nonconvex problems. For example, for over-parameterized neural networks, it has been shown that the nonconvex objective has multiple global minima \citep{kawaguchi2016deep}, only few of which can yield good generalization. In addition, \citet{allen2018convergence} show that for an over-parameterized network, SGD can converge to a global minimum in polynomial time. Combining with the empirical successes of training over-parameterized neural networks with SGD, these results strongly advocate that SGD not only can solve the nonconvex problem efficiently, but also implicitly biases towards solutions with good generalization ability. Motivated by this, this paper aims to provide more theoretical insights to the following question: \begin{center} \textbf{\emph{Does noise impose implicit bias towards certain minimizer in nonconvex optimization problems?}} \end{center} We answer this question through investigating a simple yet non-trivial problem -- nonconvex matrix factorization, which serves as an important foundation for a wide spectrum of problems such as matrix sensing \citep{bhojanapalli2016dropping,zhao2015nonconvex,chen2015fast,tu2015low}, matrix completion \citep{keshavan2010matrix,hardt2014understanding,zheng2016convergence}, and deep linear networks \citep{ji2018gradient, gunasekar2018implicit}. Given a matrix $M \in \RR^{d_1 \times d_2} $, the nonconvex matrix factorization aims to solve: \begin{align}\label{mf_obj} \min_{X\in \RR^{d_1 \times r},Y \in \RR^{d_2 \times r}} \frac{1}{2}\norm{XY^\top - M}_{\rm{F}}^2. \end{align} Despite its simplicity, \eqref{mf_obj} possesses several intriguing landscape properties: nonconvexity of the objective, all the saddle points satisfy the strict saddle property, and infinitely many global optima due to scaling and rotational invariance. Specifically, for any pair of global optimum $(X^*, Y^*)$, $(\alpha X^*, \frac{1}{\alpha} Y^*)$ and $(X^*R, Y^*R)$ are also global optima for any non-zero constant $\alpha$ and rotation matrix $R \in \RR^{r \times r}$. This is different from symmetric matrix factorization, which only possesses rotational invariance. The scaling and rotational invariance also imply that the global minima of \eqref{mf_obj} are connected, a landscape property that is also shared by deep neural networks \citep{nguyen2017loss, draxler2018essentially,nguyen2018loss,venturi2018spurious,garipov2018loss,liang2018understanding,nguyen2019connected,kuditipudi2019explaining}. Nonconvex matrix factorization \eqref{mf_obj} has been recently studied by \citet{du2018algorithmic}, with focus on the algorithmic behavior of gradient descent (GD). Their results reveal an interesting algorithmic regularization imposed by gradient descent: (i) gradient flow (GD with an infinitesimal step size) has automatic balancing property, i.e., the difference of the squared norm $\norm{X}_{\rm{F}}^2 - \norm{Y}_{\rm{F}}^2$ stays constant during training. (ii) for properly chosen step size, GD converges asymptotically for rank-r case, linearly for rank-1 case, while maintaining approximate balancing property. However, \citet{du2018algorithmic} do not consider any noise in the update, hence their results can not provide further theoretical insights on understanding the role of noise when applying first order algorithms to nonconvex problems. In this paper, we are interested in studying the algorithmic behavior of first order algorithms in the presence of noise. Specifically, we study a perturbed form of gradient descent (Perturbed GD) applied to the matrix factorization problem \eqref{mf_obj}, which injects independent noise to iterates, and then evaluates gradient at the perturbed iterates. Note that our algorithm is different from SGD in terms of the noise. For our algorithm, we inject independent noise to the iterates $(X_t, Y_t)'s$ and use the gradient evaluated at the perturbed iterates. The noise of SGD, in contrast, comes from the training sample. As a consequence, the noise of SGD has very complex dependence on the iterate, which is difficult to analyze. See more detailed discussions in Sections \ref{sec_discussion}. We further analyze the convergence properties of our Perturbed GD algorithm for the rank-1 case. At the early stage, noise helps the algorithm to escape from regions with undesired landscape, including the strict saddle point. After entering the region with benign landscape, Perturbed GD behaves similarly to gradient descent, until the loss is sufficiently small. At the early stage, noise provides additional explorations that help the algorithm to escape from the strict saddle point. Then at the later stage, the noise dominates the update of Perturbed GD, and gradually rescales the iterates to a balanced solution that is uniquely determined by the injected noise. Specifically, the ratio of the norm $\norm{x_t}_2/\norm{y_t}_2$ is completely determined by the ratio of the variance of noise injected to $(x_t, y_t)$. To the best of our knowledge, this is the first theoretical result towards understanding the implicit bias of noise in nonconvex optimization problems. Our analysis reveals an interesting characterization of the local landscape around global minima, which relates to the sharp/flat minima in deep neural networks \citep{keskar2016large}, and we will further discuss these connections in detail in Section \ref{sec_discussion}. We believe that investigating the implicit bias of the noise in nonconvex matrix factorization can serve as a fundamental building block for studying stochastic optimization for more sophiscated nonconvex problems, including training over-parameterized neural networks. \noindent{\textbf{Notations}}: $\mathbf{1}$ Given a matrix $A$, $\tr(A)$ denotes the trace of $A.$ For matrices $A,B\in \RR^{n\times m},$ we use $\inner{A}{B}$ to denote the Frobenius inner product, i.e., $\inner{A}{B}=\tr(A^\top B).$ $\norm{A}_{\rm{F}}=\sqrt{\inner{A}{A}}$ denotes the Frobenius norm of $A.$ $I_d\in \RR^{d\times d}$ denotes the identity matrix. \subsection{ Rank-1 Matrix Factorization} We consider the following nonconvex optimization problem: \begin{align}\label{mat_fct_ncvx} \min_{x\in\RR^{d_1}, y\in\RR^{d_2}}\cF(x,y)=\frac{1}{2}\|xy^\top-M\|^2_{\rm{F}}, \end{align} where $M \in \RR^{d_1\times d_2}$ is a rank-1 matrix and can be factorized as follows: $\displaystyle M=u_*v_*^\top, $ where $u_*\in\RR^{d_1}$, $v_*\in\RR^{d_2}$. Without loss of generality, we assume $\norm{u_*}_2=\norm{v_*}_2=1.$ The optimization landscape of \eqref{mat_fct_ncvx} has been well studied in the previous literature \citep{ge2016matrix, ge2017no, chi2019nonconvex, li2019symmetry}. Because of the bilinear form in $\cF,$ there exist infinitely many global minima to \eqref{mat_fct_ncvx}, including highly unbalanced ones, i.e., $xy^\top=M$ with $\norm{x}_2\gg\norm{y}_2$ or $\norm{x}_2\gg\norm{y}_2$ (see Definition \ref{def:balance}). In Section \ref{main}, we will show that such unbalancedness essentially implies global minima with a large condition number. To address this issue, \citet{tu2015low,ge2017no} propose a regularizer of the form $(\norm{x}_2^2-\norm{y}_2^2)^2$ to balance $\norm{x}_2$ and $\norm{y}_2$. Recently, \citet{du2018algorithmic} show that even without explicit regularization, gradient descent with small random initialization converges to balanced solutions with constant probability. Yet, all of the previous results assume noiseless updates. The algorithmic behavior of first order algorithms with noisy updates remains unclear for the nonconvex matrix factorization problem. \subsection{Perturbed Gradient Descent} To study the effect of noise, we consider a perturbed gradient descent algorithm (Perturbed GD). At the $t$-th iteration, we first perturb the iterate $(x_t,y_t)$ with independent Gaussian noise $\xi_{1,t}\sim N(0,\sigma_1^2I_{d_1})$ and $\xi_{2,t}\sim N(0,\sigma_2^2I_{d_2}),$ respectively. We then update $(x_t, y_t)$ with the gradient evaluated at the perturbed iterates. The detail of the Perturbed GD algorithm is summarized in Algorithm \ref{alg:Perturbed GD}. \begin{algorithm}[H] \caption{Perturbed Gradient Descent for Rank-1 Matrix Factorization.} \label{alg:Perturbed GD} \begin{algorithmic} \STATE{\textbf{Input}: step size $\eta$, noise level $\sigma_1, \sigma_2$, matrix $M \in \RR^{d_1 \times d_2}$, number of iterations $T$.} \STATE{\textbf{Initialize}: initialize $(x_0, y_0)$ arbitrarily.} \FOR{$t = 0 \ldots T-1$} \STATE{Sample $\xi_{1,t} \sim N(0, \sigma_1^2 I_{d_1})$ and $\xi_{2,t} \sim N(0, \sigma_2^2 I_{d_2})$.} \STATE{$\tilde{x}_t = x_t + \xi_{1,t}, ~ \tilde{y}_t = y_t + \xi_{2,t}.$} \STATE{$x_{t+1} = x_t - \eta (\tilde{x}_t \tilde{y}_t^\top - M ) \tilde{y}_t.$} \STATE{$y_{t+1} = y_t - \eta (\tilde{y}_t \tilde{x}_t^\top - M^\top ) \tilde{x}_t.$} \ENDFOR \end{algorithmic} \end{algorithm} \textbf{Smoothing Effect}. Using Perturbed GD to solve \eqref{mat_fct_ncvx} can also be viewed as solving the following stochastic optimization problem: \begin{align}\label{mat_fact_eq} \min_{x\in\RR^{d_1}, y\in\RR^{d_2}}\tilde\cF(x,y)=\EE_{\xi_{1},\xi_{2}}\cF(x+\xi_1,y+\xi_2), \end{align} where $\xi_{1}\sim N(0,\sigma_1^2 I_{d_1})$ and $\xi_{2}\sim N(0,\sigma_2^2 I_{d_2}).$ Throughout this paper, we will refer to problem \eqref{mat_fact_eq} as the smoothed problem. The expectation in \eqref{mat_fact_eq} can be viewed as convoluting the objective function with a Gaussian kernel. In Section \ref{main}, we show that this convolution effectively smooths out unbalanced optima and yields a benign landscape. \begin{remark} The use of random noise to convolute with the objective function is also known as randomized smoothing, which is first proposed in \citet{duchi2012randomized}. \citet{zhou2019towards,jin2017escape} further exploit this effect to explain the importance of noise in helping first order algorithms to escape from strict saddle points and spurious local optima. \end{remark} \section{Numerical Experiments}\label{sec_numerical} We present numerical results to support our theoretical findings. We compare our Perturbed GD algorithm with gradient descent (GD), and demonstrate that Perturbed GD with $\gamma-$balanced noise converges to $\gamma-$balanced optima, while the optima obtained by GD are highly sensitive to initialization and step size. We also show that the phase transition between Phase II and Phase III is not an artifact of the proof, and faithfully captures the true algorithmic behavior of Perturbed GD. \noindent\textbf{Rank-1 Matrix Factorization}. We first consider the rank-1 matrix factorization problem. Without loss of generality, the matrix $M$ to be factorized is given by $M = u_* v_*^\top$, where $u_* = (1, 0, \ldots, 0)\in \RR^{d_1}$ and $v_* = (1, 0, \ldots, 0)\in \RR^{d_2}$, with $d_1 = 20$ and $d_2 = 30$. We initialize iterates $(x_0, y_0)$ with $x_0 \sim N(0,\sigma_x^2 I_{d_1}),\ y_0 \sim N(0,\sigma_y^2 I_{d_2})$. For all experiments, we use $\gamma$-balanced noise in Perturbed GD. Specifically, we choose $\xi_{1,t} \sim N(0, \sigma_1^2 I_{d_1}), \xi_{2,t} \sim N(0, \sigma_2^2 I_{d_2})$. \noindent\textbf{(1) Balanced Noise}. We consider the case of balanced noise ($\gamma =1$), where we take $\sigma_1 = \sqrt{1.5} \times 0.05$ and $\sigma_2 = 0.05$. One can verify that $\gamma^2 = {d_1 \sigma_1^2}/{(d_2 \sigma_2^2)} = 1$. We first use balanced step size to compare with GD studied in \citet{du2018algorithmic}. Specifically, we set $\eta_x = \eta_y = 10^{-2}$ for both Perturbed GD and GD. We further consider two initialization schemes: (i) small initializations (which is also adopted in \citet{du2018algorithmic}): $\sigma_x = \sigma_y = 10^{-2}$; (ii) large initializations: $\sigma_x = \sigma_y = 10^{-1}$. Fig. \ref{mainresult}.(a, b, d, e) summarize the results of $100$ repeated experiments in a box-plot. As can be seen, regardless of initializations, Perturbed GD always converges to the balanced optima (Fig. \ref{mainresult}.(a, b)). In contrast, GD only converges to approximately balanced optima for small initializations (Fig. \ref{mainresult}.(d)), and the large initialization yields a large variance in terms of the balancedness of the obtained solution (Fig. \ref{mainresult}.(e)). \begin{figure*}[t] \centering \includegraphics[width=0.9\linewidth]{box-rk1.pdf} \caption{Perturbed GD with balanced noise (a, b and c), Perturbed GD with unbalanced noise (g) and GD (d,e and f) for the rank-1 matrix factorization problem. (a) and (d) use small initializations ($\sigma_x = \sigma_y = 10^{-2}$) and balanced step size ($\eta_x = \eta_y = 10^{-2}$). (b) and (e) use large initializations ($\sigma_x = \sigma_y = 10^{-1}$) and balanced step size. (c) and (f) use small initializations and unbalanced step size ($\eta_x = 0.5 \eta_y =5 \times 10^{-3}$).}\label{mainresult} \end{figure*} We also use unbalanced step size (different step sizes for updating $x$ and $y$) and compare the convergence properties of Perturbed GD and GD. Specifically, we set $\eta_x = 0.5 \eta_y =5 \times 10^{-3}$ for both Perturbed GD and GD. We adopt a small initialization scheme, with $\sigma_x = \sigma_y = 10^{-2}$. Fig. \ref{mainresult}.(c, f) summarize the results of $100$ repeated simulations in a box-plot. As can be seen, even with small initializations, GD with unbalanced step size converges to the approximately $\sqrt{0.5}$-balanced optima, instead of the $1$-balanced optima (Fig. \ref{mainresult}.(f)). In contrast, Perturbed GD is able to converge to the balanced optima with unbalanced step size (Fig. \ref{mainresult}.(c)). Our results suggest that the noise is the most important factor in determining the balancedness of the solutions obtained by Perturbed GD. \noindent\textbf{(2) Unbalanced Noise}. We run Perturbed GD with unbalanced noise. We take $\sigma_1 = \sqrt{0.75} \times 0.05$ and $\sigma_2 = 0.05$ with $\gamma^2 ={d_1 \sigma_1^2}/ {(d_2 \sigma_2^2)}= 0.5$. We use a small initialization: $\sigma_x = \sigma_y = 10^{-2}$, and balanced step size: $\eta_x = \eta_y = 10^{-2}$. Fig. \ref{mainresult}.(g) summarizes the results of $100$ repeated simulations in a box-plot. As can be seen, for $\gamma \neq 1$, the Perturbed GD converges to the $\gamma$-balanced optima. \noindent \textbf{Rank-10 Matrix Factorization}. We then consider rank-10 nonconvex matrix factorization problem. The matrix $M$ to be factorized is given by $M = U_* V_*^\top$, where $ U_* =\begin{pmatrix} I_{10}, 0 \end{pmatrix}_{d_1\times10}^\top, V_* =\begin{pmatrix} I_{10}, 0 \end{pmatrix}_{d_2\times10}^\top, $ with $d_1 = 20$ and $d_2 = 30$. We initialize iterates $(X_0, Y_0)$ with all entries ${X_0}^{(i,j)}$'s and ${Y_0}^{(i,j)}$'s independently sampled from $N(0,\sigma_x^2)$ and $N(0,\sigma_y^2)$, respectively. For all experiments, we use $\gamma$-balanced noise in Perturbed GD. Specifically, we choose $\xi_{1,t}$ and $\xi_{2,t}$ with i.i.d. elements drawn from $N(0,\sigma_1^2)$ and $N(0,\sigma_2^2)$ respectively. We repeat a similar set of experiments as in rank-1 case, and summarize the results in Fig. \ref{mainresultrk-10}. For each of the experiments, we use the same set of $(\eta_x, \eta_y, \sigma_x, \sigma_y)$ as their counterpart in the rank-1 case. As can be seen, Perturbed GD always converges to the $\gamma-$balanced optima, regardless of initializations. Our experiments suggest that, for the rank-r matrix factorization problem, the noise still determines the balancedness of the optima obtained by Perturbed GD. \begin{figure*}[t] \centering \includegraphics[width=0.9\linewidth]{box-rk10.pdf} \caption{Perturbed GD with balanced noise (a, b and c) and GD (d,e and f) for rank-10 matrix factorization problem. (a) and (d) use small initializations ($\sigma_x = \sigma_y = 10^{-2}$) and balanced step size ($\eta_x = \eta_y = 10^{-2}$). (b) and (e) use large initializations ($\sigma_x = \sigma_y = 10^{-1}$) and balanced step size. (c) and (f) use small initializations and unbalanced step size ($\eta_x = 0.5 \eta_y =5 \times 10^{-3}$).}\label{mainresultrk-10} \end{figure*} \noindent \textbf{Phase Transition.} We further demonstrate the transition between Phase II and Phase III in the Perturbed GD algorithm. Specifically, we consider 2-dimensional problem $f(x,y) = (1-xy)^2$ with balanced optima $\pm(1,1)$. We set $\sigma_1 = \sigma_2 = 0.05$, initialize $(x_0 ,y_0) = (3,5)$, and use balanced step size $\eta_x = \eta_y = 0.01$. We repeat the experiments $50$ times and summarize the result of one realization in Fig. \ref{2phase}, as the convergence properties of Perturbed GD are highly consistent across different realizations. We also use exponential moving average to smooth the loss trajectory to better illustrate the overall progress of the objective in Phase III. As can be seen, in around the first $30$ to $40$ iterations, Perturbed GD and GD behave similarly. Both Perturbed GD and GD iterate towards the set of global optima $\{(x,y)\big| xy=1\}$, while driving the loss to zero, and the squared norm ratio $x_t^2/y_t^2$ in Perturbed GD decreases from $0.36$ to around $0.004$. \begin{figure}[t] \includegraphics[width=0.9\linewidth]{phase.pdf} \caption{Algorithmic behaviors of Perturbed GD and GD. For Perturbed GD, phase transition happens around the first $30\!\sim\!40$ iterations, as shown in (a,b,c). GD does not show phase transitions. }\label{2phase} \end{figure} After that, GD converges to the unbalanced optimum. Since the loss is sufficiently small, the noise dominates the update of Perturbed GD. Then the squared norm ratio $x_t^2/y_t^2$ gradually increases from $0.004$ to $1$. Perturbed GD iterates towards the balanced optimum while staying close to global optima. The phase transition phenomenon can also been observed for higher dimensional problems. As shown in Figure \ref{fig:addit} for $d=4$, the loss greatly decreases to and stay around $10^{-4}$ in the first $2\times10^3$ iterations, and then the squared norm ratio ${\norm{x}_2^2}/{\norm{y}_2^2}$ gradually increases from $0.5$ to $1$. This implies the transition between Phase II and Phase III, that is Perturbed GD first approaches the set of global minima and then converges to the balanced optimum. \begin{figure}[t]\label{fig:addit} \includegraphics[width=0.48\linewidth]{PGD_loss.pdf} \includegraphics[width=0.465\linewidth]{PGD_ratio.pdf} \caption{Algorithmic behaviors of Perturbed GD and GD for $d=4$. For Perturbed GD, phase transition happens around the first $2\times10^3$ iterations. } \end{figure} \section{Preliminaries} We first introduce some important notions and results which can be used in the following proof. Assuming $\{u_*,\tilde{u}_1,\dots,\tilde{u}_{d_1-1}\}$ and $\{v_*,\tilde{v}_1,\dots,\tilde{v}_{d_2-1}\}$ are two sets of standard orthogonal basis of $\RR^{d_1}$ and $\RR^{d_2}$ respectively, we can then rewrite $\forall x\in\RR^{d_1}$ and $\forall y\in\RR^{d_2}$ as \begin{align*} x\triangleq \alpha_1u_*+\sum_{i=1}^{d _1-1} \beta_{1}^{(i)}\tilde{u}_i,\\ y\triangleq \alpha_2v_*+\sum_{j=1}^{d _2-1} \beta_2^{(j)}\tilde{v}_j, \end{align*} where $\alpha_1=x^\top u_*,$ $\alpha_2=y^\top v_*,$ $\beta_{1}^{(i)}=x^\top \tilde{u}_i$ and $\beta_{2}^{(j)}=y^\top \tilde{v}_j, ~\forall 0\leq i\leq d_1-1, 0\leq j\leq d_2-1.$ For simplicity, we denote $\beta_k=(\beta_k^{(1)},...,\beta_k^{(d_k-1)})^\top$ where $k=1,2.$ With the notions above, we can rewrite the Perturbed GD update as \begin{align*} \alpha_{1,t+1}=\alpha_{1,t}-\eta<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>,\\\alpha_{2,t+1}=\alpha_{2,t}-\eta<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>. \end{align*} Note that the optimal solutions to \eqref{mat_fct_ncvx} satisfy $x^\top My=1$. Thus, in our proof, we need to characterize the update of $x^\top My,$ which can be re-expressed as \begin{align} x_{t+1}^\top My _{t+1}&=\alpha_{1,t+1}\alpha_{2,t+1}\nonumber\\&=(\alpha_{1,t}-\eta<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>)(\alpha_{2,t}-\eta<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>)\nonumber\\&=\alpha_{1,t}\alpha_{2,t}-\eta\left(\alpha_{2,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>+\alpha_{1,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>\right)\nonumber\\&\quad\quad\quad\quad\quad\quad\quad\quad+\eta^2<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*><\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>. \end{align} For simplicity, we denote \begin{align*} &A_t\triangleq \alpha_{2,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>+\alpha_{1,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>,\\&B_t\triangleq <\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*><\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>. \end{align*} Then the update of $x^\top My=\alpha_1 \alpha_2$ can be expressed in a more compact way as follows. \begin{align}\label{*1} \alpha_{1,t+1}\alpha_{2,t+1}=\alpha_{1,t}\alpha_{2,t}-\eta A_t+\eta^2B_t. \end{align} Similarly, the update of $(x^\top My-1)^2$ can be re-expressed as \begin{align}\label{*2} (x_{t+1}^\top My_{t+1}-1)^2&=(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2 \nonumber\\&=(\alpha_{1,t}\alpha_{2,t}-1-\eta A_t+\eta^2B_t)^2 \nonumber\\&=(\alpha_{1,t}\alpha_{2,t}-1)^2+\eta^2 A_t^2+\eta^4 B_t^2\nonumber\\&~~~~~~~~~~~~-2\eta A_t(\alpha_{1,t}\alpha_{2,t}-1)-2\eta^3A_tB_t+2\eta^2B_t(\alpha_{1,t}\alpha_{2,t}-1). \end{align} Furthermore, since the balanced optima satisfy $x^\top u_*=y^\top v_*,$ we further explicitly write down the update of $\left((x^\top u_*)^2-(y^\top v_*)^2\right)^2$ as follows. \begin{align}\label{*3} \left((x_{t+1}^\top u_*)^2-(y_{t+1}^\top v_*)^2\right)^2&=(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2\nonumber\\ &=\left(\alpha_{1,t}^2-\alpha_{2,t}^2\right)^2+4\eta^2 D_t^2+\eta^4 F_t^2\nonumber\\&~~~~~~~~~~~~-4\eta D_t\left(\alpha_{1,t}^2-\alpha_{2,t}^2\right)-4\eta^3D_t F_t+2\eta^2F_t\left(\alpha_{1,t}^2-\alpha_{2,t}^2\right), \end{align} where $D_t$ and $F_t$ is defined as \begin{align*} &D_t\triangleq \alpha_{1,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>-\alpha_{2,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>,\\&F_t\triangleq <\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>^2-<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>^2. \end{align*} Next we are going to calculate $<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>$, the gradient projection along the direction of the optimum $u_*$. \begin{align}\label{*4} &<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>\nonumber\\&=u_*^\top\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}) \nonumber\\&=u_*^\top\left((x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-u_*v_*^\top\right)(y_{t}+\xi_{2,t}) \nonumber\\&=\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+2\alpha_{1,t}y_{t}^\top \xi_{2,t}-v_*^\top \xi_{2,t}+u_*^\top \xi_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)\nonumber\\&~~~~~~~~~~~~+2u_*^\top\xi_{1,t}y_{t}^\top \xi_{2,t}+\alpha_{1,t}\norm{\xi_{2,t}}^2+u_*^\top \xi_{1,t}\norm{\xi_{2,t}}^2 \nonumber\\&\triangleq \alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x, \end{align} where $g_x=2\alpha_{1,t}y_{t}^\top \xi_{2,t}-v_*^\top \xi_{2,t}+u_*^\top \xi_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)+2u_*^\top\xi_{1,t}y_{t}^\top \xi_{2,t}+\alpha_{1,t}\norm{\xi_{2,t}}^2+u_*^\top \xi_{1,t}\norm{\xi_{2,t}}^2.$ Similarly, we have \begin{align}\label{*5} &<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*> \nonumber\\&=\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+2\alpha_{2,t}x_{t}^\top \xi_{1,t}-u_*^\top \xi_{1,t}+v_*^\top \xi_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)\nonumber\\&~~~~~~~~~~~~+2v_*^\top\xi_{2,t}x_{t}^\top \xi_{1,t}+\alpha_{2,t}\norm{\xi_{1,t}}^2+v_*^\top \xi_{2,t}\norm{\xi_{1,t}}^2 \nonumber\\&\triangleq \alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y, \end{align} where $g_y=\alpha_{1,t}+2\alpha_{2,t}x_{t}^\top \xi_{1,t}-u_*^\top \xi_{1,t}+v_*^\top \xi_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)+2v_*^\top\xi_{2,t}x_{t}^\top \xi_{1,t}+\alpha_{2,t}\norm{\xi_{1,t}}^2+v_*^\top \xi_{2,t}\norm{\xi_{1,t}}^2.$ \section{Proof of Lemma \ref{lem_landscape}, Lemma \ref{lem_condition_number} , and Lemma \ref{lem_convolutional}}\label{pf_1} \subsection{Proof of Lemma \ref{lem_landscape}} \begin{proof} By setting the gradient of $ \cF$ to zero we get \begin{align} &\norm{x}_2^2y=M^\top x,\label{zero_grad_1}\\ &\norm{y}_2^2x=My.\label{zero_grad_2} \end{align} Recall that $M=u_* v_*^\top$ and \begin{align*} x\triangleq \alpha_1u_*+\sum_{i=1}^{d _1-1} \beta_{1}^{(i)}\tilde{u}_i,\\ y\triangleq \alpha_2v_*+\sum_{j=1}^{d _2-1} \beta_2^{(j)}\tilde{v}_j. \end{align*} Substitute $x$ and $y$ in \eqref{zero_grad_1} and \eqref{zero_grad_2} by their expansion, we then have \begin{align*} (\alpha_1^2+\sum_{i=1}^{d_1-1} (\beta_{1}^{(i)})^2)\alpha_2&=\alpha_1,\\ (\alpha_2^2+\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2)\alpha_1&=\alpha_2,\\ (\alpha_1^2+\sum_{i=1}^{d_1-1} (\beta_{1}^{(i)})^2)\beta_{2}^{(j)}&=0, \forall j=1,...,d_2-1,\\ (\alpha_2^2+\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2)\beta_{1}^{(j)}&=0, \forall j=1,...,d_1-1. \end{align*} The above equalities yield the following two types of stationary points. \begin{itemize} \item $(\alpha_1^2+\sum_{i=1}^{d_1-1} (\beta_{1}^{(i)})^2)(\alpha_2^2+\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2)\neq 0,\beta_1=0,\beta_2=0.$ This leads to $\alpha_1\alpha_2=1.$ Thus, $xy^\top=\alpha_1\alpha_2 u_*v_*^\top=M,$ and $\cF(\alpha_1u_*,\alpha_2v_*)=0.$ Then we have global optima $\left(\alpha u_*, \frac{1}{\alpha} v_*\right)$ for $\alpha\neq 0.$ \item Either $(\alpha_1^2+\sum_{i=1}^{d_1-1} (\beta_{1}^{(i)})^2)=0$ and $\alpha_2=0,$ or $(\alpha_2^2+\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2)=0$ and $\alpha_1=0.$ We next show that stationary points satisfy these conditions are strict saddle points. We only consider the first case, and the second case can be proved following similar lines. We first calculate the Hessian matrix as follows. \begin{align*} \nabla^2 \cF(x, y)=\begin{pmatrix} \norm{y}_2^2I_{d_1} & 2xy^\top-M \\ 2yx^\top-M^\top & \norm{x}_2^2I_{d_2} \end{pmatrix}. \end{align*} At $x=0$, $y=\sum_{j=1}^{d_2-1} \beta_2^{(j)}\tilde{v}_j,$ \begin{align*} \nabla^2 \cF(x, y)=\begin{pmatrix} \sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2I_{d_1} & -M \\ -M^\top & 0 \end{pmatrix}. \end{align*} For any $a\in \RR^{d_1},b\in \RR^{d_2},$ \begin{align*} \begin{pmatrix} a^\top, b^\top \end{pmatrix}\nabla^2 \cF(x, y)\begin{pmatrix} a \\ b \end{pmatrix}=\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2\norm{a}_2^2-2a^\top M b \end{align*} For $(a, b)=(\tilde{u}_i, \tilde{v}_j)$, this quantity is positive. For $(a, b)=(u_*, \sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2v_*),$ this quantity is negative. Thus, $x=0$, $y=\sum_{j=1}^{d_2-1} \beta_2^{(j)}\tilde{v}_j$satisfies strict saddle property. We conclude that for any $x\in \RR^{d_1},y\in\RR^{d_2}$ such that $x^\top u_*=y^\top v_*=0,$ we have strict saddle points $(x,0)$ and $(0,y).$ \end{itemize} \end{proof} \subsection{Proof of Lemma \ref{lem_condition_number}} At $x=\alpha u_*, y=\frac{1}{\alpha}v_*,$ \begin{align*} \nabla^2 \cF(\alpha u_*, \frac{1}{\alpha}v_*)=\begin{pmatrix} \frac{1}{\alpha^2}I_{d_1} &M \\ M^\top & \alpha^2I_{d_2} \end{pmatrix}. \end{align*} One can verify that $\nabla^2 \cF\left(\alpha u_*,\frac{1}{\alpha} v_*\right)$ has eigenvalues $\alpha^2+\frac{1}{\alpha^2},\alpha^2, \frac{1}{\alpha^2}.$ The largest eigenvalue is $\lambda_1=\alpha^2+\frac{1}{\alpha^2}$ and the smallest eigenvalue is $\lambda_{d_1+d_2}=\min\{\alpha^2,\frac{1}{\alpha^2}\}.$ Thus, the condition number can be easily calculated as follows. $$\kappa\left(\nabla^2 \cF\left(\alpha u_*,\frac{1}{\alpha} v_*\right)\right)=\max\{\alpha^4,\frac{1}{\alpha^4}\}+1.$$ \subsection{Proof of Lemma \ref{lem_convolutional}} \begin{proof} Recall that $M=u_* v_*^\top$ and \begin{align*} x\triangleq \alpha_1u_*+\sum_{i=1}^{d _1-1} \beta_{1}^{(i)}\tilde{u}_i,\\ y\triangleq \alpha_2v_*+\sum_{j=1}^{d _2-1} \beta_2^{(j)}\tilde{v}_j. \end{align*} By setting the gradient of $\tilde \cF$ to zero we get \begin{align*} &(\norm{x}_2^2+d_1 \sigma_1^2)y=M^\top x,\\ &(\norm{y}_2^2+d_2 \sigma_2^2)x=My. \end{align*} From the equations above, we can verify that $(x,y)=(0,0)$ is a stationary point. Furthermore, left multiplying the equations above by $\tilde{v}_j^\top$ and $\tilde{u}_i^\top$ respectively, we will get that $\beta_{1}^{(i)}$'s and $\beta_{2}^{(j)}$'s are all zeros. Similarly, by left multiplying the equations above by $v_*^\top$ and $u_*^\top$ respectively, we will get\begin{align*} &(\alpha_1^2+d_1 \sigma_1^2)\alpha_2^2=\alpha_2\alpha_1,\\ &(\alpha_2^2+d_2 \sigma_2^2)\alpha_1^2=\alpha_1\alpha_2. \end{align*} Then, with some algebraic manipulations, we get \begin{align*} &\alpha_1^2=\sqrt{\frac{d_1 \sigma_1^2}{d_2 \sigma_2^2}} - d_1 \sigma_1^2,\\ &\alpha_2^2=\sqrt{\frac{d_2 \sigma_2^2}{d_1 \sigma_1^2}} - d_2 \sigma_2^2. \end{align*} Specifically, when $d_1 \sigma_1^2=\gamma^2 d_2 \sigma_2^2\leq\gamma$, we have \begin{align*} &\alpha_1=\gamma \alpha_2= \pm \sqrt{\gamma- \gamma^2 d_2 \sigma_2^2}. \end{align*} Next, we are going to show that $(0, 0)$ is a strict saddle point and $(x_*, y_*)\triangleq\pm(\alpha_1u_*, \alpha_2v_*)$ are global optima. We first calculate the Hessian matrix as follows. \begin{align*} \nabla^2 \tilde\cF(x, y)=\begin{pmatrix} \left(\norm{y}_2^2+d_2\sigma_2^2\right)I_{d_1} & 2xy^\top-M \\ 2yx^\top-M^\top & \left(\norm{x}_2^2+d_1\sigma_1^2\right)I_{d_2} \end{pmatrix}. \end{align*} Since the injected noise is small, we have $\alpha_1\alpha_2=1-\sqrt{d_1\sigma_1^2d_2\sigma_2^2}>0$. For any $a \in \RR^{d_1}$ and $b \in \RR^{d_2}$, we have\begin{align*} &\begin{pmatrix} a^\top, b^\top \end{pmatrix}\nabla^2 \tilde\cF(x_*, y_*)\begin{pmatrix} a \\ b \end{pmatrix}\\ =&(\alpha_2^2+d_2\sigma_2^2)\norm{a}_2^2+(\alpha_1^2+d_1\sigma_1^2)\norm{b}_2^2+2(2\alpha_1\alpha_2-1)(a^\top u_*) (b^\top v_*)\\ \geq&(\alpha_2^2+d_2\sigma_2^2)\norm{a}_2^2+(\alpha_1^2+d_1\sigma_1^2)\norm{b}_2^2-2|2(1-\sqrt{d_1\sigma_1^2d_2\sigma_2^2})-1|\ \norm{a}_2\ \norm{b}_2\\ >&\sqrt{\frac{d_2 \sigma_2^2}{d_1 \sigma_1^2}}\norm{a}_2^2+\sqrt{\frac{d_1 \sigma_1^2}{d_2 \sigma_2^2}}\norm{b}_2^2-2\norm{a}_2\ \norm{b}_2, \end{align*}where the last inequality comes from the fact that $d_1\sigma_1^2$ and $d_2\sigma_2^2$ should be small enough such that $$0<1-\sqrt{d_1\sigma_1^2d_2\sigma_2^2}<1.$$Note that $$\sqrt{\frac{d_2 \sigma_2^2}{d_1 \sigma_1^2}}\norm{a}_2^2+\sqrt{\frac{d_1 \sigma_1^2}{d_2 \sigma_2^2}}\norm{b}_2^2-2\norm{a}_2\ \norm{b}_2=\left((\frac{d_2 \sigma_2^2}{d_1 \sigma_1^2})^{\frac{1}{4}}\norm{a}_2-(\frac{d_1 \sigma_1^2}{d_2 \sigma_2^2})^{\frac{1}{4}}\norm{b}_2\right)^2.$$ Thus, as long as ${d_1\sigma_1^2\ d_2\sigma_2^2}<1$, we have $\nabla^2 \cF_{reg}(x_*, y_*)$ is positive definite (PD). Thus $(x_*, y_*)$ is global minimum and so is $(-x_*,- y_*)$. Similarly, for any $a \in \RR^{d_1}$ and $b \in \RR^{d_2}$, we have \begin{align*} &\begin{pmatrix} a^\top, b^\top \end{pmatrix}\nabla^2 \cF_{reg}(0, 0)\begin{pmatrix} a \\ b \end{pmatrix} =d_2\sigma_2^2\norm{a}_2^2+d_1\sigma_1^2\norm{b}_2^2-2(a^\top u_*) (b^\top v_*). \end{align*} It is easy to show that for $(a, b)=(u_*, v_*)$, the quantity is negative when noise is small enough. But for $(a, b)=(\tilde{u}_i, \tilde{v}_j)$, this quantity is positive. Thus, (0,0) is a strict saddle point. Here, we prove Lemma \ref{lem_convolutional}. \end{proof} \section{Proof of Theorem \ref{thm_main}}\label{pf_2} \subsection{Boundedness of Trajectory} We first show that the solution trajectory of Perturbed GD is bounded with high probability, which is a sufficient condition for our following convergence analysis. \begin{lemma}[Boundedness of Trajectories]\label{lem_bounded} Given $x_0\in\RR^{d_1},$ $y_0\in\RR^{d_2},$ we choose $\sigma_1, \sigma_2>0$ such that $\sigma^2=\EE\left[\norm{\xi_1}_2^2\right]=\EE\left[\norm{\xi_2}_2^2\right]$ and $\norm{x_0}^2_2+\norm{y_0}^2_2\leq 1/\sigma^2.$ For any $\delta\in(0,1)$, we take $$\eta\leq \eta_1= C_1{\sigma^6}(\log((d_1+d_2)/\delta)\log(1/\delta))^{-1},$$ for some positive constant $C_1.$ Then with probability at least $1-\delta,$ we have $\norm{x_t}^2_2+\norm{y_t}^2_2\leq 2/\sigma^2$ for any $t\leq T_1=O(1/\eta^2).$ \end{lemma} \begin{proof} We first define the event where the injected noise for both $x$ and $y$ is bounded for the first $t$ iterations. \begin{align} &\cA_t=\left\{\big|{\xi^{(i)}_{1,\tau}}\big|,\big|{\xi^{(j)}_{2,\tau}}\big|\leq \sigma\left(\sqrt{2\log((d_1+d_2)\eta^{-2})} +\sqrt{\log(1/\delta)}\right), \forall\tau\leq t, i=1,...,d_1,j=1,..,d_2 \right\}. \end{align} By the concentration result of the maximum of Gaussian distribution, we have $\PP(\cA_{1/\eta^2})\geq 1-\delta.$ Moreover, we use $\cH_t$ to denote the event where the first $t$ iterates $\{(x_\tau,y_\tau)\}_{\tau\leq t}$ is bounded, i.e., $ \cH_t=\left\{\norm{x_\tau}^2_2+\norm{y_\tau}^2_2\leq \frac{2}{\sigma_2},\forall\tau\leq t\right\}. $ Let $\cF_t=\sigma\{(x_\tau,y_\tau),\tau\leq t\}$ denote the $\sigma-$algebra generated by that past $t$ iterations. Under the event $\cH_t$ and $\cA_t,$ we have the following inequality on the conditional expectation of $\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2.$ \begin{align} &\EE\left[(\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2)\mathds{1}_{\cH_t\cap\cA_t}\big|\cF_t\right]\nonumber\\=&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)-4\eta\left(\norm{x_{t}}_2^2\norm{y_{t}}_2^2-x_t^\top M y_t\right)\right\}\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\norm{\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t})}_2^2+\norm{\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t})}_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ \leq&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)-4\eta\left(\norm{x_{t}}_2^2\norm{y_{t}}_2^2-\norm{x_{t}}_2\norm{y_{t}}_2\right)\right\}\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[ \big|\big|\norm{y_t+\xi_{2,t}}_2^2(x_t+\xi_{1,t})-M(y_t+\xi_{2,t})\big|\big|_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[ \big|\big|\norm{x_t+\xi_{1,t}}_2^2(y_t+\xi_{2,t})-M^\top(x_t+\xi_{1,t})\big|\big|_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ \leq&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)-4\eta\left(\norm{x_{t}}_2^2\norm{y_{t}}_2^2-\norm{x_{t}}_2\norm{y_{t}}_2\right)\right\}\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+2\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[ \norm{y_t+\xi_{2,t}}_2^4\norm{x_t+\xi_{1,t}}_2^2+\norm{M(y_t+\xi_{2,t})}_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+2\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[ \norm{x_t+\xi_{1,t}}_2^4\norm{y_t+\xi_{2,t}}_2^2+\norm{M^\top(x_t+\xi_{1,t})}_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ \leq&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)+\eta+\eta^2 C_2\right\}\mathds{1}_{\cH_t\cap\cA_t},\nonumber \end{align} where the second inequality comes from the fact that $x^2-x\geq-\frac{1}{4}$ for all $x\in\RR$ and $C_2=(\sigma^2+1/\sigma^2)(2/\sigma^4+6/d_1+6/d_2+6\sigma^4)$ in the last inequality. We take $\eta\leq 1/C_2,$ then we have \begin{align} \EE\left[(\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2-1/\sigma^2)\mathds{1}_{\cH_t\cap\cA_t}\big|\cF_t\right]&\leq(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2-1/\sigma^2\right)\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &\leq(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2-1/\sigma^2\right)\mathds{1}_{\cH_{t-1}\cap\cA_{t-1}}\label{martingale1}. \end{align} If we denote $G_t=(1-2\eta\sigma^2)^{-t}(\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2-1/\sigma^2),$ $G_t\mathds{1}_{\cH_{t-1}\cap\cA_{t-1}}$ is then a super-martingale according to \eqref{martingale1}. We will apply Azuma's Inequality to prove the bound and before that we have to bound the difference between $G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}$ and $\EE[G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}\big|\cF_t].$ \begin{align*} d_{t+1}&=\big| G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}- \EE[G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}\big|\cF_t]\big|\\ &=(1-2\eta\sigma^2)^{-t-1}\Bigg|2\eta x_t^\top[(x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-M](y_{t}+\xi_{2,t}))\\&\hspace{+1.2in}-2\eta \EE_{\xi_{1,t},\xi_{2,t}}\left[x_t^\top[(x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-M](y_{t}+\xi_{2,t}))\right]\\ &\hspace{+1.2in}+2\eta y_t^\top[(x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-M]^\top(x_{t}+\xi_{1,t}))\\&\hspace{+1.2in}-2\eta \EE_{\xi_{1,t},\xi_{2,t}}\left[y_t^\top[(x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-M]^\top(x_{t}+\xi_{1,t}))\right]\\ &\hspace{+1.2in}+\eta^2\big|\big|\norm{y_t+\xi_{2,t}}_2^2(x_t+\xi_{1,t})-M(y_t+\xi_{2,t})\big|\big|_2^2\\ &\hspace{+1.2in}-\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\big|\big|\norm{y_t+\xi_{2,t}}_2^2(x_t+\xi_{1,t})-M(y_t+\xi_{2,t})\big|\big|_2^2\right]\\ &\hspace{+1.2in}+\eta^2\big|\big|\norm{x_t+\xi_{1,t}}_2^2(y_t+\xi_{2,t})-M^\top(x_t+\xi_{1,t})\big|\big|_2^2\\ &\hspace{+1.2in}-\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\big|\big|\norm{x_t+\xi_{1,t}}_2^2(y_t+\xi_{2,t})-M^\top(x_t+\xi_{1,t})\big|\big|_2^2\right]\Bigg|\mathds{1}_{\cH_{t}\cap\cA_{t}} \\&\leq {C_1}' (1-2\eta\sigma^2)^{-t-1}\eta \sigma^{-2}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right), \end{align*} where ${C_1}'$ is some positive constant. Denote $r_t=\sqrt{\sum_{i=1}^t d_i^2}.$ By Azuma's inequality, we have $$\PP\left(G_t\mathds{1}_{\cH_{t-1}\cap\cA_{t-1}}-G_0\geq O(1)r_t\left(\log\frac{1}{\eta^2\delta}\right)^{\frac{1}{2}}\right)\leq O(\eta^2\delta).$$ Then when with probability at least $1-O(\eta^2\delta),$ we have \begin{align*} (\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2)\mathds{1}_{\cH_{t}\cap\cA_{t}}&\leq 1/\sigma^2+(1-\eta \sigma^2)^t(\norm{x_0}^2_2+\norm{y_0}^2_2-1/\sigma^2)\\ &\hspace{+0.3in}+O(1)(1-\eta\sigma^2)^t r_t(\log\frac{1}{\eta^2\delta})^{\frac{1}{2}}\\ &\leq 1/\sigma^2+ 0+O\left(\sqrt{\eta} \sigma^{-2}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\left(\log\frac{1}{\eta^2\delta}\right)^{\frac{1}{2}}\\ &\leq 2/\sigma^2, \end{align*} when $\eta=O\left(\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right).$ In order to satisfy $\eta\leq1/C_2$ at the same time, we take $\eta\leq\eta_1=O\left({\sigma^6}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right)$ to make sure that all inequalities above hold. The above inequality shows that if $\cH_{t-1}\cap\cA_{t-1}$ holds, then $\cH_{t}\cap\cA_{t}$ holds with probability at least $1-O(\eta^2\delta).$ Hence with probability at least $1-\delta,$ we have $(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2)\mathds{1}_{\cA_{t-1}}\leq 2/\sigma^2$ for all $t\leq T_1=O(1/\eta^2).$ Recall that $$\PP(\cA_{1/\eta^2})\geq 1-\delta.$$ Thus, we have with probability at least $1-2\delta,$ $\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\leq 2/\sigma^2$ for all $t\leq T_1=O(1/\eta^2).$ By properly rescaling $\delta$, we prove Lemma \ref{lem_bounded}. \end{proof} \subsection{Proof of Lemma \ref{lem_b_converge}} \begin{proof} We only prove the convergence of $\norm{\beta_{1,t}}_2^2$ here. The proof of the convergence of $\norm{\beta_{2,t}}_2^2$ follows similar lines. For notational simplicity, we denote $\xi_{1,t}^{(-1)}=(\xi_{1,t}^{(2)},..., \xi_{1,t}^{(d_1)})^\top,\ \forall t\geq 0.$ We first bound the conditional expectation of $\norm{\beta_{1,t+1}}_2^2$ given $\cF_t$: \begin{align} \EE\left[\norm{\beta_{1,t+1}}_2^2\big|\cF_t\right]&=\norm{\beta_{1,t}}_2^2+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\norm{y_t+\xi_{2,t}}_2^4\norm{\beta_{1,t}+\xi_{1,t}^{(-1)}}_2^2\right]-2\eta\left(\norm{y_t}_2^2+\sigma^2\right)\norm{\beta_{1,t}}_2^2\nonumber\\ &\leq (1-2\eta\sigma^2)\norm{\beta_{1,t}}_2^2+\eta^2 C_2, \label{b_martingale} \end{align} where $C_2=(\sigma^2+1/\sigma^2)(2/\sigma^4+6/d_1+6/d_2+6\sigma^4).$ The bound on the second moment comes from the boundedness of $\norm{y_t}_2^2,$ which has been shown in Lemma \ref{lem_bounded}. Define $G_t=(1-2\eta\sigma^2)^{-t}\left(\norm{\beta_{1,t}}_2^2-\frac{\eta C_2}{2\sigma^2}\right),$ and $\cE_t=\left\{\forall \tau\leq t, \norm{\beta_{1,\tau}}_2^2\geq \frac{\eta C_2}{\sigma^2}\right\}.$ By \eqref{b_martingale}, we have $$\EE\left[G_{t+1}\mathds{1}_{\cE_t}\big|\cF_t\right]\leq G_t\mathds{1}_{\cE_t}\leq G_t\mathds{1}_{\cE_{t-1}} .$$ Hence, by Markov inequality we have \begin{align*} \PP(\cE_t)= \PP\left( \norm{\beta_{1,t}}_2^2\mathds{1}_{\cE_{t-1}}\geq \frac{\eta C_2}{\sigma^2}\right)&\leq\frac{\EE\left[ \norm{\beta_{1,t}}_2^2\mathds{1}_{\cE_{t-1}}\right]}{\frac{\eta C_2}{\sigma^2}}\\ &\leq\frac{(1-2\eta\sigma^2)^{t}(\norm{\beta_{1,0}}_2^2-\frac{\eta C_2}{2\sigma^2})+\frac{\eta C_2}{2\sigma^2}}{\frac{\eta C_2}{\sigma^2}}\\ &\leq (1-2\eta\sigma^2)^{t}\frac{2}{C_2\eta}+\frac{1}{2} \leq \frac{3}{4}, \end{align*} when $t\geq\frac{1}{2\eta\sigma^2}\log\frac{8}{C_2\eta}.$ We take $t=\frac{1}{\eta\sigma^2}\log\frac{8}{C_2\eta}$ to make sure the inequality above holds. Thus with probability at least $\frac{1}{4},$ there exists a $ \tau\leq\frac{1}{\eta\sigma^2}\log\frac{8}{C_2\eta}$, such that $ \norm{\beta_{1,\tau}}_2^2\leq \frac{\eta C_2}{\sigma^2}.$ Thus, with probability at least $1-\delta,$ we can find a $\tau$ such that $ \norm{\beta_{1,\tau}}_2^2\leq \frac{\eta C_2}{\sigma^2},$ where $$\tau\leq\tau_1=\frac{1}{\log (4/3)\eta\sigma^2}\log\frac{8}{C_2\eta}\log\frac{1}{\delta}=\frac{1}{\log (4/3)\eta\sigma^2}\left(\log\frac{8}{C_2}+\log\frac{1}{\eta}\right)\log\frac{1}{\delta}=O\left(\frac{1}{\eta\sigma^2}\log\frac{1}{\eta}\log\frac{1}{\delta}\right).$$ We next show that with probability at least $1-\delta$, for $\forall t\geq \tau_1,$ we have $ \norm{\beta_{1,t}}_2^2\leq 2\frac{\eta C_2}{\sigma^2}.$ This can be done following the similar lines to the proof of Lemma \ref{lem_bounded}. We first restart the counter of the time and assume $ \norm{\beta_{1,0}}_2^2\leq \frac{\eta C_2}{\sigma^2}.$ Denote $\cH_t=\left\{\forall \tau\leq t,\norm{\beta_{1,\tau}}_2^2\leq \frac{2\eta C_2}{\sigma^2}\right\}.$ Then by \eqref{b_martingale}, we have $$\EE\left[G_{t+1}\mathds{1}_{\cH_t\cap\cA_t}\big|\cF_t\right]\leq G_t\mathds{1}_{\cH_t\cap\cA_t}\leq G_t\mathds{1}_{\cH_{t-1}\cap\cA_{t-1}} .$$ The difference of $ G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}$ and $\EE[G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}\big|\cF_t]$ can be easily bounded as follows \begin{align*} D_{t+1}&=\big| G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}- \EE[G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}\big|\cF_t]\big|={C_2}'\eta^2\sigma^{-3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right), \end{align*} where ${C_2}'$ is some constant. Then by applying Azuma's inequality and following the similar lines to the proof of Lemma \ref{lem_bounded}, we show that when $$\eta\leq {C_3}'{\sigma^8}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1},$$ with probability at least $1-\delta,$ we have \begin{align*} \norm{\beta_{1,t}}_2^2\leq 2\frac{\eta C_2}{\sigma^2}.\end{align*} for any $\tau_1\leq t\leq T_1= O(1/\eta^2).$ \end{proof} \subsection{Proof of Lemma \ref{lem_escape}} We prove this lemma in three steps. \noindent $\bullet$ {\bf Step 1:} The following lemma shows that after polynomial time, with high probability, the algorithm can move out of the $O(\eta)$ neighborhood of the saddle point. \begin{lemma}[Escaping from the Unique Saddle Point] Suppose $\norm{\beta_{1,t}}_2^2\leq 2\frac{\eta C_2}{\sigma^2}$ and $\norm{\beta_{2,t}}_2^2\leq 2\frac{\eta C_2}{\sigma^2}$ hold for all $t>0.$ For $\forall\delta\in(0,1),$ we take $$\eta=O\left(\sigma^{12}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right).$$ Then with probability at least $1-\delta,$ there exists a $\tau\leq{\tau}_2,$ such that $$x_{\tau}^\top M y_{\tau}\geq 9\frac{\eta C_2}{\sigma^4},$$ where ${\tau}_2=\frac{5}{\eta \sigma^2}\log \frac{4 \sigma^3}{\eta C_2}\log\frac{1}{\delta}.$ \end{lemma} \begin{proof} Let's define the event $\cH_t=\{x_{\tau}^\top M y_{\tau}\leq9\frac{\eta C_2}{\sigma^4}, ~\forall \tau\leq t \}.$ Following the proof of Lemma \ref{lem_bounded}, we can refine the bound on the conditional expectation of $\norm{x_{t+1}}^2_2+\norm{y_{t+1}}^2_2$ given $\cH_t.$ Specifically, we have \begin{align*} \EE\left[(\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2)\mathds{1}_{\cH_t}\big|\cF_t\right]=&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)-4\eta\left(\norm{x_{t}}_2^2\norm{y_{t}}_2^2-x_t^\top M y_t\right)\right\}\mathds{1}_{\cH_t}\nonumber\\ &+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\norm{\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t})}_2^2+\norm{\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t})}_2^2\right]\mathds{1}_{\cH_t}\nonumber\\ \leq&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)+34C_2\eta^2 \sigma^{-4})\right\}\mathds{1}_{\cH_t}. \end{align*} Denote $G_t=(1-2\eta\sigma^2)^{-t}\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2-17C_2\eta \sigma^{-6}\right).$ Then we have $$\EE[G_{t+1}\mathds{1}_{\cH_t}]\leq G_{t}\mathds{1}_{\cH_t}\leq G_{t}\mathds{1}_{\cH_{t-1}}.$$ Thus, by Markov inequality, we have \begin{align*} \PP( (\norm{x_{t}}_2^2+\norm{y_{t}}_2^2)\mathds{1}_{\cH_{t-1}}\geq 34C_2\eta \sigma^{-6} )&\leq \frac{(1-2\eta\sigma^2)^{t}(\norm{x_{0}}_2^2+\norm{y_{0}}_2^2-17C_2\eta \sigma^{-6})+17C_2\eta \sigma^{-6}}{34C_2\eta \sigma^{-6}}\\ &\leq \frac{3}{4}, \end{align*} when $t\geq \frac{1}{2\eta \sigma^2}\log \frac{4 \sigma^3}{\eta C_2}.$ Thus with probability at least $1-\delta,$ there exists a $\tau\leq{\tau}_2=\frac{5}{2\eta \sigma^2}\log \frac{4 \sigma^3}{\eta C_2}\log\frac{1}{\delta}=O(\frac{1}{\eta \sigma^2}\log \frac{1}{\eta }\log\frac{1}{\delta}),$ such that $$(\norm{x_{\tau}}_2^2+\norm{y_{{\tau}}}_2^2)\mathds{1}_{\cH_{\tau-1}}\leq 34C_2\eta \sigma^{-6}.$$ Following the exactly same proof of Lemma \ref{lem_bounded}, we can show that with probability at least $1-\delta,$ for all $\tau_2\leq t\leq T_1=O(\frac{1}{\eta^2}),$ we have \begin{align*} (\norm{x_{t}}_2^2+\norm{y_{t}}_2^2)\mathds{1}_{\cH_{t-1}}&\leq(17+34)C_2\eta \sigma^{-6}+O\left(\sqrt{\eta} \sigma^{-2}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\left(\log\frac{1}{\eta^2\delta}\right)^{\frac{1}{2}} \\&= {C_4}'\eta\sigma^{-12}+{C_5}'\sqrt{\eta} \sigma^{-2}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\left(\log\frac{1}{\eta^2\delta}\right)^{\frac{1}{2}}\\ &\leq {C_6}' \left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}+{C_7}' \sigma^{4}, \end{align*} where ${C_4}'$, ${C_5}'$, ${C_6}'$ and ${C_7}'$ are some positive constants, and when$$\eta= O\left({\sigma^{12}}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right).$$ We next show that for large enough $t,$ with high probability, $\cH_{t-1}$ does not hold. We prove by contradiction: if $\cH_{t-1}$ holds for all $t$, the solution trajectory stays in a small neighborhood around $0.$ If so, we can show that with constant probability, $|\alpha_{1,t}+\alpha_{2,t}|$ will explode to infinity, which is in contradiction with the boundedness. Here follows the detailed proof. Assuming $\cH_{t-1}$ holds for all $t\leq\tilde{\tau}_2=O(\frac{1}{\eta}\log\frac{1}{\eta\sigma}\log\frac{1}{\delta})$, by the analysis above we have $\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\leq {C_6}' \left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}+{C_7}' \sigma^{4}$ holds for $\forall t\leq\tilde{\tau}_2$. Note that with at least some constant probability, \begin{align*}&\Big|\alpha_{1,t+1}+\alpha_{2,t+1}\Big|-\Big|\alpha_{1,t}+\alpha_{2,t}\Big|\\=&\Big|\alpha_{1,t}+\alpha_{2,t}-\eta\left(<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>+<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>\right)\Big|-\Big|\alpha_{1,t}+\alpha_{2,t}\Big| \\ \geq &{C_8}'\eta(\sigma_1+\sigma_2),\end{align*} where ${C_8}'$ is some positive constant. This means we can find a $\tau=O(\log\frac{1}{\delta})\leq\tilde{\tau}_2$, such that $\Big|\alpha_{1,\tau}+\alpha_{2,\tau}\Big|>{C_8}'\eta(\sigma_1+\sigma_2),$ with probability at least $1-\delta.$ We will use this point as our initialization for the following proof. We next give a lower bound on the conditional expectation of $|\alpha_{1,t+1}+\alpha_{2,t+1}|.$ \begin{align*} &\EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|\big|\cF_t\right]\\ \geq&\Big|\EE\left[(\alpha_{1,t+1}+\alpha_{2,t+1})\big|\cF_t\right]\Big|\\ =&\Big|\left(1+\eta(1-\sigma^2-\alpha_{1,t}\alpha_{2,t})\right)(\alpha_{1,t}+\alpha_{2,t})-\eta\left(\norm{\beta_{1,t}}_2^2\alpha_{2,t}+\norm{\beta_{2,t}}_2^2\alpha_{1,t}\right)\Big|\\ \geq& \left(1+\eta(1-\sigma^2-\alpha_{1,t}\alpha_{2,t})\right)\Big|\alpha_{1,t}+\alpha_{2,t}\Big|- 2{C_9}' \frac{\eta^{2.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\\ \geq&\left(1+\frac{1}{2}\eta\right)\Big|\alpha_{1,t}+\alpha_{2,t}\Big|- 2{C_9}'\frac{\eta^{2.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right), \end{align*} where ${C_9}'$ is some positive constant. This is equivalent to the following inequality: \begin{align*} &\EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|-4{C_9}'\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\Bigg|\cF_t\right]\\\geq &\left(1+\frac{1}{2}\eta\right)\left(\Big|\alpha_{1,t}+\alpha_{2,t}\Big|- 4{C_9}'\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\\ \geq &\left(1+\frac{1}{2}\eta\right)^{t+1}\left(\Big|\alpha_{1,0}+\alpha_{2,0}\Big|- 4{C_9}'\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\\ \geq &\left(1+\frac{1}{2}\eta\right)^{t+1}\left({C_8}'\eta(d_1+d_2)- 4{C_9}'\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{\sigma_1+\sigma_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\\ \geq &\ {C_{10}}' \eta(\sigma_1+\sigma_2)\exp{t\eta/2}\\ \geq & \ {C_{10}}' \eta(\sigma/\sqrt{d_1}+\sigma/\sqrt{d_2})\frac{1}{\eta\sigma} \\ = &\ {C_{10}}'(1/\sqrt{d_1}+1/\sqrt{d_2}) , \end{align*} where ${C_{10}}' $ is some positive constant and last inequality holds when $t\geq \frac{2}{\eta}\log\frac{1}{\eta\sigma}.$ Note that here we still take $\eta=O\left(\sigma^{12}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right)$, which makes sure that $\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)$ is small enough to make the fourth inequality hold. For fixed $d_1$ and $d_2$, if we let $\delta$ and $\sigma$ go to zero (which guarantees that $\eta$ goes to zero), we will have the conditional expectation of $|\alpha_{1,t+1}+\alpha_{2,t+1}|$ stays in a neighborhood around a positive constant. However, by our assumption, $\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2$ stays in a very small neighborhood around zero, which makes $\alpha_{1,t+1}^2+\alpha_{2,t+1}^2\leq\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2$ also very small. This implies that $|\alpha_{1,t+1}+\alpha_{2,t+1}|$ can be arbitrarily small as long as we make $\delta$ and $\sigma$ small enough, which is the desired contradiction. Thus, we know that, with probability at least $1-\delta,$ there exists a t satisfying ${\tau}_2\leq t\leq {\tau}_2+\tilde{\tau}_2\leq 2{\tau}_2,$ such that $\cH_{t}$ does not hold. Re-scale ${\tau}_2$ and we prove the result. \end{proof} \noindent $\bullet$ {\bf Step 2:} The following lemma shows that after step 1, with high probability, the algorithm will continue move away from the saddle point and escape from the saddle point at some time. \begin{lemma} \label{lem_conv_region} Suppose $x_0^\top M y_0\geq 9\frac{\eta C_2}{\sigma^4}.$ We take $\eta$ as in Lemma \ref{lem_escape}. Then there almost surely exists a $\tau\leq \tau_2'=\frac{2}{\eta}\log\frac{2\sigma^3}{\eta C_2},$ such that $$x_{\tau}^\top M y_{\tau}\geq\frac{1}{2}+\sigma^2.$$ \end{lemma} \begin{proof} Suppose $x_t^\top M y_t\leq\frac{1}{2}+\sigma^2$ holds for all $t\leq \tau_2'.$ We next give a lower bound on the conditional expectation of $|\alpha_{1,t+1}+\alpha_{2,t+1}|.$ \begin{align*} \EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|\big|\cF_t\right]&\geq\Big|\EE\left[(\alpha_{1,t+1}+\alpha_{2,t+1})\big|\cF_t\right]\Big|\\ &=\Big|\left(1+\eta(1-\sigma^2-\alpha_{1,t}\alpha_{2,t})\right)(\alpha_{1,t}+\alpha_{2,t})-\eta\left(\norm{\beta_{1,t}}_2^2\alpha_{2,t}+\norm{\beta_{2,t}}_2^2\alpha_{1,t}\right)\Big|\\ &\geq\left(1+\frac{1}{2}\eta\right)\Big|\alpha_{1,t}+\alpha_{2,t}\Big|- 4\frac{\eta^2 C_2}{\sigma^4}.\end{align*} This implies that \begin{align*} \EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|-8\frac{\eta C_2}{\sigma^4}\big|\cF_t\right]&\geq\left(1+\frac{1}{2}\eta\right)\left(\Big|\alpha_{1,t}+\alpha_{2,t}\Big|-8\frac{\eta C_2}{\sigma^4}\right).\end{align*} The above inequality further implies a lower bound on the conditional expectation of $|\alpha_{1,t+1}+\alpha_{2,t+1}|-8\frac{\eta C_2}{\sigma^4}.$ \begin{align*} \EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|-8\frac{\eta C_2}{\sigma^4}\right]&\geq\left(1+\frac{1}{2}\eta\right)^{t}\left(\Big|\alpha_{1,0}+\alpha_{2,0}\Big|-8\frac{\eta C_2}{\sigma^4}\right)\\ &\geq\left(1+\frac{1}{2}\eta\right)^{t} \frac{\eta C_2}{\sigma^4}\geq \frac{2}{\sigma},\end{align*} when $t= \tau_2'=\frac{2}{\eta}\log\frac{2\sigma^3}{\eta C_2}.$ On the other hand, $$\EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|-8\frac{\eta C_2}{\sigma^4}\right]\leq \frac{\sqrt{2}}{\sigma},$$ which leads to a contradiction unless $$\PP(x_t^\top M y_t\leq\frac{1}{2}+\sigma^2,~\forall t\leq \tau_2')=0.$$ We prove the result. \end{proof} \noindent $\bullet$ {\bf Step 3:} We then show that the algorithm will never iterate back towards the saddle point after escaping from it. Then Lemma \ref{lem_escape} is proved. \begin{lemma}\label{prop1} Suppose there exists a time step $\tau$ such that for some positive constant $c<\frac{1}{2}$ \begin{align*} x_{\tau}^\top My_{\tau}>2c, \end{align*} then $\forall \delta \in (0,1)$, with probability at least $1-\delta$, $\forall \tau\leq t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| x^\top My>c\right\}. \end{align*} \end{lemma} By taking $c=\frac{1}{4}$, we can prove Lemma \ref{lem_escape} from Lemma \ref{prop1}. \begin{proof} We consider two cases: \\ (i) if $\alpha_{1,t}\alpha_{2,t}\geq2(1-c)$, then $\alpha_{1,t+1}\alpha_{2,t+1}>2c$ w.p.1.\\ (ii) if $\alpha_{1,t}\alpha_{2,t}<2(1-c)$, then, $\forall \delta \in (0,1)$ and $\forall t \leq T_1=O(\frac{1}{\eta^2})$, w.p. at least $1-\delta$, \begin{align*}(\alpha_{1,t},\alpha_{2,t})\in\{(\alpha_1,\alpha_2)|\left(\alpha_1\alpha_2-1\right)^2<(1-c)^2\}. \end{align*} Note that $(\alpha_{1,t}\alpha_{2,t}-1)^2<(1-c)^2$ implies that $\alpha_{1,t}\alpha_{2,t}>c$, thus we can prove Lemma \ref{prop1}. Further note that the injected noise is upper bounded with high probability, we assume in the following $T_1=O(\frac{1}{\eta^2})$ steps, \begin{align*} \norm{\xi_{k,t}}_\infty\leq\Bar{\sigma}\triangleq\sigma\left(\sqrt{2\log((d_1+d_2)\eta^{-2})} +\sqrt{\log(1/\delta)}\right). \end{align*} {\bf Case (i):} By plugging \eqref{*4} and \eqref{*5} to \eqref{*1}, we get \begin{align*} \alpha_{1,t+1}\alpha_{2,t+1}=\alpha_{1,t}\alpha_{2,t}-\eta \left(\alpha_{2,t} (\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x)+\alpha_{1,t}(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y)\right) \\+\eta^2(\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x)(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y)\\=2c+(\alpha_{1,t}\alpha_{2,t}-2c)-\eta [\alpha_{2,t} (\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x)+\alpha_{1,t}(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y) \\+\eta(\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x)(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y)], \end{align*} First note that by Lemma \ref{lem_conv_region} we can always find such $0<c<\frac{1}{2}$ in the condition. As $\alpha_{1,t}\alpha_{2,t}-2c>2(1-c)-2c=2(1-2c)>0$, we can prove (i) by choosing $\eta$ small enough. By Lemma \ref{lem_bounded}, $\alpha_{1,t}^2+\alpha_{2,t}^2$ is bounded by $\frac{2}{\sigma^2}$ and we can know $g_x$ and $g_y$ are at most of order $O\left(\frac{1}{\sigma}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)\right)$ from the following upper bound: \begin{align*} |g_x|\leq\Bar{\sigma}(\alpha_{2,t}^2+2\alpha_{2,t}\alpha_{1,t}+1+\frac{2\eta C_2}{\sigma^2})+\Bar{\sigma}^2(2\alpha_{2,t}+\alpha_{1,t})+\Bar{\sigma}^3,\\|g_y|\leq\Bar{\sigma}(\alpha_{1,t}^2+2\alpha_{1,t}\alpha_{2,t}+1+\frac{2\eta C_2}{\sigma^2})+\Bar{\sigma}^2(2\alpha_{1,t}+\alpha_{2,t})+\Bar{\sigma}^3. \end{align*} It is easy to show that for properly selected $$\eta\leq\tilde{\eta}_1=O\left(\left(\frac{1}{\sigma^4}+\frac{1}{\sigma^2}\left(\sqrt{\log(d_1+d_2)}+\sqrt{\log\frac{1}{\delta}}\right)\right)^{-1}\right),$$ the last two terms are greater than zero. Thus, w.p.1, $\alpha_{1,t+1}\alpha_{2,t+1}>2c$.\\ {\bf Case (ii):} Without loss of generality, we place the time origin at t, and thus we have $2c<\alpha_1^0\alpha_2^0<2-2c$. \begin{align*} \EE_\xi[A_t] &=\EE_\xi[\alpha_{2,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>+\alpha_{1,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>]. \end{align*} By plugging \eqref{*4} and \eqref{*5} in the above equation, we have \begin{align*} \EE_\xi[A_t] &=\alpha_{2,t}(\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+\alpha_{1,t}\sigma^2)+\alpha_{1,t}(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+\alpha_{2,t}\sigma^2) \\&=(\alpha_{2,t}\alpha_{1,t}-1)(\alpha_{2,t}^2+\alpha_{1,t}^2)+\alpha_{2,t}\alpha_{1,t}(2\sigma^2+\norm{\beta_{1,t}}_2^2+\norm{\beta_{2,t}}_2^2). \end{align*} Again, by plugging \eqref{*4} and \eqref{*5} to \eqref{*2} and taking expectation conditioning on $\cF_t$, when $\alpha_{1,t}\alpha_{2,t}>c$, we will get \begin{align*} \EE[(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2|\cF_t] \nonumber&=(\alpha_{1,t}\alpha_{2,t}-1)^2-2\eta \EE_\xi[A_t](\alpha_{1,t}\alpha_{2,t}-1)\\&+\eta^2 \EE_\xi[A_t^2]+\eta^4 \EE_\xi[B_t^2]-2\eta^3\EE_\xi[A_tB_t]+2\eta^2\EE_\xi[B_t](\alpha_{1,t}\alpha_{2,t}-1)\\ &=\left(1-2\eta (\alpha_{2,t}^2+\alpha_{1,t}^2)\right)(\alpha_{1,t}\alpha_{2,t}-1)^2\\&~~~~-2\eta \alpha_{2,t}\alpha_{1,t}(\alpha_{1,t}\alpha_{2,t}-1)(2\sigma^2+\norm{\beta_{1,t}}_2^2+\norm{\beta_{2,t}}_2^2)\\&~~~~+\eta^2 \EE_\xi[A_t^2]+\eta^4 \EE_\xi[B_t^2]-2\eta^3\EE_\xi[A_tB_t]+2\eta^2\EE_\xi[B_t](\alpha_{1,t}\alpha_{2,t}-1)\\ &\leq\left(1-4\eta c\right)(\alpha_{1,t}\alpha_{2,t}-1)^2+2\eta \frac{1}{4}(2\sigma^2+\frac{2\eta C_2}{\sigma^2}+\frac{2\eta C_2}{\sigma^2})+\tilde{C}_1\eta^2\\ &\leq\left(1-4\eta c\right)(\alpha_{1,t}\alpha_{2,t}-1)^2+\eta (\sigma^2+\frac{2\eta C_2}{\sigma^2})+\tilde{C}_1\eta^2, \end{align*} where $\tilde{C}_1=O\left(\frac{1}{\sigma^4}\left(\log\frac{d_1+d_2}{\delta}\right)\right),$ if $\alpha_{1,t}\alpha_{2,t}$ is of constant order. Note that here we choose $\eta\leq\tilde{\eta}_1$ as mentioned in (i). Denote $\gamma\triangleq\frac{\eta (\sigma^2+\frac{2\eta C_2}{\sigma^2})+\tilde{C}_1\eta^2}{4\eta c}$, the inequality above can be re-expressed as \begin{align} \label{*6} \EE[\{(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2-\gamma\}|\cF_t]\leq\left(1-4\eta c\right)\{(\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma\}. \end{align} We denote $G_t\triangleq(1-4\eta c)^{-t}\{(\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma\}$ and $\cE_t\triangleq\{\forall \tau\leq t : (\alpha_{1,\tau}\alpha_{2,\tau}-1)^2<(1-c)^2\}.$ Since $\alpha_{1,\tau}\alpha_{2,\tau}>c$ can be inferred from $(\alpha_{1,\tau}\alpha_{2,\tau}-1)^2<(1-c)^2$, we can get \begin{align*} \EE[G_{t+1}\mathds{1}_{\cE_t}|\cF_t]\leq G_{t}\mathds{1}_{\cE_t}\leq G_{t}\mathds{1}_{\cE_{t-1}}. \end{align*} This means $\{G_{t}\mathds{1}_{\cE_{t-1}}\}$ is a supermartingale. To use Azuma's inequality, we need to bound the following difference \begin{align*} d_{t+1}&\triangleq|G_{t+1}\mathds{1}_{\cE_t}-\EE[G_{t+1}\mathds{1}_{\cE_t}|\cF_t]|\\ &=(1-4\eta c)^{-t-1}|(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2-\EE[(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2|\cF_t]|\mathds{1}_{\cE_t}\\ &\leq(1-4\eta c)^{-t-1}\tilde{C}_2, \end{align*} where $\tilde{C}_2=O\left(\eta\frac{1}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)\right)$. Again, here we choose $\eta\leq\tilde{\eta}_1$. We further define $r_t\triangleq\sqrt{\sum_{i=1}^t d_i^2}$, and by Azuma's inequality we have \begin{align*} \mathbb{P}\left(G_{t} \mathds{1}_{\cE_{t-1}}-G_{0} \geq O(1) r_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) \leq \exp \left(-\frac{O(1) r_{t}^{2} \log \left(\frac{1}{\eta^{2} \delta}\right)}{2 \sum_{i=0}^{t} d_{i}^{2}}\right)=O\left(\eta^{2} \delta\right). \end{align*} Thus, with probability at least $1-O(\eta^2 \delta)$, \begin{align*} ((\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma) \mathds{1}_{\cE_{t-1}} &< (1-4\eta c)^t\left((\alpha_{1,0}\alpha_{2,0}-1)^2+O(1) r_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<(\alpha_{1,0}\alpha_{2,0}-1)^2+O\left( (1-4\eta c)^t r_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<(1-c)^2+(1-2c)^2-(1-c)^2\\&~~~~~~~~+O\left( \frac{\sqrt{\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)\log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right), \end{align*} When $\cE_{t-1}$ holds, w.p. at least $1-O(\eta^2 \delta)$, we will have \begin{align*} (\alpha_{1,t}\alpha_{2,t}-1)^2&<(1-c)^2+(1-2c)^2-(1-c)^2\\&~~~~+O\left( \frac{\sqrt{\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)\log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) +\gamma<(1-c)^2, \end{align*} as $(1-2c)^2-(1-c)^2$ is some negative constant, by choosing $\eta\leq\tilde{\eta}_2=O\left(\sigma^4(\log\frac{1}{\delta})^{-1}\left(\log\frac{d_1+d_2}{\delta}\right)^{-1}\right)$ and $\sigma$ small enough, we can make sure the sum of last four terms is negative. Now we know that if $\cE_{t-1}$ holds, $\cE_{t}$ holds w.p. at least $1-O(\eta^2 \delta)$. It is easy to show that the Perturbed GD updates satisfy $(\alpha_{1,t}\alpha_{2,t}-1)^2<(1-c)^2$ in the following $O(\frac{1}{\eta^2})$ steps w.p. at least $1-\delta$. \end{proof} \subsection{Proof of Lemma \ref{lem_loss}} We partition this lemma into two parts: Lemma \ref{lma11} shows that after polynomial time, the algorithm enters $\{ (x,y)\big|(x^\top My-1)^2<4\gamma\},$ where $\gamma$ is a small constant depending on $\sigma$, and Lemma \ref{lma22} shows that the algorithm then stays in $\{ (x,y)\big|(x^\top My-1)^2<6\gamma\}.$ It is easy to prove Lemma \ref{lem_loss} from Lemmas \ref{lma11} and \ref{lma22}. \begin{lemma} \label{lma11} Suppose $\forall t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| x^\top My>c\right\}, \end{align*}then $\forall \delta \in (0,1)$, with probability at least $1-\delta$, there exists a time step $\tau\leq \tau_3\triangleq O(\frac{1}{\eta}\log\frac{1}{\sigma}\log\frac{1}{\delta})$ such that \begin{align*} (x_{\tau}^\top My_{\tau}-1)^2<4\gamma, \end{align*} where {$\gamma=O(\sigma^2)$}. \end{lemma} \begin{lemma} \label{lma22} Suppose there exists a time step $\tau\leq \tau_3\triangleq O(\frac{1}{\eta}\log\frac{1}{\sigma}\log\frac{1}{\delta})$ such that \begin{align*} (x_{\tau}^\top My_{\tau}-1)^2<4\gamma, \end{align*} and $\forall t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| x^\top My>c\right\}, \end{align*}then $\forall \delta \in (0,1)$, with probability at least $1-\delta$,$\forall t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| (x^\top My-1)^2<6\gamma\right\}. \end{align*} \end{lemma} Here follows proof of Lemmas \ref{lma11} and \ref{lma22}. \begin{proof} Define $\cH_t\triangleq\{\forall \tau\leq t : (\alpha_{1,\tau}\alpha_{2,\tau}-1)^2\geq4\gamma\}$, for $t>\frac{\log(\frac{(\alpha_{1,0}\alpha_{2,0}-1)^2}{\gamma})}{4\eta c}$, we have \begin{align*} 4\gamma\EE[\mathds{1}_{\cH_t}]\leq\EE[(\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma]\leq(1-4\eta c)^t\left((\alpha_{1,0}\alpha_{2,0}-1)^2-\gamma\right)+\gamma<2\gamma, \end{align*} where the first inequality comes from the definition of $\cH_t$ and the second one comes from \eqref{*6}. Thus, if we choose $t=O\left(\frac{\log(\frac{1}{\gamma \sigma^2})}{\eta}\right)$ and recursively applying the inequality above $O(\log(\frac{1}{\delta}))$ times, we will get, for $\tau_3=O\left(\frac{1}{\eta}\log(\frac{1}{\delta})\log(\frac{1}{\gamma \sigma^2})\right)=O\left(\frac{1}{\eta}\log\frac{1}{\sigma}\log\frac{1}{\delta}\right)$, \begin{align*} \PP(\cH_{\tau_3})<(\frac{1}{2})^{\log(\frac{1}{\delta})}=\delta. \end{align*} Thus, w.p. at least $1-\delta$, there exists a $\tau\leq \tau_3$ s.t. $(\alpha_{1,\tau}\alpha_{2,\tau}-1)^2<4\gamma$. Here, we finish the proof of Lemma \ref{lma11}. Without loss of generality, we place the time origin at $\tau$, i.e. $(\alpha_{1,0}\alpha_{2,0}-1)^2<4\gamma$. We next prove Lemma \ref{lma22}. Denote $\cA\triangleq\{(\alpha_1,\alpha_2)|(\alpha_1\alpha_2-1)^2<6\gamma\}$ and $\cA_t\triangleq\{\forall \tau\leq t:(\alpha_{1,\tau},\alpha_{2,\tau})\in \cA\}$, again note that $\alpha_{1,t}\alpha_{2,t}>c$ can be inferred from $(\alpha_{1,t}\alpha_{2,t}-1)^2<6\gamma<(1-c)^2$. Thus, without any assumption, we can get \begin{align*} \EE[G_{t+1}\mathds{1}_{\cA_t}|\cF_t]\leq G_{t}\mathds{1}_{\cA_t}\leq G_{t}\mathds{1}_{\cA_{t-1}}. \end{align*} This means $\{G_{t}\mathds{1}_{\cA_{t-1}}\}$ is a supermartingale. To use Azuma's inequality, we need to bound the following difference \begin{align*} \tilde{d}_{t+1}&\triangleq|G_{t+1}\mathds{1}_{\cA_t}-\EE[G_{t+1}\mathds{1}_{\cA_t}|\cF_t]|\\ &=(1-4\eta c)^{-t-1}|(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2-\EE[(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2|\cF_t]|\mathds{1}_{\cA_t}\\ &\leq(1-4\eta c)^{-t-1}\tilde{C}_3, \end{align*} where {$\tilde{C}_3=O(\sqrt{\gamma}\tilde{C}_2)$}. Further define $\tilde{r}_t\triangleq\sqrt{\sum_{i=1}^t \tilde{d_i}^2}$, by Azuma's inequality we have \begin{align*} \mathbb{P}\left(G_{t} \mathds{1}_{\cA_{t-1}}-G_{0} \geq O(1) \tilde{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) \leq \exp \left(-\frac{O(1) \tilde{r}_{t}^{2} \log \left(\frac{1}{\eta^{2} \delta}\right)}{2 \sum_{i=0}^{t} \tilde{d_{i}}^{2}}\right)=O\left(\eta^{2} \delta\right). \end{align*} Thus, with probability at least $1-O(\eta^2 \delta)$, \begin{align*} ((\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma) \mathds{1}_{\cA_{t-1}} &< (1-4\eta c)^t\left((\alpha_{1,0}\alpha_{2,0}-1)^2+O(1) \tilde{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<(\alpha_{1,0}\alpha_{2,0}-1)^2+O\left( (1-4\eta c)^t \tilde{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<4\gamma+O\left( {\frac{\sqrt{\gamma\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)}\log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right), \end{align*} When $\cE_{t-1}$ holds, w.p. at least $1-O(\eta^2 \delta)$ \begin{align*} (\alpha_{1,t}\alpha_{2,t}-1)^2&<5\gamma+O\left( \frac{{\tilde{C}_3}}{\sqrt{\eta}} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) <6\gamma, \end{align*} by choosing {$\eta\leq\tilde{\eta}_3=O\left(\sigma^6(\log\frac{1}{\delta})^{-1}\left(\log\frac{d_1+d_2}{\delta}\right)^{-1}\right)$} we can guarantee the last term is smaller than $\gamma$. Now we know that if $\cA_{t-1}$ holds, $\cA_{t}$ holds w.p. at least $1-O(\eta^2 \delta)$. It is easy to show that the Perturbed GD updates satisfy $(\alpha_{1,t}\alpha_{2,t}-1)^2<6\gamma$ in the following $O(\frac{1}{\eta^2})$ steps w.p. at least $1-\delta$. \end{proof} \subsection{Proof of Lemma \ref{lem_convergence}} \begin{lemma} \label{a1=a2} Suppose $(x_t^\top My_t-1)^2<6\gamma$ holds for all t, where $\gamma$ is as defined above. For any $\delta \in (0,1)$ and any $\Delta>0$, if we choose $\sigma=O\left((\log\frac{1}{\delta})^{-\frac{1}{3}}\right)$ and take step size $$\eta\leq\tilde{\eta}_4=O\left(\sigma^{10}\Delta\right),$$ then with probability at least $1-\delta$, we have \begin{align*} (x_t,y_t) \in \left\{(x,y)| \left((x^\top u_*)^2-(y^\top v_*)^2\right)^2<6\Delta \right\}, \end{align*} for all t's such that $\tau_4\leq t \leq T_1$, where $T_1\triangleq O(\frac{1}{\eta^2})$ and $\tau_4\triangleq O(\frac{1}{\eta\sigma^2}\log\frac{1}{\eta}\log\frac{1}{\delta})$. \end{lemma} Again, we partition this lemma into two parts. It is easy to prove Lemma \ref{a1=a2} from Lemmas \ref{lma3} and \ref{lma4}. \begin{lemma} \label{lma3} Suppose $\forall t\leq T_1= O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| (x^\top My-1)^2<6\gamma\right\}. \end{align*}then $\forall \delta \in (0,1)$, with probability at least $1-\delta$, there exists a time step $\tau\leq \tau_4= O(\frac{1}{\eta\sigma^2}\log\frac{1}{\eta}\log\frac{1}{\delta})$ such that \begin{align*} \left((x_{\tau}^\top u_*)^2-(y_{\tau}^\top v_*)^2\right)^2<4\Delta, \end{align*} where {$\Delta=O(\frac{\eta}{\sigma^{10}})$}. \end{lemma} \begin{proof} \begin{align*} \EE_\xi[D_t]&=\EE_\xi[\alpha_{1,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>-\alpha_{2,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>] \\&=\alpha_{1,t}(\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+\alpha_{1,t}\sigma^2)-\alpha_{2,t}(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+\alpha_{2,t}\sigma^2) \\&=\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)+(\alpha_{1,t}^2\norm{\beta_{2,t}}_2^2-\alpha_{2,t}^2\norm{\beta_{1,t}}_2^2). \end{align*} Thus, we have \begin{align*} \EE_\xi[D_t](\alpha_{1,t}^2-\alpha_{2,t}^2)&=\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)^2+(\alpha_{1,t}^2-\alpha_{2,t}^2)(\alpha_{1,t}^2\norm{\beta_{2,t}}_2^2-\alpha_{2,t}^2\norm{\beta_{1,t}}_2^2) \\&\geq\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-2\alpha_{1,t}^2\alpha_{2,t}^2\frac{2\eta C_2}{\sigma^2} \\&>\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-2(1+\sqrt{6\gamma})^2\frac{2\eta C_2}{\sigma^2} \\&>\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-3\frac{2\eta C_2}{\sigma^2}. \end{align*} The last inequality can be achieved by choosing $\gamma<\frac{(\sqrt{3/2}-1)^2}{6}$. Plugging \eqref{*4} and \eqref{*5} into \eqref{*3}, taking expectation conditioning on previous trajectory $\cF_t$ and plugging the equation above in, we get \begin{align*} \EE[(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2|\cF_t] \nonumber&=(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-4\eta \EE_\xi[D_t](\alpha_{1,t}^2-\alpha_{2,t}^2)\\&+4\eta^2 \EE_\xi[D_t^2]+\eta^4 \EE_\xi[F_t^2]-4\eta^3\EE_\xi[D_t F_t]+2\eta^2\EE_\xi[F_t]\left(\alpha_{1,t}^2-\alpha_{2,t}^2\right)\\ &\leq\left(1-4\eta \sigma^2\right)(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2+24\eta \frac{\eta C_2}{\sigma^2}+\tilde{C}_4\eta^2, \end{align*} where ${\tilde{C}_4=O(\frac{1}{\sigma^4}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)^2)}$. Denote $\Delta\triangleq\frac{24\eta \frac{\eta C_2}{\sigma^2}+\tilde{C}_4\eta^2}{4\eta \sigma^2}$. the inequality above can be re-expressed as \begin{align*} \EE[\{(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2-\Delta\}|\cF_t]\leq\left(1-4\eta \sigma^2\right)\{(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-\Delta\}. \end{align*} Denote $\cB_t\triangleq\{\forall \tau\leq t : ((\alpha_{1,\tau}^2-\alpha_{2,\tau}^2)^2\geq4\Delta\}$, for $t>\frac{\log(\frac{\left(\alpha_{1,0}^2-\alpha_{2,0}^2\right)^2}{\delta})}{4\eta \sigma^2}$, we have \begin{align*} 4\Delta\EE[\mathds{1}_{\cB_t}]\leq\EE[(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-\Delta]\leq(1-4\eta c)^t((\alpha_{1,0}^2-\alpha_{2,0}^2)^2-\Delta)+\Delta<2\Delta, \end{align*} where the first inequality comes from the definition of $\cB_t$ and the second one comes from calculation above. Thus, if we choose $t=O(\frac{\log(\frac{1}{\Delta \sigma^2})}{\eta \sigma^2})$ and recursively applying the inequality above $\log(\frac{1}{\delta})$ times, we will get, for $\tau_4=O(\frac{1}{\eta \sigma^2}\log(\frac{1}{\delta})\log(\frac{1}{\Delta \sigma^2}))=O(\frac{1}{\eta \sigma^2}\log(\frac{1}{\delta})\log(\frac{1}{\eta}))$, \begin{align*} \PP(\cB_{\tau_4})<(\frac{1}{2})^{\log(\frac{1}{\delta})}=\delta. \end{align*} Thus, w.p. at least $1-\delta$, there exists a $\tau\leq \tau_4$ s.t. $(\alpha_{1,\tau}^2-\alpha_{2,\tau}^2)^2<4\Delta$. Here, we finish the proof of Lemma \ref{lma3}. \end{proof} \begin{lemma} \label{lma4} Suppose there exists a time step $\tau\leq \tau_4= O(\frac{1}{\eta\sigma^2}\log\frac{1}{\Delta\sigma^2}\log\frac{1}{\delta})$ such that \begin{align*} \left((x_{\tau}^\top u_*)^2-(y_{\tau}^\top v_*)^2\right)^2<4\Delta, \end{align*}then $\forall \delta \in (0,1)$, with probability at least $1-\delta$, $\forall t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| \left((x^\top u_*)^2-(y^\top v_*)^2\right)^2<6\Delta \right\}. \end{align*} \end{lemma} \begin{proof} Without loss of generality, we place the time origin at $\tau$, i.e. $(\alpha_{1,0}^2-\alpha_{2,0}^2)^2<4\Delta$. Denote $\cD\triangleq\{(\alpha_1,\alpha_2)|(\alpha_1^2-\alpha_2^2)^2<6\Delta\}$ and $\cD_t\triangleq\{\forall \tau\leq t:(\alpha_{1,\tau},\alpha_{2,\tau})\in \cD\}$. Note that $(\alpha_{1,t}\alpha_{2,t}-1)^2<6\gamma$ still holds with high probability. Defining $H_t\triangleq(1-4\eta\sigma^2)^{-t}\{(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-\Delta\}$, we can get \begin{align*} \EE[H_{t+1}\mathds{1}_{\cD_t}|\cF_t]\leq H_{t}\mathds{1}_{\cD_t}\leq H_{t}\mathds{1}_{\cD_{t-1}}. \end{align*} This means $\{H_{t}\mathds{1}_{\cD_{t-1}}\}$ is a supermartingale. To use Azuma's inequality, we need to bound the following difference \begin{align*} \Bar{d}_{t+1}&\triangleq|H_{t+1}\mathds{1}_{\cD_t}-\EE[H_{t+1}\mathds{1}_{\cD_t}|\cF_t]|\\ &=(1-4\eta \sigma^2)^{-t-1}|(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2-\EE[(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2|\cF_t]|\mathds{1}_{\cD_t}\\ &\leq(1-4\eta \sigma^2)^{-t-1}\tilde{C}_5, \end{align*} where ${\tilde{C}_5=O(\eta\frac{\sqrt{\Delta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right))}$. Further define $\Bar{r}_t\triangleq\sqrt{\sum_{i=1}^t \Bar{d_i}^2}$, by Azuma's inequality we will get \begin{align*} \mathbb{P}\left(H_{t} \mathds{1}_{\cD_{t-1}}-H_{0} \geq O(1) \Bar{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) \leq \exp \left(-\frac{O(1) \Bar{r}_{t}^{2} \log \left(\frac{1}{\eta^{2} \delta}\right)}{2 \sum_{i=0}^{t} \Bar{d_{i}}^{2}}\right)=O\left(\eta^{2} \delta\right). \end{align*} Thus, with probability at least $1-O(\eta^2 \delta)$, we have \begin{align*} ((\alpha_{1,t}^2-\alpha_{2,t}^2)^2-\Delta) \mathds{1}_{\cD_{t-1}} &< (1-4\eta \sigma^2)^t\left((\alpha_{1,0}^2-\alpha_{2,0}^2)^2+O(1) \Bar{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<(\alpha_{1,0}^2-\alpha_{2,0}^2)^2+O\left( (1-4\eta \sigma^2)^t \Bar{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<4\Delta+O\left( \frac{\sqrt{\Delta\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right) \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right). \end{align*} When $\cD_{t-1}$ holds, w.p. at least $1-O(\eta^2 \delta)$, by the inequality above we have \begin{align*} (\alpha_{1,t}^2-\alpha_{2,t}^2)^2&<5\Delta+O\left( \frac{\sqrt{\Delta\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right) \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) <6\Delta. \end{align*} Note that to make sure last terms is smaller than $\Delta$, we need \begin{align*} \frac{\sqrt{\eta}}{\sigma^2\sqrt{\Delta}}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right) \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)=O(1). \end{align*} As $\Delta=O(\frac{\eta}{\sigma^{10}})$, we know that as long as $\eta$ is polynomial in $\sigma$, choosing $\sigma=O((\log\frac{1}{\delta})^{-\frac{1}{3}})$ is sufficient. Now we know that if $\cD_{t-1}$ holds, $\cD_{t}$ holds w.p. at least $1-O(\eta^2 \delta)$. It is easy to show that the Perturbed GD updates satisfy $(\alpha_{1,t}^2-\alpha_{2,t}^2)^2<6\Delta$ in the following $O(\frac{1}{\eta^2})$ steps w.p. at least $1-\delta$. \end{proof} With Lemmas \ref{lem_bounded}, \ref{lem_loss} and \ref{a1=a2}, we can prove Lemma \ref{lem_convergence}. Here follows a brief proof. \begin{proof} \begin{align*} |1-x_t^\top u_*|&<(1+x_t^\top u_*)|1-x_t^\top u_*|\\ &=|1-(x_t^\top u_*)^2+x_t^\top u_*v_*^\top y_t-x_t^\top u_*v_*^\top y_t|\\ &\leq|1-x_t^\top u_*v_*^\top y_t|+|(x_t^\top u_*)^2-x_t^\top u_*v_*^\top y_t|\\ &=|1-x_t^\top M y_t|+x_t^\top u_*|x_t^\top u_*-v_*^\top y_t|\\ &\leq|1-x_t^\top M y_t|+\frac{\sqrt{2}}{\sigma}|x_t^\top u_*-v_*^\top y_t|\\ &<\sqrt{6\gamma}+\frac{\sqrt{2}}{\sigma}\sqrt{6\Delta}. \end{align*} The last inequality comes from Lemmas \ref{lma22} and \ref{lma4}. Together with Lemma \ref{lem_bounded} we can get \begin{align*} \norm{x_t-u_*}^2=(1-x_t^\top u_*)^2+\norm{\beta_{1,t}}_2^2<(\sqrt{6\gamma}+\frac{\sqrt{2}}{\sigma}\sqrt{6\Delta})^2+2\frac{\eta C_2}{\sigma^2}=O(\sigma^2+\frac{\eta}{\sigma^{10}}). \end{align*} Note that we use $C_2=O(\frac{1}{\sigma^6})$ when calculating the order. Similarly, we have \begin{align*} \norm{y_t-v_*}^2<(\sqrt{6\gamma}+\frac{\sqrt{2}}{\sigma}\sqrt{6\Delta})^2+2\frac{\eta C_2}{\sigma^2}=O(\sigma^2+\frac{\eta}{\sigma^{10}}). \end{align*} Then, for any $\epsilon>0$, by choosing $\sigma=O(\sqrt{\epsilon})$ and $\eta\leq\eta_5=O(\sigma^{10} \epsilon)$, we will have $\norm{x_t-u_*}^2<\epsilon$ and $\norm{y_t-v_*}^2<\epsilon$ \end{proof} \section{Proof of Theorem \ref{thm_rankr}}\label{pf_3} Recall that the gradient of $\tilde\cF$ takes the following form. \begin{align*} \nabla_X\tilde\cF(X,Y)&=(XY^\top-M)Y-d_2\sigma_2^2X,\\ \nabla_Y\tilde\cF(X,Y)&=(XY^\top-M)^\top X-d_1\sigma_1^2Y. \end{align*} Suppose $(U,V)$ is a stationary point. Then we have \begin{align} (UV^\top-M)V-d_2\sigma_2^2U&=0,\label{eq_stat_1}\\ (UV^\top-M)^\top U-d_1\sigma_1^2V&=0. \label{eq_stat_2} \end{align} \noindent $\bullet$ {\bf Step 1:} To prove the first statement, simply left multiply each side of \eqref{eq_stat_1} by $U^\top$ and each side of \eqref{eq_stat_2} by $V^\top,$ and we have the following equations. \begin{align*} U^\top UV^\top V-U^\top MV-d_2\sigma_2^2U^\top U&=0,\\ V^\top VU^\top U-V^\top M^\top U-d_1\sigma_1^2V^\top V&=0. \end{align*} Note that the following equation naturally holds. $$U^\top UV^\top V-U^\top MV=\left(V^\top VU^\top U-V^\top M^\top U\right)^\top.$$ Combine these three equations together and we have $$U^\top U=\frac{d_1\sigma_1^2}{d_2\sigma_2^2}(V^\top V)^\top=\gamma^2V^\top V.$$ \noindent $\bullet$ {\bf Step 2:} We next show that $(\tilde U,\tilde V)R$ is a stationary point, where $$(\tilde U,\tilde V)=(\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}},\frac{1}{\sqrt{\gamma} }B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}),$$ where $R\in \RR^{r\times r}$ is an orthogonal matrix. We only need to check \eqref{eq_stat_1} and \eqref{eq_stat_2}. In fact, we have \begin{align*} \nabla_X\tilde\cF(\tilde UR,\tilde VR)&=-\gamma\sigma^2 AB^\top\frac{1}{\sqrt{\gamma} }B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R+\sigma^2\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R\\ &=-\sigma^2\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R+\sigma^2\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R\\ &=0, \end{align*} and \begin{align*} \nabla_Y\tilde\cF(\tilde UR,\tilde VR)&=-\gamma\sigma^2 BA^\top\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R+\sigma^2\gamma^2\frac{1}{\sqrt{\gamma} }B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R\\ &=-\sigma^2\gamma\sqrt{\gamma} B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R+\sigma^2\gamma\sqrt{\gamma} B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R\\ &=0. \end{align*} Combine the above equations together and we know that $(\tilde U,\tilde V)$ is a stationary point. \noindent $\bullet$ {\bf Step 3:} We next show that $\{(\tilde U,\tilde V)R\big| R\in\RR^{r\times r}, \text{orthogonal}\}$ are the global minima and all other stationary points enjoy strict saddle property. Without loss of generality, we assume $\gamma=1.$ We first calculate the Hessian $\nabla^2 \tilde\cF(X,Y).$ The Hessian can be viewed as a matrix that operates on vectorized matrices of dimension $(d_1+d_2)\times r.$ Then, for any $W\in\RR^{(d_1+d_2)\times r}$, the Hessian defines a quadratic form $$[\nabla^2 \tilde\cF(W)](Z_1,Z_2)=\sum_{i,j,k,l}\frac{\partial^2\tilde\cF(W)}{\partial W[i,j]\partial W[k,l]}Z_1[i,j]Z_2[k,l], ~\forall Z_1,Z_2\in\RR^{(d_1+d_2)\times r}.$$ We can then express the Hessian $\nabla^2 \tilde\cF(W)$ as follows: \begin{align*} [\nabla^2 \tilde\cF(X,Y)](\Delta,\Delta)=2<XY^\top-M,\Delta_U\Delta_V^\top>+\norm{U\Delta_V^\top+\Delta_U V^\top}_{\rm{F}}^2+\sigma^2 \norm{\Delta_U}_{\rm{F}}^2+\sigma^2 \norm{\Delta_V}_{\rm{F}}^2, \end{align*} where $\Delta=\begin{bmatrix} \Delta_U\\ \Delta_V \end{bmatrix},$ $\Delta_U\in \RR^{d_1\times r}$ and $\Delta_V\in \RR^{d_2\times r}.$ We further denote $W=\begin{bmatrix} X\\ Y \end{bmatrix},$ $\tilde W=\begin{bmatrix} \tilde U\\ \tilde V \end{bmatrix},$ and $\tilde M=\tilde U\tilde V^\top,$ $$R=\argmin_{R'\in\RR^{r\times r}, \text{orthogonal}}\norm{W-\tilde WR'}.$$ We then have the following lemma. \begin{lemma}\label{lem_negative_curv} Let $\sigma_{\min}(M)$ be the smallest singular value of $M$. Suppose $d_1\sigma_1^2=d_2 \sigma_2^2=\sigma^2,$ and $\sigma^2<\sigma_{\min}(M).$ For $\forall(U,V)\in (\RR^{d_1},\RR^{d_2})$ such that $\nabla\tilde\cF(U,V)=0,$ we denote $\Delta= \begin{bmatrix} U-\tilde U R\\ V-\tilde V R \end{bmatrix},$ then we have \begin{align}\label{negative_curv} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)\leq -\norm{UV^\top-\tilde M}_{\rm{F}}^2-3\sigma^2\norm{A^\top U-B^\top V}_{\rm{F}}^2. \end{align} Moreover, $[\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)<0$ if $$(U,V)\notin \{(\tilde U,\tilde V)R'\big|R'\in \RR^{r\times r}, R'R'^\top=R'^\top R'=I_r \}.$$\end{lemma} \begin{proof} Recall that the quadratic form defined by the Hessian can be written as follows. \begin{align}\label{quadratic} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)=2<UV^\top-M,\Delta_U\Delta_V^\top>+\norm{U\Delta_V^\top+\Delta_U V^\top}_{\rm{F}}^2+\sigma^2 \norm{\Delta_U}_{\rm{F}}^2+\sigma^2 \norm{\Delta_V}_{\rm{F}}^2. \end{align} We start from the second term $\norm{U\Delta_V^\top+\Delta_U V^\top}_{\rm{F}}^2$. Similar to the proof of Claim B.5 in \citet{du2018algorithmic}, we have \begin{align*} \norm{U\Delta_V^\top+\Delta_U V^\top}_{\rm{F}}^2&= \norm{\Delta_U\Delta_V^\top+ UV^\top-\tilde M}_{\rm{F}}^2\\ &= \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2+2 <\Delta_U\Delta_V^\top, UV^\top-\tilde M>\\ &= \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2+2 <\Delta_U\Delta_V^\top, UV^\top-M>+2 <\Delta_U\Delta_V^\top, M-\tilde M>\\ &= \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2+2 <\Delta_U\Delta_V^\top, UV^\top-M>+2 \sigma^2<\Delta_U\Delta_V^\top, AB^\top> \end{align*} Plugging this equation into \eqref{quadratic}, we have \begin{align*} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)&=4<UV^\top-M,\Delta_U\Delta_V^\top>+ \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2\\&~~~+\sigma^2 \left(\norm{\Delta_U}_{\rm{F}}^2+ \norm{\Delta_V}_{\rm{F}}^2+2 <\Delta_U\Delta_V^\top, AB^\top> \right). \end{align*} Note that using the fact $\nabla\tilde\cF(U,V)=\nabla\tilde\cF(\tilde U,\tilde V)=0,$ one can easily verify that \begin{align*} 4<UV^\top-M,\Delta_U\Delta_V^\top>&=4<UV^\top-M,\tilde M>-2\sigma^2 \left(\norm{\Delta_U}_{\rm{F}}^2+ \norm{\Delta_V}_{\rm{F}}^2\right)+2\sigma^2 \left(\norm{\tilde U}_{\rm{F}}^2+ \norm{\tilde V}_{\rm{F}}^2\right)\\ &=-4\norm{UV^\top-\tilde M}_{\rm{F}}^2-4\sigma^2<AB^\top,\tilde M-UV^\top>\\&~~~+2\sigma^2\left(\norm{\tilde U}_{\rm{F}}^2+ \norm{\tilde V}_{\rm{F}}^2-\norm{ U}_{\rm{F}}^2-\norm{ V}_{\rm{F}}^2\right). \end{align*} Thus, \begin{align} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)&=-4\norm{UV^\top-\tilde M}_{\rm{F}}^2+ \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2\nonumber\\ &~~~-\sigma^2 \Bigg[\norm{\Delta_U}_{\rm{F}}^2+ \norm{\Delta_V}_{\rm{F}}^2-2 <\Delta_U\Delta_V^\top, AB^\top>\nonumber\\ &~~~~~-2\left(\norm{\tilde U}_{\rm{F}}^2+ \norm{\tilde V}_{\rm{F}}^2-\norm{ U}_{\rm{F}}^2-\norm{ V}_{\rm{F}}^2\right) +4<AB^\top,\tilde M-UV^\top>\Bigg].\label{eq_quad} \end{align} We then have the following two claims: \begin{claim}\label{claim_a} $-4\norm{UV^\top-\tilde M}_{\rm{F}}^2+ \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2\leq -\norm{UV^\top-\tilde M}_{\rm{F}}^2.$ \end{claim} \begin{proof} Similar to the proof of Claim B.5 in \citet{du2018algorithmic}, we have \begin{align*} \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2&\leq\frac{1}{4}\norm{\Delta\Delta^\top}_{\rm{F}}^2\\ &\leq \frac{1}{2}\norm{WW^\top-\tilde W\tilde W^\top}_{\rm{F}}^2\\ &=2\norm{UV^\top-\tilde M}_{\rm{F}}^2-\norm{U^\top\tilde U-V^\top\tilde V}_{\rm{F}}^2+\frac{1}{2}\norm{U^\top U-V^\top V}_{\rm{F}}^2+\frac{1}{2}\norm{\tilde U^\top \tilde U-\tilde V^\top \tilde V}_{\rm{F}}^2\\ &=2\norm{UV^\top-\tilde M}_{\rm{F}}^2-\norm{U^\top\tilde U-V^\top\tilde V}_{\rm{F}}^2\\ &\leq 2\norm{UV^\top-\tilde M}_{\rm{F}}^2. \end{align*} Thus, \begin{align*} -4\norm{UV^\top-\tilde M}_{\rm{F}}^2+ \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2&\leq-4\norm{UV^\top-\tilde M}_{\rm{F}}^2+ 2\norm{UV^\top-\tilde M}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2\\&=-\norm{UV^\top-\tilde M}_{\rm{F}}^2. \end{align*} \end{proof} \begin{claim}\label{claim_b} $\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2 +2<AB^\top,\tilde M-UV^\top>=\norm{A^\top \Delta_U-B^\top \Delta_V}_{\rm{F}}^2 .$ \end{claim} \begin{proof} First, the LHS of the equation can be rewritten as follows. \begin{align*} &\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2 +2<AB^\top,\tilde M-UV^\top>\\ =&\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>+4<AB^\top,\tilde M>-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2\\ =&\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>+4\norm{\tilde U}_{\rm{F}}^2-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2\\ =&\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2 +\norm{\tilde U}_{\rm{F}}^2+\norm{\tilde V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>, \end{align*} where we use the fact $$<AB^\top,\tilde M>=\norm{\tilde U}_{\rm{F}}^2=\norm{\tilde V}_{\rm{F}}^2.$$ On the other hand, the RHS of the equation can be rewritten as follows. \begin{align*} \norm{A^\top \Delta_U-B^\top \Delta_V}_{\rm{F}}^2&=\norm{\Delta_U}_{\rm{F}}^2+ \norm{\Delta_V}_{\rm{F}}^2-2 <\Delta_U\Delta_V^\top, AB^\top>\\ &=\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2 +\norm{\tilde U}_{\rm{F}}^2+\norm{\tilde V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>\\ &~~~~~-2\tr(U(\tilde UR)^\top)-2\tr(V(\tilde VR)^\top)+2\tr(U(\tilde UR)^\top)+2\tr(V(\tilde VR)^\top)\\ &=\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2 +\norm{\tilde U}_{\rm{F}}^2+\norm{\tilde V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>\\ &=\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2 +2<AB^\top,\tilde M-UV^\top>. \end{align*} \end{proof} Plugging the conclusions in Claims \ref{claim_a} and \ref{claim_b} into \eqref{eq_quad}, we have \begin{align*} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)\leq-\norm{UV^\top-\tilde M}_{\rm{F}}^2-3\sigma^2\norm{A^\top \Delta_U-B^\top \Delta_V}_{\rm{F}}^2. \end{align*} Note that since $A^\top \tilde U R=B^\top \tilde V R,$ we have $A^\top \Delta_U-B^\top \Delta_V=A^\top U-B^\top V.$ To justify our last statement, we have the following claim. \begin{claim}\label{claim_v} $\norm{UV^\top-\tilde M}_{\rm{F}}^2+3\sigma^2\norm{A^\top U-B^\top V}_{\rm{F}}^2=0$ if and only if $$(U,V)=(\tilde U,\tilde V)R,$$ where $R\in \RR^{r\times r}$ is an orthogonal matrix. \end{claim} \begin{proof} $\norm{UV^\top-\tilde M}_{\rm{F}}^2+3\sigma^2\norm{A^\top U-B^\top V}_{\rm{F}}^2=0$ if and only if \begin{align} UV^\top&=\tilde M,\label{con1}\\ A^\top U&=B^\top V.\label{con2} \end{align} Left multiplying each side of \eqref{con1} by $A^\top,$ we have \begin{align*} A^\top UV^\top=(\Sigma-\sigma^2I)B^\top &\Leftrightarrow B^\top VV^\top=(\Sigma-\sigma^2I)B^\top\\ &\Leftrightarrow VV^\top=B(\Sigma-\sigma^2I)B^\top\\ &\Leftrightarrow V=B(\Sigma-\sigma^2I)^{\frac{1}{2}}R=\tilde VR' . \end{align*} The last equivalent argument comes from Theorem 4 in \citet{li2019symmetry}. Plugging $V=\tilde VR'$ back to \eqref{con2}, we have $U=\tilde UR'.$ \end{proof} As a direct result of Claim \ref{claim_v}, for stationary point $(U,V)\notin\{(\tilde U,\tilde V)R'\big|R'\in \RR^{r\times r}, R'R'^\top=R'^\top R'=I_r \},$ we have $$[\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)<0.$$ We prove the lemma. \end{proof} Lemma \ref{lem_negative_curv} directly implies that $\{(\tilde U,\tilde V)R'\big|R'\in \RR^{r\times r}, R'R'^\top=R'^\top R'=I_r \}$ contains all the global optima, and all other stationary points enjoy strict saddle property. \newpage \section{Perturbed GD}\label{alg_detail} The detail of the Perturbed GD algorithm is summarized in Algorithm \ref{alg:Perturbed GD}. \begin{algorithm}[H] \caption{Perturbed Gradient Descent for Rank-1 Matrix Factorization.} \label{alg:Perturbed GD} \begin{algorithmic} \STATE{\textbf{Input}: step size $\eta$, noise level $\sigma_1, \sigma_2$, matrix $M \in \RR^{d_1 \times d_2}$, number of iterations $T$.} \STATE{\textbf{Initialize}: initialize $(x_0, y_0)$ arbitrarily.} \FOR{$t = 0 \ldots T-1$} \STATE{Sample $\xi_{1,t} \sim N(0, \sigma_1^2 I_{d_1})$ and $\xi_{2,t} \sim N(0, \sigma_2^2 I_{d_2})$.} \STATE{$\tilde{x}_t = x_t + \xi_{1,t}, ~ \tilde{y}_t = y_t + \xi_{2,t}.$} \STATE{$x_{t+1} = x_t - \eta (\tilde{x}_t \tilde{y}_t^\top - M ) \tilde{y}_t.$} \STATE{$y_{t+1} = y_t - \eta (\tilde{y}_t \tilde{x}_t^\top - M^\top ) \tilde{x}_t.$} \ENDFOR \end{algorithmic} \end{algorithm} \subsection{ Rank-1 Matrix Factorization} We consider the following nonconvex optimization problem: \begin{align}\label{mat_fct_ncvx} \min_{x\in\RR^{d_1}, y\in\RR^{d_2}}\cF(x,y)=\frac{1}{2}\|xy^\top-M\|^2_{\rm{F}}, \end{align} where $M \in \RR^{d_1\times d_2}$ is a rank-1 matrix and can be factorized as follows: $\displaystyle M=u_*v_*^\top, $ where $u_*\in\RR^{d_1}$, $v_*\in\RR^{d_2}$. Without loss of generality, we assume $\norm{u_*}_2=\norm{v_*}_2=1.$ The optimization landscape of \eqref{mat_fct_ncvx} has been well studied in the previous literature \citep{ge2016matrix, ge2017no, chi2019nonconvex, li2019symmetry}. Because of the bilinear form in $\cF,$ there exist infinitely many global minima to \eqref{mat_fct_ncvx}, including highly unbalanced ones, i.e., $xy^\top=M$ with $\norm{x}_2\gg\norm{y}_2$ or $\norm{x}_2\gg\norm{y}_2$ (see Definition \ref{def:balance}). In Section \ref{main}, we will show that such unbalancedness essentially implies global minima with a large condition number. To address this issue, \citet{tu2015low,ge2017no} propose a regularizer of the form $(\norm{x}_2^2-\norm{y}_2^2)^2$ to balance $\norm{x}_2$ and $\norm{y}_2$. Recently, \citet{du2018algorithmic} show that even without explicit regularization, gradient descent with small random initialization converges to balanced solutions with constant probability. Yet, all of the previous results assume noiseless updates. The algorithmic behavior of first order algorithms with noisy updates remains unclear for the nonconvex matrix factorization problem. \subsection{Perturbed Gradient Descent} To study the effect of noise, we consider a perturbed gradient descent algorithm (Perturbed GD). At the $t$-th iteration, we first perturb the iterate $(x_t,y_t)$ with independent Gaussian noise $\xi_{1,t}\sim N(0,\sigma_1^2I_{d_1})$ and $\xi_{2,t}\sim N(0,\sigma_2^2I_{d_2}),$ respectively. We then update $(x_t, y_t)$ with the gradient evaluated at the perturbed iterates. The detail of the Perturbed GD algorithm is summarized in Algorithm \ref{alg:Perturbed GD}. \begin{algorithm}[H] \caption{Perturbed Gradient Descent for Rank-1 Matrix Factorization.} \label{alg:Perturbed GD} \begin{algorithmic} \STATE{\textbf{Input}: step size $\eta$, noise level $\sigma_1, \sigma_2$, matrix $M \in \RR^{d_1 \times d_2}$, number of iterations $T$.} \STATE{\textbf{Initialize}: initialize $(x_0, y_0)$ arbitrarily.} \FOR{$t = 0 \ldots T-1$} \STATE{Sample $\xi_{1,t} \sim N(0, \sigma_1^2 I_{d_1})$ and $\xi_{2,t} \sim N(0, \sigma_2^2 I_{d_2})$.} \STATE{$\tilde{x}_t = x_t + \xi_{1,t}, ~ \tilde{y}_t = y_t + \xi_{2,t}.$} \STATE{$x_{t+1} = x_t - \eta (\tilde{x}_t \tilde{y}_t^\top - M ) \tilde{y}_t.$} \STATE{$y_{t+1} = y_t - \eta (\tilde{y}_t \tilde{x}_t^\top - M^\top ) \tilde{x}_t.$} \ENDFOR \end{algorithmic} \end{algorithm} \textbf{Smoothing Effect}. Using Perturbed GD to solve \eqref{mat_fct_ncvx} can also be viewed as solving the following stochastic optimization problem: \begin{align}\label{mat_fact_eq} \min_{x\in\RR^{d_1}, y\in\RR^{d_2}}\tilde\cF(x,y)=\EE_{\xi_{1},\xi_{2}}\cF(x+\xi_1,y+\xi_2), \end{align} where $\xi_{1}\sim N(0,\sigma_1^2 I_{d_1})$ and $\xi_{2}\sim N(0,\sigma_2^2 I_{d_2}).$ Throughout this paper, we will refer to problem \eqref{mat_fact_eq} as the smoothed problem. The expectation in \eqref{mat_fact_eq} can be viewed as convoluting the objective function with a Gaussian kernel. In Section \ref{main}, we show that this convolution effectively smooths out unbalanced optima and yields a benign landscape. \begin{remark} The use of random noise to convolute with the objective function is also known as randomized smoothing, which is first proposed in \citet{duchi2012randomized}. \citet{zhou2019towards,jin2017escape} further exploit this effect to explain the importance of noise in helping first order algorithms to escape from strict saddle points and spurious local optima. \end{remark} \section{Numerical Experiments}\label{sec_numerical} We present numerical results to support our theoretical findings. We compare our Perturbed GD algorithm with gradient descent (GD), and demonstrate that Perturbed GD with $\gamma-$balanced noise converges to $\gamma-$balanced optima, while the optima obtained by GD are highly sensitive to initialization and step size. We also show that the phase transition between Phase II and Phase III is not an artifact of the proof, and faithfully captures the true algorithmic behavior of Perturbed GD. \noindent\textbf{Rank-1 Matrix Factorization}. We first consider the rank-1 matrix factorization problem. Without loss of generality, the matrix $M$ to be factorized is given by $M = u_* v_*^\top$, where $u_* = (1, 0, \ldots, 0)\in \RR^{d_1}$ and $v_* = (1, 0, \ldots, 0)\in \RR^{d_2}$, with $d_1 = 20$ and $d_2 = 30$. We initialize iterates $(x_0, y_0)$ with $x_0 \sim N(0,\sigma_x^2 I_{d_1}),\ y_0 \sim N(0,\sigma_y^2 I_{d_2})$. For all experiments, we use $\gamma$-balanced noise in Perturbed GD. Specifically, we choose $\xi_{1,t} \sim N(0, \sigma_1^2 I_{d_1}), \xi_{2,t} \sim N(0, \sigma_2^2 I_{d_2})$. \noindent\textbf{(1) Balanced Noise}. We consider the case of balanced noise ($\gamma =1$), where we take $\sigma_1 = \sqrt{1.5} \times 0.05$ and $\sigma_2 = 0.05$. One can verify that $\gamma^2 = {d_1 \sigma_1^2}/{(d_2 \sigma_2^2)} = 1$. We first use balanced step size to compare with GD studied in \citet{du2018algorithmic}. Specifically, we set $\eta_x = \eta_y = 10^{-2}$ for both Perturbed GD and GD. We further consider two initialization schemes: (i) small initializations (which is also adopted in \citet{du2018algorithmic}): $\sigma_x = \sigma_y = 10^{-2}$; (ii) large initializations: $\sigma_x = \sigma_y = 10^{-1}$. Fig. \ref{mainresult}.(a, b, d, e) summarize the results of $100$ repeated experiments in a box-plot. As can be seen, regardless of initializations, Perturbed GD always converges to the balanced optima (Fig. \ref{mainresult}.(a, b)). In contrast, GD only converges to approximately balanced optima for small initializations (Fig. \ref{mainresult}.(d)), and the large initialization yields a large variance in terms of the balancedness of the obtained solution (Fig. \ref{mainresult}.(e)). \begin{figure*}[t] \centering \includegraphics[width=0.9\linewidth]{box-rk1.pdf} \caption{Perturbed GD with balanced noise (a, b and c), Perturbed GD with unbalanced noise (g) and GD (d,e and f) for the rank-1 matrix factorization problem. (a) and (d) use small initializations ($\sigma_x = \sigma_y = 10^{-2}$) and balanced step size ($\eta_x = \eta_y = 10^{-2}$). (b) and (e) use large initializations ($\sigma_x = \sigma_y = 10^{-1}$) and balanced step size. (c) and (f) use small initializations and unbalanced step size ($\eta_x = 0.5 \eta_y =5 \times 10^{-3}$).}\label{mainresult} \end{figure*} We also use unbalanced step size (different step sizes for updating $x$ and $y$) and compare the convergence properties of Perturbed GD and GD. Specifically, we set $\eta_x = 0.5 \eta_y =5 \times 10^{-3}$ for both Perturbed GD and GD. We adopt a small initialization scheme, with $\sigma_x = \sigma_y = 10^{-2}$. Fig. \ref{mainresult}.(c, f) summarize the results of $100$ repeated simulations in a box-plot. As can be seen, even with small initializations, GD with unbalanced step size converges to the approximately $\sqrt{0.5}$-balanced optima, instead of the $1$-balanced optima (Fig. \ref{mainresult}.(f)). In contrast, Perturbed GD is able to converge to the balanced optima with unbalanced step size (Fig. \ref{mainresult}.(c)). Our results suggest that the noise is the most important factor in determining the balancedness of the solutions obtained by Perturbed GD. \noindent\textbf{(2) Unbalanced Noise}. We run Perturbed GD with unbalanced noise. We take $\sigma_1 = \sqrt{0.75} \times 0.05$ and $\sigma_2 = 0.05$ with $\gamma^2 ={d_1 \sigma_1^2}/ {(d_2 \sigma_2^2)}= 0.5$. We use a small initialization: $\sigma_x = \sigma_y = 10^{-2}$, and balanced step size: $\eta_x = \eta_y = 10^{-2}$. Fig. \ref{mainresult}.(g) summarizes the results of $100$ repeated simulations in a box-plot. As can be seen, for $\gamma \neq 1$, the Perturbed GD converges to the $\gamma$-balanced optima. \noindent \textbf{Rank-10 Matrix Factorization}. We then consider rank-10 nonconvex matrix factorization problem. The matrix $M$ to be factorized is given by $M = U_* V_*^\top$, where $ U_* =\begin{pmatrix} I_{10}, 0 \end{pmatrix}_{d_1\times10}^\top, V_* =\begin{pmatrix} I_{10}, 0 \end{pmatrix}_{d_2\times10}^\top, $ with $d_1 = 20$ and $d_2 = 30$. We initialize iterates $(X_0, Y_0)$ with all entries ${X_0}^{(i,j)}$'s and ${Y_0}^{(i,j)}$'s independently sampled from $N(0,\sigma_x^2)$ and $N(0,\sigma_y^2)$, respectively. For all experiments, we use $\gamma$-balanced noise in Perturbed GD. Specifically, we choose $\xi_{1,t}$ and $\xi_{2,t}$ with i.i.d. elements drawn from $N(0,\sigma_1^2)$ and $N(0,\sigma_2^2)$ respectively. We repeat a similar set of experiments as in rank-1 case, and summarize the results in Fig. \ref{mainresultrk-10}. For each of the experiments, we use the same set of $(\eta_x, \eta_y, \sigma_x, \sigma_y)$ as their counterpart in the rank-1 case. As can be seen, Perturbed GD always converges to the $\gamma-$balanced optima, regardless of initializations. Our experiments suggest that, for the rank-r matrix factorization problem, the noise still determines the balancedness of the optima obtained by Perturbed GD. \begin{figure*}[t] \centering \includegraphics[width=0.9\linewidth]{box-rk10.pdf} \caption{Perturbed GD with balanced noise (a, b and c) and GD (d,e and f) for rank-10 matrix factorization problem. (a) and (d) use small initializations ($\sigma_x = \sigma_y = 10^{-2}$) and balanced step size ($\eta_x = \eta_y = 10^{-2}$). (b) and (e) use large initializations ($\sigma_x = \sigma_y = 10^{-1}$) and balanced step size. (c) and (f) use small initializations and unbalanced step size ($\eta_x = 0.5 \eta_y =5 \times 10^{-3}$).}\label{mainresultrk-10} \end{figure*} \noindent \textbf{Phase Transition.} We further demonstrate the transition between Phase II and Phase III in the Perturbed GD algorithm. Specifically, we consider 2-dimensional problem $f(x,y) = (1-xy)^2$ with balanced optima $\pm(1,1)$. We set $\sigma_1 = \sigma_2 = 0.05$, initialize $(x_0 ,y_0) = (3,5)$, and use balanced step size $\eta_x = \eta_y = 0.01$. We repeat the experiments $50$ times and summarize the result of one realization in Fig. \ref{2phase}, as the convergence properties of Perturbed GD are highly consistent across different realizations. We also use exponential moving average to smooth the loss trajectory to better illustrate the overall progress of the objective in Phase III. As can be seen, in around the first $30$ to $40$ iterations, Perturbed GD and GD behave similarly. Both Perturbed GD and GD iterate towards the set of global optima $\{(x,y)\big| xy=1\}$, while driving the loss to zero, and the squared norm ratio $x_t^2/y_t^2$ in Perturbed GD decreases from $0.36$ to around $0.004$. \begin{figure}[t] \includegraphics[width=0.9\linewidth]{phase.pdf} \caption{Algorithmic behaviors of Perturbed GD and GD. For Perturbed GD, phase transition happens around the first $30\!\sim\!40$ iterations, as shown in (a,b,c). GD does not show phase transitions. }\label{2phase} \end{figure} After that, GD converges to the unbalanced optimum. Since the loss is sufficiently small, the noise dominates the update of Perturbed GD. Then the squared norm ratio $x_t^2/y_t^2$ gradually increases from $0.004$ to $1$. Perturbed GD iterates towards the balanced optimum while staying close to global optima. The phase transition phenomenon can also been observed for higher dimensional problems. As shown in Figure \ref{fig:addit} for $d=4$, the loss greatly decreases to and stay around $10^{-4}$ in the first $2\times10^3$ iterations, and then the squared norm ratio ${\norm{x}_2^2}/{\norm{y}_2^2}$ gradually increases from $0.5$ to $1$. This implies the transition between Phase II and Phase III, that is Perturbed GD first approaches the set of global minima and then converges to the balanced optimum. \begin{figure}[t]\label{fig:addit} \includegraphics[width=0.48\linewidth]{PGD_loss.pdf} \includegraphics[width=0.465\linewidth]{PGD_ratio.pdf} \caption{Algorithmic behaviors of Perturbed GD and GD for $d=4$. For Perturbed GD, phase transition happens around the first $2\times10^3$ iterations. } \end{figure} \section{Discussions}\label{sec_discussion} \vspace{-0.1in} \noindent {\bf Connections to SGD.} This paper studies the implicit bias of the noise in nonconvex optimization. Direct analysis on SGD is beyond current technical limit due to complex dependencies. Specifically, the noise in SGD comes from random sampling of the training data, and heavily depends on the iterate. This induces a complex dependency between iterate and the noise, and makes it difficult to characterize the distribution of the noise. Our Perturbed GD can be viewed as a close variant of SGD. Moreover, the noise in Perturbed GD follows Gaussian distribution and is independent of the iterates. Hence, the analysis of Perturbed GD, though still highly non-trivial, is now technically manageable. \noindent {\bf Biased Gradient Estimator.} Different from SGD, Perturbed GD implements a biased gradient estimator, i.e., $ \EE_{\xi_{1},\xi_{2}}\nabla\cF(x+\xi_1,y+\xi_2)\neq\nabla\cF(x,y).$ Such biased gradient also appears in training deep neural networks combined with computational heuristics. Specifically, \cite{luo2018towards} show that with batch normalization, the gradient estimator in SGD is also biased with respect to the original loss. The similarity between this biased gradient and our perturbed gradient is worth future investigation. \noindent {\bf Extension to Other Types of Noise.} Our work considers Gaussian noise, but can be extended to analyzing other types of noise. For example, we can show that anisotropic noise will have different smoothing effects along different directions. The implicit bias will thus depend on the covariance of noise in addition to the noise level. For another example, heavy tailed distribution of noise will affect the probability of the convergence of Perturbed GD. It may be difficult to achieve high probability convergence as we have shown for light tailed distributions. \noindent {\bf Sharp/Flat Minima and Phase Transition.} Lemma \ref{lem_condition_number} shows that the Hessian matrix of an unbalanced global optimum is ill-conditioned, and the landscape around such an optimum is sharp in some directions and flat in others. For nonconvex matrix factorization, all the global optima are connected and form a path. The landscape around the path forms a valley, which is narrow around unbalanced optima and wide around the balanced ones (See Figure \ref{fig:landscape2}). Our three-phase convergence analysis shows that the Perturbed GD first falls into the valley, and then traverses within the valley until it finds the balanced optima. \begin{figure}[t] \centering \includegraphics[width=0.95\linewidth]{landscape4} \caption{ The visualization of objective $\cF(x,y)=(1-xy)^2.$ All the global optima are connected and form a path. The landscape around the path forms a valley. Around unbalanced optima, the landscape is sharp in some directions and flat in others. Around balanced optima, the landscape only contains flat directions.} \label{fig:landscape2} \end{figure} As we have mentioned earlier, people have shown that the local optima of deep neural networks are also connected. Thus, the phase transition also provides a new perspective to explain the plateau of training curves after learning rate decay in training neural networks. Our analysis suggests that after adjusting the learning rate, the algorithm enters a new phase, where the noise slowly re-adjusts the landscape. At the beginning of this phase, the loss decreases rapidly due to the reduced noise level. By the end of this phase, the algorithm falls into a region with benign landscape that is suitable for further decreasing the step size. \noindent {\bf Related Literature.} Implicit bias of noise has also been studied in \citet{haochen2020shape,blanc2020implicit}. However, they consider perturbing labels while our work considers perturbing parameters. In a broader sense, our work is also related to \citet{li2019towards} which show that the noise scale of SGD may change the learning order of patterns. However, they do not study the implicit bias of noise towards certain optima. \section{Introduction} Nonconvex optimization has been widely adopted in various domains, including image recognition \citep{hinton2012deep, krizhevsky2012imagenet}, Bayesian graphical models \citep{jordan2004graphical, attias2000variational}, recommendation systems \citep{salakhutdinov2007restricted}, etc. Despite the fact that solving a nonconvex problem is generally difficult, empirical evidences have shown that simple first order algorithms such as stochastic gradient descent (SGD), are able to solve a majority of the aforementioned nonconvex problems efficiently. The theory behind these empirical observations, however, is still largely unexplored. In classical optimization literature, there have been fruitful results on characterizing the convergence of SGD to first-order stationary points for nonconvex problems. However, these types of results fall short of explaining the empirical evidences that SGD often converges to global minima for a wide class of nonconvex problems used in practice. More recently, understanding the role of noise in the algorithmic behavior of SGD has received significant attention. For instance, \citet{jin2017escape} show that a perturbed form of gradient descent is able to escape from strict saddle points and converge to second-order stationary points (i.e., local minima). \citet{zhou2019towards} further show that noise in the update can help SGD to escape from spurious local minima and converge to the global minima. We argue that, despite all these recent results on showing the convergence of SGD to global minima for various nonconvex problems, there is still an important question yet to be addressed. Specifically, the convergence of SGD in the presence of multiple global minima remains uncleared for many important nonconvex problems. For example, for over-parameterized neural networks, it has been shown that the nonconvex objective has multiple global minima \citep{kawaguchi2016deep}, only few of which can yield good generalization. In addition, \citet{allen2018convergence} show that for an over-parameterized network, SGD can converge to a global minimum in polynomial time. Combining with the empirical successes of training over-parameterized neural networks with SGD, these results strongly advocate that SGD not only can solve the nonconvex problem efficiently, but also implicitly biases towards solutions with good generalization ability. Motivated by this, this paper aims to provide more theoretical insights to the following question: \begin{center} \textbf{\emph{Does noise impose implicit bias towards certain minimizer in nonconvex optimization problems?}} \end{center} We answer this question through investigating a simple yet non-trivial problem -- nonconvex matrix factorization, which serves as an important foundation for a wide spectrum of problems such as matrix sensing \citep{bhojanapalli2016dropping,zhao2015nonconvex,chen2015fast,tu2015low}, matrix completion \citep{keshavan2010matrix,hardt2014understanding,zheng2016convergence}, and deep linear networks \citep{ji2018gradient, gunasekar2018implicit}. Given a matrix $M \in \RR^{d_1 \times d_2} $, the nonconvex matrix factorization aims to solve: \begin{align}\label{mf_obj} \min_{X\in \RR^{d_1 \times r},Y \in \RR^{d_2 \times r}} \frac{1}{2}\norm{XY^\top - M}_{\rm{F}}^2. \end{align} Despite its simplicity, \eqref{mf_obj} possesses several intriguing landscape properties: nonconvexity of the objective, all the saddle points satisfy the strict saddle property, and infinitely many global optima due to scaling and rotational invariance. Specifically, for any pair of global optimum $(X^*, Y^*)$, $(\alpha X^*, \frac{1}{\alpha} Y^*)$ and $(X^*R, Y^*R)$ are also global optima for any non-zero constant $\alpha$ and rotation matrix $R \in \RR^{r \times r}$. This is different from symmetric matrix factorization, which only possesses rotational invariance. The scaling and rotational invariance also imply that the global minima of \eqref{mf_obj} are connected, a landscape property that is also shared by deep neural networks \citep{nguyen2017loss, draxler2018essentially,nguyen2018loss,venturi2018spurious,garipov2018loss,liang2018understanding,nguyen2019connected,kuditipudi2019explaining}. Nonconvex matrix factorization \eqref{mf_obj} has been recently studied by \citet{du2018algorithmic}, with focus on the algorithmic behavior of gradient descent (GD). Their results reveal an interesting algorithmic regularization imposed by gradient descent: (i) gradient flow (GD with an infinitesimal step size) has automatic balancing property, i.e., the difference of the squared norm $\norm{X}_{\rm{F}}^2 - \norm{Y}_{\rm{F}}^2$ stays constant during training. (ii) for properly chosen step size, GD converges asymptotically for rank-r case, linearly for rank-1 case, while maintaining approximate balancing property. However, \citet{du2018algorithmic} do not consider any noise in the update, hence their results can not provide further theoretical insights on understanding the role of noise when applying first order algorithms to nonconvex problems. In this paper, we are interested in studying the algorithmic behavior of first order algorithms in the presence of noise. Specifically, we study a perturbed form of gradient descent (Perturbed GD) applied to the matrix factorization problem \eqref{mf_obj}, which injects independent noise to iterates, and then evaluates gradient at the perturbed iterates. Note that our algorithm is different from SGD in terms of the noise. For our algorithm, we inject independent noise to the iterates $(X_t, Y_t)'s$ and use the gradient evaluated at the perturbed iterates. The noise of SGD, in contrast, comes from the training sample. As a consequence, the noise of SGD has very complex dependence on the iterate, which is difficult to analyze. See more detailed discussions in Sections \ref{sec_discussion}. We further analyze the convergence properties of our Perturbed GD algorithm for the rank-1 case. At the early stage, noise helps the algorithm to escape from regions with undesired landscape, including the strict saddle point. After entering the region with benign landscape, Perturbed GD behaves similarly to gradient descent, until the loss is sufficiently small. At the early stage, noise provides additional explorations that help the algorithm to escape from the strict saddle point. Then at the later stage, the noise dominates the update of Perturbed GD, and gradually rescales the iterates to a balanced solution that is uniquely determined by the injected noise. Specifically, the ratio of the norm $\norm{x_t}_2/\norm{y_t}_2$ is completely determined by the ratio of the variance of noise injected to $(x_t, y_t)$. To the best of our knowledge, this is the first theoretical result towards understanding the implicit bias of noise in nonconvex optimization problems. Our analysis reveals an interesting characterization of the local landscape around global minima, which relates to the sharp/flat minima in deep neural networks \citep{keskar2016large}, and we will further discuss these connections in detail in Section \ref{sec_discussion}. We believe that investigating the implicit bias of the noise in nonconvex matrix factorization can serve as a fundamental building block for studying stochastic optimization for more sophiscated nonconvex problems, including training over-parameterized neural networks. \noindent{\textbf{Notations}}: $\mathbf{1}$ Given a matrix $A$, $\tr(A)$ denotes the trace of $A.$ For matrices $A,B\in \RR^{n\times m},$ we use $\inner{A}{B}$ to denote the Frobenius inner product, i.e., $\inner{A}{B}=\tr(A^\top B).$ $\norm{A}_{\rm{F}}=\sqrt{\inner{A}{A}}$ denotes the Frobenius norm of $A.$ $I_d\in \RR^{d\times d}$ denotes the identity matrix. \section{Main Results}\label{main} We study the algorithmic behavior of our proposed perturbed gradient descent (Perturbed GD) algorithm. We primarily focus on the rank-1 nonconvex matrix factorization. We first characterize the landscape of the original problem \eqref{mat_fct_ncvx}, and show that noise effectively smooths the original problem, yielding a smoothed problem \eqref{mat_fact_eq} with benign landscape. We then provide a non-asymptotic convergence analysis of the Perturbed GD, and demonstrate the implicit bias induced by the noise. In particular, we consider the case where noise is balanced, see more details in Theorem \ref{thm_main}. Due to space limit, we defer all the proofs to the appendix. We analyze the landscape of the original problem \eqref{mat_fct_ncvx} and the smoothed problem \eqref{mat_fact_eq}. Note that due to the bilinear form in $\cF(x,y)$, we have for $\alpha \neq 0,$ $\displaystyle\cF\left(\alpha x,\alpha^{-1} y\right)=\cF( x,y).$ The scaling invariance nature of \eqref{mat_fct_ncvx} results in undesired landscape properties, and makes the analysis of first order algorithms particularly difficult. To facilitate further discussions, below we characterize the landscape of \eqref{mat_fct_ncvx}. \begin{lemma}[\bf Landscape Analysis]\label{lem_landscape} The gradients of $\cF$ with respect to $x$ and $y$ take the form: \begin{align*} \nabla_{x}\cF(x,y)=(xy^\top-M)y,~ \nabla_{y}\cF(x,y)=(xy^\top-M)^{\top}x. \end{align*} Then $\cF$ has two types of stationary points: (i) For any $\alpha\neq 0,$ $\left(\alpha u_*, \alpha^{-1} v_*\right)$ is a global optimum; (ii) For any $x\in \RR^{d_1},y\in\RR^{d_2}$ such that $x^\top u_*=y^\top v_*=0,$ $(x,0)$ and $(0,y)$ are strict saddle. \end{lemma} The scaling-invariance leads to infinitely many global optima for \eqref{mat_fct_ncvx}, each taking the form $(\alpha u_*, \alpha^{-1} v_*)$. However, Lemma \ref{lem_condition_number} shows that different values of $\alpha$ lead to significantly different local landscape around the global minima. \begin{lemma}\label{lem_condition_number} The condition number of the Hessian matrix of $\cF$ at the global optimum $(\alpha u_*,\alpha^{-1} v_*)$ is $$\kappa\left(\nabla^2 \cF\left(\alpha u_*,\alpha^{-1} v_*\right)\right)=\max\left\{\alpha^4,\alpha^{-4}\right\}+1.$$ \end{lemma} \begin{figure*} \centering \includegraphics[width=\linewidth]{landscape1} \vspace{-0.2in} \caption{ The visualization of objective functions $\cF(x,y)=(1-xy)^2$ and $\tilde\cF(x,y)$ with $x,y\in \RR$ and $\sigma_1^2=\sigma_2^2=0.0975.$ For $\cF(x,y),$ any $(x,y)$ that satisfies $xy=1$ is a global minimum (as shown in (a)). $\tilde\cF(x,y)$ only has global minima close to $\pm(1,1)$ (as shown in (c)).}\label{fig:landscape} \vspace{-0.1in} \end{figure*} From Lemma \ref{lem_condition_number} we know that when $|\alpha|$ is extremely large or small, the global minimum $(\alpha u_*, \alpha^{-1} v_*)$ is ill-conditioned (large condition number). The condition number also relates to the sharp/flat minima in neural networks, which we will discuss in detail in Section \ref{sec_discussion}. Our previous discussion implies that any global optimum $(x_*, y_*)$ to \eqref{mat_fct_ncvx} will be ill-conditioned if the norm ratio $\norm{x_*}_2/\norm{y_*}_2$ is close to zero or infinity. To facilitate further discussions, we define the balancedness property as follows. \begin{definition}[$\gamma$-balancedness]\label{def:balance} We say that $(x,y)\in(\RR^{d_1},\RR^{d_2})$ is $\gamma-$balanced for some positive number $\gamma$ if $\norm{x}_2=\gamma\norm{y}_2.$ Informally, we say $(x,y)$ is unbalanced if $\gamma$ is close to zero or infinity, and $(x,y)$ is balanced if $\gamma$ is close to 1.\end{definition} Our next lemma shows that, in the presence of noise, the smoothed problem \eqref{mat_fact_eq} has only balanced optima, and the balancedness is completely determined by the ratio of the variance of the noise injected to the iterates. \begin{lemma}\label{lem_convolutional} We say the noise in the Perturbed GD is $\gamma-$balanced if $\EE\left[\norm{\xi_1}_2^2\right]=\gamma^2\EE\left[\norm{\xi_2}_2^2\right]$. For $\gamma = 1$, we say the noise is balanced. For any noise ratio $\gamma>0,$ if we take $\sigma^2=\EE\left[\norm{\xi_1}_2^2\right]=\gamma^2\EE\left[\norm{\xi_2}_2^2\right]\leq \min\{\gamma,1\},$ \eqref{mat_fact_eq} has two global optima $(\tilde u_*(\gamma),\tilde v_*(\gamma))=\pm\sqrt{\gamma-\sigma^2}\left(u_*,\gamma^{-1}v_*\right)$ and one saddle point $(0,0)$ with $\lambda_{\min}(\nabla^2 \tilde F(0,0))<0$. \end{lemma} We see that when the noise is $\gamma$-balanced and the noise level is sufficiently small, the global minima of the smoothed problem \eqref{mat_fact_eq} can be arbitrarily close to the $\gamma$-balanced global minima of the original rank-1 matrix factorization problem \eqref{mat_fct_ncvx}. Compared with Lemma \ref{lem_landscape}, Lemma \ref{lem_convolutional} shows that noise effectively smooths the landscape of the original problem \eqref{mat_fct_ncvx}, the unbalanced optima (with ill-conditioned Hessian) to \eqref{mat_fct_ncvx} are no longer optima after convolution. See Figure \ref{fig:landscape} for detailed illustration of the landscape for the original problem \eqref{mat_fct_ncvx} and the smoothed problem \eqref{mat_fact_eq}. Moreover, one can verify that the saddle point $(0,0)$ of problem \eqref{mat_fact_eq} enjoys the strict saddle property, i.e., the smallest eigenvalue of Hessian at $(0,0)$ is negative. By the stable manifold theory proposed in \citet{lee2019first}, Perturbed GD with an infinitesimal step size and vanishing noise avoids all the strict saddle points and converges to the global optima, asymptotically. However, this asymptotic result does not provide any finite time guarantee. We next present the non-asymptotic convergence analysis of our proposed Perturbed GD algorithm. Our analysis shows that Perturbed GD has implicit bias determined by the injected noise. Specifically, we consider balanced noise, i.e., $\EE\left[\norm{\xi_1}_2^2\right]=\EE\left[\norm{\xi_2}_2^2\right]$, and show that Perturbed GD converges to the balanced $\pm(u_*,v_*)$ in polynomial time in Theorem \ref{thm_main}. \begin{theorem}[\bf Convergence Analysis]\label{thm_main} Suppose $x_0\in \RR^{d_1},$ $y_0\in \RR^{d_2}.$ For any $\epsilon>0$ and for any $\delta\in(0,1),$ we take $\sigma^2=\EE\left[\norm{\xi_1}_2^2\right]=\EE\left[\norm{\xi_2}_2^2\right]=\rm{poly}(\epsilon,(\log(1/\delta))^{-1}),$ $$\eta={\rm{poly}}( \sigma,\epsilon, (d_1+d_2)^{-1},(\log{(d_1+d_2)})^{-1},(\log(1/\delta))^{-1}).$$ With probability at least $1-\delta,$ we have $\norm{x_t-u_*}_2\leq \epsilon$ and $\norm{y_t-v_*}_2\leq \epsilon.$ for all $t_1\leq t\leq T_1= O\left({1}/{\eta^2}\right),$ where $t_1=O\left(\eta^{-1} \sigma^{-2}\log(1/\eta )\log(1/\delta)\right).$ \end{theorem} Theorem \ref{thm_main} differs from the convergence analysis of GD in \citet{du2018algorithmic} in the following aspects: (i) Perturbed GD converges regardless of initializations, while GD requires a random initialization near $(0,0)$; (ii) Perturbed GD guarantees convergence with high probability, while GD converges with only constant probability over randomness of initializations; (iii) Perturbed GD converges to the balanced solutions ($\norm{x}_2/\norm{y}_2=1$) while GD only maintains approximate balancedness, i.e., $c_0\leq\norm{x}_2/\norm{y}_2\leq C_0$ for some absolute constants $c_0, C_0>0.$ We provide a proof sketch that contains the essential ingredients of characterizing the convergence, since the proof of Theorem \ref{thm_main} is very technical and highly involved. See more details and the proof of technical lemmas in Appendix \ref{pf_2}. \begin{proof}[Proof Sketch] The convergence of Perturbed GD consists of three phases: {\bf Phase I}: Regardless of initialization, within polynomial time the noise encourages Perturbed GD to escape from region with undesired landscape (e.g., strict saddle points) and enter the region with benign landscape. {\bf Phase II}: Perturbed GD drives the loss to zero and approaches the set of global minima $\left\{(x,y)\big|xy^\top=M\right\}.$ {\bf Phase III}: After the loss is sufficiently small, the injected noise dominates the update, which helps the Perturbed GD to balance $x$ and $y,$ and converge to a balanced optimum. Before we proceed with our proof, we first define some notations. Let $U$ and $V$ denote the linear span of $u_*$ and $v_*$, respectively, i.e., $ \cU=\{\alpha_1 u_*:\alpha_1\in\RR\},~~\cV=\{\alpha_1 v_*:\alpha_1\in\RR\}. $ The corresponding orthogonal complement of $\cU$ (or $\cV$) in $\RR^{d_1}$ (or $\RR^{d_2}$) is denoted as $\cU^\perp$ (or $\cV^\perp$). Our analysis considers the convergence of Perturbed GD in $(\cU,\cV)$ and $(\cU^\perp,\cV^\perp),$ respectively. Specifically, we take the following orthogonal decomposition of $x_t$ and $y_t$: \begin{align*} x_t&=u_*^\top x_t u_*+(x_t-u_*^\top x_t u_*),\\y_t&=v_*^\top y_t v_*+(y_t-v_*^\top y_t v_*). \end{align*} One can check that $(x_t-u_*^\top x_t u_*)\in\cU^\perp$ and $(y_t-v_*^\top y_t v_*)\in \cV^\perp,$ respectively. \noindent $\bullet$ {\bf Phase I}: Regardless of initializations, the following lemma shows that $(x_t-u_*^\top x_t u_*)$ and $(y_t-v_*^\top y_t v_*)$ vanish after polynomial time. \begin{lemma}\label{lem_b_converge} Suppose $\norm{x_t}^2_2+\norm{y_t}^2_2\leq 2/\sigma^2$ holds for all $t>0.$ For any $\delta\in(0,1)$ we take $$\eta\leq \eta_2=C_4{\sigma^8}\left(\log((d_1+d_2)/\delta)\log(1/\delta)\right)^{-1},$$ where $C_4$ is some positive constant. Then with probability at least $1-\delta,$ we have \begin{align*} \norm{x_t-u_*^\top x_t u_*}_2^2&\leq 2\eta C_2\sigma^{-2},\\ \norm{y_t-v_*^\top y_t v_*}_2^2&\leq 2\eta C_2\sigma^{-2} \end{align*} for any $\tau_1\leq t\leq T_1= O(1/\eta^2),$ where $\tau_1=O(\eta^{-1}\sigma^{-2}\log(1/\eta)\log(1/\delta))$ and $C_2=(\sigma^2+1/\sigma^2)(2/\sigma^4+6/d_1+6/d_2+6\sigma^4).$ \end{lemma} Lemma \ref{lem_b_converge} shows that with an arbitrary initialization, the projection of $x_t$ (and $y_t$) onto $\cU^\perp$ (and $\cV^\perp$) vanishes. Thus, we only need to characterize the algorithmic behavior of Perturbed GD in subspace $(\cU,\cV).$ However, as the projection of $x_t$ (and $y_t$) onto $\cU^\perp$ (and $\cV^\perp$) vanishes, Perturbed GD can possibly approach the region with undesired landscape, e.g., the small neighborhood around the saddle point. The next lemma shows that the Perturbed GD will escape such regions within polynomial time. \begin{lemma}\label{lem_escape} Suppose $\norm{x_t-u_*^\top x_t u_*}_2^2\leq 2\eta C_2\sigma^{-2}$ and $\norm{y_t-v_*^\top y_t v_*}_2^2\leq 2\eta C_2\sigma^{-2}$ hold for all $t>0.$ For any $\delta\in(0,1),$ we take $\eta\leq\eta_3=C_3\sigma^{12}\left(\log((d_1+d_2)/\delta)\log(1/\delta)\right)^{-1},$ where $C_3$ is some positive constant. Then with probability at least $1-\delta,$ we have \begin{align}\label{eq_escape} x_{t}^\top u_* v_*^\top y_{t}=x_{t}^\top M y_{t}\geq 1/4, \end{align} for all $\tau_2\leq t\leq T_1=O\left({1}/{\eta^2}\right),$ where $\tau_2=O\left(\eta^{-1}\sigma^{-2}\log(1/\eta)\log(1/\delta)\right).$ \end{lemma} Since the saddle point $(x,y)=(0,0)$ satisfies $x^\top M y = 0,$ Lemmas \ref{lem_b_converge} and \ref{lem_escape} together imply that regardless of initializations, Perturbed GD is bounded away from the saddle point after polynomial time. The algorithm then enters Phase II and approaches the set of global optima $\left\{(x,y)\big|xy^\top=M\right\}.$ \noindent $\bullet$ {\bf Phase II}: In this phase, $(x_t,y_t)$ is still away from the region $\left\{(x,y)\big|xy^\top=M\right\}.$ Thus, $\nabla\cF(x_{t},y_{t})$ dominates the update of Perturbed GD. Perturbed GD behaves similarly to gradient descent while driving the loss to zero. The next lemma formally characterizes this behavior, showing that $xy^ \top$ converges to $M.$ \begin{lemma}\label{lem_loss} Suppose $x_{t}^\top u_* v_*^\top y_{t}\geq\frac{1}{4}$ holds for all $t>0.$ For any $\epsilon>0$ and for any $\delta\in(0,1)$, we choose $\sigma\leq\sigma'_1=C_4\sqrt{\epsilon}$ and take $\eta\leq \eta_4={C}_5{\sigma^6}\left(\log((d_1+d_2)/\delta)\log(1/\delta)\right)^{-1},$ where $C_4,C_5$ are some positive constants. Then with probability at least $1-\delta,$ we have $$\norm{x_ty_t^\top-M}_{\rm{F}}\leq \epsilon,$$ for all $\tau_3\leq t\leq T_1= O\left({1}/{\eta^2}\right),$ where $\tau_3=O\left(\eta^{-1}\log\frac{1}{\sigma}\log(1/\delta)\right).$ \end{lemma} \noindent $\bullet$ {\bf Phase III}: After Phase II, Perturbed GD enters the region where $x_ty_t^\top\approx M.$ Thus, the noise will dominate the update of the Perturb GD. Recall in Lemma \ref{lem_condition_number}, we show that the Hessian of unbalanced optima has a large condition number, hence a small perturbation will significantly change the gradient of the objective. This implies that the unbalanced optima would be unstable against the noise, and Perturbed GD will escape from such optima and continue iterating towards the balanced optima. The next lemma shows that $x_t $ and $y_t $ converge to $u_*, v_*,$ respectively. \begin{lemma}\label{lem_convergence} For $\forall\epsilon>0$, suppose $\norm{x_ty_t^\top-M}_{\rm{F}}\leq \epsilon$ holds for all $t>0.$ For any $\delta\in(0,1)$, we choose $\sigma\leq\sigma'_2={C}_6\left(\log(1/\delta)\right)^{-1/3}$ and take $\eta\leq \eta_5=C_7\sigma^{10}\epsilon,$ where $C_6, C_7$ are some positive constants. Then with probability at least $1-\delta,$ we have \begin{align*} \norm{x_t -u_*}_2\leq \epsilon,~\norm{y_t-v_*}_2\leq \epsilon, \end{align*} for all $\tau_3\leq t\leq T_1= O\left({1}/{\eta^2}\right),$ where $\tau_4=O\left(\eta^{-1}\sigma^{-2}\log\eta^{-1}\log(1/\delta)\right).$ \end{lemma} With all the lemmas in place, we take $\sigma\leq \min\{\sigma'_1,\sigma'_2\},$ $\eta\leq \min\{\eta_1,\eta_2,\eta_3,\eta_4,\eta_5\}$ and $t_1=\tau_1+\tau_2+\tau_3+\tau_4$, then the claim of Theorem \ref{thm_main} follows immediately. \end{proof} \vspace{-0.12in} \section{Extension to Rank-r Matrix Factorization}\label{sec_extend} \vspace{-0.1in} We extend our results to the rank-$r$ matrix factorization, which solves the following problem: \begin{align}\label{rankr_mf} \min_{X \in \RR^{d_1 \times r},Y \in \RR^{d_2 \times r}} \cF(X,Y) = \frac{1}{2}\norm{XY^\top - M}_{\rm{F}}^2, \end{align} where $M\in \RR^{d_1\times d_2}$ is a rank-$r$ matrix. Let $M=A\Sigma B^\top$ be the SVD of $M.$ Let $U_*=A\Sigma^{\frac{1}{2}}$ and $V_*=B\Sigma^{\frac{1}{2}},$ then $(U_*, V_*)$ is a global minimum for problem \eqref{rankr_mf}. Similar to the rank-1 case, using Perturbed GD to solve problem \eqref{rankr_mf} can be viewed as solving the following smoothed problem: \begin{align}\label{rankr_eq} \min_{X,Y}\tilde\cF(X,Y)=\EE_{\xi_{1},\xi_{2}}\cF(X+\xi_1,Y+\xi_2), \end{align} where $\xi_{1}\in \RR^{d_1\times r},$ $\xi_{2}\in \RR^{d_2\times r}$ have i.i.d. elements drawn from $N(0,\sigma_1^2)$ and $N(0,\sigma_2^2),$ respectively. The next theorem shows that the noise in Perturbed GD effectively addresses the scaling invariance issue of \eqref{rankr_mf}. The smoothed problem \eqref{rankr_eq} only has balanced global minima. \begin{theorem}\label{thm_rankr} Let $\sigma_{\min}(M)$ be the smallest singular value of $M$. Suppose $$\EE\left[\norm{\xi_1}_{\rm{F}}^2\right]=\gamma^2\EE\left[\norm{\xi_2}_{\rm{F}}^2\right]=r\gamma^2\sigma^2,$$ and $\gamma\sigma^2<\sigma_{\min}(M).$ Then for $\forall(U,V)\in (\RR^{d_1},\RR^{d_2})$ such that $\nabla\tilde\cF(U,V)=0,$ we have $U^\top U=\gamma^2 V^\top V.$ Moreover, denote $(\tilde U,\tilde V)=\left(\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}},\gamma^{-1/2}B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}\right),$ then the set $\{(\tilde U,\tilde V)R \big|R \in \RR^{r\times r}, RR^\top=R^\top R=I_r \}$ contains all the global optima. All other stationary points are strict saddles, i.e., $\lambda_{\min}\left(\nabla^2 \tilde\cF(U,V)\right)<0.$ \end{theorem} Theorem \ref{thm_rankr} shows that when noise is balanced, \eqref{rankr_eq} only has balanced global optima and strict saddle points. We can invoke \citet{lee2019first} again and show Perturbed GD with an infinitesimal step size and vanishing noise converges to the balanced global optima, asymptotically. Note that compared to the rank-1 case, in addition to scaling invariance, the objective is also rotation invariant, i.e., $\tilde\cF(X,Y)= \tilde\cF(XR,YR),$ where $R\in \RR^{r\times r}$ is an orthogonal matrix. Thus, we can only recover $U_*$ and $V_*$ up to a rotation factor. To establish the non-asymptotic convergence result, the optimization error should be measured by the following metric that is rotation invariant: \begin{align*} \mathrm{dist}_\cR(D_1,D_2)=\min_{R\in \RR^{r\times r}:RR^\top=I_r}\norm{D_1-D_2R}_{\rm{F}}. \end{align*} However, it is more challenging and involved to handle the complex nature of the distance $\mathrm{dist}_\cR(\cdot, \cdot)$. As our results for the rank-1 case already provide insights on understanding the implicit bias of noise, we leave the non-asymptotic analysis of rank-r matrix factorization for future investigation. \section{Preliminaries} We first introduce some important notions and results which can be used in the following proof. Assuming $\{u_*,\tilde{u}_1,\dots,\tilde{u}_{d_1-1}\}$ and $\{v_*,\tilde{v}_1,\dots,\tilde{v}_{d_2-1}\}$ are two sets of standard orthogonal basis of $\RR^{d_1}$ and $\RR^{d_2}$ respectively, we can then rewrite $\forall x\in\RR^{d_1}$ and $\forall y\in\RR^{d_2}$ as \begin{align*} x\triangleq \alpha_1u_*+\sum_{i=1}^{d _1-1} \beta_{1}^{(i)}\tilde{u}_i,\\ y\triangleq \alpha_2v_*+\sum_{j=1}^{d _2-1} \beta_2^{(j)}\tilde{v}_j, \end{align*} where $\alpha_1=x^\top u_*,$ $\alpha_2=y^\top v_*,$ $\beta_{1}^{(i)}=x^\top \tilde{u}_i$ and $\beta_{2}^{(j)}=y^\top \tilde{v}_j, ~\forall 0\leq i\leq d_1-1, 0\leq j\leq d_2-1.$ For simplicity, we denote $\beta_k=(\beta_k^{(1)},...,\beta_k^{(d_k-1)})^\top$ where $k=1,2.$ With the notions above, we can rewrite the Perturbed GD update as \begin{align*} \alpha_{1,t+1}=\alpha_{1,t}-\eta<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>,\\\alpha_{2,t+1}=\alpha_{2,t}-\eta<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>. \end{align*} Note that the optimal solutions to \eqref{mat_fct_ncvx} satisfy $x^\top My=1$. Thus, in our proof, we need to characterize the update of $x^\top My,$ which can be re-expressed as \begin{align} x_{t+1}^\top My _{t+1}&=\alpha_{1,t+1}\alpha_{2,t+1}\nonumber\\&=(\alpha_{1,t}-\eta<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>)(\alpha_{2,t}-\eta<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>)\nonumber\\&=\alpha_{1,t}\alpha_{2,t}-\eta\left(\alpha_{2,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>+\alpha_{1,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>\right)\nonumber\\&\quad\quad\quad\quad\quad\quad\quad\quad+\eta^2<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*><\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>. \end{align} For simplicity, we denote \begin{align*} &A_t\triangleq \alpha_{2,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>+\alpha_{1,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>,\\&B_t\triangleq <\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*><\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>. \end{align*} Then the update of $x^\top My=\alpha_1 \alpha_2$ can be expressed in a more compact way as follows. \begin{align}\label{*1} \alpha_{1,t+1}\alpha_{2,t+1}=\alpha_{1,t}\alpha_{2,t}-\eta A_t+\eta^2B_t. \end{align} Similarly, the update of $(x^\top My-1)^2$ can be re-expressed as \begin{align}\label{*2} (x_{t+1}^\top My_{t+1}-1)^2&=(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2 \nonumber\\&=(\alpha_{1,t}\alpha_{2,t}-1-\eta A_t+\eta^2B_t)^2 \nonumber\\&=(\alpha_{1,t}\alpha_{2,t}-1)^2+\eta^2 A_t^2+\eta^4 B_t^2\nonumber\\&~~~~~~~~~~~~-2\eta A_t(\alpha_{1,t}\alpha_{2,t}-1)-2\eta^3A_tB_t+2\eta^2B_t(\alpha_{1,t}\alpha_{2,t}-1). \end{align} Furthermore, since the balanced optima satisfy $x^\top u_*=y^\top v_*,$ we further explicitly write down the update of $\left((x^\top u_*)^2-(y^\top v_*)^2\right)^2$ as follows. \begin{align}\label{*3} \left((x_{t+1}^\top u_*)^2-(y_{t+1}^\top v_*)^2\right)^2&=(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2\nonumber\\ &=\left(\alpha_{1,t}^2-\alpha_{2,t}^2\right)^2+4\eta^2 D_t^2+\eta^4 F_t^2\nonumber\\&~~~~~~~~~~~~-4\eta D_t\left(\alpha_{1,t}^2-\alpha_{2,t}^2\right)-4\eta^3D_t F_t+2\eta^2F_t\left(\alpha_{1,t}^2-\alpha_{2,t}^2\right), \end{align} where $D_t$ and $F_t$ is defined as \begin{align*} &D_t\triangleq \alpha_{1,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>-\alpha_{2,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>,\\&F_t\triangleq <\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>^2-<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>^2. \end{align*} Next we are going to calculate $<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>$, the gradient projection along the direction of the optimum $u_*$. \begin{align}\label{*4} &<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>\nonumber\\&=u_*^\top\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}) \nonumber\\&=u_*^\top\left((x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-u_*v_*^\top\right)(y_{t}+\xi_{2,t}) \nonumber\\&=\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+2\alpha_{1,t}y_{t}^\top \xi_{2,t}-v_*^\top \xi_{2,t}+u_*^\top \xi_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)\nonumber\\&~~~~~~~~~~~~+2u_*^\top\xi_{1,t}y_{t}^\top \xi_{2,t}+\alpha_{1,t}\norm{\xi_{2,t}}^2+u_*^\top \xi_{1,t}\norm{\xi_{2,t}}^2 \nonumber\\&\triangleq \alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x, \end{align} where $g_x=2\alpha_{1,t}y_{t}^\top \xi_{2,t}-v_*^\top \xi_{2,t}+u_*^\top \xi_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)+2u_*^\top\xi_{1,t}y_{t}^\top \xi_{2,t}+\alpha_{1,t}\norm{\xi_{2,t}}^2+u_*^\top \xi_{1,t}\norm{\xi_{2,t}}^2.$ Similarly, we have \begin{align}\label{*5} &<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*> \nonumber\\&=\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+2\alpha_{2,t}x_{t}^\top \xi_{1,t}-u_*^\top \xi_{1,t}+v_*^\top \xi_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)\nonumber\\&~~~~~~~~~~~~+2v_*^\top\xi_{2,t}x_{t}^\top \xi_{1,t}+\alpha_{2,t}\norm{\xi_{1,t}}^2+v_*^\top \xi_{2,t}\norm{\xi_{1,t}}^2 \nonumber\\&\triangleq \alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y, \end{align} where $g_y=\alpha_{1,t}+2\alpha_{2,t}x_{t}^\top \xi_{1,t}-u_*^\top \xi_{1,t}+v_*^\top \xi_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)+2v_*^\top\xi_{2,t}x_{t}^\top \xi_{1,t}+\alpha_{2,t}\norm{\xi_{1,t}}^2+v_*^\top \xi_{2,t}\norm{\xi_{1,t}}^2.$ \section{Proof of Lemma \ref{lem_landscape}, Lemma \ref{lem_condition_number} , and Lemma \ref{lem_convolutional}}\label{pf_1} \subsection{Proof of Lemma \ref{lem_landscape}} \begin{proof} By setting the gradient of $ \cF$ to zero we get \begin{align} &\norm{x}_2^2y=M^\top x,\label{zero_grad_1}\\ &\norm{y}_2^2x=My.\label{zero_grad_2} \end{align} Recall that $M=u_* v_*^\top$ and \begin{align*} x\triangleq \alpha_1u_*+\sum_{i=1}^{d _1-1} \beta_{1}^{(i)}\tilde{u}_i,\\ y\triangleq \alpha_2v_*+\sum_{j=1}^{d _2-1} \beta_2^{(j)}\tilde{v}_j. \end{align*} Substitute $x$ and $y$ in \eqref{zero_grad_1} and \eqref{zero_grad_2} by their expansion, we then have \begin{align*} (\alpha_1^2+\sum_{i=1}^{d_1-1} (\beta_{1}^{(i)})^2)\alpha_2&=\alpha_1,\\ (\alpha_2^2+\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2)\alpha_1&=\alpha_2,\\ (\alpha_1^2+\sum_{i=1}^{d_1-1} (\beta_{1}^{(i)})^2)\beta_{2}^{(j)}&=0, \forall j=1,...,d_2-1,\\ (\alpha_2^2+\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2)\beta_{1}^{(j)}&=0, \forall j=1,...,d_1-1. \end{align*} The above equalities yield the following two types of stationary points. \begin{itemize} \item $(\alpha_1^2+\sum_{i=1}^{d_1-1} (\beta_{1}^{(i)})^2)(\alpha_2^2+\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2)\neq 0,\beta_1=0,\beta_2=0.$ This leads to $\alpha_1\alpha_2=1.$ Thus, $xy^\top=\alpha_1\alpha_2 u_*v_*^\top=M,$ and $\cF(\alpha_1u_*,\alpha_2v_*)=0.$ Then we have global optima $\left(\alpha u_*, \frac{1}{\alpha} v_*\right)$ for $\alpha\neq 0.$ \item Either $(\alpha_1^2+\sum_{i=1}^{d_1-1} (\beta_{1}^{(i)})^2)=0$ and $\alpha_2=0,$ or $(\alpha_2^2+\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2)=0$ and $\alpha_1=0.$ We next show that stationary points satisfy these conditions are strict saddle points. We only consider the first case, and the second case can be proved following similar lines. We first calculate the Hessian matrix as follows. \begin{align*} \nabla^2 \cF(x, y)=\begin{pmatrix} \norm{y}_2^2I_{d_1} & 2xy^\top-M \\ 2yx^\top-M^\top & \norm{x}_2^2I_{d_2} \end{pmatrix}. \end{align*} At $x=0$, $y=\sum_{j=1}^{d_2-1} \beta_2^{(j)}\tilde{v}_j,$ \begin{align*} \nabla^2 \cF(x, y)=\begin{pmatrix} \sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2I_{d_1} & -M \\ -M^\top & 0 \end{pmatrix}. \end{align*} For any $a\in \RR^{d_1},b\in \RR^{d_2},$ \begin{align*} \begin{pmatrix} a^\top, b^\top \end{pmatrix}\nabla^2 \cF(x, y)\begin{pmatrix} a \\ b \end{pmatrix}=\sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2\norm{a}_2^2-2a^\top M b \end{align*} For $(a, b)=(\tilde{u}_i, \tilde{v}_j)$, this quantity is positive. For $(a, b)=(u_*, \sum_{i=1}^{d_2-1} (\beta_{2}^{(i)})^2v_*),$ this quantity is negative. Thus, $x=0$, $y=\sum_{j=1}^{d_2-1} \beta_2^{(j)}\tilde{v}_j$satisfies strict saddle property. We conclude that for any $x\in \RR^{d_1},y\in\RR^{d_2}$ such that $x^\top u_*=y^\top v_*=0,$ we have strict saddle points $(x,0)$ and $(0,y).$ \end{itemize} \end{proof} \subsection{Proof of Lemma \ref{lem_condition_number}} At $x=\alpha u_*, y=\frac{1}{\alpha}v_*,$ \begin{align*} \nabla^2 \cF(\alpha u_*, \frac{1}{\alpha}v_*)=\begin{pmatrix} \frac{1}{\alpha^2}I_{d_1} &M \\ M^\top & \alpha^2I_{d_2} \end{pmatrix}. \end{align*} One can verify that $\nabla^2 \cF\left(\alpha u_*,\frac{1}{\alpha} v_*\right)$ has eigenvalues $\alpha^2+\frac{1}{\alpha^2},\alpha^2, \frac{1}{\alpha^2}.$ The largest eigenvalue is $\lambda_1=\alpha^2+\frac{1}{\alpha^2}$ and the smallest eigenvalue is $\lambda_{d_1+d_2}=\min\{\alpha^2,\frac{1}{\alpha^2}\}.$ Thus, the condition number can be easily calculated as follows. $$\kappa\left(\nabla^2 \cF\left(\alpha u_*,\frac{1}{\alpha} v_*\right)\right)=\max\{\alpha^4,\frac{1}{\alpha^4}\}+1.$$ \subsection{Proof of Lemma \ref{lem_convolutional}} \begin{proof} Recall that $M=u_* v_*^\top$ and \begin{align*} x\triangleq \alpha_1u_*+\sum_{i=1}^{d _1-1} \beta_{1}^{(i)}\tilde{u}_i,\\ y\triangleq \alpha_2v_*+\sum_{j=1}^{d _2-1} \beta_2^{(j)}\tilde{v}_j. \end{align*} By setting the gradient of $\tilde \cF$ to zero we get \begin{align*} &(\norm{x}_2^2+d_1 \sigma_1^2)y=M^\top x,\\ &(\norm{y}_2^2+d_2 \sigma_2^2)x=My. \end{align*} From the equations above, we can verify that $(x,y)=(0,0)$ is a stationary point. Furthermore, left multiplying the equations above by $\tilde{v}_j^\top$ and $\tilde{u}_i^\top$ respectively, we will get that $\beta_{1}^{(i)}$'s and $\beta_{2}^{(j)}$'s are all zeros. Similarly, by left multiplying the equations above by $v_*^\top$ and $u_*^\top$ respectively, we will get\begin{align*} &(\alpha_1^2+d_1 \sigma_1^2)\alpha_2^2=\alpha_2\alpha_1,\\ &(\alpha_2^2+d_2 \sigma_2^2)\alpha_1^2=\alpha_1\alpha_2. \end{align*} Then, with some algebraic manipulations, we get \begin{align*} &\alpha_1^2=\sqrt{\frac{d_1 \sigma_1^2}{d_2 \sigma_2^2}} - d_1 \sigma_1^2,\\ &\alpha_2^2=\sqrt{\frac{d_2 \sigma_2^2}{d_1 \sigma_1^2}} - d_2 \sigma_2^2. \end{align*} Specifically, when $d_1 \sigma_1^2=\gamma^2 d_2 \sigma_2^2\leq\gamma$, we have \begin{align*} &\alpha_1=\gamma \alpha_2= \pm \sqrt{\gamma- \gamma^2 d_2 \sigma_2^2}. \end{align*} Next, we are going to show that $(0, 0)$ is a strict saddle point and $(x_*, y_*)\triangleq\pm(\alpha_1u_*, \alpha_2v_*)$ are global optima. We first calculate the Hessian matrix as follows. \begin{align*} \nabla^2 \tilde\cF(x, y)=\begin{pmatrix} \left(\norm{y}_2^2+d_2\sigma_2^2\right)I_{d_1} & 2xy^\top-M \\ 2yx^\top-M^\top & \left(\norm{x}_2^2+d_1\sigma_1^2\right)I_{d_2} \end{pmatrix}. \end{align*} Since the injected noise is small, we have $\alpha_1\alpha_2=1-\sqrt{d_1\sigma_1^2d_2\sigma_2^2}>0$. For any $a \in \RR^{d_1}$ and $b \in \RR^{d_2}$, we have\begin{align*} &\begin{pmatrix} a^\top, b^\top \end{pmatrix}\nabla^2 \tilde\cF(x_*, y_*)\begin{pmatrix} a \\ b \end{pmatrix}\\ =&(\alpha_2^2+d_2\sigma_2^2)\norm{a}_2^2+(\alpha_1^2+d_1\sigma_1^2)\norm{b}_2^2+2(2\alpha_1\alpha_2-1)(a^\top u_*) (b^\top v_*)\\ \geq&(\alpha_2^2+d_2\sigma_2^2)\norm{a}_2^2+(\alpha_1^2+d_1\sigma_1^2)\norm{b}_2^2-2|2(1-\sqrt{d_1\sigma_1^2d_2\sigma_2^2})-1|\ \norm{a}_2\ \norm{b}_2\\ >&\sqrt{\frac{d_2 \sigma_2^2}{d_1 \sigma_1^2}}\norm{a}_2^2+\sqrt{\frac{d_1 \sigma_1^2}{d_2 \sigma_2^2}}\norm{b}_2^2-2\norm{a}_2\ \norm{b}_2, \end{align*}where the last inequality comes from the fact that $d_1\sigma_1^2$ and $d_2\sigma_2^2$ should be small enough such that $$0<1-\sqrt{d_1\sigma_1^2d_2\sigma_2^2}<1.$$Note that $$\sqrt{\frac{d_2 \sigma_2^2}{d_1 \sigma_1^2}}\norm{a}_2^2+\sqrt{\frac{d_1 \sigma_1^2}{d_2 \sigma_2^2}}\norm{b}_2^2-2\norm{a}_2\ \norm{b}_2=\left((\frac{d_2 \sigma_2^2}{d_1 \sigma_1^2})^{\frac{1}{4}}\norm{a}_2-(\frac{d_1 \sigma_1^2}{d_2 \sigma_2^2})^{\frac{1}{4}}\norm{b}_2\right)^2.$$ Thus, as long as ${d_1\sigma_1^2\ d_2\sigma_2^2}<1$, we have $\nabla^2 \cF_{reg}(x_*, y_*)$ is positive definite (PD). Thus $(x_*, y_*)$ is global minimum and so is $(-x_*,- y_*)$. Similarly, for any $a \in \RR^{d_1}$ and $b \in \RR^{d_2}$, we have \begin{align*} &\begin{pmatrix} a^\top, b^\top \end{pmatrix}\nabla^2 \cF_{reg}(0, 0)\begin{pmatrix} a \\ b \end{pmatrix} =d_2\sigma_2^2\norm{a}_2^2+d_1\sigma_1^2\norm{b}_2^2-2(a^\top u_*) (b^\top v_*). \end{align*} It is easy to show that for $(a, b)=(u_*, v_*)$, the quantity is negative when noise is small enough. But for $(a, b)=(\tilde{u}_i, \tilde{v}_j)$, this quantity is positive. Thus, (0,0) is a strict saddle point. Here, we prove Lemma \ref{lem_convolutional}. \end{proof} \section{Proof of Theorem \ref{thm_main}}\label{pf_2} \subsection{Boundedness of Trajectory} We first show that the solution trajectory of Perturbed GD is bounded with high probability, which is a sufficient condition for our following convergence analysis. \begin{lemma}[Boundedness of Trajectories]\label{lem_bounded} Given $x_0\in\RR^{d_1},$ $y_0\in\RR^{d_2},$ we choose $\sigma_1, \sigma_2>0$ such that $\sigma^2=\EE\left[\norm{\xi_1}_2^2\right]=\EE\left[\norm{\xi_2}_2^2\right]$ and $\norm{x_0}^2_2+\norm{y_0}^2_2\leq 1/\sigma^2.$ For any $\delta\in(0,1)$, we take $$\eta\leq \eta_1= C_1{\sigma^6}(\log((d_1+d_2)/\delta)\log(1/\delta))^{-1},$$ for some positive constant $C_1.$ Then with probability at least $1-\delta,$ we have $\norm{x_t}^2_2+\norm{y_t}^2_2\leq 2/\sigma^2$ for any $t\leq T_1=O(1/\eta^2).$ \end{lemma} \begin{proof} We first define the event where the injected noise for both $x$ and $y$ is bounded for the first $t$ iterations. \begin{align} &\cA_t=\left\{\big|{\xi^{(i)}_{1,\tau}}\big|,\big|{\xi^{(j)}_{2,\tau}}\big|\leq \sigma\left(\sqrt{2\log((d_1+d_2)\eta^{-2})} +\sqrt{\log(1/\delta)}\right), \forall\tau\leq t, i=1,...,d_1,j=1,..,d_2 \right\}. \end{align} By the concentration result of the maximum of Gaussian distribution, we have $\PP(\cA_{1/\eta^2})\geq 1-\delta.$ Moreover, we use $\cH_t$ to denote the event where the first $t$ iterates $\{(x_\tau,y_\tau)\}_{\tau\leq t}$ is bounded, i.e., $ \cH_t=\left\{\norm{x_\tau}^2_2+\norm{y_\tau}^2_2\leq \frac{2}{\sigma_2},\forall\tau\leq t\right\}. $ Let $\cF_t=\sigma\{(x_\tau,y_\tau),\tau\leq t\}$ denote the $\sigma-$algebra generated by that past $t$ iterations. Under the event $\cH_t$ and $\cA_t,$ we have the following inequality on the conditional expectation of $\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2.$ \begin{align} &\EE\left[(\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2)\mathds{1}_{\cH_t\cap\cA_t}\big|\cF_t\right]\nonumber\\=&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)-4\eta\left(\norm{x_{t}}_2^2\norm{y_{t}}_2^2-x_t^\top M y_t\right)\right\}\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\norm{\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t})}_2^2+\norm{\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t})}_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ \leq&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)-4\eta\left(\norm{x_{t}}_2^2\norm{y_{t}}_2^2-\norm{x_{t}}_2\norm{y_{t}}_2\right)\right\}\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[ \big|\big|\norm{y_t+\xi_{2,t}}_2^2(x_t+\xi_{1,t})-M(y_t+\xi_{2,t})\big|\big|_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[ \big|\big|\norm{x_t+\xi_{1,t}}_2^2(y_t+\xi_{2,t})-M^\top(x_t+\xi_{1,t})\big|\big|_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ \leq&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)-4\eta\left(\norm{x_{t}}_2^2\norm{y_{t}}_2^2-\norm{x_{t}}_2\norm{y_{t}}_2\right)\right\}\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+2\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[ \norm{y_t+\xi_{2,t}}_2^4\norm{x_t+\xi_{1,t}}_2^2+\norm{M(y_t+\xi_{2,t})}_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &+2\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[ \norm{x_t+\xi_{1,t}}_2^4\norm{y_t+\xi_{2,t}}_2^2+\norm{M^\top(x_t+\xi_{1,t})}_2^2\right]\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ \leq&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)+\eta+\eta^2 C_2\right\}\mathds{1}_{\cH_t\cap\cA_t},\nonumber \end{align} where the second inequality comes from the fact that $x^2-x\geq-\frac{1}{4}$ for all $x\in\RR$ and $C_2=(\sigma^2+1/\sigma^2)(2/\sigma^4+6/d_1+6/d_2+6\sigma^4)$ in the last inequality. We take $\eta\leq 1/C_2,$ then we have \begin{align} \EE\left[(\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2-1/\sigma^2)\mathds{1}_{\cH_t\cap\cA_t}\big|\cF_t\right]&\leq(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2-1/\sigma^2\right)\mathds{1}_{\cH_t\cap\cA_t}\nonumber\\ &\leq(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2-1/\sigma^2\right)\mathds{1}_{\cH_{t-1}\cap\cA_{t-1}}\label{martingale1}. \end{align} If we denote $G_t=(1-2\eta\sigma^2)^{-t}(\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2-1/\sigma^2),$ $G_t\mathds{1}_{\cH_{t-1}\cap\cA_{t-1}}$ is then a super-martingale according to \eqref{martingale1}. We will apply Azuma's Inequality to prove the bound and before that we have to bound the difference between $G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}$ and $\EE[G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}\big|\cF_t].$ \begin{align*} d_{t+1}&=\big| G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}- \EE[G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}\big|\cF_t]\big|\\ &=(1-2\eta\sigma^2)^{-t-1}\Bigg|2\eta x_t^\top[(x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-M](y_{t}+\xi_{2,t}))\\&\hspace{+1.2in}-2\eta \EE_{\xi_{1,t},\xi_{2,t}}\left[x_t^\top[(x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-M](y_{t}+\xi_{2,t}))\right]\\ &\hspace{+1.2in}+2\eta y_t^\top[(x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-M]^\top(x_{t}+\xi_{1,t}))\\&\hspace{+1.2in}-2\eta \EE_{\xi_{1,t},\xi_{2,t}}\left[y_t^\top[(x_{t}+\xi_{1,t})(y_{t}+\xi_{2,t})^\top-M]^\top(x_{t}+\xi_{1,t}))\right]\\ &\hspace{+1.2in}+\eta^2\big|\big|\norm{y_t+\xi_{2,t}}_2^2(x_t+\xi_{1,t})-M(y_t+\xi_{2,t})\big|\big|_2^2\\ &\hspace{+1.2in}-\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\big|\big|\norm{y_t+\xi_{2,t}}_2^2(x_t+\xi_{1,t})-M(y_t+\xi_{2,t})\big|\big|_2^2\right]\\ &\hspace{+1.2in}+\eta^2\big|\big|\norm{x_t+\xi_{1,t}}_2^2(y_t+\xi_{2,t})-M^\top(x_t+\xi_{1,t})\big|\big|_2^2\\ &\hspace{+1.2in}-\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\big|\big|\norm{x_t+\xi_{1,t}}_2^2(y_t+\xi_{2,t})-M^\top(x_t+\xi_{1,t})\big|\big|_2^2\right]\Bigg|\mathds{1}_{\cH_{t}\cap\cA_{t}} \\&\leq {C_1}' (1-2\eta\sigma^2)^{-t-1}\eta \sigma^{-2}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right), \end{align*} where ${C_1}'$ is some positive constant. Denote $r_t=\sqrt{\sum_{i=1}^t d_i^2}.$ By Azuma's inequality, we have $$\PP\left(G_t\mathds{1}_{\cH_{t-1}\cap\cA_{t-1}}-G_0\geq O(1)r_t\left(\log\frac{1}{\eta^2\delta}\right)^{\frac{1}{2}}\right)\leq O(\eta^2\delta).$$ Then when with probability at least $1-O(\eta^2\delta),$ we have \begin{align*} (\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2)\mathds{1}_{\cH_{t}\cap\cA_{t}}&\leq 1/\sigma^2+(1-\eta \sigma^2)^t(\norm{x_0}^2_2+\norm{y_0}^2_2-1/\sigma^2)\\ &\hspace{+0.3in}+O(1)(1-\eta\sigma^2)^t r_t(\log\frac{1}{\eta^2\delta})^{\frac{1}{2}}\\ &\leq 1/\sigma^2+ 0+O\left(\sqrt{\eta} \sigma^{-2}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\left(\log\frac{1}{\eta^2\delta}\right)^{\frac{1}{2}}\\ &\leq 2/\sigma^2, \end{align*} when $\eta=O\left(\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right).$ In order to satisfy $\eta\leq1/C_2$ at the same time, we take $\eta\leq\eta_1=O\left({\sigma^6}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right)$ to make sure that all inequalities above hold. The above inequality shows that if $\cH_{t-1}\cap\cA_{t-1}$ holds, then $\cH_{t}\cap\cA_{t}$ holds with probability at least $1-O(\eta^2\delta).$ Hence with probability at least $1-\delta,$ we have $(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2)\mathds{1}_{\cA_{t-1}}\leq 2/\sigma^2$ for all $t\leq T_1=O(1/\eta^2).$ Recall that $$\PP(\cA_{1/\eta^2})\geq 1-\delta.$$ Thus, we have with probability at least $1-2\delta,$ $\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\leq 2/\sigma^2$ for all $t\leq T_1=O(1/\eta^2).$ By properly rescaling $\delta$, we prove Lemma \ref{lem_bounded}. \end{proof} \subsection{Proof of Lemma \ref{lem_b_converge}} \begin{proof} We only prove the convergence of $\norm{\beta_{1,t}}_2^2$ here. The proof of the convergence of $\norm{\beta_{2,t}}_2^2$ follows similar lines. For notational simplicity, we denote $\xi_{1,t}^{(-1)}=(\xi_{1,t}^{(2)},..., \xi_{1,t}^{(d_1)})^\top,\ \forall t\geq 0.$ We first bound the conditional expectation of $\norm{\beta_{1,t+1}}_2^2$ given $\cF_t$: \begin{align} \EE\left[\norm{\beta_{1,t+1}}_2^2\big|\cF_t\right]&=\norm{\beta_{1,t}}_2^2+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\norm{y_t+\xi_{2,t}}_2^4\norm{\beta_{1,t}+\xi_{1,t}^{(-1)}}_2^2\right]-2\eta\left(\norm{y_t}_2^2+\sigma^2\right)\norm{\beta_{1,t}}_2^2\nonumber\\ &\leq (1-2\eta\sigma^2)\norm{\beta_{1,t}}_2^2+\eta^2 C_2, \label{b_martingale} \end{align} where $C_2=(\sigma^2+1/\sigma^2)(2/\sigma^4+6/d_1+6/d_2+6\sigma^4).$ The bound on the second moment comes from the boundedness of $\norm{y_t}_2^2,$ which has been shown in Lemma \ref{lem_bounded}. Define $G_t=(1-2\eta\sigma^2)^{-t}\left(\norm{\beta_{1,t}}_2^2-\frac{\eta C_2}{2\sigma^2}\right),$ and $\cE_t=\left\{\forall \tau\leq t, \norm{\beta_{1,\tau}}_2^2\geq \frac{\eta C_2}{\sigma^2}\right\}.$ By \eqref{b_martingale}, we have $$\EE\left[G_{t+1}\mathds{1}_{\cE_t}\big|\cF_t\right]\leq G_t\mathds{1}_{\cE_t}\leq G_t\mathds{1}_{\cE_{t-1}} .$$ Hence, by Markov inequality we have \begin{align*} \PP(\cE_t)= \PP\left( \norm{\beta_{1,t}}_2^2\mathds{1}_{\cE_{t-1}}\geq \frac{\eta C_2}{\sigma^2}\right)&\leq\frac{\EE\left[ \norm{\beta_{1,t}}_2^2\mathds{1}_{\cE_{t-1}}\right]}{\frac{\eta C_2}{\sigma^2}}\\ &\leq\frac{(1-2\eta\sigma^2)^{t}(\norm{\beta_{1,0}}_2^2-\frac{\eta C_2}{2\sigma^2})+\frac{\eta C_2}{2\sigma^2}}{\frac{\eta C_2}{\sigma^2}}\\ &\leq (1-2\eta\sigma^2)^{t}\frac{2}{C_2\eta}+\frac{1}{2} \leq \frac{3}{4}, \end{align*} when $t\geq\frac{1}{2\eta\sigma^2}\log\frac{8}{C_2\eta}.$ We take $t=\frac{1}{\eta\sigma^2}\log\frac{8}{C_2\eta}$ to make sure the inequality above holds. Thus with probability at least $\frac{1}{4},$ there exists a $ \tau\leq\frac{1}{\eta\sigma^2}\log\frac{8}{C_2\eta}$, such that $ \norm{\beta_{1,\tau}}_2^2\leq \frac{\eta C_2}{\sigma^2}.$ Thus, with probability at least $1-\delta,$ we can find a $\tau$ such that $ \norm{\beta_{1,\tau}}_2^2\leq \frac{\eta C_2}{\sigma^2},$ where $$\tau\leq\tau_1=\frac{1}{\log (4/3)\eta\sigma^2}\log\frac{8}{C_2\eta}\log\frac{1}{\delta}=\frac{1}{\log (4/3)\eta\sigma^2}\left(\log\frac{8}{C_2}+\log\frac{1}{\eta}\right)\log\frac{1}{\delta}=O\left(\frac{1}{\eta\sigma^2}\log\frac{1}{\eta}\log\frac{1}{\delta}\right).$$ We next show that with probability at least $1-\delta$, for $\forall t\geq \tau_1,$ we have $ \norm{\beta_{1,t}}_2^2\leq 2\frac{\eta C_2}{\sigma^2}.$ This can be done following the similar lines to the proof of Lemma \ref{lem_bounded}. We first restart the counter of the time and assume $ \norm{\beta_{1,0}}_2^2\leq \frac{\eta C_2}{\sigma^2}.$ Denote $\cH_t=\left\{\forall \tau\leq t,\norm{\beta_{1,\tau}}_2^2\leq \frac{2\eta C_2}{\sigma^2}\right\}.$ Then by \eqref{b_martingale}, we have $$\EE\left[G_{t+1}\mathds{1}_{\cH_t\cap\cA_t}\big|\cF_t\right]\leq G_t\mathds{1}_{\cH_t\cap\cA_t}\leq G_t\mathds{1}_{\cH_{t-1}\cap\cA_{t-1}} .$$ The difference of $ G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}$ and $\EE[G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}\big|\cF_t]$ can be easily bounded as follows \begin{align*} D_{t+1}&=\big| G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}- \EE[G_{t+1}\mathds{1}_{\cH_{t}\cap\cA_{t}}\big|\cF_t]\big|={C_2}'\eta^2\sigma^{-3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right), \end{align*} where ${C_2}'$ is some constant. Then by applying Azuma's inequality and following the similar lines to the proof of Lemma \ref{lem_bounded}, we show that when $$\eta\leq {C_3}'{\sigma^8}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1},$$ with probability at least $1-\delta,$ we have \begin{align*} \norm{\beta_{1,t}}_2^2\leq 2\frac{\eta C_2}{\sigma^2}.\end{align*} for any $\tau_1\leq t\leq T_1= O(1/\eta^2).$ \end{proof} \subsection{Proof of Lemma \ref{lem_escape}} We prove this lemma in three steps. \noindent $\bullet$ {\bf Step 1:} The following lemma shows that after polynomial time, with high probability, the algorithm can move out of the $O(\eta)$ neighborhood of the saddle point. \begin{lemma}[Escaping from the Unique Saddle Point] Suppose $\norm{\beta_{1,t}}_2^2\leq 2\frac{\eta C_2}{\sigma^2}$ and $\norm{\beta_{2,t}}_2^2\leq 2\frac{\eta C_2}{\sigma^2}$ hold for all $t>0.$ For $\forall\delta\in(0,1),$ we take $$\eta=O\left(\sigma^{12}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right).$$ Then with probability at least $1-\delta,$ there exists a $\tau\leq{\tau}_2,$ such that $$x_{\tau}^\top M y_{\tau}\geq 9\frac{\eta C_2}{\sigma^4},$$ where ${\tau}_2=\frac{5}{\eta \sigma^2}\log \frac{4 \sigma^3}{\eta C_2}\log\frac{1}{\delta}.$ \end{lemma} \begin{proof} Let's define the event $\cH_t=\{x_{\tau}^\top M y_{\tau}\leq9\frac{\eta C_2}{\sigma^4}, ~\forall \tau\leq t \}.$ Following the proof of Lemma \ref{lem_bounded}, we can refine the bound on the conditional expectation of $\norm{x_{t+1}}^2_2+\norm{y_{t+1}}^2_2$ given $\cH_t.$ Specifically, we have \begin{align*} \EE\left[(\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2)\mathds{1}_{\cH_t}\big|\cF_t\right]=&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)-4\eta\left(\norm{x_{t}}_2^2\norm{y_{t}}_2^2-x_t^\top M y_t\right)\right\}\mathds{1}_{\cH_t}\nonumber\\ &+\eta^2 \EE_{\xi_{1,t},\xi_{2,t}}\left[\norm{\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t})}_2^2+\norm{\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t})}_2^2\right]\mathds{1}_{\cH_t}\nonumber\\ \leq&\left\{(1-2\eta\sigma^2)\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\right)+34C_2\eta^2 \sigma^{-4})\right\}\mathds{1}_{\cH_t}. \end{align*} Denote $G_t=(1-2\eta\sigma^2)^{-t}\left(\norm{x_{t}}_2^2+\norm{y_{t}}_2^2-17C_2\eta \sigma^{-6}\right).$ Then we have $$\EE[G_{t+1}\mathds{1}_{\cH_t}]\leq G_{t}\mathds{1}_{\cH_t}\leq G_{t}\mathds{1}_{\cH_{t-1}}.$$ Thus, by Markov inequality, we have \begin{align*} \PP( (\norm{x_{t}}_2^2+\norm{y_{t}}_2^2)\mathds{1}_{\cH_{t-1}}\geq 34C_2\eta \sigma^{-6} )&\leq \frac{(1-2\eta\sigma^2)^{t}(\norm{x_{0}}_2^2+\norm{y_{0}}_2^2-17C_2\eta \sigma^{-6})+17C_2\eta \sigma^{-6}}{34C_2\eta \sigma^{-6}}\\ &\leq \frac{3}{4}, \end{align*} when $t\geq \frac{1}{2\eta \sigma^2}\log \frac{4 \sigma^3}{\eta C_2}.$ Thus with probability at least $1-\delta,$ there exists a $\tau\leq{\tau}_2=\frac{5}{2\eta \sigma^2}\log \frac{4 \sigma^3}{\eta C_2}\log\frac{1}{\delta}=O(\frac{1}{\eta \sigma^2}\log \frac{1}{\eta }\log\frac{1}{\delta}),$ such that $$(\norm{x_{\tau}}_2^2+\norm{y_{{\tau}}}_2^2)\mathds{1}_{\cH_{\tau-1}}\leq 34C_2\eta \sigma^{-6}.$$ Following the exactly same proof of Lemma \ref{lem_bounded}, we can show that with probability at least $1-\delta,$ for all $\tau_2\leq t\leq T_1=O(\frac{1}{\eta^2}),$ we have \begin{align*} (\norm{x_{t}}_2^2+\norm{y_{t}}_2^2)\mathds{1}_{\cH_{t-1}}&\leq(17+34)C_2\eta \sigma^{-6}+O\left(\sqrt{\eta} \sigma^{-2}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\left(\log\frac{1}{\eta^2\delta}\right)^{\frac{1}{2}} \\&= {C_4}'\eta\sigma^{-12}+{C_5}'\sqrt{\eta} \sigma^{-2}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\left(\log\frac{1}{\eta^2\delta}\right)^{\frac{1}{2}}\\ &\leq {C_6}' \left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}+{C_7}' \sigma^{4}, \end{align*} where ${C_4}'$, ${C_5}'$, ${C_6}'$ and ${C_7}'$ are some positive constants, and when$$\eta= O\left({\sigma^{12}}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right).$$ We next show that for large enough $t,$ with high probability, $\cH_{t-1}$ does not hold. We prove by contradiction: if $\cH_{t-1}$ holds for all $t$, the solution trajectory stays in a small neighborhood around $0.$ If so, we can show that with constant probability, $|\alpha_{1,t}+\alpha_{2,t}|$ will explode to infinity, which is in contradiction with the boundedness. Here follows the detailed proof. Assuming $\cH_{t-1}$ holds for all $t\leq\tilde{\tau}_2=O(\frac{1}{\eta}\log\frac{1}{\eta\sigma}\log\frac{1}{\delta})$, by the analysis above we have $\norm{x_{t}}_2^2+\norm{y_{t}}_2^2\leq {C_6}' \left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}+{C_7}' \sigma^{4}$ holds for $\forall t\leq\tilde{\tau}_2$. Note that with at least some constant probability, \begin{align*}&\Big|\alpha_{1,t+1}+\alpha_{2,t+1}\Big|-\Big|\alpha_{1,t}+\alpha_{2,t}\Big|\\=&\Big|\alpha_{1,t}+\alpha_{2,t}-\eta\left(<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>+<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>\right)\Big|-\Big|\alpha_{1,t}+\alpha_{2,t}\Big| \\ \geq &{C_8}'\eta(\sigma_1+\sigma_2),\end{align*} where ${C_8}'$ is some positive constant. This means we can find a $\tau=O(\log\frac{1}{\delta})\leq\tilde{\tau}_2$, such that $\Big|\alpha_{1,\tau}+\alpha_{2,\tau}\Big|>{C_8}'\eta(\sigma_1+\sigma_2),$ with probability at least $1-\delta.$ We will use this point as our initialization for the following proof. We next give a lower bound on the conditional expectation of $|\alpha_{1,t+1}+\alpha_{2,t+1}|.$ \begin{align*} &\EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|\big|\cF_t\right]\\ \geq&\Big|\EE\left[(\alpha_{1,t+1}+\alpha_{2,t+1})\big|\cF_t\right]\Big|\\ =&\Big|\left(1+\eta(1-\sigma^2-\alpha_{1,t}\alpha_{2,t})\right)(\alpha_{1,t}+\alpha_{2,t})-\eta\left(\norm{\beta_{1,t}}_2^2\alpha_{2,t}+\norm{\beta_{2,t}}_2^2\alpha_{1,t}\right)\Big|\\ \geq& \left(1+\eta(1-\sigma^2-\alpha_{1,t}\alpha_{2,t})\right)\Big|\alpha_{1,t}+\alpha_{2,t}\Big|- 2{C_9}' \frac{\eta^{2.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\\ \geq&\left(1+\frac{1}{2}\eta\right)\Big|\alpha_{1,t}+\alpha_{2,t}\Big|- 2{C_9}'\frac{\eta^{2.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right), \end{align*} where ${C_9}'$ is some positive constant. This is equivalent to the following inequality: \begin{align*} &\EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|-4{C_9}'\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\Bigg|\cF_t\right]\\\geq &\left(1+\frac{1}{2}\eta\right)\left(\Big|\alpha_{1,t}+\alpha_{2,t}\Big|- 4{C_9}'\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\\ \geq &\left(1+\frac{1}{2}\eta\right)^{t+1}\left(\Big|\alpha_{1,0}+\alpha_{2,0}\Big|- 4{C_9}'\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\\ \geq &\left(1+\frac{1}{2}\eta\right)^{t+1}\left({C_8}'\eta(d_1+d_2)- 4{C_9}'\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{\sigma_1+\sigma_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)\right)\\ \geq &\ {C_{10}}' \eta(\sigma_1+\sigma_2)\exp{t\eta/2}\\ \geq & \ {C_{10}}' \eta(\sigma/\sqrt{d_1}+\sigma/\sqrt{d_2})\frac{1}{\eta\sigma} \\ = &\ {C_{10}}'(1/\sqrt{d_1}+1/\sqrt{d_2}) , \end{align*} where ${C_{10}}' $ is some positive constant and last inequality holds when $t\geq \frac{2}{\eta}\log\frac{1}{\eta\sigma}.$ Note that here we still take $\eta=O\left(\sigma^{12}\left(\log\frac{d_1+d_2}{\delta}\log\frac{1}{\delta}\right)^{-1}\right)$, which makes sure that $\frac{\eta^{1.2} C_2}{\sigma^3}\left(\left(\log\frac{d_1+d_2}{\eta}\right)^{\frac{1}{2}}+\left(\log\frac{1}{\delta}\right)^{\frac{1}{2}}\right)$ is small enough to make the fourth inequality hold. For fixed $d_1$ and $d_2$, if we let $\delta$ and $\sigma$ go to zero (which guarantees that $\eta$ goes to zero), we will have the conditional expectation of $|\alpha_{1,t+1}+\alpha_{2,t+1}|$ stays in a neighborhood around a positive constant. However, by our assumption, $\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2$ stays in a very small neighborhood around zero, which makes $\alpha_{1,t+1}^2+\alpha_{2,t+1}^2\leq\norm{x_{t+1}}_2^2+\norm{y_{t+1}}_2^2$ also very small. This implies that $|\alpha_{1,t+1}+\alpha_{2,t+1}|$ can be arbitrarily small as long as we make $\delta$ and $\sigma$ small enough, which is the desired contradiction. Thus, we know that, with probability at least $1-\delta,$ there exists a t satisfying ${\tau}_2\leq t\leq {\tau}_2+\tilde{\tau}_2\leq 2{\tau}_2,$ such that $\cH_{t}$ does not hold. Re-scale ${\tau}_2$ and we prove the result. \end{proof} \noindent $\bullet$ {\bf Step 2:} The following lemma shows that after step 1, with high probability, the algorithm will continue move away from the saddle point and escape from the saddle point at some time. \begin{lemma} \label{lem_conv_region} Suppose $x_0^\top M y_0\geq 9\frac{\eta C_2}{\sigma^4}.$ We take $\eta$ as in Lemma \ref{lem_escape}. Then there almost surely exists a $\tau\leq \tau_2'=\frac{2}{\eta}\log\frac{2\sigma^3}{\eta C_2},$ such that $$x_{\tau}^\top M y_{\tau}\geq\frac{1}{2}+\sigma^2.$$ \end{lemma} \begin{proof} Suppose $x_t^\top M y_t\leq\frac{1}{2}+\sigma^2$ holds for all $t\leq \tau_2'.$ We next give a lower bound on the conditional expectation of $|\alpha_{1,t+1}+\alpha_{2,t+1}|.$ \begin{align*} \EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|\big|\cF_t\right]&\geq\Big|\EE\left[(\alpha_{1,t+1}+\alpha_{2,t+1})\big|\cF_t\right]\Big|\\ &=\Big|\left(1+\eta(1-\sigma^2-\alpha_{1,t}\alpha_{2,t})\right)(\alpha_{1,t}+\alpha_{2,t})-\eta\left(\norm{\beta_{1,t}}_2^2\alpha_{2,t}+\norm{\beta_{2,t}}_2^2\alpha_{1,t}\right)\Big|\\ &\geq\left(1+\frac{1}{2}\eta\right)\Big|\alpha_{1,t}+\alpha_{2,t}\Big|- 4\frac{\eta^2 C_2}{\sigma^4}.\end{align*} This implies that \begin{align*} \EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|-8\frac{\eta C_2}{\sigma^4}\big|\cF_t\right]&\geq\left(1+\frac{1}{2}\eta\right)\left(\Big|\alpha_{1,t}+\alpha_{2,t}\Big|-8\frac{\eta C_2}{\sigma^4}\right).\end{align*} The above inequality further implies a lower bound on the conditional expectation of $|\alpha_{1,t+1}+\alpha_{2,t+1}|-8\frac{\eta C_2}{\sigma^4}.$ \begin{align*} \EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|-8\frac{\eta C_2}{\sigma^4}\right]&\geq\left(1+\frac{1}{2}\eta\right)^{t}\left(\Big|\alpha_{1,0}+\alpha_{2,0}\Big|-8\frac{\eta C_2}{\sigma^4}\right)\\ &\geq\left(1+\frac{1}{2}\eta\right)^{t} \frac{\eta C_2}{\sigma^4}\geq \frac{2}{\sigma},\end{align*} when $t= \tau_2'=\frac{2}{\eta}\log\frac{2\sigma^3}{\eta C_2}.$ On the other hand, $$\EE\left[|\alpha_{1,t+1}+\alpha_{2,t+1}|-8\frac{\eta C_2}{\sigma^4}\right]\leq \frac{\sqrt{2}}{\sigma},$$ which leads to a contradiction unless $$\PP(x_t^\top M y_t\leq\frac{1}{2}+\sigma^2,~\forall t\leq \tau_2')=0.$$ We prove the result. \end{proof} \noindent $\bullet$ {\bf Step 3:} We then show that the algorithm will never iterate back towards the saddle point after escaping from it. Then Lemma \ref{lem_escape} is proved. \begin{lemma}\label{prop1} Suppose there exists a time step $\tau$ such that for some positive constant $c<\frac{1}{2}$ \begin{align*} x_{\tau}^\top My_{\tau}>2c, \end{align*} then $\forall \delta \in (0,1)$, with probability at least $1-\delta$, $\forall \tau\leq t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| x^\top My>c\right\}. \end{align*} \end{lemma} By taking $c=\frac{1}{4}$, we can prove Lemma \ref{lem_escape} from Lemma \ref{prop1}. \begin{proof} We consider two cases: \\ (i) if $\alpha_{1,t}\alpha_{2,t}\geq2(1-c)$, then $\alpha_{1,t+1}\alpha_{2,t+1}>2c$ w.p.1.\\ (ii) if $\alpha_{1,t}\alpha_{2,t}<2(1-c)$, then, $\forall \delta \in (0,1)$ and $\forall t \leq T_1=O(\frac{1}{\eta^2})$, w.p. at least $1-\delta$, \begin{align*}(\alpha_{1,t},\alpha_{2,t})\in\{(\alpha_1,\alpha_2)|\left(\alpha_1\alpha_2-1\right)^2<(1-c)^2\}. \end{align*} Note that $(\alpha_{1,t}\alpha_{2,t}-1)^2<(1-c)^2$ implies that $\alpha_{1,t}\alpha_{2,t}>c$, thus we can prove Lemma \ref{prop1}. Further note that the injected noise is upper bounded with high probability, we assume in the following $T_1=O(\frac{1}{\eta^2})$ steps, \begin{align*} \norm{\xi_{k,t}}_\infty\leq\Bar{\sigma}\triangleq\sigma\left(\sqrt{2\log((d_1+d_2)\eta^{-2})} +\sqrt{\log(1/\delta)}\right). \end{align*} {\bf Case (i):} By plugging \eqref{*4} and \eqref{*5} to \eqref{*1}, we get \begin{align*} \alpha_{1,t+1}\alpha_{2,t+1}=\alpha_{1,t}\alpha_{2,t}-\eta \left(\alpha_{2,t} (\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x)+\alpha_{1,t}(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y)\right) \\+\eta^2(\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x)(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y)\\=2c+(\alpha_{1,t}\alpha_{2,t}-2c)-\eta [\alpha_{2,t} (\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x)+\alpha_{1,t}(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y) \\+\eta(\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+g_x)(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+g_y)], \end{align*} First note that by Lemma \ref{lem_conv_region} we can always find such $0<c<\frac{1}{2}$ in the condition. As $\alpha_{1,t}\alpha_{2,t}-2c>2(1-c)-2c=2(1-2c)>0$, we can prove (i) by choosing $\eta$ small enough. By Lemma \ref{lem_bounded}, $\alpha_{1,t}^2+\alpha_{2,t}^2$ is bounded by $\frac{2}{\sigma^2}$ and we can know $g_x$ and $g_y$ are at most of order $O\left(\frac{1}{\sigma}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)\right)$ from the following upper bound: \begin{align*} |g_x|\leq\Bar{\sigma}(\alpha_{2,t}^2+2\alpha_{2,t}\alpha_{1,t}+1+\frac{2\eta C_2}{\sigma^2})+\Bar{\sigma}^2(2\alpha_{2,t}+\alpha_{1,t})+\Bar{\sigma}^3,\\|g_y|\leq\Bar{\sigma}(\alpha_{1,t}^2+2\alpha_{1,t}\alpha_{2,t}+1+\frac{2\eta C_2}{\sigma^2})+\Bar{\sigma}^2(2\alpha_{1,t}+\alpha_{2,t})+\Bar{\sigma}^3. \end{align*} It is easy to show that for properly selected $$\eta\leq\tilde{\eta}_1=O\left(\left(\frac{1}{\sigma^4}+\frac{1}{\sigma^2}\left(\sqrt{\log(d_1+d_2)}+\sqrt{\log\frac{1}{\delta}}\right)\right)^{-1}\right),$$ the last two terms are greater than zero. Thus, w.p.1, $\alpha_{1,t+1}\alpha_{2,t+1}>2c$.\\ {\bf Case (ii):} Without loss of generality, we place the time origin at t, and thus we have $2c<\alpha_1^0\alpha_2^0<2-2c$. \begin{align*} \EE_\xi[A_t] &=\EE_\xi[\alpha_{2,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>+\alpha_{1,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>]. \end{align*} By plugging \eqref{*4} and \eqref{*5} in the above equation, we have \begin{align*} \EE_\xi[A_t] &=\alpha_{2,t}(\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+\alpha_{1,t}\sigma^2)+\alpha_{1,t}(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+\alpha_{2,t}\sigma^2) \\&=(\alpha_{2,t}\alpha_{1,t}-1)(\alpha_{2,t}^2+\alpha_{1,t}^2)+\alpha_{2,t}\alpha_{1,t}(2\sigma^2+\norm{\beta_{1,t}}_2^2+\norm{\beta_{2,t}}_2^2). \end{align*} Again, by plugging \eqref{*4} and \eqref{*5} to \eqref{*2} and taking expectation conditioning on $\cF_t$, when $\alpha_{1,t}\alpha_{2,t}>c$, we will get \begin{align*} \EE[(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2|\cF_t] \nonumber&=(\alpha_{1,t}\alpha_{2,t}-1)^2-2\eta \EE_\xi[A_t](\alpha_{1,t}\alpha_{2,t}-1)\\&+\eta^2 \EE_\xi[A_t^2]+\eta^4 \EE_\xi[B_t^2]-2\eta^3\EE_\xi[A_tB_t]+2\eta^2\EE_\xi[B_t](\alpha_{1,t}\alpha_{2,t}-1)\\ &=\left(1-2\eta (\alpha_{2,t}^2+\alpha_{1,t}^2)\right)(\alpha_{1,t}\alpha_{2,t}-1)^2\\&~~~~-2\eta \alpha_{2,t}\alpha_{1,t}(\alpha_{1,t}\alpha_{2,t}-1)(2\sigma^2+\norm{\beta_{1,t}}_2^2+\norm{\beta_{2,t}}_2^2)\\&~~~~+\eta^2 \EE_\xi[A_t^2]+\eta^4 \EE_\xi[B_t^2]-2\eta^3\EE_\xi[A_tB_t]+2\eta^2\EE_\xi[B_t](\alpha_{1,t}\alpha_{2,t}-1)\\ &\leq\left(1-4\eta c\right)(\alpha_{1,t}\alpha_{2,t}-1)^2+2\eta \frac{1}{4}(2\sigma^2+\frac{2\eta C_2}{\sigma^2}+\frac{2\eta C_2}{\sigma^2})+\tilde{C}_1\eta^2\\ &\leq\left(1-4\eta c\right)(\alpha_{1,t}\alpha_{2,t}-1)^2+\eta (\sigma^2+\frac{2\eta C_2}{\sigma^2})+\tilde{C}_1\eta^2, \end{align*} where $\tilde{C}_1=O\left(\frac{1}{\sigma^4}\left(\log\frac{d_1+d_2}{\delta}\right)\right),$ if $\alpha_{1,t}\alpha_{2,t}$ is of constant order. Note that here we choose $\eta\leq\tilde{\eta}_1$ as mentioned in (i). Denote $\gamma\triangleq\frac{\eta (\sigma^2+\frac{2\eta C_2}{\sigma^2})+\tilde{C}_1\eta^2}{4\eta c}$, the inequality above can be re-expressed as \begin{align} \label{*6} \EE[\{(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2-\gamma\}|\cF_t]\leq\left(1-4\eta c\right)\{(\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma\}. \end{align} We denote $G_t\triangleq(1-4\eta c)^{-t}\{(\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma\}$ and $\cE_t\triangleq\{\forall \tau\leq t : (\alpha_{1,\tau}\alpha_{2,\tau}-1)^2<(1-c)^2\}.$ Since $\alpha_{1,\tau}\alpha_{2,\tau}>c$ can be inferred from $(\alpha_{1,\tau}\alpha_{2,\tau}-1)^2<(1-c)^2$, we can get \begin{align*} \EE[G_{t+1}\mathds{1}_{\cE_t}|\cF_t]\leq G_{t}\mathds{1}_{\cE_t}\leq G_{t}\mathds{1}_{\cE_{t-1}}. \end{align*} This means $\{G_{t}\mathds{1}_{\cE_{t-1}}\}$ is a supermartingale. To use Azuma's inequality, we need to bound the following difference \begin{align*} d_{t+1}&\triangleq|G_{t+1}\mathds{1}_{\cE_t}-\EE[G_{t+1}\mathds{1}_{\cE_t}|\cF_t]|\\ &=(1-4\eta c)^{-t-1}|(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2-\EE[(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2|\cF_t]|\mathds{1}_{\cE_t}\\ &\leq(1-4\eta c)^{-t-1}\tilde{C}_2, \end{align*} where $\tilde{C}_2=O\left(\eta\frac{1}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)\right)$. Again, here we choose $\eta\leq\tilde{\eta}_1$. We further define $r_t\triangleq\sqrt{\sum_{i=1}^t d_i^2}$, and by Azuma's inequality we have \begin{align*} \mathbb{P}\left(G_{t} \mathds{1}_{\cE_{t-1}}-G_{0} \geq O(1) r_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) \leq \exp \left(-\frac{O(1) r_{t}^{2} \log \left(\frac{1}{\eta^{2} \delta}\right)}{2 \sum_{i=0}^{t} d_{i}^{2}}\right)=O\left(\eta^{2} \delta\right). \end{align*} Thus, with probability at least $1-O(\eta^2 \delta)$, \begin{align*} ((\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma) \mathds{1}_{\cE_{t-1}} &< (1-4\eta c)^t\left((\alpha_{1,0}\alpha_{2,0}-1)^2+O(1) r_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<(\alpha_{1,0}\alpha_{2,0}-1)^2+O\left( (1-4\eta c)^t r_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<(1-c)^2+(1-2c)^2-(1-c)^2\\&~~~~~~~~+O\left( \frac{\sqrt{\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)\log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right), \end{align*} When $\cE_{t-1}$ holds, w.p. at least $1-O(\eta^2 \delta)$, we will have \begin{align*} (\alpha_{1,t}\alpha_{2,t}-1)^2&<(1-c)^2+(1-2c)^2-(1-c)^2\\&~~~~+O\left( \frac{\sqrt{\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)\log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) +\gamma<(1-c)^2, \end{align*} as $(1-2c)^2-(1-c)^2$ is some negative constant, by choosing $\eta\leq\tilde{\eta}_2=O\left(\sigma^4(\log\frac{1}{\delta})^{-1}\left(\log\frac{d_1+d_2}{\delta}\right)^{-1}\right)$ and $\sigma$ small enough, we can make sure the sum of last four terms is negative. Now we know that if $\cE_{t-1}$ holds, $\cE_{t}$ holds w.p. at least $1-O(\eta^2 \delta)$. It is easy to show that the Perturbed GD updates satisfy $(\alpha_{1,t}\alpha_{2,t}-1)^2<(1-c)^2$ in the following $O(\frac{1}{\eta^2})$ steps w.p. at least $1-\delta$. \end{proof} \subsection{Proof of Lemma \ref{lem_loss}} We partition this lemma into two parts: Lemma \ref{lma11} shows that after polynomial time, the algorithm enters $\{ (x,y)\big|(x^\top My-1)^2<4\gamma\},$ where $\gamma$ is a small constant depending on $\sigma$, and Lemma \ref{lma22} shows that the algorithm then stays in $\{ (x,y)\big|(x^\top My-1)^2<6\gamma\}.$ It is easy to prove Lemma \ref{lem_loss} from Lemmas \ref{lma11} and \ref{lma22}. \begin{lemma} \label{lma11} Suppose $\forall t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| x^\top My>c\right\}, \end{align*}then $\forall \delta \in (0,1)$, with probability at least $1-\delta$, there exists a time step $\tau\leq \tau_3\triangleq O(\frac{1}{\eta}\log\frac{1}{\sigma}\log\frac{1}{\delta})$ such that \begin{align*} (x_{\tau}^\top My_{\tau}-1)^2<4\gamma, \end{align*} where {$\gamma=O(\sigma^2)$}. \end{lemma} \begin{lemma} \label{lma22} Suppose there exists a time step $\tau\leq \tau_3\triangleq O(\frac{1}{\eta}\log\frac{1}{\sigma}\log\frac{1}{\delta})$ such that \begin{align*} (x_{\tau}^\top My_{\tau}-1)^2<4\gamma, \end{align*} and $\forall t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| x^\top My>c\right\}, \end{align*}then $\forall \delta \in (0,1)$, with probability at least $1-\delta$,$\forall t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| (x^\top My-1)^2<6\gamma\right\}. \end{align*} \end{lemma} Here follows proof of Lemmas \ref{lma11} and \ref{lma22}. \begin{proof} Define $\cH_t\triangleq\{\forall \tau\leq t : (\alpha_{1,\tau}\alpha_{2,\tau}-1)^2\geq4\gamma\}$, for $t>\frac{\log(\frac{(\alpha_{1,0}\alpha_{2,0}-1)^2}{\gamma})}{4\eta c}$, we have \begin{align*} 4\gamma\EE[\mathds{1}_{\cH_t}]\leq\EE[(\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma]\leq(1-4\eta c)^t\left((\alpha_{1,0}\alpha_{2,0}-1)^2-\gamma\right)+\gamma<2\gamma, \end{align*} where the first inequality comes from the definition of $\cH_t$ and the second one comes from \eqref{*6}. Thus, if we choose $t=O\left(\frac{\log(\frac{1}{\gamma \sigma^2})}{\eta}\right)$ and recursively applying the inequality above $O(\log(\frac{1}{\delta}))$ times, we will get, for $\tau_3=O\left(\frac{1}{\eta}\log(\frac{1}{\delta})\log(\frac{1}{\gamma \sigma^2})\right)=O\left(\frac{1}{\eta}\log\frac{1}{\sigma}\log\frac{1}{\delta}\right)$, \begin{align*} \PP(\cH_{\tau_3})<(\frac{1}{2})^{\log(\frac{1}{\delta})}=\delta. \end{align*} Thus, w.p. at least $1-\delta$, there exists a $\tau\leq \tau_3$ s.t. $(\alpha_{1,\tau}\alpha_{2,\tau}-1)^2<4\gamma$. Here, we finish the proof of Lemma \ref{lma11}. Without loss of generality, we place the time origin at $\tau$, i.e. $(\alpha_{1,0}\alpha_{2,0}-1)^2<4\gamma$. We next prove Lemma \ref{lma22}. Denote $\cA\triangleq\{(\alpha_1,\alpha_2)|(\alpha_1\alpha_2-1)^2<6\gamma\}$ and $\cA_t\triangleq\{\forall \tau\leq t:(\alpha_{1,\tau},\alpha_{2,\tau})\in \cA\}$, again note that $\alpha_{1,t}\alpha_{2,t}>c$ can be inferred from $(\alpha_{1,t}\alpha_{2,t}-1)^2<6\gamma<(1-c)^2$. Thus, without any assumption, we can get \begin{align*} \EE[G_{t+1}\mathds{1}_{\cA_t}|\cF_t]\leq G_{t}\mathds{1}_{\cA_t}\leq G_{t}\mathds{1}_{\cA_{t-1}}. \end{align*} This means $\{G_{t}\mathds{1}_{\cA_{t-1}}\}$ is a supermartingale. To use Azuma's inequality, we need to bound the following difference \begin{align*} \tilde{d}_{t+1}&\triangleq|G_{t+1}\mathds{1}_{\cA_t}-\EE[G_{t+1}\mathds{1}_{\cA_t}|\cF_t]|\\ &=(1-4\eta c)^{-t-1}|(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2-\EE[(\alpha_{1,t+1}\alpha_{2,t+1}-1)^2|\cF_t]|\mathds{1}_{\cA_t}\\ &\leq(1-4\eta c)^{-t-1}\tilde{C}_3, \end{align*} where {$\tilde{C}_3=O(\sqrt{\gamma}\tilde{C}_2)$}. Further define $\tilde{r}_t\triangleq\sqrt{\sum_{i=1}^t \tilde{d_i}^2}$, by Azuma's inequality we have \begin{align*} \mathbb{P}\left(G_{t} \mathds{1}_{\cA_{t-1}}-G_{0} \geq O(1) \tilde{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) \leq \exp \left(-\frac{O(1) \tilde{r}_{t}^{2} \log \left(\frac{1}{\eta^{2} \delta}\right)}{2 \sum_{i=0}^{t} \tilde{d_{i}}^{2}}\right)=O\left(\eta^{2} \delta\right). \end{align*} Thus, with probability at least $1-O(\eta^2 \delta)$, \begin{align*} ((\alpha_{1,t}\alpha_{2,t}-1)^2-\gamma) \mathds{1}_{\cA_{t-1}} &< (1-4\eta c)^t\left((\alpha_{1,0}\alpha_{2,0}-1)^2+O(1) \tilde{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<(\alpha_{1,0}\alpha_{2,0}-1)^2+O\left( (1-4\eta c)^t \tilde{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<4\gamma+O\left( {\frac{\sqrt{\gamma\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)}\log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right), \end{align*} When $\cE_{t-1}$ holds, w.p. at least $1-O(\eta^2 \delta)$ \begin{align*} (\alpha_{1,t}\alpha_{2,t}-1)^2&<5\gamma+O\left( \frac{{\tilde{C}_3}}{\sqrt{\eta}} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) <6\gamma, \end{align*} by choosing {$\eta\leq\tilde{\eta}_3=O\left(\sigma^6(\log\frac{1}{\delta})^{-1}\left(\log\frac{d_1+d_2}{\delta}\right)^{-1}\right)$} we can guarantee the last term is smaller than $\gamma$. Now we know that if $\cA_{t-1}$ holds, $\cA_{t}$ holds w.p. at least $1-O(\eta^2 \delta)$. It is easy to show that the Perturbed GD updates satisfy $(\alpha_{1,t}\alpha_{2,t}-1)^2<6\gamma$ in the following $O(\frac{1}{\eta^2})$ steps w.p. at least $1-\delta$. \end{proof} \subsection{Proof of Lemma \ref{lem_convergence}} \begin{lemma} \label{a1=a2} Suppose $(x_t^\top My_t-1)^2<6\gamma$ holds for all t, where $\gamma$ is as defined above. For any $\delta \in (0,1)$ and any $\Delta>0$, if we choose $\sigma=O\left((\log\frac{1}{\delta})^{-\frac{1}{3}}\right)$ and take step size $$\eta\leq\tilde{\eta}_4=O\left(\sigma^{10}\Delta\right),$$ then with probability at least $1-\delta$, we have \begin{align*} (x_t,y_t) \in \left\{(x,y)| \left((x^\top u_*)^2-(y^\top v_*)^2\right)^2<6\Delta \right\}, \end{align*} for all t's such that $\tau_4\leq t \leq T_1$, where $T_1\triangleq O(\frac{1}{\eta^2})$ and $\tau_4\triangleq O(\frac{1}{\eta\sigma^2}\log\frac{1}{\eta}\log\frac{1}{\delta})$. \end{lemma} Again, we partition this lemma into two parts. It is easy to prove Lemma \ref{a1=a2} from Lemmas \ref{lma3} and \ref{lma4}. \begin{lemma} \label{lma3} Suppose $\forall t\leq T_1= O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| (x^\top My-1)^2<6\gamma\right\}. \end{align*}then $\forall \delta \in (0,1)$, with probability at least $1-\delta$, there exists a time step $\tau\leq \tau_4= O(\frac{1}{\eta\sigma^2}\log\frac{1}{\eta}\log\frac{1}{\delta})$ such that \begin{align*} \left((x_{\tau}^\top u_*)^2-(y_{\tau}^\top v_*)^2\right)^2<4\Delta, \end{align*} where {$\Delta=O(\frac{\eta}{\sigma^{10}})$}. \end{lemma} \begin{proof} \begin{align*} \EE_\xi[D_t]&=\EE_\xi[\alpha_{1,t}<\nabla_{x}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),u_*>-\alpha_{2,t}<\nabla_{y}\cF(x_{t}+\xi_{1,t},y_{t}+\xi_{2,t}),v_*>] \\&=\alpha_{1,t}(\alpha_{1,t}(\alpha_{2,t}^2+\norm{\beta_{2,t}}_2^2)-\alpha_{2,t}+\alpha_{1,t}\sigma^2)-\alpha_{2,t}(\alpha_{2,t}(\alpha_{1,t}^2+\norm{\beta_{1,t}}_2^2)-\alpha_{1,t}+\alpha_{2,t}\sigma^2) \\&=\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)+(\alpha_{1,t}^2\norm{\beta_{2,t}}_2^2-\alpha_{2,t}^2\norm{\beta_{1,t}}_2^2). \end{align*} Thus, we have \begin{align*} \EE_\xi[D_t](\alpha_{1,t}^2-\alpha_{2,t}^2)&=\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)^2+(\alpha_{1,t}^2-\alpha_{2,t}^2)(\alpha_{1,t}^2\norm{\beta_{2,t}}_2^2-\alpha_{2,t}^2\norm{\beta_{1,t}}_2^2) \\&\geq\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-2\alpha_{1,t}^2\alpha_{2,t}^2\frac{2\eta C_2}{\sigma^2} \\&>\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-2(1+\sqrt{6\gamma})^2\frac{2\eta C_2}{\sigma^2} \\&>\sigma^2(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-3\frac{2\eta C_2}{\sigma^2}. \end{align*} The last inequality can be achieved by choosing $\gamma<\frac{(\sqrt{3/2}-1)^2}{6}$. Plugging \eqref{*4} and \eqref{*5} into \eqref{*3}, taking expectation conditioning on previous trajectory $\cF_t$ and plugging the equation above in, we get \begin{align*} \EE[(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2|\cF_t] \nonumber&=(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-4\eta \EE_\xi[D_t](\alpha_{1,t}^2-\alpha_{2,t}^2)\\&+4\eta^2 \EE_\xi[D_t^2]+\eta^4 \EE_\xi[F_t^2]-4\eta^3\EE_\xi[D_t F_t]+2\eta^2\EE_\xi[F_t]\left(\alpha_{1,t}^2-\alpha_{2,t}^2\right)\\ &\leq\left(1-4\eta \sigma^2\right)(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2+24\eta \frac{\eta C_2}{\sigma^2}+\tilde{C}_4\eta^2, \end{align*} where ${\tilde{C}_4=O(\frac{1}{\sigma^4}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right)^2)}$. Denote $\Delta\triangleq\frac{24\eta \frac{\eta C_2}{\sigma^2}+\tilde{C}_4\eta^2}{4\eta \sigma^2}$. the inequality above can be re-expressed as \begin{align*} \EE[\{(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2-\Delta\}|\cF_t]\leq\left(1-4\eta \sigma^2\right)\{(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-\Delta\}. \end{align*} Denote $\cB_t\triangleq\{\forall \tau\leq t : ((\alpha_{1,\tau}^2-\alpha_{2,\tau}^2)^2\geq4\Delta\}$, for $t>\frac{\log(\frac{\left(\alpha_{1,0}^2-\alpha_{2,0}^2\right)^2}{\delta})}{4\eta \sigma^2}$, we have \begin{align*} 4\Delta\EE[\mathds{1}_{\cB_t}]\leq\EE[(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-\Delta]\leq(1-4\eta c)^t((\alpha_{1,0}^2-\alpha_{2,0}^2)^2-\Delta)+\Delta<2\Delta, \end{align*} where the first inequality comes from the definition of $\cB_t$ and the second one comes from calculation above. Thus, if we choose $t=O(\frac{\log(\frac{1}{\Delta \sigma^2})}{\eta \sigma^2})$ and recursively applying the inequality above $\log(\frac{1}{\delta})$ times, we will get, for $\tau_4=O(\frac{1}{\eta \sigma^2}\log(\frac{1}{\delta})\log(\frac{1}{\Delta \sigma^2}))=O(\frac{1}{\eta \sigma^2}\log(\frac{1}{\delta})\log(\frac{1}{\eta}))$, \begin{align*} \PP(\cB_{\tau_4})<(\frac{1}{2})^{\log(\frac{1}{\delta})}=\delta. \end{align*} Thus, w.p. at least $1-\delta$, there exists a $\tau\leq \tau_4$ s.t. $(\alpha_{1,\tau}^2-\alpha_{2,\tau}^2)^2<4\Delta$. Here, we finish the proof of Lemma \ref{lma3}. \end{proof} \begin{lemma} \label{lma4} Suppose there exists a time step $\tau\leq \tau_4= O(\frac{1}{\eta\sigma^2}\log\frac{1}{\Delta\sigma^2}\log\frac{1}{\delta})$ such that \begin{align*} \left((x_{\tau}^\top u_*)^2-(y_{\tau}^\top v_*)^2\right)^2<4\Delta, \end{align*}then $\forall \delta \in (0,1)$, with probability at least $1-\delta$, $\forall t\leq T_1\triangleq O(\frac{1}{\eta^2})$, \begin{align*} (x_t,y_t) \in \left\{(x,y)| \left((x^\top u_*)^2-(y^\top v_*)^2\right)^2<6\Delta \right\}. \end{align*} \end{lemma} \begin{proof} Without loss of generality, we place the time origin at $\tau$, i.e. $(\alpha_{1,0}^2-\alpha_{2,0}^2)^2<4\Delta$. Denote $\cD\triangleq\{(\alpha_1,\alpha_2)|(\alpha_1^2-\alpha_2^2)^2<6\Delta\}$ and $\cD_t\triangleq\{\forall \tau\leq t:(\alpha_{1,\tau},\alpha_{2,\tau})\in \cD\}$. Note that $(\alpha_{1,t}\alpha_{2,t}-1)^2<6\gamma$ still holds with high probability. Defining $H_t\triangleq(1-4\eta\sigma^2)^{-t}\{(\alpha_{1,t}^2-\alpha_{2,t}^2)^2-\Delta\}$, we can get \begin{align*} \EE[H_{t+1}\mathds{1}_{\cD_t}|\cF_t]\leq H_{t}\mathds{1}_{\cD_t}\leq H_{t}\mathds{1}_{\cD_{t-1}}. \end{align*} This means $\{H_{t}\mathds{1}_{\cD_{t-1}}\}$ is a supermartingale. To use Azuma's inequality, we need to bound the following difference \begin{align*} \Bar{d}_{t+1}&\triangleq|H_{t+1}\mathds{1}_{\cD_t}-\EE[H_{t+1}\mathds{1}_{\cD_t}|\cF_t]|\\ &=(1-4\eta \sigma^2)^{-t-1}|(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2-\EE[(\alpha_{1,t+1}^2-\alpha_{2,t+1}^2)^2|\cF_t]|\mathds{1}_{\cD_t}\\ &\leq(1-4\eta \sigma^2)^{-t-1}\tilde{C}_5, \end{align*} where ${\tilde{C}_5=O(\eta\frac{\sqrt{\Delta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right))}$. Further define $\Bar{r}_t\triangleq\sqrt{\sum_{i=1}^t \Bar{d_i}^2}$, by Azuma's inequality we will get \begin{align*} \mathbb{P}\left(H_{t} \mathds{1}_{\cD_{t-1}}-H_{0} \geq O(1) \Bar{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) \leq \exp \left(-\frac{O(1) \Bar{r}_{t}^{2} \log \left(\frac{1}{\eta^{2} \delta}\right)}{2 \sum_{i=0}^{t} \Bar{d_{i}}^{2}}\right)=O\left(\eta^{2} \delta\right). \end{align*} Thus, with probability at least $1-O(\eta^2 \delta)$, we have \begin{align*} ((\alpha_{1,t}^2-\alpha_{2,t}^2)^2-\Delta) \mathds{1}_{\cD_{t-1}} &< (1-4\eta \sigma^2)^t\left((\alpha_{1,0}^2-\alpha_{2,0}^2)^2+O(1) \Bar{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<(\alpha_{1,0}^2-\alpha_{2,0}^2)^2+O\left( (1-4\eta \sigma^2)^t \Bar{r}_{t} \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right)\\ &<4\Delta+O\left( \frac{\sqrt{\Delta\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right) \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right). \end{align*} When $\cD_{t-1}$ holds, w.p. at least $1-O(\eta^2 \delta)$, by the inequality above we have \begin{align*} (\alpha_{1,t}^2-\alpha_{2,t}^2)^2&<5\Delta+O\left( \frac{\sqrt{\Delta\eta}}{\sigma^2}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right) \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)\right) <6\Delta. \end{align*} Note that to make sure last terms is smaller than $\Delta$, we need \begin{align*} \frac{\sqrt{\eta}}{\sigma^2\sqrt{\Delta}}\left(\sqrt{\log\frac{d_1+d_2}{\eta^2}}+\sqrt{\log\frac{1}{\delta}}\right) \log ^{\frac{1}{2}}\left(\frac{1}{\eta^{2} \delta}\right)=O(1). \end{align*} As $\Delta=O(\frac{\eta}{\sigma^{10}})$, we know that as long as $\eta$ is polynomial in $\sigma$, choosing $\sigma=O((\log\frac{1}{\delta})^{-\frac{1}{3}})$ is sufficient. Now we know that if $\cD_{t-1}$ holds, $\cD_{t}$ holds w.p. at least $1-O(\eta^2 \delta)$. It is easy to show that the Perturbed GD updates satisfy $(\alpha_{1,t}^2-\alpha_{2,t}^2)^2<6\Delta$ in the following $O(\frac{1}{\eta^2})$ steps w.p. at least $1-\delta$. \end{proof} With Lemmas \ref{lem_bounded}, \ref{lem_loss} and \ref{a1=a2}, we can prove Lemma \ref{lem_convergence}. Here follows a brief proof. \begin{proof} \begin{align*} |1-x_t^\top u_*|&<(1+x_t^\top u_*)|1-x_t^\top u_*|\\ &=|1-(x_t^\top u_*)^2+x_t^\top u_*v_*^\top y_t-x_t^\top u_*v_*^\top y_t|\\ &\leq|1-x_t^\top u_*v_*^\top y_t|+|(x_t^\top u_*)^2-x_t^\top u_*v_*^\top y_t|\\ &=|1-x_t^\top M y_t|+x_t^\top u_*|x_t^\top u_*-v_*^\top y_t|\\ &\leq|1-x_t^\top M y_t|+\frac{\sqrt{2}}{\sigma}|x_t^\top u_*-v_*^\top y_t|\\ &<\sqrt{6\gamma}+\frac{\sqrt{2}}{\sigma}\sqrt{6\Delta}. \end{align*} The last inequality comes from Lemmas \ref{lma22} and \ref{lma4}. Together with Lemma \ref{lem_bounded} we can get \begin{align*} \norm{x_t-u_*}^2=(1-x_t^\top u_*)^2+\norm{\beta_{1,t}}_2^2<(\sqrt{6\gamma}+\frac{\sqrt{2}}{\sigma}\sqrt{6\Delta})^2+2\frac{\eta C_2}{\sigma^2}=O(\sigma^2+\frac{\eta}{\sigma^{10}}). \end{align*} Note that we use $C_2=O(\frac{1}{\sigma^6})$ when calculating the order. Similarly, we have \begin{align*} \norm{y_t-v_*}^2<(\sqrt{6\gamma}+\frac{\sqrt{2}}{\sigma}\sqrt{6\Delta})^2+2\frac{\eta C_2}{\sigma^2}=O(\sigma^2+\frac{\eta}{\sigma^{10}}). \end{align*} Then, for any $\epsilon>0$, by choosing $\sigma=O(\sqrt{\epsilon})$ and $\eta\leq\eta_5=O(\sigma^{10} \epsilon)$, we will have $\norm{x_t-u_*}^2<\epsilon$ and $\norm{y_t-v_*}^2<\epsilon$ \end{proof} \section{Proof of Theorem \ref{thm_rankr}}\label{pf_3} Recall that the gradient of $\tilde\cF$ takes the following form. \begin{align*} \nabla_X\tilde\cF(X,Y)&=(XY^\top-M)Y-d_2\sigma_2^2X,\\ \nabla_Y\tilde\cF(X,Y)&=(XY^\top-M)^\top X-d_1\sigma_1^2Y. \end{align*} Suppose $(U,V)$ is a stationary point. Then we have \begin{align} (UV^\top-M)V-d_2\sigma_2^2U&=0,\label{eq_stat_1}\\ (UV^\top-M)^\top U-d_1\sigma_1^2V&=0. \label{eq_stat_2} \end{align} \noindent $\bullet$ {\bf Step 1:} To prove the first statement, simply left multiply each side of \eqref{eq_stat_1} by $U^\top$ and each side of \eqref{eq_stat_2} by $V^\top,$ and we have the following equations. \begin{align*} U^\top UV^\top V-U^\top MV-d_2\sigma_2^2U^\top U&=0,\\ V^\top VU^\top U-V^\top M^\top U-d_1\sigma_1^2V^\top V&=0. \end{align*} Note that the following equation naturally holds. $$U^\top UV^\top V-U^\top MV=\left(V^\top VU^\top U-V^\top M^\top U\right)^\top.$$ Combine these three equations together and we have $$U^\top U=\frac{d_1\sigma_1^2}{d_2\sigma_2^2}(V^\top V)^\top=\gamma^2V^\top V.$$ \noindent $\bullet$ {\bf Step 2:} We next show that $(\tilde U,\tilde V)R$ is a stationary point, where $$(\tilde U,\tilde V)=(\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}},\frac{1}{\sqrt{\gamma} }B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}),$$ where $R\in \RR^{r\times r}$ is an orthogonal matrix. We only need to check \eqref{eq_stat_1} and \eqref{eq_stat_2}. In fact, we have \begin{align*} \nabla_X\tilde\cF(\tilde UR,\tilde VR)&=-\gamma\sigma^2 AB^\top\frac{1}{\sqrt{\gamma} }B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R+\sigma^2\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R\\ &=-\sigma^2\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R+\sigma^2\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R\\ &=0, \end{align*} and \begin{align*} \nabla_Y\tilde\cF(\tilde UR,\tilde VR)&=-\gamma\sigma^2 BA^\top\sqrt{\gamma} A(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R+\sigma^2\gamma^2\frac{1}{\sqrt{\gamma} }B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R\\ &=-\sigma^2\gamma\sqrt{\gamma} B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R+\sigma^2\gamma\sqrt{\gamma} B(\Sigma-\gamma\sigma^2I_r)^{\frac{1}{2}}R\\ &=0. \end{align*} Combine the above equations together and we know that $(\tilde U,\tilde V)$ is a stationary point. \noindent $\bullet$ {\bf Step 3:} We next show that $\{(\tilde U,\tilde V)R\big| R\in\RR^{r\times r}, \text{orthogonal}\}$ are the global minima and all other stationary points enjoy strict saddle property. Without loss of generality, we assume $\gamma=1.$ We first calculate the Hessian $\nabla^2 \tilde\cF(X,Y).$ The Hessian can be viewed as a matrix that operates on vectorized matrices of dimension $(d_1+d_2)\times r.$ Then, for any $W\in\RR^{(d_1+d_2)\times r}$, the Hessian defines a quadratic form $$[\nabla^2 \tilde\cF(W)](Z_1,Z_2)=\sum_{i,j,k,l}\frac{\partial^2\tilde\cF(W)}{\partial W[i,j]\partial W[k,l]}Z_1[i,j]Z_2[k,l], ~\forall Z_1,Z_2\in\RR^{(d_1+d_2)\times r}.$$ We can then express the Hessian $\nabla^2 \tilde\cF(W)$ as follows: \begin{align*} [\nabla^2 \tilde\cF(X,Y)](\Delta,\Delta)=2<XY^\top-M,\Delta_U\Delta_V^\top>+\norm{U\Delta_V^\top+\Delta_U V^\top}_{\rm{F}}^2+\sigma^2 \norm{\Delta_U}_{\rm{F}}^2+\sigma^2 \norm{\Delta_V}_{\rm{F}}^2, \end{align*} where $\Delta=\begin{bmatrix} \Delta_U\\ \Delta_V \end{bmatrix},$ $\Delta_U\in \RR^{d_1\times r}$ and $\Delta_V\in \RR^{d_2\times r}.$ We further denote $W=\begin{bmatrix} X\\ Y \end{bmatrix},$ $\tilde W=\begin{bmatrix} \tilde U\\ \tilde V \end{bmatrix},$ and $\tilde M=\tilde U\tilde V^\top,$ $$R=\argmin_{R'\in\RR^{r\times r}, \text{orthogonal}}\norm{W-\tilde WR'}.$$ We then have the following lemma. \begin{lemma}\label{lem_negative_curv} Let $\sigma_{\min}(M)$ be the smallest singular value of $M$. Suppose $d_1\sigma_1^2=d_2 \sigma_2^2=\sigma^2,$ and $\sigma^2<\sigma_{\min}(M).$ For $\forall(U,V)\in (\RR^{d_1},\RR^{d_2})$ such that $\nabla\tilde\cF(U,V)=0,$ we denote $\Delta= \begin{bmatrix} U-\tilde U R\\ V-\tilde V R \end{bmatrix},$ then we have \begin{align}\label{negative_curv} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)\leq -\norm{UV^\top-\tilde M}_{\rm{F}}^2-3\sigma^2\norm{A^\top U-B^\top V}_{\rm{F}}^2. \end{align} Moreover, $[\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)<0$ if $$(U,V)\notin \{(\tilde U,\tilde V)R'\big|R'\in \RR^{r\times r}, R'R'^\top=R'^\top R'=I_r \}.$$\end{lemma} \begin{proof} Recall that the quadratic form defined by the Hessian can be written as follows. \begin{align}\label{quadratic} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)=2<UV^\top-M,\Delta_U\Delta_V^\top>+\norm{U\Delta_V^\top+\Delta_U V^\top}_{\rm{F}}^2+\sigma^2 \norm{\Delta_U}_{\rm{F}}^2+\sigma^2 \norm{\Delta_V}_{\rm{F}}^2. \end{align} We start from the second term $\norm{U\Delta_V^\top+\Delta_U V^\top}_{\rm{F}}^2$. Similar to the proof of Claim B.5 in \citet{du2018algorithmic}, we have \begin{align*} \norm{U\Delta_V^\top+\Delta_U V^\top}_{\rm{F}}^2&= \norm{\Delta_U\Delta_V^\top+ UV^\top-\tilde M}_{\rm{F}}^2\\ &= \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2+2 <\Delta_U\Delta_V^\top, UV^\top-\tilde M>\\ &= \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2+2 <\Delta_U\Delta_V^\top, UV^\top-M>+2 <\Delta_U\Delta_V^\top, M-\tilde M>\\ &= \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2+2 <\Delta_U\Delta_V^\top, UV^\top-M>+2 \sigma^2<\Delta_U\Delta_V^\top, AB^\top> \end{align*} Plugging this equation into \eqref{quadratic}, we have \begin{align*} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)&=4<UV^\top-M,\Delta_U\Delta_V^\top>+ \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2\\&~~~+\sigma^2 \left(\norm{\Delta_U}_{\rm{F}}^2+ \norm{\Delta_V}_{\rm{F}}^2+2 <\Delta_U\Delta_V^\top, AB^\top> \right). \end{align*} Note that using the fact $\nabla\tilde\cF(U,V)=\nabla\tilde\cF(\tilde U,\tilde V)=0,$ one can easily verify that \begin{align*} 4<UV^\top-M,\Delta_U\Delta_V^\top>&=4<UV^\top-M,\tilde M>-2\sigma^2 \left(\norm{\Delta_U}_{\rm{F}}^2+ \norm{\Delta_V}_{\rm{F}}^2\right)+2\sigma^2 \left(\norm{\tilde U}_{\rm{F}}^2+ \norm{\tilde V}_{\rm{F}}^2\right)\\ &=-4\norm{UV^\top-\tilde M}_{\rm{F}}^2-4\sigma^2<AB^\top,\tilde M-UV^\top>\\&~~~+2\sigma^2\left(\norm{\tilde U}_{\rm{F}}^2+ \norm{\tilde V}_{\rm{F}}^2-\norm{ U}_{\rm{F}}^2-\norm{ V}_{\rm{F}}^2\right). \end{align*} Thus, \begin{align} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)&=-4\norm{UV^\top-\tilde M}_{\rm{F}}^2+ \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2\nonumber\\ &~~~-\sigma^2 \Bigg[\norm{\Delta_U}_{\rm{F}}^2+ \norm{\Delta_V}_{\rm{F}}^2-2 <\Delta_U\Delta_V^\top, AB^\top>\nonumber\\ &~~~~~-2\left(\norm{\tilde U}_{\rm{F}}^2+ \norm{\tilde V}_{\rm{F}}^2-\norm{ U}_{\rm{F}}^2-\norm{ V}_{\rm{F}}^2\right) +4<AB^\top,\tilde M-UV^\top>\Bigg].\label{eq_quad} \end{align} We then have the following two claims: \begin{claim}\label{claim_a} $-4\norm{UV^\top-\tilde M}_{\rm{F}}^2+ \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2\leq -\norm{UV^\top-\tilde M}_{\rm{F}}^2.$ \end{claim} \begin{proof} Similar to the proof of Claim B.5 in \citet{du2018algorithmic}, we have \begin{align*} \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2&\leq\frac{1}{4}\norm{\Delta\Delta^\top}_{\rm{F}}^2\\ &\leq \frac{1}{2}\norm{WW^\top-\tilde W\tilde W^\top}_{\rm{F}}^2\\ &=2\norm{UV^\top-\tilde M}_{\rm{F}}^2-\norm{U^\top\tilde U-V^\top\tilde V}_{\rm{F}}^2+\frac{1}{2}\norm{U^\top U-V^\top V}_{\rm{F}}^2+\frac{1}{2}\norm{\tilde U^\top \tilde U-\tilde V^\top \tilde V}_{\rm{F}}^2\\ &=2\norm{UV^\top-\tilde M}_{\rm{F}}^2-\norm{U^\top\tilde U-V^\top\tilde V}_{\rm{F}}^2\\ &\leq 2\norm{UV^\top-\tilde M}_{\rm{F}}^2. \end{align*} Thus, \begin{align*} -4\norm{UV^\top-\tilde M}_{\rm{F}}^2+ \norm{\Delta_U\Delta_V^\top}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2&\leq-4\norm{UV^\top-\tilde M}_{\rm{F}}^2+ 2\norm{UV^\top-\tilde M}_{\rm{F}}^2+ \norm{UV^\top-\tilde M}_{\rm{F}}^2\\&=-\norm{UV^\top-\tilde M}_{\rm{F}}^2. \end{align*} \end{proof} \begin{claim}\label{claim_b} $\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2 +2<AB^\top,\tilde M-UV^\top>=\norm{A^\top \Delta_U-B^\top \Delta_V}_{\rm{F}}^2 .$ \end{claim} \begin{proof} First, the LHS of the equation can be rewritten as follows. \begin{align*} &\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2 +2<AB^\top,\tilde M-UV^\top>\\ =&\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>+4<AB^\top,\tilde M>-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2\\ =&\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>+4\norm{\tilde U}_{\rm{F}}^2-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2\\ =&\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2 +\norm{\tilde U}_{\rm{F}}^2+\norm{\tilde V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>, \end{align*} where we use the fact $$<AB^\top,\tilde M>=\norm{\tilde U}_{\rm{F}}^2=\norm{\tilde V}_{\rm{F}}^2.$$ On the other hand, the RHS of the equation can be rewritten as follows. \begin{align*} \norm{A^\top \Delta_U-B^\top \Delta_V}_{\rm{F}}^2&=\norm{\Delta_U}_{\rm{F}}^2+ \norm{\Delta_V}_{\rm{F}}^2-2 <\Delta_U\Delta_V^\top, AB^\top>\\ &=\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2 +\norm{\tilde U}_{\rm{F}}^2+\norm{\tilde V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>\\ &~~~~~-2\tr(U(\tilde UR)^\top)-2\tr(V(\tilde VR)^\top)+2\tr(U(\tilde UR)^\top)+2\tr(V(\tilde VR)^\top)\\ &=\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2 +\norm{\tilde U}_{\rm{F}}^2+\norm{\tilde V}_{\rm{F}}^2-2<AB^\top,\tilde M+UV^\top>\\ &=\norm{ U}_{\rm{F}}^2+\norm{ V}_{\rm{F}}^2-\norm{\tilde U}_{\rm{F}}^2- \norm{\tilde V}_{\rm{F}}^2 +2<AB^\top,\tilde M-UV^\top>. \end{align*} \end{proof} Plugging the conclusions in Claims \ref{claim_a} and \ref{claim_b} into \eqref{eq_quad}, we have \begin{align*} [\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)\leq-\norm{UV^\top-\tilde M}_{\rm{F}}^2-3\sigma^2\norm{A^\top \Delta_U-B^\top \Delta_V}_{\rm{F}}^2. \end{align*} Note that since $A^\top \tilde U R=B^\top \tilde V R,$ we have $A^\top \Delta_U-B^\top \Delta_V=A^\top U-B^\top V.$ To justify our last statement, we have the following claim. \begin{claim}\label{claim_v} $\norm{UV^\top-\tilde M}_{\rm{F}}^2+3\sigma^2\norm{A^\top U-B^\top V}_{\rm{F}}^2=0$ if and only if $$(U,V)=(\tilde U,\tilde V)R,$$ where $R\in \RR^{r\times r}$ is an orthogonal matrix. \end{claim} \begin{proof} $\norm{UV^\top-\tilde M}_{\rm{F}}^2+3\sigma^2\norm{A^\top U-B^\top V}_{\rm{F}}^2=0$ if and only if \begin{align} UV^\top&=\tilde M,\label{con1}\\ A^\top U&=B^\top V.\label{con2} \end{align} Left multiplying each side of \eqref{con1} by $A^\top,$ we have \begin{align*} A^\top UV^\top=(\Sigma-\sigma^2I)B^\top &\Leftrightarrow B^\top VV^\top=(\Sigma-\sigma^2I)B^\top\\ &\Leftrightarrow VV^\top=B(\Sigma-\sigma^2I)B^\top\\ &\Leftrightarrow V=B(\Sigma-\sigma^2I)^{\frac{1}{2}}R=\tilde VR' . \end{align*} The last equivalent argument comes from Theorem 4 in \citet{li2019symmetry}. Plugging $V=\tilde VR'$ back to \eqref{con2}, we have $U=\tilde UR'.$ \end{proof} As a direct result of Claim \ref{claim_v}, for stationary point $(U,V)\notin\{(\tilde U,\tilde V)R'\big|R'\in \RR^{r\times r}, R'R'^\top=R'^\top R'=I_r \},$ we have $$[\nabla^2 \tilde\cF(U,V)](\Delta,\Delta)<0.$$ We prove the lemma. \end{proof} Lemma \ref{lem_negative_curv} directly implies that $\{(\tilde U,\tilde V)R'\big|R'\in \RR^{r\times r}, R'R'^\top=R'^\top R'=I_r \}$ contains all the global optima, and all other stationary points enjoy strict saddle property. \newpage \section{Perturbed GD}\label{alg_detail} The detail of the Perturbed GD algorithm is summarized in Algorithm \ref{alg:Perturbed GD}. \begin{algorithm}[H] \caption{Perturbed Gradient Descent for Rank-1 Matrix Factorization.} \label{alg:Perturbed GD} \begin{algorithmic} \STATE{\textbf{Input}: step size $\eta$, noise level $\sigma_1, \sigma_2$, matrix $M \in \RR^{d_1 \times d_2}$, number of iterations $T$.} \STATE{\textbf{Initialize}: initialize $(x_0, y_0)$ arbitrarily.} \FOR{$t = 0 \ldots T-1$} \STATE{Sample $\xi_{1,t} \sim N(0, \sigma_1^2 I_{d_1})$ and $\xi_{2,t} \sim N(0, \sigma_2^2 I_{d_2})$.} \STATE{$\tilde{x}_t = x_t + \xi_{1,t}, ~ \tilde{y}_t = y_t + \xi_{2,t}.$} \STATE{$x_{t+1} = x_t - \eta (\tilde{x}_t \tilde{y}_t^\top - M ) \tilde{y}_t.$} \STATE{$y_{t+1} = y_t - \eta (\tilde{y}_t \tilde{x}_t^\top - M^\top ) \tilde{x}_t.$} \ENDFOR \end{algorithmic} \end{algorithm}

https://arxiv.org/abs/2212.00228

Gated Recurrent Neural Networks with Weighted Time-Delay Feedback

We introduce a novel gated recurrent unit (GRU) with a weighted time-delay feedback mechanism in order to improve the modeling of long-term dependencies in sequential data. This model is a discretized version of a continuous-time formulation of a recurrent unit, where the dynamics are governed by delay differential equations (DDEs). By considering a suitable time-discretization scheme, we propose $\tau$-GRU, a discrete-time gated recurrent unit with delay. We prove the existence and uniqueness of solutions for the continuous-time model, and we demonstrate that the proposed feedback mechanism can help improve the modeling of long-term dependencies. Our empirical results show that $\tau$-GRU can converge faster and generalize better than state-of-the-art recurrent units and gated recurrent architectures on a range of tasks, including time-series classification, human activity recognition, and speech recognition.

\section{Illustrations of Differences Between ODE and DDE Dynamics} \label{sect:appA} Of particular interest to us are the differences between ODEs and DDEs that are driven by an input. To illustrate the differences in the context of RNNs in terms of how they map the input signal into an output signal, we consider the simple examples: \begin{itemize} \item[(a)] DDE based RNNs with the hidden states $h \in \RR$ satisfying $\dot{h} = -h(t-\tau) + u(t)$, with $\tau = 0.5$ and $\tau = 1$, and $h(t) = 0$ for $t \in [-\tau,0]$, and \item[(b)] an ODE based RNN with the hidden states $h \in \RR$ satisfying $\dot{h} = -h(t) + u(t)$, \end{itemize} where $u(t) = \cos(t)$ is the driving input signal. Figure \ref{fig:diff} shows the difference between the dynamics of the hidden state driven by the input signal for (a) and (b). We see that, when compared to the ODE based RNN, the introduced delay causes time lagging in the response of the RNNs to the input signal. The responses are also amplified. In particular, using $\tau = 0.5$ makes the response of the RNN closely matches the input signal. In other words, simply fine tuning the delay parameter $\tau$ in the scalar RNN model is sufficient to replicate the dynamics of the input signal. To further illustrate the differences, we consider the following examples of RNN with a nonlinear activation: \begin{itemize} \item[(c)] DDE based RNNs with the hidden states $h \in \RR$ satisfying $\dot{h} = - h + \tanh(-h(t-\tau) + s(t))$, with $\tau > 0$, and $h(t) = 0$ for $t \in [-\tau,0]$, and \item[(d)] an ODE based RNN with the hidden states $h \in \RR$ satisfying $\dot{h} = - h + \tanh(-h(t) + s(t))$, \end{itemize} where the driving input signal $s$ is taken to be the truncated Weierstrass function: \begin{equation} s(t) = \sum_{n=0}^3 a^{-n} \cos(b^n \cdot \omega t ), \end{equation} where $a=3$, $b=4$ and $\omega=2$. Figure \ref{fig:diff2} shows the difference between the input-driven dynamics of the hidden states for (c) and (d). We see that, when compared to the ODE based RNN, the introduced delay causes time lagging in the response of the RNNs to the input signal. Even though the response of both RNNs does not match the input signal precisely (since we consider RNNs with one-dimensional hidden states here and therefore their expressivity is limited), we see that using $\tau = 0.5$ produces a response that tracks the seasonality of the input signal better than the ODE based RNN. \begin{figure}[!t] \centering \includegraphics[width=0.85\textwidth]{figs/diff_plot.pdf} \caption{Hidden state dynamics of the DDE based RNNs with $\tau=0.5$ and $\tau=1$, and the ODE based RNN ($\tau=0$). All RNNs are driven by the same cosine input signal. } \label{fig:diff} \end{figure} \begin{figure}[!h] \centering \includegraphics[width=0.85\textwidth]{figs/simple_plot2.pdf} \caption{Hidden state dynamics of the DDE based RNN with $\tau=0.5$ and the ODE based RNN ($\tau = 0$). All RNNs are driven by the same input signal $s(t)$. } \label{fig:diff2} \end{figure} \section{Proof of Theorem \ref{thm_exist2main}} \label{sect:appB} In this section, we provide a proof of Theorem \ref{thm_exist2main}. To start, note that one can view the solution of the DDE as a mapping of functions on the interval $[t -\tau,t]$ into functions on the interval $[t,t + \tau]$, i.e., as a sequence of functions defined over a set of contiguous time intervals of length $\tau$. This perspective makes it more straightforward to prove existence and uniqueness theorems analogous to those for ODEs than by directly viewing DDEs as an evolution over the state space $\RR^d$. The following theorem, adapted from Theorem 3.7 in \cite{smith2011introduction}, provides sufficient conditions for existence and uniqueness of solution through a point $(t_0, \phi) \in \mathbb{R} \times C$ for the IVP \eqref{eq_genddemain}. Recall that $C := C([-\tau, 0], \mathbb{R}^d)$, the Banach space of continuous functions from $[-\tau, 0]$ into $\mathbb{R}^d$ with the topology of uniform convergence. It is equipped with the norm $\|\phi \| := \sup \{ |\phi(\theta)| : \theta \in [-\tau, 0] \}$. \begin{theorem}[Adapted from Theorem 3.7 in \cite{smith2011introduction}] \label{thm_existence} Let $t_0 \in \RR$ and $\phi \in C$ be given. Assume that $f$ is continuous and satisfies the Lipschitz condition on each bounded subset of $\mathbb{R} \times C$, i.e., for all $a, b \in \mathbb{R}$, there exists a constant $K > 0$ such that \begin{equation} |f(t, \phi) - f(t, \psi)| \leq K \| \phi - \psi \|, \ t \in [a,b], \ \|\phi\|, \|\psi \| \leq M, \end{equation} with $K$ possibly dependent on $a, b, M$. There exists $A > 0$, depending only on $M$ such that if $\phi \in C$ satisfies $\| \phi \| \leq M$, then there exists a unique solution $h(t) = h(t, \phi)$ of Eq. \eqref{eq_genddemain}, defined on $[t_0 - \tau, t_0 + A]$. Moreover, if $K$ is the Lipschitz constant for $f$ corresponding to $[t_0, t_0 + A]$ and M, then \begin{equation} \max_{\eta \in [t_0 - \tau, t_0 + A]} |h(\eta, \phi) - h(\eta, \psi)| \leq \|\phi - \psi \| e^{KA}, \ \|\phi\|, \| \psi \| \leq M. \end{equation} \end{theorem} We now provide existence and uniqueness result for the continuous-time $\tau$-GRU model, assuming that the input $x$ is continuous in $t$. As before, we define the state $h_t \in C$ as: \begin{equation} h_t(\theta) := h(t+\theta), \ -\tau \leq \theta \leq 0. \end{equation} Then the DDE defining the model can be formulated as the following IVP for the nonautonomous system: \begin{equation} \label{eq_gendde} \dot{h}(t) = -h(t) + u(t, h(t)) + a(t, h(t)) \odot z(t, h_t), \ t \geq t_0, \end{equation} and $h_{t_0} = \phi \in C$ for some initial time $t_0 \in \mathbb{R}$, with the dependence on $x(t)$ casted as dependence on $t$ for notational convenience. Now we restate Theorem \ref{thm_exist2main} from the main text and provide the proof. \begin{theorem}[Existence and uniqueness of solution for continuous-time $\tau$-GRU] \label{thm_exist2} Let $t_0 \in \RR$ and $\phi \in C$ be given. There exists a unique solution $h(t) = h(t, \phi)$ of Eq. \eqref{eq_gendde}, defined on $[t_0 - \tau, t_0 + A]$ for any $A > 0$. In particular, the solution exists for all $t \geq t_0$, and \begin{equation} \| h_t(\phi) - h_t(\psi) \| \leq \| \phi - \psi \| e^{K(t-t_0)}, \end{equation} for all $t \geq t_0$, where $K = 1 + \|W_1\| + \|W_2\| + \|W_4\|/4$. \end{theorem} \begin{proof} We shall apply Theorem \ref{thm_existence}. To verify the Lipschitz condition: for any $\phi, \psi \in C$, \begin{align} & \hspace{-7mm} |(u(t, \phi) + a(t, \phi) \odot z(t, \phi) - \phi) - (u(t, \psi) + a(t, \psi) \odot z(t, \psi) - \psi)| \nonumber \\ &\leq |u(t, \phi) - u(t, \psi)| + |\phi - \psi| + |(a(t,\phi) - a(t, \psi)) \odot z(t,\phi) | + | a(t, \psi) \odot (z(t, \phi) - z(t, \psi) )| \\ &\leq \|W_1\| \cdot |\phi - \psi| + |\phi - \psi| + \frac{1}{4} \|W_4\| \cdot |\phi - \psi| + \|W_2\| \cdot |\phi - \psi| \\ &=: K |\phi - \psi|, \end{align} where we have used the fact that the tanh and sigmoid are Lipschitz continuous, both bounded by one in absolute value, and they have positive derivatives of magnitude no larger than one and $1/4$ respectively in the last inequality~above. Therefore, we see that the right hand side function satisfies a global Lipschitz condition and so the result follows from Theorem \ref{thm_existence}. \end{proof} \section{Proof of Proposition \ref{prop_delaymain}} \label{sect:appC} In this section, we restate Proposition \ref{prop_delaymain}, and provide its proof and some remarks. \begin{proposition} \label{prop_delay} Consider the linear time-delayed RNN whose hidden states are described by the update equation: \begin{equation} \label{eq_recursion} h_{n+1} = Ah_n + Bh_{n-m} + Cu_n, \ n=0,1,\dots, \end{equation} and $h_n = 0$ for $n =-m, -m+1, \dots, 0$ with $m > 0$. Then, assuming that $A$ and $B$ commute, we have: \begin{equation} \frac{\partial h_{n+1}}{\partial u_i} = A^{n-i} C, \end{equation} for $n=0,1, \dots, m$, $i=0,\dots, n$, and \begin{align} \frac{\partial h_{m+1+j}}{\partial u_i} &= A^{m+j-i} C + \delta_{i,j-1} BC + 2 \delta_{i,j-2} ABC \nonumber \\ &\ \ \ \ + 3 \delta_{i,j-3} A^2 B C + \cdots + j \delta_{i,0} A^{j-1} B C, \end{align} for $j = 1,2,\dots, m+1$, $i=0,1,\dots, m+j$, where $\delta_{i,j}$ denotes the Kronecker delta. \end{proposition} \begin{proof} Note that by definition $h_i = 0$ for $i=-m, -m+1, \dots, 0$, and, upon iterating the recursion \eqref{eq_recursion}, one~obtains: \begin{equation} h_{n+1} = A^n C u_0 + A^{n-1} C u_1 + \cdots + AC u_{n-1} + C u_n, \end{equation} for $n = 0, 1, \dots, m$. Now, applying the above formula for $h_1$, we obtain \begin{align} h_{m+2} &= A h_{m+1} + B h_1 + Cu_{m+1} \nonumber \\ &= (B + A^{m+1} ) C u_0 + A^{m} C u_1 + \cdots + A^2 C u_{m-1} + ACu_m + Cu_{m+1}. \end{align} Likewise, we obtain: \begin{align} h_{m+3} &= A h_{m+2} + B h_2 + Cu_{m+2} \nonumber \\ &= (BA + A^{m+2} + AB) C u_0 + (B+A^{m+1})C u_1 + A^m C u_2 + \cdots + AC u_{m+1} + C u_{m+2} \nonumber \\ &= (2AB + A^{m+2}) C u_0 + (B+A^{m+1})C u_1 + A^m C u_2 + \cdots + AC u_{m+1} + C u_{m+2}, \end{align} where we have used commutativity of $A$ and $B$ in the last line above. Applying the above procedure repeatedly and using commutativity of $A$ and $B$ give: \begin{equation} \label{eq_linear} h_{m+1+j} = (A^{m+j} + j A^{j-1} B) C u_0 + ( A^{m+j-1} + (j-1) A^{j-2} B) C u_1 + \cdots + ACu_{m+j-1} + C u_{m+j}, \end{equation} for $j = 1,2, \dots, m+1$. The formula in the proposition then follows upon taking the derivative with respect to the $u_i$ in the above formula for the hidden states $h_k$. \end{proof} We remark that one could also derive formula for the gradients $\frac{\partial h_{n+1+j}}{\partial u_i}$ for $n \geq 2m+1$ as well as those for our $\tau$-GRU architecture analogously, albeit the resulting expression is quite complicated. In particular, the dependence on higher powers of $B$ for the coefficients in front of the Kronecker deltas would appear in the formula for the former case (with much more complicated expressions without assuming commutativity of the matrices). However, we emphasize that the qualitative conclusion derived from the analysis remains the same: that introduction of delays places more emphasis on gradient information due to input elements in the past, so they act as buffers to propagate the gradient information more effectively than the counterpart models without delays. \section{Gradient Bounds for $\tau$-GRU} \label{app:gradbound} \noindent {\bf On the exploding and vanishing gradient problem.} For simplicity of our discussion here, we consider the loss~function: \begin{equation} \mathcal{E}_n = \frac{1}{2} \| y_n - \overline{y}_n \|^2, \end{equation} where $n = 1, \dots, N$ and $\overline{y}_n$ denotes the underlying growth truth. The training of $\tau$-GRU involves computing gradients of this loss function with respect to its underlying parameters $\theta \in \Theta = [W_{1,2,3,4}, U_{1,2,3,4}, V]$ at each iteration of the gradient descent algorithm. Using chain rule, we obtain \cite{pascanu2013difficulty}: \begin{equation} \frac{\partial \mathcal{E}_n}{\partial \theta} = \sum_{k=1}^n \frac{\partial \mathcal{E}_n^{(k)}}{\partial \theta}, \end{equation} where \begin{equation} \frac{\partial \mathcal{E}_n^{(k)}}{\partial \theta} = \frac{\partial \mathcal{E}_n}{\partial h_n} \frac{\partial h_n}{\partial h_k} \frac{\partial^+ h_k}{\partial \theta}, \end{equation} with $\frac{\partial^+ h_k}{\partial \theta}$ denoting the ``immediate'' partial derivative of the state $h_k$ with respect to $\theta$, i.e., where $h_{k-1}$ is taken as a constant with respect to $\theta$ \cite{pascanu2013difficulty}. The partial gradient $\frac{\partial \mathcal{E}_n^{(k)}}{\partial \theta}$ measures the contribution to the hidden state gradient at step $n$ due to the information at step $k$. It can be shown that this gradient behaves as \begin{equation} \frac{\partial \mathcal{E}_n^{(k)}}{\partial \theta} \sim \gamma^{n-k}, \end{equation} for some $\gamma > 0$ \cite{pascanu2013difficulty}. If $\gamma > 1$, then the gradient grows exponentially in sequence length, for long-term dependencies where $k \ll n$, causing the exploding gradient problem. On the other hand, if $\gamma < 1$, then the gradient decays exponentially for $k \ll n$, causing the vanishing gradient problem. Therefore, we can investigate how $\tau$-GRU deals with these problems by deriving bounds on the gradients. In particular, we are interested in the behavior of the gradients for long-term dependencies, i.e., $k \ll n$, and shall show that the delay mechanism in $\tau$-GRU slows down the exponential decay rate, thereby reducing the sensivity to the vanishing gradient problem. Recall that the update equations defining $\tau$-GRU are given by $h_n = 0$ for $n=-m, -m+1, \dots, 0$, \begin{equation} h_n = (1-g(A_{n-1})) \odot h_{n-1} + g(A_{n-1}) \odot [u(B_{n-1}) + a(C_{n-1}) \odot z(D_{n-m-1})], \end{equation} for $n=1,2,\dots, N$, where $m := \lfloor \tau/\Delta t \rfloor \in \{1,2,\dots, N-1\}$, $A_{n-1} = W_3 h_{n-1} + U_3 x_{n-1}$, $B_{n-1} = W_1 h_{n-1} + U_1 x_{n-1}$, $C_{n-1} = W_4 h_{n-1} + U_4 x_{n-1}$, and $D_{n-m-1} = W_2 h_{n-m-1} + U_2 x_{n-1}$. In the sequel, we shall denote the $i$th component of a vector $v$ as $v^i$ and the $(i,j)$ entry of a matrix $A$ as $A^{ij}$. We start with the following lemma. \begin{lemma} \label{app_lem} For every $i$, we have $ h_n^i = 0$, for $n = -m, -m+1, \dots, 0$, and $ |h_n^i| \leq 2$, for $n=1,2,\dots, N$. \end{lemma} \begin{proof} The $i$th component of the hidden states of $\tau$-GRU are given by: $h_n^i = 0$ for $n=-m, -m+1, \dots, 0$, and \begin{equation} h_n^i = (1-g(A^i_{n-1})) h^i_{n-1} + g(A^i_{n-1}) [u(B^i_{n-1}) + a(C^i_{n-1}) z(D^i_{n-m-1})], \end{equation} for $n=1,2,\dots, N$. Using the fact that $g(x), a(x) \in (0,1)$ and $u(x), z(x) \in (-1,1)$ for all $x$, we can bound the $h_n^i$ as: \begin{align} h_n^i &\leq (1-g(A^i_{n-1})) \max(h^i_{n-1},2) + g(A^i_{n-1}) \max(h^i_{n-1},2) \nonumber \\ &\leq \max(h^i_{n-1},2), \end{align} for all $i$ and $n = 1,2,\dots, N$. Similarly, we have: \begin{align} h_n^i &\geq (1-g(A^i_{n-1})) \min(-2, h^i_{n-1}) + g(A^i_{n-1}) \min(-2, h^i_{n-1}) \nonumber \\ &\geq \min(-2, h^i_{n-1}), \end{align} for all $i$ and $n = 1,2,\dots, N$. Thus, \begin{equation} \min(-2, h^i_{n-1}) \leq h_n^i \leq \max(h^i_{n-1},2), \end{equation} for all $i$ and $n = 1,2,\dots, N$. Now, iterating over $n$ and using $h_0^i = 0$ for all $i$, we obtain $-2 \leq h_n^i \leq 2$ for all $i$ and $n = 1,2,\dots, N$. \end{proof} We now provide the gradients bound for $\tau$-GRU in the following proposition and proof. \begin{proposition} \label{app_prop} Assume that there exists an $\epsilon > 0$ such that $\max_n g(A^i_{n-1}) \geq \epsilon$ and $\max_n a(C^i_{n-1}) \geq \epsilon$ for all $i$. Then \begin{equation} \left \| \frac{\partial h_n}{\partial h_k} \right\|_\infty \leq (1+C-\epsilon)^{n-k} + \| W_2\|_\infty \cdot \left( (1+C-\epsilon)^{n-k-2} \delta_{m,1} + \dots + (1+C-\epsilon) \delta_{m, n-k-2} + \delta_{m, n-k-1} \right), \end{equation} for $n=1, \dots, N$ and $k < n$, where $C = \|W_1\|_\infty + \|W_3\|_\infty + \frac{1}{4}\|W_4\|_\infty$. \end{proposition} \begin{proof} Recall $h_n = 0$ for $n=-m, -m+1, \dots, 0$, and \begin{equation} h_n = (1-g(A_{n-1})) \odot h_{n-1} + g(A_{n-1}) \odot [u(B_{n-1}) + a(C_{n-1}) \odot z(D_{n-m-1})] =: F(h_{n-1}, h_{n-m-1}), \end{equation} for $n=1,2,\dots, N$. Denote $q_{n,l} := \frac{\partial F}{\partial h_{n-l}}$, where $F := F(h_{n-1}, h_{n-l})$ for $l>1$. The gradients $\frac{\partial h_n}{\partial h_k}$ can be computed recursively as~follows. \begin{align} p_n^{(1)} &:= \frac{\partial h_n}{\partial h_{n-1}}, \\ p_n^{(2)} &:= \frac{\partial h_n}{\partial h_{n-2}} = p_n^{(1)} p_{n-1}^{(1)} + q_{n,2} \delta_{m,1}, \\ p_n^{(3)} &:= \frac{\partial h_n}{\partial h_{n-3}} = p_n^{(1)} p_{n-1}^{(2)} + q_{n,3} \delta_{m,2}, \\ \vdots \\ p_n^{(n-k)} &:= \frac{\partial h_n}{\partial h_{k}} = p_n^{(1)} p_{n-1}^{(n-k-1)} + q_{n,n-k} \delta_{m,n-k-1}. \end{align} As $\|p_n^{(n-k)}\| \leq \|p_n^{(1)} \| \cdot \| p_{n-1}^{(n-k-1)}\| + \|q_{n,n-k}\| \delta_{m,n-k-1}$, it remains to upper bound the $p_n^{(1)}$ and $q_{n,n-k}$. The $i$th component of the hidden states can be written as: \begin{equation} h_n^i = (1-g(A^i_{n-1})) h^i_{n-1} + g(A^i_{n-1}) [u(B^i_{n-1}) + a(C^i_{n-1}) z(D^i_{n-m-1})], \end{equation} where $A^i_{n-1} = W_3^{iq} h_{n-1}^q + U_3^{ir} x_{n-1}^{r}$, $B^i_{n-1} = W_1^{iq} h_{n-1}^q + U_1^{ir} x_{n-1}^{r}$, $C^i_{n-1} = W_4^{iq} h_{n-1}^q + U_4^{ir} x_{n-1}^{r}$, and $D^i_{n-m-1} = W_2^{iq} h_{n-m-1}^q + U_2^{ir} x_{n-1}^{r}$, using Einstein's summation notation for repeated indices. Therefore, applying chain rule and using Einstein's summation for repeated indices in the following, we obtain: \begin{align} \frac{\partial h_n^i}{\partial h_{n-1}^j} &= (1-g(A_{n-1}^i)) \frac{\partial h_{n-1}^i}{\partial h_{n-1}^j} - \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^l} \frac{\partial A_{n-1}^l}{\partial h_{n-1}^j} h_{n-1}^i + g(A_{n-1}^i) \frac{\partial u(B_{n-1}^i)}{\partial B_{n-1}^l} \frac{\partial B_{n-1}^l}{\partial h_{n-1}^j} \nonumber \\ &\ \ \ \ \ + \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^l} \frac{\partial A_{n-1}^l}{\partial h_{n-1}^j} u(B_{n-1}^i) + g(A_{n-1}^i) \frac{\partial a(C_{n-1}^i)}{\partial C_{n-1}^l} \frac{\partial C_{n-1}^l}{\partial h_{n-1}^j} z(D_{n-m-1}^i) \nonumber \\ &\ \ \ \ \ + \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^l} \frac{\partial A_{n-1}^l}{\partial h_{n-1}^j} a(C_{n-1}^i) z(D_{n-m-1}^i). \end{align} Noting that $\frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^l} = 0$, $\frac{\partial u(B_{n-1}^i)}{\partial B_{n-1}^l} = 0$ and $\frac{\partial a(C_{n-1}^i)}{\partial C_{n-1}^l} = 0$ for $i \neq l$, we have: \begin{align} \frac{\partial h_n^i}{\partial h_{n-1}^j} &= (1-g(A_{n-1}^i)) \frac{\partial h_{n-1}^i}{\partial h_{n-1}^j} - \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^i} \frac{\partial A_{n-1}^i}{\partial h_{n-1}^j} h_{n-1}^i + g(A_{n-1}^i) \frac{\partial u(B_{n-1}^i)}{\partial B_{n-1}^i} \frac{\partial B_{n-1}^i}{\partial h_{n-1}^j} \nonumber \\ &\ \ \ \ \ + \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^i} \frac{\partial A_{n-1}^i}{\partial h_{n-1}^j} u(B_{n-1}^i) + g(A_{n-1}^i) \frac{\partial a(C_{n-1}^i)}{\partial C_{n-1}^i} \frac{\partial C_{n-1}^i}{\partial h_{n-1}^j} z(D_{n-m-1}^i) \nonumber \\ &\ \ \ \ \ + \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^i} \frac{\partial A_{n-1}^i}{\partial h_{n-1}^j} a(C_{n-1}^i) z(D_{n-m-1}^i) \\ &= (1-g(A_{n-1}^i)) \delta_{i,j} - \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^i} W_3^{ij} h_{n-1}^i + g(A_{n-1}^i) \frac{\partial u(B_{n-1}^i)}{\partial B_{n-1}^i} W_1^{ij} \nonumber \\ &\ \ \ \ \ + \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^i} W_3^{ij} u(B_{n-1}^i) + g(A_{n-1}^i) \frac{\partial a(C_{n-1}^i)}{\partial C_{n-1}^i} W_4^{ij} z(D_{n-m-1}^i) \nonumber \\ &\ \ \ \ \ + \frac{\partial g(A_{n-1}^i)}{\partial A_{n-1}^i} W_3^{ij} a(C_{n-1}^i) z(D_{n-m-1}^i). \end{align} Using the assumption that $\max_n g(A^i_{n-1}), \max_n a(C^i_{n-1}) \geq \epsilon$ for all $i$, the fact that $|z(x)|, |u(x)| \leq 1$, $g(x), a(x) \in (0,1)$, $g'(x), a'(x) \leq 1/4$, $u'(x) \leq 1$ for all $x \in \RR$, and Lemma \ref{app_lem}, we obtain: \begin{align} \left|\frac{\partial h_n^i}{\partial h_{n-1}^j} \right| &\leq (1-\epsilon) \delta_{i,j} + \frac{1}{4} |W_3^{ij}| \cdot |h_{n-1}^i| + |W_1^{ij}| + \frac{1}{4} |W_3^{ij}| + \frac{1}{4} |W_4^{ij}| + \frac{1}{4} |W_3^{ij}| \\ &\leq (1-\epsilon) \delta_{i,j} + |W_1^{ij}| + |W_3^{ij}| + \frac{1}{4} |W_4^{ij}|. \end{align} Therefore, \begin{equation} \left \| \frac{\partial h_n}{\partial h_{n-1}} \right\|_\infty := \max_{i=1,\dots, d} \sum_{j=1}^d \left| \frac{\partial h_n^i}{\partial h_{n-1}^j } \right| \leq (1-\epsilon ) + \|W_1\|_\infty + \|W_3 \|_\infty + \frac{1}{4} \|W_4\|_\infty =: 1-\epsilon + C. \end{equation} Likewise, we obtain, for $l>1$: \begin{align} \frac{\partial F^i}{\partial h_{n-l}^j} &= g(A_{n-1}^i) a(C_{n-1}^i) \frac{\partial z(D_{n-l}^i)}{\partial D_{n-l}^i} \frac{\partial D_{n-l}^i}{\partial h_{n-l}^j} \\ &= g(A_{n-1}^i) a(C_{n-1}^i) \frac{\partial z(D_{n-l}^i)}{\partial D_{n-l}^i} W_2^{ij}. \end{align} Using the fact that $|g(x)|, |a(x)| \leq 1$ and $|z'(x)| \leq 1$ for all $x \in \RR$, we obtain: \begin{align} \left|\frac{\partial F^i}{\partial h_{n-l}^j}\right| &\leq |W_2^{ij}|, \end{align} for $l > 1$, and thus $\|q_{n,n-k}\|_\infty = \| \frac{\partial F}{\partial h_k} \|_\infty \leq \|W_2\|_\infty$ for $k > 1$. The upper bound in the proposition follows by using the above bounds for $\|p_{n}^{(1)}\|_\infty$ and $\|q_{n,n-k}\|_\infty$, and iterating the recursion $\|p_n^{(n-k)}\| \leq \|p_n^{(1)} \| \cdot \| p_{n-1}^{(n-k-1)}\| + \|q_{n,n-k}\| \delta_{m,n-k-1}$ over $k$. \end{proof} From Proposition \ref{app_prop}, we see that if $\epsilon > C$, then the gradient norm decays exponentially as $k$ becomes large. However, the delay in $\tau$-GRU introduces jump-ahead connections (buffers) to slow down the exponential decay. For instance, choosing $m=1$ for the delay, we have $\left \| \frac{\partial h_n}{\partial h_k} \right\| \sim (1+C-\epsilon)^{n-k-2}$ as $k \to \infty$ (instead of $\left \| \frac{\partial h_n}{\partial h_k} \right\| \sim (1+C-\epsilon)^{n-k-1}$ as $k \to \infty$ in the case when no delay is introduced into the model). The larger the $m$ is, the more effective the delay is able to slow down the exponential decay of the gradient norm. These qualitative conclusions can already be derived by studying the linear time-delayed RNN, which we consider in the main text for simplicity. \section{Approximation Capability of Time-Delayed RNNs} \label{app:uat} RNNs (without delay) have been shown to be universal approximators of a large class of open dynamical systems \cite{schafer2006recurrent}. Analogously, RNNs with delay can be shown to be universal approximators of open dynamical systems with~delay. Let $m > 0$ (time lag) and consider the state space models (which, in this section, we shall simply refer to as delayed RNNs) of the form: \begin{align} \label{eq_gendRNN} s_{n+1} &= f(As_n + B s_{n-m} + Cu_n + b), \nonumber \\ r_n &= Ds_n, \end{align} and dynamical systems of the form \begin{align} \label{ds_approx} x_{n+1} &= g(x_n, x_{n-m}, u_n), \nonumber \\ o_n &= o(x_n), \end{align} for $n = 0,1,\dots, N$. Here $u_n \in \RR^{d_u}$ is the input, $o_n \in \RR^{d_o}$ is the target output of the dynamical systems to be learned, $s_n \in \RR^d$ is the hidden state of the learning model, $r_n \in \RR^{q}$ is the model output, $f$ is the tanh function applied component-wise, the maps $g$ and $o$ are Lipschitz continuous, and the matrices $A$, $B$, $C$, $D$ and the vector $b$ are learnable parameters. For simplicity, we take the initial functions to be $s_n = y_n = 0$ for $n = -m, -m+1, \dots, 0$. The following theorem shows that the delayed RNNs \eqref{eq_gendRNN} are capable of approximating a large class of time-delay dynamical systems, of the form \eqref{ds_approx}, to arbitrary accuracy. \begin{theorem} \label{thm_uat} Assume that there exists a constant $R > 0$ such that $\max(\|x_{n+1}\|, \|u_n\|) < R$ for $n = 0,1,\dots, N$. Then, for a given $\epsilon > 0$, there exists a delayed RNN of the form \eqref{eq_gendRNN} such that the following holds for some $d$: \begin{equation} \label{thm_bd} \|r_n - o_n \| \leq \epsilon, \end{equation} for $n = 0,1,\dots, N$. \end{theorem} \begin{proof} The proof proceeds along the line of \cite{schafer2006recurrent, rusch2022long}, using the universal approximation theorem (UAT) for feedforward neural network maps and with straightforward modification to deal with the extra delay variables $s_{n-n}$ and $x_{n-m}$ here. The proof proceeds in a similar manner as the one provided in Section E.4 in \cite{rusch2022long}. The goal is to construct hidden states, output state, weight matrices and bias vectors such that an output of the delayed RNN approximates the dynamical system \eqref{ds_approx}. Let $\epsilon > \epsilon^* > 0, R^* > R \gg 1$ be parameters to be defined later. Then, using the UAT for continuous functions with neural networks with the tanh activation function \cite{barron1993universal}, we can obtain the following statements. Given an $\epsilon^*$, there exist weight matrices $W_1$, $W_2$, $W_3$, $V_1$ and a bias vector $b_1$ of appropriate dimensions such that the neural network defined by $\mathcal{N}_1(h, \tilde{h}, u) := W_3 \tanh(W_1 h + W_2 \tilde{h} + V_1 u + b_1)$ approximates the underlying function $g$ as follows: \begin{equation} \max_{\max (\|h\|, \|\tilde{h} \|, \|u\|) < R^* } \| g(h, \tilde{h}, u) - \mathcal{N}_1(h, \tilde{h}, u) \| \leq \epsilon^*. \end{equation} Now, we define the dynamical system: \begin{align} p_{n} &= W_3 \tanh(W_1 p_{n-1} + W_2 p_{n-m-1} + V_1 u_{n-1} + b_1), \end{align} with $p_i = 0$ for $i = -m, -m+1, \dots, 0$. Then using the above approximation bound, we obtain, for $n=1, \dots, N+1$, \begin{align} \|x_{n} - p_{n} \| &= \| g(x_{n-1}, x_{n-m-1}, u_{n-1}) - p_n \| \\ &\leq \|g(x_{n-1}, x_{n-m-1}, u_{n-1}) - W_3 \tanh(W_1 p_{n-1} + W_2 p_{n-m-1} + V_1 u_{n-1} + b_1)\| \\ &\leq \|g(x_{n-1}, x_{n-m-1}, u_{n-1}) - g(p_{n-1}, p_{n-m-1}, u_{n-1}) \| \nonumber \\ &\ \ \ \ \ + \| g(p_{n-1}, p_{n-m-1}, u_{n-1}) - W_3 \tanh(W_1 p_{n-1} + W_2 p_{n-m-1} + V_1 u_{n-1} + b_1)\| \\ &\leq Lip(g) (\| x_{n-1} - p_{n-1} \| + \|x_{n-m-1} - p_{n-m-1}\|) + \epsilon^*, \end{align} where $Lip(g)$ is the Lipschitz constant of $g$ on the compact set $\{ (h,\tilde{h}, u): \| h\|, \|\tilde{h}\|, \|u\| < R^* \}$. Iterating the above inequality over $n$ leads to: \begin{equation} \|x_n - p_n \|\leq \epsilon^* C_1(n, m, Lip(g)), \end{equation} for some constant $C_1>0$ that is dependent on $n, m, Lip(g)$. Using the Lipschitz continuity of the output function $o$, we obtain: \begin{equation} \label{out_b} \| o_n - o(p_n) \| \leq \epsilon^*C_2(n, m, Lip(g), Lip(o)), \end{equation} for some constant $C_2$ that is dependent on $n, m, Lip(g), Lip(o)$, where $Lip(o)$ is the Lipschitz constant of $o$ on the compact set $\{ h : \| h\| < R^* \}$. Next we use the UAT for neural networks again to obtain the following approximation result. Given an $\overline{\epsilon}$, there exists weight matrices $W_4, W_5$ and bias vector $b_2$ of appropriate dimensions such that the tanh neural network, $\mathcal{N}_2(h) := W_5 \tanh(W_4 h + b_2)$ approximates the underlying output function $o$ as: \begin{equation} \max_{\|h\| < R^*} \| o(h) - \mathcal{N}_2(h) \| \leq \overline{\epsilon}. \end{equation} Defining $\overline{o}_n = W_5 \tanh(W_4 p_n + b_2)$, we obtain, using the above approximation bound and the inequality \eqref{out_b}: \begin{equation} \label{o_b_2} \|o_n - \overline{o}_n \| = \|o_n - o(p_n) \| + \| o(p_n) - \overline{o}_n \| \leq \epsilon^* C_2(n, m, Lip(g), Lip(o)) + \overline{\epsilon}. \end{equation} Now, let us denote: \begin{align} \tilde{p}_{n} &= \tanh(W_1 p_{n-1} + W_2 p_{n-m-1} + V_1 u_{n-1} + b_1), \end{align} so that $p_n = W_3 \tilde{p}_n$. With this notation, we have: \begin{equation} \overline{o}_n = W_5 \tanh(W_4 W_3 \tanh(W_1 W_3 \tilde{p}_{n-1} + W_2 W_3 \tilde{p}_{n-m-1} + V_1 u_{n-1} + b_1) + b_2). \end{equation} Since the function $R(y) = W_5 \tanh(W_4 W_3 \tanh(W_1 W_3 y + W_2 W_3 \tilde{p}_{n-m-1} + V_1 u_{n-1} + b_1) + b_2)$ is Lipschitz continuous in $y$, we can apply the UAT again to obtain: for any $\tilde{\epsilon}$, there exists weight matrices $W_6, W_7$ and bias vector $b_3$ of appropriate dimensions such that \begin{equation} \max_{\|y\| < R^*} \| R(y) - W_7 \tanh(W_6 y + b_3) \| \leq \tilde{\epsilon}. \end{equation} Denoting $\tilde{o}_n := W_7 \tanh(W_6 p_{n-1} + b_3)$ and using the above approximation bound, we obtain $\|\overline{o}_n - \tilde{o}_n \| \leq \tilde{\epsilon}$. Finally, we collect all the ingredients above to construct a delayed RNN that can approximate the dynamical system \eqref{ds_approx}. To this end, we define the hidden states (in an enlarged state space): $s_n := (\tilde{p}_n, \hat{p}_n)$, with $\tilde{p}_n$, $\hat{p}_n$ sharing the same dimension. These hidden states evolve according to the dynamical system: \begin{align} s_n &= \tanh\left( \left[ {\begin{array}{cc} W_1 W_3 & 0 \\ W_6 W_3 & 0 \\ \end{array} } \right] s_{n-1} + \left[ {\begin{array}{cc} W_2 W_3 & 0 \\ 0 & 0 \\ \end{array} } \right] s_{n-m-1} + \left[ {\begin{array}{c} V_1 u_{n-1} \\ 0 \\ \end{array} } \right] + \left[ {\begin{array}{c} b_1 \\ 0 \\ \end{array} } \right] \right). \end{align} Defining the output state as $r_n := [0, W_7] s_n$, with the $s_n$ satisfying the above system, we arrive at a delayed RNN that approximates the dynamical system \eqref{ds_approx}. In fact, we can verify that $r_n = \tilde{o}_n$. Setting $\overline{\epsilon} < \epsilon/2$ and $\epsilon^* < \epsilon/(2 C_2(n, m, Lip(g), Lip(o)))$ give us the bound \eqref{thm_bd} in the theorem. \end{proof} \section{Additional Details and Experiments} \label{sect:appD} In this section, we provide additional empirical results and details to demonstrate the advantages of $\tau$-GRU when compared to other RNN architectures. \begin{table}[!b] \caption{Results for Google12. Results indicated by $^*$ are produces by us, results indicated by $^+$ are from~\cite{rusch2022long}. } \label{tab:results_google12} \centering \scalebox{0.9}{ \begin{tabular}{l c c ccccccc} \toprule Model & Test Accuracy ($\%$) & \# units & \# param \\ \midrule tanh RNN~\cite{rusch2022long}$^+$ & 73.4 & 128 & 27k \\ LSTM~\cite{rusch2022long}$^+$ & 94.9 & 128 & {107k}\\ GRU~\cite{rusch2022long}$^+$ & 95.2 & 128 & 80k \\ AsymRNN~\cite{chang2018antisymmetricrnn}$^+$ & 90.2 & 128 & 20k \\ expRNN~\cite{lezcano2019cheap}$^+$ & 92.3 & 128 & 19k \\ coRNN~\cite{rusch2021coupled}$^+$ & 94.7 & 128 & 44k \\ Fast GRNN~\cite{kusupati2018fastgrnn}$^+$ & 94.8 & 128 & 27k \\ LEM~\cite{rusch2022long} & 95.7 & 128 & {107k} \\ Lipschitz RNN~\cite{erichson2020lipschitz}$^*$ & 95.6 & 128 & 34k\\ Noisy RNN~\cite{lim2021noisy}$^*$ & 95.7 & 128 & 34k\\ iRNN~\cite{Kag2020RNNs}$^*$ & 95.1 & - & {8.5k} \\ TARNN~\cite{kag2021time}$^*$ & 95.9 & 128 & {107k} \\ \midrule \textbf{ours} & \textbf{96.2} & 128 & {107k} \\ \bottomrule \end{tabular}} \end{table} \subsection{Speech Recognition: Google 12} Here, we consider the Google Speech Commands data set V2~\cite{warden2018speech} to demonstrate the performance of our model for speech recognition. The aim of this task is to learn a model that can classify a short audio sequence, which is sampled at a rate of 16 kHz from 1 second utterances of $2,618$ speakers. We consider the Google 12-label task (Google12) which is composed of 10 keyword classes, and in addition one class that corresponds to `silence', and a class corresponding to `unknown' keywords. We adopt the standard train/validation/test set split for evaluating our model, and we use dropout, applied to the inputs, with rate 0.03 to reduce overfitting. Table~\ref{tab:results_google12} presents the results for our $\tau$-GRU and a number of competitive RNN architectures. We adopt the results for the competitive RNNs from~\cite{rusch2022long}. Our proposed $\tau$-GRU shows the best performance on this task, i.e., $\tau$-GRU is able to outperform gated and continuous-time RNNs on this task that requires an expressive recurrent unit. \subsection{Learning the Dynamics of Mackey-Glass System} Here, we consider the task of learning the Mackey-Glass equation, originally introduced in \cite{mackey1977oscillation} to model the variation in the relative quantity of mature cells in the blood: \begin{equation} \dot{x} = a \frac{x(t-\delta)}{1 + x^n(t-\delta)} - b x(t), \ t \geq \delta, \end{equation} where $\delta \geq 17$, $a, b, n > 0$, with $x$ satisfying $\dot{x} = a x(0)/(1 + x(0)^n) - b x$ for $t \in [0, \delta]$. It is a scalar equation with chaotic dynamics, with infinite-dimensional state space. Increasing the value of $\delta$ increases the dimension of the~attractor. For data generation, we choose $a = 0.2$, $b = 0.1$, $n=10$, $\delta = 17$, $x(0) \sim \rm{Unif}(0,1)$, and use the classical Runge-Kutta method (RK4) to integrate the system numerically from $t=0$ to $t = 1000$ with a step-size of 0.25. The training and testing samples are the time series (of length 2000) generated by the RK4 scheme on the interval $[500,1000]$ for different realizations of $x(0)$. Figure \ref{fig:data_plot} shows a realization of the trajectory produced by the Mackey-Glass system (and also the DDE based ENSO system considered in the main text). \begin{figure}[!h] \centering \includegraphics[width=0.44\textwidth]{figs/mg_plot.png} \includegraphics[width=0.44\textwidth]{figs/mz_plot.png} \caption{A realization of the Mackey-Glass dynamics (left) and the DDE based ENSO dynamics (right). } \label{fig:data_plot} \end{figure} Table \ref{tab:results_MG} shows that our $\tau$-GRU model (with $\alpha = \beta = 1$ and using $\tau = 10$) is more effective in learning the Mackey-Glass system when compared to other RNN architectures. We also see that the predictive performance deteriorates without making full use of the combination of the standard recurrent unit and delay recurrent unit (setting either $\alpha$ or $\beta$ to zero). Moreover, $\tau$-GRU demonstrates improved performance when compared to the simple delay GRU model (Eq. \eqref{eq_simpleDRNN}) and the counterpart model without using the gating. Similar observation also holds for the ENSO prediction task; see Table \ref{tab:add_results_ENSO}. Figure \ref{fig:mg_traintest} shows that our model converges much faster than other RNN models during training. In particular, our model is able to achieve both lower training and testing error (as measured by the root mean square error (RMSE)) with fewer epochs, demonstrating the effectiveness of the delay mechanism in improving the performance on the problem of long-term dependencies. This is consistent with our analysis on how the gradient information is propagated through the delay buffers in the network (see Proposition \ref{prop_delaymain}), suggesting that the delay buffers can propagate gradient more efficiently. Similar behavior is also observed for the ENSO task; see Figure \ref{fig:mz_traintest}. \begin{table}[!h] \caption{Additional results for the ENSO model prediction.} \label{tab:add_results_ENSO} \centering \scalebox{0.9}{ \begin{tabular}{l c c ccccccc} \toprule Model & MSE ($\times 10^{-2}$) & \# units & \# parameters \\ \midrule simple delay GRU (Eq. \eqref{eq_simpleDRNN}) & 0.2317 & 16 & 0.897k \\ ablation (no gating) & 0.4289 & 16 & 0.929k \\ \midrule \textbf{$\tau$-GRU (ours)} & \textbf{0.17} & 16 & 1.2k \\ \bottomrule \end{tabular}} \end{table} \begin{table}[!t] \caption{Results for the Mackey-Glass system prediction.} \label{tab:results_MG} \centering \scalebox{0.9}{ \begin{tabular}{l c c ccccccc} \toprule Model & MSE ($\times 10^{-2}$) & \# units & \# parameters \\ \midrule Vanilla RNN & 0.3903 & 16 & 0.321k \\ LSTM & 0.6679 & 16 & 1.233k \\ GRU & 0.4351 & 16 & 0.929k \\ Lipschitz RNN & 8.9718 & 16 & 0.561k \\ coRNN & 1.6835 & 16 & 0.561k \\ LEM & 0.1430 & 16 & 1.233k \\ \midrule simple delay GRU (Eq. \eqref{eq_simpleDRNN}) & 0.2772 & 16 & 0.897k \\ ablation (no gating) & 0.2765 & 16 & 0.929k \\ ablation ($\alpha = 0$) & 0.1553 & 16 & 0.625k \\ ablation ($\beta = 0$) & 0.2976 & 16 & 0.929k \\ \midrule \textbf{$\tau$-GRU (ours)} & \textbf{0.1358} & 16 & 1.233k \\ \bottomrule \end{tabular}} \end{table} \begin{figure}[!h] \centering \includegraphics[width=0.44\textwidth]{figs/mg_trainrmse.png} \includegraphics[width=0.44\textwidth]{figs/mg_testrmse.png} \caption{Train RMSE (left) and test RMSE (right) vs. epoch for the Mackey-Glass learning task.} \label{fig:mg_traintest} \end{figure} \begin{figure}[!h] \centering \includegraphics[width=0.44\textwidth]{figs/mz_trainrmse.png} \includegraphics[width=0.44\textwidth]{figs/mz_testrmse.png} \caption{Train RMSE (left) and test RMSE (right) vs. epoch for the ENSO learning task.} \label{fig:mz_traintest} \end{figure} \clearpage \section{Tuning Parameters} To tune our $\tau$-GRU, we use a non-exhaustive random search within the following plausible ranges for $\tau={5,\dots,200}$. We used Adam as our optimization algorithm for all of the experiments. For the synthetic data sets generated by the ENSO and Mackey-Glass system, we used learning rate of 0.01. For the other experiments we considered learning rates between 0.01 and 0.0005. We used dropout for the IMDB and Google12 task to avoid overfitting. Table~\ref{tab:tuning} is listing the tuning parameters for the different tasks that we considered in this work. \begin{table}[!h] \caption{Summary of tuning parameters.} \label{tab:tuning} \centering \scalebox{0.85}{ \begin{tabular}{l c c c c c c c c } \toprule Name & d & lr & $\tau$ & dropout & epochs \\ \midrule Adding Task $N=2000$ & 128 & 0.0026 & 900 & - & 200 \\ Adding Task $N=5000$ & 128 & 0.002 & 2000 & - & 200 \\ \midrule IMDB & 128 & 00012 & 1 & 0.04 & 30 \\ \midrule HAR-2 & 128 & 0.00153 & 10 & - & 100 \\ \midrule sMNIST & 128 & 0.0018 & 50 & - & 60 \\ \midrule psMNIST & 128 & 0.0055 & 65 & - & 80 \\ \midrule sCIFAR & 128 & 0.0035 & 30 & - & 50 \\ \midrule nCIFAR & 128 & 0.0022 & 965 & - & 50 \\ \midrule Google12 & 128 & 0.00089 & 5 & 0.03 & 60 \\ \midrule ENSO & 16 & 0.01 & 20 & - & 400 \\ \midrule Mackey-Glass system & 16 & 0.01 & 10 & - & 400 \\ \bottomrule \end{tabular}} \end{table} \paragraph{Sensitivity to Random Initialization.} We evaluate our models for each tasks using 8 seeds. The maximum, minimum, average values, and standard deviations obtained for each task are tabulated in Table~\ref{tab:minmax}. \begin{table}[h] \caption{Sensitivity to random initialization evaluated over 8 runs with different seeds.} \label{tab:minmax} \centering \scalebox{0.91}{ \begin{tabular}{l c c c c c c c c} \toprule Task & Maximum & Minimum & Average & standard dev. & d \\ \midrule IMDB & 88.7 & 86.2 & 87.9 & 0.82 & 128 \\ HAR-2 & 97.4 & 96.6 & 96.9 & 0.41 & 128 \\ sMNIST & 99.4 & 99.1 & 99.3 & 0.08 & 128 \\ psMNIST & 97.2 & 96.0 & 96.8 & 0.39 & 128 \\ sCIFAR & 74.9 & 72.65 & 73.54 & 0.90 & 128 \\ nCIFAR & 62.7 & 61.7 & 62.3 & 0.32 & 128 \\ Google12 & 96.2 & 95.7 & 95.9 & 0.17 & 128 \\ \bottomrule \end{tabular}} \end{table} \section{Introduction} Recurrent neural networks (RNNs) and their variants are flexible gradient-based methods specially designed to model sequential data. Models of this type can be viewed as dynamical systems whose temporal evolution is governed by a system of differential equations driven by an external input. Indeed, there is a long-standing tradition to formulate continuous-time variants of RNNs~\cite{pineda1988dynamics}. In this setting, the data are formulated in continuous-time, i.e., inputs are defined by the function ${\bf x} = {\bf x}(t) \in \mathbb{R}^p$ and targets are defined as ${\bf y} = {\bf y}(t) \in \mathbb{R}^q$. In this way, one can, for instance, employ a nonautonomous ordinary differential equation (ODE) to model the dynamics of the hidden states ${\bf h}(t)\in\mathbb{R}^d$, where $t$ denotes continuous time, as \begin{equation*} \frac{d {\bf h}(t)}{d t}\,\,\, = \,\,\, f({\bf h}(t),\, {\bf x}(t);\, \boldsymbol{\theta}). \end{equation*} Here, $f: \mathbb{R}^d\times \mathbb{R}^p \rightarrow \mathbb{R}^d$ is a function which is parameterized by a neural network (NN) with the learnable weights $\boldsymbol{\theta}$. A prototypical choice for $f$ is the $\tanh$ recurrent unit: \begin{equation*} f({\bf h}(t),\, {\bf x}(t); \boldsymbol{\theta}) := \tanh({\bf W} {\bf h}(t) + {\bf U} {\bf x}(t) + {\bf b}), \end{equation*} where ${\bf W} \in \mathbb{R}^{d\times d}$ denotes a hidden-to-hidden weight matrix, ${\bf U}\in \mathbb{R}^{d \times p}$ an input-to-hidden weight matrix, and ${\bf b}$ a bias term. With this continuous-time formulation in hand, one can then use tools from dynamical systems theory to study the dynamical behavior of the model as well as to motivate mechanisms that can prevent rapidly diverging or converging dynamics. For instance,~\cite{chang2018antisymmetricrnn} proposed a parametrization of the hidden-to-hidden matrix as an antisymmetric matrix to ensure stable hidden state dynamics, and~\cite{erichson2020lipschitz} relaxed this idea to improve model expressivity. More recently,~\cite{rusch2022long} has proposed an RNN architecture based on a suitable time-discretization of a set of coupled multiscale ODEs. \begin{figure*}[!t] \begin{CatchyBox}{$\tau$-GRU}\vspace{+0.3cm} \textbf{Continuous-time formulation of $\tau$-GRU:}\\ \begin{minipage}[h]{0.95\linewidth \begin{align \label{eq:tdRNN_de} \frac{d \, {\bf h}(t)}{dt} = \underbrace{u({\bf h}(t),\, {\bf x}(t))}_\text{instantaneous dynamics} \, + \,\,\, \underbrace{\textcolor{wine}{a({\bf h}(t), {\bf x}(t))} \odot \textcolor{darkcyan}{z({\bf h}(t-\tau),\, {\bf x}(t))}}_\text{weighted time-delayed feedback} \,\,\, - \,\,\, {\bf h}(t) \end{align} \end{minipage}\vspace{+0.3cm} \textbf{Discrete-time formulation of $\tau$-GRU:}\\ \begin{minipage}[h]{0.55\linewidth \begin{equation} {\bf h}_{n+1} = (1-{\bf g}_n) \odot {\bf h}_n + {\bf g}_n \odot ({\bf u}_n + \textcolor{wine}{{\bf a}_n} \odot \textcolor{darkcyan}{{\bf z}_n}) \label{eq:ourdRNN} \end{equation} % with % \vspace{-0.3cm} \begin{align} {\bf u}_n & = u({\bf h}_n, {\bf x}_n) := \text{tanh}({\bf W}_1 {\bf h}_n + {\bf U}_1 {\bf x}_n) \label{eq:3} \\ \textcolor{darkcyan}{{\bf z}_n} & = \textcolor{darkcyan}{z({\bf h}_{l}, {\bf x}_n)} := \text{tanh}({\bf W}_2 {\bf h}_{l} + {\bf U}_2 {\bf x}_{n}) \\ {\bf g}_n & = g({\bf h}_n, {\bf x}_n) := \text{sigmoid}({\bf W}_3 {\bf h}_n + {\bf U}_3 {\bf x}_n) \\ \textcolor{wine}{{\bf a}_n} & = \textcolor{wine}{a({\bf h}_n, {\bf x}_n)} :=\text{sigmoid}({\bf W}_4 {\bf h}_{n} + {\bf U}_4 {\bf x}_{n}) \label{eq:6} \end{align} % \hfill \end{minipage} % \hfill \begin{minipage}[h]{.51\linewidth}\small \centering \begin{tabular}{l|c|c} input & ${\bf x}$ & $\mathbb{R}^{p}$\\\hline time index & $t$ & $\mathbb{R}$ \\\hline time delay & $\tau$ & $\mathbb{R}$ \\\hline hidden state & ${\bf h}$ & $\mathbb{R}^{d}$\\\hline hidden-to-hidden matrix & ${\bf W}_i$ & $\mathbb{R}^{d \times d}$\\\hline input-to-hidden matrix & ${\bf U}_i$ & $\mathbb{R}^{d\times p}$\\\hline decoder matrix & ${\bf V}$ & $\mathbb{R}^{q\times d}$\\\hline \end{tabular} \\ \vspace{0.2cm} ${\bf h}_n \approx {\bf h}(t_n)$, $t_n = n \Delta t$, $n=0,1,\dots$ \\ \vspace{0.1cm} $l := n - \lfloor \tau/\Delta t \rfloor$ \end{minipage} \end{CatchyBox} \end{figure*} In this work, we consider using input-driven delay differential equations (DDEs) to model the dynamics of the hidden states: \begin{align*} \frac{d {\bf h}(t)}{d t} \,\,\, = \,\,\, f({\bf h}(t),\, {\bf h}(t-\tau),\, {\bf x}(t);\, \boldsymbol{\theta}), \end{align*} where $\tau$ is a constant that indicates the delay (i.e., time-lag). Here, the time derivative is described by a function $f: \mathbb{R}^d\times \mathbb{R}^d\times \mathbb{R}^p \rightarrow \mathbb{R}^d$ that explicitly depends on states from the past. In prior work~\cite{lin1996learning}, it has been shown that delay units can greatly improve performance on long-term dependency problems \cite{pascanu2013difficulty}, i.e., problems for which the desired model output depends on inputs presented at times far in the past. In more detail, we propose a novel continuous-time recurrent unit, given in Eq.~\eqref{eq:tdRNN_de}, that is composed of two parts: (i) a component $u({\bf h}(t),{\bf x}(t))$ that explicitly models instantaneous dynamics; and (ii) a component $z({\bf h}(t-\tau), {\bf x}(t))$ that provides time-delayed feedback to account for non-instantaneous dynamics. The feedback also helps to propagate gradient information more efficiently, thus lessening the issue of vanishing gradients. In addition, we introduce $a({\bf h}(t),{\bf x}(t))$ to weight the importance of the feedback component-wise, which helps to better model different time scales. By considering a suitable time-discretization scheme of this continuous-time setup, we obtain a gated recurrent unit (GRU), given in Eq. \eqref{eq:ourdRNN}, which we call $\tau$-GRU. The individual parts are described by Eq.~\eqref{eq:3}-\eqref{eq:6}, where ${\bf g}_n$ and ${\bf a}_n$ resemble commonly used gating functions. \paragraph{Main Contributions.} Here are our main contributions. \begin{itemize} \item {\bf Design.} We introduce a novel gated recurrent unit, which we call $\tau$-GRU, that incorporates a weighted time-delay feedback mechanism to lessen the vanishing gradients issue. This model is motivated by DDEs, and it is obtained by discretizing the continuous Eq. \eqref{eq:tdRNN_de}. \item {\bf Theory.} We show that the continuous-time $\tau$-GRU model has a unique well-defined solution (see Theorem~\ref{thm_exist2main}). Moreover, we provide intuition and analysis to understand how the introduction of delays in $\tau$-GRU can act as a buffer to help lessen the vanishing gradients problem, thus improving the ability to retain information far in the past. See Proposition \ref{prop_delaymain} for a simplified setting and Proposition \ref{app_prop} for $\tau$-GRU. \item {\bf Experiments.} We provide empirical results to demonstrate the superior performance of $\tau$-GRU, when compared to other RNN architectures, on a variety of benchmark tasks. In particular, we show that $\tau$-GRU converges faster during training and can achieve improved generalization performance. Moreover, we demonstrate that it provides favorable trade-offs between effectiveness in dealing with long-term dependencies and expressivity in the considered tasks. See Figure \ref{fig:intro_fig} for an illustration of this. \end{itemize} \begin{figure}[!t] \centering \includegraphics[width=0.45\textwidth]{figs/e_vs_l_overview.pdf \caption{Test accuracy for nCIFAR~\cite{chang2018antisymmetricrnn} versus Google-12~\cite{warden2018speech}. nCIFAR requires a recurrent unit with long-term dependency capabilities, while Google-12 requires a highly expressive unit. Our proposed $\tau$-GRU is able to improve performance on both tasks, relative to existing state-of-the-art alternatives, including LEM~\cite{rusch2022long}.} \label{fig:intro_fig} \end{figure} \section{Related Work} Here, we discuss recent RNN advances that have been demonstrated to outperform classic architectures such as Long Short Term Memory (LSTM) networks~\cite{hochreiter1997long} and Gated Recurrent Units (GRUs)~\cite{cho2014properties}. We also briefly discuss prior works on incorporating delays into neural architectures. (We omit a detailed discussion of other recent deep learning models for sequential problems that leverage dynamical system theory:~\cite{erichson2019physics,azencot2020forecasting,gu2021combining,smith2022simplified,hasani2022liquid}.) \\ \noindent \textbf{Unitary and orthogonal RNN.} \ The seminal work by~\cite{arjovsky2016unitary} introduced a recurrent unit where the hidden weight matrix is constructed as the product of unitary matrices. This enforces that the eigenvalues lie on the unit circle. This, in turn, prevents vanishing and exploding gradients, thereby enabling the learning of long-term dependencies. However, such unitary RNNs suffer from limited expressivity, since the construction of the hidden matrix is restrictive~\cite{azencot2021differential}. Work by~\cite{wisdom2016full} and~\cite{vorontsov2017orthogonality} partially addressed this issue by considering the Cayley transform on skew-symmetric matrices; and work by~\cite{lezcano2019cheap,lezcano2019trivializations} leveraged skew-Hermitian matrices to parameterize the orthogonal group to improve expressiveness. The expressiveness of RNNs has been further improved by considering nonnormal hidden matrices~\cite{kerg2019non}. \\ \noindent \textbf{Continuous-time RNNs.} \ The recent work on Neural ODEs~\cite{chen2018neural} and variants~\cite{kidger2020neural,queiruga2021stateful,xia2021heavy,hasani2022closed} have motivated the formulation of several modern continuous-time RNNs, which are expressive and have good long-term memory. The work by~\cite{chang2018antisymmetricrnn} used an antisymmetric matrix to parameterize the hidden-to-hidden matrix in order to obtain stable dynamics. This was later relaxed by~\cite{erichson2020lipschitz}. In~\cite{Kag2020RNNs}, a modified differential equation was considered, which allows one to update the hidden states based on the difference between predicted and previous states. Work by \cite{rusch2021coupled} demonstrated that long-term memory can be improved by modeling the hidden dynamics by a second-order system of ODEs, which models a coupled network of controlled forced and damped nonlinear oscillators. Another approach for improving long-term memory was motivated by a time-adaptive discretization of an ODE~\cite{kag2021time}. The expressiveness of continuous-time RNNs has been further improved by introducing a suitable time-discretization of a set of multiscale ODEs~\cite{rusch2022long}. It has also been shown that noise-injected RNNs can be viewed as discretizations of stochastic differential equations driven by input data~\cite{lim2021noisy}. In this case, the noise can help to stabilize the hidden dynamics during training and improve robustness to input perturbations. \\ \noindent \textbf{Using delays in NNs.} The idea of introducing delays into NNs goes back to \cite{waibel1989phoneme, lang1990time}, where a time-delay NN was proposed to tackle the phoneme recognition problem. Several works followed: \cite{kim1998time} considered a time-delayed RNN model that is suitable for temporal correlations and prediction of chaotic and financial time series; and delays were also incorporated into and studied within the nonlinear autoregressive with exogenenous inputs (NARX) RNNs \cite{lin1996learning}. More recently, \cite{zhu2021neural} introduced delay terms in Neural ODEs and demonstrated their outstanding approximation capacities. In particular, the model of \cite{zhu2021neural} is able to learn delayed dynamics where the trajectories in the lower-dimensional phase space could be mutually intersected, while the Neural ODEs~\cite{chen2018neural} are not able to do so. \section{Method} In this section, we first provide an introduction to DDEs; then, we motivate the formulation of our DDE-based models, in continuous as well as discrete time; and, finally, we propose a weighted time-delay feedback architecture. \paragraph{Notation.} $\odot$ denotes Hadamard product, $| v |$ denotes vector norm for the vector $v$, $\| A \|$ denotes operator norm for the matrix $A$, $\sigma$ and $\hat{\sigma}$ (or sigmoid) denote the $\tanh$ and sigmoid function, respectively; and $\lceil x \rceil$ and $\lfloor x \rfloor$ denote the ceiling function and floor function in $x$, respectively. \subsection{Background on Delay Differential Equations} DDEs are an important class of dynamical systems that arise in natural and engineering sciences~\cite{smith2011introduction, erneux2009applied, keane2017climate}. In these systems, a feedback term is introduced to adjust the system non-instantaneously, resulting in delays in time. In mathematical terms, the derivative of the system state depends explicitly on the past value of the state variable. Here, we focus on DDEs with a single discrete delay \begin{equation} \label{eq_gen_dde} \dot{h} = F(t, h(t), h(t-\tau)), \end{equation} with $\tau > 0$, where $F$ is a continuous function. Due to the presence of the delay term, we need to {\it specify an initial function} which describes the behavior of the system prior to the initial time 0. For the DDE, it would be a function $\phi$ defined on $[-\tau, 0]$. Hence, a DDE numerical solver must save all the information needed to approximate delayed~terms. Instead of thinking the solution of the DDE as consisting of a sequence of values of $h$ at increasing values of $t$, as one would do for ODEs, it is more fruitful to view it as a mapping of functions on the interval $[t -\tau,t]$ into functions on the interval $[t,t + \tau]$, i.e., as a sequence of functions defined over a set of contiguous time intervals of length $\tau$. Since the state of the system at time $t \geq 0$ must contain all the information necessary to determine the solution for future times $s \geq t$, it should contain the initial condition $\phi$. More precisely, the DDE is a functional differential equation with the state space $C := C([-\tau, 0], \mathbb{R}^d)$. This state space is the Banach space of continuous functions from $[-\tau, 0]$ into $\mathbb{R}^d$, with the topology of uniform convergence. It is equipped with the norm $\|\phi \| := \sup \{ |\phi(\theta)| : \theta \in [-\tau, 0] \}$. In contrast to the ODEs (with $\tau = 0)$ whose state space is finite-dimensional, DDEs are generally infinite-dimensional dynamical systems. Various aspects of DDEs have been studied, including their solution properties \cite{hale2013introduction, asl2003analysis}, dynamics \cite{lepri1994high, baldi1994delays} and stability \cite{marcus1989stability, belair1993stability, liao2002delay, yang2014exponential, park2019dynamic}. \subsection{Formulation of Continuous-Time $\tau$-GRUs} The basic form of a time-delayed RNN is \begin{align}\label{eq:simple_model} \begin{split} \dot{{\bf h}} =\sigma({\bf W}_1 {\bf h}(t) + {\bf W}_2 {\bf h}(t-\tau)+ {\bf U} {\bf x}(t) ) - {\bf h}(t), \end{split} \end{align} for $t \geq 0$, and ${\bf h}(t) = 0$ for $t \in [-\tau, 0]$, with the output ${\bf y}(t) = {\bf V} {\bf h}(t)$. In this expression, ${\bf h} \in \RR^d$ denotes the hidden states, $f: \RR^d \times \RR^d \times \RR^{p} \to \RR^d$ is a nonlinear function, and $\sigma: \RR \to (-1,1)$ denotes the tanh activation function applied component-wise. The matrices ${\bf W}_1, {\bf W}_2 \in \RR^{d \times d}$, ${\bf U} \in \RR^{d \times p}$ and ${\bf V} \in \RR^{q \times d}$ are learnable parameters, and $\tau \geq 0$ denotes the discrete time-lag. For notational brevity, we omit the bias term here, assuming it is included in ${\bf W}_1$. It is important to equip RNNs with mechanism to better represent a large number of scales, as discussed by~ \cite{tallec2018can} and more recently by~\cite{rusch2022long}. Therefore, we follow \cite{tallec2018can} and consider a time warping function $c:\mathbb{R}^d\rightarrow \mathbb{R}^d$ which we define to be a parametric function $c(t)$. Using the reasoning in~\cite{tallec2018can}, and denoting $t_\tau := t-\tau$, we can formulate the following continuous-time delay recurrent unit \begin{align*}\label{eq:simple_model} \dot{{\bf h}} = \frac{d c(t)}{d t}\left[ \sigma({\bf W}_1 {\bf h}(t) + {\bf W}_2 {\bf h}(t_\tau) + {\bf U}_1 {\bf x}(t)) - {\bf h}(t)\right]. \end{align*} Now, we need a learnable function to model $\frac{d c(t)}{d t}$. A natural choice is to consider a standard gating function, which is a universal approximator, taking the form \begin{equation} \frac{d c(t)}{d t} = \hat{\sigma}({\bf W}_3 {\bf h}_t + {\bf U}_3 {\bf x}_t) =: {\bf g}(t), \end{equation} where ${\bf W}_3 \in \RR^{d \times d}$ and ${\bf U}_3 \in \RR^{d \times p}$ are learnable parameters, and where $\hat{\sigma}: \RR \to (0,1)$ is the component-wise sigmoid function. \subsection{From Continuous to Discrete-Time $\tau$-GRUs} To learn the weights of the recurrent unit, a numerical integration scheme can be used to discretize the continuous model. Specifically, we discretize the time as $t_n = n \Delta t$ for $n = -\lfloor \tau/\Delta t \rfloor, \dots, -1, 0, 1, \dots$, and approximate the solution $({\bf h}(t))$ to Eq. \eqref{eq:simple_model} by the sequence $({\bf h}_n = {\bf h}(t_n))$, given by ${\bf h}_n = 0$ for $n = -\lfloor \tau/\Delta t \rfloor, \dots, 0$, and \begin{align} {\bf h}_{n+1} &= {\bf h}_n + \int_{t_n}^{t_n+\Delta t} f({\bf h}(s),\, {\bf h}(s-\tau), {\bf x}(s)) \,\mathrm{d}s \\ &\approx {\bf h}_n + \mathtt{scheme}[f,\,{\bf h}_n,\,{\bf h}_{l}, \Delta t], \end{align} for $n=0,1,\dots$, where the subscript $n$ denotes discrete time indices, $l := n - \lfloor \tau/\Delta t \rfloor$, and $\Delta t$ represents the time difference between a pair of consecutive elements in the input sequence. In addition, $\mathtt{scheme}$ refers to a numerical integration scheme whose application yields an approximate solution for the integral. Using the explicit forward Euler scheme and choosing $\Delta t = 1$ gives: \begin{equation} \label{eq_simpleDRNN} {\bf h}_{n+1} = (1-{\bf g}_n) \odot {\bf h}_n + {\bf g}_n \odot \sigma({\bf W}_1 {\bf h}_n + {\bf W}_2 {\bf h}_{l} + {\bf U} {\bf x}_n). \end{equation} Note that this discretization corresponds to the leaky-integrator described by~\cite{jaeger2007optimization}. It can be shown that \eqref{eq_simpleDRNN} is a universal approximator of a large class of open dynamical systems with delay (see Theorem \ref{thm_uat}). However, the performance of this architecture is not able to outperform state-of-the-art RNN architectures on a number of tasks. To improve the performance, we propose a modified gated recurrent unit next. \subsection {Discrete-Time $\tau$-GRUs with a Weighted Time-Delay Feedback Architecture} In this work, we propose to model the hidden dynamics using a mixture of a standard recurrent unit and a delay recurrent unit. To this end, we replace the $\sigma$ in Eq. \eqref{eq_simpleDRNN} by \begin{equation} \label{eq_motivate} {\bf u}_n + {\bf a}_n \odot {\bf z}_n, \end{equation} so that we yield a new GRU that takes the form \begin{equation} {\bf h}_{n+1} = (1-{\bf g}_n) \odot {\bf h}_n + {\bf g}_n \odot \left( {\bf u}_n + {\bf a}_n \odot {\bf z}_n \right). \end{equation} Here ${\bf u}_n$ describes the standard recurrent unit $${\bf u}_n=\text{tanh}({\bf W}_1 {\bf h}_n + {\bf U}_1 {\bf x}_n),$$ and ${\bf z}_n$ describes the delay recurrent unit $${\bf z}_n = \text{tanh}({\bf W}_2 {\bf h}_{l} + {\bf U}_2 {\bf x}_{n}).$$ Further, the gate ${\bf g}_n$ is a learnable vector-valued coefficient $${\bf g}_n =\text{sigmoid}({\bf W}_3 {\bf h}_{n} + {\bf U}_3 {\bf x}_{n}).$$ The weighting term ${\bf a}_n$ is also a vector-valued coefficient $${\bf a}_n =\text{sigmoid}({\bf W}_4 {\bf h}_{n} + {\bf U}_4 {\bf x}_{n}),$$ with the learnable parameters ${\bf W}_4 \in \RR^{d \times d}$ and ${\bf U}_4 \in \RR^{d \times p}$ that weights the importance of the time-delay feedback component-wise for the task on hand. From the design point of view, Eq. \eqref{eq_motivate} can be motivated by the sigmoidal coupling used in Hodgkin-Huxley type neural models (see Eq. (1)-(2) and Eq. (4) in \cite{campbell2007time}). \section{Theory} In this section, we define the notion of solution for DDEs and show that the continuous-time $\tau$-GRU has a unique solution. Moreover, we provide intuition and analysis to understand how the delay mechanism can help improving long-term dependencies. \subsection{Existence and Uniqueness of Solution for Continuous-Time $\tau$-GRUs} Since we must know $h(t + \theta)$, $\theta \in [-\tau, 0]$ in order to determine the solution of the DDE \eqref{eq_gen_dde} for $s > t$, we call the state of the dynamical system at time $t$ the element of $C$ which we denote as $h_t$, defined as $h_t(\theta) := h(t+\theta)$ for $\theta \in [-\tau, 0]$. The trajectory of the solution can thus be viewed as the curve $t \to h_t$ in the state space $C$. In general, DDEs can be formulated as the following initial value problem (IVP) for the nonautonomous system \cite{hale2013introduction}: \begin{equation} \label{eq_genddemain} \dot{h}(t) = f(t, h_t), \ t \geq t_0, \end{equation} where $h_{t_0} = \phi \in C$ for some initial time $t_0 \in \mathbb{R}$, and $f: \mathbb{R} \times C \to \mathbb{R}^d$ is a given continuous function. The above equation describes a general type of systems, including ODEs ($\tau = 0$) and DDEs of the form $\dot{h}(t) = g(t, h(t), h(t-\tau))$ for some continuous function $g$. We say that a function $h$ is a solution of Eq. (\ref{eq_genddemain}) on $[t_0 - \tau, t_0 + A]$ if there exists $t_0 \in \mathbb{R}$ and $A > 0$ such that $h \in C([t_0 - \tau, t_0 + A), \mathbb{R}^d)$, $(t, h_t) \in \mathbb{R} \times C$ and $h(t)$ satisfies Eq. (\ref{eq_genddemain}) for $t \in [t_0, t_0 + A)$. It can be shown that (see, for instance, Lemma 1.1 in \cite{hale2013introduction}) if $f(t,\phi)$ is continuous, then finding a solution of Eq. (\ref{eq_genddemain}) is equivalent to solving the integral equation: $h_{t_0} = \phi$, \begin{equation} h(t) = \phi(0) + \int_{t_0}^t f(s, h_s) \ \rm{d}s, \ t \geq t_0. \end{equation} We now provide existence and uniqueness result for the continuous-time $\tau$-GRU model, assuming that the input $x$ is continuous in $t$. Defining the state $h_t \in C$ as $h_t(\theta) := h(t+\theta)$ for $\theta \in [-\tau, 0]$ as before, the DDE describing the $\tau$-GRU model can be formulated as the following IVP: \begin{equation} \label{eq_gendde} \dot{h} = - h(t) + u(t, h(t)) + a(t, h(t)) \odot z(t, h_t), \ t \geq t_0, \end{equation} and $h_{t_0} = \phi \in C$ for some initial time $t_0 \in \mathbb{R}$, with the dependence on $x(t)$ viewed as dependence on $t$. By applying Theorem 3.7 in \cite{smith2011introduction}, we can obtain the following result. See App.~\ref{sect:appB} for a proof of this theorem. \begin{theorem}[Existence and uniqueness of solution for continuous-time $\tau$-GRU] \label{thm_exist2main} Let $t_0 \in \RR$ and $\phi \in C$ be given. There exists a unique solution $h(t) = h(t, \phi)$ of Eq. \eqref{eq:tdRNN_de}, defined on $[t_0 - \tau, t_0 + A]$ for any $A > 0$. In particular, the solution exists for all $t \geq t_0$, and \begin{equation} \| h_t(\phi) - h_t(\psi) \| \leq \| \phi - \psi \| e^{K(t-t_0)}, \end{equation} for all $t \geq t_0$, where $K = 1 + \|W_1\| + \|W_2\| + \|W_4\|/4$. \end{theorem} Theorem \ref{thm_exist2main} guarantees that the continuous-time $\tau$-GRU model, as a functional differential equation, has a well-defined unique solution that does not blow up in finite time. \subsection{The Delay Mechanism in $\tau$-GRUs Can Help Improve Long-Term Dependencies} RNNs suffer from the problem of vanishing and exploding gradients, leading to the problem of long-term dependencies. While the gating mechanisms could mitigate the problem to some extent, the delays introduced in $\tau$-GRUs can further help reduce the sensitivity to long-term dependencies. To understand the reason for this, we consider how gradients are computed using the backpropagation through time (BPTT) algorithm~\cite{pascanu2013difficulty}. BPTT involves the two stages of unfolding the network in time and backpropagating the training error through the unfolded network. When $\tau$-GRUs are unfolded in time, the delays in the hidden state will appear as jump-ahead connections (buffers) in the unfolded network. These buffers provide a shorter path for propagating gradient information, and therefore reducing the sensitivity of the network to long-term dependencies. Such intuition is also used to explain the behavior of the NARX RNNs in \cite{lin1996learning}. We now make this intuition precise in the following simplified setting. See App.~\ref{sect:appC} for a proof of this proposition. We also provide analogous results (bounds for the gradient norm) and discussions for $\tau$-GRU (in App.~\ref{app:gradbound}, see Proposition \ref{app_prop}). \begin{proposition} \label{prop_delaymain} Consider the linear time-delayed RNN whose hidden states are described by the update equation: \begin{equation} h_{n+1} = Ah_n + Bh_{n-m} + Cu_n, \ n=0,1,\dots, \end{equation} and $h_n = 0$ for $n=-m, -m+1, \dots, 0$ with $m > 0$. Then, assuming that $A$ and $B$ commute, we have: \begin{equation} \frac{\partial h_{n+1}}{\partial u_i} = A^{n-i} C, \end{equation} for $n=0,1, \dots, m$, $i=0,\dots, n$, and \begin{align} \frac{\partial h_{m+1+j}}{\partial u_i} &= A^{m+j-i} C + \delta_{i,j-1} BC + 2 \delta_{i,j-2} ABC \nonumber \\ &\ \ \ \ + 3 \delta_{i,j-3} A^2 B C + \cdots + j \delta_{i,0} A^{j-1} B C, \end{align} for $j = 1,2,\dots, m+1$, $i=0,1,\dots, m+j$, where $\delta_{i,j}$ denotes the Kronecker delta. \end{proposition} We remark that the commutativity assumption is not necessary. It is used here only to simplify the expression for the gradients. Analogous formula for the gradients can be derived without such assumption, at the cost of displaying more complicated formulae. From Proposition \ref{prop_delaymain}, we see that the presence of the delay allows the model to place more emphasis on the gradients due to input information in the past (as can be seen in Eq. \eqref{eq_linear} in the proof of the proposition, where additional terms dependent on $B$ appear in the coefficients in front of the past inputs). In particular, if $\|A\| < 1$ and $B=0$, then the gradients decay exponentially as $i$ becomes large. Introducing the delay term ($B\neq 0$) makes the rates at which this exponential decays lower by perturbing the gradients of hidden states (dependent on the delay parameter $m$) with respect to the past inputs with nonzero values. Similar qualitative conclusion can also be drawn for our $\tau$-GRU (see Proposition \ref{app_prop} and the discussions in App. \ref{app:gradbound}). Therefore, one expects that these networks would be able to more effectively deal with long-term dependencies than the counterpart models without delays. \section{Experimental Results} \label{sect:exp} In this section, we consider several benchmark datasets to demonstrate the performance of our proposed $\tau$-GRU. We use standard protocols for training and initialization, and validation sets for parameter tuning (see App.~\ref{sect:appD} for~details). \begin{figure}[!b] \centering \begin{subfigure}[b]{0.48\textwidth} \centering \includegraphics[width=0.99\textwidth]{figs/adding_task_2000.pdf} \caption{Sequence length $N=2000$.} \label{fig:adding_task_2000} \end{subfigure ~ \begin{subfigure}[b]{0.48\textwidth} \centering \includegraphics[width=0.99\textwidth]{figs/adding_task_5000.pdf} \caption{Sequence length $N=5000$.} \label{fig:adding_task_5000} \end{subfigure} \caption{Results for the adding task on very long sequences.} \label{fig:adding_task} \end{figure \subsection{The Adding Task} The adding task is a classic problem for testing the ability of models to learn long-term dependencies, originally proposed by~\cite{hochreiter1997long}. The inputs of this problem are composed of two stacked random vectors $u$ and $v$ of length $N$. The elements of $u$ are drawn from the uniform distribution $\mathcal{U}(0,1)$. The vector $v$ has two non-zero elements (both set to 1), one placed in a random location $i$ sampled from the index set $i\in\{1,\dots,\lfloor\frac{N}{2}\rfloor\}$, and the other element is placed at location $j$ sampled from the index set $j\in\{\lceil\frac{N}{2}\rceil,\dots,N\}$. The target value for each sequence is constructed as the sum $\sum(u\odot v)$, i.e., the sum of the two elements in $u$ that correspond to the the non-negative entries $v$. Here, we follow the work by~\cite{rusch2022long} and consider two challenging settings with very long input sequences $N=\{2000,5000\}$. Figure~\ref{fig:adding_task} shows results for our $\tau$-GRU and several state-of-the-art RNN models which are designed to solve long-term dependencies tasks, such as LEM~\cite{rusch2022long}, coRNN~\cite{rusch2021coupled}, DTRIV$_\infty$~\cite{lezcano2019trivializations}, fastGRNN~\cite{kusupati2018fastgrnn}, and LSTM with chrono initialization~\cite{tallec2018can}. It can be seen that the DTRIV$_\infty$ and fastGRNN are performing poorly in both cases. In contrast, our $\tau$-GRU shows a favorable performance, i.e., $\tau$-GRU is converging faster, and it achieves lower mean squared errors than other methods. \subsection{Sentiment Analysis: IMDB} Next, we consider the IMDB dataset~\cite{maas2011learning} to study the expressiveness of our proposed model on a sentiment analysis task. The aim of this task is to predict whether a movie review is positive or negative. This dataset is composed of $50$k movie reviews, with an average length of 240 words per review. Each review is annotated by a label that indicates a positive, or negative sentiment. The dataset is split evenly into a training and test set, so that both sets contain $25$k reviews. We use 15\% of the training data for validation. Further, we use standard preprocessing schemes, following ~\cite{pennington2014glove}, to restrict the dictionary to $25$k words, and to embed the data with a pretrained GloVe model~\cite{pennington2014glove}. Table~\ref{tab:results_imbd} shows that our $\tau$-GRU achieves a substantially higher test accuracy than LSTM, and GRU models. Further, $\tau$-GRU is also able to outperform the continuous-time coRNN~\cite{rusch2021coupled} and the highly expressive LEM~\cite{rusch2022long}. \begin{table}[!h] \caption{Results for IMDB sentiment analysis task.} \label{tab:results_imbd} \centering \scalebox{0.9}{ \begin{tabular}{l c c ccccccc} \toprule Model & Test Accuracy ($\%$) & \# units & \# param \\ \midrule LSTM~\cite{campos2018skip} & 86.8 & 128 & {220k} \\ Skip LSTM~\cite{campos2018skip} & 86.6 & 128 & {220k} \\ GRU~\cite{campos2018skip} & 86.2 & 128 & 164k \\ Skip GRU~\cite{campos2018skip} & 86.6 & 128 & 164k \\ ReLU GRU~\cite{dey2017gate} & 84.8 & 128 & 99k \\ coRNN~\cite{rusch2021coupled} & 87.4 & 128 & 46k \\ LEM & 88.1 & 128 & {220k} \\ \midrule \textbf{$\tau$-GRU (ours)} & \textbf{88.6} & 128 & {220k} \\ \bottomrule \end{tabular}} \end{table} \begin{table}[!b] \caption{Results for HAR2 task.} \label{tab:results_har2} \centering \scalebox{0.9}{ \begin{tabular}{l c c ccccccc} \toprule Model & Test Accuracy ($\%$) & \# units & \# param \\ \midrule GRU~\cite{kusupati2018fastgrnn} & 93.6 & 75 & {19k}\\ LSTM~\cite{Kag2020RNNs} & 93.7 & 64 & {19k}\\ FastRNN~\cite{kusupati2018fastgrnn} & 94.5 & 80 & 7k\\ FastGRNN~\cite{kusupati2018fastgrnn} & 95.6 & 80 & 7k\\ AsymRNN~\cite{Kag2020RNNs} & 93.2 & 120 & 8k\\ iRNN~\cite{Kag2020RNNs} & 96.4 & 64 & 4k\\ DIRNN~\cite{zhang2021deep} & 96.5 & 64 & -\\ coRNN~\cite{rusch2021coupled} & 97.2 & 64 & 9k\\ LipschitzRNN & 95.4 & 64 & 9k\\ LEM & 97.1 & 64 & {19k}\\ \midrule \textbf{$\tau$-GRU (ours)} & \textbf{97.4} & 64 & {19k} \\ \bottomrule \end{tabular}} \end{table} \begin{table*}[!t] \caption{Test accuracies on sMNIST, and psMNIST.} \label{tab:image_mnist} \centering \scalebox{0.85}{ \begin{tabular}{lcccc|cccc} \toprule Model& sMNIST & psMNIST & \# units & \# parameters & sCIFAR & nCIFAR & \# units & \# parameters \\ \midrule LSTM~\cite{kag2021time} & 97.8 & 92.6 & 128 & {68k} & 59.7 & 11.6 & 128 & 69k / 117k \\ r-LSTM~\cite{trinh2018learning} & 98.4 & 95.2 & - & 100K & 72.2 & - & - & 101k / -\\ chrono-LSTM~\cite{rusch2022long} & 98.9 & 94.6 & 128 & {68k} & - & 55.9 & 128 & - / 116k\\ Antisymmetric RNN~\cite{chang2018antisymmetricrnn} & 98.0 & 95.8 & 128 & 10k & 62.2 & 54.7 & 256 & 37k / 37k\\ Lipschitz RNN~\cite{erichson2020lipschitz} & 99.4 & 96.3 & 128 & 34k & 64.2 & 59.0 & 256 & 134k / 158k\\ expRNN~\cite{lezcano2019cheap} & 98.4 & 96.2& 360 & {68k} & - & 49.0 & 128 & - / 47k\\ iRNN~\cite{Kag2020RNNs} & 98.1 & 95.6 & 128 & 8k & - & 54.5 & 128 & - / 12k \\ TARNN~\cite{kag2021time} & 98.9 & 97.1 & 128 & {68k} & - & 59.1 & 128 & - / 100K \\ coRNN~\cite{rusch2021coupled} & 99.3 & {96.6} & 128 & 34k & - & 59.0 & 128 & - / 46k\\ {LEM}~\cite{rusch2022long} & \textbf{99.5} & {96.6} & 128 & {68k}& - & 60.5 & 128 & - / 117k\\ \midrule Simple delay GRU (Eq.~\eqref{eq_simpleDRNN}) & 98.7 & 94.1 & 128 & 51k & 57.2 & 59.8 & 128 & 52k / 75k \\ \textbf{$\tau$-GRU (ours)} & {99.4} & \textbf{97.3} & 128 & {68k} & \textbf{74.9} & \textbf{62.2} & 128 & 69k / 117k \\ \bottomrule \end{tabular}} \end{table*} \subsection{Human Activity Recognition: HAR-2} Here, we study the performance of our model for human activity recognition using the HAR dataset provided by~\cite{anguita2012human}. This dataset consists of measurements by an accelerometer and gyroscope on a Samsung Galaxy S3 smartphone which are tracking six activities for 30 volunteers within an age bracket of 19-48 years. For learning, the sequences are divided into shorter sequences of length $N=128$, and the raw measurements are summarized by 9 features per time step. Kusupati et al.~\cite{kusupati2018fastgrnn} proposed the HAR-2 dataset which groups the activities into two categories \{Sitting, Laying, Walking Upstairs\} and \{Standing, Walking, Walking Downstairs\}. We use $7,352$ sequences for training, $900$ for validation and $2,947$ for testing. Table~\ref{tab:results_har2} shows that our $\tau$-GRU is able to outperform traditional gated architectures on this task. The most competitive model is coRNN~\cite{rusch2021coupled}, which achieves a test accuracy of $97.2\%$ with just $9$k parameters. LEM~\cite{rusch2022long} achieves $97.1\%$ with the the same number of parameters as our $\tau$-GRU. \subsection{Sequential Image Classification} Next, we consider four sequential variations of the MNIST~\cite{lecun1998gradient} and CIFAR-10~\cite{CIFAR10} image classification datasets. These sequence classification tasks aim to evaluate the capability of RNNs to learn long-term dependencies. The sequential MNIST (sMNIST) task, originally proposed by~\cite{le2015simple}, is sequentially presenting the $N=784$ pixels of each thumbnail to the recurrent unit. The final hidden state is used to predict the class membership probability of the flattened image. A variation of the sMNIST task is the permuted sMNIST task (psMNIST), which presents the model with a fixed random permutation of the pixel-by-pixel sequence. This task removes any natural patterns in the sequence and requires that models can learn long-term dependencies between pixels that are possibly far apart. Since the standard sMNIST task has essentially been solved by state-of-the-art RNNs, \cite{chang2018antisymmetricrnn} has proposed to consider the sequential CIFAR-10 (sCIFAR) task instead. This task is more challenging due to the increased sequence length, $N=1024$, of the flattened input images. Each element of the sequence is a 3-dimensional vector that contains the pixels for each color channel. To solve this task, models require to have long-term memory and sufficient~expressivity. \begin{figure}[!b] \centering \includegraphics[width=0.48\textwidth]{figs/psmnist_acc.pdf} \caption{Test accuracy on permuted sequential MNIST as a function of the number of epochs. Our $\tau$-GRU requires substantially fewer number of epochs to reach peak performance.} \label{fig:psmnist_acc} \end{figure} Furthermore, \cite{chang2018antisymmetricrnn} has also proposed a noise-padded CIFAR-10 (nCIFAR) task, which requires that the RNN has the ability to memorize information from far in the past, and be able to suppress noisy segments that contain no relevant information. Specifically, we construct a sequence of length $N=1000$ where each element is a $96$-dimensional vector. The first $32$ two elements are the rows of an image, where the channels are stacked. The remaining $968$ elements of the sequence are random vectors drawn from the standard normal distribution. Table~\ref{tab:image_mnist} shows results for the four described tasks. Our $\tau$-GRU is competitive and able to outperform other models on the psMNIST, sCIFAR and nCIFAR task. The proposed weighted time-delay feedback mechanisms demonstrates a clear advantage in particular for the CIFAR tasks. Figure~\ref{fig:psmnist_acc} shows that our $\tau$-GRU is converging faster than other recently introduced continuous-time models, such as LEM~\cite{rusch2022long}, coRNN~\cite{rusch2021coupled}, and LipschitzRNN~\cite{erichson2020lipschitz}. This could help reduce training times when using these~methods. \subsection{Ablation Study using psMNIST} We use the psMNIST task to perform an ablation study. To do so, we consider the following model $${\bf h}_{t+1} = (1-{\bf g}_t) \odot {\bf h}_t + {\bf g}_t \odot (\beta \cdot {\bf u}_t + \alpha \cdot {\bf a}_t \odot {\bf z}_t)$$ where $\alpha\in [0,1]$ and $\beta\in [0,1]$ are constants that can be used to control the effect of different components. We are interested in the cases where $\alpha$ and $\beta$ are either 0 or 1, i.e., a component is switched off or on. Table~\ref{table:ablation} shows the results for different ablation configurations. By setting $\alpha=0$ we yield a simple gated RNN. Second, for $\beta=0$, we yield a $\tau$-GRU without instantaneous dynamics. Third, we show how different values of $\tau$ affect the performance. Setting $\tau=0$ leads to a $\tau$-GRU without time-delay feedback. We also show that a model without the weighting function ${\bf a}_t$ is not able to achieve peak performance. \begin{table}[!h] \caption{Ablation study.} \label{table:ablation} \centering \scalebox{0.9}{ \begin{tabular}{l c c ccccccc} \toprule Model & $\alpha$ & $\beta$ & $\tau$ & ${\bf a}_t$ & Test Accuracy (\%) \\ \midrule ablation & 0 & 1 & - & yes & 94.6 \\ ablation & 1 & 0 & 65 & yes &94.9 \\ \midrule ablation & 1 & 1 & 0 & yes & 95.1 \\ ablation & 1 & 1 & 20 & yes & 96.4 \\ ablation & 1 & 1 & 65 & no & 96.8 \\ \midrule \textbf{$\tau$-GRU (ours)} & 1 & 1 & 65 & yes &\textbf{97.3}\\ \bottomrule \end{tabular}} \end{table} \begin{figure}[!h] \centering \includegraphics[width=0.48\textwidth]{figs/tau_ablation.pdf} \caption{Sensitivity analysis of $\tau$-GRU on psMNIST. The light green envelopes represent $\pm 1$ standard deviation around the mean.} \label{fig:tau_ablation_main} \end{figure} Figure~\ref{fig:tau_ablation_main} further investigates the effect of $\tau$. The performance of $\tau$-GRU is increasing as a function of $\tau$ and peaking around $\tau=65$. It can be seen that the performance in the range $50-150$ is relatively constant for this task. Thus, the model is relatively insensitive as long as $\tau$ is sufficiently large, but not too large. The performance is starting to decrease for $\tau>150$. Since $\tau$ takes discrete values, tuning is easier as compared to continuous tuning parameters, for instance, parameters used by LEM~\cite{rusch2022long}, coRNN~\cite{rusch2021coupled}, or LipschitzRNN~\cite{erichson2020lipschitz}. \subsection{Learning Climate Dynamics} We consider the task of learning the dynamics of the DDE model for El Ni\~{n}o Southern Oscillation (ENSO) of \cite{falkena2019derivation} (see Eq. (46) there). It models the sea surface temperature $T$ in the eastern Pacific Ocean, and is described by: \begin{equation} \dot{T} = T-T^3 - c T(t-\delta) (1-\gamma T^2(t-\delta)), \ t > \delta, \end{equation} where $\gamma < 1$, $c > 0$, with $T(t)$ satisfying $\dot{T} = T-T^3 - c T(0) (1-\gamma T(0)^2)$ with $T(0) \sim \rm{Unif}(0,1)$ for $t \in [0, \delta]$. For data generation, we follow \cite{falkena2019derivation}, and choose $c = 0.93$, $\gamma = 0.49$ and $\delta = 4.8$. We use the classical Runge-Kutta method (RK4) to numerically integrate the system from $t=0$ to $t=400$ using a step-size of 0.1. The training and testing samples are the time series (of length 2000) generated by the RK4 scheme on the interval $[200,400]$ for different realizations of $T(0)$. Table \ref{tab:results_ENSO} shows that our model (with $\alpha = \beta = 1$ and using $\tau = 20$) is more effective in learning the ENSO dynamics when compared to other RNN architectures. We also see that the predictive performance deteriorates without using appropriate combination of standard and delay recurrent units (setting either $\alpha$ or $\beta$ to zero). \begin{table}[!h] \caption{Results for the ENSO model prediction.} \label{tab:results_ENSO} \centering \scalebox{0.9}{ \begin{tabular}{l c c ccccccc} \toprule Model & MSE ($\times 10^{-2}$) & \# units & \# parameters \\ \midrule Vanilla RNN & 0.45 & 16 & 0.3k \\ LSTM & 0.92 & 16 & 1.2k \\ GRU & 0.53 & 16 & 0.9k \\ Lipschitz RNN & 10.6 & 16 & 0.6k \\ coRNN & 4.00 & 16 & 0.6k \\ LEM & 0.31 & 16 & 1.2k \\ \midrule ablation ($\alpha = 0$) & 0.31 & 16 & 0.6k \\ ablation ($\beta = 0$) & 0.38 & 16 & 0.9k \\ \midrule \textbf{$\tau$-GRU (ours)} & \textbf{0.17} & 16 & 1.2k \\ \bottomrule \end{tabular}} \end{table} \section{Conclusion and Future Work} Starting from a continuous-time formulation, we derive a discrete-time gated recurrent unit with delay, $\tau$-GRU. We also provide intuition and analysis to understand how the delay can act as a buffer to improve modeling of long-term dependencies. Importantly, we demonstrate the superior performance of $\tau$-GRU in several challenging sequential learning tasks. We now discuss some future directions. On the one hand, using multiple delays could lead to improved models, and therefore it is of interest to study the extension of $\tau$-GRU to include suitable distributed delay mechanism. On the other hand, achieving model robustness with respect to adversarial perturbations and common corruptions is critical for sensitive applications, but it is largely unexplored in the sequential learning setting. It has been shown in \cite{lim2021noisy, lim2021nfm, erichson2022noisymix} that noise injection can be effective in improving robustness. Therefore, it is also of interest to consider and study noise-injected versions of $\tau$-GRU to improve the trade-offs between accuracy and robustness. \section*{Acknowledgments} N. B. Erichson would like to acknowledge IARPA (contract W911NF20C0035). S. H. Lim would like to acknowledge the WINQ Fellowship, the Swedish Research Council (VR/2021-03648), and the computational resources provided by the Swedish National Infrastructure for Computing (SNIC) at Chalmers Centre for Computational Science and Engineering (C3SE) partially funded by the Swedish Research Council through grant agreement no. 2018-05973. M. W. Mahoney would also like to acknowledge NSF, and ONR for providing partial support of this work. Our conclusions do not necessarily reflect the position or the policy of our sponsors, and no official endorsement should be inferred. {\normalsize \bibliographystyle{ieee_fullname}

https://arxiv.org/abs/1911.03012

Counting extensions revisited

We consider rooted subgraphs in random graphs, i.e., extension counts such as (i) the number of triangles containing a given vertex or (ii) the number of paths of length three connecting two given vertices. In 1989, Spencer gave sufficient conditions for the event that, with high probability, these extension counts are asymptotically equal for all choices of the root vertices. For the important strictly balanced case, Spencer also raised the fundamental question as to whether these conditions are necessary. We answer this question by a careful second moment argument, and discuss some intriguing problems that remain open.

\section{Introduction} Subgraph counts and their many natural generalizations are central topics in random graph theory: since the~1960's they are a constant source of beautiful problems and~conjectures, which have repeatedly inspired the development of important new probabilistic~techniques and~insights (see~\cite{BB,AS,JLR,FK}). In this paper we consider \mbox{rooted subgraph counts} in the binomial random graph~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$, i.e., so-called \linebreak[4] \mbox{extension counts}~\cite{SS1988,S90b,LS1991,Vu2001} such as (i)~the number of triangles containing a given~vertex or (ii)~the number of paths of length three connecting two given~vertices. In combinatorics and related areas, the need for studying such extension~counts arises frequently in probabilistic proofs and applications, including zero-one~laws in random graphs~\cite{SS1988,LS1991,S2001}, games on random graphs~\cite{LP2010,N2017}, random graph processes~\cite{bohman2010early,BFL2015, bohman2013dynamic, pontiveros2013triangle, BW2018}, sparse random analogues of classical extremal and Ramsey results~\cite{NS2017,SS2018,BK2019}, and many more, such as~\cite{S90a,R92,Vu2001,ST2002,YR2007,JR11,spohel2013general,M2015,TBD}. Consequently the investigation of extension~counts is not only a natural problem in probabilistic combinatorics, but also an important issue from the applications point of~view. After initial groundwork of Shelah and Spencer~\cite{SS1988} as well as Spencer~\cite{S90a} on (rooted~subgraph) extension~counts, in~1989 Spencer~\cite{S90b} proved sufficient conditions for the event that, with high probability\footnote{As usual, we say that an event holds~\emph{whp} (with~high~probability) if it holds with probability tending to~$1$ as~$n\to \infty$.}, these extension counts are asymptotically equal in~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$ for all choices of the root~vertices. For the important strictly balanced case, he also raised the fundamental question whether these sufficient conditions (see~\eqref{eq_Spencer_SB} below) are qualitatively necessary. In this paper we answer Spencer's 30-year old question by a careful second moment argument, see~Theorem~\ref{thm_strictly_balanced} below, rectifying a surprising gap in the random graph literature. We also discuss some further partial results and intriguing open~problems (see~Sections~\ref{ss_partial}--\ref{sec:intro:general} below). \subsection{Main result}\label{sec:intro:main} To fix notation, by a \emph{rooted graph}~${(G,H)}$ we mean a graph~${H=(V(H),E(H))}$ and an induced subgraph~${G \subseteq H}$ with labeled `root' vertices~${V(G)=\{1, \ldots, v_G\}}$. Given a tuple~${\xx =(x_1, \ldots, x_{v_G})}$ consisting of distinct vertices, a~{\emph{$(G,H)$-extension} of $\xx$ is a copy of the graph~$H_G:= (V(H), E(H)\setminus E(G))$ in which each vertex~$j \in V(G)$ is mapped onto~$x_j$. Note that if~$\xx$ spans a copy of~$G$ in which each vertex~$j \in V(G)$ is mapped onto~$x_j$, then every~$(G,H)$-extension of~$\xx$ corresponds to a copy of~$H$. Since the edges between root vertices do not affect the definition of a~$(G,H)$-extension, the reader may without loss of generality assume~$V(G)$ is an independent set~of~$H$ in the results below, cf.~\cite{JLR,JR11} (allowing for~$G$ that are not independent will be convenient in some proofs, though). For brevity, we write~$\oset{v_G}$ for the set of all \emph{roots}, i.e., tuples~$\xx = (x_1, \dots, x_{v_G})$ of distinct vertices from~$[n] := \left\{ 1, \dots, n \right\}$. Let~$X_{\xx} = X_{G,H}(\xx)$ denote the number of~$(G,H)$-extensions of~$\xx$ in the binomial random graph~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$. Note that the expected~value \begin{equation}\label{def:muGH} \mu=\mu_{G,H} := \E X_{\xx} \asymp n^{v_H-v_G}p^{e_H-e_G} \end{equation} does not depend on the particular choice of~$\xx$. To avoid trivialities, we henceforth assume that~$H$ has more edges than~$G$, i.e., that~$e_H>e_G$. Similarly as for (unrooted) subgraph counts, we define \begin{equation}\label{def:mGH} m(G,H) := \max_{G \subsetneq J \subseteq H} d(G,J) \quad \text{ with } \quad d(G,J):=\frac{e_J-e_G}{v_J-v_G}, \end{equation} and say that~$(G,H)$ is \emph{strictly~balanced} if~$d(G,J) < d(G,H)$ for all~$G \subsetneq J \subsetneq H$. We also call~$(G,H)$~\emph{grounded} if at least one root vertex~$j \in V(G)$ is connected to a non-root vertex~$w \in V(H) \setminus V(G)$. Spencer derived in~1989 sufficient conditions for the event that, with high probability, all extension counts satisfy~$X_{\xx} \sim \mu$, i.e., are asymptotically equal. In the important case when~$(G,H)$ is strictly balanced, \mbox{\cite[Theorem~2]{S90b}} states that for every fixed~$\eps \in (0,1]$ there is a constant~$K(\eps) > 0$ such that \begin{equation}\label{eq_Spencer_SB} \lim_{n \to \infty} \Pr\Bigpar{\max_{\xx \in \oset{v_G}}|X_\xx - \mu| < \eps\mu} = 1 \quad \text{ if } \mu \ge K(\eps) \log n. \end{equation} Spencer remarked that in~\eqref{eq_Spencer_SB} his constant satisfies~$K(\eps) \to \infty$ as~$\eps \to 0$, and speculated that this is probably also necessary, see~\cite[Remark~on~p.249]{S90b}. In other words, he raised the question whether his sufficient condition is qualitatively best possible. Our main result answers this fundamental question: \eqref{eq:main:strbal}~shows that the `correct' dependence is~$K(\eps)=\Theta(\eps^{-2})$ in the grounded case, even when~$\eps=\eps(n) \to 0$ at some polynomial~rate. For completeness, \eqref{eq:main:strbal:non}~also shows that the logarithm in the sufficient condition~\eqref{eq_Spencer_SB} is~unnecessary in the less interesting ungrounded~case (where extension counts are essentially unrooted subgraph counts, cf.~example~(b) in~\refF{fig_primal}). \begin{theorem}[Main result: strictly balanced case]\label{thm_strictly_balanced Let~$(G,H)$ be a rooted graph that is strictly balanced. There are constants~$c, C, \alpha > 0$ such that, for all~$p=p(n) \in [0,1]$ and~$\eps=\eps(n) \in [n^{-\alpha},1]$, the following~holds: \begin{romenumerate2} \ite If the rooted graph~$(G,H)$ is grounded, then \begin{align} \label{eq:main:strbal} \vspace{-0.125em} \lim_{n \to \infty} \Pr\Bigpar{\max_{\xx \in \oset{v_G}}|X_\xx - \mu| < \eps\mu} &= \begin{cases} 0 & \text{if $\eps^2\mu \le c \log n$,} \\ 1 & \text{if $\eps^2\mu \ge C \log n$.} \end{cases}\hspace{2.5em}\vspace{-0.125em} \intertext{\ite If the rooted graph~$(G,H)$ is not grounded, then} \label{eq:main:strbal:non} \vspace{-0.25em} \lim_{n \to \infty} \Pr\Bigpar{\max_{\xx \in \oset{v_G}}|X_\xx - \mu| < \eps\mu} &= \begin{cases} 0 & \text{if $\eps^2\mu \to 0$,} \\ 1 & \text{if $\eps^2\mu \to \infty$.} \end{cases}\hspace{2.5em}\vspace{-0.125em}% \end{align}% \end{romenumerate2}\vspace{-0.125em}% \end{theorem} \noindent In concrete words, \eqref{eq:main:strbal}--\eqref{eq:main:strbal:non} of~\refT{thm_strictly_balanced} give thresholds for the concentration of extension counts in terms of~$\eps^2\mu$, similar to the thresholds in terms of the edge probability~$p$ that are well-known for many properties of~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$. The role of the expression~$\eps^2 \mu$ in~\eqref{eq:main:strbal}--\eqref{eq:main:strbal:non} can heuristically be explained via Chernoff-type bounds of form~$\e^{-\Omega(\eps^2\mu)}$ on the tails~$\Pr(|X_\xx - \mu| \ge \eps \mu)$ of~$X_\xx$. Indeed, considering the union bound over the~$\Theta(n^{v_G})$ roots~$\xx$, it then seems plausible that the $1$-statement follows when~$\eps^2\mu$ is at least a large enough multiple of~$\log n$. An intuitive reason why the~$\log n$ factor is absent in the ungrounded threshold~\eqref{eq:main:strbal:non} is that here the~$X_{\xx}$ are strongly correlated and in fact almost equal (e.g., in example~(b) from~\refF{fig_primal} each~$X_{\xx}$ is well-approximately by the total number of triangles), so there should be no need to use a union~bound. The main contribution of \refT{thm_strictly_balanced} is the $0$-statement in the grounded threshold~\eqref{eq:main:strbal}, which was missing in previous work: our proof uses a careful second moment argument (combining correlation inequalities and counting arguments with Janson's inequality) in order to establish that, with high probability, there exists a root~$\xx$ with~${X_\xx \ge (1+\eps)\mu}$, i.e., with too~many $(G,H)$-extensions. This is closely related to the task of obtaining good lower bounds on ${\Pr(X_\xx \ge (1+\eps)\mu)}$, which are not so well understood as upper bounds; see~\cite{JR2002,JW,Ch19,SW18}. To sidestep this conceptual obstacle, in \refS{s_strictly_balanced} we therefore work with (easier to estimate) auxiliary events that enforce~${X_\xx \ge (1+\eps)\mu}$ via `disjoint' extensions, and we believe that our approach might also be useful for establishing `lower bounds' in other~problems. \pagebreak[3] \subsection{Partial results: beyond the strictly balanced case}\label{ss_partial} We also establish some threshold results for extension counts of rooted graphs~$(G,H)$ that are not necessarily strictly balanced. Here things are more complicated, since we now need to take into account all subgraphs of~$J \subseteq H$ containing the root~$G$, in particular those that satisfy~$d(G,J) = m(G,H)$; cf.~\cite{S90a,S90b,R92,JLR}. We call such subgraphs~$J$~\emph{primal}, and for brevity also say that~$J$~is~\emph{grounded} if~$(G,J)$~is~grounded. The partial results Theorems~\ref{thm_unique}--\ref{thm_nogrounded} below cover all strictly balanced~$(G,H)$, and they in particular imply that \refT{thm_strictly_balanced} also holds with~$\eps^2\Phi$ instead of~$\eps^2\mu$ (possibly after modifying the constants~$c,C,\alpha$), where \begin{equation}\label{eq_PhiGH} \Phi = \Phi_{G,H} := \min_{G \subseteq J \subseteq H : e_J > e_G} \mu_{G,J}. \end{equation} There is no contradiction here: the extra assumption~$\eps \ge n^{-\alpha}$ ensures that the conclusions of the {$0$-}~and {$1$-statements} of \refT{thm_strictly_balanced} coincide regardless of whether we use~$\eps^2\Phi$ or~$\eps^2\mu$ (cf.~\refS{sec:ext:non}). It thus comes at no surprise that in our main result \refT{thm_strictly_balanced} the technical assumption~$\eps \ge n^{-\alpha}$ is indeed\footnote{For examples~(a) and~(b) from \refF{fig_primal} with~$\eps \asymp n^{-1/2}$ and~$\eps \asymp n^{-1}$, when~$p \asymp n^{-1/4}$ it is routine to check that~$\Phi \to \infty$, $\eps^2\Phi \to 0$ and~$\eps^2 \mu \gg \log n$ in both cases. Hence the~$0$-statement holds by~\eqref{eq:general} of \refT{thm_general}, showing that \refT{thm_strictly_balanced}~fails.} necessary. The following result covers the case where the unique primal subgraph of~$(G,H)$ is grounded, such as in examples~(a) and~(c) from~\refF{fig_primal}; this case includes the graphs in~\refT{thm_strictly_balanced}~(i). \begin{theorem}[Unique and grounded primal case]\label{thm_unique}% Let~$(G,H)$ be a rooted graph with a unique primal subgraph~$J$. If~$(G,J)$ is grounded, then there are constants $c, C, \alpha > 0$ such that, for all~$p=p(n) \in [0,1]$ and $\eps=\eps(n) \in [n^{-\alpha},1]$, \begin{equation}\label{eq:thm:unique} \lim_{n \to \infty} \Pr\Bigpar{\max_{\ii \in \oset{v_G}} |X_{\ii} - \mu| < \eps\mu} = \begin{cases} 0 &\text{if } \eps^2 \Phi \le c \log n, \\ 1 &\text{if } \eps^2 \Phi \ge C \log n. \end{cases}\vspace{-0.125em}% \end{equation}% \end{theorem} \noindent The heuristic idea is that main contribution to deviations of~$X_\xx=X_{G,H}(\xx)$ comes from those of~$X_{G,J}(\xx)$, and, since~$(G,J)$ is strictly balanced and grounded, the problem thus intuitively reduces to \refT{thm_strictly_balanced}~(i). \begin{figure} \begin{center} \hspace*{\fill} \begin{tikzpicture}[thick,scale=1] \foreach \x in {1, 2, ..., 3} { \draw[ultra thick] (-120*\x+30:0.5) node[vertex] (n\x) {} -- (-120*\x + 150:0.5); } \draw (n1) circle (4pt); \node at (1,-0.5) () {(a)}; \end{tikzpicture} \hspace*{\fill} \begin{tikzpicture}[thick,scale=1] \foreach \x in {1, 2, ..., 4} { \node[vertex] at (-90*\x:0.5) (n\x) {}; } \foreach \x/\y in {2/3,3/4,4/2} { \draw[ultra thick] (n\x) -- (n\y); } \draw (n1) circle (4pt); \node at (1,-0.5) () {(b)}; \end{tikzpicture} \hspace*{\fill} \begin{tikzpicture}[thick] \foreach \x in {1, 2, ..., 3} { \draw[ultra thick] (-120*\x+30:0.5) node[vertex] (n\x) {} -- (-120*\x + 150:0.5); } \draw (n1) circle (4pt); \draw (n3) -- +(0.7,0) node[vertex] (){}; \node at (1,-0.5) () {(c)}; \end{tikzpicture} \hspace*{\fill} \begin{tikzpicture}[thick] \begin{scope} \foreach \x in {1, 2, ..., 4} { \draw[ultra thick] (-90*\x:0.5) node[vertex] (n\x) {} -- (-90*\x + 90:0.5); } \end{scope} \foreach \x/\y in {1/3,2/4} { \draw[ultra thick] (n\x) -- (n\y); } \foreach \x/\y in {4/5, 5/6, 6/7, 7/8, 8/9} { \draw (n\x) -- (0.5*\x - 1.5, 0) node[vertex] (n\y) {}; } \draw (n9) circle (4pt); \node at (1.5,-0.5) () {(d)}; \end{tikzpicture}\vspace{-1.125em} \hspace*{\fill} \end{center} \caption{Examples of rooted graphs, with the root vertex circled and primal subgraphs marked in bold: (a)~strictly balanced and grounded, (b)~strictly balanced and not~grounded, (c)~with a unique primal that is~grounded, and (d)~with a unique primal that is not~grounded. Our main result \refT{thm_strictly_balanced} applies to~(a),(b), \refT{thm_unique} applies to~(a),(c), \refT{thm_nogrounded} applies to~(b),(d), and \refT{thm_general} applies to all of~them.} \label{fig_primal} \end{figure} The following result covers the case where no primal subgraph of~$(G,H)$ is grounded, such as in examples~(b) and~(d) from~\refF{fig_primal}; this case includes the graphs in~\refT{thm_strictly_balanced}~(ii). \begin{theorem}[No grounded primals case]\label{thm_nogrounded}% Let~$(G,H)$ be a rooted graph with no grounded primal subgraphs. There is a constant~$\alpha > 0$ such that, for all~$p=p(n) \in [0,1]$ and~$\eps=\eps(n) \in [n^{-\alpha},1]$, \begin{equation}\label{eq:main:nogrounded} \lim_{n \to \infty} \Pr\Bigpar{\max_{\ii \in \oset{v_G}} |X_{\ii} - \mu| < \eps\mu} = \begin{cases} 0 &\text{if } \eps^2 \Phi \to 0, \\ 1 &\text{if } \eps^2 \Phi \to \infty. \end{cases}\vspace{-0.125em}% \end{equation}% \end{theorem} \noindent Similarly to~\refT{thm_strictly_balanced}~(ii), the intuition is that all~$X_\xx$ are approximately equal once we know the number of unrooted copies of a certain subgraph of~$H$ (e.g., in example~(d) from \refF{fig_primal} this special subgraph~is~$K_4$). Theorems~\ref{thm_unique}--\ref{thm_nogrounded} give thresholds for the concentration of extension counts in terms of~$\eps^2\Phi$. For general~${(G,H)}$ we do not have such a threshold, but the following result intuitively states that the transition from the \mbox{$0$-statement} to the \mbox{$1$-statement} always happens at some point as~$\eps^2 \Phi$ changes from~$o(1)$ to~$n^{\Omega(1)}$. \begin{theorem}[General case: approximate conditions]\label{thm_general Let~$(G,H)$ be a rooted graph. For all~$p=p(n) \in [0,1]$ and~$\eps = \eps(n) \in (0,1]$ with~$1-p = \Omega(1)$ and~$\Phi \to \infty$, \begin{equation}\label{eq:general} \lim_{n \to \infty} \Pr\Bigpar{\max_{\xx \in \oset{v_G}}|X_\xx - \mu| < \eps\mu} = \begin{cases} 0 &\text{if } \eps^2 \Phi \to 0, \\ 1 &\text{if } \eps^2 \Phi = n^{\Omega(1)}. \end{cases}\vspace{0.75em \end{equation}% \end{theorem} \noindent The $1$-statement in~\eqref{eq:general} implies~\cite[Corollary~4]{S90b}, which in turn strengthens a result that played a key role in the study of zero-one laws~\cite{SS1988} due to Shelah and Spencer (since the `safe' assumptions from~\cite{S90b,SS1988} imply~$\Phi = n^{\Omega(1)}$ via~\refR{rem:Phibig}~\ref{eq:Phibig:iv} from \refS{s_prelim}). \pagebreak[3] \subsection{Discussion: open problems and cautionary examples}\label{sec:intro:general} For rooted subgraph extension counts, the main open problem is to fully determine the thresholds for concentration, i.e., to close the gap in~\eqref{eq:general} of \refT{thm_general} (and to weaken the conditions of Theorems~\ref{thm_strictly_balanced}--\ref{thm_nogrounded}). \begin{problem}\label{prb:open}% Determine the `correct' conditions for the $0$-~and $1$-statements of any rooted graph~$(G,H)$. \end{problem} Our understanding of~\refPr{prb:open} is still far from satisfactory. Indeed, even for fixed~$\eps \in (0,1]$ the correct {$1$-statement} condition remains open, which we now illustrate for the rooted graph~(e) from \refF{counterexample}. In this case, any {$(G,H)$-extension} can be viewed as a combination of a {$(G,K_4)$-extension} and a {$(K_4,H)$-extension}. The proof of Spencer's general $1$-statement \mbox{result~\cite[Theorem~3]{S90b}} combines this decomposition with his strictly balanced result~\eqref{eq_Spencer_SB} for~$(G,K_4)$ and~$(K_4,H)$, leading to a sufficient condition of form~$\min\{\mu_{G,K_4}, \mu_{K_4,H}\} \ge K'(\eps) \log n$ (cf.~\cite[Section~2]{S90b}). The following result shows that this sufficient condition can be weakened in some range, demonstrating that Spencer's general $1$-statement condition is not always optimal. \begin{proposition}\label{prop:counterexample:2}% Let~$(G,H)$ be the rooted graph~(e) depicted in \refF{counterexample}. Set~$\omega := np^2$. For all~$p=p(n) \in [0,1]$ and $\eps=\eps(n) \in (0,1]$ such that~$\omega \ll \log n$ and~$\eps^2 \omega^3 \gg \log n$, we have~$\eps^2\mu_{G,K_4} \gg \log n \gg \eps^2 \mu_{K_4,H}$ but~$\Pr(\max_{\xx \in \oset{v_G}}|X_\xx - \mu| < \eps\mu) \to 1$ as~$n \to \infty$. \end{proposition} \begin{figure} \begin{center} \hspace*{\fill} \begin{tikzpicture}[thick,scale=0.5] \draw (-1,0) node[vertex] (root) { } -- (1, 1) node[vertex] (top) { } -- (1, -1) node[vertex] (bot) { } -- (root) (top) -- (0.2, 0) node[vertex] (mid) {} -- (bot) (mid) -- (root) (top) -- (2, 0) node[vertex] (add1) {} -- (bot); \draw (root) circle (8pt); \node at (3.5,-1) () {(e)}; \end{tikzpicture} \hspace*{\fill} \begin{tikzpicture}[thick,scale=0.5] \draw (-1,0) node[vertex] (root) { } -- (1, 1) node[vertex] (top) { } -- (1, -1) node[vertex] (bot) { } -- (root) (top) -- (0.2, 0) node[vertex] (mid) {} -- (bot) (mid) -- (root) (top) -- (2, 0.5) node[vertex] (add1) {} -- (bot) (top) -- (2, -0.5) node[vertex] (add2) {} -- (bot); \draw (root) circle (8pt); \node at (3.5,-1) () {(f)}; \end{tikzpicture}\vspace{-1.125em} \hspace*{\fill} \end{center} \caption{The rooted graphs used in~Propositions~\ref{prop:counterexample:2}--\ref{prop:counterexample:1}, with the root vertex circled: for~(e) Spencer's general $1$-statement is not~optimal, and for~(f) the natural condition~$\eps^2\Phi \gg \log n$ does~not imply the~$1$-statement.} \label{counterexample} \end{figure} \noindent It is not hard to see that in the setting of \refP{prop:counterexample:2} we have $\eps^2\Phi \asymp \eps^2 \mu_{G, K_4} \gg \log n$, which together with Theorems~\ref{thm_unique}--\ref{thm_nogrounded} suggests that maybe~$\eps^2 \Phi \gg \log n$ is always a sufficient condition\footnote{Further support comes from the fact that~$X_\xx$ is asymptotically normal, see~\refCl{cl:mom}~\ref{cl:mom:asymp} in \refApp{apx:general} and the variance estimate~\eqref{eq:Variance} from \refS{s_prelim}, which makes it plausible that~$\Pr(|X_\xx - \mu| \ge \eps\mu) \le \e^{-\Omega((\eps \mu)^2/\Var X_\xx)} \le \e^{-\Omega(\eps^2 \Phi)} \ll n^{-v_G}$, which in turn would then establish the $1$-statement by taking the union bound over all~$\Theta(n^{v_G})$ roots~$\xx$.} for the $1$-statement (which would sharpen~\refT{thm_general}). However, the following cautionary result shows that this speculation is false for the rooted graph~(f) depicted in~\refF{counterexample}, indicating that Problem~\ref{prb:open} is more tricky than one might~think. \begin{proposition}\label{prop:counterexample:1} Let~$(G,H)$ be the rooted graph~(f) depicted in \refF{counterexample}. Set~$\omega := np^2$. For all~$p=p(n) \in [0,1]$ and~$\eps=\eps(n) \in (0,1]$ such that~$\omega \ll (\log n)^{0.39}$ and~$\eps^2 \omega^3 \gg \log n$, we have~$\eps^2\Phi \asymp \eps^2\mu_{G,K_4} \gg \log n$ but~$\Pr(\max_{\xx \in \oset{v_G}}|X_\xx - \mu| < \eps\mu) \to 0$ as~$n \to \infty$. \end{proposition} Overall, we hope that the above intriguing examples and open problems will stimulate more research into rooted subgraph counts. When~$(G,H)$ is strictly balanced and grounded, then we conjecture that~\eqref{eq:thm:unique} holds for suitable~$c,C>0$ under the natural assumptions~$\mu \to \infty$ and~$1-p=\Omega(1)$, i.e., without assuming~$\eps \ge n^{-\alpha}$. We leave it as an open problem to formulate a conjecture for the general solution to~\refPr{prb:open}, which in many cases is closely related to determining the regime where $\Pr(|X_\xx - \mu| \ge \eps\mu)$ changes from~$n^{-o(1)}$ to~$n^{-\omega(1)}$, say. In the concluding remarks we also discuss a potential connection to extreme value theory (see \refS{sec:conclusion}). \subsection{Organization of the paper} In \refS{s_prelim} we introduce some auxiliary results, which also imply~\refT{thm_general}. In \refS{s_strictly_balanced} we prove our main result \refT{thm_strictly_balanced}~(i) for strictly balanced~$(G,H)$ that are grounded. In Sections~\ref{s_nogrounded} and~\ref{s_unique} we prove Theorems~\ref{thm_unique}--\ref{thm_nogrounded} for the no grounded primal case, and the unique and grounded primal case. In \refS{sec:ext:non} we then prove \refT{thm_strictly_balanced}~(ii) for strictly balanced~$(G,H)$ that are not grounded. In \refS{sec:counterexample} we prove the cautionary examples from Propositions~\ref{prop:counterexample:2}--\ref{prop:counterexample:1}. Finally, \refS{sec:conclusion} contains some concluding remarks and~problems. \pagebreak[3] \section{Preliminaries}\label{s_prelim} In this preliminary section we collect some useful basic observations, and a partial result which implies Theorem~\ref{thm_general}. First, by adapting the textbook argument~\cite[Lemma~3.5]{JLR} for (unrooted) subgraph counts, for any rooted graph~$(G,H)$ it is standard to see that the variance of~$X_{G,H}(\xx)$ satisfies \begin{equation}\label{eq:Variance} \sigma^2 = \sigma_{G,H}^2 := \Var X_{G,H}(\xx) \asymp (1 - p)\mu_{G,H}^2 /\Phi_{G,H} \end{equation} for any edge probability~$p=p(n) \in (0,1]$, where~$\mu=\mu_{G,H}$ and~$\Phi=\Phi_{G,H}$ are defined as in~\eqref{def:muGH} and~\eqref{eq_PhiGH};~cf.~\cite{Matas2012phd}. Next, inspired by similar statements for subgraph counts~\cite[Lemma~3.6]{JLR}, using that~$\mu_{G,J} \asymp \xpar{n^{1/d(G,J)}p}^{e_J-e_G}$ for all~$G \subseteq J \subseteq H$ with~$e_J > e_G$, it is straightforward to establish the following useful~properties. \begin{remark}\label{rem:Phibig}% For any rooted graph~$(G,H)$, the following holds for all~$p=p(n)\in [0,1]$: \begin{romenumerate} \item\label{eq:Phibig:i $\Phi \to \infty$ is equivalent to~$p \gg n^{-1/m(G,H)}$. \item\label{eq:Phibig:ii $\Phi = \Omega(1)$ is equivalent to~$p = \Omega(n^{-1/m(G,H)})$. \item\label{eq:Phibig:iii If~$\Phi \asymp 1$, then~$\mu_{G,J} \asymp 1$ for any~$G \subseteq J \subseteq H$ that is primal for~$(G,H)$. \item\label{eq:Phibig:iv If~$p = \Omega(n^{-1/m(G,H) + \eta})$ for some constant~$\eta \ge 0$, then~$\Phi = \Omega(n^{\eta})$. \end{romenumerate} \end{remark} \noindent Finally, the approximate result Theorem~\ref{thm_general} immediately follows from the following slightly more general theorem, whose technical statement will be convenient in several later proofs. In particular, in some ranges of the parameters, we will be able to deduce the desired $1$-~or $0$-statements directly from~\eqref{eq:thm_generaltail:1}--\eqref{eq:thm_generaltail:0} below. \begin{theorem}\label{thm_generaltail}% For any rooted graph~$(G,H)$, the following holds for all~$p=p(n)\in [0,1]$: \begin{romenumerate} \item\label{thm_tail1}% If~$\Phi = \Omega(1)$ and $(t/\mu)^2\Phi\ge n^{\Omega(1)}$, then \begin{equation}\label{eq:thm_generaltail:1} \lim_{n \to \infty} \Pr\Bigpar{\max_{\xx \in \oset{v_G}}|X_\xx - \mu| < t} = 1. \hspace{2.5em} \end{equation}% \item\label{thm_tail0}% If $\eps = \eps(n) \in (0,1]$ and either (a)~$\Phi(1-p) \to \infty$ and $\eps^2 \Phi/(1-p) \to 0$, or~(b)~$\Phi \to 0$, then \begin{equation}\label{eq:thm_generaltail:0} \lim_{n \to \infty} \Pr\Bigpar{\max_{\xx \in \oset{v_G}}|X_\xx - \mu| \ge \eps \mu} = 1 . \hspace{2.5em} \vspace{-0.125em}% \end{equation}% \end{romenumerate} \end{theorem} \begin{remark}\label{rem:thm_generaltail}% In~\ref{thm_tail1}, the conclusion~\eqref{eq:thm_generaltail:1} holds with probability~$1 - o(n^{-\tau})$ for any constant~$\tau > 0$. \end{remark} \noindent We defer the simple proof of Theorem~\ref{thm_generaltail} to Appendix~\ref{apx:general}, and only mention the main ideas here. Claim~\ref{thm_tail0} exploits that~$X_\xx$ is asymptotically normal in a wide range. Claim~\ref{thm_tail1} is based on Markov's inequality and a central moment estimate $\E (X_\xx - \mu)^{2m} \le C_m \sigma^{2m} \le D_m (\mu^2/\Phi)^m$ that is a by-product of the usual asymptotic normality proof via the method of moments (see Claim~\ref{cl:mom} in Appendix~\ref{apx:general}). This approach for obtaining tail estimates `without much effort' does not seem to be as widely known in probabilistic combinatorics, and we believe that it will be useful in other applications (e.g., it yields a simple direct proof of~\cite[Corollary~4]{S90b}). % \section{Strictly balanced and grounded case (Theorem~\ref{thm_strictly_balanced})}\label{s_strictly_balanced} In this section we prove the threshold~\eqref{eq:main:strbal} of Theorem~\ref{thm_strictly_balanced}~(i) for strictly balanced rooted graphs~$(G,H)$ that are grounded (see \refS{sec:ext:non} for the less interesting ungrounded case). The~$0$-statement in~\eqref{eq:main:strbal} is the main difficulty, and here the plan is to use a second moment argument to show the existence of a root~$\xx \in \oset{v_G}$ with too many $(G,H)$-extensions, i.e., with~$X_{\xx} \ge (1+\eps)\mu$. Unfortunately, even an asymptotic estimate of the relevant first moment is challenging, since the upper tail proba\-bi\-li\-ty~$\Pr(X_{\xx} \ge (1+\eps)\mu)$ is hard to estimate up to a~$1+o(1)$ factor (this is an instance of the `infamous' upper tail problem~\cite{JR2002,SW18}). To sidestep this technical difficulty, we instead show the existence of a root~${\xx \in \oset{v_G}}$ which attains~$X_{\xx}=\ceil{(1+\eps)\mu}$ due to exactly~$\ceil{(1+\eps)\mu}$ extensions that are vertex-disjoint outside of~$\xx$. The crux is that these auxiliary events are more tractable: we can estimate the relevant first and second moments up to the required~$1+o(1)$ factors via a careful mix of Harris' inequality~\cite{Harris}, Janson's inequality~\cite{J90,BS,RiordanWarnke2015}, and counting arguments. It turns out that here the extra assumption~$\eps \ge n^{-\alpha}$ is helpful: it will allow us to focus on fairly small edge probabilities~$p=p(n)$ close to~$n^{-1/d(G,H)}$, which intuitively makes it easier to show that various events are approximately independent (as tacitly required by the second moment method); see~\refS{sec:0statement} for the~details. The~$1$-statement in~\eqref{eq:main:strbal} is simpler (and nowadays fairly routine). For edge probabilities~$p=p(n)$ that are close to~$n^{-1/d(G,H)}$, we use a standard union bound argument, estimating the lower tail~$\Pr(X_{\xx} \le (1-\eps)\mu)$ via Janson's inequality~\cite{J90,JLR,RiordanWarnke2015} and the upper tail~$\Pr(X_{\xx} \ge (1+\eps)\mu)$ via an inequality of Warnke~\cite{WUT}. For edge probabilities~$p=p(n)$ much larger than~$n^{-1/d(G,H)}$, it turns out that we can simply use the partial result \refT{thm_generaltail}~\ref{thm_tail1} due to the extra assumption~$\eps \ge n^{-\alpha}$; see~\refS{sec:1statement} for the~details. \subsection{Technical preliminaries}\label{s_strictly_balanced:prelim} Our upcoming arguments exploit two standard properties of strictly balanced rooted graphs: (i)~that, for fairly small edge probabilities~$p=p(n)$, the expectation~$\mu=\mu_{G,H}$ is significantly smaller than any other expectation~$\mu_{G,J}$ with~$G \subsetneq J \subsetneq H$ (note that~$\mu_{G,H}/\mu_{G,J} \asymp n^{v_H - v_J}p^{e_H - e_J} \ll 1$ via~\eqref{eq:lem:density:subs} below), and (ii)~that, after removing the root vertices~$G$ from~$H$, the remaining graph~${H - V(G)}$ is connected. Both mimic well-known properties from the unrooted case, so we defer the routine proof of \refL{lem:StrBal} to~\refS{s_strictly_balanced:prelim:deferred}. \begin{lemma}\label{lem:StrBal}% For any strictly balanced rooted graph~$(G,H)$, the following holds: \begin{romenumerate} \item\label{eq:StrBal:density There is a constant $\beta > 0$ such that, for all~$p=p(n) \in [0,1]$ with~$p= O(n^{-1/d(G,H) + \beta})$, \begin{equation} \label{eq:lem:density:subs} \max_{G \subsetneq J \subsetneq H} n^{v_H - v_J}p^{e_H - e_J} \ll n^{-\beta}. \hspace{2.5em} \vspace{-0.125em}% \end{equation}% \item\label{eq:StrBal:connected The graph~${H - V(G)}$, obtained from~$H$ by deleting the vertices of~$G$, is connected. \end{romenumerate} \end{lemma} \subsection{The $0$-statement}\label{sec:0statement} Our second moment based proof of the $0$-statement{} in~\eqref{eq:main:strbal} of Theorem~\ref{thm_strictly_balanced} hinges on the following key lemma. Given a root~$\xx \in \oset{v_G}$, let~$\cE_{\xx}$ denote the event that, in~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$, the root~$\xx$ has exactly~$\dex:= \ceil{(1 + \eps)\mu}$ pairwise \emph{vertex-disjoint} $(G,H)$-extensions (i.e., sharing no vertices outside~$\xx$), and no~other $(G,H)$-extensions. We also say that two roots~$\xx_1, \xx_2 \in \oset{v_G}$ are \emph{disjoint} if they share no elements as (unordered) sets. \begin{lemma}\label{lem:main}% Let $(G,H)$ be a rooted graph that is strictly balanced and grounded. There are constants~$c, \gamma > 0$ such that, for all~$\eps=\eps(n) \in (0,1]$ and~$p=p(n) \in [0,1]$ with~$p \le n^{-1/d(G,H) + \gamma}$, $\mu \ge 1/2$ and~$\eps^2\mu \le c \log n$, the following holds: for all roots~$\xx \in \oset{v_G}$ we have \begin{equation}\label{eq:pr:lb} \Pr(\cE_{\xx}) \gg n^{-1/2}, \end{equation} and for all disjoint roots~$\xx_1,\xx_2 \in \oset{v_G}$ we have \begin{equation}\label{eq:pr:ub} \Pr(\cE_{\xx_1}, \: \cE_{\xx_2}) \le (1 + o(1)) \Pr(\cE_{\xx_1})\Pr(\cE_{\xx_2}). \end{equation} \end{lemma} \begin{proof}[Proof of the $0$-statement{} in~\eqref{eq:main:strbal} of Theorem~\ref{thm_strictly_balanced} (assuming Lemma~\ref{lem:main})] Let~$c, \gamma>0$ be the constants given by Lemma~\ref{lem:main}. Fix arbitrary $0 < \alpha < \gamma/2$. First, when~$p > n^{-1/d(G,H) + \gamma}$, then~$\eps \ge n^{-\alpha}$ and \refR{rem:Phibig}~\ref{eq:Phibig:iv} imply~$\eps^2\mu \ge n^{-2\alpha} \cdot \Phi_{G,H} = \Omega(n^{\gamma - 2\alpha}) \gg \log n$, so the condition of the $0$-statement{} cannot be satisfied and hence there is nothing to prove. Next, when~$\mu < 1/2$, then~$(1+\eps) \mu \le 2 \mu < 1$ and~$\eps \le 1$ imply that the interval~$\left((1-\eps)\mu, (1 + \eps)\mu\right)$ contains no integers, and so the $0$-statement{} again holds trivially. Henceforth we thus can assume~$\mu \ge 1/2$ and~$p \le n^{-1/d(G,H) + \gamma}$, as required by Lemma~\ref{lem:main}. For convenience, we set~$s := \lfloor n / v_G \rfloor \asymp n$, and choose disjoint roots~$\xx_1, \dots, \xx_s \in \oset{v_G}$. Writing~$Y := |\left\{ i \in [s] : \cE_{\xx_i} \text { holds} \right\}|$, to prove the $0$-statement{} of Theorem~\ref{thm_strictly_balanced} we shall now show that~$Y > 0$~\whp{}. Namely, using~\eqref{eq:pr:lb} we obtain~$\E Y = \sum_{1 \le i \le s}\Pr(\cE_{\xx_i}) \gg s \cdot n^{-1/2} \asymp n^{1/2} \to \infty$, and together with~\eqref{eq:pr:ub} it follows~that \begin{equation*}\label{eq:mu2:ub} \begin{split} \E Y^2 & \le \sum_{1 \le i ,j \le s: \; i \neq j} \Pr(\cE_{\xx_i}, \: \cE_{\xx_j}) + \sum_{1 \le i \le s} \Pr(\cE_{\xx_i}) \; \le \; (1 + o(1)) \cdot (\E Y)^2 + \E Y \; \sim \; (\E Y)^2. \end{split} \end{equation*} Now Chebyshev's inequality readily yields $\Pr(Y=0) \le \Var Y/(\E Y)^2 \to 0$, completing the proof. \end{proof} The remainder of Section~\ref{sec:0statement} is dedicated to the proof of Lemma~\ref{lem:main}. For concreteness, for~$\beta>0$ as given by \refL{lem:StrBal}~\ref{eq:StrBal:density}, we choose the constants~$\gamma, c \in (0,1)$ such that \begin{equation}\label{eq:gammadef} \gamma e_H \; < \; \min\bigcpar{\beta/v_H, \: \beta/2, \: 1/2, \: 1-c}. \end{equation} Recalling~$\mu \asymp n^\vGH p^\eGH$ and~$\eps \le 1$, using the assumptions~$\mu \ge 1/2$ and~$p \le n^{-1/d(G,H) + \gamma}$ we infer \begin{equation}\label{eq:mumupper} 1/2 \: \le \: \mu \: \le \: \dex = \ceil{(1 + \eps)\mu} \: \le \: O(n^{\gamma e_H}) \ll \min\bigcpar{n^{1/2},n^{\beta/2}} , \end{equation} with room to spare. With foresight, given~$\xx \in \oset{v_G}$, we denote by~$N = N_{G,H}(\xx)$ the number of $(G,H)$-extensions of~$\xx$ in~$K_n$. Note that $N \asymp n^{\vGH}$ does not depend on the particular choice of~$\xx$. \subsubsection{The first moment: inequality~\eqref{eq:pr:lb}}\label{sec:first} We start with~\eqref{eq:pr:lb}, i.e., a lower bound for~$\Pr(\cE_{\xx})$. Recall that every $\xx \in \oset{v_G}$ has~$N$ extensions in~$K_n$. The plan is to show that~$\Pr(\cE_{\xx})$ is comparable to $\Pr(\Bin(N,p^\eGH) = \dex)$, more precisely~that \begin{equation}\label{eq:pr:lb:bin} \Pr(\cE_{\xx}) \; \ge \; (1+o(1)) \cdot \binom{N}{\dex} p^{(\eGH)\dex} (1-p^\eGH)^{N - \dex} . \end{equation} In view of $\dex \approx (1+\eps)\mu = (1 + \eps) Np^\eGH$, using a standard local limit theorem (or alternatively Stirling's formula) it then will be routine to deduce that~\eqref{eq:pr:lb:bin} is~$\Theta(\mu^{-1/2}) \cdot \e^{-\Theta(\eps^2\mu)}$, which together with~\eqref{eq:gammadef}--\eqref{eq:mumupper} and the assumption~$\eps^2\mu \le c\log n$ eventually establishes the desired inequality~\eqref{eq:pr:lb}; see \eqref{eq:MoivreLaplace}--\eqref{eq:MoivreLaplace2}~below. Turning to the technical details, given $\xx \in \oset{v_G}$, let~$\fH(\xx)$ denote the set of all (unordered) collections of~$\dex$ vertex-disjoint $(G,H)$-extensions of~$\xx$. Given~$\cC \in \fH(\xx)$, let~$\cC^c$ denote the remaining~${N - \dex}$ extensions of~$\xx$. Writing~$\cI_{\cS}$ for the event that all extensions in~$\cS$ are present in~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$, and~$\cD_{\cS}$ for the event that all extensions in~$\cS$ are not present in~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$, note that \begin{equation}\label{eq:er} \Pr(\cE_{\xx}) = \sum_{\cC \in \fH(\xx)} \Pr(\cI_{\cC} , \: \cD_{\cC^c}) = \sum_{\cC \in \fH(\xx)} \Pr(\cI_{\cC}) \Pr(\cD_{\cC^c} \mid \cI_{\cC}) \ge |\fH(\xx)| \min_{\cC \in \fH(\xx)} \Pr(\cI_{\cC})\Pr(\cD_{\cC^c} \mid \cI_{\cC}). \end{equation} Since~$N \asymp n^{\vGH}$ and~$\dex \ll n^{1/2}$ (see~\eqref{eq:mumupper}), routine calculations give \begin{equation}\label{eq:Hr} |\fH(\xx)| = \frac{\left(N- O(\dex n^{\vGH-1})\right)^\dex}{\dex!} = \frac{N^\dex}{\dex!} \cdot \e^{O(\dex^2/n)} \sim \frac{N^\dex}{\dex!} \sim \binom{N}{\dex} . \end{equation} Since the extensions in $\cC \in \fH(\xx)$ are disjoint, we have \begin{equation}\label{eq:I} \Pr(\cI_{\cC})= p^{(\eGH)\dex}. \end{equation} For the remaining lower bound on~$\Pr(\cD_{\cC^c} | \cI_{\cC})$, the idea is to apply Harris' inequality~\cite{Harris}, and then use \refL{lem:StrBal}~\ref{eq:StrBal:density} to show that the effect of `overlapping' pairs of extensions is negligible. \begin{claim}\label{cl:D:lower}% Let~$\xx \in \oset{v_G}$. Then, for all~$\cC \in \fH(\xx)$, we have \begin{equation}\label{eq:D:lower} \Pr(\cD_{\cC^c} | \cI_{\cC}) \; \ge \; (1+o(1)) \cdot (1-p^\eGH)^{N-\dex} . \end{equation} \end{claim} \begin{proof}% Defining the auxiliary graph~$F := \bigpar{[n], \; \bigcup\{E(H_1): H_1 \in \cC\}}$, note that every extension $H_2 \in \cC^c$ contains at least one edge not in~$F$ (since by \refL{lem:StrBal}~\ref{eq:StrBal:connected}, after deleting the root vertices~$\xx$, all graphs in~$\{H_1 -\xx: H_1 \in \cC\}$ are vertex-disjoint and connected). Since~$1-x \ge \e^{-2x}$ for~$x \le 1/2$, using Harris' inequality it routinely follows~that \begin{equation} \label{eq:Harris} \Pr(\cD_{\cC^c} | \cI_{\cC}) \ge \prod_{H_2 \in \cC^c} (1-p^{\eGH - e(H_2 \cap F)}) \ge (1-p^{\eGH})^{N - \dex} \cdot \exp \Big( -2 \hspace{-0.25em} \sum_{\substack{H_2 \in \cC^c: \\ e(H_2 \cap F) \ge 1}} \hspace{-0.5em} p^{\eGH - e(H_2 \cap F)}\Big) . \end{equation} To estimate the sum in \eqref{eq:Harris}, note that if~$H_2 \in \cC^c$ shares an edge with~$F$, then~$E(H_2 \cap F)$ corresponds to a $(G,J)$-extension of~$\xx$ for some~$G \subsetneq J \subsetneq H$. The number of such extensions is at most~$(v_H\dex)^{v_J-v_G} = O(\dex^{v_H})$, with room to spare. Given a $(G,J)$-extension, it can be further extended to some~$H_2 \in \cC^c$ in at most~$n^{v_H-v_J}$ ways. Using~$e_H-e_G-(e_J-e_G)=e_H-e_J$ together with~\eqref{eq:mumupper} and \eqref{eq:lem:density:subs}, it follows that \begin{equation}\label{eq:Harris:overlap} \sum_{\substack{H_2 \in \cC^c: \\ e(H_2 \cap F) \ge 1}} \hspace{-0.5em} p^{\eGH - e(H_2 \cap F)} \le \sum_{G \subsetneq J \subsetneq H} \hspace{-0.375em} O\Bigpar{\dex^{v_H} n^{v_H-v_J} \cdot p^{e_H-e_J}} \ll n^{\gamma e_H v_H - \beta} = o(1), \end{equation} which together with~\eqref{eq:Harris} establishes inequality~\eqref{eq:D:lower}. \end{proof} Combining estimates~\eqref{eq:er}--\eqref{eq:D:lower}, we readily obtain inequality~\eqref{eq:pr:lb:bin}. To establish~\eqref{eq:pr:lb}, it remains to estimate the right-hand side of~\eqref{eq:pr:lb:bin} via a standard local limit theorem for the binomial distribution, namely~\cite[Theorem~1 in~Section~VII.3]{Feller}. Number~$k$ in~\cite{Feller} translates in our setting to~${k := {\dex - \lfloor(N + 1)p^\eGH\rfloor}={\eps\mu+O(1)}}$ (what is~$m$ in~\cite{Feller} is~$\lfloor(N+1)p^\eGH \rfloor$ in our case), and thus~$k \le \dex \ll N^{2/3}$ by~\eqref{eq:mumupper}. Hence the aforementioned local limit theorem from~\cite{Feller} applies, which in view of~$\mu=Np^\eGH$ gives \begin{equation}\label{eq:MoivreLaplace} \binom{N}{\dex} p^{(\eGH)\dex} (1-p^\eGH)^{N-\dex} \; \sim\; \frac{1}{\sqrt{2\pi \mu(1-p^\eGH)}} \cdot \exp\biggpar{- \frac{k^2}{\mu(1 - p^\eGH)}} . \end{equation} Using~\eqref{eq:mumupper} we readily infer~$k^2/\mu = \eps^2\mu + O(1)$. Note that~\eqref{eq:gammadef} implies~$p^\eGH \le n^{-(v_H-v_G) + \gamma(\eGH)} \le n^{-1 + 1/2}= n^{-1/2} \to 0$. Using the estimates~\eqref{eq:mumupper} and~$\eps^2\mu \le c \log n$ together with~$\gamma e_H + c < 1$ (see~\eqref{eq:gammadef}), it now follows that~\eqref{eq:MoivreLaplace} is at least \begin{equation}\label{eq:MoivreLaplace2} \Omega\bigpar{\mu^{-1/2}} \cdot \exp \Bigpar{-\bigpar{1+O(p^\eGH)} \tfrac{1}{2}\eps^2\mu} \; \ge \; \Omega(1) \cdot \exp \Bigpar{-\tfrac{\gamma e_H + c}{2}\log n} \gg n^{-1/2}, \end{equation} which together with~\eqref{eq:pr:lb:bin} completes the proof of inequality~\eqref{eq:pr:lb} from Lemma~\ref{lem:main}. \subsubsection{The second moment: inequality~\eqref{eq:pr:ub}} Now we turn to~\eqref{eq:pr:ub}, i.e., an upper bound for $\Pr\left( \cE_{\xx_1}, \cE_{\xx_2} \right)$ when~$\xx_1$, $\xx_2$ are disjoint. Recalling~\eqref{eq:er}, note~that \begin{equation}\label{eq:EE00} \Pr(\cE_{\xx_1} , \: \cE_{\xx_2}) = \sum_{\cC_1 \in \fH(\xx_1)}\sum_{\cC_2 \in \fH(\xx_2, \cC_1)} \Pr(\cI_{\cC_1 \cup \cC_2} , \: \cD_{\cC_1^c \cup \cC_2^c}) , \end{equation} where we (with foresight) define \begin{equation}\label{eq:HRC} \fH(\xx_2, \cC_1) := \bigcpar{ \cC_2 \in \fH(\xx_2) : \ \Pr(\cI_{\cC_1 \cup \cC_2} , \: \cD_{\cC_1^c \cup \cC_2^c})>0 }. \end{equation} Guided by the heuristic that the various events are approximately independent, the plan is to show~that \begin{equation}\label{eq:EE01} \Pr(\cI_{\cC_1 \cup \cC_2} , \: \cD_{\cC_1^c \cup \cC_2^c}) \; \le \; (1+o(1)) \Pr(\cI_{\cC_1} , \: \cD_{\cC_1^c}) \cdot \Pr(\cI_{\cC_2}, \: \cD_{\cC_2^c}) , \end{equation} though the actual details will be slightly more involved. Ignoring these complications for now, note that~\eqref{eq:EE01} would together with~\eqref{eq:EE00}, \eqref{eq:er} and~$\fH(\xx_2, \cC_1) \subseteq \fH(\xx_2)$ indeed imply the desired inequality~\eqref{eq:pr:ub}. Turning to the technical details, since $\cI_{\cC_1 \cup \cC_2}$ is an increasing event and $\cD_{\cC_1^c \cup \cC_2^c}$ is a decreasing event, using~\eqref{eq:EE00} and Harris' inequality~\cite{Harris} we obtain \begin{equation}\label{eq:EE} \begin{split} \Pr(\cE_{\xx_1}, \: \cE_{\xx_2}) \le \sum_{\cC_1 \in \fH(\xx_1)}\sum_{\cC_2 \in \fH(\xx_2, \cC_1)} \Pr(\cI_{\cC_1 \cup \cC_2}) \Pr(\cD_{\cC_1^c \cup \cC_2^c}) . \end{split} \end{equation} Recalling that every~$\xx \in \oset{v_G}$ has~$N$ extensions in~$K_n$, Harris' inequality also readily gives~$\Pr(\cD_{\cC_1^c \cup \cC_2^c}) \ge (1 - p^\eGH)^{2(N-\dex)}$. We will now prove an asymptotically matching upper bound that does \emph{not} depend on the choice of $\cC_1$ and~$\cC_2$ (similarly as in Claim~\ref{cl:D:lower}). Here the idea is to apply a form of Janson's inequality~\cite{BS,JLR,AS}, and then again use \refL{lem:StrBal}~\ref{eq:StrBal:density} to argue that `overlaps' have negligible contribution. \begin{claim}\label{cl:D2:upper}% Let~$\xx_1,\xx_2 \in \oset{v_G}$ be disjoint. Then, for all~$\cC_1 \in \fH(\xx_1)$ and~$\cC_2 \in \fH(\xx_2)$, we~have \begin{equation}\label{eq:D2:upper} \Pr(\cD_{\cC_1^c \cup \cC_2^c}) \; \le \; (1 + o(1)) \cdot (1-p^\eGH)^{2(N - \dex)} . \end{equation} \end{claim} \begin{proof}% Let~$\cF$ be the family of~{{$(e_H$}{$-$}{$e_G)$}}-element edge-sets corresponding to extensions in~$\cC_1^c \cup \cC_2^c$ (each extension of~$\xx_i$ is uniquely determined by its edge-set, since~$H$ has no isolated vertices outside of~$V(G)$ by \refL{lem:StrBal}~\ref{eq:StrBal:connected}). Note that if an extension in~$\cC_1^c$ is also an extension in~$\cC_2^c$, then it must contain some vertex from~$\xx_2$ (because~$(G,H)$ is grounded). Since~$\xx_1, \xx_2$ are disjoint, the number of such duplicate extensions is~$O(n^{v_H - v_G - 1})$, so that~$|\cF| \ge 2(N - \dex) - O(n^{v_H - v_G - 1})$. Note that the event~$\cD_{\cC_1^c \cup \cC_2^c}$ implies~$\sum_{E \in \cF} \indic{E \subseteq {\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)} = 0$. Since~$(1-x)^{-1} \le \e^{2x}$ for~$x \le 1/2$, by invoking the Boppana--Spencer~\cite{BS} variant of Janson's inequality (see, e.g.,~\cite[Remark~2.20]{JLR} or~\cite[Theorem~8.1.1]{AS}) it then follows~that \begin{equation} \label{eq:Janson} \Pr(\cD_{\cC_1^c \cup \cC_2^c}) \le (1-p^\eGH)^{|\cF|} \cdot \e^{2\Delta} \le (1-p^\eGH)^{2(N - \dex)} \cdot \e^{O(n^{v_H - v_G - 1}p^{\eGH}+\Delta)}, \end{equation} where \begin{equation}\label{eq:Janson:Delta} \Delta := \sum_{\substack{(E_1, E_2) \in \cF\times\cF:\\ 1 \le |E_1 \cap E_2| < \eGH}} \hspace{-0.75em} p^{|E_1 \cup E_2|}. \end{equation} Using~$\mu=Np^\eGH \asymp n^{\vGH}p^\eGH$ together with~\eqref{eq:mumupper}, it follows that \begin{equation}\label{eq:Janson:approx} n^{v_H - v_G - 1}p^\eGH \asymp \mu \cdot n^{-1} \ll n^{1/2 - 1} = o(1). \end{equation} Turning to the~$\Delta$-term, note that~$|\cF|p^\eGH \le 2(N - \dex)p^\eGH \le 2 \mu$. By proceeding analogously to the estimates in~\eqref{eq:Harris}--\eqref{eq:Harris:overlap}, using \eqref{eq:mumupper} and \eqref{eq:lem:density:subs} it routinely follows~that \begin{equation}\label{eq:Janson:Delta:bound} \Delta \le \sum_{E_1 \in \cF} p^{\eGH}\hspace{-0.75em}\sum_{\substack{E_2 \in \cF:\\ 1 \le |E_1 \cap E_2| < \eGH}}\hspace{-1.0em} p^{\eGH-|E_1 \cap E_2|} \le O\Bigpar{\mu \cdot \sum_{G \subsetneq J \subsetneq H} n^{v_H - v_J}p^{e_H - e_J}} = o(1), \end{equation} which together with~\eqref{eq:Janson}--\eqref{eq:Janson:approx} establishes inequality~\eqref{eq:D2:upper}. \end{proof} To sum up, by inserting the estimates~\eqref{eq:I} and~\eqref{eq:D2:upper} into~\eqref{eq:EE}, we readily arrive at \begin{equation}\label{eq:EE1} \Pr(\cE_{\xx_1},\cE_{\xx_2}) \; \le \; (1+o(1)) \cdot p^{(\eGH)\dex}(1-p^\eGH)^{2(N - \dex)} \sum_{\cC_1 \in \fH(\xx_1)}\sum_{\cC_2 \in \fH(\xx_2, \cC_1)} \Pr(\cI_{\cC_2} \mid \cI_{\cC_1}) . \end{equation} Anticipating that the main contribution comes from pairs~$\cC_1, \cC_2$ of `disjoint' collections, we~partition \begin{equation}\label{eq:part} \fH(\xx_1) := \fH_0(\xx_1,\xx_2) \cup \fH_{\ge 1}(\xx_1,\xx_2) , \end{equation} where $\fH_0(\xx_1,\xx_2)$ contains the collections~$\cC_1 \in \fH(\xx_1)$ for which the auxiliary~graph \begin{equation}\label{eq:F} F = F(\cC_1) := \Bigl([n], \; \bigcup\bigl\{E(H') : H' \in \cC_1\bigr\}\Bigr) \end{equation} contains no extensions of~$\xx_2$, and~$\fH_{\ge 1}(\xx_1,\xx_2)$ contains the remaining ones. Since~$\xx_1, \xx_2$ are disjoint and~$(G,H)$ is grounded, every $\cC_1 \in \fH_{\ge 1}(\xx_1,\xx_2)$ must contain at least one extension overlapping with~$\xx_2$ (in at least one vertex). From \eqref{eq:Hr}, $N \asymp n^{\vGH}$ and~$\dex \ll n$ (see~\eqref{eq:mumupper}) it follows~that, for some constant~$A = A(G,H) > 0$, \begin{equation}\label{eq:negligbleC1} \left|\fH_{\ge 1}(\xx_1,\xx_2)\right| \le A n^{\vGH-1} \cdot \binom{N}{\dex-1} \asymp n^{\vGH-1}\cdot \frac{\dex}{N} \cdot |\fH(\xx_1)| \ll |\fH(\xx_1)| . \end{equation} Exploiting the groundedness assumption, we next show that pairs~$\cC_1, \cC_2$ can only overlap in at most~$v_G=O(1)$ extensions (see~Claim~\ref{cl:finiteOverlaps}), and that overlapping pairs effectively have negligible contribution (see~Claim~\ref{cl:conditionalsum}). \begin{claim}\label{cl:finiteOverlaps}% Let~$\xx_1,\xx_2 \in \oset{v_G}$ be disjoint. Then, for all~$\cC_1 \in \fH(\xx_1)$, the graph~$F = F(\cC_1)$ defined in~\eqref{eq:F} contains at most~$v_G$ vertex-disjoint extensions of~$\xx_2$. \end{claim} \begin{proof}% The graph~${F - \xx_1}$ (obtained by removing the vertices~$\xx_1$ from~$F$), consists of isolated vertices and vertex-disjoint copies of the graph~${H - V(G)}$, which, by \refL{lem:StrBal}~\ref{eq:StrBal:connected}, is connected. Let~$H'$ be obtained from~${H - E(G)}$ by removing isolated root vertices (if any). Since~$(G,H)$ is grounded, we have ${e_{{H - V(G)}} < e_{H'}}$. Note that~$H'$ is connected (since it equals~${H - V(G)}$ with some root vertices connected to it) and therefore~${F - \xx_1}$ is~$H'$-free. It follows that any extension of~$\xx_2$ that is present in~$F$ must intersect~$\xx_1$, so there are at most~$|\xx_1| = v_G$ such vertex-disjoint extensions of~$\xx_2$. \end{proof} \begin{claim}\label{cl:conditionalsum}% Let~$\xx_1,\xx_2 \in \oset{v_G}$ be disjoint. Then \begin{equation}\label{eq:sumP} \sum_{\cC_1 \in \fH(\xx_1)}\sum_{\cC_2 \in \fH(\xx_2, \cC_1)}\Pr(\cI_{\cC_2} \mid \cI_{\cC_1}) \; \le \; (1+o(1)) \sum_{\cC_1 \in \fH(\xx_1)}\sum_{\cC_2 \in \fH(\xx_2)}\Pr(\cI_{\cC_2}) . \end{equation} \end{claim} \begin{proof}[Proof of Claim~\ref{cl:conditionalsum}]% In the first step we estimate~$\sum_{\cC_2 \in \fH(\xx_2, \cC_1)}\Pr(\cI_{\cC_2} \mid \cI_{\cC_1})$ using a counting argument that accounts for the different kinds of overlaps of~$\cC_2$ with the graph~$F = F(\cC_1)$ defined in~\eqref{eq:F}. Turning to the details, as in the proof of Claim \ref{cl:D2:upper} we will think of~$(G,H)$-extensions as~{{$(e_H$}{$-$}{$e_G)$}}-element edge-sets. Suppose that~$F$ contains~$k$ extensions of~$\xx_2$. If $\cC_2 \in \fH(\xx_2,\cC_1)$ then all these~$k$ extensions must be present in $\cC_2$, since otherwise $\Pr(\cI_{\cC_1 \cup \cC_2} , \cD_{\cC_1^c \cup \cC_2^c}) \le \Pr(\cI_{\cC_1}, \cD_{\cC_2^c})= 0$ contradicting $\cC_2 \in \fH(\xx_2, \cC_1)$. Each of the remaining~${\dex - k}$ extensions~$E_i$ in~$\cC_2$ is not fully contained in~$F$, and thus intersects~$F$ in an edge-set that corresponds to a $(G,J_i)$-extension of~$\xx_2$ for some~$G \subseteq J_i \subsetneq H$ (the case~$J_i = G$ occurs when the extension~$E_i$ is edge-disjoint from~$F$). When these intersections are given by~${J_1, \dots, J_{\dex-k}}$, then we clearly have \begin{equation*} \prob{ \cI_{\cC_2} \mid \cI_{\cC_1} } = \prod_{i = 1}^{\dex - k} p^{e_H - e_G - (e_{J_i}-e_G)} = \prod_{i = 1}^{\dex - k} p^{e_H - e_{J_i}}. \end{equation*} Furthermore, the number of collections~$\cC_2$ with such intersections~${J_1, \dots, J_{\dex-k}}$ is bounded by \begin{equation* \frac{1}{(\dex-k)!} \cdot \prod_{i = 1}^{\dex - k} \bigpar{v_G + (\vGH)\dex}^{v_{J_i} - v_G}\extcount{J_i,H}, \end{equation*} where~$\extcount{G,H} := N_{G,H}=N$ and~$\extcount{J,H} := n^{v_H - v_J}$ otherwise (to clarify: we divided by~$(\dex - k)!$ since we count unordered collections~$\cC_2$). Hence, summing over all possible choices of~$J_1, \dots, J_{\dex-k}$, it follows that \begin{align} \sum_{\cC_2 \in \fH(\xx_2, \cC_1)}\Pr(\cI_{\cC_2} \mid \cI_{\cC_1}) & \le \notag \frac{1}{(\dex-k)!} \sum_{\substack{J_1, \dots, J_{\dex-k}:\\ G \subseteq J_i \subsetneq H}} \prod_{i = 1}^{\dex - k} \bigpar{v_G + (\vGH)\dex}^{v_{J_i} - v_G} \extcount{J_i,H} p^{e_H - e_{J_i}} \\ &\le \frac{\dex^k}{\dex!} \cdot \biggpar{\sum_{G \subseteq J \subsetneq H}\bigpar{v_G + (\vGH)\dex}^{v_J - v_G} \extcount{J,H} p^{e_H - e_J}}^{\dex - k} \label{eq:rhs}. \end{align} Noting that $\extcount{G,H}p^{e_H - e_G} = \mu$, using \eqref{eq:Harris:overlap} and~$\mu \asymp \dex$ we bound the sum in~\eqref{eq:rhs} from above by, say, \begin{equation}\label{eq:muerror} \mu + O\bigpar{\sum_{G \subsetneq J \subsetneq H} \dex^{v_H}n^{v_H - v_J}p^{e_H - e_J}} \le \mu + o(1) = \mu \cdot \Bigpar{1 + o\bigpar{\dex^{-1}}}. \end{equation} From the assumptions~$\eps \le 1$ and~$\mu \ge 1/2$ it follows that~$\dex \le (1+\eps)\mu + 1 \le 4 \mu$, say. Therefore, in view of \eqref{eq:rhs}--\eqref{eq:muerror}, using~$\mu=Np^\eGH$ and~\eqref{eq:Hr} it follows~that \begin{equation}\label{eq:intermediate} \sum_{\cC_2 \in \fH(\xx_2, \cC_1)}\Pr(\cI_{\cC_2} \mid \cI_{\cC_1}) \le \left( \frac{\dex}{\mu}\right)^k\frac{(Np^\eGH)^{\dex}}{\dex!} \Bigpar{1 + o\bigpar{\dex^{-1}}}^{\dex-k} \: \le \: (1+o(1)) \cdot 4^k|\fH(\xx_2)|p^{(\eGH)\dex} , \end{equation} whenever the graph~$F$ defined in~\eqref{eq:F} contains exactly~$k$ extensions of~$\xx_2$. In the second step we sum the above estimate~\eqref{eq:intermediate} over all~$\cC_1 \in \fH(\xx_1)$. Recalling the partition~\eqref{eq:part}, note that~$k =0$ when~$\cC_1 \in \fH_0(\xx_1,\xx_2)$, and that~$k \le v_G$ otherwise (see Claim~\ref{cl:finiteOverlaps}). From~\eqref{eq:intermediate} it follows that \[ \sum_{\cC_1 \in \fH(\xx_1)}\sum_{\cC_2 \in \fH(\xx_2, \cC_1)}\Pr(\cI_{\cC_2} \mid \cI_{\cC_1}) \le (1+o(1)) \cdot \Bigpar{|\fH_0(\xx_1,\xx_2)| + 4^{v_G}|\fH_{\ge 1}(\xx_1,\xx_2)|} \cdot |\fH(\xx_2)|p^{(\eGH)\dex} . \] In view of~\eqref{eq:negligbleC1}, the factor in the above parentheses is at most~$(1+o(1)) \cdot |\fH(\xx_1)|$, say, which together with~$p^{(\eGH)\dex}=\Pr(\cI_{\cC_2})$ from~\eqref{eq:I} then completes the proof of inequality~\eqref{eq:sumP}. \end{proof} Finally, inserting the estimates~\eqref{eq:sumP}, $p^{(\eGH)\dex}=\Pr(\cI_{\cC_1})$, and~\eqref{eq:D:lower} into~\eqref{eq:EE1}, it follows that \begin{equation* \Pr(\cE_{\xx_1},\cE_{\xx_2}) \; \le \; (1+o(1)) \sum_{\cC_1 \in \fH(\xx_1)}\Pr(\cI_{\cC_1})\Pr(\cD_{\cC_1^c} | \cI_{\cC_1})\sum_{\cC_2 \in \fH(\xx_2)} \Pr(\cI_{\cC_2})\Pr(\cD_{\cC_2^c} | \cI_{\cC_2}) , \end{equation*} which together with~\eqref{eq:er} completes the proof of inequality~\eqref{eq:pr:ub} and thus Lemma~\ref{lem:main} (which in turn implies the $0$-statement{} in~\eqref{eq:main:strbal} of \refT{thm_strictly_balanced}, as discussed). \noproof \subsection{The $1$-statement}\label{sec:1statement} Our proof of the $1$-statement{} in~\eqref{eq:main:strbal} of Theorem~\ref{thm_strictly_balanced} is based on a fairly standard union bound argument. \begin{proof}[Proof of the $1$-statement in~\eqref{eq:main:strbal} of Theorem~\ref{thm_strictly_balanced} Fix an arbitrary constant~$\tau > 0$. For~$\beta > 0$ as given by \refL{lem:StrBal}~\ref{eq:StrBal:density}, fix constants~$0 < \gamma \le \beta$ and $0 < \alpha < \gamma/2$ as in the proof of the $0$-statement (see \refS{sec:0statement}). If~$p > n^{-1/d(G,H) + \gamma}$, then \refR{rem:Phibig}~\ref{eq:Phibig:iv} implies~$\Phi_{G,H} = \Omega(n^\gamma)$, and using~$\eps^2 \Phi_{G,H} = \Omega(n^{\gamma - 2\alpha}) = n^{\Omega(1)}$ we see that the $1$-statement of Theorem~\ref{thm_strictly_balanced} follows from Theorem~\ref{thm_generaltail}~\ref{thm_tail1} with~$t = \eps \mu$. In the remaining main case~$p \le n^{-1/d(G,H) + \gamma}$, we fix a root~$\xx \in \oset{v_G}$. Since there are~$O(n^{v_G})$ many such roots, for the $1$-statement of Theorem~\ref{thm_strictly_balanced} it suffices to show that, for~$C>0$ large enough, \begin{equation}\label{eq:thm_ext_01:goal} \prob{|X_\xx - \mu| \ge \eps \mu} = o\bigl( n^{- (v_G + \tau)} \bigr) \qquad \text{ if~$\eps^2 \mu \ge C \log n$.} \end{equation} To avoid clutter, we shall use the convention that all implicit constants~$c_i$ may depend on~$(G,H)$. For the lower tail we shall apply Janson's inequality~\cite[Theorem~1]{RiordanWarnke2015}, which in view of~\eqref{eq:lem:density:subs} from \refL{lem:StrBal}~\ref{eq:StrBal:density} routinely (similarly to the textbook argument~\cite{JLR,JW} for unrooted subgraph counts) gives \begin{equation}\label{eq:1:LT:Janson} \prob{X_\xx \le (1- \eps) \mu} \; \le \; \exp\Bigl(-c_1 \eps^2 \mu\Bigr) \le n^{-c_1 C} = o\bigl( n^{- (v_G + \tau)} \bigr) \end{equation} for~$C > (v_G+\tau)/c_1$ (analogous to~\eqref{eq:Janson:Delta:bound}, the relevant~$\Delta$-term from~\cite{RiordanWarnke2015} is here again~$o(1)$ by~\eqref{eq:mumupper} and~\eqref{eq:lem:density:subs}). For the upper tail we shall apply~\cite[Theorem~32]{WUT} in the setting described in~\cite[Example~20]{WUT} (the conditions (H$\ell$), (P), (P$q$) are defined in~\cite[Section~4.1]{WUT}). The underlying hypergraph $\cH = \fH(\xx)$ consists of the edge-sets of extensions of $\xx$, thus having vertex-set~$V(\cH) = E(K_n)$. We set the parameters to~$N = n^2$, $\ell = 1$, $q = k = \eGH$, and~$K=v_G+2\tau$. The quantity~$\mu_j$ from~\cite[Example~20]{WUT} satisfies~$\max_{1 \le j < q}\mu_j \le \max_{G \subsetneq J \subsetneq H} n^{v_H - v_J}p^{e_H - e_J} \ll n^{-\beta}$ by \refL{lem:StrBal}~\ref{eq:StrBal:density}. Invoking~\cite[Theorem~32]{WUT}, it then follows~that \begin{equation}\label{eq:1:UT:Wsmallp} \prob{X_\xx \ge (1+\eps) \mu} \; \le \; \bigl(1 + o(1)\bigr) \cdot \exp\Bigl(-\min\bigl\{c_2\eps^2 \mu, \: (v_G + 2\tau)\log n\bigr\}\Bigr) = o\bigl( n^{- (r + \tau)} \bigr) \end{equation} for~$C > (v_G+\tau)/c_2$, completing the proof of~\eqref{eq:thm_ext_01:goal} and thus the $1$-statement in~\eqref{eq:main:strbal} of Theorem~\ref{thm_strictly_balanced}. \end{proof} \begin{remark}[Theorem~\ref{thm_strictly_balanced}: stronger $1$-statement]\label{rem:thm_strictly_balanced The above proof yields, in view of~\refR{rem:thm_generaltail}, the following stronger conclusion: for any fixed~$\tau > 0$ there is a constant~$C=C(\tau,G,H)>0$ such that the~$1$-statement in~\eqref{eq:main:strbal} of Theorem~\ref{thm_strictly_balanced} holds with probability~$1 - o(n^{-\tau})$. \end{remark} \subsection{Deferred proof of \refL{lem:StrBal}}\label{s_strictly_balanced:prelim:deferred} For completeness, we now give the routine proof of \refL{lem:StrBal} deferred from \refS{s_strictly_balanced:prelim}. \begin{proof}[Proof of \refL{lem:StrBal} \ref{eq:StrBal:density}: Set~$\Psi_{J,H} := n^{v_H - v_J}p^{e_H - e_J}$. In the case $v_J = v_H$, for any~$\beta > 0$ satisfying~$1/d(G,H) > 2\beta$ we have~$\Psi_{J,H} = p^{e_H-e_J} \ll n^{-(e_H - e_J)\beta} \le n^{-\beta}$. Henceforth we can thus assume $v_J < v_H$. Since~$G$ is an induced subgraph of~$H$ and thus of~$J$, we also have~$v_G < v_J$. Since~$(G,H)$ is strictly balanced we have~$d(G,J)< d(G,H)$, which implies \begin{equation}\label{eq:lem:density:subs:3} d(J,H) = \frac{(e_H-e_G)-(e_J-e_G)}{(v_H-v_G)-(v_J-v_G)} = \frac{(v_H-v_G)d(G,H)-(v_J-v_G)d(G,J)}{(v_H-v_G)-(v_J-v_G)} > d(G,H) . \end{equation} Hence~$1/d(G,H) > 1/d(J,H) + 2\beta$ for~$\beta>0$ sufficiently small, so that~$p = O(n^{-1/d(G,H)+\beta}) \ll n^{-1/d(J,H) - \beta}$. Observe that~$e_H > e_J$, since otherwise~$e_H = e_J$ and~$v_H > v_J$ imply~$d(G,J) > d(G,H)$, contradicting that~$(G,H)$ is strictly~balanced. Hence $\Psi_{J,H} = (n^{1/d(J,H)} p)^{e_H-e_J} \ll n^{-\beta}$, completing the proof of~\eqref{eq:lem:density:subs}. \ref{eq:StrBal:connected}: Assume the contrary. Then we can split $V(H) \setminus V(G)$ into two nonempty sets~$V_1$ and~$V_2$ such that there are no edges between $V_1$ and $V_2$. Writing~$H_i := H[V(G) \cup V_i]$, we readily obtain \begin{equation*} d(G,H) = \frac{e_H-e_G}{v_H-v_G} = \frac{\sum_{i \in [2]}(e_{H_i}-e_G)}{\sum_{i \in [2]}(v_{H_i}-v_G)} = \frac{\sum_{i \in [2]}(v_{H_i}-v_G)d(G,H_i)}{\sum_{i \in [2]}(v_{H_i}-v_G)} \le \max_{i \in [2]}d(G,H_i) . \end{equation*} Since~$(G, H)$ is strictly balanced we have~$d(G,H_i)< d(G,H)$, yielding the desired contradiction. \end{proof} \section{No grounded primals case (\refT{thm_nogrounded})}\label{s_nogrounded} In this section we prove Theorem~\ref{thm_nogrounded} by focusing on a maximal primal subgraph~$J_{\max}$ of~$(G, H)$; we remark that~$J_{\max}$ is in fact unique (the union of all primal subgraphs), but we do not need this. Our arguments hinge on the basic observation that, since~$J_{\max}$ is by assumption not grounded (i.e.,~there are no edges between~$V(G)$ and~$V(J_{\max}) \setminus V(G)$), extension counts~$X_{G,J_{\max}}(\xx)$ are essentially the same as the number of \emph{unrooted} copies of the graph~$K := {J_{\max} - V(G)}$, where the vertices of~$G$ are~deleted from~$J_{\max}$. For the $1$-statement this heuristically means that if~$X_{G,J_{\max}}(\xx)$ is concentrated for \emph{some}~$\xx$, then $X_{G,J_{\max}}(\xx)$ is concentrated for \emph{all}~$\xx$ (the reason being that not too many copies of~$K$ can overlap with any root~$\xx'$, see Lemma~\ref{lem:scattered} below). Furthermore, using Theorem~\ref{thm_generaltail}~\ref{thm_tail1} it turns out that \whp{} each copy of~$J_{\max}$ extends to the `right' number of $H$-copies (here the crux will be that~$\Phi_{J_{\max}, H} = n^{\Omega(1)}$ follows from \refR{rem:Phibig}~\ref{eq:Phibig:iv} and Lemma~\ref{lem:nogroundeddensity} below). Combining these two estimates then allows us to deduce that~\whp{}~$X_{G,H}(\xx)$ is concentrated for all~$\xx$; see Section~\ref{sec:ext:1} for~the~details. For the $0$-statement we shall proceed similarly, the main difference is that, for a \emph{fixed}~$\xx$, we start by arguing that~$X_{G,J_{\max}}(\xx)$ is not concentrated, i.e., \whp{} far away from its expected value. This allows us to deduce that~$\xx$ has~\whp{} the wrong number of $(G,H)$-extensions (since by Theorem~\ref{thm_generaltail}~\ref{thm_tail1} \whp{} each copy of~$J_{\max}$ again extends to the right number of copies of~$H$); see Section~\ref{sec:ext:0} for~the~details. \subsection{Setup and technical preliminaries} In the upcoming arguments it will, as in~\cite{S90b}, often be convenient to treat extensions as sequences of vertices. Given a rooted graph~$(G,H)$ with labeled vertices~$V(G) = {\left\{ 1, \dots, v_G \right\}}$ and~${V(H) \setminus V(G)} = {\{v_G + 1, \dots, v_H\}}$, \emph{an ordered $(G,H)$-extension of} $\xx = {(x_1, \dots, x_{v_G})} \in \oset{v_G}$ is a sequence $\yy = {(y_{v_G + 1}, \dots, y_{v_H})}$ of distinct vertices from~$[n] \setminus \{x_1, \dots, x_{v_G}\}$ such that the injection which maps each vertex~${j \in V(G)}$ onto $x_j$ and each vertex~${i \in {V(H)\setminus V(G)}}$ onto~$y_i$, also maps every edge~${f \in {E(H) \setminus E(G)}}$ onto an edge. Given a root~$\xx \in \oset{v_G}$, let~$Y_{G,H}(\xx)$ denote the number of ordered~$(G,H)$-extensions of~$\xx$ in~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$. Note that \begin{equation}\label{def:nuGH} \nu_{G,H} := \E Y_{G,H}(\xx) = (n-v_G) (n-v_G-1) \cdots (n-v_H+1) \cdot p^{\eGH} \end{equation} does not depend on the particular choice of~$\xx$. Let~$\operatorname{aut}(G,H)$ denote the number of automorphisms of~$H$ that fix the set~$V(G)$. Since each extension corresponds to~$\operatorname{aut}(G,H)$ many ordered extensions, we~obtain \begin{align} \label{eq:YX} Y_{G,H}(\xx) &= \operatorname{aut}(G,H) \cdot X_{G,H}(\xx) , \\ \label{eq:numu} \nu_{G,H} &= \operatorname{aut}(G,H) \cdot \mu_{G,H}. \end{align} One further useful elementary observation is that, for any induced~$G \subseteq J \subseteq H$, we have \begin{equation}\label{eq:muprod} \nu_{G,J} \cdot \nu_{J,H} = \nu_{G,H}. \end{equation} Our arguments will also exploit the following technical property of maximal primal~subgraphs. \begin{lemma}\label{lem:nogroundeddensity}% If~$J_{\max} \subsetneq H$ is a maximal primal of the rooted graph~$(G,H)$, \linebreak[3] then~$m(J_{\max},H) < m(G,H)$. \end{lemma} \begin{proof} Fix~$J_{\max} \subsetneq J \subseteq H$. Using maximality of~$J_{\max} \supsetneq G$, we infer~$d(G, J) < m(G,H)$ and~$d(G, J_{\max} ) = m(G,H)$. Proceeding analogous to inequality~\eqref{eq:lem:density:subs:3}, it routinely follows that \[ d(J_{\max},J) = \frac{(\vr{G}{J})d(G,J)-(\vr{G}{J_{\max}})d(G,J_{\max})}{(\vr{G}{J})-(\vr{G}{J_{\max}})} < m(G,H), \] which completes the proof by maximizing over all feasible~$J$. \end{proof} \subsection{The $0$-statement}\label{sec:ext:0} As discussed, for the $0$-statement of Theorem~\ref{thm_nogrounded} the core idea is to show that~$X_{G,J_{\max}}(\xx)$ is not concentrated for some~$\xx \in \oset{v_G}$, and that~$X_{J_{\max},H}(\y)$ is concentrated for all~$\y \in \oset{v_{J_{\max}}}$, see~\eqref{eq:badJ}--\eqref{eq:sparser}~below. \begin{proof}[Proof of the $0$-statement of Theorem~\ref{thm_nogrounded}] Assuming~$\eps \ge n^{-\alpha}$ with~$\alpha < 1/2$ (as we may), we have $\eps^2\Phi_{G,H} = \Omega(n^{1-2\alpha}p^{\eGH}) \gg p^{\eGH}$, so the assumption~$\eps^2\Phi_{G,H} \to 0$ implies~$p \to 0$ and thus~$1-p=\Theta(1)$. Since $(G,H)$ has no grounded primals, the desired $0$-statement now follows by combining the conclusions of Theorem~\ref{thm_generaltail}~\ref{thm_tail0} for~$\Phi_{G,H} \to 0$ and~$\Phi_{G,H} \to \infty$ with the conclusion of Lemma~\ref{lem:generalzero} below for~$\Phi_{G,H} \asymp 1$ (formally using, as usual, the subsubsequence principle~\cite[Section~1.2]{JLR}). \end{proof} \begin{lemma}\label{lem:generalzero}% Let~$(G,H)$ be a rooted graph with no grounded primal subgraphs. Then, for all~$p=p(n) \in [0,1]$ and~$\eps=\eps(n) \in (0,1]$ with~$\Phi_{G,H} \asymp 1$ and~$\eps \to 0$, \begin{equation}\label{eq:lem:generalzero} \lim_{n \to \infty} \Pr\Bigpar{\max_{\xx \in \oset{v_G}}|X_\xx - \mu| \ge \eps \mu} = 1 . \end{equation} \end{lemma} \begin{proof}% Note that we may assume $\eps \ge n^{-\alpha}$ for any $\alpha > 0$ (since increasing~$\eps$ gives a stronger conclusion). Let~$J_{\max}$ be a maximal primal subgraph of~$(G,H)$. By \refR{rem:Phibig}~\ref{eq:Phibig:ii}--\ref{eq:Phibig:iii}, the assumption~$\Phi_{G,H} \asymp 1$~implies \begin{gather} \label{eq:muGJmax} \mu_{G,J_{\max}} \asymp 1 ,\\ \label{eq:conditionPhip} p = \Omega\bigl(n^{-1/m(G,H)}\bigr) . \end{gather} Turning to the details, we start with the claim that, \whp{}, \begin{align} \label{eq:badJ} \max_{\xx \in \oset{v_G}} |X_{G,J_{\max}}(\xx) - \mu_{G,J_{\max}}| &> 3\eps \mu_{G,J_{\max}}, \\ \label{eq:sparser} \max_{\y \in \oset{v_{J_{\max}}}}|X_{J_{\max},H}(\y) - \mu_{J_{\max},H}| &< \tfrac{1}{2}\eps \mu_{J_{\max},H} . \end{align} To show that this claim implies the desired $0$-statement, we consider ordered extensions and note that multiplying~\eqref{eq:badJ} and~\eqref{eq:sparser} by $\operatorname{aut}(G, J_{\max})$ and $\operatorname{aut}(J_{\max},H)$, respectively, we can replace~$X$~by~$Y$ and~$\mu$~by~$\nu$, cf.~\eqref{eq:YX} and \eqref{eq:numu}. Observe that each ordered $(G,H)$-extension corresponds to a unique pair of extensions: one of~$\xx$ with respect to~$(G,J_{\max})$ and one of~$\y$ (which consists of~$\xx$ plus the vertices of the first extension) with respect to $(J_{\max},H)$. Consequently, recalling the identity~\eqref{eq:muprod}, inequalities~\eqref{eq:badJ}--\eqref{eq:sparser} imply that there is~$\xx \in \oset{v_G}$ such that either \begin{equation}\label{eq:YGHR:1} Y_{G,H}(\xx) > (1 + 3\eps)\nu_{G,J_{\max}} \cdot (1 - \eps/2) \nu_{J_{\max},H} > (1 + \eps) \nu_{G,H} \end{equation} or \begin{equation}\label{eq:YGHR:2} Y_{G,H}(\xx) < (1 - 3\eps)\nu_{G,J_{\max}} \cdot (1 + \eps/2) \nu_{J_{\max},H} < (1 - \eps) \nu_{G,H} , \end{equation} which in view of~\eqref{eq:YX} and \eqref{eq:numu} establishes the desired $0$-statement (after rescaling by~$\operatorname{aut}(G,H)$). It remains to show that~\eqref{eq:badJ} and \eqref{eq:sparser} hold \whp{}, and we start with~\eqref{eq:badJ}. Consider the unrooted graph~$K := {J_{\max} - V(G)}$, where the vertices of~$G$ are deleted from~$J_{\max}$. By construction, we have~$v_K = \vr{G}{ J_{\max}}$. Since~$J_{\max}$ is not grounded, we also have~$e_K = \er{G}{ J_{\max}}$. Using~\eqref{eq:muGJmax} we infer \begin{equation}\label{eq:muK:Jmax} \mu_K \asymp n^{v_K}p^{e_K} = n^{\vr{G}{J_{\max}}}p^{\er{G}{J_{\max}}} \asymp \mu_{G,J_{\max}} \asymp 1 , \end{equation} which by Markov's inequality implies that the number of $K$-copies is \whp{} at most~$n/(2v_K)$, say (with room to spare). This means that either (i)~there are no $K$-copies, in which case $X_{G,J_{\max}}(\xx) = 0$ for all $\xx \in \oset{v_G}$, or (ii)~the $K$-copies span at most~$n/2$ vertices, in which case there is one $\xx_1 \in \oset{v_G}$ that is disjoint from all $K$-copies and another set~$\xx_2 \in \oset{v_G}$ that intersects at least one $K$-copy, so that~$X_{G,J_{\max}}(\xx_1) = X_K > X_{G,J_{\max}}(\xx_2)$. In both cases it follows that~\eqref{eq:badJ} holds \whp{}, since~\eqref{eq:muGJmax} and~$\eps \to 0$ imply that the interval~$(1 \pm 3\eps)\mu_{G,J_{\max}}$ does not contain zero, and moreover contains at most one~integer. Turning to~\eqref{eq:sparser}, note that~\eqref{eq:sparser} holds trivially when~$J_{\max} = H$. Otherwise~$m(J_{\max},H) < m(G,H)$ by Lemma~\ref{lem:nogroundeddensity}, so that~\eqref{eq:conditionPhip} implies $p = \Omega( n^{\gamma-1/m(J_{\max},H)} )$ for some constant $\gamma > 0$. Using \refR{rem:Phibig}~\ref{eq:Phibig:iv}, it follows that~$\Phi_{J_{\max},H} = \Omega(n^\gamma)$. Assuming~$\eps \ge n^{-\alpha}$ with~$\alpha < \gamma / 2$ (as we may), we infer $\eps^2\Phi_{J_{\max},H} = \Omega(n^{\gamma/ 2 -\alpha}) = n^{\Omega(1)}$. Applying Theorem~\ref{thm_generaltail}~\ref{thm_tail1} with~$t = \tfrac{1}{2}\eps \mu_{J_{\max},H}$, now~\eqref{eq:sparser} holds~\whp{}. \end{proof} \subsection{The $1$-statement}\label{sec:ext:1} As discussed, for the $1$-statement of Theorem~\ref{thm_nogrounded} we rely on the fact that no vertex is contained in too many copies of the (unrooted) graph~${J_{\max} - V(G)}$, which is formalized by \refL{lem:scattered} below. As usual, given a graph~$K$ with~$v_K \ge 1$, subgraphs~$J \subseteq K$ with~$v_J \ge 1$ that maximize the density $d_J := d(\emptyset, J) = e_J/v_J$ are called~\emph{primal} (consistently with rooted graphs terminology), and~$K$ is called~\emph{balanced} when~$K$ itself is~primal. \begin{lemma}\label{lem:scattered}% Let $K$ be a balanced graph with $e_K \ge 1$. There are constants $\beta, C > 0$ such that, for all $p=p(n) \in [0,1]$ with~$n^{\beta - 1/d_K} \ll p = O(n^{\beta - 1/d_K})$, in ${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$ \whp{} every vertex~$x \in [n]$ is contained in at most~$C\lambda^{\vr{G_{\min}}{K}}$ copies of~$K$, where~$\lambda := np^{d_K}$ and~$G_{\min} \subseteq K$ is a primal subgraph with the smallest number of vertices. \end{lemma} \noindent We defer the density based proof to \refApp{apx:scattered} (which is rather tangential to the main~argument here), and now use \refL{lem:scattered} to prove the desired $1$-statement~of~Theorem~\ref{thm_nogrounded}. \begin{proof}[Proof of the $1$-statement of~\refT{thm_nogrounded}]% The assumptions~$\eps \le 1$ and~$\eps^2\Phi_{G,H} \to \infty$ imply~$\Phi_{G,H} \to \infty$, so \refR{rem:Phibig}~\ref{eq:Phibig:i} implies~$p\gg~n^{-1/m(G,H)}$. If~$\eps^2\Phi_{G,H} = n^{\Omega(1)}$, then the desired $1$-statement follows from Theorem~\ref{thm_generaltail}~\ref{thm_tail1}, so we may further assume~$\eps^2 \Phi_{G,H} \le n^c$ for any constant~$c>0$ of our choice, which together with the assumption~$\eps \ge n^{-\alpha}$ implies $\Phi_{G,H} \le n^{c + 2\alpha}$. Using the contrapositive of \refR{rem:Phibig}~\ref{eq:Phibig:iv}, by choosing~$\alpha,c>0$ sufficiently small (as we may) we thus henceforth can~assume \begin{equation} \label{eq:conditionPhi} n^{-1/m(G,H)} \ll p \ll n^{\beta - 1/m(G,H)}, \end{equation} where the constant~$\beta > 0$ is as given by~Lemma~\ref{lem:scattered}. Turning to the details, let~$J_{\max}$ be a maximal primal subgraph of~$(G,H)$. For convenience we use ordered extensions, as before. Note that~$Y_{G,J_{\max}}(\xx)$ is the number of (unrooted) copies of graph~$K := {J_{\max} - V(G)}$ that are disjoint from~$\xx$. Let~$Z_K(x)$ be the number of copies of~$K$ containing vertex~$x \in [n]$. We fix some~$\xx' \in \oset{v_G}$, and start with the claim that there exists a constant $D>0$ such that, \whp{}, \begin{align} \label{eq:XRGprime} |Y_{G,J_{\max}}(\xx') - \nu_{G,J_{\max}}| &< \tfrac 1 8\eps\nu_{G,J_{\max}}. \\ \label{eq:goodcount} \max_{\y \in \oset{v_{J_{\max}}}} |Y_{J_{\max},H}(\y) - \nu_{J_{\max},H}| &< \tfrac{1}{2}\eps \nu_{J_{\max},H},\\ \label{eq:vertexbounded} \max_{x \in [n]} Z_K(x) & \le D\frac{\eps\nu_{G, J_{\max}}}{\eps^2\Phi_{G,H}}. \end{align} We now show that this claim implies the desired $1$-statement. In view of~\eqref{eq:XRGprime}, the first step is to use~\eqref{eq:vertexbounded} to show that $Y_{G,J_{\max}}(\xx)$ is also concentrated for the remaining roots $\xx \neq \xx'$. Namely, using~\eqref{eq:vertexbounded} to bound the number of $(G, J_{\max})$-extensions of $\xx$ that overlap with $\xx'$ (and those of $\xx'$ overlapping with $\xx$), in view of~$\eps^2\Phi_{G,H} \to \infty$ it follows that, for every~$\xx \in \oset{v_G}$, \begin{equation*} |Y_{G,J_{\max}}(\xx) - Y_{G,J_{\max}}(\xx')| \le O\Bigpar{\sum_{x \in \xx \cup \xx'} Z_K(x)} \ll \tfrac 1 8 \eps\nu_{G,J_{\max}} , \end{equation*} which together with \eqref{eq:XRGprime} implies that, say, \begin{equation}\label{eq:XRG} \max_{\xx \in \oset{v_G}} |Y_{G,J_{\max}}(\xx) - \nu_{G,J_{\max}}| < \tfrac 1 4 \eps\nu_{G,J_{\max}} . \end{equation} The second step exploits that by~\eqref{eq:goodcount} each copy of $J_{\max}$ extends to the `right' number of copies of~$H$. Indeed, with analogous reasoning as for~\eqref{eq:YGHR:1}--\eqref{eq:YGHR:2} from \refS{sec:ext:0}, by combining~\eqref{eq:goodcount} with~\eqref{eq:XRG} it now follows (in view of~\eqref{eq:muprod}) that \[ \max_{\xx \in \oset{v_G}} Y_{G,H}(\xx) < (1 + \eps/4)\nu_{G,J_{\max}} \cdot (1 + \eps/2)\nu_{J_{\max},H} < (1 + \eps)\nu_{G,H} , \] and similarly, \[ \min_{\xx \in \oset{v_G}} Y_{G,H}(\xx) > (1 - \eps)\nu_{G,H} , \] which in view of~\eqref{eq:YX}--\eqref{eq:numu} establishes the $1$-statement of Theorem~\ref{thm_nogrounded} (by rescaling by $\operatorname{aut}(G,H)$). It remains to show that~\eqref{eq:XRGprime}--\eqref{eq:vertexbounded} hold \whp{}, and we start with~\eqref{eq:XRGprime}. Since $\Phi_{G,J_{\max}} \ge \Phi_{G,H}$ by definition, using Chebyshev's inequality together with the variance estimate~\eqref{eq:Variance} and $\eps^2 \Phi_{G,H} \to \infty$, it follows that \begin{equation* \begin{split} \prob{|X_{G,J_{\max}}(\xx') - \mu_{G,J_{\max}}| \ge \tfrac 1 8\eps\mu_{G,J_{\max}}} & \le \frac{\Var X_{G,J_{\max}}(\xx)}{(\eps/8)^2\mu_{G,J_{\max}}^2} \asymp \frac{1 - p}{\eps^2\Phi_{G,J_{\max} }} \le \frac{1}{\eps^2\Phi_{G,H}} \to 0 , \end{split} \end{equation*} which in view of~\eqref{eq:YX}--\eqref{eq:numu} then implies that~\eqref{eq:XRGprime} holds \whp{} (by rescaling by $\operatorname{aut}(G,J_{\max})$). Next we establish~\eqref{eq:goodcount}. Note that the proof of~\eqref{eq:sparser} only relies on~\eqref{eq:conditionPhip} (which here holds by \eqref{eq:conditionPhi}), and that we may assume~$\eps \ge n^{-\alpha}$ for sufficiently small~$\alpha>0$. Hence by the same argument as for~\eqref{eq:sparser}, in view of~\eqref{eq:YX}--\eqref{eq:numu} it here follows that~\eqref{eq:goodcount} holds~\whp{} (after rescaling by~$\operatorname{aut}(J_{\max},H)$). Finally, we turn to the auxiliary estimate~\eqref{eq:vertexbounded}. Note that every subgraph~$J \subseteq K$ with~$v_J \ge 1$ satisfies~$d_J = d(G,G \cup J)$. Hence~$J$ is primal for~$K$ if and only if~$G \cup J$ is primal for~$(G,J_{\max})$. Since~$J_{\max}= G \cup K$ is primal for~$(G,H)$, it follows that~$K$ is balanced, with~$d_K=d(G,J_{\max})=m(G,H)$. Using assumption~\eqref{eq:conditionPhi}, we thus have~$n^{-1/d_K} \ll p \ll n^{\beta-1/d_K}$. Invoking Lemma~\ref{lem:scattered} there is a constant~$C>0$ such that,~\whp{}, \[ \max_{x \in [n]} Z_K(x) \le C \lambda^{\vr{G_{\min}}{K}}, \] where $G_{\min}$ is a smallest primal for~$K$, which in turn gives~$d_K=d_{G_{\min}}$. Since~$J_{\max}$ is a vertex-disjoint union of the graphs~$K$ and~$G$, and~$G_{\min} \subseteq K$ we infer that~$e_{G_{\min}} = \er{G}{G \cup G_{\min}}$ and~$v_{G_{\min}} =\vr{G}{G \cup G_{\min}}$. Recalling~$\lambda = np^{d_K} = np^{d_{G_{\min}}}$, it now follows analogously to~\eqref{eq:muK:Jmax} that \begin{equation*} \lambda^{\vr{G_{\min}}{K}} = \frac{n^{v_K} p^{e_K}}{n^{v_{G_{\min}}} p^{e_{G_{\min}}}} \asymp \frac{\mu_K}{\mu_{G_{\min}}} \asymp \frac{\mu_{G, J_{\max}}}{\mu_{G,G \cup G_{\min}}} \le \frac{\mu_{G, J_{\max}}}{\Phi_{G,H}} , \end{equation*} which together with~$1 \le 1/\eps = \eps/\eps^2$ and~\eqref{eq:numu} completes the proof of~\eqref{eq:vertexbounded} for suitable~$D>0$. \end{proof} \section{Further cases}\label{sec:further} \subsection{Unique and grounded primal case (\refT{thm_unique})}\label{s_unique} In this section we prove \refT{thm_unique} by adapting the arguments from \refS{s_nogrounded} (focusing on the unique primal~$J=J_{\max}$). The key difference is that here we can use the $0$- and $1$-statements of our main result \refT{thm_strictly_balanced} to deduce that~$X_{G,J}(\xx)$ is not concentrated for some~$\xx$, or concentrated for all~$\xx$, respectively. This then allows us to prove the desired $0$- and $1$-statements, since each copy of~$J$ again extends to the `right' number of copies of~$H$ (by Theorem~\ref{thm_generaltail}~\ref{thm_tail1}, as in Section~\ref{s_nogrounded}); see \eqref{eq:JHconc}--\eqref{eq:GJconc} and~\eqref{eq:GJoff} below. \begin{proof}[Proof of \refT{thm_unique}]% If~$\Phi_{G,H} \to 0$, then the $0$-statement holds by \refT{thm_generaltail}~\ref{thm_tail0}. Therefore we henceforth can assume~$\Phi_{G,H} = \Omega(1)$, which by \refR{rem:Phibig}~\ref{eq:Phibig:ii} is equivalent to \begin{equation}\label{eq:overthreshold} p = \Omega\bigl(n^{-1/m(G,H)}\bigr). \end{equation} Note that the proof of~\eqref{eq:sparser} relies only on~\eqref{eq:conditionPhip} (which is the same as \eqref{eq:overthreshold}), the fact that $J_{\max}$ is the maximal primal (which also holds trivially for~$J$ in the current setting), and that we may assume $\eps \ge n^{-\alpha}$ for sufficiently small~$\alpha>0$ (which we may also assume here). Hence by the same argument as for~\eqref{eq:sparser}, after rescaling by~$\operatorname{aut}(J,H)$ (see~\eqref{eq:YX}--\eqref{eq:numu}) we here obtain that, \whp{}, \begin{equation}\label{eq:JHconc} \max_{\y \in \oset{v_J}}|Y_{J,H}(\y) - \nu_{J,H}| < \tfrac{1}{2} \eps \nu_{J,H}. \end{equation} We start with the~$1$-statement. Since~$\mu_{G,J} \ge \Phi_{G,H}$ by definition, the assumption~$\eps^2\Phi_{G,H} \ge C \log n$ implies~$\eps^2\mu_{G,J} \ge C \log n$. By uniqueness of the primal~$J$, the rooted graph $(G,J)$ is strictly balanced. Therefore~\eqref{eq:main:strbal} of~\refT{thm_strictly_balanced} implies (after rescaling by~$\operatorname{aut}(G,J)$) for suitable~$\alpha, C > 0$ that, \whp{}, \begin{equation}\label{eq:GJconc} \max_{\xx \in \oset{v_G}} |Y_{G,J}(\xx) - \nu_{G,J}| < \tfrac{1}{4}\eps \nu_{G,J}. \end{equation} The $1$-statement of \refT{thm_unique} now follows from~\eqref{eq:JHconc} and~\eqref{eq:GJconc} by exactly the same reasoning with which~\eqref{eq:goodcount} and~\eqref{eq:XRG} from Section~\ref{sec:ext:1} implied the $1$-statement of \refT{thm_nogrounded}. We now turn to the~$0$-statement. We again plan to apply~\eqref{eq:main:strbal} of~\refT{thm_strictly_balanced} to the strictly balanced rooted graph~$(G,J)$, for which we need to check that the assumption~$\eps^2 \Phi_{G,H} \le c \log n$ implies the required condition~$\eps^2 \mu_{G,J} \le c\log n$. We will do this by showing that~$\Phi_{G,H} = \mu_{G,J}$ for~$n$ large enough. First, note that the assumptions~$\eps \ge n^{-\alpha}$ and~$\eps^2 \Phi_{G,H} \le c \log n$ imply~$\Phi_{G,H} = O(n^{2\alpha}\log n)$. By~\eqref{eq:overthreshold} and the contrapositive of \refR{rem:Phibig}~\ref{eq:Phibig:iv} we can thus assume that, say, \begin{equation}\label{eq:closetohreshold} p \asymp n^{\theta - 1/m(G,H)} \quad \text{ with } \quad \theta = \theta(n,p) \in [0,3\alpha]. \end{equation} Since the primal~$J$ is unique, we have~$d(G,J)=m(G,H)$, and~$d(G,K)< m(G,H)$ when~$G \subsetneq K \subseteq H$ satisfies $J \neq K$. Hence there exists a constant~$\gamma=\gamma(G,J,H)>0$ such that, for any $G \subsetneq K \subseteq H$, \begin{equation}\label{eq:mu:minimal} \mu_{G,K} \asymp \left( np^{d(G,K)} \right)^{\vr{G}{K}} \asymp \left( n^{1 - \frac{d(G,K)}{m(G,H)} + \theta d(G,K)} \right)^{\vr{G}{K}} = \begin{cases} \Omega(n^{\gamma}) & \text{if~$K \neq J$,} \\ O(n^{3\alpha \er{G}{J}}) & \text{if~$K = J$.} \end{cases} \end{equation} By taking~$\alpha > 0$ small enough, it follows that~$\Phi_{G,H} = \mu_{G,J}$ for~$n$ large enough, which (as discussed) establishes~$\eps^2 \mu_{G,J} \le c \log n$. Therefore~~\eqref{eq:main:strbal} of~\refT{thm_strictly_balanced} implies (after rescaling by $\operatorname{aut}(G,J)$) that, \whp{}, \begin{equation}\label{eq:GJoff} \max_{\xx \in \oset{v_G}} |Y_{G,J}(\xx) - \nu_{G,J}| \ge 3\eps \nu_{G,J}. \end{equation} The $0$-statement of \refT{thm_unique} now follows from~\eqref{eq:GJoff} and~\eqref{eq:JHconc} by the same (routine) reasoning with which~\eqref{eq:badJ}--\eqref{eq:sparser} from Section~\ref{sec:ext:0} implied the $0$-statement of \refT{thm_nogrounded}. \end{proof} \begin{remark}[Theorem~\ref{thm_unique}: stronger $1$-statement]\label{rem:thm_unique1} The above proof yields, in view of Remarks~\ref{rem:thm_generaltail}--\ref{rem:thm_strictly_balanced}, the following stronger conclusion: for any fixed~$\tau > 0$ there is a constant~$C=C(\tau,G,H)>0$ such that the~$1$-statement in~\eqref{eq:thm:unique} of Theorem~\ref{thm_unique} holds with probability~$1 - o(n^{-\tau})$. \end{remark} \subsection{Strictly balanced and ungrounded case (\refT{thm_strictly_balanced})}\label{sec:ext:non} In this section we prove the threshold~\eqref{eq:main:strbal:non} of~\refT{thm_strictly_balanced}~(ii) for strictly balanced rooted graphs~$(G,H)$ that are not grounded, which turns out to be a simple corollary of~\refT{thm_nogrounded}. The crux is that, by decreasing~$\alpha>0$ (if~necessary), we can ensure that the {$0$-}~and {$1$-statement} conditions in~\eqref{eq:main:strbal:non} and~\eqref{eq:main:nogrounded} coincide. \begin{proof}[Proof of~\eqref{eq:main:strbal:non} of~\refT{thm_strictly_balanced}] By assumption the unique primal~$H$ is not grounded, so \refT{thm_nogrounded} applies. Decreasing the constant~$\alpha>0$ from \refT{thm_nogrounded}, we can assume that~$\beta \ge 3\alpha$, where~$\beta>0$ is the constant given by \refL{lem:StrBal}~\ref{eq:StrBal:density}. We now distinguish two ranges of~$p=p(n)$. First, when~$p \le n^{-1/d(G,H)+\beta}$, then~\eqref{eq:lem:density:subs} from \refL{lem:StrBal} implies that~$\mu=\Phi$ for~$n$ large enough (since~\eqref{eq:lem:density:subs} implies~$\mu_{G,H}/\mu_{G,J} \asymp n^{v_H - v_J}p^{e_H - e_J} \ll 1$ for all~$G \subseteq J \subsetneq H$ with~$e_J > e_G$). Second, when~$p \ge n^{-1/d(G,H)+\beta} \ge n^{-1/m(G,H)+3\alpha}$, then~$\eps \ge n^{-\alpha}$ and \refR{rem:Phibig}~\ref{eq:Phibig:iv} imply that~$\min\{\eps^2 \mu, \eps^2\Phi\} \ge n^{-2\alpha} \cdot \Phi = \Omega(n^{\alpha}) \to \infty$. Since in both ranges the {$0$-}~and {$1$-statement} conditions in~\eqref{eq:main:strbal:non} and~\eqref{eq:main:nogrounded} coincide, it follows that \refT{thm_nogrounded} implies~\eqref{eq:main:strbal:non}. \end{proof} \section{Cautionary examples (\refP{prop:counterexample:2} and~\ref{prop:counterexample:1})}\label{sec:counterexample} In this section we prove Propositions~\ref{prop:counterexample:2}--\ref{prop:counterexample:1} for the rooted graphs~(e) and~(f) depicted in~\refF{counterexample}. The proof idea for~\refP{prop:counterexample:2} is to proceed in two rounds for a fixed vertex~$x$: using~\refT{thm_strictly_balanced} we first find about~$\mu_{G,K_4}$ many~$(G,K_4)$-extensions of~$\xx=(x)$, which we then extend to about~$\mu_{G,H}$ many~$(G,H)$-extensions of~$\xx$. The crux is that most of the relevant~$(K_4,H)$-extensions from the second round evolve nearly independently, which ultimately allows us to surpass the conditions of Spencer's result~\eqref{eq_Spencer_SB} and~\refT{thm_strictly_balanced} for~$(K_4,H)$. \begin{proof}[Proof of \refP{prop:counterexample:2} Recalling~$\omega = np^2$, by assumption we have~$\eps^2 \mu_{G,K_4} \asymp \eps^3\omega^3 \gg \log n$ and~$\eps^2 \mu_{K_4,H} \le \mu_{K_4,H} \asymp \omega \ll \log n$, which readily implies~$\log \omega \asymp \log \log n$ and~$p=n^{-1/2+o(1)}$. Now it is not difficult to verify that~$\Phi_{G,H} \asymp \mu_{G,K_4} \asymp \omega^3$ (either directly, or similarly as for~\eqref{eq:mu:minimal} from \refS{s_unique}). Turning to the details of the $1$-statement, we start with the auxiliary claim that, whp, for each vertex~$x$ the following event~$\cP_x$~holds: \begin{romenumerate2} \item\label{eq:Pv:i The vertex-neighbourhood~$\Gamma_x$ of~$x$ has size~$|\Gamma_x| \le 9np$. \item\label{eq:Pv:ii The collection~$\cT_x$ of all triangles spanned by~$\Gamma_x$ has size~$|\cT_x| = (1\pm \eps/9) \tbinom{n-1}{3}p^6$. \item\label{eq:Pv:iii Every vertex~$y \in \Gamma_x$ is contained in at most~$D:=15$ triangles from~$\cT_x$. \end{romenumerate2} Indeed, invoking the $1$-statement of \refT{thm_strictly_balanced} with~$H$ equal to~$K_4$ and~$G$ being the root vertex~$v$, from $\eps ^2 \mu_{G,K_4} \asymp \eps^2\omega^3 \gg \log n$ it follows that, whp,~\ref{eq:Pv:ii} holds for all vertices~$x$. Since~$np = n^{1/2+o(1)} \gg \log n$, using standard Chernoff bounds it is routine to see that, whp,~\ref{eq:Pv:i} holds for all vertices~$x$. We claim that if~\ref{eq:Pv:iii}~fails for some~$y \in \Gamma_x$, then there are $4$~triangles in~$\cT_x$ containing $y$ that form either a flower (share no vertices other than~$y$) or a book (all contain~$yz$ for some~$z \in \Gamma_x$): to see this, note that if we assume the contrary, then for a maximal flower (with at most~$3$ triangles) each edge of it is contained in at most~$2$ other $\cT_x$-triangles, whence there are at most~$3 + 6 \cdot 2 = 15$ triangles in~$\cT_x$ containing~$y$. The probability that there is either a \mbox{$4$-flower} or \mbox{$4$-book} with all vertices connected to some extra vertex~$x$ is at most $n^{10}p^{21} + n^7p^{16} = n^{-1/2 + o(1)} \to 0$. It follows that, whp, properties~\ref{eq:Pv:i}--\ref{eq:Pv:iii} hold for all vertices~$x$, establishing the~claim. We now fix a root vertex~$x$, and expose the edges of~${\mathbb G}_{n,p}} % \newcommand{\Gnp}{\G(n,p)$ in two rounds: in the first round we expose all edges incident to~$x$ and all edges inside~$\Gamma_x$, and then in the second round we expose all remaining edges. We henceforth condition on the outcome of the first exposure round, and assume that~$\cP_x$ holds. As usual, to avoid clutter we shall omit this conditioning from our notation. Given distinct vertices~$a,b \in \Gamma_x$, let~$Y_{a,b}$ denote the number of common neighbours of~$a$ and~$b$ in~$[n] \setminus (\{x\} \cup \Gamma_x)$. Note that~$\eps \omega \gg \sqrt{\log n/\omega} \gg 1$ by assumption. Since, by~\ref{eq:Pv:ii}, $|\cT_x| \asymp n^3p^6 = \omega ^3 \ll \eps \omega^4 \asymp \eps \mu$, using~\ref{eq:Pv:iii} it is not difficult to see that \begin{equation}\label{eq:Xx} Z_x := \sum_{abc \in \cT_x}(Y_{a,b}+Y_{b,c} + Y_{a,c}) \qquad \text{satisfies} \qquad \bigl|X_{(x)}-Z_x\bigr| \: \le \: 3|\cT_x| \cdot D \ll \eps \mu/2 . \end{equation} Using~\ref{eq:Pv:i} and~$\eps \omega \gg 1$ (see above) we infer $1+|\Gamma_x| \le 10np = n^{1/2+o(1)} \ll n/\omega \ll \eps n$, and together with~\ref{eq:Pv:ii} it then follows that, say, \begin{equation}\label{eq:EZv} \E Z_x = 3|\cT_x| \cdot \bigpar{n-1-|\Gamma_x|} p^2 = (1\pm \eps/8) \mu . \end{equation} In~\eqref{eq:Xx} we now write each~$Y_{a',b'}$ as a sum of indicators of length~$2$~paths, which enables us to estimate the lower tail of~$Z_x$ via Janson's inequality. By distinguishing between pairs of edge-overlapping paths that share one or two endpoints, using~\ref{eq:Pv:iii} it is standard to see that the relevant~$\Delta$ term is at most~$\E Z_x \cdot (2D p + 2D) = O(\E Z_x)$, say. Using~$\eps^2 \E Z_x \asymp \eps^2 \omega^4 \gg \log n$, by invoking~\cite[Theorem~1]{RiordanWarnke2015} it then follows~that \begin{equation}\label{eq:Zv:upper} \Pr(Z_x \le (1-\eps/8) \E Z_x) \le \exp\bigpar{- \Omega(\eps^2 \E Z_x)} = o(n^{-1}) . \end{equation} Using~\ref{eq:Pv:iii} we also see that any path shares an edge with a total of at most~$2 D = O(1)$ paths, which enables us to estimate the upper tail of~$Z_v$ via concentration inequalities for random variables with `controlled dependencies'. In particular, by invoking~\cite[Proposition~2.44]{JLR} (see also~\cite[Theorem~9]{W14}) it follows~that \begin{equation}\label{eq:Zv:lower} \Pr(Z_x \ge (1+\eps/8) \E Z_x) \le \exp\bigpar{- \Omega(\eps^2 \E Z_x)} = o(n^{-1}) . \end{equation} To sum up, \eqref{eq:Xx}--\eqref{eq:Zv:lower} and~$1-\eps/2 < (1 \pm \eps/8)^2 < 1 + \eps/2$ imply~$\Pr(|X_{(x)}-\mu| \ge \eps \mu \mid \cP_x) = o(n^{-1})$, which readily completes the proof of the desired $1$-statement (since, whp,~$\cP_x$ holds for all~$n$ vertices~$x$). \end{proof} The proof idea for \refP{prop:counterexample:1} is to find a copy of~$K_4$ containing an edge with extremely large codegree. To this end we proceed in two steps, inspired by~\cite[Lemma~3]{SW18}: in the first steps we find~$\Theta(n)$ many vertex-disjoint copies of~$K_4$, and in the second step we then find the desired edge with large codegree. \begin{proof}[Proof of \refP{prop:counterexample:1}]% Note that~$\mu \asymp \omega^5$. As in the proof of \refP{prop:counterexample:2}, we again have~$\log \omega \asymp \log \log n$ and $\Phi_{G,H} \asymp \mu_{G,K_4} \asymp \omega^3$, so~$\eps^2 \Phi_{G,H}\asymp \eps^2 \mu_{G, K_4} \gg \log n$ by assumption. Noting~$0.39< 2/5$, we~define \begin{equation}\label{eq:CE:defz} z := \Bigceil{2 \bigpar{(1+\eps) \mu}^{1/2}} \asymp \omega^{5/2} = o(\log n/\log \omega) . \end{equation} Turning to the details of the desired $0$-statement, let~$X^v_{K_4}$ denote the size of the largest collection of vertex-disjoint copies of~$K_4$ spanned by the vertices in~$W:=\{1, \ldots, \lfloor n/2 \rfloor \}$. It is routine to check that the minimum of~${|W|^{v_G}p^{e_G}}$, taken over all~${G \subseteq K_4}$ with~${v_G \ge 1}$, equals~$|W| \approx n/2$ for~$n$ large enough. Since~${\mathbb G}_{n,p}[W]$ has the same distribution as~${\mathbb G}_{|W|,p}$, by invoking~\cite[Theorem~3.29]{JLR} there is a constant~$c>0$ such~that \begin{equation}\label{eq:XK4v} \Pr(X^v_{K_4} \ge cn) = 1 - o(1). \end{equation} We now condition on the edges spanned by~$W$, and assume that~$X^v_{K_4} \ge cn$. To avoid clutter, we shall again omit this conditioning from our notation (as in the proof of \refP{prop:counterexample:1}). We henceforth fix~$\ceil{cn}$ vertex-disjoint copies of~$K_4$ spanned by~$W$, and from the~$i$-th such copy we pick an edge~${\{v_i,w_i\}}$ and a further vertex~${x_i \not\in \{v_i,w_i\}}$. Defining~$Z_i$ as the number of vertices in~$[n] \setminus W$ that are common neighbours of~$v_i$ and~$w_i$, using~$\tbinom{m}{z} \ge (m/z)^z$ for~$m \ge z$ together with~$np^2 = \omega = o(z)$ and~\eqref{eq:CE:defz}, it routinely follows~that \begin{equation*} \Pr(Z_i \ge z) \ge \binom{\lceil n/2 \rceil}{z}p^{2z} \bigpar{1-p^2}^{\lceil n/2 \rceil-z} \ge \Bigpar{\frac{np^2}{2z}}^{z} \e^{-np^2} = \e^{-\Theta(z\log \omega)} \ge n^{-o(1)} . \end{equation*} Note that~$Z_i \ge z$ implies $X_{(x_i)} \ge \binom{z}{2} \ge \bigpar{z/2}^2 \ge (1+\eps) \mu$. Since the random variables~$Z_i$ depend on disjoint sets of independent edges, it then follows that \begin{equation* \Pr(\max_{x \in [n]}X_{(x)} < (1+\eps)\mu) \le \Pr(\max_{1 \le i \le \lceil cn \rceil}\hspace{-0.25em}Z_i < z) = \hspace{-0.125em}\prod_{1 \le i \le \lceil cn \rceil} \hspace{-0.375em} \Pr(Z_i < z) \le \Bigpar{1-n^{-o(1)}}^{\lceil cn \rceil} = o(1) . \end{equation*} Hence~$\Pr(\max_{x \in [n]}X_{(x)} \ge (1+\eps)\mu \mid X^v_{K_4} \ge cn ) = 1-o(1)$, which together with~\eqref{eq:XK4v} completes the~proof. \end{proof} \pagebreak[3] \section{Concluding remarks}\label{sec:conclusion} The results and problems of this paper can also be viewed through the lens of \emph{extreme value theory}, where a standard goal is to show that a (suitably shifted and normalized) maximum converges to a non-degenerate distribution. To see the connection, note that the proof of \refT{thm_strictly_balanced}~(i) describes an interval on which~$\max_{\xx \in \oset{v_G}} X_\xx$ is \whp{} concentrated. Our setting concerns discrete random variables (which can have complicated behaviour, cf.~\cite[Section 8.5]{FHR}), with a correlation structure that seems quite unusual for the field. Hence, as a first step, it would already be interesting to establish a `law of large numbers' result (even for a restricted class of~$(G,H)$, such as strictly balanced~ones), which is the content of the following~problem. \begin{problem}\label{prb:extreme_val} Determine for what rooted graphs~$(G,H)$ and edge probabilities~$p = p(n)$ there is a sequence~$(a_n)$ of real positive numbers such that~${(\max_{\xx} X_{\xx} - \mu)/a_n}$ converges to~$1$ in probability (as~$n \to \infty$). \end{problem} \pagebreak[3] \small \bibliographystyle{plain}

https://arxiv.org/abs/1706.08052

Mean perimeter and mean area of the convex hull over planar random walks

We investigate the geometric properties of the convex hull over $n$ successive positions of a planar random walk, with a symmetric continuous jump distribution. We derive the large $n$ asymptotic behavior of the mean perimeter. In addition, we compute the mean area for the particular case of isotropic Gaussian jumps. While the leading terms of these asymptotics are universal, the subleading (correction) terms depend on finer details of the jump distribution and describe a "finite size effect" of discrete-time jump processes, allowing one to accurately compute the mean perimeter and the mean area even for small $n$, as verified by Monte Carlo simulations. This is particularly valuable for applications dealing with discrete-time jumps processes and ranging from the statistical analysis of single-particle tracking experiments in microbiology to home range estimations in ecology.

\section{Introduction} Consider a set of $n$ points with position vectors $\{\vec r_1, \vec r_2,\ldots, \vec r_n\}$ in a $d$-dimensional space. The most natural and perhaps the simplest way to characterize the {\it shape} of this set of points is by drawing the convex hull around this set: a convex hull is the unique minimal convex polytope that encloses all the points. This unique polytope is convex since the line segment joining any two points on the surface of the polytope is fully contained within the polytope. Properties of such convex polytopes have been widely studied in mathematics, computer science (image processing and patter recognition) and in the physics of crystallography (the Wulff construction). In two dimension, where the convex hull is a polygon, there are many other applications, most notably in ecology where the home range of animals or the spread of an epidemics are typically estimated by convex hulls. For a review on the history and applications of convex hulls, see Ref.~\cite{Majumdar2010a}. When the points $\{\vec r_1, \vec r_2,\ldots, \vec r_n\}$ are drawn randomly from a joint distribution $P(\vec r_1, \vec r_2,\ldots, \vec r_n)$, the associated convex hull also becomes random and characterizing its statistical properties is a challenging problem, since the convex hull is a highly nontrivial functional of the random variables $\{\vec r_1, \vec r_2,\ldots, \vec r_n\}$. For instance, what can one say about the statistics of the surface area $S_d$ or the volume $V_d$ of the convex hull, for a given joint distribution $P(\vec r_1, \vec r_2,\ldots, \vec r_n)$? Even finding the mean surface area $\langle S_d\rangle$ or the mean volume $\langle V_d\rangle$, for arbitrary joint distribution, is a formidably difficult problem. In the special case when the points are independent and identically distributed, i.e., when the joint distribution factorizes as $P(\vec r_1, \vec r_2,\ldots, \vec r_n)= \prod_{k=1}^n P(\vec r_k)$ (with $P(\vec r_k)$ representing the marginal distribution), several results on the statistics of the surface and volume of the convex hull are known (see~\cite{Majumdar2010a} for a historical review). However, for {\it correlated} points where the joint distribution does not factorize, very few results are available. The simplest example of a set of correlated points corresponds to the case of a random walk in $d$-dimensional continuous space, where $\vec r_k$ represents the position of the walker at step $k$, starting at the origin at step $0$. The position evolves via the Markov rule, $\vec r_k= \vec r_{k-1}+ \vec \eta_k$, where $\vec \eta_k$ represents the jump at step $k$, and one assumes that $\vec \eta_k$ are independent and identically distributed random variables, each drawn from some prescribed distribution $p(\vec \eta_k)$. The walk evolves up to $n$ steps generating the vertices $\{\vec r_1, \vec r_2,\ldots, \vec r_n\}$ of its trajectory. There is a unique convex hull for each sample of this trajectory and what can one say about the mean surface area or the mean volume of this convex hull, given the jump distribution $p(\vec \eta_k)$? This is the basic problem that we address in this paper. We show that at least for $d=2$ (planar random walks), it is possible to obtain precise {\it explicit} results for all $n$ for the mean perimeter and the mean area of the convex hull of the walk, for a large class of jump distributions $p(\vec \eta)$, including in particular L\'evy flights where the jump distribution has a fat tail. We also obtain similar results for the mean area of the convex hull but under additional assumptions on the jump distribution. This problem concerning the convex hull of a random walk becomes somewhat simpler in the special case of the Brownian limit, where several results are known. Consider, for example a jump distribution $p(\vec \eta_k)$ with zero mean and a finite variance $\sigma^2$. In this case, the walk converges in the large $n$ limit to the Brownian motion. In other words, one can consider the continuous-time limit, as $\sigma^2\to 0$ and $n\to \infty$ with $n\sigma^2= 2\, D\, t$ being fixed (here $D$ is called the diffusion constant and $t$ is the duration of the walk). In this Brownian limit and for $d=2$, the mean perimeter and the mean area have been known exactly for a while. Tak\'acs~\cite{Takacs1980} computed the mean perimeter \begin{equation} \langle S_2\rangle = \sqrt{16 \pi\, D\, t}\, , \label{eq:perim_BM} \end{equation} while El Bachir~\cite{ElBachir83} and Letac~\cite{Letac93} computed the mean area \begin{equation} \langle V_2\rangle = \pi\, D\, t\, . \label{eq:area_BM} \end{equation} For a planar Brownian bridge of duration $t$ (where the walker returns to the origin after time $t$), the mean perimeter $\langle S_2\rangle_{\rm bridge} = \sqrt{\pi^3\, D\, t}$ was computed by Goldman~\cite{Goldman96}, while the mean area $\langle V_2\rangle_{\rm bridge}=(2\pi/3)\, D\, t$ was computed relatively recently by Randon-Furling {\it et al.}~\cite{Randon09}. An interesting extension of this problem in $d=2$ is to the case of $N$ independent planar Brownian motions (or Brownian bridges)~\cite{Randon09,Majumdar2010a}. This is relevant in the context of the home range of animals, where $N$ represents the size of an animal population and the trajectory of each animal is approximated by a Brownian motion during their foraging period. For a fixed population size $N$, the mean perimeter and the mean area of the convex hull was computed exactly: $\langle S_2\rangle = \alpha_N \, \sqrt{D\,t}$ and $\langle V_2 \rangle = \beta_N\, D\, t$, where the prefactors $\alpha_N$ and $\beta_N$ were found to have nontrivial $N$ dependence \cite{Randon09,Majumdar2010a}. For $d>2$, very few exact results are known for this problem. For a single $(N=1)$ Brownian motion, the mean surface area and the mean volume of the convex hull was recently computed by Eldan~\cite{Eldan2014}: $\langle S_d \rangle=\frac{2(4\pi D t)^{(d-1)/2}}{\Gamma(d)}$ and $\langle V_d\rangle= \frac{(\pi D t)^{d/2}}{\Gamma(d/2+1)^2}$ (see also \cite{Kabluchko16b} for another derivation and extension to Brownian bridges). However, for $N>1$ and $d>2$, no exact result is available. Finally, going beyond the mean surface and the mean volume, very few results are known for higher moments (see the review \cite{Majumdar2010a} for results on variance) or even the full distribution of the surface or the volume of the convex hull of Brownian motion (see Refs.~\cite{Wade15,Wade15b} for a recent discussion on the distribution of the perimeter in $d=2$ and $N=1$). Very recently, the full distribution (including the large deviation tails) of the perimeter and the area of $N\ge 1$ planar Brownian motions were calculated numerically~\cite{Claussen15,Dwenter16}. Some rigorous results on the convex hulls of L\'evy processes were recently derived \cite{Molchanov12,Molchanov16}. If one is interested only in the mean area or the mean volume of the convex hull of a generic stochastic process (not necessarily just a random walk), a particular simplification occurs in $d=2$ (planar case) where several analytical results can be derived by adapting Cauchy's formula~\cite{Cauchy1832,Santalo} for arbitrary closed convex curves in $d=2$. Indeed by employing Cauchy's formula for every realization of a random planar convex hull, it was shown in Refs.~\cite{Randon09,Majumdar2010a} that the problem of computing the mean perimeter and the mean area of an {\it arbitrary} two dimensional stochastic process (can in general be non-Markovian and in discrete-time) can be mapped to computing the extremal statistics associated with the one dimensional component of the process (see Section 3 for the precise mapping). This mapping was introduced originally in~\cite{Randon09} to compute $\langle S_2\rangle$ and $\langle V_2\rangle$ exactly for $N\ge 1$ planar Brownian motions. Since then, it has been used for a number of continuous-time planar processes: random acceleration process~\cite{Reymbaut11}, branching Brownian motion with applications to animal epidemic outbreak~\cite{Dumonteil13}, anomalous diffusion processes~\cite{Lukovic13} and also to a single Brownian motion confined to a half-space~\cite{Chupeau2015a}. The objective of this paper is to go beyond the continuous-time limit and obtain results for the convex hull of a discrete-time planar random walk of $n$ steps (with $n$ large but finite) with arbitrary jump distribution, including for instance L\'evy flights. Indeed, in any realistic experiment or simulation, the points of the trajectory are always discrete. For example, recently proposed local convex hull estimators \cite{Lanoiselee17} are based on a relatively small number of points, where we can not apply the Brownian limiting results reviewed above. The first rigorous result for a two-dimensional discrete random walk, modeled as a sum of independent random variables in the complex plane, was derived for the mean perimeter of the convex hull by Spitzer and Widom \cite{Spitzer1961}, \begin{equation} \label{eq:Spitzer} \langle L_n\rangle = 2\sum_{k=1}^n\frac{\langle \vert x_k + i y_k \vert \rangle}{k} \end{equation} (here $x_k+iy_k$ is the complex-valued position of the walker after $k$ steps). Although the formula (\ref{eq:Spitzer}) looks deceptively simple, an {\it explicit} computation of the mean $\langle L_n \rangle$ is difficult using Eq. (\ref{eq:Spitzer}), in particular its behavior for large but finite $n$. For the case of zero mean and finite variance jump distributions, the leading $n^{\frac12}$ term in $\langle L_n\rangle$ was identified in \cite{Spitzer1961}, but the relevant subleading terms were not known, to the best of our knowledge. For other results on the statistics of $L_n$, see Ref.~\cite{Snyder93}. The Spitzer-Widom formula (\ref{eq:Spitzer}) was extended to generic $d$-dimensional random walks in Ref. \cite{Vysotsky15}, in which exact combinatorial expressions for the expected surface area and the expected volume of the convex hull were derived. However, these expressions are not suitable for the asymptotic analysis at large $n$. Several other geometrical properties of the convex hull of random walks are known: for example, the exact formula for the mean number of facets of the convex polytope of a $d$-dimensional random walk, for $d=2$~\cite{Baxter61} and $d > 2$~\cite{Kabluchko16,Randon17}. But in this paper, we will restrict ourselves only to the mean perimeter $\langle L_n\rangle$ and the mean area $\langle A_n\rangle$ of a planar random walk of $n$ steps and our main goal is to derive explicitly not only the leading term in $\langle L_n\rangle$ and $\langle A_n\rangle$ for large $n$, but also the subleading terms. Our strategy is to adapt the mapping between the convex hull of a $2$-d process and the extreme statistics of the $1$-d component process, mentioned above, to the case of a single discrete-time planar random walk with generic jump distributions. Using this strategy, we are able to compute explicitly the leading and subleading terms of the mean perimeter of the convex hull of a planar random walk of $n$ steps with arbitrary symmetric continuous jump distributions for large but finite $n$. The mean area is also computed but only for the particular case of isotropic Gaussian jumps. The rest of the paper is organized as follows. Section \ref{sec:outline} outlines the class of considered planar random walks and the main results. In Sec. \ref{sec:theory}, we explain the derivation steps. In Sec. \ref{sec:simu}, several explicit examples are presented and used to illustrate the accuracy of the derived asymptotic relations by comparison with Monte Carlo simulations. In Sec. \ref{sec:discussion}, we discuss the main results, their applications, and conclusions. \ref{sec:Aderivation} and \ref{sec:Aexamples} contain some technical details of the derivation and exactly solvable examples, respectively. \section{The model and the main results} \label{sec:outline} We consider a discrete-time random walker in the plane whose jumps are random, independent, and identically distributed. Starting from the origin, the walker produces a sequence of $(n+1)$ points $\{(x_0,y_0), (x_1,y_1), \ldots, (x_n,y_n)\}\subset \mathbb R^2$ after $n$ jumps such that \begin{equation} \label{eq:RW_def} \fl (x_0,y_0) = (0,0), \qquad (x_k,y_k) = (x_{k-1}, y_{k-1}) + (\eta_k^x,\eta_k^y) \quad (k=1,2,\ldots,n), \end{equation} where the jumps $\vec \eta_k = (\eta_k^x,\eta_k^y)$ at the $k$-th step are independent from step to step, and at each step they are drawn from a prescribed joint probability density function (PDF) $p(x,y)$, i.e., \begin{equation} \mathbb P\{\eta_k^x \in (x, x+dx), \eta_k^y \in (y, y+dy)\} = p(x,y)\, dx\,dy\, . \label{noisepdf} \end{equation} We emphasize that the starting point $(x_0,y_0)$ is not random and for convenience, we choose $(x_0= y_0=0)$ to be the origin. The convex hull constructed over these $(n+1)$ points is the minimal convex polygon that encloses all these points (see Fig. \ref{fig:conhull_rw} for an illustration). We are interested in the perimeter $L_n$ and the area $A_n$ of the convex hull which are random variables given that the points are generated as successive positions of a planar random walk. We aim at computing {\it exactly} the leading and subleading terms of the mean perimeter, $\langle L_n\rangle$, and the mean area, $\langle A_n\rangle$, of the convex hull for large $n$. \begin{figure} \begin{center} \includegraphics[width=80mm]{figure1.pdf} \end{center} \caption{ Illustration of the convex hull of a $7$-stepped planar random walk. The walk starts at the origin $O$ and makes independent jumps at each step (shown by arrows). after $7$ steps, the convex hull (shown by dashed red lines) is constructed around the points of the trajectory.} \label{fig:conhull_rw} \end{figure} As explained in Sec. \ref{sec:theory}, our computation relies on two key results: (i) Cauchy's formula for the perimeter and the area of a closed convex curve, that allows one to reduce the original planar problem to the analysis of one-dimensional projections, and (ii) the Pollaczek-Spitzer formula describing the distribution of the maximum of partial sums of independent symmetric continuously distributed random variables~\cite{Pollaczek52,Spitzer56}. To use the Pollaczek-Spitzer formula, we need thus to assume that the joint probability density $p(x,y)$ is continuous and centrally symmetric: \begin{equation} \label{eq:p_symm} p(-x,-y) = p(x,y) . \end{equation} In particular, our results will not be applicable to a classical random walk on the square lattice because its distribution is not continuous. In the following, we outline the main results that will be derived in Sec. \ref{sec:theory}. The mean perimeter $\langle L_n\rangle$ is computed for a very general class of symmetric continuous jump distributions. Writing the Fourier transform of $p(x,y)$ as \begin{equation} \label{eq:hatrho_xi_general} \hat{\rho}_\theta(k) = \int\limits_{-\infty}^\infty dx \int\limits_{-\infty}^\infty dy \, p(x,y) \, e^{ik (x\cos\theta + y \sin\theta)} , \end{equation} one can characterize the behavior of the mean perimeter according to the asymptotic properties of $\hat{\rho}_\theta(k)$ as $k\to 0$. We assume a general expansion \begin{equation} \label{eq:hatrho_mu} \hat{\rho}_\theta(k) \simeq 1 - |a_\theta k|^\mu + o(|k|^\mu) \qquad (k\to 0) , \end{equation} with the scaling exponent $0 < \mu \leq 2$, and a scale $a_\theta > 0$. When $0 < \mu \leq 1$, the mean perimeter of the convex hull is infinite. We therefore focus on the case $1 < \mu \leq 2$. First, we derive an exact formula for the generating function of $\langle L_n\rangle$ which is valid for any $1 < \mu \leq 2$. Extracting the asymptotic large $n$ behavior of $\langle L_n\rangle$ from this general formula is, however, nontrivial. We distinguish the following cases. \begin{enumerate} \item When the jump variance is finite ($\mu = 2$), the mean perimeter is shown to behave as \begin{equation} \label{eq:Lmean_asympt0} \langle L_n\rangle \simeq C_0 \, n^{\frac12} + C_1 + o(1) \qquad (n \gg 1), \end{equation} with \begin{equation} \label{eq:CLn} C_0 = \frac{\sqrt{2}}{\sqrt{\pi}} \int\limits_0^{2\pi} d\theta \, \sigma_\theta , \qquad C_1 = \int\limits_0^{2\pi} d\theta \, \sigma_\theta \, \gamma_\theta , \quad \end{equation} where $\sigma_\theta$ and $\gamma_\theta$ are given by \begin{eqnarray} \label{eq:sigma_theta} \sigma_\theta^2 &=& - \lim\limits_{k\to 0} \frac{\partial^2 \hat{\rho}_\theta(k)}{\partial k^2} = \langle (\eta^x \cos \theta + \eta^y \sin\theta)^2 \rangle = \frac{a_\theta^2}{2} \,, \\ \label{eq:gamma} \gamma_\theta &=& \frac{1}{\pi \sqrt{2}} \int\limits_0^\infty \frac{dk}{k^2} \ln \left(\frac{1 - \hat{\rho}_\theta\bigl(\sqrt{2}\,k/\sigma_\theta\bigr)}{k^2}\right) . \end{eqnarray} If in addition the fourth-order moment of the jump distribution is finite, one gains the second subleading term, \begin{equation} \label{eq:Lmean_asympt} \langle L_n\rangle \simeq C_0 \, n^{\frac12} + C_1 + C_2 \, n^{-\frac12} + o(n^{-\frac12}) \qquad (n \gg 1), \end{equation} with \begin{equation} C_2 = \frac{C_0}{8} + \frac{\sqrt{2}}{24\sqrt{\pi}} \int\limits_0^{2\pi} d\theta \, \sigma_\theta \, {\mathcal K}_\theta \end{equation} and \begin{equation} \label{eq:a4} {\mathcal K}_\theta = \frac{1}{\sigma_\theta^4} \lim\limits_{k\to 0} \frac{\partial^4 \hat{\rho}_\theta(k)}{\partial k^4} = \frac{\langle (\eta^x \cos \theta + \eta^y \sin\theta)^4 \rangle}{\langle (\eta^x \cos \theta + \eta^y \sin\theta)^2 \rangle^2} \,. \end{equation} Higher-order corrections can also be derived under further moments assumptions. Note that the integral expression for the coefficient $C_0$ in front of the leading term $n^{1/2}$ first appeared in \cite{Spitzer1961}. In Sec. \ref{sec:simu}, we will show that the asymptotic formula (\ref{eq:Lmean_asympt}) is very accurate even for small $n$. \item When the jump variance is infinite (i.e., $1 < \mu < 2$), one needs to consider the subleading term in the small $k$ asymptotics of $\hat{\rho}_\theta(k)$: \begin{equation} \label{eq:hatrho_nu} \hat{\rho}_\theta(k) \simeq 1 - |a_\theta k|^\mu + b_\theta |k|^\nu + o(|k|^\nu) \qquad (k\to 0) , \end{equation} with the subleading exponent $\nu > \mu$ and a coefficient $b_\theta$. Depending on the subleading exponent $\nu$, we distinguish two cases: \subitem (1) if $\mu < \nu < \mu+1$, one has \begin{equation} \label{eq:Ln_Levy1} \langle L_n\rangle \simeq C_0 \, n^{1/\mu} + C_1 \, n^{1-(\nu-1)/\mu} + o(n^{1-(\nu-1)/\mu}) \qquad (n\gg 1), \end{equation} with \begin{eqnarray} \label{eq:C0_Levy1} C_0 &=& \frac{\mu \,\Gamma(1- 1/\mu)}{\pi} \int\limits_0^{2\pi} d\theta \, a_\theta , \\ \label{eq:C1_Levy1} C_1 &=& - \frac{\Gamma((\nu-1)/\mu)}{\pi (\mu+1-\nu)} \int\limits_0^{2\pi} d\theta \, a_\theta^{1-\nu} \, b_\theta . \end{eqnarray} Note that the coefficient $C_0$ also appears in the mean perimeter of the convex hull of continuous-time symmetric stable processes \cite{Molchanov12}. \subitem (2) if $\nu > \mu+1$, one has \begin{equation} \label{eq:Ln_Levy2} \langle L_n\rangle \simeq C_0 \, n^{1/\mu} + C_1 + o(1) \qquad (n\gg 1), \end{equation} with $C_0$ from Eq. (\ref{eq:C0_Levy1}) and \begin{equation} \label{eq:C1_Levy2} C_1 = \int\limits_0^{2\pi} d\theta \, \gamma_\theta, \end{equation} where \begin{equation} \label{eq:gamma_Levy2} \gamma_\theta = \frac{1}{\pi} \int\limits_0^{\infty} \frac{dk}{k^2} \ln \left(\frac{1- \hat{\rho}_\theta(k)}{(ak)^\mu}\right) . \end{equation} For instance, for a L\'evy symmetric alpha-stable distribution with $\hat{\rho}(k) = \exp(-|ak|^{\mu})$, one gets \cite{Comtet05} \begin{equation} \label{eq:gamma_Levystable} \gamma = a \, \frac{\zeta(1/\mu)}{(2\pi)^{1/\mu} \sin(\pi/(2\mu))} \, , \end{equation} where $\zeta(z)$ is the Riemann zeta function. \end{enumerate} The obtained results are indeed very general. In turn, our method of computation of the mean area requires two additional strong assumptions: (a) the independence of the jumps along $x$ and $y$ coordinates, i.e., $p(x,y)= p(x) p(y)$ and (b) the isotropy of the jump PDF, i.e., $p(x,y)$ should depend only on the distance $r=\sqrt{x^2+y^2}$ but not on the direction of the jump. According to Porter-Rosenzweig theorem~\cite{Porter60}, only the Gaussian jump distribution with identical variance $\sigma^2$ along $x$ and $y$ directions, i.e., $p(x,y)= \frac{1}{2\pi \sigma^2}\,\exp[-(x^2+y^2)/{2\sigma^2}]$, satisfies both properties (a) and (b). Our result for the mean area is thus only valid for this Gaussian distribution: \begin{equation} \label{eq:Amean_asympt} \sigma^{-2} \langle A_n \rangle = \frac{\pi}{2} n + \gamma \sqrt{8\pi} \, n^{\frac12} + \pi ({\mathcal K}/12 + \gamma^2) + o(1) , \end{equation} with ${\mathcal K} = 3$ and \begin{equation} \label{eq:gamma_Gauss} \gamma = \frac{\zeta(1/2)}{\sqrt{2\pi}} = - 0.58259 \ldots \end{equation} Recently, an exact formula the mean area of the convex hull of a Gaussian random walk was derived \cite{Kabluchko16b}. In the isotropic case, the formula reads \begin{equation} \label{eq:sum_Gauss} \sigma^{-2} \langle A_n \rangle = \frac12 \sum\limits_{i=1}^n \sum\limits_{j=1}^{n-i} \frac{1}{\sqrt{ij}} \,. \end{equation} While the result in Eq. (\ref{eq:sum_Gauss}) is very useful for finite $n$, deriving the large $n$ asymptotics of this double sum (including up to two subleading terms as in Eq. (\ref{eq:Amean_asympt})) seems somewhat complicated. Our method, in contrast, gives a more direct access to the asymptotics. Moreover, one can check numerically that our asymptotic formula (\ref{eq:Amean_asympt}) agrees accurately with the exact expression (\ref{eq:sum_Gauss}) even for moderately large $n$. The leading term of Eq. (\ref{eq:Amean_asympt}) was shown to be valid for a generic random walk with increments of a finite variance (see Proposition 3.3 in \cite{Wade15}). Moreover, our numerical simulations (see Sec. \ref{sec:exp_radial} and Fig. \ref{fig:exp_radial}) suggest that the obtained formula (\ref{eq:Amean_asympt}) (including the subleading terms) may be applicable for some other isotropic processes. In other words, the technical assumption about the independence of the jumps along $x$ and $y$ might be relaxed in future. This statement, which is uniquely based on numerical simulations for some jump distributions, is conjectural. In turn, the isotropy assumption is important, as illustrated by numerical simulations. The large $n$ asymptotic relations (\ref{eq:Lmean_asympt}, \ref{eq:Ln_Levy1}, \ref{eq:Ln_Levy2}, \ref{eq:Amean_asympt}) are the main results of the paper. Setting $t = n\tau$ and $D = 2\sigma^2/(4\tau)$ with a time step $\tau$, one recovers from Eqs. (\ref{eq:Lmean_asympt}, \ref{eq:Amean_asympt}) the same leading terms as in Eq. (\ref{eq:perim_BM}, \ref{eq:area_BM}) for Brownian motion (note that we write $2\sigma^2$ in $D$ because $\sigma^2$ is the variance of jumps along one direction). It is thus not surprising that the leading term in Eq. (\ref{eq:Lmean_asympt}) is universal because its derivation is valid for any planar random walk with a symmetric and continuous jump distribution and having a finite variance $\sigma^2$. Thinking of Brownian motion as a limit of random walks, the subleading terms in Eq. (\ref{eq:Lmean_asympt}) can be understood as ``finite size'' corrections. The first subleading term is valid under the same assumptions as the leading term, although the coefficient $\gamma_\theta$ depends on the jump distribution (see examples in Sec. \ref{sec:simu}). In turn, the second subleading term depends on the kurtosis ${\mathcal K}_\theta$ and thus requires an additional assumption that ${\mathcal K}_\theta$ is finite. \section{Main steps leading to the derivation of results} \label{sec:theory} \subsection{Reduction to a one-dimensional problem} \begin{figure} \begin{center} \includegraphics[width=120mm]{figure2.pdf} \end{center} \caption{ The support function $M(\theta)$ for a closed convex curve (left) and for a set of points $\{(x_0,y_0),(x_1,y_1),\ldots,(x_n,y_n)\}$ (right). $M(\theta)$ is the distance between two open circles.} \label{fig:domain} \end{figure} We start with Cauchy's formula for the perimeter $L$ and the area $A$ of an arbitrary convex domain ${\cal C}$ with a reasonably smooth boundary $\gamma_{\cal C}$~\cite{Cauchy1832,Santalo}. Let the boundary $\gamma_{\cal C}$ be parameterized as $(X(s),Y(s))$ with a curvilinear coordinate $s$ ranging from $0$ to $1$. Setting the origin of coordinates inside the domain, one defines the support function $M(\theta)$ as the distance from the origin to the closest straight line that does not cross the domain and is perpendicular to the vector from the origin in direction $\theta$ (Fig. \ref{fig:domain}). In other words, \begin{equation} M(\theta) = \max\limits_{0\leq s \leq 1} \bigl\{ X(s) \cos \theta + Y(s) \sin \theta\bigr\} . \end{equation} Cauchy showed that~\cite{Cauchy1832} \begin{eqnarray} L &=& \int\limits_0^{2\pi } d\theta \, M(\theta) , \\ A &=& \frac12\int\limits_0^{2\pi } d\theta \, \bigl(M^2(\theta) - [M'(\theta)]^2\bigr) . \end{eqnarray} For a simple derivation of this formula see Ref.~\cite{Majumdar2010a}. A straightforward calculation of $M(\theta)$ for a convex hull over a set of points may seem to be hopeless, as one would need first to construct the convex hull by identifying and ordering its vertices among the given set of points and then to compute $M(\theta)$. The key idea is that $M(\theta)$ can be found directly from the vertices of the trajectory as \cite{Spitzer1961,Randon09} \begin{equation} M(\theta) = \max\limits_{0\leq k \leq n} \bigl\{ x_k \cos\theta + y_k \sin \theta \bigr\} . \end{equation} Moreover, given that the maximum for a fixed $\theta$ is realized by a certain vertex (with index $k^*$ which discretely changes with $\theta$), one also obtains the derivative: \begin{equation} \label{eq:Mprime} M'(\theta) = - x_{k^*} \sin\theta + y_{k^*} \cos \theta . \end{equation} When the points $(x_k,y_k)$ are random, the perimeter and the area of the convex hull are random variables. We focus on the mean values $\langle L_n\rangle$ and $\langle A_n\rangle$: \begin{eqnarray} \label{eq:Lmean} \langle L_n \rangle & = & \int\limits_0^{2\pi } d\theta \, \langle M(\theta)\rangle , \\ \label{eq:Amean} \langle A_n \rangle & = & \frac12 \int\limits_0^{2\pi } d\theta \, \bigl(\langle M^2(\theta)\rangle - \langle [M'(\theta)]^2 \rangle\bigr), \end{eqnarray} i.e., the computation is reduced to the first two moments of $M(\theta)$ and to the mean $\langle [M'(\theta)]^2 \rangle$. The important observation is that, for a fixed direction $\theta$, one needs to characterize the maximum of the projection of points $(x_k,y_k)$ onto that direction \begin{equation} M(\theta) = \max\limits_{0\leq k \leq n} \{ z_k^\theta\}, \qquad z_k^\theta = x_k \cos \theta + y_k \sin \theta . \end{equation} The projection of a random walk is also a random walk. In fact, we can write according to Eq. (\ref{eq:RW_def}) \begin{equation} \label{eq:zk} z_0^\theta = 0 , \qquad z_k^\theta = z_{k-1}^\theta + \xi_k^\theta \quad (k=1,2,\ldots,n), \end{equation} with \begin{equation} \label{eq:xik} \xi_{k}^\theta = \eta_k^x \cos \theta + \eta_k^y \sin \theta. \end{equation} The probability density of $\xi_k^\theta$, $\rho_\theta(z)$, is fully determined by that of the jump $(\eta_k^x,\eta_k^y)$. In particular, its Fourier transform $\hat{\rho}_\theta(k)$ is related to $p(x,y)$ by Eq. (\ref{eq:hatrho_xi_general}). The symmetry (\ref{eq:p_symm}) implies that $\hat{\rho}_\theta(-k) = \hat{\rho}_\theta(k)$ and thus the density $\rho_\theta(z)$ is symmetric. Having discussed the general jump distributions, let us mention two particular cases that will be important later. \vskip 0.3cm \noindent (a) in the case of independent jumps along $x$ and $y$ coordinates, one has $p(x,y) = p_x(x) p_y(y)$, and thus \begin{equation} \label{eq:hatrho_xi} \hat{\rho}_\theta(k) = \hat{p}_x(k\cos\theta) \, \hat{p}_y(k\sin\theta) , \end{equation} where $\hat{p}_x$ and $\hat{p}_y$ are the Fourier transforms of $p_x(x)$ and $p_y(y)$, respectively. \vskip 0.3cm \noindent (b) For isotropic jumps, $p(x,y)$ depends only on the radial coordinate, $p(x,y)dxdy = p_r(r) dr \, d\phi/(2\pi)$, where $p_r(r)$ is the radial density (that includes the factor $r$ from the Jacobian). From Eq. (\ref{eq:hatrho_xi_general}), one gets \begin{equation} \label{eq:hatrho_xi_iso} \hat{\rho}(k) = \int\limits_0^\infty dr \, p_r(r) \, J_0(|k|r), \end{equation} in which the integration over the angular coordinate $\phi$ eliminated the dependence on $\theta$ and resulted in the Bessel function of the first kind, $J_0(|k|r)$. \subsection{Formal solution of the one-dimensional problem} The formal exact solution of the one-dimensional problem can be obtained via the Pollaczek-Spitzer formula \cite{Pollaczek52,Spitzer56}. This formula characterizes the maximum of partial sums of independent identically distributed random variables $\xi_k$ with a symmetric and continuous density $\rho(z)$: \begin{equation} M_n = \max\{ 0, \xi_1, \xi_1 + \xi_2, \ldots, \xi_1 + \xi_2 + \ldots + \xi_n\} \end{equation} (in this subsection, we temporarily drop the subscript and superscript $\theta$ from all the variables; it will be restored at the end). Considering $\xi_k$ as jumps of a random walker, $z_{k} = z_{k-1} + \xi_k$ (with $z_0 = 0$), one can also write \begin{equation} M_n = \max\{ z_0, z_1, z_2, \ldots, z_n\}. \end{equation} Pollaczek and later Spitzer showed that the cumulative distribution $Q_n(z) = \mathbb P\{ M_n \leq z\}$ of $M_n$ satisfies the following identity for $0 \leq s \leq 1$ and $\lambda \geq 0$ \begin{equation} \label{eq:PS_identity} \sum\limits_{n=0}^\infty s^n \, \langle e^{-\lambda M_n} \rangle = \sum\limits_{n=0}^\infty s^n \int\limits_0^\infty dz \, e^{-\lambda z} Q'_n(z) = \frac{\phi(s,\lambda)}{\sqrt{1-s}} \, , \end{equation} with \begin{equation} \label{eq:phi} \phi(s,\lambda) = \exp\left( - \frac{\lambda}{\pi} \int\limits_0^\infty dk \, \frac{\ln(1 - s \hat{\rho}(k))}{\lambda^2 + k^2} \right) , \end{equation} where $Q'_n(z) = dQ_n(z)/dz$ is the probability density of the maximum \cite{Pollaczek52,Spitzer56}. In principle, all the moments of $M_n$ can be obtained from the formula in Eq. (\ref{eq:PS_identity}). However, in practice, deriving explicitly the moments of $M_n$ by inverting this formula is highly nontrivial~\cite{Majumdar09}. For example, the expected maximum of a discrete-time random walk $\langle M_n\rangle$ appears in a number of different problems, from packing algorithms in computer science~\cite{Coffman98}, all the way to the survival probability of a single or multiple walkers in presence of a trap~\cite{Comtet05,MCZ2006,ZMC2007,ZMC2009,Franke2012,MMS2017}. It has also appeared in the context of the order, gap and record statistics of random walks~\cite{SM2012,SM_review,MMS2013,WMS2012,GMS2017} and $\langle M_n\rangle$ has been analyzed for large $n$ (for the leading and the next subleading term) in detail using the Pollaczek-Spitzer formula in Eq. (\ref{eq:PS_identity}). Here, in addition to calculating the first three terms in the asymptotic expansion of $\langle M_n\rangle$ for $n\gg 1$, we also calculate the large $n$ behavior of the second moment $\langle M_n^2 \rangle$, that we need for the computation of the mean area of the convex hull. In fact, the Pollaczek-Spitzer formula also determines the generating functions for all moments of $M_n$: \begin{equation} \label{eq:hm} h_m(s) = \sum\limits_{n=0}^\infty s^n \, \langle M_n^m \rangle = (-1)^m \lim\limits_{\lambda\to 0} \frac{\partial^m}{\partial \lambda^m} \frac{\phi(s,\lambda)}{\sqrt{1-s}} \, . \end{equation} By considering a general asymptotic expansion \begin{equation} \hat{\rho}(k) \simeq 1 - |ak|^\mu + o(|k|^\mu) \qquad (k\to 0) \end{equation} with an exponent $0 < \mu \leq 2$ and a scale $a > 0$, we derive in \ref{sec:Agenerating} the exact expressions \begin{equation} \label{eq:h1} h_1(s) = \frac{1}{\pi(1-s)} \int\limits_0^\infty \frac{dk}{k^2} \ln \left(\frac{1- s \hat{\rho}(k)}{1-s}\right) \qquad (0 \leq s < 1), \end{equation} which is valid for any $1<\mu \le 2$ (note that $\langle M_n\rangle = \infty$ for $0 < \mu \leq 1$), and \begin{equation} \label{eq:h2} h_2(s) = (1-s) [h_1(s)]^2 + \frac{a^2 s}{(1-s)^2} \qquad (0 \leq s < 1), \end{equation} which is valid for $\mu = 2$ (note that $\langle M_n^2\rangle = \infty$ for $0 < \mu < 2$). The exact relations (\ref{eq:h1}, \ref{eq:h2}) are new results which allow one to study the first two moments of the maximum $M_n$. In the next subsection, we will analyze the expansion of Eqs. (\ref{eq:h1}, \ref{eq:h2}) as $s\to 1$ in order to determine the asymptotic behavior of the moments $\langle M_n\rangle$ and $\langle M_n^2\rangle$ as $n\to \infty$. We consider separately jumps with a finite variance, and L\'evy flights. \subsection{Mean perimeter and mean area of the convex hull} \subsubsection{Mean perimeter for jumps with a finite variance.} Given a generic continuous jump distribution $p(x,y)$ satisfying the property in Eq. (\ref{eq:p_symm}), we determine $\hat{\rho}_\theta(k)$ using Eq. (\ref{eq:hatrho_xi_general}). Furthermore, by examining the small $k$ behavior of $\hat{\rho}_\theta(k)$, we determine the $\theta$-dependent variance $\sigma_\theta^2$ and the $\theta$-dependent kurtosis ${\mathcal K}_\theta$, using respectively Eqs. (\ref{eq:sigma_theta}) and (\ref{eq:a4}). In addition, knowing $\hat{\rho}_\theta(k)$ from Eq. (\ref{eq:hatrho_xi_general}), we also determine $\gamma_\theta$ in Eq. (\ref{eq:gamma}). Equipped with these three quantities $\sigma_\theta$, ${\mathcal K}_\theta$ and $\gamma_\theta$, we show in \ref{sec:Aderivation} that the leading large $n$ terms of the first two moments of $M_n$ are given by \begin{eqnarray} \label{eq:Mn} \frac{\langle M_n \rangle}{\sigma_\theta} &\simeq& \frac{\sqrt{2}}{\sqrt{\pi}}\, n^{\frac12} + \gamma_\theta + \frac{{\mathcal K}_\theta + 3}{12\sqrt{2\pi}}\, n^{-\frac12} + o(n^{-\frac12}), \\ \label{eq:Mn2} \frac{\langle M_n^2\rangle}{\sigma_\theta^2} &\simeq& n + \frac{\sqrt{8}\, \gamma_\theta}{\sqrt{\pi}}\, n^{\frac 12} + ({\mathcal K}_\theta/12 + \gamma_\theta^2) + o(1) \, . \end{eqnarray} For the mean perimeter of the convex hull, we will only need the first moment in Eq. (\ref{eq:Mn}). Indeed, using Eq. (\ref{eq:Lmean}), the integration of the expansion (\ref{eq:Mn}) over $\theta$ from $0$ to $2\pi$ yields the announced result (\ref{eq:Lmean_asympt}) for the mean perimeter of the convex hull. The result for the second moment in Eq. (\ref{eq:Mn2}) will be needed later to determine the mean area $\langle A_n\rangle$. \subsubsection{Mean perimeter for L\'evy flights.} \label{sec:Levy} When the variance of jumps is infinite, one gets the Taylor expansion Eq. (\ref{eq:hatrho_mu}), with the scaling exponent $0 < \mu < 2$. When $0 < \mu \leq 1$, the mean perimeter of the convex hull is infinite. Throughout this section, we focus on the case $1 < \mu < 2$, in which the first moment of jumps is finite (and zero due to the assumption of a symmetric distribution), whereas the variance is infinite. In this case, the leading behavior of the mean maximum of partial sums is universal \cite{Comtet05} \begin{equation} \label{eq:Mn_mu} \langle M_n \rangle \simeq a_\theta \frac{ \mu \Gamma(1- 1/\mu)}{\pi} \, n^{1/\mu} + o(n^{1/\mu}) \qquad (n\gg 1). \end{equation} However, the subleading term depends on finer details of the jump distribution. In order to determine the subleading term, we consider the expansion (\ref{eq:hatrho_nu}) with the subleading term $b_\theta |k|^\nu$ such that $\nu > \mu$. We distinguish two cases: $\mu < \nu < \mu+1$ and $\nu > \mu + 1$. In \ref{sec:ALevy}, we derive the following asymptotics results: \begin{equation} \label{eq:Mn_Levy1} \fl \langle M_n\rangle \simeq a_\theta \, \frac{\mu \Gamma(1- 1/\mu)}{\pi} \, n^{1/\mu} - a_\theta^{1-\nu} \, b_\theta \, \frac{\Gamma((\nu-1)/\mu)}{\pi (\mu+1-\nu)} \, n^{1-(\nu-1)/\mu} + o(n^{1-(\nu-1)/\mu}) \quad (n\gg 1) \end{equation} for $\mu < \nu < \mu+1$, and \begin{equation} \label{eq:Mn_Levy2} \langle M_n\rangle \simeq a_\theta \, \frac{\mu \Gamma(1- 1/\mu)}{\pi} \, n^{1/\mu} + \gamma_\theta + o(1) \quad (n\gg 1) \end{equation} for $\nu > \mu + 1$, with $\gamma_\theta$ given by Eq. (\ref{eq:gamma_Levy2}). The asymptotic relation (\ref{eq:Mn_Levy2}) was first derived in \cite{Comtet05} for the particular case $\nu = 2\mu$. One can see that for $\mu < \nu < \mu+1$, the subleading term of $\langle M_n\rangle$ grows with $n$, whereas for when $\nu > \mu + 1$, the subleading term is constant. Higher-order corrections can be derived as well. Finally, using the Cauchy formula (\ref{eq:Lmean}), the integration of Eqs. (\ref{eq:Mn_Levy1}, \ref{eq:Mn_Levy2}) over $\theta$ from $0$ to $2\pi$ yields Eqs. (\ref{eq:Ln_Levy1}, \ref{eq:Ln_Levy2}) for the mean perimeter of the convex hull, announced in Section \ref{sec:outline}. \subsubsection{Mean area for isotropic Gaussian jumps.} According to Eq. (\ref{eq:Amean}), the expansion (\ref{eq:Mn2}) determines the first contribution to the mean area. This contribution was calculated for an arbitrary symmetric continuous jump distribution with a finite variance. The second contribution to the mean area comes from $\langle [M'(\theta)]^2\rangle$ that has to be computed separately. We recall that $M'(\theta)$ is given by Eq. (\ref{eq:Mprime}). Our computation of this contribution relies on two additional simplifying assumptions: (a) the jumps along $x$ and $y$ coordinates are independent and (b) the jump process is isotropic, i.e., the distribution of jumps does not depend on their direction. In this case, using the isotropy condition (b), we get \begin{eqnarray} \langle [M(\theta)]^2\rangle &=& \langle [M(0)]^2\rangle = \langle x_{k^*}^2 \rangle , \\ \langle [M'(\theta)]^2\rangle &=& \langle [M'(0)]^2\rangle = \langle y_{k^*}^2 \rangle . \end{eqnarray} The disentanglement of $\langle [M(\theta)]^2\rangle$ and $\langle [M'(\theta)]^2\rangle$ allows one to compute the latter one by using the following argument. We recall that $k^*$ is the index of the maximal position among $x_k$, i.e., its statistics is fully determined by the jumps along $x$ coordinate. Once this statistics is known, the mean $\langle [M'(0)]^2\rangle$ can be found by taking the conditional expectation of $y^2_{k^*}$ at any fixed value $k^*$ and then the expectation with respect to the distribution of $k^*$. Now, once we condition $k^*$, i.e., the time step at which the $x_k$'s achieve their maximum, the $y_k$ process will, in general, be affected by this conditioning. However, if $x_k$ and $y_k$ are independent (which happens when the jump process satisfies property (a) above), we get \begin{equation} \langle [M'(0)]^2\rangle = \sigma^2 \, \langle k^* \rangle , \end{equation} where $\langle [\eta_k^y]^2\rangle = \sigma^2$. It remains to find $\langle k^* \rangle$. For symmetric and continuous jump distributions, it follows from symmetry that $\langle k^* \rangle=n/2$ independent of the details of the jump PDF. This can be deduced formally also, by noting that the time step $k^*$, at which the maximum of $x_k$ is achieved, has a universal distribution, independent of the jump distribution (given that the latter is continuous and symmetric)~\cite{Majumdar09}: \begin{equation} P_n(k^* = k) = \left(\begin{array}{c} 2k \\ k \end{array}\right) \left(\begin{array}{c} 2(n-k) \\ n-k \end{array}\right) 2^{-2n} . \end{equation} This is the direct consequence of the Sparre Andersen theorem. From this distribution, one easily computes the mean value as \begin{equation} \langle k^* \rangle = \frac{n}{2} \, , \end{equation} and thus \begin{equation} \label{eq:Mprime_average} \langle [M'(\theta)]^2\rangle = \sigma^2 \, \frac{n}{2} \,. \end{equation} This yields Eq. (\ref{eq:Amean_asympt}) for the mean area of the convex hull for an isotropic jump process with independent jumps along $x$ and $y$ coordinates. As discussed in Sec. \ref{sec:outline}, the only process satisfying two requirements (a) and (b) is the isotropic Gaussian process. Our formula (\ref{eq:Amean_asympt}) is thus provided only in this case. For a more general case, one would need to compute the joint distribution of the maximum position $k^*$ and both values $x_{k^*}$ and $y_{k^*}$ which is more complicated and remains an open problem. \section{Examples and simulations} \label{sec:simu} In this section, we illustrate the above general results on several examples of symmetric planar random walks. We derive the explicit values of the relevant parameters that determine the mean perimeter and the mean area. We also investigate the effect of anisotropy of the jump distribution on the convex hull properties. Finally, we compare our theoretical predictions to the results of Monte Carlo simulations. For this purpose, we generate $10^5$ planar random walks, compute the convex hull for each generated trajectory with $n+1$ points by using the Matlab function 'convhull', and determine its perimeter and area. These simulations yield the representative statistics of perimeters and areas from which the mean values are computed. \subsection{Gaussian jumps} We first consider the basic example of Gaussian jumps which are independent along $x$ and $y$ coordinates and characterized by variances $\sigma_x^2$ and $\sigma_y^2$. Substituting the jump probability density, \begin{equation} p(x,y) = \frac{\exp\bigl(-\frac{x^2}{2\sigma^2_x}\bigr)}{\sqrt{2\pi}\, \sigma_x} \, \frac{\exp\bigl(-\frac{y^2}{2\sigma^2_y}\bigr)}{\sqrt{2\pi} \, \sigma_y} \, , \end{equation} into Eq. (\ref{eq:hatrho_xi_general}) yields $\hat{\rho}_\theta(k) = e^{-k^2 \sigma^2_\theta/2}$, with $\sigma_\theta^2 = \sigma_x^2 \cos^2\theta + \sigma_y^2 \sin^2\theta$. One can see that the anisotropy only affects the variance $\sigma_\theta^2$, whereas the two other relevant parameters, $\gamma$ and ${\mathcal K}$, which are rescaled by variance, do not depend on $\theta$. One finds ${\mathcal K} = 3$, whereas the integral in Eq. (\ref{eq:gamma}) was computed exactly in \cite{Comtet05} and provided in Eq. (\ref{eq:gamma_Gauss}). Assuming (without loss of generality) that $\sigma_x \geq \sigma_y$, we set \begin{equation} \label{eq:sigma_Gauss} \sigma \equiv \frac{1}{2\pi} \int\limits_0^{2\pi} d\theta \, \sigma_\theta = \frac{2}{\pi} E\biggl(\sqrt{1 - (\sigma_y/\sigma_x)^2}\biggr) \, \sigma_x , \end{equation} where $E(k)$ is the complete elliptic integral of the second kind (for the isotropic case, $\sigma_x = \sigma_y = \sigma$). With this notation, we get the expansion coefficients \begin{equation} \label{eq:Cj_Gauss} \sigma^{-1} C_0 = \sqrt{8\pi} , \qquad \sigma^{-1} C_1 = 2\pi \gamma = -3.6605 \ldots , \qquad \sigma^{-1} C_2 = \frac{\sqrt{\pi}}{\sqrt{2}} \,. \end{equation} Figure \ref{fig:Gauss_iso} shows the rescaled mean perimeter, $\langle L_n\rangle/n^{1/2}$, and the rescaled mean area, $\langle A_n\rangle/n$, for isotropic planar random walk with independent Gaussian jumps ($\sigma_x = \sigma_y = 1$). The results of Monte Carlo simulations are in perfect agreement with our theoretical predictions (\ref{eq:Lmean_asympt}, \ref{eq:Amean_asympt}). One can see that the subleading terms play an important role. In fact, if one kept only the leading term and omitted the subleading terms, one would get the horizontal dotted line. This corresponds to the case of Brownian motion, in which only the leading term is present, see Eqs. (\ref{eq:perim_BM}, \ref{eq:area_BM}). The subleading terms account for the discrete-time character of random walks which is particularly important for moderate values of $n$. In order to highlight the role of the third term in the asymptotic expansions, we also draw by dashed line the asymptotics without this term. As expected, the third term improves the quality of the theoretical prediction at small $n$. Note also that the asymptotic relation is slightly less accurate for the mean area. \begin{figure} \begin{center} \includegraphics[width=75mm]{figure3a.pdf} \includegraphics[width=75mm]{figure3b.pdf} \end{center} \caption{ The rescaled mean perimeter, $\langle L_n\rangle/n^{1/2}$, (left), and the rescaled mean area, $\langle A_n\rangle/n$, (right), for isotropic planar random walks with independent Gaussian jumps, with $\sigma_x = \sigma_y = 1$. The results of Monte Carlo simulations (shown by circles) are in perfect agreement with our theoretical predictions (\ref{eq:Lmean_asympt}, \ref{eq:Amean_asympt}) (shown by solid line). The dotted horizontal line presents the coefficients $\sqrt{8\pi}$ and $\pi/2$ of the leading term, whereas the dashed line illustrates the theoretical predictions with only two principal terms. } \label{fig:Gauss_iso} \end{figure} In Fig. \ref{fig:Gauss_ani}, we consider the convex hull for anisotropic random walk with independent Gaussian jumps, with $\sigma_x = 5$ and $\sigma_y = 1$. In this case, one can use the asymptotic formula (\ref{eq:Lmean_asympt}), in which the expansion coefficients $C_j$ are given by Eq. (\ref{eq:Cj_Gauss}), with an effective variance $\sigma^2$ computed in Eq. (\ref{eq:sigma_Gauss}). In this example, $\sigma = 5 \frac{2}{\pi} E\bigl(\sqrt{1-1/25}\bigr) = 3.3439\ldots$. For the mean perimeter, one observes a perfect agreement between the theoretical predictions and Monte Carlo simulations. In turn, our asymptotic formula (\ref{eq:Amean_asympt}) for the mean area is not applicable for anisotropic case, as also confirmed by simulations (not shown). \begin{figure} \begin{center} \includegraphics[width=75mm]{figure4.pdf} \end{center} \caption{ The rescaled mean perimeter, $\langle L_n\rangle/n^{1/2}$, for anisotropic planar random walk with independent Gaussian jumps, with $\sigma_x = 5$ and $\sigma_y = 1$. The results of Monte Carlo simulations (shown by circles) are in perfect agreement with our theoretical prediction (\ref{eq:Lmean_asympt}) (shown by solid line). The dotted horizontal line presents the coefficient $\sigma \sqrt{8\pi}$ of the leading term (with $\sigma = 5 \frac{2}{\pi} E\bigl(\sqrt{1-1/25}\bigr) = 3.3439\ldots$), whereas the dashed line illustrates the theoretical predictions with only two principal terms.} \label{fig:Gauss_ani} \end{figure} \subsection{Exponentially distributed radial jumps} \label{sec:exp_radial} The next common model has exponentially distributed radial jumps with uniform angular distribution. This is a particular realization of a ``run-and-tumble'' model of bacterial motion \cite{Berg72,Berg,Lauga09}. Substituting the radial density $p_r(r) = \sigma^{-1}\, e^{-r/\sigma}$ into Eq. (\ref{eq:hatrho_xi_iso}) yields $\hat{\rho}(k) = \bigl(1 + (k\sigma)^2\bigr)^{-1/2}$. One gets ${\mathcal K} = 9$ and \begin{equation} \gamma = \frac{1}{\pi \sqrt{2}} \int\limits_0^\infty \frac{dk}{k^2} \ln \biggl(\frac{1 - \bigl(1 + 2k^2\bigr)^{-1/2}}{k^2}\biggr) = -0.8183\ldots \end{equation} from which the expansion coefficients are \begin{equation} \sigma^{-1} C_0 = \sqrt{8\pi} , \qquad \sigma^{-1} C_1 = - 5.1416\ldots , \qquad \sigma^{-1} C_2 = \sqrt{2\pi} . \end{equation} Figure \ref{fig:exp_radial} illustrates the obtained results. As for isotropic Gaussian jumps, there is a perfect agreement between theory and simulations for the mean perimeter. We also present the results for the mean area. We recall that the theoretical formula (\ref{eq:Amean_asympt}) was derived under the assumption of independent jumps along $x$ and $y$ coordinates, which evidently fails for the exponential radial jumps. In spite of this failure, the theoretical formula (\ref{eq:Amean_asympt}) is in perfect agreement with Monte Carlo simulations, except for very small $n$. This empirical observation suggests a possibility to relax this technical assumption in future, at least for large $n$. \begin{figure} \begin{center} \includegraphics[width=75mm]{figure5a.pdf} \includegraphics[width=75mm]{figure5b.pdf} \end{center} \caption{ The rescaled mean perimeter, $\langle L_n\rangle/n^{1/2}$, (left), and the rescaled mean area, $\langle A_n\rangle/n$, (right), for isotropic planar random walk with exponentially distributed radial jumps, with $\sigma = 1$. The results of Monte Carlo simulations (shown by circles) are in perfect agreement with our theoretical predictions (\ref{eq:Lmean_asympt}, \ref{eq:Amean_asympt}) (shown by solid line). The dotted horizontal line presents the coefficient $\sqrt{8\pi}$ and $\pi/2$ of the leading term, whereas the dashed line illustrates the theoretical predictions with only two principal terms.} \label{fig:exp_radial} \end{figure} \subsection{Independent exponentially distributed jumps} We also consider the case, when the jumps along $x$ and $y$ coordinates are independent and exponentially distributed, with two densities $p_x(x) = \frac12 \sigma_x^{-1} e^{-|x|/\sigma_x}$ and $p_y(y) = \frac12 \sigma_y^{-1} e^{-|y|/\sigma_y}$. The Fourier transforms are $\hat{p}_x(k) = (1 + (k\sigma_x)^2)^{-1}$ and $\hat{p}_y(k) = (1 + (k\sigma_y)^2)^{-1}$ so that \begin{equation} \hat{\rho}_\theta(k) = \frac{1}{1 + (k\sigma_x)^2 \cos^2 \theta} \, \frac{1}{1 + (k\sigma_y)^2 \sin^2 \theta} \,. \end{equation} We get $\sigma^2_\theta = 2(\sigma_x^2 \cos^2\theta + \sigma_y^2 \sin^2\theta)$ and \begin{equation} {\mathcal K}_\theta = 24 \frac{\sigma_x^4 \cos^4\theta + \sigma_x^2 \sigma_y^2 \sin^2\theta \cos^2\theta + \sigma_y^4 \sin^4\theta}{\sigma_\theta^4} \,. \end{equation} Using the identity (\ref{eq:auxil1}), we compute explicitly \begin{equation} \gamma_\theta = \frac{\sqrt{2} \sigma_x \sigma_y |\sin\theta \cos\theta| - \sigma_\theta (\sigma_x |\cos\theta| + \sigma_y |\sin\theta|)}{\sigma_\theta^2} . \end{equation} In the particular case $\sigma_x = \sigma_y = \sigma$, one gets $\sigma^2_\theta = 2\sigma^2$, ${\mathcal K}_\theta = 6(1 - \sin^2\theta \cos^2\theta)$, and \begin{equation} \gamma_\theta = \frac{|\sin\theta \cos\theta| - |\cos\theta| - |\sin\theta|}{\sqrt{2}} \,. \end{equation} For this case, we obtain from Eqs. (\ref{eq:CLn}) \begin{equation} \sigma^{-1} C_0 = 4\sqrt{\pi} , \qquad \sigma^{-1} C_1 = -6 , \qquad \sigma^{-1} C_2 = \frac{11\sqrt{\pi}}{8}\,. \end{equation} Figure \ref{fig:exp_indep} illustrates the excellent agreement between theory and simulations. \begin{figure} \begin{center} \includegraphics[width=75mm]{figure6.pdf} \end{center} \caption{ The rescaled mean perimeter, $\langle L_n\rangle/n^{1/2}$, for anisotropic planar random walk with independent exponentially distributed jumps, with $\sigma_x = \sigma_y = 1$. The results of Monte Carlo simulations (shown by circles) are in perfect agreement with our theoretical prediction (\ref{eq:Lmean_asympt}) (shown by solid line). The dotted horizontal line presents the coefficient $4\sqrt{\pi}$ of the leading term, whereas the dashed line illustrates the theoretical predictions with only two principal terms.} \label{fig:exp_indep} \end{figure} \subsection{Radial L\'evy jumps} Now, we investigate the example of radial L\'evy flights with infinite variance (but finite mean) and uniform angle distribution. Among various heavy-tailed jump distributions (e.g., Pareto distributions), we choose for our illustrative purposes the distribution \begin{equation} \label{eq:Levy_distrib} \mathbb P\{ \xi > r\} = \bigl(1 + (r/R)^2\bigr)^{-\alpha} , \end{equation} with a scale $R > 0$ and the scaling exponent $\mu = 2\alpha$, with $\frac12 < \alpha < 1$. For this distribution, Eq. (\ref{eq:hatrho_xi_iso}) yields a simple closed formula \cite{Gradshteyn} \begin{equation} \hat{\rho}(k) = \frac{2^{1-\alpha}}{\Gamma(\alpha)} (|k|R)^{\alpha} K_{\alpha}(|k|R) , \end{equation} where $K_\alpha(z)$ is the modified Bessel function of the second kind. The asymptotic behavior of $K_\alpha(z)$ near $z$ implies as $k\to 0$ \begin{equation} \hat{\rho}(k) \simeq 1 - \frac{\pi \, (|k|R)^{2\alpha}}{2^{2\alpha} \sin(\pi\alpha) \Gamma(\alpha)\Gamma(\alpha+1)} + \frac{(kR)^2}{4(1-\alpha)} + O(|k|^{2+2\alpha}) . \end{equation} Comparing this expansion to Eq. (\ref{eq:hatrho_nu}), one can identify \begin{equation} \label{eq:aR} \fl a = R \left(\frac{\pi}{2^{2\alpha} \sin(\pi\alpha) \Gamma(\alpha)\Gamma(\alpha+1)}\right)^{\frac{1}{2\alpha}} \,, \qquad b = \frac{R^2}{4(1-\alpha)} \,, \qquad \nu = 2. \end{equation} In Fig. \ref{fig:levy_radial}, the mean perimeter computed by Monte Carlo simulations for $\mu = 1.5$ and $R = 1$ is compared to the theoretical prediction (\ref{eq:Ln_Levy1}). One can see that the agreement is good but worse than for the earlier examples with a finite variance. One obvious reason is that here we have determined only two terms, whereas Eq. (\ref{eq:Lmean_asympt}) contains three terms. \begin{figure} \begin{center} \includegraphics[width=85mm]{figure7.pdf} \end{center} \caption{ The rescaled mean perimeter, $\langle L_n\rangle/n^{1/\mu}$, for isotropic planar random walk with radial L\'evy flights whose lengths are distributed according to Eq. (\ref{eq:Levy_distrib}) with $\mu = 1.5$ and $R = 1$. The results of Monte Carlo simulations (shown by circles) agree well with the theoretical prediction (\ref{eq:Ln_Levy1}) (shown by solid line). } \label{fig:levy_radial} \end{figure} \subsection{Independent L\'evy $\alpha$-stable symmetric jumps} Finally, we investigate L\'evy $\alpha$-stable symmetric jumps, with independent displacements along $x$ and $y$ coordinates given by $\hat{p}_x(k) = \hat{p}_y(k) = \exp(-|ak|^\mu)$, with $1 < \mu <2$. Using Eq. (\ref{eq:hatrho_xi}), one gets thus $\hat{\rho}_\theta(k) = \exp(-|a_\theta k|^\mu)$, with \begin{equation} a_\theta = a \bigl(|\cos \theta|^\mu + |\sin \theta|^\mu\bigr)^{1/\mu} . \end{equation} so that $\nu = 2\mu$, and $b_\theta = a_\theta^{2}/2$. The mean perimeter is determined by Eq. (\ref{eq:Ln_Levy2}), with \begin{equation} C_0 = a \, \frac{\mu \Gamma(1-1/\mu)}{\pi} \int\limits_0^{2\pi} d\theta \, \bigl(|\cos \theta|^\mu + |\sin \theta|^\mu\bigr)^{1/\mu} , \end{equation} and $\gamma$ given by Eq. (\ref{eq:gamma_Levystable}) which is independent of $\theta$. For $\mu = 3/2$, we obtain numerically $a^{-1} C_0 = 8.6275\ldots$ and $a^{-1} C_1 = -5.2151\ldots$. Figure \ref{fig:levy_stable} shows the good agreement between the theoretical prediction (\ref{eq:Ln_Levy2}) and Monte Carlo simulations. \begin{figure} \begin{center} \includegraphics[width=85mm]{figure8.pdf} \end{center} \caption{ The rescaled mean perimeter, $\langle L_n\rangle/n^{1/\mu}$, for anisotropic planar random walk with independent L\'evy $\alpha$-stable jump distribution with $\mu = 1.5$ and $a = 1$. The results of Monte Carlo simulations (shown by circles) agree well with the theoretical prediction (\ref{eq:Ln_Levy2}) (shown by solid line). } \label{fig:levy_stable} \end{figure} \section{Discussion and conclusion} \label{sec:discussion} To summarize, we have presented exact asymptotic results for the mean perimeter of the convex hull of an $n$-step discrete-time random walk in a plane, with a generic continuous jump distribution satisfying the central symmetry assumption in Eq. (\ref{eq:p_symm}). Explicit results, along with simulations confirming them have been presented for several examples of such jump distributions. For the mean area of the convex hull, we have derived exact results for isotropic Gaussian jump distributions. For jumps with a finite variance, our results provide precise estimates of the deviations from the Brownian limit and explain the discrepancies between the asymptotic Brownian limit results and observed simulations for finite but large $n$. The obtained results are particularly valuable for applications dealing with discrete-time random processes, e.g., home range estimation in ecology. Given that the tracks of animal displacements are typically recorded at discrete time steps (e.g., daily observations) and relatively short, the subleading terms play an important role. The asymptotic formulas can also be used for calibrating new estimators, based on the local convex hull, that were proposed for the analysis of intermittent processes in microbiology \cite{Lanoiselee17}. Finally, the knowledge of the mean perimeter of the convex hull can be used to estimate the scaling exponent and the scale of symmetric L\'evy flights, for which the conventional mean and variance estimators are useless. There are many interesting open problems that may possibly be addressed using the methods presented here. For example, the numerical evidence suggests a possible extension of the derived asymptotic formula for the mean area to other isotropic processes, beyond the Gaussian case. Also, it would be interesting to extend our results to the case of the convex hull of planar discrete-time random bridges (where the walker is constrained to come back to the starting point after $n$ discrete jumps). For such discrete-time bridges, there are recent exact results on the statistics of the first two maximum and the gap between them~\cite{MMS2014} which may be useful for the convex hull problem. One can also consider the problem with many independent discrete-time walkers. Finally, it would be interesting to study the statistics of the perimeter and the area for random walks with jump distributions that violate the reflection property in Eq. (\ref{eq:p_symm}), for example, for walks in presence of a drift or a potential. \section*{Acknowledgments} DG acknowledges the support under Grant No. ANR-13-JSV5-0006-01 of the French National Research Agency.

https://arxiv.org/abs/2101.03402

Summability characterizations of positive sequences

In this paper, we propose extensions for the classical Kummer test, which is a very far-reaching criterion that provides sufficient and necessary conditions for convergence and divergence of series of positive terms. Furthermore, we present and discuss some interesting consequences and examples such as extensions of the Olivier's theorem and Raabe, Bertrand and Gauss's test.

\section{Introduction} The Kummer's test is an advanced theoretical test which provides necessary and sufficient conditions that ensures convergence and divergence of series of positive terms. Below we present the statement of this result. Its proof and some additional historical background may be found in~\cite{Ludmila}, \cite{Knopp}, \cite{Tong:2004}. \begin{theorem} (Kummer's test)\label{Kummer} Consider the series \(\sum a_{n}\) where \(\{a_{n}\}\) is a~sequence of positive real numbers. \begin{enumerate} \item[(i)] The series \(\sum a_{n}\) converges if and only if there exist a~sequence \(\{q_{n}\}\), a~real number \(c>0\) and an integer \(N\geq 1\) for which \[ q_{n}\frac{a_{n}}{a_{n+1}}-q_{n+1}\geq c, \qquad n \geq N. \] \item[(ii)] The series \(\sum a_{n}\) diverges if, and only if there exist a~sequence \(\{q_{n}\}\) and an integer \(N\geq 1\) for which \(\sum \frac{1}{q_{n}}\) is a~divergent series and \[ q_{n}\frac{a_{n}}{a_{n+1}}-q_{n+1}\leq 0, \qquad n \geq N. \] \end{enumerate} \end{theorem} Besides providing an extremely far-reaching characterization of convergence and divergence of series with positive terms, the importance of Kummer's test it is mostly ratified by its implications. For instance, Bertrand's test, Gauss's test, Raabe's test~\cite{Tong:2004} are all special cases of Theorem~\ref{Kummer}. Kummer's test may be also usefull to characterize convergence in normed vector spaces~\cite[p. 7]{Muscat} and applications of this test can be found in other branches of Analysis, such as difference equations~\cite{Gyori}, as well. On the other hand, turning our focus to series of the form \(\sum c_n a_n\), there are only few results dealing with this type of series. The Abel's test and test of Dedekind and Du-Bois Reymond (see for instance,~\cite[p. 315]{Knopp},~\cite{Hadamard}) are probably the most famous, since they deal with general series of complex numbers. These tests provide conditions that ensure convergence by means of independent assumptions on \(\{c_{n}\}\) and \(\{a_{n}\}\). In this context, the main feature of our results (Theorem~\ref{thm1} and Theorem~\ref{thm2}) is that they characterize the relation between the sequences \(\{a_{n}\}\) and \(\{c_{n}\}\) in order to ensure necessary and sufficient conditions for the convergence and divergence of the series \(\sum c_n a_n\), respectively. Moreover, we present some examples and interesting consequences of this characterization. In particular, generalized versions of Raabe's, Bertrand's and Gauss's test for convergence and divergence of series of the form \(\sum c_n a_n\) are obtained. Another important consequence of Theorem~\ref{thm1} is that it is possible to show that Olivier's theorem (see, for instance~\cite[p. 124]{Knopp} or~\cite{Const}, \cite{Salat} for more information) still holds when the monotonicity assumption on the sequence of positive terms \(\{a_n\}\) is replaced by an additional assumption on a~auxiliary sequence. We also present consequences of Theorem~\ref{thm1} when it is combined to the well-know Abel summation formula and the Cauchy condensation theorem. We refer to~\cite[pp. 120 and 313]{Knopp} for more details on these results. The rest of the paper is organized as follows. In Section~\ref{subsum}, we present necessary and sufficient conditions for convergence/divergence of series generated by subsequences by extending Theorem~\ref{Kummer}. In Section~\ref{cnan} we present the results dealing with convergence and divergence of series of the form \(\sum c_n a_n\). The main ideia is to obtain necessary and sufficient conditions by means of an extension of Theorems~\ref{conv} and Theorem~\ref{div}. As we show, we characterize the relation between the sequences \(\{c_n\}\) and \(\{a_n\}\) that ensures convergence and divergence of the series. In Section~\ref{EC} we present some consequences of the results obtained. \section{An extension of Kummer's test: I}\label{subsum} In this section we present a~first extension of Theorem~\ref{Kummer}. Its main feature is that it showns that is possible to obtain information about the summability of a~sequence of positive real numbers based on the relation between non-consecutive elements of this sequence. In partiular, the idea is to characterize the summability of a~sequence by comparing it to the elements of the translated sequence \(\{a_{n+m}, \ n\geq 1\}\), for some~\(m \ge 1\). The first main result of this section is presented below. \begin{theorem}\label{conv} Let \(\{a_n\}\) be a~sequence of positive real numbers and \(m\geq 1\) any fixed positive integer. If there exists a~positive sequence \(\{q_n\}\) such that \[ q_n\frac{a_n}{a_{n+m}}-q_{n+m}\geq c, \] for some \(c>0\), for all \(n\) sufficiently large, then \(\sum a_{n}\) converges. The converse holds as well. \end{theorem} \begin{proof} From the assumption we get that \[ q_n a_n-a_{n+m}q_{n+m}\geq ca_{n+m}, \] for all \(n>N\), for some \(N\) large. Hence \[ \sum_{n = N+1}^{N+k} q_n a_n-a_{n+m}q_{n+m}\geq c \sum_{n = N+1}^{N+k} a_{n+m}, \] for all \(k\geq 1\). That is, by the telescopic sum and considering without loss of generality \(k > m\), we have \begin{align*} q_{N+1}a_{N+1} + \dots + q_{N+m}a_{N+m} - a_{N+k+1}q_{N+k+1} - \dots -{ }&{ }a_{N+k+m}q_{N+k+m} \\ &{ }\geq c \sum_{n = N+1}^{N+k} a_{n+m}, \end{align*} for all \(k>m\). Since \(\{a_n\}\) and \(\{q_n\}\) are positive, the left side of previous inequality is less than \(q_{N+1} a_{N+1}+ \dots + q_{N+m}a_{N+m}\) and then the series \(\sum a_{n+m}\) converges. Therefore, \(\sum a_{n}\) also converges. Conversely, if \(\sum a_n\) converges, \(\sum a_n = S\) say, then let us write \(\sum a_{n+m-1} = S_m\), for \(m\geq 1\), positive integer. Let us define \(\{q_n\}\) as \[ q_n = \frac{S_{m}-\sum_{i = 1}^{n}a_{i+m-1}}{a_{n}}, \qquad n = 1, 2, 3, \dots , \] thus, for this \(\{q_n\}\) we have that \begin{align*} q_n \frac{a_n}{a_{n+m}}-q_{n+m} { }&{ }= \frac{\sum_{i = n+1}^{n+m}a_{i+m-1}}{a_{n+m}}\\ { }&{ }= 1+\frac{a_{n+m+1}+ \dots +a_{n+2m-1}}{a_{n+m}}\\ { }&{ }> 1, \end{align*} for all \(n\geq 1\). The proof is concluded. \end{proof} We proceed by presenting a~divergence version for the previous theorem. \begin{theorem}\label{div} Let \(\{a_n\}\) be a~sequence of positive real and \(m\geq 1\) a~fixed positive integer. If there exists a~positive sequence \(\{q_n\}\) such that \(\sum \frac{1}{q_{n}}\) diverges, \(q_n a_n\geq c>0\), and \[ q_n\frac{a_n}{a_{n+m}}-q_{n+m}\leq 0, \] for all \(n\) sufficiently large, then \(\sum_{n = 1}^{\infty} a_{n}\) diverges. The converse holds, as well. \end{theorem} \begin{proof} From the assumptions we obtain that there exists \(N>0\) such that \[ q_n\frac{a_n}{a_{n+m}}-q_{n+m}\leq 0, \] for all \(n\geq N\). As so, \[ c\frac{1}{q_{n+m}}\leq a_{n+m}, \] for all \(n>N\). Since \(\sum 1/q_n\) diverges, we obtain from the comparsion test that \(\sum a_n\) diverges. Conversely, suppose that \(\sum a_{n}\) diverges. Define for each \(n\geq 1\) \[ q_n = \frac{\sum_{i = 1}^{n}a_i}{a_n}. \] Note that the definition implies \(q_1 = 1\), hence \(a_n q_n = \sum_{i = 1}^{n}a_{i}\geq a_1\), for all \(n\geq 1\), that is, \(a_n q_n\geq a_1 q_1>0\) for all \(n\geq 1\). Clearly \[ q_{n}\frac{a_{n}}{a_{n+m}}-q_{n+m}\leq 0, \] for all \(n\geq 1\). Let us now show that \(\sum \frac{1}{q_{n}}\) diverges. From the divergence of \(\sum a_{n}\), given any positive integer \(k\) there exists a~positive integer \(n\geq k\) such that \begin{equation}\label{aux11} a_{k}+\dots+a_{n}\geq a_1+\dots+a_{k-1}. \end{equation} Due to~\eqref{aux11}, \begin{align*} \sum_{j = k}^{n}\frac{1}{q_{j}} { }&{ }= \frac{a_{k}}{a_{1}+ \dots +a_{k}}+\dots+\frac{a_{n}}{a_{1}+ \dots +a_{n}}\\ { }&{ }\geq \frac{a_{k}}{a_{1}+ \dots +a_{n}}+\dots+\frac{a_{n}}{a_{1}+ \dots +a_{n}}\\ { }&{ }= \frac{1}{\frac{a_{1}+ \dots +a_{k-1}}{a_{k}+ \dots +a_{n}} +1}\\ { }&{ }> \frac{1}{2}. \end{align*} Hence, \(\sum_{j = 1}^{n}\frac{1}{q_{j}}\) is not a~Cauchy sequence. Therefore the series \(\sum \frac{1}{q_{n}}\) diverges. \end{proof} \section{Extension of Kummer's test: II}\label{cnan} Let us now turn our atention to series of the form \(\sum c_{n}a_{n}\) with positive terms. The central idea in the following result is that it characterizes the relation between the sequences \(\{c_{n}\}\) and \(\{a_{n}\}\) in order to ensure the convergence of the series. The reader will note that the proof follows the same lines as the proof of Theorem~\ref{conv} and also, that it could be obtained by some changes in the proof of Theorem~\ref{Kummer}, nevertherless, as the reader will also note, our proof provides important informations about the relation between the sequences \(\{a_n\}\) and \(\{c_n\}\). \begin{theorem}\label{thm1} Consider the series \(\sum c_{n}a_{n}\) with \(\{a_{n}\}\) \(\{c_{n}\}\) sequences of positive real numbers. The series \(\sum c_{n}a_{n}\) converges if and only if that there exist a~sequence \(\{q_{n}\}\) of positive real numbers and a~positive integer \(N\geq 1\) for which \[ q_{n}\frac{a_{n}}{a_{n+1}}-q_{n+1}\geq c_{n+1}, \quad n\geq N. \] \end{theorem} \begin{proof} Let us show that \(\sum c_{n}a_{n}\) converges. For this, note that the condition \[ q_{n}\frac{a_{n}}{a_{n+1}}-q_{n+1}\geq c_{n+1}, \quad n\geq N \] implies that \begin{equation}\label{auxx} a_{n}q_{n}\geq a_{n+1}(q_{n+1}+ c_{n+1}), \quad n\geq N. \end{equation} That is, \begin{align*} a_{N}q_{N} { }&{ }\geq a_{N+1}(q_{N+1}+c_{N+1})\\ { }&{ }\geq a_{N+2}(q_{N+2}+c_{N+2})+a_{N+1}c_{N+1}\\ { }&{\ \,}\vdots\\ { }&{ }\geq a_{N+k}q_{N+k} + \sum_{i = 1}^{k}c_{N+i}a_{N+i}\\ { }&{ }\geq \sum_{i = 1}^{k}c_{N+i}a_{N+i}>0, \end{align*} for all integer \(k\geq 0\). This implies the convergence of \(\sum c_{n}a_{n}\). For the converse, suppose that \(S:= \sum c_{n}a_{n}\) and let us define \begin{equation}\label{pn} q_{n} = \,\frac{S-\sum_{i = 1}^{n}c_{i}a_{i}}{a_{n}}, \quad n\geq N. \end{equation} For this \(\{q_{n}\}\), clearly \(q_{n}>0\) for all \(n\geq 1\) and it is easy to check that \[ q_{n}\frac{a_{n}}{a_{n+1}}-q_{n+1} = c_{n+1}, \quad n\geq N. \qedhere \] \end{proof} Some remarks: \begin{enumerate} \item[(i)] One can observe that it is, of course, possible to reduce any series to this form, as any number can be expressed as the product of two other numbers. Success in applying the above theorem will depend on the skill with which the terms are so split up. \item[(ii)] Note that in the first part of Theorem~\ref{thm1}, the assumption of positivity of the sequences \(\{a_{n}\}\) and \(\{c_{n}\}\) can be replaced by the following assumptions: \(\{a_{n}\}\) is positive and \(\{c_n\}\) is such that \(\sum_{i = 1}^{k}c_{i}a_{i}>0 \) for all \(k\) sufficiently large. \end{enumerate} Next, we presente a~version of Kummer's test for divergent series of the form \(\sum c_{n}a_{n}\). The reader will note that it is more restrictive when it is compared to Theorem~\ref{Kummer}-\((ii)\) however it may be suitable in some cases. \begin{theorem}\label{thm2} Consider the series \(\sum c_{n}a_{n}\) with \(\{a_{n}\}\) \(\{c_{n}\}\) sequences of positive real numbers. \begin{enumerate} \item[(i)] Suppose that there exist a~sequence \(\{q_{n}\}\) and a~positive integer \(N\) for which \[ q_{n}\frac{a_{n}}{a_{n+1}}-q_{n+1}\leq -c_{n+1}, \quad n\geq N \] with \(\sum \frac{1}{q_{n}}\) being a~divergent series. Then \(\sum a_{n}\), \(\sum \frac{1}{c_{n}}\), \(\sum (q_{n}-c_{n})a_{n}\) and \(\sum q_{n}a_{n}\) diverge. If, in addition, \(\sum \frac{c_{n}}{q_{n}}\) diverges then \(\sum c_{n}a_{n}\) diverges. \item[(ii)] Suppose that both series \(\sum c_{n}a_{n}\) and \(\sum a_{n}\) diverge. Also, suppose that for every \(m\in\mathbb{N}\) there exists \(r\geq m\), \(r\in\mathbb{N}\), such that \[ a_{m}+\dots+a_{r}\geq c_{m}a_{m}+\dots+c_{r}a_{r}. \] Then there exist a~sequence \(\{q_{n}\}\) and a~positive integer \(N\geq 1\) such that \(\sum\frac{1}{q_{n}}\) diverges and \[ q_{n}\frac{a_{n}}{a_{n+1}}-q_{n+1}\leq -c_{n+1}, \quad n\geq N. \] \end{enumerate} \end{theorem} \begin{proof} To prove \((i)\) note that \(\{q_{n}\}\) satisfies \begin{gather} a_{n+1}\geq \frac{q_{n}a_{n}}{q_{n+1}-c_{n+1}}, \quad n\geq N,\label{aux111} \\ 0<q_{n+1}-c_{n+1}<q_{n+1},\quad n\geq N \label{aux1} \\ \textup{and} \nonumber \\ 0<c_{n+1}<q_{n+1}, \quad n\geq N. \nonumber \end{gather} By last inequality and comparsion test we see that \(\sum \frac{1}{c_{n}}\) diverges. Next, using~\eqref{aux111} successively we see that \begin{gather*} a_{N+1}\geq \frac{q_{N}a_{N}}{q_{N+1}-c_{N+1}}, \\ a_{N+2}\geq \frac{q_{N+1}a_{N+1}}{q_{N+2}-c_{N+2}}\geq \frac{a_{N}q_{N}q_{N+1}}{(q_{N+2}-c_{N+2})(q_{N+1}-c_{N+1})}, \end{gather*} and in general, \begin{equation}\label{aux} a_{N+k+1} \geq \frac{a_{N}q_{N}q_{N+1} \dots q_{N+k}}{(q_{N+1}-c_{N+1}) \dots (q_{N+k+1}-c_{N+k+1})}, \quad k\geq 0. \end{equation} From~\eqref{aux1} and~\eqref{aux} we get \begin{equation}\label{auxDiverg} a_{N+k+1}> \frac{a_{N}q_N}{q_{N+k+1}}, \quad k\geq 0. \end{equation} Thus \[ \sum_{k = 0}^{\infty}a_{N+k+1}> a_{N}q_N\sum_{k = 0}^{\infty}\frac{1}{q_{N+k+1}} \] and therefore \(\sum a_{n}\) diverges. From~\eqref{aux} \[ (q_{N+k+1}-c_{N+k+1})a_{N+k+1}\geq \frac{a_{N}q_{N}q_{N+1} \dots q_{N+k}}{(q_{N+1}-c_{N+1}) \dots (q_{N+k}-c_{N+k})}, \quad k\geq 0, \] and applying once again~\eqref{aux1} we obtain that \[ q_{N+k+1}a_{N+k+1}>(q_{N+k+1}-c_{N+k+1})a_{N+k+1}\geq a_{N}q_{N}>0, \quad k\geq 0. \] This last set of inequalities implies that \[ \lim_{n\to\infty} q_{N+k+1}a_{N+k+1}\neq 0 \quad \textup{ and } \quad \lim_{k\to\infty} (q_{N+k+1}-c_{N+k+1})a_{N+k+1}\neq 0, \] so both series \(\sum q_{n}a_{n}\) and \(\sum (q_{n}-c_{n})a_{n}\) diverge. Note that from~\eqref{auxDiverg} we obtain that \begin{equation}\label{auxDiverg1} c_{N+k+1}a_{N+k+1}> a_{N}q_N\frac{c_{N+k+1}}{q_{N+k+1}}, \quad k\geq 0. \end{equation} Therefore, if \(\sum \frac{c_{n}}{q_{n}} \) diverges, then it is clear that \(\sum c_{n}a_{n}\) diverges. In order to prove \((ii)\) define \[ q_{n} = \frac{\sum_{i = 1}^{n}c_{i}a_{i}}{a_{n}}, \quad n\geq 1. \] Clearly, this is a~sequence of positive real numbers that satisfies \[ q_{n}\frac{a_{n}}{a_{n+1}}-q_{n+1}\leq -c_{n+1}, \quad n\geq 1. \] Let us show that \(\sum \frac{1}{q_{n}}\) diverges by concluding that the sequence \(\{s_{k}\}\), defined as \(s_{k} = \sum_{i = 1}^{k}\frac{1}{q_{i}}\), for each \(k\geq 1\), is not a~Cauchy sequence. Since \(\sum c_{n}a_{n}\) is divergent, given \(m\in \mathbb{N}\) there exists \(k > m\), \(k\in\mathbb{N}\), such that \begin{equation}\label{aux2} c_{m}a_{m}+\dots+c_{k}a_{k}>c_{1}a_{1}+\dots+c_{m-1}a_{m-1}. \end{equation} Also, from the hypothesis, there exists \(r\geq m\) such that \begin{equation}\label{aux3} a_{m}+\dots+a_{r}\geq c_{m}a_{m}+\dots+c_{r}a_{r}. \end{equation} Next, we split the proof in two cases: \(k\leq r\) and \(k>r\). If \(k\leq r\), from~\eqref{aux2} we see that \begin{equation}\label{aux112} c_{m}a_{m}+\dots+c_{k}a_{k}+\dots+c_{r}a_{r}\geq c_{m}a_{m}+\dots+c_{k}a_{k}>c_{1}a_{1}+\dots+c_{m-1}a_{m-1}. \end{equation} Thus, by~\eqref{aux112} and~\eqref{aux3} \begin{align*} \sum_{n = m}^{r}\frac{1}{q_{n}} { }&{ }= \frac{a_{m}}{c_{1}a_{1}+ \dots +c_{m}a_{m}}+\dots+\frac{a_{r}}{c_{1}a_{1}+ \dots +c_{r}a_{r}}\\ { }&{ }\geq \frac{a_{m}+ \dots +a_{r}}{c_{1}a_{1}+ \dots +c_{r}a_{r}}\\ { }&{ }\geq \frac{c_{m}a_{m}+ \dots +c_{r}a_{r}}{c_{1}a_{1}+ \dots +c_{r}a_{r}}\\ { }&{ }= \frac{1}{\frac{c_{1}a_{1}+ \dots +c_{m-1}a_{m-1}}{c_{m}a_{m}+ \dots +c_{r}a_{r}}+ 1}\\ { }&{ }>\frac{1}{2} \end{align*} and \(\{s_k\}\) is not a~Cauchy sequence. On the other hand, if \(k>r\) we can use hypothesis again (now applied to \(m_1 = r+1\)) and to obtain \(r_{1}\geq r+1\) such that \[ a_{r+1}+\dots+a_{r_{1}}\geq c_{r+1}a_{r+1}+\dots+c_{r_{1}}a_{r_1}. \] Again, we can use the same argument to conclude that there exists \(r_{2}\geq r_{1}+1\) such that \[ a_{r_{1}+1}+\dots+a_{r_{2}}\geq c_{r_{1}+1}a_{r_{1}+1}+\dots+c_{r_{2}}a_{r_{2}}. \] This procedure can be applied a~finite number of times in order to obtain \(r_{j} \geq k\) for which \[ a_{r_{(j-1)}+1}+\dots+a_{r_{j}}\geq c_{r_{(j-1)}+1}a_{r_{(j-1)}+1}+\dots+c_{r_{j}}a_{r_{j}}. \] Summing up~\eqref{aux3} with all these previous inequalities we obtain that \[ a_{m}+\dots+a_{r_{j}}\geq c_{m}a_{m}+\dots+c_{r_{j}}a_{r_{j}} \] with \(k\leq r_j\). This reduces the proof to the previous case which we have already proved. \end{proof} \section{Some examples and consequences}\label{EC} The main goal in this section is to present some of the implications of the main results of this paper. The next three theorems are extensions of the Raabe, Bertrand and Gauss test derived from Theorem~\ref{thm1} and Theorem~\ref{thm2}. For more information about these tests we refer to~\cite{Ludmila}, \cite{Knopp} and references therein. Consider the sequences \[ R_n^{-} = n \frac{a_n}{a_{n+1}}-(n+1)- c_{n+1} \quad \textrm{and} \quad \,R_n^{+} = n \frac{a_n}{a_{n+1}}-(n+1)+ c_{n+1}, \] for all positive integer \(n\). \begin{theorem}[Raabe's test] Let \(\sum c_n a_n\) be a~series of positive terms and suppose that \(\liminf R_n^{-} = R_1\) and \(\limsup R_n^{-} = R_2\). If \begin{enumerate} \item[(i)] \(R_1> 0\), then \(\sum c_n a_n\) converges; \item[(ii)] \(R_2< 0\) and \(\sum c_n/n \) diverges, then \(\sum c_n a_n\) diverges. \end{enumerate} \end{theorem} \begin{proof} \begin{enumerate} \item[(i)] If \(R_1> 0\), then for all \(n\) sufficiently large we have that \[ n\frac{a_n}{a_{n+1}}-(n+1)- c_{n+1}\geq 0, \] hence Theorem~\ref{thm1}, with \(q_n = n\) for all \(n\geq 1\), implies that the series \(\sum c_n a_n\) converges. \item[(ii)] If \(R_2< 0\), then for all \(n\) sufficiently large \[ n\frac{a_n}{a_{n+1}}-(n+1)+ c_{n+1}\leq 0. \] Again, we have \(q_{n} = n\) for all \(n\geq1\). So, due to the divergence of \(\sum c_n/n \), Theorem~\ref{thm2} implies that \(\sum c_n a_n\) diverges. \qedhere \end{enumerate} \end{proof} \begin{theorem}[Bertrand's test] Let \(\sum c_n a_n\) be a~series of positive terms. \begin{enumerate} \item[(i)] If \[ \frac{a_n}{a_{n+1}}> 1+\frac{1}{n}+\frac{\theta_n +c_{n+1}}{n\ln(n)}, \] for some sequence \(\{\theta_n\}\), such that \(\theta_{n}\geq \theta>1\), for all \(n\geq 1\), then \(\sum c_n a_n\) converges. \item[(ii)] If \[ \frac{a_n}{a_{n+1}}\leq 1+\frac{1}{n}+\frac{\theta_n -c_{n+1}}{n\ln(n)}, \] for some sequence \(\{\theta_n\}\), such that \(\theta_{n}\leq \theta<1\), for all \(n\geq 1\), and \(\sum \frac{c_n}{n\ln(n)} \) diverges, then \(\sum c_n a_n\) diverges. \end{enumerate} \end{theorem} \begin{proof} \begin{enumerate} \item[(i)] From the assumption we get \[ n\ln(n)\frac{a_n}{a_{n+1}}\geq n\ln(n)+\ln(n)+ c_{n+1}+\theta_n, \] for all \(n\) sufficiently large. That is, \[ n\ln(n)\frac{a_n}{a_{n+1}}-(n+1)\ln(n+1)\geq (n+1)\ln\left(\frac{n}{n+1}\right)+\theta_n+ c_{n+1}, \] for all \(n\) sufficiently large. It follows from the assumption on \(\{\theta_n\}\) that \[ (n+1)\ln\left(\frac{n}{n+1}\right)+\theta_n> 0, \] for all \(n>1\) sufficiently large hence we conclude that \[ n\ln(n)\frac{a_n}{a_{n+1}}-(n+1)\ln(n+1)> c_{n+1}, \] for all \(n\) sufficiently large. Therefore, the convergence of \(\sum c_{n}a_{n}\) follows from an application of Theorem~\ref{thm1}. \item[(ii)] It suffices to note that \[ n\ln(n)\frac{a_n}{a_{n+1}}-(n+1)\ln(n+1)\leq (n+1)\ln\left(\frac{n}{n+1}\right)+\theta_n- c_{n+1}, \] for all \(n\) sufficiently large. Since \((n+1)\ln\left(\frac{n}{n+1}\right)+\theta_n< 0\) for all \(n>1\) sufficiently large we obtain \[ n\ln(n)\frac{a_n}{a_{n+1}}-(n+1)\ln(n+1)< - c_{n+1}, \] for all \(n\) sufficiently large. The conclusion follows from Theorem~\ref{thm2}. \qedhere \end{enumerate} \end{proof} \begin{theorem}[Gauss's test]\label{gauss} Let \(\sum c_n a_n\) be a~series of positive terms, \(\gamma\geq 1\) and \(\{\theta_n\}\) a~bounded sequence of real numbers. \begin{enumerate} \item[(i)] Suppose that there exists a~\(\mu \in\mathbb{R}\) such that \(\theta_{n}\geq (1-\mu)n^{\gamma-1}\) holds for all \(n\) sufficiently large. If \[ \frac{a_n}{a_{n+1}}\geq 1+ \frac{c_{n+1}}{n}+\frac{\mu}{n}+\frac{\theta_n}{n^{\gamma}}, \] holds for all \(n\) sufficiently large, then \(\sum c_n a_n\) converges. \item[(ii)] Suppose that there exists a~\(\mu \in\mathbb{R}\) such that \(\theta_{n}\leq (1-\mu)n^{\gamma-1}\) holds for all \(n\) sufficiently large. If \(\sum c_n/n \) diverges and \[ \frac{a_n}{a_{n+1}}\leq 1- \frac{c_{n+1}}{n}+\frac{\mu}{n}+\frac{\theta_n}{n^{\gamma}}, \] for all \(n\) sufficiently large, then \(\sum c_n a_n\) diverges. \end{enumerate} \end{theorem} \begin{proof} \begin{enumerate} \item[(i)] From the assumption we obtain that \[ n\frac{a_n}{a_{n+1}}-(n+1)\geq c_{n+1}+(\mu-1)+\frac{\theta_n}{n^{\gamma-1}}, \] for all \(n\) sufficiently large. Taking \(N>0\) such that \(\mu-1+\frac{\theta_n}{n^{\gamma-1}}\geq 0\), for all \(n>N\), we concude that \[ n\frac{a_n}{a_{n+1}}-(n+1)\geq c_{n+1}, \] for all \(n>N\). Therefore, by Theorem~\ref{thm1}, the series \(\sum c_n a_n\) converges. \item[(ii)] Due to the assumptions on \((ii)\), we have that \(\mu-1 +\frac{\theta_n}{n^{\gamma-1}}<0\) and \[ n\frac{a_n}{a_{n+1}}-(n+1)\leq -c_{n+1}+(\mu-1)+\frac{\theta_n}{n^{\gamma-1}}\leq -c_{n+1}, \] for all \(n\) sufficiently large. The conclusion follows from an application of Theorem~\ref{thm2}. \qedhere \end{enumerate} \end{proof} Theorem~\ref{thm1} also allows us to provide a~different approach for the well-know Cauchy's condensation test, which we present in the next lemma. \begin{lemma}\label{CC}\cite[p. 120]{Knopp}(Cauchy's condensation test) Let \(\{a_n\}\) be a~decreasing sequence of positive numbers. Then \(\sum a_{n}\) converges if, and only if, \(\sum 2^{n}a_{2^{n}}\) converges. \end{lemma} For a~decreasing sequence \(\{a_{n}\}\) of positive real numbers, combining Lemma~\ref{CC} with Theorem~\ref{thm1}, we obtain a~the following characterization of convergence. \begin{theorem} Let \(\sum a_{n}\) be a~series with \(\{a_n\}\) being a~decreasing sequence. Then \(\sum a_{n}\) converges if, and only if, there exists a~sequence \(\{q_{n}\}\) of positive numbers such that \[ q_{n}-2 q_{n+1}\geq 2 a_{2^{n+1}}, \] for all \(n\) sufficiently large. \end{theorem} \begin{proof} By Lemma~\ref{CC}, \(\sum a_n\) coverges if, and only if, \(\sum 2^n a_{2^n}\) converges. On the other hand, an application of Theorem~\ref{thm1} with \(a_{n} = 2^n\) and \(c_n = a_{2^{n}}\) show us that \(\sum 2^n a_{2^n}\) converges if, and only if, there exists a~sequence \(\{q_n\}\) of positive real numbers such that \[ q_n-2 q_{n+1}\geq 2 a_{2^{n+1}}, \] for all \(n\) sufficiently large. The proof is concluded. \end{proof} To close this section of applications we present a~result related to the Olivier's Theorem, which is stated below. \begin{lemma}[{\cite[p. 124]{Knopp} or \cite{Const}}] \label{Olivier} Let \(\{a_{n}\}\) be summable decreasing sequence of positive real numbers. Then \(\lim n\,a_{n} = 0\). \end{lemma} We are going to show that it possible to recover the same Olivier's asymptotic behavior for \(\{a_n\}\) without the decreasigness assumption on \(\{a_n\}\). Instead of using the monotonicity, we consider an additional assumption on the sequence \(\{q_{n}\}\) (that auxiliary sequence of Theorem~\ref{thm1}). \begin{theorem} Suppose that \(\{a_{n}\}\) is a~sequence of positive numbers. We have that \(\sum a_n\) converges if, and only if, there exists a~sequence \(\{q_n\}\) of positive numbers such that \[ q_{n}\frac{n+1}{n}-q_{n+1}\geq (n+1)a_{n+1}, \] for all \(n\) sufficiently large. Moreover, if \(\{q_n\}\) satisfies \[ \lim q_{n}\frac{n+1}{n}-q_{n+1} = 0, \] then \(\lim n a_{n} = 0\). \end{theorem} \begin{proof} It is clear that \(\sum a_{n}\) converges if and only if \(\sum \frac{1}{n} na_{n}\) also converges. From Theorem~\ref{thm1}, with \(a_n = 1/n\) and \(c_n = n a_n\), we can conclude that \(\sum a_n\) converges if, and only if, there exists a sequence \(\{q_{n}\}\) such that \[ q_{n}\frac{n+1}{n}-q_{n+1}\geq (n+1)a_{n+1}, \] for all \(n\) sufficiently large. Hence, \(\lim n a_n = 0\) certainly occurs when the sequence \(\{q_n\}\) above is such that \[ \lim q_{n}\frac{n+1}{n}-q_{n+1} = 0. \qedhere \] \end{proof} For more information on this asymptotic behavior of summable sequences of positive numbers we refer to~\cite{Lifly}, \cite{Const}, \cite{Salat} and references therein.

https://arxiv.org/abs/2005.05500

Binary polynomial power sums vanishing at roots of unity

Let $c_1(x),c_2(x),f_1(x),f_2(x)$ be polynomials with rational coefficients. With obvious exceptions, there can be at most finitely many roots of unity among the zeros of the polynomials $c_1(x)f_1(x)^n+c_2(x)f_2(x)^n$ with $n=1,2\ldots$. We estimate the orders of these roots of unity in terms of the degrees and the heights of the polynomials $c_i$ and $f_i$.

\section{Introduction} \label{sintr} Let $c_1(x),c_2(x),f_1(x),f_2(x)$ be non-zero polynomials in ${\mathbb Q}[x]$. We denote by ${\bf u}:=\{u_n(x)\}_{n\ge 1}\subset {\mathbb Q}[x]$ the sequence of polynomials given by \begin{equation} \label{eq:1} u_n(x)=c_1(x)f_1(x)^n+c_2(x)f_2(x)^n\quad {\text{\rm for~all}}\quad n\ge 1. \end{equation} We study roots of unity $\zeta$ such that ${u_n(\zeta)=0}$ for some~$n$. It can happen accidentally that $u_n(x)$ is the zero polynomial for some~$n$. We ignore these~$n$. We would like to show that aside from some exceptional situations, the following holds true: there exist at most finitely many roots of unity~$\zeta$ such that for some~$n$ the polynomial $u_n(x) $ is not identically zero but ${u_n(\zeta)=0}$. The following example shows that we indeed have to exclude some exceptional cases. \begin{example} \label{exinf} Let $a,b$ be integers with $b$ non-zero, and assume that $$ c_2(x)/c_1(x)=\delta x^a, \qquad f_2(x)/f_1(x)=\varepsilon x^b, \qquad \delta,\varepsilon\in \{1,- 1\}. $$ We then get $$ u_n(x)=c_1(x)f_1(x)^n(1+\delta \varepsilon^n x^{a+bn}) $$ and we see that if $x=\zeta$ is such that $\zeta^{a+bn}=-\delta\varepsilon^n$, then ${u_n(\zeta)=0}$. The condition that $b\ne 0$ insures that $u_n(x)$ is non-zero for $n$ sufficiently large (in fact, for all $n$ except eventually one of them, namely $n=-a/b$), and every $u_n(x)$ vanishes at the roots of unity of order ${|a+bn|}$ or ${2|a+bn|}$ depending on the sign of $\delta\varepsilon^n$. \end{example} It turns out that this example is the only case when the polynomials $u_n(x)$ vanish at infinitely many roots of unity. We have the following theorem. \begin{theorem} \label{thm:thmmain} Let ${c_1(x),c_2(x),f_1(x),f_2(x)\in \Q[x]}$ be non-zero polynomials. For a positive integer~$n$ define $u_n(x)$ as in~\eqref{eq:1}. Then the following two conditions are equivalent. \begin{enumerate} \item \label{iinf} There exist infinitely many roots of unity~$\zeta$ such that for some~$n$ the polynomial $u_n(x) $ is not identically zero but ${u_n(\zeta)=0}$. \item \label{iex} There exist ${a,b \in \Z}$ with ${b\ne 0}$ and ${\delta,\varepsilon \in \{1,-1\}}$ such that $$ c_2(x)/c_1(x)=\delta x^a, \qquad f_2(x)/f_1(x)=\varepsilon x^b. $$ \end{enumerate} \end{theorem} It is not hard to derive this theorem from classical results on unlikely intersection like the Theorem of Bombieri-Masser-Zannier-Maurin~\cite{BHMZ10,BMZ99,Ma08}. See also the recent work of Ostafe and Shparlinski~\cite{Os16,OS20}, especially Theorem~2.11 and Corollary~2.14 in~\cite{OS20}. However, we are mainly interested in a quantitative statement: when condition~\ref{iex} of Theorem~\ref{thm:thmmain} is not satisfied, we want to bound the orders of the roots of unity~$\zeta$ such that ${u_n(\zeta)=0}$ for some~$n$, in terms of the degrees and the heights of our polynomials $f_i,c_i$. To the best of our knowledge, no quantitative version of the Bombieri-Masser-Zannier-Maurin theorem is available which would imply such a bound. To state our result, let us recall the definition of the height of a non-zero polynomial in $\Q[x]$. The height of a primitive vector ${{\mathbf a}=(a_1, \ldots,a_k)\in \Z^{k}}$ (\textit{primitive} means that ${\gcd(a_1,\ldots, a_k)=1}$) is defined by $$ {\mathrm{h}}({\mathbf a}):=\log\max\{|a_1|, \ldots, |a_k|\}. $$ In general, given a non-zero vector ${{\mathbf a}\in \Q^{k+1}}$, there exists ${\lambda \in \Q^\times}$, well defined up to multiplication by~$\pm1$, such that ${{\mathbf a}^\ast=\lambda {\mathbf a}}$ is primitive, and we set ${{\mathrm{h}}({\mathbf a}):={\mathrm{h}}({\mathbf a}^\ast)}$. We define the height of a non-zero polynomial ${g(x)\in \Q[x]}$ as the height of the vector of its coefficients. More generally, we define the height of a non-zero vector ${(g_1, \ldots, g_k)\in \Q[x]^k}$ as the height of the vector formed of the coefficients of all polynomials $g_1, \ldots, g_k$. We have the following theorem. \begin{theorem} \label{thm:thmmain1} Let ${c_1(x),c_2(x),f_1(x),f_2(x)\in \Q[x]}$ be non-zero polynomials such that condition~\ref{iex} of Theorem~\ref{thm:thmmain} is not satisfied. Set \begin{align*} D&:=\max\{\deg c_1,\deg c_2,\deg f_1, \deg f_2\},\\ X&:=\max\{3,{\mathrm{h}}(c_1,c_2),{\mathrm{h}}(f_1,f_2)\}. \end{align*} Let~$m$ be a positive integer and~$\zeta$ a primitive $m$th root of unity such that for some~$n$ the polynomial $u_n(x) $ is not identically zero but ${u_n(\zeta)=0}$. Then \begin{equation} \label{eupperm} m\le e^{100D(X+D)}. \end{equation} \end{theorem} The numerical constant~$100$ here is rather loose; probably, one can replace it by~$4$ or so. One may ask whether there is a bound for~$m$ which depends only on one of the parameters~$D$ or~$X$. The following examples show that this is not the case. \begin{example} Consider ${u_n(x)=(2x)^n-2^m}$, for which $$ (c_1(x),c_2(x),f_1(x),f_2(x))=(1,-2^m,2x,1), \qquad X=\max\{3,m\log 2\}. $$ Then ${u_m(x)=2^m(x^m-1)}$ vanishes at primitive $m$th roots of unity, and we have ${m\ge X/\log2}$ (provided ${m\ge 5}$). Hence no bound independent of~$X$ is possible. \end{example} \begin{example} \label{exd} Consider ${u_n(x)=x^n+x^D+1}$, for which $$ (c_1(x),c_2(x),f_1(x),f_2(x))=(1,x^D+1,x,1). $$ Then $u_{2D}=(x^{3D}-1)/(x^D-1)$ vanishes at primitive $3D$th roots of unity, so we have ${m\ge 3D}$. Hence no bound independent of~$D$ is possible. \end{example} One may also ask whether in Theorem~\ref{thm:thmmain1} one can bound~$n$ such that $u_n(x)$ vanishes at a root of unity. The answer is ``no'' in general. Indeed, if polynomials ${c_1(x)f_1(x)}$ and ${c_2(x)f_2(x)}$ have a common root, then every $u_n(x)$ will vanish at that root. But even if ${c_1(x)f_1(x)}$ and ${c_2(x)f_2(x)}$ do not simultaneously vanish at some root of unity, it is still possible that $u_n(x)$ vanishes at a root of unity for infinitely many~$n$. This is, for instance, the case for the sequence ${u_n(x)=x^n+x^D+1}$ from Example~\ref{exd}: it vanishes at primitive $3D$th roots of unity whenever ${n\equiv 2D\bmod 3D}$. Nevertheless, we can bound the \textit{smallest}~$n$ with this property. Here is the precise statement. \begin{theorem} \label{thsmallestn} In the set-up of Theorem~\ref{thm:thmmain}, assume that, for a given~$m$, the set of positive integers~$n$ with the property ``the polynomial $u_n(x)$ is not identically~$0$ but vanishes at an $m$th root of unity'' is not empty. Then the smallest~$n$ in this set satisfies $$ n \le m(\log m)^3(X+\log D). $$ More precisely, either there exists~$n$ in this set satisfying ${n\le 2m}$, or every~$n$ in this set satisfies ${n\le m(\log m)^3(X+\log D)}$. \end{theorem} Throughout the article we use standard notation. We denote $\varphi(n)$ the Euler function, $\mu(n)$ the Möbius function, $\Lambda(n)$ the von Mangoldt function and ${\omega(n)}$ the number of prime divisors of~$n$ counted without multiplicities. Theorems~\ref{thm:thmmain} and~\ref{thm:thmmain1} are proved in Section~\ref{sproofs}, and Theorem~\ref{thsmallestn} is proved in Section~\ref{ssmallestn}. In Sections~\ref{sheights},~\ref{scyclo} and~\ref{sprim} we collect various auxiliary facts used in the proof. In particular, in Section~\ref{sprim} we revisit Schinzel's classical Primitive Divisor Theorem~\cite{Sc74}. We obtain a version of this theorem fully explicit in all parameters, which is key ingredient in our proof of Theorem~\ref{thm:thmmain1}. \section{Heights} \label{sheights} All results of this section are well-known, but sometimes we prefer to give a short proof than to look for a bibliographical reference. Recall the definition of the absolute logarithmic (projective) height. Let $$ \bar\alpha=(\alpha_0,\alpha_1, \ldots, \alpha_k)\in \bar\Q^{k+1} $$ be a non-zero vector of algebraic numbers. Pick a number field~$K$ containing all~$\alpha_i$ and normalize the absolute values of~$K$ to extend the standard absolute values of~$\Q$. With this normalization, the height of~$\bar\alpha$ is defined by \begin{equation} \label{eheight} {\mathrm{h}}(\bar\alpha)=d^{-1}\sum_{v\in M_K}d_v\log\max\{|\alpha_0|_v,\ldots, |\alpha_k|_v\}, \end{equation} where ${d=[K:\Q]}$ and ${d_v=[K_v:\Q_v]}$ is the local degree. This definition is known to be independent of the choice of~$K$ and invariant under multiplication of~$\bar\alpha$ by a non-zero algebraic number: ${{\mathrm{h}}(\lambda\bar\alpha)={\mathrm{h}}(\bar\alpha)}$ for ${\lambda \in \bar\Q^\times}$. When ${\alpha \in \Q^{n+1}}$ this definition coincides with the definition of height from Section~\ref{sintr}. Separating the contributions of infinite and finite places, we can rewrite equation~\eqref{eheight} as \begin{equation} \label{eheightold} \begin{aligned} {\mathrm{h}}(\bar\alpha)&= d^{-1}\sum_{K\stackrel\sigma\hookrightarrow \C} \log \max\{|\alpha_0^\sigma|,\ldots, |\alpha_k^\sigma|\}\\ &+ d^{-1}\sum_{{\mathfrak{p}}}\max\{-\nu_{\mathfrak{p}}(\alpha_0), \ldots, -\nu_{\mathfrak{p}}(\alpha_k)\}\log{\mathcal{N}}{\mathfrak{p}}, \end{aligned} \end{equation} where the first sum is over the complex embeddings of~$K$, the second sum is over the finite primes of~$K$, and ${\mathcal{N}}{\mathfrak{p}} $ denotes the absolute norm of~${\mathfrak{p}}$. Now we define the height ${\mathrm{h}}(g)$ of a non-zero polynomial~$g$ with algebraic coefficients (in one or in several variables), or, more generally, the height ${{\mathrm{h}}(g_1,\ldots, g_k)}$ of a vector of such polynomials as the height of the vector of all coefficients of those polynomials (ordered somehow). With a standard abuse of notation, for ${\alpha\in \bar\Q}$ we write ${{\mathrm{h}}(\alpha)}$ for ${{\mathrm{h}}(1,\alpha)}$. If~$\alpha$ belongs to a number field~$K$ then \begin{align} \label{ehplus} {\mathrm{h}}(\alpha)&=d^{-1}\sum_{v\in M_K}d_v\log^+|\alpha|_v\\ \label{ehmin} &=d^{-1}\sum_{v\in M_K}-d_v\log^-|\alpha|_v \qquad (\alpha\ne 0), \end{align} where ${\log^+=\max\{\log, 0\}}$ and ${\log^-=\min \{\log,0\}}$. \begin{lemma} \label{lhpol} Let ${\alpha\in \bar\Q}$ and ${f(x)\in\bar\Q[x]}$ a polynomial of degree less or equal to~$D$. Then \begin{equation} \label{ehfa} {\mathrm{h}}(f(\alpha))\le D{\mathrm{h}}(\alpha)+ {\mathrm{h}}(1,f)+\log(D+1). \end{equation} More generally, if ${g(x)\in\bar\Q[x]}$ is another polynomial of degree less or equal to~$D$ and ${g(\alpha) \ne0}$ then \begin{equation} \label{ehfga} {\mathrm{h}}(f(\alpha)/g(\alpha))\le D{\mathrm{h}}(\alpha)+ {\mathrm{h}}(g,f)+\log(D+1). \end{equation} If ${f(\alpha) =0}$ then \begin{equation} \label{ehroot} {\mathrm{h}}(\alpha)\le {\mathrm{h}}(f)+\log 2. \end{equation} Furthermore, let~$r$ be a non-negative integer. Then \begin{equation} \label{ehder} {\mathrm{h}}(1,f^{(r)}/r!) \le {\mathrm{h}}(1,f)+D\log2. \end{equation} \end{lemma} \begin{proof} We start by proving~\eqref{ehfga}. By definition, $$ {\mathrm{h}}(f(\alpha)/g(\alpha))={\mathrm{h}}(1,f(\alpha)/g(\alpha))={\mathrm{h}}(g(\alpha), f(\alpha)). $$ Write $$ f(x)=a_Dx^D+\cdots+a_0, \qquad g(x)=b_Dx^D+\cdots+b_0. $$ Let~$K$ be a number field containing~$\alpha$ and the coefficients of $f,g$. We set ${d=[K:\Q]}$. For ${v\in M_K}$ we have $$ |f(\alpha)|_v \le \begin{cases} (D+1)|f|_v\max\{1,|\alpha|_v\}^D,& v\mid \infty,\\ |f|_v|\max\{1,|\alpha|_v\}^D,& v<\infty, \end{cases} $$ where ${|f|_v =\max\{|a_0|_v, \ldots,|a_D|_v\}}$, and similarly for $g(\alpha)$. Hence \begin{align*} {\mathrm{h}}(g(\alpha), f(\alpha))&\le d^{-1}\sum_{v\in M_K}d_v\log\max\{|g(\alpha)|_v,|f(\alpha)|_v\}\\ &\le d^{-1}\sum_{v\in M_K}d_v(\log\max\{|f_v|,|g|_v\}+D\log^+|\alpha|_v) \\ &+d^{-1}\sum_{\substack{v\in M_K\\v\mid \infty}}d_v\log(D+1)\\ &= {\mathrm{h}}(g,f)+D{\mathrm{h}}(\alpha)+ \log(D+1), \end{align*} which proves~\eqref{ehfga}. For~\eqref{ehroot} see \cite[Proposition~3.6(1)]{BB13}. Finally, we have $$ \frac{f^{(r)}}{r!}(x)=\sum_{k=r}^D\binom kr a_k x^{k-r}. $$ Since $$ \binom kr \le 2^k\le 2^D, $$ we have $$ \left|\frac{f^{(r)}}{r!}\right|_v \le \begin{cases} 2^D|f|_v, & v\mid \infty,\\ |f|_v, & v<\infty. \end{cases} $$ Hence \begin{align*} {\mathrm{h}}\left(1, \frac{f^{(r)}}{r!}\right) &= d^{-1}\sum_{v\in M_K}d_v\log^+\left|\frac{f^{(r)}}{r!}\right|_v\\ &\le d^{-1}\sum_{v\in M_K}d_v\log^+|f_v| +d^{-1}\sum_{\substack{v\in M_K\\v\mid \infty}}d_vD\log2\\ &= {\mathrm{h}}(1,f)+D\log2. \end{align*} The lemma is proved. \end{proof} \begin{lemma} \label{lhdivide} Let ${f_1(x), \ldots, f_k(x)\in \bar \Q[x]}$ be non-zero polynomials of degrees not exceeding~$D$, and let ${g(x) \in \bar\Q[x]}$ be a common divisor of ${f_1, \ldots, f_k}$ (in the ring ${\bar\Q[x]}$). Then $$ {\mathrm{h}}(f_1/g, \ldots, f_k/g) \le {\mathrm{h}}(f_1, \ldots, f_k) +(D+k-1)\log2. $$ \end{lemma} \begin{proof} Consider the polynomial $$ f(x,y_1, \ldots, y_{k-1}):=f_1(x)y_1+\cdots +f_{k-1}(x)y_{k-1}+ f_k(x) \in \bar\Q[x,y_1, \ldots, y_{k-1}]. $$ Applying Theorem~1.6.13 from \cite{BG06}, we obtain $$ {\mathrm{h}}(f/g) \le {\mathrm{h}}(f/g)+{\mathrm{h}}(g) \le {\mathrm{h}}(f) +(D+k-1)\log2. $$ Since $$ {\mathrm{h}}(f_1/g, \ldots, f_k/g)= {\mathrm{h}}(f/g), \qquad {\mathrm{h}}(f_1, \ldots, f_k)={\mathrm{h}}(f), $$ the result follows. \end{proof} \begin{lemma} \label{lvals} Let~$K$ be a number field of degree~$d$ and ${\alpha\in K}$. Then \begin{equation} \label{evalsone} \sum_{\nu_{\mathfrak{p}}(\alpha)<0}\log{\mathcal{N}}{\mathfrak{p}} \le d{\mathrm{h}}(\alpha), \qquad \sum_{\nu_{\mathfrak{p}}(\alpha)>0}\log{\mathcal{N}}{\mathfrak{p}} \le d{\mathrm{h}}(\alpha), \end{equation} where the first sum is over (finite) primes~${\mathfrak{p}}$ of~$K$ with ${\nu_{\mathfrak{p}}(\alpha)<0}$, the second sum over those with ${\nu_{\mathfrak{p}}(\alpha)>0}$, and in the second sum we assume ${\alpha\ne 0}$. More generally, let ${\alpha_1, \ldots, \alpha_k\in K}$. Then \begin{equation} \label{evalsmany} \sum_{\substack{\nu_{\mathfrak{p}}(\alpha_i)<0\ \text{for} \\\text{some}\ i\in \{1,\ldots,k\}}}\log{\mathcal{N}}{\mathfrak{p}} \le d{\mathrm{h}}(\bar\alpha), \qquad \bar\alpha=(1,\alpha_1, \ldots, \alpha_k). \end{equation} \end{lemma} \begin{proof} Inequality~\eqref{evalsmany} is immediate from~\eqref{eheightold} (note that ${\alpha_0=1}$), and both statements in~\eqref{evalsone} are special cases of~\eqref{evalsmany}. \end{proof} \begin{lemma}[``Liouville's inequality''] \label{lliouv} Let~$K$ and~$\alpha$ be as in Lemma~\ref{lvals}, ${\alpha \ne 0}$. Let ${S\subset M_K}$ be any set of places of~$K$ (finite or infinite). Then $$ e^{-d{\mathrm{h}}(\alpha)}\le \prod_{v\in M_K}|\alpha|_v^{d_v}\le e^{d{\mathrm{h}}(\alpha)}. $$ In particular, if ${\sigma_1, \ldots \sigma_r:K\hookrightarrow\C}$ are some distinct complex embeddings of~$K$ then $$ \prod_{i=1}^r |\alpha^{\sigma_i}|\ge e^{-d{\mathrm{h}}(\alpha)}. $$ \end{lemma} We omit the proof, which is well-known and easy. \section{Cyclotomic polynomials} \label{scyclo} We denote ${\Phi_m(T)}$ the $m$th cyclotomic polynomial. We will systematically use the identity \begin{equation} \label{ecyclomu} \Phi_m(T) = \prod_{d\mid m}(T^d-1)^{\mu(m/d)}, \end{equation} In this section we study values of cyclotomic polynomials at algebraic points. We give an asymptotic expression for the height of $\Phi_m(\gamma)$ as ${\gamma \in \bar\Q}$ is fixed and ${m\to\infty}$. We also estimate the absolute value of $\Phi_m(\gamma)$ from below. The results of this section can be viewed as totally explicit versions of some results from \cite[Section~3]{BBL13}, and we follow~\cite{BBL13} rather closely. We note however that all this goes back to the 1974 work of Schinzel~\cite{Sc74} or even earlier. \subsection{The height} \begin{theorem} \label{tas} Let~$\gamma$ be an algebraic number. Then $$ {\mathrm{h}}(\Phi_m(\gamma))=\varphi(m){\mathrm{h}}(\gamma)+O_1\bigl(2^{\omega(m)}\log (\pi m)\bigr). $$ \end{theorem} Recall that ${A=O_1(B)}$ means that ${|A|\le B}$. To prove this theorem we need some preparations. We follow \cite[Section~3]{BBL13} with some changes. \begin{proposition} \label{pcyc} For a positive integer~$m$ we have \begin{equation} \label{ephinz} \max_{|z|\le1}\log|\Phi_m(z)|\le 2^{\omega(m)}\log (\pi m), \end{equation} the maximum being over the unit disc on the complex plane. (We use the convention ${\log0=-\infty}$.) For ${0<\varepsilon\le 1/2}$ we also have \begin{equation} \label{ephitriv} \min_{|z|\le1-\varepsilon}\log|\Phi_m(z)|\ge -2^{\omega(m)}\log \frac1\varepsilon. \end{equation} \end{proposition} \begin{proof} By the maximum principle, it suffices to prove that~\eqref{ephinz} holds for complex~$z$ with ${|z|=1}$. Thus, fix such~$z$. We will actually prove a slightly sharper bound \begin{equation} \label{ecircle} \log|\Phi_m(z)|\le (2^{\omega(m)-1}+1)\log m+2^{\omega(m)}\log\pi. \end{equation} We can write~$z$ in a unique way as ${z=\zeta e^{2\pi i\theta/m}}$, where~$\zeta$ is an $m$th root of unity (not necessarily primitive) and ${-1/2<\theta\le 1/2}$. We may assume ${\theta\ne 0}$, because for the finitely many~$z$ with ${\theta=0}$ the bound extends by continuity. Let~$\ell$ be the exact order of~$\zeta$; thus, ${\ell\mid m}$ and~$\zeta$ is a primitive $\ell$th root of unity. Let~$d$ be any other divisor of~$m$. If ${\ell\nmid d}$ then ${d\le m/2}$ and $$ 2\ge |z^d-1| \ge 2\sin(\pi d/2m)\ge 2d/m. $$ (We use the inequality ${|\sin x|\ge (2/\pi)x}$ which holds for ${|x|\le\pi/2}$.) This implies that \begin{equation} \label{elndivd} \bigl|\log |z^d-1| \bigr|\le\log (m/d). \end{equation} And if ${\ell\mid d}$ then we have ${|z^m-1| = 2\sin(\pi \theta d/m)}$, which implies that $$ 2\pi\theta d/m \ge |z^d-1|\ge 4\theta d/m. $$ Writing ${d=d'\ell}$, this implies that \begin{equation} \label{eldivd} \log |z^{d'\ell}-1| = \log d'-\log \frac m{2\ell\theta}+O_1\left(\log\pi\right). \end{equation} Using~\eqref{ecyclomu} we obtain \begin{align*} \log|\Phi_m(z)| &= \sum_{\genfrac{}{}{0pt}{}{d\mid m}{\ell\nmid d}} \mu\left(\frac md\right) \log |z^d-1|+ \sum_{d'\mid m/\ell}\mu\left(\frac{m/\ell}{d'}\right) \log |z^{\ell d'}-1|\\ &\le \sum_{d\mid m} \left|\mu\left(\frac md\right)\right| \log \frac md+ \sum_{d'\mid m/\ell}\mu\left(\frac{m/\ell}{d'}\right) \left(\log d'-\log \frac m{2\ell\theta}\right)\\ &\hphantom{\le{}}+O_1(2^{\omega(n/\ell)}\log\pi)\\ &=2^{\omega(m)-1}\sum_{p\mid m}\log p +\Lambda\left(\frac m\ell\right) +\delta\log(2\theta)+ O_1(2^{\omega(m/\ell)}\log\pi), \end{align*} where ${\delta=0}$ if ${\ell<m}$ and ${\delta=1}$ if ${\ell=m}$. Since ${\log(2\theta)\le 0}$, this proves~\eqref{ecircle}. The proof of~\eqref{ephitriv} is much easier. When ${|z|\le 1-\varepsilon}$, we have $$ 2\ge |z^d-1|\ge 1-|z|^d\ge 1-|z|\ge \varepsilon. $$ Since ${0<\varepsilon\le 1/2}$ this implies that ${\bigl|\log|z^d-1|\bigr|\le \log (1/\varepsilon)}$. We obtain $$ \bigl|\log |\Phi_m(z)|\bigr|=\left|\sum_{d\mid m}\mu\left(\frac md\right) \log|z^d-1|\right| \le 2^{\omega(m)}\log \frac1\varepsilon. $$ In particular,~\eqref{ephitriv} holds. \end{proof} \begin{corollary} \label{ccyc} Let~$m$ be a positive integer and ${z\in \C}$. Then $$ \log^+|\Phi_m(z)|= \varphi(m)\log^+|z|+O_1\bigl(2^{\omega(m)}\log (\pi m)\bigr), $$ where ${\log^+=\max\{\log, 0\}}$. \end{corollary} \begin{proof} For ${|z|\le 1}$ this is Proposition~\ref{pcyc}. If ${ |z|>1}$ then \begin{equation} \label{ephire} \log|\Phi_m(z)|=\varphi(m)\log|z|+\log|\Phi_m(z^{-1})|, \end{equation} and ${\log|\Phi_m(z^{-1})|\le 2^{\omega(m)}\log (\pi m)}$ by Proposition~\ref{pcyc}. This already implies the upper bound $$ \log^+|\Phi_m(z)|\le \varphi(m)\log^+|z|+2^{\omega(m)}\log (\pi m). $$ The lower bound \begin{equation} \label{ephilo} \log^+|\Phi_m(z)|\ge \varphi(m)\log^+|z|-2^{\omega(m)}\log (\pi m) \end{equation} is trivial when ${m=1}$, so we will assume ${m\ge 2}$ in the sequel. In the case ${1<|z|\le m/(m-1)}$ we have $$ \log^+|\Phi_m(z)| \ge 0 \ge \varphi(m)\log\frac{m}{m-1}-1\ge \varphi(m)\log^+|z|-1, $$ which is much better than wanted. Finally, if ${|z|\ge m/(m-1)}$, then $$ \log|\Phi_m(z^{-1})|\ge -2^{\omega(m)}\log m $$ by~\eqref{ephitriv} with ${\varepsilon=1/m}$. Hence~\eqref{ephilo} follows from~\eqref{ephire} in this case. \end{proof} \paragraph{Proof of Theorem~\ref{tas}.} We use~\eqref{ehplus} with ${\alpha=\Phi_m(\gamma)}$. For ${v\in M_K}$ we have $$ \log^+|\Phi_m(\gamma)|_v= \begin{cases} \varphi(m)\log^+|\gamma|_v+O_1\bigl(2^{\omega(m)}\log (\pi m)\bigr), & v\mid \infty,\\ \varphi(m)\log^+|\gamma|_v, & v<\infty. \end{cases} $$ Indeed, the archimedean case is Corollary~\ref{ccyc}, and the non-archimedean case is obvious. Summing up, the result follows. \qed \subsection{The lower bound} The following result is proved in \cite[Corollary~4.2]{BL20} as a consequence of Baker's theory of logarithmic forms. \begin{proposition} \label{pabs} Let~$\gamma$ be a complex algebraic number of degree~$d$, not a root of unity, and~$n$ a positive integer. Then \begin{equation*} |\gamma^n-1|\ge e^{-10^{12}d^4({\mathrm{h}}(\gamma)+1)\log (n+1)}. \end{equation*} \end{proposition} \begin{corollary} \label{carch} Let~$\gamma$ and~$m$ be as in Proposition~\ref{pabs}. Then \begin{equation} \label{elowerreal} \log |\Phi_m(\gamma)|\ge -10^{12}d^4({\mathrm{h}}(\gamma)+1)\cdot 2^{\omega(m)}\log (m+1). \end{equation} \end{corollary} \begin{proof} If ${|\gamma|\ge 1}$ then $$ \log|\Phi_m(\gamma)|=\varphi(m)\log|\gamma|+\log|\Phi(\gamma^{-1})|\ge \log|\Phi(\gamma^{-1})|. $$ Hence, replacing, if necessary,~$\gamma$ by~$\gamma^{-1}$, we may assume ${|\gamma|\le 1}$. We have \begin{equation} \label{esumagain} \log|\Phi_m(\gamma)|=\sum_{n\mid m}\mu\left(\frac mn\right) \log |\gamma^n-1|. \end{equation} Proposition~\ref{pabs} implies that $$ 2\ge |\gamma^n-1|\ge e^{-10^{12}d^4({\mathrm{h}}(\gamma)+1)\log (n+1)}. $$ Hence for ${1\le n\le m}$ we have $$ \bigl|\log |\gamma^n-1|\bigr|\le 10^{12}d^4({\mathrm{h}}(\gamma)+1)\log (m+1). $$ Substituting this to~\eqref{esumagain}, we obtain $$ \bigl|\log |\Phi_m(\gamma)|\bigr|\le 10^{12}d^4({\mathrm{h}}(\gamma)+1)\cdot2^{\omega(m)}\log (m+1). $$ In particular, we proved~\eqref{elowerreal}. \end{proof} \section{Schinzel's Primitive Divisor Theorem} \label{sprim} Let~$\gamma$ be a non-zero algebraic number, not a root of unity. We consider the sequence $$ u_n=u_n(\gamma)=\gamma^n-1. $$ (Note that in this section $(u_n)$ is a numerical sequence, while in the other sections it is a sequence of polynomials.) A prime ${\mathfrak{p}}$ of the number field ${K=\Q(\gamma)}$ is called \textit{primitive divisor} for~$u_n$ if $$ \nu_{\mathfrak{p}}(u_n) >0, \qquad \nu_{\mathfrak{p}}(u_k)=0 \quad (k=1, \ldots, n-1). $$ For further use, let us fix here some basic properties of primitive divisors. Recall that ${\Phi_n(T)}$ denotes the $n$th cyclotomic polynomial, and~${\mathcal{N}}{\mathfrak{p}}$ is the absolute norm of~${\mathfrak{p}}$. \begin{proposition} \label{pprim} Assume that~${\mathfrak{p}}$ is a primitive divisor of~$u_n$. Then~$n$ divides ${{\mathcal{N}}{\mathfrak{p}}-1}$ and ${\nu_{\mathfrak{p}}(\Phi_n(\gamma))\ge 1}$. In particular, ${n<{\mathcal{N}}{\mathfrak{p}}}$. \end{proposition} The proofs are very easy and we omit them. Schinzel~\cite{Sc74} proved that~$u_n$ admits a primitive divisor for ${n\ge n_0(d)}$, where~$d$ is the degree of~$\gamma$. This was an improvement upon the earlier work~\cite{PS68}, where the same was proved under the assumption ${n\ge n_0(\gamma)}$. Stewart~\cite{St77} made Schinzel's result explicit, but he imposed an additional hypothesis ${\gamma=\alpha/\beta}$, where ${\alpha,\beta\in {\mathcal{O}}_K}$ are coprime algebraic integers. Here we obtain a fully explicit version of Schinzel's result without any extra hypothesis. \begin{theorem} \label{thschin} Let~$\gamma$ be an algebraic number of degree~$d$, not a root of unity. Assume that \begin{equation} \label{ehypo} n\ge \max\{2^{d+1},10^{30}d^{9}\}. \end{equation} Then ${u_n=\gamma^n-1}$ admits a primitive divisor. \end{theorem} Theorem~\ref{thschin} is a consequence of the following result, appearing, albeit in a different setting, in Schinzel's work. \begin{proposition} \label{pup} In the above set-up, assume that~$u_n$ does not admit a primitive divisor. Then \begin{equation} \label{eup} {\mathrm{h}}(\Phi_n(\gamma)) \le 10^{13}d^4 ({\mathrm{h}}(\gamma)+1)\cdot2^{\omega(n)}\log (n+1). \end{equation} \end{proposition} \subsection{Proof of Proposition~\ref{pup}} We start from the following well-known fact. \begin{lemma} \label{lwellknown} Let~$K$ be a number field of degree~$d$ and~$p$ a prime number. Let~${\mathfrak{p}}$ be a prime of~$K$ above~$p$ of ramification index~$e_{\mathfrak{p}}$ (that is, ${e_{\mathfrak{p}}=\nu_{\mathfrak{p}}(p)}$). Let ${\xi\in K}$ satisfy $$ \nu_{\mathfrak{p}}(\xi-1) >\frac {e_{\mathfrak{p}}}{p-1}. $$ Then for any positive integer~$n$ we have $$ \nu_{\mathfrak{p}}(\xi^n-1)=\nu_{\mathfrak{p}}(\xi-1)+\nu_{\mathfrak{p}}(n). $$ \end{lemma} The proof of the lemma can be found, for instance, in \cite[Lemma~1]{PS68}. \begin{lemma} \label{lschin} Let~$\gamma$ be an algebraic number of degree~$d$, not a root of unity, and~$n$ an integer satisfying ${n\ge 2^{d+1}}$. Let~${\mathfrak{p}}$ be a prime of the field $\Q(\gamma)$ which is not a primitive divisor of ${u_n=\gamma^n-1}$. Then ${\nu_{\mathfrak{p}}(\Phi_n(\gamma))\le \nu_{\mathfrak{p}}(n)}$. \end{lemma} This is Schinzel's~\cite{Sc74} crucial ``Lemma~4''. Since his set-up is slightly different, we reproduce the proof here. \begin{proof} We may assume that ${\nu_{\mathfrak{p}}(\gamma^n-1)>0}$, since there is nothing to prove otherwise. In particular, ${\nu_{\mathfrak{p}}(\gamma)=0}$. For ${k=0,1,2\ldots}$ denote~$\ell_k$ the multiplicative order of ${\gamma \bmod {\mathfrak{p}}^k}$; that is,~$\ell_k$ is the smallest positive integer~$\ell$ with the property ${\nu_{\mathfrak{p}}(\gamma^\ell- 1)\ge k}$. Clearly, ${\nu_{\mathfrak{p}}(\gamma^n -1)\ge k}$ if and only if ${\ell_k\mid n}$. Together with~\eqref{ecyclomu} this implies that for every~$k$ the following holds: \begin{equation} \label{eschinzel} \nu_{\mathfrak{p}}\bigl(\Phi_n(\gamma)\bigr)= \sum_{i=1}^k \sum_{\ell_i\mid m\mid n}\mu\left(\frac nm\right)+\sum_{\ell_{k+1}\mid m\mid n}\mu\left(\frac nm\right)\bigl(\nu_{\mathfrak{p}}(\gamma^m-1) -k\bigr) \end{equation} Let~$p$ be the rational prime below~${\mathfrak{p}}$ and ${e_{\mathfrak{p}}=\nu_{\mathfrak{p}}(p)}$ the ramification index. We will apply~\eqref{eschinzel} with $$ k=\left\lfloor \frac {e_{\mathfrak{p}}}{p-1}\right\rfloor, $$ which will be our choice of~$k$ from now on. We claim that \begin{equation} \label{eclaim} n>\ell_{k+1}. \end{equation} We postpone the proof of~\eqref{eclaim} (which is a bit messy) until later, and now complete the proof of the lemma assuming validity of~\eqref{eclaim}. Since ${n>\ell_{k+1}\ge\ell_i}$ for ${i=1, \ldots, k}$, the double sum in~\eqref{eschinzel} vanishes. Also, if ${\ell_{k+1}\mid m}$ then $$ \nu_{\mathfrak{p}}(\gamma^m-1)=\nu_{\mathfrak{p}}(\gamma^{\ell_{k+1}}-1)+\nu_{\mathfrak{p}}\left(\frac{m}{\ell_{k+1}}\right) $$ by Lemma~\ref{lwellknown}. Hence~\eqref{eschinzel} can be rewritten as \begin{equation} \label{eschinzelbis} \nu_{\mathfrak{p}}\bigl(\Phi_n(\gamma)\bigr)= \sum_{\ell_{k+1}\mid m\mid n}\mu\left(\frac nm\right)\bigl(\nu_{\mathfrak{p}}(\gamma^{\ell_{k+1}}-1) -k\bigr)+ \sum_{\ell_{k+1}\mid m\mid n}\mu\left(\frac nm\right)\nu_{\mathfrak{p}}\left(\frac{m}{\ell_{k+1}}\right). \end{equation} Since ${n>\ell_{k+1}}$, the first sum in~\eqref{eschinzelbis} vanishes. As for the second sum, it vanishes (just being empty) if ${\ell_{k+1}\nmid n}$. From now on assume that ${\ell_{k+1}\mid n}$ and set ${n'=n/\ell_{k+1}}$. We obtain \begin{equation*} \nu_{\mathfrak{p}}\bigl(\Phi_n(\gamma)\bigr)= e_{\mathfrak{p}}\sum_{m'\mid n'}\mu\left(\frac {n'}{m'}\right)\nu_p\left(m'\right)= \begin{cases} e_{\mathfrak{p}},& \text{$n'$ is a power of~$p$},\\ 0,& \text{otherwise}. \end{cases} \end{equation*} In any case we obtain ${\nu_{\mathfrak{p}}\bigl(\Phi_n(\gamma)\bigr)\le \nu_{\mathfrak{p}}(n)}$. This proves the lemma. We are left with the claim~\eqref{eclaim}. Note first of all that \begin{equation} \label{eellone} n>\ell_1 \end{equation} because~${\mathfrak{p}}$ is not a primitive divisor of~$u_n$. Another useful observation is that \begin{equation} \label{epelli} \ell_{i+1}\le p\ell_i\qquad (i=1,2, \ldots). \end{equation} Indeed, $$ \gamma^{p\ell_i}-1=\sum_{j=1}^{p-1}\binom pj(\gamma^{\ell_i}-1)^j+(\gamma^{\ell_i}-1)^p, $$ which implies that ${\nu_{\mathfrak{p}}(\gamma^{p\ell_i}-1)>\nu_{\mathfrak{p}}(\gamma^{\ell_i}-1)}$, proving~\eqref{epelli}. If ${k=0}$ then~\eqref{eclaim} is~\eqref{eellone}. Now assume that ${k\ge 1}$. In this case \begin{equation} \label{ebet} p-1\le e_{\mathfrak{p}}\le d. \end{equation} On the other hand, let ${p^{f_{\mathfrak{p}}}={\mathcal{N}}{\mathfrak{p}}}$ be the absolute norm of~${\mathfrak{p}}$. Clearly, $$ \ell_1\le p^{f_{\mathfrak{p}}}-1\le p^{d/e_{\mathfrak{p}}}-1. $$ In the special case ${p=3}$, ${e_{\mathfrak{p}}=d=2}$ we have ${k=1}$ and ${\ell_2\le p\ell_1 \le 6}$. Since ${n\ge 2^{d+1}=8}$ by the hypothesis, this proves~\eqref{eclaim} in this special case. From now on we assume that ${d\ge 3}$ for ${p=3}$. Using~\eqref{epelli} iteratively, we obtain $$ \ell_{k+1}\le p^k\ell_1 <p^{e_{\mathfrak{p}}/(p-1)+d/e_{\mathfrak{p}}}\le \max_{p-1\le t\le d}p^{t/(p-1)+d/t}=p^{1+d/(p-1)}. $$ We have to show that \begin{equation*} p^{1+d/(p-1)}\le 2^{d+1}. \end{equation*} This is true by inspection in the cases $$ p=2, \qquad p=3,\ d\ge 3, \qquad p=5, \ d\ge 4. $$ Now assume that ${p\ge 7}$, in which case ${d\ge 6}$. Since ${p\le d+1}$, we have $$ p^{1+d/(p-1)} \le (d+1) \cdot7^{d/6}. $$ A calculation shows that ${(d+1) \cdot7^{d/6} \le 2^{d+1}}$ for ${d\ge 6}$. This completes the proof of~\eqref{eclaim}. \end{proof} \paragraph{Proof of Proposition~\ref{pup}} We use~\eqref{ehmin} with ${\alpha=\Phi_n(\gamma)}$. For ${v\in M_K}$ we have $$ -\log^-|\Phi_n(\gamma)|_v\le \begin{cases} 10^{12}d^4({\mathrm{h}}(\gamma)+1)\cdot2^{\omega(n)}\log (n+1), & v\mid \infty,\\ -\log|n|_v, & v<\infty. \end{cases} $$ Indeed, the archimedean case is Corollary~\ref{carch}, and the non-archimedean case is Lemma~\ref{lschin}. Summing up, we obtain $$ {\mathrm{h}}(\Phi_n(\gamma))\le 10^{12}d^4({\mathrm{h}}(\gamma)+1)\cdot2^{\omega(n)}\log (n+1)+\log n, $$ which is sharper than~\eqref{eup}. \qed \subsection{Proof of Theorem~\ref{thschin}} Assume~$u_n$ does not have a primitive divisor, but~$n$ satisfies~\eqref{ehypo}. We have, in particular, ${n\ge 10^{30}}$. Comparing Proposition~\ref{pup} and Theorem~\ref{tas}, we obtain \begin{align*} \varphi(n){\mathrm{h}}(\gamma)&\le 10^{13}d^4 % ({\mathrm{h}}(\gamma)+1)\cdot2^{\omega(n)}\log (n+1)+2^{\omega(n)}\log (\pi n)\\ &\le 10^{14}d^4 ({\mathrm{h}}(\gamma)+1)\cdot2^{\omega(n)}\log (n+1). \end{align*} Since~$\gamma$ is not a root of unity, we have \begin{equation} \label{evout} d{\mathrm{h}}(\gamma) \ge 2(\log(3d))^{-3}, \end{equation} see \cite[Corollary~2]{Vo96}. Hence $$ \varphi(n){\mathrm{h}}(\gamma)\le 10^{15}d^5 (\log(3d))^3 {\mathrm{h}}(\gamma)\cdot2^{\omega(n)}\log (n+1), $$ which implies \begin{equation} \label{ealmost} \varphi(n)\le 10^{15}d^5 (\log(3d))^3 \cdot2^{\omega(n)}\log (n+1). \end{equation} For ${n\ge 10^{30}}$ we have \begin{equation} \label{eboundphomega} \varphi(n) \ge 0.5 \frac n{\log\log n},\qquad \omega(n)\le \frac{\log n}{\log\log n-1.2}, \end{equation} see \cite[Theorem~15]{RS62} and~\cite[Theorem~13]{Ro83}. Hence for ${n\ge 10^{30}}$ \begin{align*} 2^{\omega(n)}\frac{n}{\varphi(n)}\log(n+1) &\le n^{(\log2)/(\log\log(10^{30})-1.2)} \cdot 2(\log\log n) \cdot \log(n+1)\\ &\le n^{1/3}. \end{align*} Using this, we deduce from~\eqref{ealmost} the inequality ${n^{2/3}\le 10^{15}d^5 (\log(3d))^3 }$. A quick calculation shows that this inequality is incompatible with~\eqref{ehypo}. \qed \section{Proof of Theorems~\ref{thm:thmmain} and~\ref{thm:thmmain1}} \label{sproofs} Since condition~\ref{iex} of Theorem~\ref{thm:thmmain} trivially implies condition~\ref{iinf} (see Example~\ref{exinf}) it suffices to prove Theorem~\ref{thm:thmmain1}. Thus, in the sequel: \begin{itemize} \item $c_i(x)$ and $f_i(x)$ are polynomials not satisfying condition~\ref{iex} of Theorem~\ref{thm:thmmain} and $$ u_n(x)=c_1(x)f_1(x)^n+c_2(x)f_2(x)^n \qquad (n=1,2,\ldots); $$ \item $m$ and~$n$ are positive integers such that ${u_n(\zeta)=0}$ for a primitive $m$th root of unity~$\zeta$; since ${u_n(x)\in \Q[x]}$, this is equivalent to \begin{equation} \Phi_m(x)\mid u_n(x). \end{equation} \end{itemize} \subsection{Some reductions} We start by some general observations. \begin{itemize} \item We may assume that \begin{equation} \label{enonv} c_1(\zeta)c_2(\zeta)f_1(\zeta)f_2(\zeta) \ne 0. \end{equation} Otherwise ${\varphi(m)\le D}$, and, using \begin{equation} \label{ephgeroot} \varphi(m)\ge m^{1/2} \qquad (m\ne 2,6) \end{equation} (see~\cite{Va67}), we obtain ${m\le \max\{6, D^2\}}$, which is much sharper than what we want to prove. \item We may assume that at least one of $f_1,f_2$ is a non-constant polynomial. Otherwise ${\deg u_n(x)\le D}$, and we again obtain ${\varphi(m)\le D}$. \item We may assume that ${n>D}$. Otherwise ${\deg u_n(x) \le D+D^2}$, and, using~\eqref{ephgeroot} we obtain ${m\le \max\{6, (D+D^2)^2\}}$, again much sharper than the wanted result. \item Replacing ${c_i(x)}$ and ${f_i(x)}$ by $$ \tilde{c}_i(x):=c_i(x)/\gcd(c_1(x),c_2(x)), \qquad \tilde{f}_i(x):=f_i(x)/\gcd(f_1(x),f_2(x)), $$ respectively, we may assume that the polynomials $c_1,c_2$ are coprime in the ring $\Q[x]$, and so are $f_1,f_2$: \begin{equation} \label{ecoprime} \gcd(c_1(x),c_2(x))=\gcd(f_1(x),f_2(x)) =1. \end{equation} Lemma~\ref{lhdivide} implies that $$ {\mathrm{h}}(\tilde{c}_1,\tilde{c}_2) \le {\mathrm{h}}(c_1,c_2)+(D+1)\log2 \le X+(D+1)\log2, $$ and similarly for ${{\mathrm{h}}(\tilde{f}_1,\tilde{f}_2)}$. Hence, to prove~\eqref{eupperm} in the general case, it suffices to prove \begin{equation} \label{eupperngam} m\le e^{30D(X+D)} \end{equation} in the ``coprime case'', that is, assuming~\eqref{ecoprime}. \end{itemize} We distinguish several cases according to the nature of roots of our polynomials: \begin{enumerate} \item $f_1(x)f_2(x)$ admits a root which is non-zero and not a root of unity; \item $f_1(x)f_2(x)$ vanishes at a root of unity; \item $f_1(x)f_2(x)$ vanishes only at~$0$. \end{enumerate} These cases are treated separately in the subsequent subsections. \subsection{The polynomial $f_1(x)f_2(x)$ admits a root~$\gamma$ which is non-zero and not a root of unity} \label{ssgamma} By symmetry, we may assume that~$\gamma$ is a root of $f_1(x)$. Since the statement of Theorem~\ref{thm:thmmain1} is invariant under multiplication of the polynomials $c_1,c_2$ by the same non-zero rational number, we may assume that the polynomial $c_1(x)$ is monic. Similarly, we may assume that $f_1(x)$ is monic. Denote ${K=\Q(\gamma)}$. Then $$ d:=[K:\Q]\le D. $$ Since ${X\ge 3}$, the right-hand side of~\eqref{eupperngam} exceeds ${10^{30}D^9}$. Hence we may assume that $$ m>\max\{ 2^{d+1}, 10^{30}d^9\}. $$ Theorem~\ref{thschin} together with Proposition~\ref{pprim} implies now that there exists a prime~${\mathfrak{p}}$ of~$K$ such that ${\nu_{\mathfrak{p}}(\Phi_m(\gamma))>0}$ and \begin{equation*} m<{\mathcal{N}}{\mathfrak{p}}. \end{equation*} So we only have to bound ${\mathcal{N}}{\mathfrak{p}}$. \subsubsection{The numbers~$\beta$ and~$\delta$} We have ${f_2(\gamma)\ne 0}$ by~\eqref{ecoprime}. However, it it possible that ${c_2(\gamma)=0}$. Denote~$r$ the order of~$\gamma$ as a root of $c_2(x)$, and set $$ \beta=\frac{c_2^{(r)}(\gamma)}{r!}, \qquad \delta=f_2(\gamma), $$ These are non-zero elements of the number field~$K$. We claim that one of the following holds: \begin{align} \label{elocal} \nu_{\mathfrak{p}}(\alpha)&<0 \quad\text{for some coefficient $\alpha$ of~$c_1$ or~$f_1$ or~$c_2$ or~$f_2$};\\ \label{ebeta} \nu_{\mathfrak{p}}(\beta) &>0;\\ \label{edelta} \nu_{\mathfrak{p}}(\delta)&>0. \end{align} Indeed, since ${\nu_{\mathfrak{p}}(\Phi_m(\gamma))>0}$, there exists a primitive $m$th rooth of unity~$\zeta$ and a prime ${{\mathfrak{P}}\mid {\mathfrak{p}}}$ of the field $K(\zeta)$ such that $$ \nu_{\mathfrak{P}}(\zeta-\gamma) >0. $$ Now, if~\eqref{elocal} does not hold, then our four polynomials belong to ${{\mathcal{O}}_{\mathfrak{P}}[x]}$, where~${\mathcal{O}}_{\mathfrak{P}}$ is the local ring of~${\mathfrak{P}}$. Moreover, since~$f_1$ is monic, ${\gamma\in {\mathcal{O}}_{\mathfrak{P}}}$. Hence the polynomials $$ F(x):=\frac{c_1(x)f_1(x)^n}{(x-\gamma)^r}, \qquad G(x) :=\frac{c_2(x)}{(x-\gamma)^r} $$ belong to ${\mathcal{O}}_{\mathfrak{P}}[x]$ as well. Note that $F(x)$ is indeed a polynomial, and moreover $$ F(\gamma) =0, $$ because ${n>D\ge r}$. We have ${\beta=G(\gamma)}$ and ${F(\zeta) =-G(\zeta)f_2(\zeta)^n}$ (because ${u_n(\zeta)=0}$). This implies the following congruences in the ring~${\mathcal{O}}_{\mathfrak{P}}$: \begin{align*} \beta\delta^n \equiv G(\zeta)f_2(\zeta)^n \equiv -F(\zeta)\equiv -F(\gamma) \equiv 0\mod{\mathfrak{P}}. \end{align*} Hence either ${\beta\equiv0\bmod {\mathfrak{P}}}$ or ${\delta\equiv0\bmod {\mathfrak{P}}}$, which means that one of~\eqref{ebeta} or~\eqref{edelta} holds true. \subsubsection{Estimates} Now we are ready to estimate ${\mathcal{N}}{\mathfrak{p}}$. Using Lemma~\ref{lvals}, we obtain \begin{equation} \log{\mathcal{N}}{\mathfrak{p}} \le \max\{{\mathrm{h}}(1,c_1),{\mathrm{h}}(1,c_2),{\mathrm{h}}(1,f_1), {\mathrm{h}}(1,f_2), {\mathrm{h}}(\beta), {\mathrm{h}}(\delta)\}. \end{equation} Since $f_1(x)$ is a monic polynomial, we have \begin{equation} \label{ehfoneftwo} {\mathrm{h}}(1,f_1),{\mathrm{h}}(1,f_2)\le {\mathrm{h}}(f_1,f_2) \le X, \end{equation} and similarly for $c_1,c_2$. Furthermore, using Lemma~\ref{lhpol}, we find \begin{align*} {\mathrm{h}}(\gamma) & \le {\mathrm{h}}(f_1)+\log2\\ & \le X+\log2,\\ {\mathrm{h}}(\delta) &\le {\mathrm{h}}(1,f_2)+ D{\mathrm{h}}(\gamma) + \log(D+1)\\ &\le (D+1)X+2D,\\ {\mathrm{h}}(\beta) &\le {\mathrm{h}}(1, c_2^{(r)}/r!) + D{\mathrm{h}}(\gamma) + \log(D+1) \\ & \le {\mathrm{h}}(1,c_2) + D\log2 +DX+D\log2+\log(D+1) \\ &\le (D+1)X+2D. \end{align*} This implies that $$ \log{\mathcal{N}}{\mathfrak{p}} \le (D+1)X+2D <3DX. $$ Since ${m< {\mathcal{N}}{\mathfrak{p}}}$, this proves~\eqref{eupperngam}. \subsection{The polynomial $f_1(x)f_2(x)$ vanishes at a root of unity~$\xi$} We may assume that ${f_1(\xi)=0}$. Then ${f_2(\xi)\ne 0}$ by~\eqref{ecoprime}. Let us describe our argument informally. Since ${f_1(\xi)/f_2(\xi)=0}$, there exists ${\varepsilon>0}$ such that ${|f_1(z)/f_2(z)|\le 1/2}$ when ${|z-\xi|\le\varepsilon}$. Now assume that ${u_n(\zeta)=0}$ for some primitive $m$th root of unity~$\zeta$. Using~\eqref{enonv}, we may write \begin{equation} \label{ealphanow} 0\ne\alpha:=\frac{c_2(\zeta)}{c_1(\zeta)}=-\left(\frac{f_1(\zeta)}{f_2(\zeta)}\right)^n. \end{equation} Let ${\Q(\zeta)\stackrel\sigma\hookrightarrow \C}$ be a complex embedding of the field~$\Q(\zeta)$ such that~$\zeta^\sigma$ belongs to the $\varepsilon$-neighborhood of~$\xi$. Then ${|\alpha^\sigma| \le (1/2)^n}$. Define \begin{equation} \label{ebetanow} \beta:=\prod_{|\zeta^\sigma-\xi|\le \varepsilon} \alpha^\sigma, \end{equation} the product being over all~$\sigma$ as above. Since the $\varepsilon$-neighborhood of~$\xi$ contains a positive proportion of primitive $m$th roots of unity, we have $$ -\log|\beta|\gg n\varphi(m), $$ where the implied constant depends on our polynomials $c_i$ and $f_i$ and on our choice of~$\varepsilon$. On the other hand, ${\alpha\ne 0}$, and ${{\mathrm{h}}(\alpha) \ll1}$ by Lemma~\ref{lhpol}. Hence Liouville's inequality (Lemma~\ref{lliouv}) implies that $$ -\log|\beta| = \sum_{|\zeta^\sigma-\xi|\le \varepsilon}-\log |\alpha^\sigma| \ll [\Q(\zeta):\Q]=\varphi(m). $$ This bounds~$n$. This all will be made explicit in Subsection~\ref{sssexpl}. But first, we establish some simple lemmas. \subsubsection{Some lemmas} \begin{lemma} \label{lcountcoprime} Let ${a,b\in \R}$, ${a<b}$, and~$m$ a positive integer. Denote ${\varphi(m,a,b)}$ the number of integers~$k$ coprime with~$m$ and satisfying ${a\le k\le b}$. Then $$ \varphi(m,a,b) = (b-a)\varphi(m)+O_1(2^{\omega(m)}). $$ \end{lemma} For the proof, see \cite[Lemma~2.3]{FGL17}. \begin{lemma} \label{lcountroots} Let~$\varepsilon$ satisfy ${0<\varepsilon\le 1}$ and let~$\xi$ be a complex number on the unit circle; that is, ${|\xi|=1}$. Let~$m$ be a positive integer. Then there exist at least ${\pi^{-1}\varepsilon\varphi(m)- 2^{\omega(m)}}$ primitive $m$th roots of unity~$\zeta$ satisfying ${|\zeta-\xi|\le \varepsilon}$. \end{lemma} \begin{proof} Write ${\xi=e^{2\pi \theta i}}$ with ${\theta\in \R}$, and let ${\eta>0}$ be the smallest positive real number with the property ${2\sin (\pi \eta)=\varepsilon}$. Note that ${1/6\ge \eta>(2\pi)^{-1}\varepsilon}$. If~$k$ is an integer satisfying $$ m(\theta-\eta) \le k\le m(\theta+\eta), \qquad \gcd(m,k)=1, $$ then ${\zeta:=e^{2\pi i k/m}}$ is a primitive $m$th root of unity satisfying ${|\zeta-\xi|\le \varepsilon}$. Lemma~\ref{lcountcoprime} implies that there is at least ${2\eta\varphi(m)- 2^{\omega(m)}}$ choices for~$k$, with distinct~$k$ giving rise to distinct~$\zeta$ (this is because ${\eta\le 1/6}$). Since ${\eta \ge (2\pi)^{-1}\varepsilon}$, the result follows. \end{proof} \begin{lemma} \label{leps} Let ${f_1(x),f_2(x)\in \C[x]}$ be polynomials of degrees bounded by~$D$, and with coefficients bounded by ${H\ge 1}$ in absolute value. Let ${\xi\in \C}$ be such that $$ |\xi|\le 1, \qquad f_1(\xi)=0, \qquad f_2(\xi)=\delta\ne 0. $$ Set $$ \varepsilon= \frac{\min\{|\delta|,1\}}{3D^2H}. $$ Then for ${z\in \C}$ satisfying ${|z-\xi|\le \varepsilon}$ we have ${|f_1(z)/f_2(z)|\le 1/2}$. \end{lemma} \begin{proof} Since ${|\xi|\le 1}$ and ${\varepsilon \le 1/3D}$, we have for ${|z-\xi|\le\varepsilon}$ trivial estimates $$ |f_i'(z)|\le \frac12D(D+1)H(1+\varepsilon)^{D-1}\le D^2H \qquad (i=1,2). $$ Hence for ${|z-\xi|\le\varepsilon}$ we have \begin{equation*} |f_1(z)|\le D^2H\varepsilon \le \frac13|\delta|, \qquad |f_2(z)|\ge |\delta|-D^2H\varepsilon \ge \frac23|\delta|. \end{equation*} This proves the lemma. \end{proof} \subsubsection{The estimates} \label{sssexpl} As in Subsection~\ref{ssgamma} we may assume that~$f_1$ is monic, which implies that we have~\eqref{ehfoneftwo}. In particular, the coefficients of~$f_1$ and~$f_2$ are bounded in absolute value by ${H:=e^X}$. Set ${\delta=f_2(\xi)}$. Note that the degree of~$\xi$ is at most~$D$ and the height is~$0$, because it is a root of unity. Using Lemmas~\ref{lhpol} and~\ref{lliouv}, we estimate $$ |\delta|\ge e^{-{\mathrm{h}}(f_2(\xi))} \ge e^{-{\mathrm{h}}(1,f_2)-\log(D+1)}\ge ((D+1)H)^{-1}. $$ Setting ${\varepsilon =(6D^3H^2)^{-1}}$, Lemma~\ref{leps} implies that $$ \left|\frac{f_1(z)}{f_2(z)}\right|\le 1/2 $$ for ${z\in \C}$ with ${|z-\xi|\le \varepsilon}$. Now define~$\alpha$ and~$\beta$ as in~\eqref{ealphanow},~\eqref{ebetanow}. Then \begin{equation} \label{ebetasmall} -\log|\beta|\ge nr\log2, \end{equation} where~$r$ is the number of embeddings ${\Q(\zeta)\stackrel\sigma\hookrightarrow\C}$ such that ${|\zeta^\sigma-\xi|\le \varepsilon}$. Denote ${\sigma_1, \ldots, \sigma_r}$ all those~$\sigma$. Lemmas~\ref{lliouv} and~\ref{lhpol} imply that \begin{align*} -\log|\beta|&=\sum_{i=1}^r-\log|\alpha^{\sigma_i}|\\ &\le [\Q(\zeta):\Q] {\mathrm{h}}(\alpha) \\ &\le \varphi(m) ({\mathrm{h}}(c_1,c_2)+\log(D+1))\\ &\le \varphi(m) (X+\log(D+1)). \end{align*} Together with~\eqref{ebetasmall} this implies that \begin{equation} \label{erphm} n\le \frac{\varphi(m)}{r\log2}(X+\log(D+1)), \end{equation} so we only have to bound~$r$ from below. Lemma~\ref{lcountroots} implies that $$ r\ge \pi^{-1}\varepsilon \varphi(m)-2^{\omega(m)}, $$ where we recall that ${\varepsilon=(6D^3H^2)^{-1}}$ with ${H=e^X}$. Using~\eqref{eboundphomega} with~$n$ replaced by~$m$, a messy but trivial calculation shows that either ${m\le e^{30D(X+D)}}$ (as we want) or ${2^{\omega(m)} \le (2\pi)^{-1}\varepsilon \varphi(m)}$. Thus, ${r\ge (2\pi)^{-1}\varepsilon \varphi(m)}$, which, substituted to~\eqref{erphm}, gives $$ n\le 100 D^4e^{3X}. $$ Then $$ \varphi(m) \le \deg u_n(x) \le 200D^5e^{3X}, $$ and, using~\eqref{ephgeroot}, we deduce from this an estimate much sharper than~\eqref{eupperngam}. \subsection{The only root of $f_1(x)f_2(x)$ is~$0$} We may assume that ${f_1(x)=1}$ and ${f_2(x)=\kappa x^b}$, where ${\kappa\in \Q^\times}$ and $$ 1\le b\le D<n. $$ We recall the following theorem of Mann~\cite{Ma65}. \begin{theorem} Let ${a_0,a_1,\ldots,a_k\in\Q^\times}$ and $x_0=1,x_1,\ldots,x_k$ be roots of unity such that \begin{equation} \label{eq:Mann} a_0x_0+a_1x_1+\cdots+a_kx_k=0. \end{equation} Assume that \begin{equation} \label{eq:nondeg} \sum_{i\in I} a_i x_i\ne 0 \end{equation} for every non-empty proper subset ${I\subset \{0,\ldots,k\}}$. Then $x_i^m=1$ where $$ m=\prod_{p\le k+1} p. $$ \end{theorem} For us, we label $$ c_i(x)=\sum_{j=0}^D c_{i,j} x^j\quad {\text{\rm for}}\quad i=1,2, $$ and we get \begin{equation} \label{eq:Mann1} \sum_{j=0}^D c_{1,j} \zeta^j+\sum_{j=0}^D c_{2,j}\kappa^n \zeta^{j+nb}=0. \end{equation} This almost looks like the equation from Mann's theorem \eqref{eq:Mann} except that the non-degeneracy condition \eqref{eq:nondeg} might fail. So, let us study~\eqref{eq:Mann1}. Let $C$ be the set of non-zero coefficients among $c_{1,j}$ and $c_{2,j}\kappa^n$ for $0\le j\le D$. If $c\in C$ then ${c=c_{\ell,j}\kappa^{\delta n}}$ for some $\ell\in \{1,2\}$ and $j\in \{0,\ldots,j\}$, then put $x_c=\zeta^{j+\delta nb}$. Here, we take $\delta=0$ if $\ell=1$ and $\delta=1$ if $\ell=2$. With these conventions, equation \eqref{eq:Mann1} is $$ \sum_{c\in C} cx_c=0. $$ This splits into a certain number of non-degenerate equations. That is, there is a partition $C_1\cup C_2\cup \cdots \cup C_t=C$ such that $\sum_{c\in C_i} cx_c=0$ for ${i=1,\ldots,t}$ and each of these sub-equations is non-degenerate in the sense that it has no zero proper sub-sums. Clearly, ${\#C_i\ge 2}$ for each~$i$. We analyze two sub-cases. \subsubsection{We have $\#C_i\ge 3$ for some ${i\in \{1,\ldots,t\}}$} \label{ssgethree} Then $C_i$ contains two coefficients with the same $\ell$. We assume that ${\ell=1}$ (the case ${\ell=2}$ reduces to ${\ell=1}$ replacing~$\zeta$ by~$\zeta^{-1}$) and let $j_1<j_2$ be the smallest such that $c_{1,j_1}$,~$c_{1,j_2}$ belong to $C_i$. Then the equation is $$ c_{1,j_1}\zeta^{j_1}+c_{1,j_2}\zeta^{j_2}+\sum_{\substack{c_{\ell,j}\kappa^{\delta n}\in C_i\\ \ell=2 ~{\text{\rm or}}~j>j_2}} c_{\ell,j} \kappa^{n\delta} \zeta^{j+n\delta b}=0. $$ Dividing by $\zeta^{j_1}$, we get $$ c_{1,j_1}+c_{1,j_2}\zeta^{j_2-j_1}+\sum_{\substack{c_{\ell,j}\kappa^{\delta n}\in C_i\\ \ell=2 ~{\text{\rm or}}~j>j_2}} c_{\ell,j} \kappa^{n\delta} \zeta^{j-j_1+n\delta b}=0. $$ We are now in the position to apply Mann's theorem to conclude that $$ \zeta^{(j_2-j_1)m_1}=1, \qquad m_1\mid \prod_{p\le \#C_i} p \mid \prod_{p\le 2D+2}p, $$ because ${\#C_i\le 2D+2}$. Since ${|j_2-j_1|\le D}$, we have \begin{equation} \label{emzero} m\le D\prod_{p\le 2D+2}p. \end{equation} The inequality $ {\sum_{p\le x}\log p \le 1.02x} $ holds for all ${x>0}$, see \cite[Theorem~9]{RS62}. Hence $$ \log m\le \log D+\sum_{p\le 2D+2}\log p \le 4D, $$ which is much sharper than what we need. \subsubsection{We have $\#C_i=2$ for all $i=1,\ldots,t$} In fact, we may assume not only that $\#C_i=2$ but also that each $C_i$ contains exactly one $c_{1,j_1}$ and one $c_{2,j_2}\kappa^n$; otherwise the argument from Subsection~\ref{ssgethree} applies, and we again have~\eqref{emzero}. So, let $$ c_{1,j_1}\zeta^{j_1}+c_{2,j_2}\kappa^n\zeta^{j_2+nb}=0. $$ We then get $\zeta^{j_2-j_1+nb}=-c_{1,j_1}/c_{2,j_2} \kappa^{-n}$. The pair $(j_1,j_2)$ depends on $i$. Assume first that, as we loop over~$i$, the differences $j_2-j_1$ are not the same over all $i$; that is, there are two values of $i$ corresponding to say $(j_1,j_2)$ and $(j_1',j_2')$ such that ${j_2'-j_1'\ne j_2-j_1}$. We obtain $$ \zeta^{(j_2-j_1)-(j_2'-j_1')}=\frac{c_{1,j_1}/c_{2,j_2}}{c_{1,j_1'}/c_{2,j_2'}} $$ and the number on the right is a root of unity belonging to~$\Q$. Hence it is $\pm1$. The exponent on the left satisfies $$ 0\ne \bigl|(j_2-j_1)-(j_2'-j_1')\bigr|\le 2D. $$ Hence ${m\le 4D}$, again better than wanted. Now let us assume that ${j_2=j_1+a}$ with the same~$a$ for all $i$. In this case $c_{2,j_1+a}=\lambda c_{1,j_1}$ with the same ${\lambda \in \Q^\times}$ holds for all the~$i$ as well. This makes the rational function $c_2(x)/c_1(x)$ equal to $\lambda x^{a}$, and so $$ u_n(x)=c_1(x)(1+\lambda\kappa^n x^{a+nb}). $$ Since ${u_n(\zeta)=0}$ but ${c_1(\zeta)\ne 0}$, we must have ${1+\lambda\kappa^n \zeta^{a+nb}=0}$, which means that ${\lambda\kappa^n}$ is a root of unity, so~$\pm1$. Now we have two options: either both~$\lambda$ and~$\kappa$ are $\pm1$, or none is. The first option means that condition~\ref{iex} of Theorem~\ref{thm:thmmain} is satisfied, which is against our hypothesis. Hence ${\lambda\kappa^n=\pm1}$, but ${\lambda,\kappa\ne \pm1}$. We have clearly ${{\mathrm{h}}(\kappa)={\mathrm{h}}(f_1,f_2)\le X}$ and ${{\mathrm{h}}(\lambda)={\mathrm{h}}(c_1,c_2)\le X}$. Since~$\kappa$ is a rational number, distinct from~$0$ and from $\pm1$, its numerator or denominator (say, the former) is at least~$2$ in absolute value. It follows that the denominator of ${\lambda=\pm\kappa^{-n}}$ is at least $2^n$ in absolute value. But the denominator of~$\lambda$ cannot exceed ${e^{{\mathrm{h}}(\lambda)}\le e^X}$. We obtain ${2^n\le e^X}$, which implies ${n\le \log X}$. Hence $$ \varphi(m)\le \deg u_n(x) \le D+D\log X, $$ which implies a much sharper estimate for~$m$ than the wanted~\eqref{eupperngam}. Theorem~\ref{thm:thmmain1} is proved. \section{Proof of Theorem~\ref{thsmallestn}} \label{ssmallestn} Let~$\zeta$ be an $m$th primitive root of unity such that the set \begin{equation} \label{eprop} \{n\in \Z_{>0}: \text{$u_n(x)$ is not identically~$0$, but ${u_n(\zeta)=0}$}\} \end{equation} is not empty. If ${c_1(\zeta)f_1(\zeta)=c_2(\zeta)f_2(\zeta)=0}$ then set~\eqref{eprop} consists of all positive integers, and includes~$1$ in particular. If, say, ${c_1(\zeta)f_1(\zeta)\ne 0}$, and set~\eqref{eprop} is non-empty, then $$ c_1(\zeta)f_1(\zeta)c_2(\zeta)f_2(\zeta)\ne0. $$ Denoting $$ \eta=\frac{f_1(\zeta)}{f_2(\zeta)}, \qquad \theta=-\frac{c_2(\zeta)}{c_1(\zeta)}, $$ set~\eqref{eprop} consists of~$n$ with the property ${\eta^n=\theta}$. If~$\eta$ is a root of unity, then its order divides $2m$, and there exists a positive ${n\le 2m}$ such that ${\eta^n=\theta}$. If~$\eta$ is not a root of unity, then ${n={\mathrm{h}}(\theta)/{\mathrm{h}}(\eta)}$. We have ${{\mathrm{h}}(\theta) \le X+\log(D+1)}$ by Lemma~\ref{lhpol}, and ${\varphi(m){\mathrm{h}}(\eta) \ge 2(\log\varphi(m))^{-3}}$, see~\eqref{evout}. Hence $$ n\le m(\log m)^3(X+\log D). $$ Theorem~\ref{thsmallestn} is proved. \subsection*{Acknowledgements} Yu.~B. was partially supported by the Indian Government SPARC Project P445. F. L. was supported in part by Grant NUM2020 from the Wits CoEMaSS. Part of this work was done while F. L. was visiting the Max Planck Institute for Mathematics in Bonn from September 2019 to February 2020. He thanks this institution for its support, hospitality and excellent working conditions. We thank Yann Bugeaud, Philipp Habegger, Alina Ostafe and Igor Shpar\-linski for helpful discussions. We also thank the referee for the encouraging report and many useful suggestions that helped us to improve the presentation. {\footnotesize \bibliographystyle{amsplain}

https://arxiv.org/abs/2010.00277

Jordan Algebras of Symmetric Matrices

We study linear spaces of symmetric matrices whose reciprocal is also a linear space. These are Jordan algebras. We classify such algebras in low dimensions, and we study the associated Jordan loci in the Grassmannian.

\section{Introduction} Let $\mathbb{S}^n$ be the space of symmetric $n \times n$ matrices over the complex numbers $\mathbb{C}$. The $m$-dimensional subspaces $\mathcal{L} \subset \mathbb{S}^n$ are points in the Grassmannian ${\rm Gr}(m,\mathbb{S}^n)$. The terms {\em pencil of quadrics} \cite{FMS} resp.~{\em net of quadrics}~\cite{Wall77} are used when $m=2,3$. We here consider {\em regular} $\mathcal{L}$, i.e. $\mathcal{L}_{\rm inv} = \{ X \in \mathcal{L} \mid {\rm det}(X) \not = 0 \}$ is nonempty. The {\em reciprocal variety} $\mathcal{L}^{-1}$ is the closure in $\mathbb{S}^n$ of the set $\{X^{-1} \mid X \in \mathcal{L}_{\rm inv}\}$. We wish to know under which circumstances the variety $\mathcal{L}^{-1}$ is a linear space in $\mathbb{S}^n$. The motivation for this question arises in optimization, namely in the theory of semidefinite programming \cite{Fay, pablo}, and in statistics, namely from Gaussian models that are linear in covariance matrices and concentration matrices \cite{Jen, STZ}. \smallskip The following theorem furnishes a complete answer to our question.\vspace{-0.2cm} \begin{theorem} \label{thm:characterization} For $\mathcal{L} \in {\rm Gr}(m,\mathbb{S}^n)$ and $U\in\mathcal{L}_{\rm inv}$, the following are equivalent: \vspace{-0.17cm} \begin{itemize} \item[(a)] The reciprocal variety $\mathcal{L}^{-1}$ is also a linear space in $\mathbb{S}^n$. \vspace{-0.17cm} \item[(b)] $\mathcal{L}$ is a subalgebra of the Jordan algebra $(\mathbb{S}^n,\bullet_U)$.\vspace{-0.17cm} \item[(c)] $\mathcal{L}^{-1}$ equals $\mathcal{L}$ up to congruence; namely $\mathcal{L}^{-1}=U^{-1}\mathcal{L} U^{-1}$. \end{itemize} \end{theorem}\vspace{-0.05cm} Theorem \ref{thm:characterization} is due to Jensen \cite{Jen} in the special case when $\mathcal{L}$ contains the identity matrix ${\bf 1}_n$. He then uses $U = {\bf 1}_n$ for the Jordan algebra structure on $\mathbb{S}^n$.\vspace{-0.2cm} \begin{example}[$m=n=4$] Consider the two linear subspaces of $\mathbb{S}^4$ given by \vspace{-0.2cm} $$ \mathcal{L}_1 \,\,=\,\, \begin{small} \begin{pmatrix} x & y & 0 & 0 \\ y & z & 0 & 0 \\ 0 & 0 & w & 0 \\ 0 & 0 & 0 & w \end{pmatrix}\end{small} \quad {\rm and} \quad \mathcal{L}_2 \,\,= \,\, \begin{small} \begin{pmatrix} x & y & 0 & w \\ y & z & -w & 0 \\ 0 & -w & x & y \\ w & 0 & y & z \end{pmatrix} \end{small} \quad\, {\rm with}\,\,\,\, U = {\bf 1}_4. \vspace{-0.2cm} $$ They satisfy (a),(b),(c) in Theorem \ref{thm:characterization}. Note that $\mathcal{L}_2$ appears in \cite[eqn (39)]{Jen}. It illustrates case (ii) in \cite[Theorem 1]{Jen}. The Jordan algebra property fails when $-w$ is replaced by $w$ in $\mathcal{L}_2$. The reciprocal variety is a threefold of degree five. \end{example}\vspace{-0.1cm} The proof of Theorem~\ref{thm:characterization} is relatively short. It will be presented in Section~\ref{sec2}, along with all relevant definitions. We also review classification results for Jordan algebras \cite{Jordan_1934}. This lays the foundation for studying the {\em Jordan locus} ${\rm Jo}(m,\mathbb{S}^n)$. This is the subvariety of the Grassmannian ${\rm Gr}(m,\mathbb{S}^n)$ whose points are the subspaces $\mathcal{L}$ that satisfy the closed conditions (a), (b) and (c) above. To be precise, let $X_1,\ldots,X_m$ represent a basis of $\mathcal{L}$ and write $\adj(U)$ for the adjoint of $U$. From condition (b) it can be seen that $\Jo(m,\mathbb{S}^n)$ is the subvariety set-theoretically defined (in dual Stiefel coordinates) by the linear dependence~of \vspace{-0.24cm} \[ X_1,X_2,\ldots,X_m, \,\,X_i\adj(U)X_j+X_j\adj(U)X_i\vspace{-0.1cm} \] for all matrices $U\in\mathcal{L}$ and all indices $1\leq i\leq j\leq m$. The main goal of this article is the classification of low-dimensional Jordan subalgebras of $\mathbb{S}^n$ and the description of the corresponding Jordan loci for small~$n$. In Section \ref{sec3} we determine the loci of Jordan pencils $(m=2)$ and we study their equations. Section \ref{sec4} establishes a lower bound on the codimension of any proper subalgebra of $(\mathbb{S}^n ,\bullet_U)$, and it identifies the Jordan copencils $(m=4,n=3)$. In Section~\ref{sec5} we classify Jordan subalgebras of dimension $m=3$. This is applied in Section~\ref{sec6} to study the Jordan loci ${\rm Jo}(3,\mathbb{S}^n)$. We emphasize the explicit computation of their irreducible components and defining polynomials. One finding of independent interest is the Chow matrix in Theorem \ref{thm:chow} which represents the Chow form of the determinantal variety $\{ X \in \mathbb{S}^n \mid \rk X \leq n-2 \}$. \section{Jordan algebras} \label{sec2} Fix an invertible matrix $U \in \mathbb{S}^n$. We define an algebra structure on $\mathbb{S}^n$ by setting\vspace{-0.25cm} \begin{equation} \label{eq:bulletdef} X\bullet_UY \,\,\,:= \,\,\, \frac{1}{2}(XU^{-1}Y\,+\,YU^{-1}X) \qquad \hbox{for all $X,Y \in \mathbb{S}^n$}.\vspace{-0.25cm} \end{equation} When $U$ is clear from context, we drop the subscript. The product $\bullet$ is bilinear and commutative, with unit $U$. Associativity fails, but, setting $\,X^{\bullet 2} \,:= \, X \bullet X$, the following weaker version holds. This is the axiom for a (unital) {\em Jordan algebra}:\vspace{-0.15cm} \begin{equation} \label{eq:jordanaxiom} X^{\bullet2} \bullet (X \bullet Y) \quad = \quad X \bullet (X^{\bullet2} \bullet Y) .\vspace{-0.15cm} \end{equation} The role of the unit can be played by any matrix $U \in \mathbb{S}^n_{\rm inv}$. The resulting Jordan algebra structure is denoted by $(\mathbb{S}^n, \,\bullet_U)$. Especially important for applications, notably in statistics, optimization and physics, is the case when $U$ is real and positive definite. Such algebras over $\mathbb{R}$ are known as {\em Euclidean Jordan algebras}. \begin{remark} \label{rmk:Knotclose} Over the field $\mathbb{R}$, the isomorphism type of the Jordan algebra $(\mathbb{S}^n,\,\bullet_U)$ depends on the choice of $U$. To see this, let $n=2$ and write $E_{ij}$ for the matrix units. If $U = {\bf 1}_2 = E_{11} + E_{22}$ then $(\mathbb{S}^2,\,\bullet_U)$ has no nilpotent elements. However, if $\,U' = E_{12} + E_{21}\, $ then the matrices $E_{11}$ and $E_{22}$ are nilpotent in $(\mathbb{S}^2,\,\bullet_{U'})$. \end{remark} We return to this issue later: over $\mathbb{C}$, the isomorphism type of $(\mathbb{S}^n,\,\bullet_U)$ is in fact independent of $U$. First, let us prove the result in the introduction. \begin{proof}[Proof of Theorem \ref{thm:characterization}] Clearly, (c) implies (a). Suppose that (a) holds, so $\mathcal{L}^{-1}$ is a linear space. We shall prove that (b) holds. Consider invertible matrices $U,X\in\mathcal{L}$. For small $t \in \mathbb{R}$, the inverse of $X-tU$ is given by the Neumann series\vspace{-0.25cm} $$ (X-tU)^{-1}\,=\,X^{-1}({\bf 1}_n -tUX^{-1})^{-1}\,=\, X^{-1}\sum_{k=0}^\infty t^k(UX^{-1})^k \,\,\,\,\in\, \,\, \mathcal{L}^{-1}.\vspace{-0.25cm} $$ Since $\mathcal{L}^{-1}$ is linear and closed, the following matrix is in $\mathcal{L}^{-1}$ as well:\vspace{-0.25cm} $$ X^{-1}UX^{-1} \,\, = \,\,\, \lim_{t\to 0} \, \frac{1}{t}\bigl((X-tU)^{-1}-X^{-1}\bigr) \,\,\,\,\in\, \,\, \mathcal{L}^{-1}.\vspace{-0.25cm} $$ Its inverse $X\bullet_{U}X$ is in $\mathcal{L}$. Since invertible $X$ are dense in $\mathcal{L}$, this conclusion holds for all $X\in\mathcal{L}$. Condition (b) states that $X,Y \in \mathcal{L}$ implies $X \bullet_U Y \in \mathcal{L}$. Let $X,Y\in\mathcal{L}$. Then the Jordan algebra squares $ (X+Y)^{\bullet2}, X^{\bullet2}, Y^{\bullet2}$ are in $\mathcal{L}$, and we conclude that $ \,X \bullet_U Y = \frac{1}{2} ((X+Y)^{\bullet2} - X^{\bullet2} - Y^{\bullet2} )\,$ is also in $\mathcal{L}$. So (b) holds. Assume (b) and consider any invertible $X\in\mathcal{L} U^{-1}$. Then $XU\bullet_{U}XU = X^2 U$ is in $\mathcal{L}$. Next, we see that $X^3 U =X^2U\bullet_{U}XU\in \mathcal{L}$. By iterating this, every power $X^k$ lies in $\mathcal{L} U^{-1}$. Since the characteristic polynomial of $X$ has nonzero constant term, the Cayley-Hamilton Theorem implies $X^{-1} \in\mathcal{L} U^{-1}$. Since invertible matrices are dense in $\mathcal{L} U^{-1}$, we get $\mathcal{L} U^{-1} = (\mathcal{L} U^{-1})^{-1} = U\mathcal{L}^{-1}$, i.e.~(c) holds. \end{proof} We next review material from the theory of abstract Jordan algebras. Already in 1934, Jordan, von Neumann and Wigner~\cite{Jordan_1934} established a structure theorem for Euclidean Jordan algebras. We state a variant that holds over $\mathbb{C}$. An {\em ideal} $\mathcal{I}$ in a Jordan algebra $\mathcal{A}$ is a subspace such that $X\bullet Y\in\mathcal{I}$ for all $X \in\mathcal{I}, \,Y \in\mathcal{A}$. The {\em radical} of $\mathcal{A}$ is the ideal $\rad \mathcal{A}:=\{X\in \mathcal{A}\mid X\bullet Y \textit{ is nilpotent for all $Y \in \mathcal{A}$}\}$. \begin{theorem}\label{thm: strucure Jordan Algebras} For a finite-dimensional complex Jordan algebra $\mathcal{A}$, the quotient $\mathcal{A}/\rad \mathcal{A}$ is a direct sum of simple Jordan algebras. Every simple Jordan algebra of dimension $\leq 8$ is isomorphic to either $(\mathbb{C}, \, \cdot \,)$, or to $(\mathbb{S}^3, \bullet_{{\bf 1}_3})$, or to a {\em spin factor} $\mathbb{C} \times \mathbb{C}^d$, $d\geq 2$, with product $(\lambda,a)\bullet (\mu,b)=(\lambda\mu+a^\top b,\lambda b+\mu a)$ and unit $(1,0)$. \end{theorem} \begin{proof} We find this by applying known structure theorems to semi-simple Jordan algebras of dimension $\leq 8$. See \cite[Ch.~X, \S 3.3]{BraunKoecher1966} or \cite[Part~I, \S 2.13]{McCrimmon2004}. \end{proof} Theorem~\ref{thm: strucure Jordan Algebras} will be used to classify low-dimensional Jordan algebras. First, however, we highlight the difference between $\mathbb{R}$ and $\mathbb{C}$ concerning Remark~\ref{rmk:Knotclose}. \begin{lemm}\label{lemma: squareroot} Square roots exist in any complex Jordan subalgebra $\mathcal{L}\subset\mathbb{S}^n$ with unit $U$. To be precise, for every matrix $X\in\mathcal{L}_{\rm inv}$ there exists $Y\in\mathcal{L}$ with $Y^{\bullet2}=X$. \end{lemm} \begin{proof} The subalgebra $\mathcal{B}=\spa\{X^{\bullet\ell}\mid \ell\geq0\}$ of $\mathcal{L}$ is associative as $Y\bullet Z= YU^{-1}Z$ $=ZU^{-1}Y$ for all $Y,Z\in\mathcal B$. By \cite[Part~I, \S 2.13]{McCrimmon2004}, the only associative simple Jordan algebra in $\mathbb{C}$. So by Theorem~\ref{thm: strucure Jordan Algebras}, the quotient $\mathcal{B}/\rad \mathcal{B}$ is a direct sum of copies of $\mathbb{C}$. In particular, every element of $\mathcal{B}/\rad \mathcal{B}$ is a square. So there exist $Z \in \mathcal{B}$ and $W\!\in \rad \mathcal{B}$ with $Z^{\bullet2}+W=X$. We have $X^{-1}\in U^{-1}\mathcal{B}U^{-1}$ and so, since $W\in\rad\mathcal{B}$, the matrix $WX^{-1}$ is nilpotent. Hence the matrix $X-W = Z^{\bullet2}$ is invertible and using Theorem~\ref{thm:characterization} we can take $Z^{\bullet-2}:= U\bullet_{Z^{\bullet 2}}U\in\mathcal{B}$ which satisfies $Z^{\bullet2}\bullet Z^{\bullet -2}=U$. Since $W\in\rad \mathcal{B}$, the element $Z^{\bullet-2}\bullet W$ is nilpotent, say $(Z^{\bullet-2}\bullet W)^{\bullet p}=0$. Using the Taylor series $\sqrt{1+\lambda}= \sum_{\ell=0}^\infty a_\ell \lambda^\ell $, we~see that\vspace{-0.35cm} $$ \begin{matrix} X \,\,\,=\,\,\,Z^{\bullet2}\bullet\left(U+Z^{\bullet-2}\bullet W\right)\,\,\,=\,\,\, Z^{\bullet2}\bullet \left(\,\sum_{\ell=0}^{p-1} a_\ell \left(Z^{\bullet-2}\bullet W\right)^{\bullet\ell}\right)^{\bullet2} . \end{matrix}\vspace{-0.25cm} $$ Hence $X$ has a square root $Y$ in the subalgebra $\mathcal{B}$ of $\mathcal{L}$. \end{proof} Having access to a square root within the algebra, we can find isomorphisms between $(\mathcal{L},\bullet_U)$ and $(\mathcal{L},\bullet_V)$ for different choices of matrices $U,V\in\mathcal{L}_{\rm inv}$. \begin{proposition} \label{prop:compU} Consider a complex Jordan algebra $(\mathcal{L},\bullet_U)$ as in Theorem~\ref{thm:characterization}. There is an isomorphism $(\mathcal{L},\bullet_U)\cong(\mathcal{L},\bullet_V)$ of Jordan algebras for all $V \in \mathcal{L}_{\rm inv}$. \end{proposition} \begin{proof} By Lemma \ref{lemma: squareroot}, there exists $W\in\mathcal{L}_{\rm inv}$ with $WU^{-1}W=V$. Consider the~map\vspace{-0.15cm} \[ \phi : \mathcal{L} \rightarrow \mathcal{L} \,, \,\,\, X\,\,\mapsto \,\,VW^{-1}XW^{-1}V\,\,=\,\,2V\bullet_W(X\bullet_WV)-(V\bullet_WV)\bullet_W X. \vspace{-0.15cm} \] This linear automorphism of $\mathcal{L}$ is invertible since $VW^{-1}$ is invertible. We have\vspace{-0.15cm} \[ \phi(XU^{-1}X) \,=\,\phi(X W^{-1} V W^{-1} X) \,=\,VW^{-1}XW^{-1}VW^{-1}XW^{-1}V \,=\,\phi(X)V^{-1}\phi(X). \vspace{-0.15cm} \] This shows that $\phi$ is a Jordan algebra isomorphism from $(\mathcal{L},\bullet_U)$ to $(\mathcal{L},\bullet_V)$. \end{proof} We now come to the classification of low-dimensional Jordan algebras. \begin{lemm} \label{lem:k+1} If $k=\dim(\rad\mathcal{A})$, then $X^{\bullet k+1}=0$ for all $X\in\rad\mathcal{A}$. \end{lemm} \begin{proof} Let $X\in\rad\mathcal{A}$. As $X,X^{\bullet2},\ldots,X^{\bullet k+1}\in\rad\mathcal{A}$ are linearly dependent, the minimum polynomial of $X$ has degree\,$\leq k+1$. As $X$ is nilpotent, $X^{\bullet k+1}=0$. \end{proof} \begin{proposition}\label{thm:2d_classification} Every Jordan algebra $\mathcal A$ of dimension two over $\mathbb{C}$ is isomorphic to the Jordan algebra $\mathbb{C}\{U,X\}$ with unit $U$, where the product is given by $ \begin{array}{llllll} 1\colon\!\!\!\!&X^{\bullet2} =X,&\mbox{\!\!when }\mathcal{A}\cong\mathbb{C}\times\mathbb{C}; \mbox{ or}\\ 2\colon\!\!\!\!&X^{\bullet2} =0,&\mbox{\!\!when }\mathcal{A}/\rad\mathcal{A}\cong\mathbb{C}. \end{array} $ \end{proposition} \begin{proof} By Theorem~\ref{thm: strucure Jordan Algebras}, either $\mathcal{A}\cong\mathbb{C}\times\mathbb{C}$ or $\mathcal{A}/\rad\mathcal{A}\cong\mathbb{C}$. We seek $X$ such that $\{U,X\}$ is a basis of $\mathcal{A}$. In the first case, take the idempotent $X=(1,0)$ in $\mathbb{C}\times\mathbb{C}$. In the second case, we choose $X$ to span $\rad\mathcal{A}$. Then $X^{\bullet 2}=0$ by Lemma \ref{lem:k+1}. \end{proof} \begin{theorem}\label{thm:3d_classification} Every Jordan algebra $\mathcal A$ of dimension three over $\mathbb{C}$ is isomorphic to the Jordan algebra $\mathbb{C}\{U,X,Y\}$ with unit $U$, where the product is given~by $ \begin{array}{llllll} 1\!\!\!\!&{\rm (a)}\colon\!\!\!\!&X^{\bullet2} =X,&\!\!\!\! Y^{\bullet2}=Y& \mbox{\!\!\!\!and } X\bullet Y=0,&\mbox{\!\!when }\mathcal{A}\cong\mathbb{C}\times\mathbb{C}\times\mathbb{C}; \\ &{\rm (b)}\colon\!\!\!\!&X^{\bullet2} =U,&\!\!\!\! Y^{\bullet2}=U& \mbox{\!\!\!\!and } X\bullet Y=0,&\mbox{\!\!when }\mathcal{A}\cong\mathbb{C}\times\mathbb{C}^2; \\ 2\!\!\!\!&{\rm (a)}\colon\!\!\!\!&X^{\bullet2} =X,&\!\!\!\! Y^{\bullet2}=0& \mbox{\!\!\!\!and } X\bullet Y=0\mbox{ or}&\\ &{\rm (b)}\colon\!\!\!\!&X^{\bullet2} =X,&\!\!\!\! Y^{\bullet2}=0& \mbox{\!\!\!\!and } X\bullet Y=Y/2,&\mbox{\!\!when }\mathcal{A}/\rad\mathcal{A}\cong\mathbb{C}\times\mathbb{C}; \mbox{ or}\\ 3\!\!\!\!&{\rm (a)}\colon\!\!\!\!&X^{\bullet2} =Y,&\!\!\!\! Y^{\bullet2}=0& \mbox{\!\!\!\!and } X\bullet Y=0\mbox{ or} \\ &{\rm (b)}\colon\!\!\!\!&X^{\bullet2} =0,&\!\!\!\! Y^{\bullet2}=0& \mbox{\!\!\!\!and } X\bullet Y=0,&\mbox{\!\!when }\mathcal{A}/\rad\mathcal{A}\cong\mathbb{C}. \end{array} $ \end{theorem} \begin{proof} By Theorem~\ref{thm: strucure Jordan Algebras}, there are four possible cases to consider: \vspace{-0.25cm} \[ \mathcal{A}\,\cong\,\mathbb{C}\times\mathbb{C}\times\mathbb{C},~ \mathcal{A}\,\cong\,\mathbb{C}\times\mathbb{C}^2, ~\mathcal{A}/\rad\mathcal{A}\,\cong\,\mathbb{C}\times\mathbb{C} \,\, \mbox{ or } \,\,\mathcal{A}/\rad\mathcal{A}\,\cong\,\mathbb{C}.\vspace{-0.25cm} \] We seek $X,Y$ such that $\{U,X,Y\}$ is a basis of $\mathcal{A}$. In the first case, the choices $X=(1,0,0)$ and $Y=(0,1,0)$ in $\mathbb{C}\times\mathbb{C}\times\mathbb{C}$ have the desired properties. In the second case, we can choose $X,Y$ to be elements of order two with product zero. Two such elements in $\mathbb{C}\times\mathbb{C}^2$ are $(0,e_1),(0,e_2)$. So, case 1(a) or case 1(b) holds. Now suppose $\mathcal{A}/\rad\mathcal{A}\cong\mathbb{C}\times\mathbb{C}$. Let $X \in \mathcal{A}$ be a preimage of $(1,0)$, so $X^{\bullet2}\equiv X\!\!\mod\rad\mathcal{A}$. Choose $Y$ to span $\rad\mathcal{A}$. Then $Y^{\bullet2}=0$ and $X\bullet Y\in\rad\mathcal{A}$. Here we use that $\rad\mathcal{A}$ is an ideal. So $X^{\bullet2}=X+\lambda Y$ and $X\bullet Y=\mu Y$ for some $\lambda,\mu\in\mathbb{C}$. We determine all $\lambda,\mu \in \mathbb{C}$ such that the commutative bilinear form $\bullet$ satisfies \vspace{-0.20cm} $$ Z^{\bullet2}\bullet(Z\bullet W)=Z\bullet(Z^{\bullet2}\bullet W)\vspace{-0.10cm} $$ with $Z=c_1U+c_2X+c_3Y$ and $W=d_1U+d_2X+d_3Y$ for all $c_1,c_2,c_3,d_1,d_2,d_3\in\mathbb{C}$. Using computer algebra, we find that either $\mu=0$, $\mu=1$ or $(\lambda,\mu)=(0,1/2)$. The first two cases are isomorphic via the base change that replaces $X$ by $U-X$ or $X\pm\lambda Y$. From this we can conclude that case 2(a) holds or case 2(b) holds. If $\mathcal{A}/\rad\mathcal{A}\cong\mathbb{C}$, we choose $X,Y$ to span $\rad\mathcal{A}$. In this case, either the square of every element in $\rad\mathcal{A}$ is $0$, or we can choose $X,Y$ such that $X^{\bullet2}=Y$, since $X$ is nilpotent. In the first case, we find that $X^{\bullet2}=Y^{\bullet2}=X\bullet Y=0$. In the second case, we find that $Y^{\bullet2}=X\bullet Y=0$ as $X^{\bullet 3}=0$. So case 3(a) or case 3(b) holds. \end{proof} \section{Jordan Pencils} \label{sec3} We now embark on the study of the Jordan locus ${\rm Jo}(m,\mathbb{S}^n)$ in ${\rm Gr}(m,\mathbb{S}^n)$. It is convenient to work in the sub-Grassmannian ${\rm Gr}_{\bf 1}(m,\mathbb{S}^n)$ of subspaces $\mathcal{L}$ that contain the identity matrix ${\bf 1}_n$. The {\em restricted Jordan locus} is the intersection \begin{equation} \label{eq:restrictedJordan} {\rm Jo}_{\bf 1}(m,\mathbb{S}^n) \,\,\, =\,\,\, {\rm Gr}_{\bf 1}(m,\mathbb{S}^n) \, \cap \, {\rm Jo}(m,\mathbb{S}^n). \end{equation} We next examine the variety (\ref{eq:restrictedJordan}) for $m=2$. Pencils in ${\rm Gr}_{\bf 1}(2,\mathbb{S}^n)$ have the form $\mathcal{L} = \mathbb{C} \{ {\bf 1}_n, X \}$, where ${\rm trace}(X) = 0$. This identifies ${\rm Gr}_{\bf 1}(2,\mathbb{S}^n)$ with the projective space $ \mathbb{P}^{\binom{n+1}{2}-2}$ of traceless symmetric matrices. With these conventions, $\mathcal{L}$ is a Jordan algebra if and only if the minimal polynomial of $X$ has degree two. \begin{proposition} The restricted Jordan locus ${\rm Jo}_{\bf 1}(2,\mathbb{S}^n)$ is a variety with $\lfloor n/2 \rfloor$ irreducible components $V_i$ for $1 \leq i \leq n/2$. The component $V_i$ has codimension $\binom{i+1}{2} + \binom{n-i+1}{2} - 2$. It is parametrized by diagonalizable matrices that have two distinct eigenvalues of multiplicities $i$ and $n-i$. This is case 1 in Proposition \ref{thm:2d_classification}. \end{proposition} \begin{proof} Following \cite{FMS}, regular pencils are classified by their Segre symbols $\sigma$. By \cite[Theorem 3.2]{FMS}, a pencil $\mathcal{L}$ has ${\rm deg}(\mathcal{L}^{-1}) = 1$ if and only if its Segre symbol is $\sigma = [(1,\ldots, 1), (1 ,\ldots, 1)]\,$ or $\,\sigma = [(2,\ldots, 2 ,1 ,\ldots, 1)]$. The varieties $V_i$ are the closures of the former strata. These contain the latter in their closure, by \cite[Theorem 5.1]{FMS}. The formula for ${\rm codim}(V_i)$ appears in \cite[Proposition~5.4]{FMS}. \end{proof} \begin{remark} Nondiagonalizable pencils in ${\rm Jo}_{\bf 1}(2,\mathbb{S}^n)$ have only one eigenvalue and Jordan blocks of sizes one or two. This is case 2 in Proposition \ref{thm:2d_classification}. \end{remark} Equations defining $V_i$ are computed as follows. Fix a symmetric $n \times n$ matrix of unknowns, $X = (x_{ij}) $. Consider the entries of the product $(X-a {\bf 1}_n)(X-b {\bf 1}_n)$, the $(i+1)$-minors of $X-a {\bf 1}_n$, and the $(n-i+1)$-minors of $X-b {\bf 1}_n$. By eliminating $a$ and $b$ from these equations, we obtain homogeneous polynomials in the entries of $X$ that cut out $V_i$ set-theoretically. We discuss these varieties for $n \leq 5$. All pencils of binary quadrics $(n=2)$ are Jordan: ${\rm Jo}_{\bf 1}(2,\mathbb{S}^2) = V_1 ={\rm Gr}_{\bf 1}(2,\mathbb{S}^2)$. \bigskip We thus start with Jordan pencils of plane conics ($n=3$). \begin{example}[$m=2,n=3$] The variety ${\rm Jo}_{\bf 1}(2,\mathbb{S}^3) = V_1$ is irreducible of codimension $2$ and degree $4$. Its ideal is generated by the following seven cubics: \vspace{-0.17cm} $$ \begin{small} \begin{matrix} x_{11} x_{13} x_{22}-x_{11} x_{13} x_{33}-x_{12}^2 x_{13}+x_{12} x_{22} x_{23} -x_{12} x_{23} x_{33}+x_{13}^3-x_{13} x_{22}^2+x_{13} x_{22} x_{33}, \\ x_{11} x_{12} x_{22}-x_{11} x_{12} x_{33}-x_{12}^3+x_{12} x_{13}^2 -x_{12} x_{22} x_{33}+x_{12} x_{33}^2+x_{13} x_{22} x_{23}-x_{13} x_{23} x_{33}, \\ x_{11}^2 x_{23}-x_{11} x_{12} x_{13}-x_{11} x_{22} x_{23}-x_{11} x_{23} x_{33} +x_{12} x_{13} x_{22}+x_{13}^2 x_{23}+x_{22} x_{23} x_{33}-x_{23}^3, \\ x_{12}^2 x_{23}-x_{12} x_{13} x_{22}+x_{12} x_{13} x_{33}-x_{13}^2 x_{23}, \quad x_{11} x_{13} x_{23}-x_{12} x_{13}^2+x_{12} x_{23}^2-x_{13} x_{22} x_{23}, \\ \!\!\!\! x_{11} x_{12} x_{23}-x_{12}^2 x_{13}-x_{12} x_{23} x_{33}+x_{13} x_{23}^2\,, \quad x_{11}^2 x_{22}-x_{11}^2 x_{33}-x_{11} x_{12}^2 +x_{11} x_{13}^2 \quad \\ \qquad -x_{11} x_{22}^2 +x_{11} x_{33}^2+x_{12}^2 x_{22}-x_{13}^2 x_{33}+x_{22}^2 x_{33} -x_{22} x_{23}^2-x_{22} x_{33}^2+x_{23}^2 x_{33}. \end{matrix} \end{small} \vspace{-0.13cm} $$ These are the expressions in the sum of squares representation of the discriminant of the characteristic polynomial of $X$, seen in \cite[page 97]{CBMS}. Indeed, $V_1$ is also the Zariski closure of all \underbar{real} matrices in $\mathbb{S}^3$ with a double eigenvalue. \end{example} \begin{example}[$m=2,n=4$] We compute ${\rm Jo}_{\bf 1}(2,\mathbb{S}^4) $ in {\tt Macaulay2} as follows: \vspace{-0.13cm} \begin{small} \begin{verbatim} R = QQ[a,b,x11,x12,x13,x14,x22,x23,x24,x33,x34,x44]; X = matrix {{x11,x12,x13,x14},{x12,x22,x23,x24}, {x13,x23,x33,x34},{x14,x24,x34,x44}}; I = eliminate({a,b},minors(1,(X-a)*(X-b))) \end{verbatim} \end{small} \vspace{-0.13cm} The ideal ${\tt I}$ is generated by $30$ cubics. It is the intersection of two prime ideals, corresponding to ${\rm Jo}_{\bf 1}(2,\mathbb{S}^4) = V_1 \cup V_2$. The component $V_1$ has codimension $5$ and degree $8$. Its prime ideal is generated by $10$ quadrics, including $\,x_{13} x_{24}-x_{12} x_{34},\, x_{14} x_{23}-x_{12} x_{34}$ and $x_{13} x_{23}-x_{14} x_{24}-x_{12} x_{33}+x_{12} x_{44}$. The component $V_2$ has codimension $4$ and degree $6$. Its prime ideal is generated by $9$ quadrics. \end{example} \begin{example}[$m=2,n=5$] The ideal of ${\rm Jo}_{\bf 1}(2,\mathbb{S}^5) = V_1 \cup V_2$ is generated by $81$ cubics. The variety $V_1$ has codimension $9$ and degree $16$. Its ideal is generated by $35$ quadrics, including ten $2 \times 2$-minors, like $x_{12} x_{34}-x_{13} x_{24}$. The variety $V_2$ has codimension $7$ and degree $40$. Its ideal is generated by $95$ cubics, e.g.~$ \begin{small} x_{12} x_{13} x_{35} - x_{13} x_{15} x_{23} + x_{22} x_{23} x_{35} - x_{23}^2 x_{25} - x_{23} x_{34} x_{45} - x_{23} x_{35} x_{55} + x_{24} x_{34} x_{35} + x_{25} x_{35}^2. \end{small} $ \end{example} We now observe that ${\rm Jo}(m,\mathbb{S}^n)$ is the orbit of (\ref{eq:restrictedJordan}) under the congruence action by ${\rm GL}(n)$. This implies the following result in the Grassmannian ${\rm Gr}(2,\mathbb{S}^n)$. \begin{corollary} \label{cor:n2floor} The Jordan locus ${\rm Jo}(2,\mathbb{S}^n)$ has $\lfloor n/2\rfloor$ irreducible components. Using the notation from \cite[Section 5]{FMS}, these components are the Grassmann strata $\overline{{\rm Gr}_\sigma}$ that are associated with the Segre symbols $\sigma = [(1,\ldots, 1), (1 ,\ldots, 1)]$. \end{corollary} \begin{remark} Jordan algebras can be viewed as a nonabelian generalization of {\em partition matroids}. These are the matroids that are direct sums of uniform matroids. Indeed, the diagonalizable types in Corollary \ref{cor:n2floor} correspond to the rank-$2$ partition matroids, and those in Theorem \ref{thm:classification m=3,n=4} to the rank-$3$ partition matroids. \end{remark} \section{Jordan Copencils} \label{sec4} This section concerns Jordan algebras of low codimension. These are quite rare: \begin{theorem} \label{Thm: codimension} Let $\mathcal{L}\subset\mathbb{S}^n$ be a proper Jordan subalgebra. Then $\codim(\mathcal{L})\geq n-1$. \end{theorem} To prove this theorem, we need the following result and some lemmas.\vspace{-0.08cm} \begin{theorem}[Peirce decomposition, {\cite[Part II, Chapter 8]{McCrimmon2004}}] \label{thm:peirce} Let $X_1,\dots X_d$ be orthogonal idempotents in $\mathcal{L}$ with $U=X_1+\dots+X_d$. Then $\mathcal{L}=\bigoplus_{1\leq i\leq j\leq d} \mathcal{X}_{i,j}$ where $\,\mathcal X_{i,i}=\{\,Y\in \mathcal{L}\,\mid \,Y= X_i\bullet Y\}\,$ and $\,\mathcal X_{i,j}=\{\,Y\in \mathcal{L}\,\mid \, Y = 2 X_i\bullet Y= 2 X_j\bullet Y\}\,$ for $\,i<j$. \end{theorem} Here, being {\em orthogonal} means that $X_i \bullet X_j =0$ whenever $i \not= j$. An idempotent $X$ is called {\em primitive} if there are no nonzero orthogonal idempotents $X_1$ and $ X_2$ with $X=X_1+X_2$. We denote by $J_n$ the $n\times n$ matrix with ones on its anti-diagonal. \begin{lemm}[{\cite[Lemma 1]{bukovsek-omladic}}]\label{lemma:conjugation0} If $X,Y\in\mathbb{S}^n$ are similar, then they are orthogonally congruent. \end{lemm} \begin{lemm}\label{lemm:antitriagle} If the unit $U$ of a Jordan subalgebra $\mathcal{L} \subset \mathbb{S}^n$ is primitive, then $\mathcal{L}$ is congruent to $\,\mathbb{C} J_n\oplus \mathcal{L}'$ where $\mathcal{L}'$ consists of upper anti-triangular matrices. \end{lemm} \begin{proof} After congruence, we may assume $U = {\bf 1}_n$. Let $X\in\mathcal{L}$. By considering the Jordan canonical form of $X$, adding multiples of ${\bf 1}_n$ to $X$, and taking powers, we see that ${\bf 1}_n$ cannot be primitive if $X$ has distinct eigenvalues. So $\mathcal{L}=\mathbb{C} {\bf 1}_n\oplus \mathcal{L}'$ where $\mathcal{L}'\subset\mathcal{L}$ is the subset of nilpotent matrices. By Lemma~\ref{lemma:conjugation0}, we see that $XY+YX\in\mathcal{L}'$ for $X,Y\in\mathcal{L}'$. So $\mathcal{L}'$ is anti-triangularizable by \cite[Theorem 9]{bukovsek-omladic}. The transformation given there takes the unit matrix ${\bf 1}_n$ to the matrix $J_n$. \end{proof} Two spaces $\mathcal{L},\mathcal{L}'\subset\mathbb{S}^n$ are congruent if $\mathcal{L}'=P\mathcal{L} P^\top$ for some $P\in\GL(n)$. We say that $\mathcal{L},\mathcal{L}'$ are {\em orthogonally congruent} if $P$ lies in the orthogonal group $O(n)$.\vspace{-0.08cm} \begin{lemm}\label{lemma:conjugation} If $\mathcal{L},\mathcal{L}'\subset\mathbb{S}^n$ are similar, then they are orthogonally congruent. \end{lemm} \begin{proof} We have $\mathcal{L} P=P\mathcal{L}'$ for some $P\in\GL(n)$. Consider $A\in\mathcal{L}$ and $B\in\mathcal{L}'$ with $AP=PB$. By the proof of \cite[Lemma 1]{bukovsek-omladic}, we get $AU=UB$ for $U=(PP^\top)^{-1/2}P$. The matrix $U$ is orthogonal, and it maps $\mathcal{L}$ into $\mathcal{L'}$ under congruence. \end{proof} \begin{lemm}\label{lemma:idempotent} Let $X\in\mathbb{S}^n$ be an idempotent matrix of rank $r$. Then $X$ is orthogonally congruent to $\Diag({\bf 1}_r,{\bf 0}_{n-r})$. If $X=\Diag({\bf 1}_r,{\bf 0}_{n-r})$ and $Y\in\mathbb{S}^n$ is a matrix with $XY+YX=0$, then $Y=\Diag({\bf 0}_r,Z)$ for some $Z\in\mathbb{S}^{n-r}$. \end{lemm} \begin{proof} This follows from Lemma~\ref{lemma:conjugation0}. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm: codimension}] Either $\mathcal{L}$ contains a primitive idempotent of rank $>1$, or the unit $U$ of $\mathcal{L}$ is a sum of orthogonal rank-$1$ idempotents. We shall examine the associated Peirce decompositions (Theorem \ref{thm:peirce}). These have $d=2$ and $d=n$. Suppose that $\mathcal{L}$ contains a primitive idempotent $X_1$ of rank $r>1$ and take $X_2=U-X_1$. Then $\mathcal{L}=\mathcal{X}_{1,1}\oplus\mathcal{X}_{1,2}\oplus\mathcal{X}_{2,2}$. By Lemmas \ref{lemm:antitriagle} and \ref{lemma:idempotent}, we assume that $X_1=\Diag(J_r,{\bf 0}_{n-r})$, $X_2=\Diag({\bf 0}_r,{\bf 1}_{n-r})$ and $\mathcal{X}_{1,1}$ consists of upper anti-triangular $r \times r$ matrices. So $\mathcal{X}_{1,1}$ has codimension $\geq r-1$ in $\mathbb{S}^r$. Elements in $\mathcal X_{1,2}$ are of the form $Z={\footnotesize \begin{pmatrix} {\bf 0}_r & V^\top\\ V & {\bf 0}_{n-r} \end{pmatrix} }$ with $Z^{\bullet 2}=\Diag(V^\top V,VJ_rV^\top)$. It follows that $V^\top V\in\mathcal{X}_{1,1}$. So, the last column $v$ of $V$ satisfies $v^\top v=0$. Hence $v$ is contained in a subspace $\mathcal Y\subset\mathbb{C}^{n-r}$ of codimension $c\geq (n-r)/2$. We have $\mathcal{X}_{2,2}=\Diag({\bf 0}_r,\mathcal{Z})$ where $\mathcal{Z}\subset\mathbb{S}^{n-r}$ consists of matrices that map $\mathcal Y$ into itself. Therefore $\mathcal{Z}$ has codimension $\geq c(n-r-c)$. So $\mathcal{L}$ has codimension $\geq r-1+c+c(n-r-c)\geq n-1$. If all primitive idempotents of $\mathcal{L}$ have rank $1$, then $U=X_1+\dots+X_n$ for rank-$1$ orthogonal idempotents $X_1,\ldots,X_n$. By Lemma~\ref{lemma:idempotent}, we may assume $X_i=E_{ii}$. We note that the spaces of the Peirce decomposition are either $\mathcal{X}_{i,j} =\mathbb{C}\{E_{i,j}+E_{j,i}\}$ or $\mathcal{X}_{i,j}=\{0\}$. Since $\mathcal{L}$ is a proper subalgebra, at least one of these spaces is $\{0\}$, say $\mathcal X_{1,2}=\{0\}$. Since $2(E_{i,j}+E_{j,i})\bullet (E_{i,k}+E_{k,i})=E_{k,j}+E_{j,k}$, this implies either $\mathcal X_{1,i}=\{0\}$ or $\mathcal X_{2,i}=\{0\}$ for each $i \in \{3,\dots, n\}$. Therefore $\codim (\mathcal{L})\geq n-1$. \end{proof} A linear subspace $\mathcal{L}\subset\mathbb{S}^n$ is said to be a {\em copencil} if $\codim (\mathcal{L}) = 2$, i.e.~the orthogonal complement $ \mathcal{L}^\perp = \{\, X \in \mathbb{S}^n \mid {\rm trace}(XZ) = 0 \,\,\hbox{for} \,\, Z \in \mathcal{L} \} $ is a pencil. We say that $\mathcal{L}$ is a {\em Jordan copencil} if $\mathcal{L}$ is a copencil that satisfies the equivalent conditions (a), (b) and (c) from Theorem \ref{thm:characterization}. Theorem \ref{Thm: codimension} implies: \begin{corollary} There are no Jordan copencils unless $n \leq 3$. Every copencil with $n=2$ is a Jordan copencil since it is the span of one invertible matrix $U $ in $\mathbb{S}^2$. \end{corollary} It remains to study $n=3$. We represent Jordan copencils by the variety \begin{equation} \label{eq:copencils3} \left\{\, (X,Y) \in \mathbb{S}^3 \times \mathbb{S}^3 \,\,\middle|\,\, (\mathbb{C} \{X,Y\} )^\perp \,\,\hbox{is a point in} \,\, \,{\rm Jo}(4,\mathbb{S}^3) \, \right\} . \end{equation} \vspace{5pt} \begin{proposition}\label{thm:classification m=4,n=3} There are two congruence orbits of Jordan copencils in $\mathbb{S}^3$:\vspace{-0.15cm} $$ \mathcal{L}_1 \,\,=\,\, \begin{small} \begin{pmatrix} \, x & 0 & 0 \, \\ \,0 &y&w \, \\ \, 0 &w&z\, \end{pmatrix} \end{small} \quad \mbox{ and } \quad \mathcal{L}_2\, \,= \,\, \begin{small} \begin{pmatrix} \, x&y&w \, \\ \, y&z & 0 \,\\ \, w & 0 & 0 \, \end{pmatrix}. \end{small}\vspace{-0.15cm} $$ The orbit of $\mathcal{L}_2$ is in the closure of the orbit of $\mathcal{L}_1$. These orbits have codimensions $5$ and $4$, so the Jordan locus ${\rm Jo}(4, \mathbb{S}^3)$ is irreducible of codimension~$4$. The prime ideal of (\ref{eq:copencils3}) has degree $21$ and is generated by $4$ cubics and $15$ quartics. \end{proposition} \begin{proof} Theorem~\ref{thm: strucure Jordan Algebras} gives a finite list of cases to examine. By multiple applications of Lemma~\ref{lemma:idempotent}, we see that $\mathbb{C}\times\mathbb{C}\times\mathbb{C}\times\mathbb{C}$ cannot be embedded into~$\mathbb{S}^3$. By Theorem~\ref{thm:classification m=3,n=n}, $\mathbb{C}\times\mathbb{C}^2$ cannot be embedded into $\mathbb{S}^3$. And, a computation shows that the Jordan algebra $\mathbb{C}\{X,Y\}$ with zero bilinear form $\bullet$ cannot be embedded into $\mathbb{S}^3$. Using the help of a computer as in the proof of Theorem~\ref{thm:3d_classification}, we generate a finite list of Jordan algebras that do not contain any of these Jordan algebras as a subalgebra. Using Lemma~\ref{lemma:idempotent}, we find that all embeddings of these Jordan algebras are congruent to $\mathcal{L}_1$ or $\mathcal{L}_2$. The last two statements are proved by computer algebra. Note that $\mathcal{L}_1^\perp $ and $\mathcal{L}_2^\perp$ are singular pencils. The four cubics in the ideal for $\mathcal{L}_1$ are the four coefficients of ${\rm det}(x X + y Y)$. The $15$ quartics can also be derived using linear algebra. For the containment of orbits, we check that $\mathcal{L}_2$ satisfies these $19$ equations. \end{proof} \section{Three-Dimensional Jordan Algebras} \label{sec5} The conditions in Theorem~\ref{thm:characterization} are invariant under congruence. Ideally, one would classify Jordan subalgebras of $\mathbb{S}^n$ up to congruence and describe the poset of orbit closures. We tackle this problem for $m=3$. The following two theorems are our main results in Section \ref{sec5}. The first gives a complete solution for $4 \times 4$ matrices. The second gives a solution for Jordan subalgebras without a radical. \begin{theorem}\label{thm:classification m=3,n=4} Every Jordan net in $\mathbb{S}^4$ is congruent to one of the following nets: \noindent$ \begin{array}{llllll} 1\!\!\!\!&{\rm (a)}\!\!\!\!\!\!\!\!\!\!&\colon\!\!\!\!&\Diag(x{\bf 1}_2,y,z);\\ &{\rm (b)}\!\!\!\!\!\!\!\!\!\!&\colon\!\!\!\!&\mbox{the Kronecker product of ${\bf 1}_2$ with $\mathbb{S}^2$; \qquad (see $\mathcal{L}_2$ in Example \ref{ex:drei})} \\ 2\!\!\!\!&{\rm (a1)}\!\!\!\!\!\!\!\!\!\!&\colon\!\!\!\!&\Diag(xJ_2+yE_{11},z{\bf 1}_2);\\ &{\rm (a2)}\!\!\!\!\!\!\!\!\!\!&\colon\!\!\!\!&\Diag(xJ_3+yE_{11},z);\\ &{\rm (b)}\!\!\!\!\!\!\!\!\!\!&\colon\!\!\!\!&\Diag(xJ_2,yJ_2)+z(E_{13}+E_{31});\\ 3\!\!\!\!&{\rm (a)}\!\!\!\!\!\!\!\!\!\!&\colon\!\!\!\!&x\Diag(J_3,1)+y(E_{12}+E_{21})+zE_{11};\\ &{\rm (b1)}\!\!\!\!\!\!\!\!\!\!&\colon\!\!\!\!&xJ_4+yE_{11}+zE_{22};\mbox{ or}\\ &{\rm (b2)}\!\!\!\!\!\!\!\!\!\!&\colon\!\!\!\!&xJ_4+yE_{11}+z(E_{13}+E_{31}). \end{array} $ \smallskip \noindent The numberings of the congruence classes are the same as in Theorem \ref{thm:3d_classification}. Their orbit closures are subvarieties in the Grassmannian ${\rm Gr}(3,\mathbb{S}^4)$ of dimension~$21$. \begin{samepage} The codimensions of these Jordan loci and their containments are as follows: \vspace*{-0.07cm} \begin{center} \begin{tikzpicture} \node () at (-3,2.4) {$\codim$}; \node () at (-3,1.6) {$\codim$}; \node () at (-3,.8) {$\codim$}; \node () at (-3,0) {$\codim$}; \node () at (-2,2.4) {$9$}; \node () at (-2,1.6) {$10$}; \node () at (-2,.8) {$11$}; \node () at (-2,0) {$12$}; \node (1a) at (1,2.4) {1(a)}; \node (1b) at (4,1.6) {1(b)}; \node (2a1) at (0,1.6) {2(a1)}; \node (2a2) at (2,1.6) {2(a2)}; \node (2b) at (5.,.8) {2(b)}; \node (3a) at (1.,.8) {3(a)}; \node (3b1) at (3.,.8) {3(b1)}; \node (3b2) at (3.,0) {3(b2)}; \draw [thick] (1a) -- (2a1); \draw [thick] (1a) -- (2a2); \draw [thick] (1b) -- (2b); \draw [thick] (1b) -- (3b1); \draw [thick] (2a1) -- (3a); \draw [thick] (2a2) -- (3b1); \draw [thick] (2b) -- (3b2); \draw [thick] (2a2) -- (3a); \draw [thick] (3a) -- (3b2); \draw [thick] (3b1) -- (3b2); \end{tikzpicture} \end{center} \end{samepage} \end{theorem} \begin{corollary} \label{cor:Jo34} The Jordan locus ${\rm Jo}(3,\mathbb{S}^4)$ has two irreducible components. \end{corollary} Jordan algebras of types 1(a) and 1(b) are of primary interest also for $n \geq 5$. \begin{theorem}\label{thm:classification m=3,n=n} Every Jordan net in $\mathbb{S}^n$ of type 1(a) is congruent to a diagonal~net $ \Diag(x{\bf 1}_{k_1},y{\bf 1}_{k_2},z{\bf 1}_{k_3})$ for $k_1\geq k_2\geq k_3\geq1$ with sum $n$. Jordan nets of type 1(b)~exist only for even $n$. They are congruent to the Kronecker product of ${\bf 1}_{n/2}$ with~$\mathbb{S}^2$. \end{theorem} We note that Theorem \ref{thm:classification m=3,n=n} does not account for all irreducible components of ${\rm Jo}(3,\mathbb{S}^n)$ when $n \geq 5$. At present we do not know these decompositions. \begin{proposition} \label{prop:tau} The Jordan net $\mathcal{L}_* =uJ_5+x(E_{15}+E_{51})+y(E_{14}+E_{41}) $ in $\mathbb{S}^5$ has type 2(b). It is not in the closure of the diagonalizable locus 1(a) in ${\rm Jo}(3,\mathbb{S}^5)$. \end{proposition} \begin{proof} For any linear space $\mathcal{L} \subset \mathbb{S}^n$ we consider $\tau(\mathcal{L}):=\min\{\rk X\mid X\in\mathcal{L} \setminus \{0\}\}$. This invariant equals $1$ for all Jordan algebras $\mathcal{L}$ of type 1(a), since $k_3 = 1$ for $n=5$. For the special net above, we see that $\tau(\mathcal{L}_*) = 2$. It now suffices to note that $\{\mathcal{L}\in\Gr(m,\mathbb{S}^n)\mid \tau(\mathcal{L})\leq k\}$ is a closed set because it is the projection along $\mathbb{P}(\mathbb{S}^n)$ of the incidence variety $\{(\mathcal{L},X)\in\Gr(m,\mathbb{S}^n)\times\mathbb{P}(\mathbb{S}^n)\mid X\in\mathcal{L},\rk X\leq k\}$. \end{proof} The situation is easier if we restrict to the case of interest in applications. Let $\mathcal{L}$ be a real Jordan subalgebra of symmetric matrices whose unit $U$ is positive definite. Then $\mathcal{L}$ is {\em formally real}, i.e. if $X_1^{\bullet2}+\cdots+X_k^{\bullet 2}=0$, then $X_1,\ldots,X_k=0$. Formally real Jordan algebras have zero radical \cite{Jordan_1934} and this persists if we view them as Jordan algebras over $\mathbb{C}$. We define the {\em Euclidean Jordan locus} ${\rm EJo}(3,\mathbb{S}^n)$ to be the closure in ${\Gr}(3,\mathbb{S}^n) $ of the set of such Jordan algebras. For our study with $m=3$, we get that formally real Jordan nets have type 1(a) or~1(b). \begin{corollary} The number of irreducible components of ${\rm EJo}(3,\mathbb{S}^n)$ is the coefficient of the $t^n$ in the generating function \[ t^3/((1-t)(1-t^2)(1-t^3)) + t^2/(1-t^2). \] \end{corollary} \begin{proof} The first summand is the generating function of triples of $a,b,c\geq0$ with $a+2b+3c=n-3$. These correspond to triples of $k_1\geq k_2\geq k_3\geq 1$ with sum $n$ via \vspace{-0.16cm} \[ (a,b,c)\mapsto(a+b+c+1,b+c+1,c+1). \vspace{-0.16cm} \] The second generating function accounts for the Jordan nets of type 1(b). \end{proof} In the remainder of this section we present the proofs of our two theorems. \begin{proof}[Proof of Theorem \ref{thm:classification m=3,n=n}] Let $\mathcal{L}$ be a Jordan net of type 1(a). Then $\mathcal{L}$ is congruent to $\,\Diag(x{\bf 1}_{k_1},y{\bf 1}_{k_2},z{\bf 1}_{k_3}) $. This is seen by two applications of Lemma~\ref{lemma:idempotent}. Suppose $\mathcal{L}$ has unit $U = {\bf 1}_n$ and type 1(b). After applying an orthogonal congruence, we may assume that $X=\Diag({\bf 1}_r,-{\bf 1}_{n-r})$ for some $1\leq r\leq n-1$. The condition $XY+YX=0$ now implies $ Y= \begin{small} \begin{pmatrix} 0&P\\ P^\top&0\end{pmatrix} \end{small} $ for some $P\in\mathbb{C}^{r\times(n-r)}$. The condition $Y^2={\bf 1}_n$ implies $PP^\top={\bf 1}_r$ and $P^\top P={\bf 1}_{n-r}$. This is only possible when $2r=n$. By applying the orthogonal congruence $\Diag(P^\top, {\bf 1}_r)$, we get the desired form. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:classification m=3,n=4}] We proceed case-by-case, starting from the classification of Jordan algebras in Theorem~\ref{thm:3d_classification}. Let $U,X,Y$ be a basis of $\mathcal{L}$, where we assume $U={\bf 1}_4$. The types 1(a) and 1(b) are already covered by the previous proof. Suppose that $\mathcal{L}$ has type 2(a). We can assume $X=\Diag({\bf 1}_{4-k},{\bf 0}_k)$ and $Y=\Diag({\bf 0}_{4-k},Z)$ with $Z^2={\bf 0}_k$ for some $k\in\{2,3\}$. Up to orthogonal congruence, there is only one $Z \in \mathbb{S}^k$ with $Z^2 = {\bf 0}_k\neq Z$, as there is only one option for its Jordan canonical form. This uses Lemma~\ref{lemma:conjugation0}. So, we get two classes of this type. Suppose that $\mathcal{L}$ has type 2(b). Then we can assume that $X=\Diag({\bf 1}_k,{\bf 0}_{4-k})$. The condition $XY+YX=Y$ implies $ Y= \begin{small} \begin{pmatrix} 0&P\\ P^\top&0\end{pmatrix} \end{small} $ for some matrix $P\in\mathbb{C}^{k\times(4-k)}$. The condition $Y^2={\bf 0}_4$ implies $PP^\top={\bf 0}_k$ and $P^\top P={\bf 0}_{4-k}$. As $P$ is nonzero, this is only possible when $k=2$. So $P=ab^\top$ for some nonzero vectors $a,b\in\mathbb{C}^2$ with $a^\top a=b^\top b=0$. Such matrices $Y$ form a single orbit of $\Diag({\rm O}(2),{\rm O}(2))$. So we get one class of this type. Suppose that $\mathcal{L}$ has type 3(a). Then $X^2=Y$ and $X^3=0$. By Lemma~\ref{lemma:conjugation0} and appealing to Jordan canonical forms as above, there is a unique $X$ up to orthogonal congruence. So we get one class of this type. Consider $\mathcal{L}$ of type 3(b). We have $X^2=Y^2=XY+YX=0$. Since $X$, $Y$ or~$X+Y$ has rank~$2$, we may assume $\rk X=2$. One can show, using Lemma~\ref{lemma:conjugation0}, that we may assume $\rk Y=1$. Using Lemma~\ref{lemma:conjugation} and conjugating $\mathcal{L}$ in $\mathbb{C}^{n \times n}$, we~get\vspace{-0.15cm} $$ X=\begin{pmatrix} 0&{\bf 1}_2\\ 0&0 \end{pmatrix}\,\,\,\mbox{ and } \,\,\, Y=\begin{pmatrix} 0&Z\\0&0 \end{pmatrix}\vspace{-0.15cm} \,\,\quad \hbox{where $\,\rk Z =1$.} $$ By conjugating with a matrix $\Diag(P,P)$ and scaling, we find $Z=E_{22}$ or $Z=E_{12}$. So we get two classes. In one of them, $\rad\mathcal{L}$ is spanned by two rank-$1$ matrices. To identify the poset of inclusions, consider the associated nets of quadrics $ (a,b,c,d)\mathcal{L}(a,b,c,d)^T$. In this notation, type 1(a) is written as $x(a^2+b^2)+yc^2+zd^2$. Replacing $(a,b,c,d)$ by $(a,b,c,c+td)$ in 1(a), where $t \in \mathbb{R} \setminus \{0\}$, we get \vspace{-0.15cm} $$ x(a^2+b^2)+y'c^2+z'(2cd+td^2)\vspace{-0.15cm} $$ for $y'=y+z$ and $z'=tz$. As $t\to0$, we obtain type 2(a1). Similarly: $ \begin{array}{llclcl} \bullet\mbox{ We get ~1(a)}&\!\!\!\to\!\!\!&\mbox{2(a2)}&\mbox{using }(c+t^2a,tb,c,d).\\ \bullet\mbox{ We get ~1(b)}&\!\!\!\to\!\!\!&\mbox{2(b)}&\mbox{using }(-a,b,i(a+tc),i(b+td)). \end{array} $ $ \begin{array}{llclcl} \bullet\mbox{ We get ~1(b)}&\!\!\!\to\!\!\!&\mbox{3(b1)}&\mbox{using }(a,tb,tc,d).\\ \bullet\mbox{ We get ~2(a1)}&\!\!\!\to\!\!\!&\mbox{3(a)}&\mbox{using }(a,b,a+tb+t^2c,td).\\ \bullet\mbox{ We get ~2(a2)}&\!\!\!\to\!\!\!&\mbox{3(a)}&\mbox{using }(a,td,t(tc-b),a+tb).\\ \bullet\mbox{ We get ~2(a2)}&\!\!\!\to\!\!\!&\mbox{3(b1)}&\mbox{using }(a,d+tb,c,d).\\ \bullet\mbox{ We get ~2(b)}&\!\!\!\to\!\!\!&\mbox{3(b2)}&\mbox{using }(c+ta,d+tb,c,d).\\ \bullet\mbox{ We get ~3(a)}&\!\!\!\to\!\!\!&\mbox{3(b2)}&\mbox{using }(a,b,tc,i(b-td)).\\ \bullet\mbox{ We get ~3(b1)}&\!\!\!\to\!\!\!&\mbox{3(b2)}&\mbox{using }(a,a+tb,c,-c+td). \end{array} $ \noindent The codimensions were found by direct computation. It remains to show that there is no inclusion of orbit closures when there is no edge in our Hasse diagram. The arrow 1(a) $\rightarrow$ 2(b) is missing because $\tau$ cannot go up in limits; see Proposition \ref{prop:tau}. The absence of other inclusions is certified by explicit polynomials that vanish on one orbit, but not on another. For example, the polynomial for the orbit 1(b) displayed in (\ref{eq:onerep}) does not vanish on the orbits 1(a), 2(a1), 2(a2) and 3(a). Similarly, $4 p_{012} p_{146} - 4 p_{013} p_{145} - p_{014} p_{056} - p_{014} p_{126} + p_{014} p_{135} - p_{014} p_{234} + p_{024} p_{046} - p_{034} p_{045} \, $ vanishes on 2(a1) but not on 3(b1). \end{proof} \section{Nets of Quadrics} \label{sec6} In this section we offer a more detailed study of Jordan loci in the case $m=3$. \begin{proposition} \label{prop:33} The Jordan locus ${\rm Jo}(3,\mathbb{S}^3) $ consists of nets that are generated by three rank-$1$ conics in $\mathbb{P}^2$. This is an irreducible variety of codimension $3$ in the Grassmannian ${\rm Gr}(3,\mathbb{S}^3) $, and it has degree $57$ in the Pl\"ucker embedding in $ \mathbb{P}^{19}$. Its prime ideal is generated by $62$ quadrics in the $20$ Pl\"ucker coordinates. \end{proposition} \begin{proof} Here we use the Wall's classification \cite{Wall77} of nets of conics. For each net $\mathcal{L}$ in his list, we computed $\mathcal{L}^{-1}$. The surface $\mathcal{L}^{-1}$ is linear only for Wall's types E, G and H. According to \cite[Figure 5]{Wall77}, the strata G and H are contained in the closure of stratum E, which has codimension $3$, by \cite[Table 2]{Wall77}. We see in \cite[Table 1]{Wall77} that E corresponds to nets spanned by rank-$1$ conics. Hence the closure of stratum E is an irreducible variety. It equals ${\rm Jo}(3,\mathbb{S}^3) $. The results about the ideal and degree of this Jordan locus are in \cite[Proposition 3.1]{BS}. \end{proof} The ideal in Proposition \ref{prop:33} simplifies greatly if we restrict to $ {\rm Jo}_{\bf 1}(3,\mathbb{S}^3)$. Namely, for the net $\mathcal{L} = {\rm span} \{{\bf 1}_3,\, X,\, Y \}$, with $X = (x_{ij})$ and $Y= (y_{ij})$, we obtain \vspace{-0.14cm} \begin{equation} \label{eq:buchloe} \begin{small} \begin{matrix} \!\! \bigl\langle \, (x_{11} y_{12}-x_{12} y_{11})+(x_{12} y_{22}-x_{22} y_{12})+(x_{13} y_{23}-x_{23} y_{13}), \\ \quad (x_{11} y_{23} - x_{23} y_{11} )+(x_{12} y_{13}-x_{13} y_{12})+2 (x_{22} y_{23}-x_{23} y_{22}), \\ \qquad 2 (x_{11} y_{13}-x_{13} y_{11})+(x_{12} y_{23}-x_{23} y_{12})-(x_{13} y_{22}-x_{22} y_{13}) \bigr\rangle . \end{matrix} \end{small} \vspace{-0.14cm} \end{equation} This prime ideal of codimension $3$ defines the Jordan locus in ${\rm Gr}_{\bf 1}(3, \mathbb{S}^3)$. We now assume $n \geq 3$. For any net $\mathcal{L} \in {\rm Gr}(3,\mathbb{S}^n)$, the reciprocal $\mathcal{L}^{-1}$ is a surface in $\mathbb{P} (\mathbb{S}^n) = \mathbb{P}^{\binom{n+1}{2}-1}$. This is a linear plane precisely when $\mathcal{L}$ is a Jordan algebra. We now examine the other extreme case, when $\mathcal{L}$ is generic in~${\rm Gr}(3,\mathbb{S}^n)$. \begin{theorem} \label{thm:chow} The following conditions are equivalent for a net of quadrics $\mathcal{L}$: \vspace{-0.15cm} \begin{enumerate} \item Every nonzero complex matrix in $\mathcal{L}$ has rank $\geq n-1$. \vspace{-0.15cm} \item The linear span of the surface $\mathcal{L}^{-1}$ is the ambient projective space $\mathbb{P}(\mathbb{S}^n)$.\vspace{-0.15cm} \item $\mathcal{L}^{-1}$ is linearly isomorphic to the $(n-1)$st Veronese embedding of $\mathbb{P}^2$. \vspace{-0.15cm} \item The Chow matrix of $\mathcal{L}$ is invertible. \end{enumerate} \end{theorem} To explain this result, we must first define the {\em Chow matrix}. Fix a basis $X,Y,Z$ of $\mathcal{L}$, and let $x,y,z$ be unknowns. The Chow matrix is square of size $\binom{n+1}{2} \times \binom{n+1}{2}$, and its entries are homogeneous polynomials of degree $n-1$ in the entries of $X,Y,Z$. The columns are labeled by the monomials of degree $n-1$ in $x,y,z$, and the rows are labeled by pairs $(i,j)$ where $1 \leq i \leq j \leq n$. Consider the entry in position $(i,j)$ of the adjoint matrix of $xX+yY+zZ$. This is a polynomial of degree $n-1$ in $x,y,z$. Its coefficients form the row indexed $(i,j)$ of the Chow matrix. The entry in a given column is the coefficient of the monomial label. \begin{example}[$n=3$] The Chow matrix of three plane conics has format $6 \times 6$: \vspace{-0.15cm} $$ \begin{tiny} \begin{pmatrix} x_{22} x_{33}-x_{23}^2 & x_{22} y_{33}-2 x_{23} y_{23}+x_{33} y_{22} & x_{22} z_{33}-2 x_{23} z_{23}+x_{33} z_{22} & y_{22} y_{33}-y_{23}^2 & \cdots & \cdots \phantom{spa} \\ x_{13} x_{23} -x_{12} x_{33} &x_{13} y_{23}-x_{12} y_{33}+x_{23} y_{13}-x_{33} y_{12} & x_{13} z_{23}-x_{12} z_{33}+x_{23} z_{13}-x_{33} z_{12} & y_{13} y_{23} -y_{12} y_{33}& \cdots & \cdots \phantom{spa} \\ x_{12} x_{23}-x_{13} x_{22} & x_{12} y_{23}-x_{13} y_{22}-x_{22} y_{13}+x_{23} y_{12} & x_{12} z_{23}-x_{13} z_{22}-x_{22} z_{13}+x_{23} z_{12} & y_{12} y_{23}-y_{13} y_{22} & \cdots & \cdots \phantom{spa} \\ x_{11} x_{33}-x_{13}^2 & x_{11} y_{33}-2 x_{13} y_{13}+x_{33} y_{11} & x_{11} z_{33}-2 x_{13} z_{13}+x_{33} z_{11} & y_{11} y_{33}-y_{13}^2 & \cdots & \cdots \phantom{spa} \\ x_{12} x_{13} -x_{11} x_{23} & x_{12} y_{13} -x_{11} y_{23}+x_{13} y_{12}-x_{23} y_{11} & x_{12} z_{13} -x_{11} z_{23}+x_{13} z_{12}-x_{23} z_{11} & y_{12} y_{13} -y_{11} y_{23} & \cdots & \cdots \phantom{spa} \\ x_{11} x_{22}-x_{12}^2 & x_{11} y_{22}-2 x_{12} y_{12}+x_{22} y_{11} & x_{11} z_{22}-2 x_{12} z_{12}+x_{22} z_{11} & y_{11} y_{22}-y_{12}^2 & \cdots & \cdots \phantom{spa} \\ \end{pmatrix}. \end{tiny} \vspace{-0.13cm} $$ The row labels $(1,1),(1,2),(1,3),(2,2),(2,3),(3,3)$ refer to the entries of the adjoint of $xX+yY+zZ$. The column labels are the monomials $x^2, xy, xz, y^2, yz, z^2$. The determinant of the Chow matrix is an irreducible homogeneous polynomial of degree $12$ with $22659$ terms. It vanishes if and only if the net contains a rank-$1$ matrix. Thus, it equals the Chow form \cite{FKO} of the Veronese surface in~$\mathbb{P}^5$. \end{example} The {\em Chow form} of a projective variety is a hypersurface in the Grassmannian. Its points are linear spaces of complementary dimension that unexpectedly intersect the variety. If the linear space is expressed as the row space of a matrix then the entries of the matrix are the {\em dual Stiefel coordinates} of that linear space. \begin{lemm} \label{lem:stiefel} Consider the net $\mathcal{L} = \{xX + yY+zZ\}$ spanned by $X,Y,Z \in \mathbb{S}^n$. The Chow matrix is singular if and only if $\mathcal{L}\setminus\{{\bf 0}_n\}$ contains a matrix of rank $\leq n-2$. So the determinant of the Chow matrix is the Chow form, written in dual Stiefel coordinates, of the subvariety of $\mathbb{P}(\mathbb{S}^n)$ defined by the $(n-1) \times (n-1)$-minors. \end{lemm} \begin{proof} If $X$ has rank $\leq n-2$ then the coefficient of $x^{n-1}$ is zero in every entry of the adjoint of $xX + yY+zZ$, so the first row of the Chow matrix is zero. Such a rank drop happens whenever some matrix in $\mathcal{L}$ has rank $\leq n-2$ since the determinant of the Chow matrix is invariant of the choice of basis in $\mathcal{L}$. That determinant is a nonzero homogeneous polynomial in the entries of $X,Y,Z$ of degree $(n-1)\binom{n+1}{2}$. We already argued that it vanishes on the hypersurface defined by the Chow form of the rank-$( n-2)$ variety. That variety has degree $\binom{n+2}{3}$, so its Chow form in dual Stiefel coordinates is an irreducible polynomial of degree $3 \binom{n+2}{3}$. This irreducible polynomial divides our determinant. Since $(n-1)\binom{n+1}{2} = 3 \binom{n+2}{3}$, the two polynomials agree up to a nonzero constant. \end{proof} \begin{remark} The Chow matrix is new and interesting even for $n=4$. It gives the Chow form for symmetric $4 \times 4$ matrices of rank~$\leq~2$. A formula for that Chow form was the main result in \cite[Section 3]{FKO}. Our construction is much simpler. \end{remark} \begin{proof}[Proof of Theorem \ref{thm:chow}] Lemma \ref{lem:stiefel} says that conditions (1) and (4) are equivalent. The surface $\mathcal{L}^{-1}$ is parametrized by the entries of the adjoint of $xX+yY+zZ$. By inverting the Chow matrix, we can express every monomial of degree $n-1$ in $x,y,z$ as a linear combination of those entries. Hence $\mathcal{L}^{-1}$ is linearly isomorphic to the Veronese surface, so (4) implies (3). Clearly (3) implies (2). Finally, if (4) fails then $\mathcal{L}^{-1}$ lies in a hyperplane in $\mathbb{P}(\mathbb{S}^n)$, so (2) fails. \end{proof} \begin{corollary} \label{cor:chowrank} Given any regular net $\mathcal{L}$ of symmetric $n \times n$ matrices, the rank of its Chow matrix equals the dimension of the linear span of $\mathcal{L}^{-1}$ inside $\mathbb{S}^n$. \end{corollary} \begin{proof} A linear form vanishes on $\mathcal{L}^{-1}$ if and only if it vanishes on the adjoint of $xX + yY+zZ$. These are the vectors in the left kernel of the Chow matrix. \end{proof} Every regular subspace $\mathcal{L}\subset\mathbb{S}^n$ generates a unique Jordan subalgebra ${\rm Jor}(\mathcal{L})$ inside $\mathbb{S}^n$. The dimension of ${\rm Jor}(\mathcal{L})$ is bounded below by the rank of the Chow matrix of $\mathcal{L}$. However, this bound is not tight. To see this, consider the~net \begin{equation} \label{eq:netrank8} \mathcal{L} \,\,=\,\,\, \begin{small} \begin{pmatrix} \, x +y \, & z & z & 0 \,\, \\ z & \!\! x-y \!\! & 0 & z \,\, \\ z & 0 & x & z \,\,\\ 0 & z & z & x \,\, \\ \end{pmatrix} \end{small}. \end{equation} Here, ${\rm Jor}(\mathcal{L})= \mathbb{S}^4$ but the $10 \times 10$ Chow matrix has rank~$8$. Its left kernel gives the two linear forms $\, z_{14}-z_{23}-z_{33}+z_{44} \,$ and $\, 2 z_{12}-z_{13}-z_{24} \,$ which vanish on $\mathcal{L}^{-1}$. \smallskip We now return to the polynomial equations that vanish on the Jordan locus in the Grassmannian. The following result is immediate from Corollary \ref{cor:chowrank}. \begin{proposition} \label{prop:choweqns} Using dual Stiefel coordinates on ${\rm Gr}(m,\mathbb{S}^n)$, the Jordan locus $\,{\rm Jo}(m,\mathbb{S}^n)\,$ is cut out by the $(m+1) \times (m+1)$-minors of the Chow matrix. \end{proposition} \begin{proof} For any regular $\mathcal{L}\in\Gr(m,\mathbb{S}^n)$, the reciprocal variety $\mathcal{L}^{-1}$ is irreducible of dimension $m$ and contained in its span, which is itself an irreducible variety. These two varieties coincide if and only if the latter has dimension~$\leq m$. \end{proof} \begin{remark} These $4 \times 4$-minors do not generate the prime ideal of ${\rm Jo}(3,\mathbb{S}^n)$. For instance, if $n=3$ and we fix ${\bf 1}_3 \in \mathcal{L}$ then Proposition \ref{prop:choweqns} yields $30$ quartics in $x_{ij},y_{ij}$. These cut out the same variety ${\rm Jo}_{\bf 1}(3,\mathbb{S}^3)$ as the three quadrics in (\ref{eq:buchloe}). \end{remark} We know from Theorem \ref{thm:classification m=3,n=4} that ${\rm Jo}(3,\mathbb{S}^4)$ has two irreducible components, of types 1(a) and 1(b). They are given by the nets $\mathcal{L}_1$ and $\mathcal{L}_2$ in the next example. \begin{example} \label{ex:drei} We consider the congruence orbits of the following nets in $\mathbb{S}^4$: \vspace{-0.1cm} $$\mathcal{L}_1 \,\,= \,\, \begin{small} \begin{pmatrix} x & 0 & 0 & 0 \\ 0 & x & 0 & 0 \\ 0 & 0 & y & 0 \\ 0 & 0 & 0 & z \end{pmatrix} \end{small}\, , \,\,\, \mathcal{L}_2 \,\,= \,\, \begin{small} \begin{pmatrix} x & y & 0 & 0 \\ y & z & 0 & 0 \\ 0 & 0 & x & y \\ 0 & 0 & y & z \end{pmatrix} \end{small} \, , \,\,\, \mathcal{L}_3 \,\,= \,\, \begin{small} \begin{pmatrix} 0 & 0 & x & y \\ 0 & -2x & -y & z \\ x & -y & -2z & 0 \\ y & z & 0 & 0 \end{pmatrix} \end{small}. \vspace{-0.1cm} $$ They have codimensions $9$, $10$ and $9$ in ${\rm Gr}(3,\mathbb{S}^4)$. The first two fill out ${\rm Jo}(3,\mathbb{S}^4)$. It is interesting to compare the non-diagonalizable nets $\mathcal{L}_2$ and $\mathcal{L}_3$. Both represent double conics in $\mathbb{P}^2$, since ${\rm det}(\mathcal{L}_2) = {\rm det}(\mathcal{L}_3) = (xz-y^2)^2$. But, while $\mathcal{L}_2^{-1}$ is linear, the Chow matrix of $\mathcal{L}_3$ is invertible, so $\mathcal{L}_3^{-1}$ is a Veronese surface. The congruence orbit of $\mathcal{L}_3$ is the smooth variety of codimension $9$ studied~in \cite{EPS}. It is essentially the Hilbert scheme of twisted cubic curves in $\mathbb{P}^3$. \end{example} We end this section by reporting a challenging computation. It is predicated on the idea that Pl\"ucker coordinates are best when working with Grassmannians. \begin{example}\label{example: quadrics in plucker} The $21$-dimensional Grassmannian ${\rm Gr}(3,\mathbb{S}^4)$ lives in $\mathbb{P}^{119}$. This space has $\binom{10}{3} = 120$ dual Pl\"ucker coordinates $p_{ijk}$, one for each $3 \times 3$-minor~of \vspace{-0.14cm} $$ \begin{small} \begin{pmatrix} \, u_{11} & u_{12} & u_{13} & u_{14} & u_{22} & u_{23} & u_{24} & u_{33} & u_{34} & u_{44} \, \\ \, x_{11} & x_{12} & x_{13} & x_{14} & x_{22} & x_{23} & x_{24} & x_{33} & x_{34} & x_{44} \, \\ \, y_{11} & y_{12} & y_{13} & y_{14} & y_{22} & y_{23} & y_{24} & y_{33} & y_{34} & y_{44}\ \\ \end{pmatrix}. \end{small} \vspace{-0.14cm} $$ We computed the quadrics that vanish on subvarieties of ${\rm Gr}(3,\mathbb{S}^4)$. Modulo Pl\"ucker relations, there are $4950$ linearly independent quadrics. Each quadric has the form $\sum \lambda^{ijk}_{lmn} p_{ijk} p_{lmn}$ where the $\lambda^{ijk}_{lmn}$ are unknown coefficients and $i \leq l, j \leq m, k \leq n$. For each of the three orbits in Example \ref{ex:drei}, we set up a linear system of equations in the $\lambda^{ijk}_{lmn}$ whose solutions are quadrics that vanish on that variety. We then solved these equations using {\tt Maple}. Here are the results: We found a basis of $85$ quadrics for the orbit of $\mathcal{L}_1$. One representative is \vspace{-0.16cm} $$ \begin{small} \begin{matrix} p_{124} p_{789}-p_{125} p_{589}+p_{125} p_{679} -p_{126} p_{678} + p_{127} p_{569} - p_{128} p_{568} - p_{234} p_{579} \\ + p_{234} p_{678} + p_{235} p_{479} + p_{235} p_{568} - p_{236} p_{478} - p_{236} p_{567} - p_{237} p_{459} + p_{238} p_{458}. \end{matrix} \end{small} \vspace{-0.16cm} $$ We found a basis for $189$ quadrics for the orbit of $\mathcal{L}_2$. One representative is \vspace{-0.16cm} \begin{equation} \label{eq:onerep} \begin{small} \begin{matrix} p_{012} p_{018} - p_{012} p_{026} - p_{012} p_{035} - 2 p_{012} p_{123} - p_{013} p_{017} + 2 p_{013} p_{025} - p_{023} p_{024}. \end{matrix} \end{small} \vspace{-0.16cm} \end{equation} Each of these $189$ quadrics also vanishes on $\mathcal{L}_3$. But, there are $918$ quadrics vanishing on the orbit of $\mathcal{L}_3$. One representative is $\, 8 p_{012} p_{457} \,-\, 4 p_{045} p_{057} \,+\, p_{047}^2 $. To see that $\mathcal{L}_3$ is not contained in $\mathcal{L}_2$, we can use the quartics in Proposition~\ref{prop:choweqns}. \end{example} \begin{small}

https://arxiv.org/abs/physics/0503214

Optimal supply against fluctuating demand

Sornette et al. claimed that the optimal supply does not agree with the average demand, by analyzing a bakery model where a daily demand fluctuates with a uniform distribution. In this note, we extend the model to general probability distributions, and obtain the formula of the optimal supply for Gaussian distribution, which is more realistic. Our result is useful in a real market to earn the largest income on average.

\section{Introduction} Sornette et al.\ (1999) claimed that the optimal supply does not agree with the average demand, contrary to the common sense in economy. They considered a bakery model where a daily demand fluctuates with a uniform distribution, and derived the formula of the optimal supply. Although their result is reasonable and meaningful, it is not clear how the result will change if we consider a different distribution. In this note, we extend the model to general probability distributions, and calculate the optimal supply for the Gaussian distribution, which is more realistic in a market. \section{Model and analysis of Sornette, Stauffer and Takayasu} In this section we review the model and analysis of Sornette et al.\ Let us consider a bakery shop where baked croissants are sold every day. A question is how many croissants should be baked a day to make the maximal profit. We define the variables as follows. \begin{itemize} \item $x$: the selling price of a croissant. \item $y$: its production cost. \item $s$: the production number of croissants per day (supply). \item $n$: the number of croissants requested by customers per day (demand). \item $D$: the average demand, i.e., $D\equiv<n>$. \end{itemize} The expectation of the total profit $L(s)$ is given by \begin{equation}\label{L} L(s)\equiv<x~{\rm min}(n,s)-ys> =x\int^s_0nP(n)dn+xs\int^{\infty}_sP(n)dn-ys, \end{equation} where $P(n)$ is the probability distribution of $n$. Sornette et al.\ assumed, for simplicity, a uniform distribution, \begin{equation}\label{uniform} P_u(n)\equiv \left\{\begin{array}{ll} 1/2\delta ~& {\rm for} ~~~D-\delta\le n\le D+\delta\\ 0 ~& {\rm for} ~~~n<D-\delta,~D+\delta<n. \end{array}\right. \end{equation} Then one can integrate (\ref{L}) as \begin{equation} L(s)=-{x\over4\delta}\left\{s-D-\delta\left(1-{2y\over x}\right)\right\}^2 +(x-y)\left(D-{\delta y\over x}\right). \end{equation} $L(s)$ takes the maximam value when $s$ takes \begin{equation}\label{su} s_{{\rm max}}\equiv D+\delta\left(1-{2y\over x}\right). \end{equation} This shows that, if the cost per price, $y/x$, is larger (smaller) than half, the optimal demand, $s_{{\rm max}}$, is smaller (larger) than the average demand, $D$. \section{Optimal supply for Gaussian distribution} Let us re-analyze (\ref{L}) for general probability distributions. We do not have to integrate (\ref{L}) directly, because what we want to know is the optimal supply $s_{{\rm max}}$, which is given by \begin{equation}\label{dL} {dL\over ds}(s_{{\rm max}})=x\int^{\infty}_{s_{{\rm max}}}P(n)dn-y=0. \end{equation} This simple equation gives the optimal supply for general probability distributions. If we assume Gaussian distribution, \begin{equation} P_G(n)\equiv{1\over\sqrt{2\pi}\sigma} \exp\left[-{(n-D)^2\over2\sigma^2}\right], \end{equation} the above integration is expressed as \begin{equation} \int^{\infty}_{s_{{\rm max}}}P_G(n)dn=\frac12-\frac12 {\rm Erf}\left({s_{{\rm max}}-D\over\sqrt{2}\sigma}\right), \end{equation} where Erf is the error function, which is defined as \begin{equation} {\rm Erf}~z\equiv{2\over\sqrt{\pi}}\int^z_0e^{-t^2}dt. \end{equation} Then we arrive at the formula of the optimal supply for the Gaussian distribution, \begin{eqnarray}\label{sG} {s_{{\rm max}}-D\over\sigma} &=&\sqrt{2}{\rm Erf}^{-1}\left(1-\frac{2y}x\right) \nonumber\\ &=&\sqrt{{\pi\over2}}\left(1-\frac{2y}x\right) +{\sqrt{2}\pi^{\frac32}\over24}\left(1-\frac{2y}x\right)^3 +O\left[\left(1-\frac{2y}x\right)^5\right] ~~~({\rm Gaussian}). \end{eqnarray} Because the cost is usually in the range $0.3x<y<0.7x$, which reads $|1-2y/x|<0.4$, the first-order approximation in (\ref{sG}) is sufficient in most cases. For reference, we rewrite the result for the uniform distribution (\ref{su}). Because the variance of the uniform distribution (\ref{uniform}) is evaluated as $\sigma^2=\delta^2/3$, (\ref{su}) is rewritten as \begin{equation}\label{su2} {s_{{\rm max}}-D\over\sigma}=\sqrt{3}\left(1-\frac{2y}x\right) ~~~({\rm uniform}). \end{equation} We see that the difference between $\sqrt{\pi/2}\approx1.25$ in (\ref{sG}) and $\sqrt{3}\approx1.73$ in (\ref{su2}) is not negligible. Contrary to the speculation of Sornette {\it et al.}, however, the critical value of the cost-to-price ratio, $y/x=1/2$, is unchanged. Because Gaussian distribution is more realistic, our simple formula (\ref{sG}) is useful in a real market to earn the largest income on average.

https://arxiv.org/abs/2211.03255

Minimal Area of a Voronoi Cell in a Packing of Unit Circles

We present a new self-contained proof of the well-known fact that the minimal area of a Voronoi cell in a unit circle packing is equal to $2\sqrt{3}$, and the minimum is achieved only on a perfect hexagon. The proof is short and, in our opinion, instructive.

\section{Introduction} This work originated from attempts to rely on the proof in \cite{H} of the fact that the minimal area of a Voronoi cell in a unit circle packing in $\bbR^2$ is equal to $2\sqrt{3}$. Unfortunately, this proof contains some gaps (see Appendix), and our efforts to recover the proof resulted in an alternative approach presented below. The result itself immediately implies the strong/local version of the theorem of Thue and Fejes T\'oth \cite{Th, To} on dense unit circle packings in $\bbR^2$. Additional applications include the derivation of so-called Peierls estimates in two-dimensional lattice hard-core models of statistical mechanics \cite{MSS1}. \section{Definitions and basic properties} For any $\bx \in \bbR^2$ and $r > 0$ denote by $B_r(\bx)$ an open disk of radius $r$ centered at $\bx$. A shorthand notation $B(\bx)$ is used for the {\it unit} disk having $r=1$. A collection of non-overlapping open unit disks is called {\it admissible} and is denoted by $\{B(\bx_i)\}$. An admissible collection $\{B(\bx_i)\}$ represents a {\it unit circle packing}. (It is clear that an admissible collection is finite or countable, as inside each disk there exists a point with rational coordinates.) An admissible collection $\{B(\bx_i)\}$ can be identified with the corresponding collection of centers $\{\bx_i\}$ which is also called admissible. Each element of $\{\bx_i\}$ is called an {\it occupied} point in $\bbR^2$; respectively, we speak about admissible collections of occupied points. Clearly, $|\bx_{i'}-\bx_{i''}| \ge 2$ for any distinct $\bx_{i'}, \bx_{i''} \in \{\bx_i\}$, where $|\bx_{i'}-\bx_{i''}|$ denotes the Euclidean distance between $\bx_{i'},\bx_{i''} \in \bbR^2$. For each $\bx_{i'} \in \{\bx_i\}$ the corresponding {\it Voronoi cell} is defined as $$V(\bx_{i'}) = \left\{\bz \in \bbR^2:\;\;|\bx_{i'}-\bz| \le \inf_{i'' \not = i'}|\bx_{i''}-\bz|\right\}.$$ If for $\bx_{i''}$ the intersection $V(\bx_{i'}) \cap V(\bx_{i''})$ contains more than one point of $\bbR^2$ then we say that $\bx_{i''}$ is a {\it neighboring occupied point} for $\bx_{i'}$ or simply $\bx_{i''}$ is a {\it neighbor} of $\bx_{i'}$. Observe that $V(\bx)$ can be unbounded. If $V(\bx)$ is bounded, i.e., $V(\bx) \subset B_{r}(\bx)$ for some $0 < r <\infty$, then $B(\by) \subset B_{4r}(\bx)$ for any neighboring occupied point $\by$. Due to the admissibility requirement the number of such $\by$ cannot exceed $|B_{4r}(\cdot)|/|B(\cdot)| < \infty$, where $|\cdot|$ denotes the area of the corresponding disk. Our aim is to find the minimal possible area of a Voronoi cell among all Voronoi cells in all admissible collections of occupied points. Suppose that a Voronoi cell with minimal or close to minimal area can be found in an admissible collection containing unbounded Voronoi cells. Then this collection can be completed without breaking the admissibility by a finite or countable set of additional occupied points such that the resulting admissible collection has bounded Voronoi cells only. For that reason from now on we mainly consider admissible collections without unbounded Voronoi cells. The rest of this section verifies some standard properties of Voronoi cells which makes the entire argument self-contained. \medskip {\bf Lemma 1.} {\sl For an occupied point $\bx$ in an admissible collection the corresponding Voronoi cell $V(\bx)$ contains the closure of $B(\bx)$.} \medskip {\bf Proof.} If for a point $\bz \in B(\bx)$ there exists an occupied point $\by \not = \bx$ with $|\by-\bz| < |\bx-\bz| < 1$ then by the triangle inequality $|\by-\bx| < 2$ which contradicts the admissibility of the collection. \qed \medskip {\bf Lemma 2.} {\sl For an occupied point $\bx$ in an admissible collection the corresponding Voronoi cell $V(\bx)$ is a convex subset of $\bbR^2$.} \medskip {\bf Proof.} For an occupied point $\by \not = \bx$ and any $\bz \in V(\bx)$ one has $|\bz-\bx|^2 \le |\bz-\by|^2$ and consequently $$|\bx|^2 - 2\bz\cdot \bx \le |\by|^2 - 2\bz\cdot \by.$$ Consider two distinct points $\bz', \bz'' \in V(\bx)$ and suppose that $\bz = \lambda \bz' + (1-\lambda) \bz''$, where $0 \le \lambda \le 1$. Then $$\beacl |\bz -\bx|^2 &= |\bz|^2 + \lambda(|\bx|^2-2 \bz'\cdot \bx) + (1-\lambda)(|\bx|^2-2 \bz''\cdot \bx)\cr \\ &\le |\bz|^2 + \lambda(|\by|^2-2 \bz'\cdot \by) + (1-\lambda)(|\by|^2-2 \bz''\cdot \by) \cr &= |\bz -\by|^2,\ena $$ which establishes the lemma. \qed \medskip {\bf Lemma 3.} {\sl For an occupied point $\bx$ in an admissible collection the boundary of the corresponding Voronoi cell $V(\bx)$ is piecewise linear.} \medskip {\bf Proof.} Consider two different occupied points $\bx, \by$ with $V(\bx)\cap V(\by)$ containing more than one $\bbR^2$ point. Let $\bz', \bz'' \in V(\bx)\cap V(\by)$ and $\bz' \not= \bz''$. Now suppose that $\bz = \lambda \bz' + (1-\lambda) \bz''$, $0 \le \lambda \le 1$. Then $$\beacl |\bz -\bx|^2 &= |\bz|^2 + \lambda(|\bx|^2-2 \bz'\cdot \bx) + (1-\lambda)(|\bx|^2-2 \bz''\cdot \bx)\cr &= |\bz|^2 + \lambda(|\by|^2-2 \bz'\cdot \by) + (1-\lambda)(|\by|^2-2 \bz''\cdot \by) \cr &= |\bz -\by|^2,\ena $$ i.e., $\bz \in V(\bx)\cap V(\by)$. Also note that for any two distinct occupied points $\bx, \by$ the set $$\left\{ \bz \in \bbR^2:\;\; |\bz -\bx| \le |\bz -\by| \right\}$$ is a closed half-plane. Consequently, $V(\bx)$ is an intersection of a finite or countable number of closed half-planes. \qed \medskip Lemmas~1-3 are valid for both bounded and unbounded Voronoi cells in an admissible collection of occupied points. If $V(\bx)$ is bounded then these lemmas imply that $V(\bx)$ is a convex polygon containing the unit disk $B(\bx)$. Furthermore, each neighboring occupied point $\by$ is the reflection of $\bx$ with respect to the common side of $V(\bx)$ and $V(\by)$. The clockwise circular order of sides of a bounded polygon $V(\bx)$ generates the clockwise circular order of the corresponding neighboring occupied points. Connecting the neighboring points in this circular order we obtain a so-called {\it polygon of neighbors}. Together with $\bx$, each side of this polygon uniquely defines a {\it constituting triangle} such that the entire polygon of neighbors is partitioned into constituting triangles. By construction, the vertices of $V(\bx)$ are the centers of the circumcircles of these constituting triangles. The angle of the constituting triangle at vertex $\bx$ is called the {\it constituting angle}. \medskip {\bf Lemma 4.} {\sl The circumradius of a constituting triangle is not shorter than $2/\sqrt{3}$.} \medskip {\bf Proof.} At least one angle of a constituting triangle, say angle $\alpha$, is not larger than $\pi / 3$. By the admissibility requirement the length $a$ of the opposite triangle side is not shorter than $2$. According to the sine theorem for triangles the triangle circumradius $$r = {a \over 2 \sin \alpha} \le {2 \over 2 \sin {\pi \over 3}} = {2\over\sqrt{3}},$$ which establishes the lemma. \qed \section{Results} Consider an admissible collection $\{\bx_i\}$ which forms a triangular lattice with the shortest distance between sites equal to $2$. Then for each $\bx_i$ the corresponding polygon of neighbors is a perfect hexagon with the side length equal to $2$. Correspondingly, $V(\bx_i)$ is a perfect hexagon with the side length equal to $2 / \sqrt{3}$ and the area $|V(\bx_i)|= 2 \sqrt{3}$. \medskip {\bf Theorem.} {\sl The minimal area of a Voronoi cell in an admissible configuration of occupied points is equal to $2 \sqrt{3}$. The minimum is achieved only on occupied points $\bx$ having a perfect hexagon with side length $2/\sqrt{3}$ as its Voronoi cell, or equivalently, a perfect hexagon with side length} $2$ as the corresponding polygon of neighbors. \medskip {\bf Proof.} In view of properties of a Voronoi cell presented in Lemmas~1-4 the problem is reduced to finding the polygon of minimal area among all polygons $P$ having the following properties: \begin{description} \item{(i)} $P$ is convex. \item{(ii)} $P$ contains a unit disk. \item{(iii)} The distances from the vertices of $P$ to the center of this disk are not shorter than~${2 \over \sqrt{3}}$.\end{description} \noindent The last property is a weaker replacement of the requirement for a Voronoi cell to have the corresponding neighboring occupied points at distances not shorter than $2$ from each other. It turns out that this weaker requirement is enough to establish the theorem. For a convex polygon of area $a$ and one of its angles of measure $\alpha$ define the {\it angular density} (with respect to this angle) as the ratio $E={a \over \alpha}$. Now take any polygon satisfying (i)-(iii) and consider several rays originating at the center $\bo$ of the contained unit disk. Let the angles between the clockwise consecutive rays be smaller than $\pi$. These rays partition the polygon into several convex polygons each located inside the angle $\alpha_j$ between the corresponding two clockwise consecutive rays. Clearly, the area $a$ of the original polygon $P$ can be calculated as $a= \sum_j E_j \alpha_j$, where $E_j$ is the corresponding angular density (with respect to angle $\alpha_j$). For any vertex $\bv$ of $P$ the convex hull of this vertex and the unit disk centered at $\bo$ belongs to $P$. If $|\bo-\bv|=r$ then this convex hull is bounded by two straight segments $[\bv\ba]$ and $[\bv\bc]$ of length $\sqrt{r^2-1}$ and the arc of the unit circle connecting $\ba$ with $\bc$ and having length $2\pi - 2 \arctan\sqrt{r^2-1}$. The area of the quadrilateral $[\bo\ba\bv\bc]$ is equal to $\sqrt{r^2-1}$ and $|\angle \ba\bo\bc|=2\arctan\sqrt{r^2-1}$. Therefore, the corresponding angular density is $E_{\bv} = {\sqrt{r^2-1}\over 2 \arctan\sqrt{r^2-1}}$. Take any point $\bb$ inside the segment $[\bv\bc]$ and consider the angular density $\overline E_{\bv}$ of the smaller quadrilateral $[\bo\ba\bb\bv]$ with respect to the angle $\angle \ba\bo\bb$. Let $|\bb-\bc| = x < \sqrt{r^2-1}$. Then $$\overline E_{\bv} = {\sqrt{r^2-1} - {x \over 2} \over 2\arctan \sqrt{r^2-1} - \arctan x} > {\sqrt{r^2-1} \over 2\arctan \sqrt{r^2-1}} = E_{\bv}.$$ For two clockwise consecutive polygon vertices $\bv_i$ and $\bv_{i+1}$ consider the corresponding convex hulls and two corresponding quadrilaterals $[\bo\ba_i\bv_i\bc_i]$ and $[\bo\ba_{i+1}\bv_{i+1}\bc_{i+1}]$. Observe that consecutive open quadrilaterals are either adjacent (have a common side) or intersecting. (In the case when they are separated by a non-zero angle $\angle \bc_i\bo\ba_{i+1}$ the segment $[\bv_i \bv_{i+1}]$ intersects the interior of the unit disk which contradicts property (ii).) With this observation at hand, denote by $\bb_i$ the intersection point of segments $[\bv_i\bc_i]$ and $[\bv_{i+1}\ba_{i+1}]$. According to the displayed equation above, the angular density of $[\bo\ba_i\bv_i\bb_i]$ is larger than the angular density of $[\bo\ba_i\bv_i\bc_i]$ and the angular density of $[\bo\bb_i\bv_{i+1}\bc_{i+1}]$ is larger than the angular density of $[\bo\ba_{i+1}\bv_{i+1}\bc_{i+1}]$. \def\cQ{\mathcal Q} Consider now the union of the quadrilaterals $[\bo\ba_i\bv_i\bc_i]$ over all vertices $\bv_i$. It is a polygon $Q$ (generally, non-convex) contained in $P$. Between each two vertices $\bv_i$ and $\bv_{i+1}$ the polygon $Q$ may contain an additional vertex $\bb_i$ that is the intersection point introduced above. The angular density of $[\bo\bb_{i-1}\bv_i\bb_i]$ is not smaller than the angular density of $[\bo\ba_i\bv_i\bc_i]$, and the two angular densities are equal only if $\bb_{i-1}=\ba_i$ and $\bb_i=\bc_i$. For $r \ge {2 \over \sqrt{3}}$ the minimum of $E_{\bv} = {\sqrt{r^2-1}\over 2\arctan\sqrt{r^2-1}}$ equals $\sqrt{3} \over \pi$ and is achieved only at $r = {2 \over \sqrt{3}}$. Thus, the total area of the union of the quadrilaterals $[\bo\ba_i\bv_i\bc_i]$ (or equivalently the union of mutually disjoint open quadrilaterals $[\bo\bb_{i-1}\bv_i\bb_{i}]$) is not smaller than ${\sqrt{3} \over \pi} 2\pi = 2\sqrt{3}$. Obviously, this minimum is achieved only when the interiors of the quadrilaterals $[\bo\ba_i\bv_i\bc_i]$ are disjoint (i.e., $[\bo\ba_i\bv_i\bc_i] = [\bo\bb_{i-1}\bv_i\bb_{i}]$) and $|\bo-\bv_i| = {2 \over \sqrt{3}}$ for all $i$. In this case the corresponding polygon is the perfect hexagon with side length ${2 \over \sqrt{3}}$. Indeed, if $V(\bx)$ is a hexagon with $r_i > {2 \over \sqrt{3}}$ for some $i$ then $|\angle \ba_i \bo \bc_i| > {2 \pi \over 6}$. Therefore, such a hexagon has the angular density larger than minimal for some non-zero angle. Consequently, its area is larger than $2\sqrt{3}$. Any polygon with $n > 6$ vertices contains at least two overlapping quadrilaterals $[\bo \ba_i\bv_i\bc_i]$ even if $r_i = {2 \over \sqrt{3}}$ for all $i$ because $n {2 \pi \over 6} > 2\pi$. Therefore, the angular density becomes larger than minimal for some non-zero angle. Hence, any polygon satisfying (i)-(iii) with more than 6 vertices has a non-minimal area. A polygon with $n = 3,4$ or 5 vertices necessarily has at least one $|\angle \ba_i \bo \bc_i| \ge {2\pi \over n}$ and therefore $r_i > {1 \over \cos{\pi \over n}} > {2 \over \sqrt{3}}$. Thus, the corresponding angular density $E_{\bv_i}$ is again greater than $\sqrt{3} \over \pi$ for some non-zero angle; consequently, the area of the entire polygon is larger than the minimal area $2\sqrt{3}$.~\qed \section{Appendix} The problem with the argument in \cite{H} is that at some point the proof considers the ``most critical case'', but it is not specified what quantity is optimized at this ``most critical case'' and why this quantity is important. The desired quantity seems to be the area excess in $V(\bx)$ over the area of the contained unit disk which can be attributed to a single non-close neighbor. (In the terminology of \cite{H} a neighbor $\by$ is non-close to the center $\bx$ of $V(\bx)$ iff $|\bx-\by| > 2.3$.) By considering this ``most critical case'' for a single non-close neighbor the proof in \cite{H} concludes that the corresponding area excess is at least $0.21$. After that the proof claims that the presence of two non-close neighbors implies that the area excess is at least $0.42$. This additivity assumption is actually wrong as individual excesses can overlap, and one can give a counterexample showing two non-close neighbors with the total attributed area excess smaller than $0.42$.

https://arxiv.org/abs/2101.02995

The set of ratios of derangements to permutations in digraphs is dense in $[0, 1/2]$

A permutation in a digraph $G=(V, E)$ is a bijection $f:V \rightarrow V$ such that for all $v \in V$ we either have that $f$ fixes $v$ or $(v, f(v)) \in E$. A derangement in $G$ is a permutation that does not fix any vertex. In [1] it is proved that in any digraph, the ratio of derangements to permutations is at most $1/2$. Answering a question posed in [1], we show that the set of possible ratios of derangements to permutations in digraphs is dense in the interval $[0, 1/2]$.

\section{Introduction} A {\em permutation} in a digraph $G=(V, E)$ is a bijection $f:V \rightarrow V$ such that for all $v \in V$ we either have that $f$ fixes $v$ or $(v, f(v)) \in E$. A {\em derangement} in $G$ is a permutation that does not fix any vertex. We define the parameter $(d/p)_G$ to be the ratio of derangements to permutations in $G$. Bucic, Devlin, Hendon, Horne and Lund \cite{BDHHL} showed that $(d/p)_G \le 1/2$ for all digraphs $G$, with equality if and only if $G$ is a directed cycle. They also gave a construction (the blow-up of a directed cycle) that can achieve a ratio arbitrarily close to but not equal to $1/2$. Let $S=\{(d/p)_G\; :\; G \;\;\text{is a digraph}\}$ be a set of values arising as a ratio $(d/p).$ In \cite{BDHHL} they analyzed the ratio $(d/p)_G$ for the random graph $G=G(n, m)$, and as a corollary of this analysis they showed that $S$ is dense in $[0, 1/e]$. This corollary follows from two facts: $(d/p)_G$ is concentrated around its mean, and by choosing a suitable value of $m$ one can make the expected ratio $(d/p)_G$ close to any given value in $[0, 1/e]$. At the end of the paper \cite{BDHHL} they ask whether $S$ is dense in $[0, 1/2]$. Our main theorem, below, answers this question in the positive. \begin{maintheorem}\label{thm:main} The set of possible ratios of derangements to permutations in digraphs is dense in $[0, 1/2]$. \end{maintheorem} The construction we use, described in more detail later, is a random subgraph of the blow-up of a directed cycle. The main part of the proof is an application of the second moment method (see \cite{FK}, for example, for an introduction to the method) to show that the number of derangements and permutations are concentrated around their expectations. \section{Proof of Theorem \ref{thm:main}} \subsection{Outline} First we outline the proof. Suppose we are given a fixed real number $r \in [0, 1/2]$. We will show that there exists a sequence of digraphs $G_k$ such that the ratio of derangements to permutations in $G_k$ is $r+o(1)$ as $k \rightarrow \infty$ (which proves Theorem \ref{thm:main}). If $r=0$ or $1/2$ this is trivial. Indeed, for $r=0$ observe that a digraph with one vertex and no edges has no derangements and one permutation, and for $r=1/2$ observe that any directed cycle has one derangement and two permutations. So we assume $0 < r < 1/2$. Our construction is as follows. As defined in \cite{BDHHL}, let the digraph $D_{k,\l}$ where $k\geq 1$ and $\l\geq 2$ have vertices $v_{ij}$ for $i\in [k]$ and $j\in[\l]$ such that $(v_{ij}, v_{lm})\in E(D_{k,\l})$ if and only if $m=j+1\, \text{mod}\, l$. In other words, $D_{k, \l}$ is the blow-up of a directed $\l$-cycle where each vertex is expanded to a set of $k$ vertices. We let $V_i = \{v_{ij}: j \in [\l]\}$. As was shown in \cite{BDHHL}, the number of derangements on $D_{k,\l}$ is $(k!)^\l$ and the number of permutations on $D_{k,\l}$ is $\sum_{i=0}^k\left(\binom{k}{i} (k-i)!\right)^\l$. Hence, $(d/p)_{k,\l}=\left(\sum_{i=0}^k \left(\frac{1}{i!} \right)^\l\right)^{-1}$ can be made arbitrarily close to 1/2 by choosing $\ell$ large enough (even for large $k$). This construction yields a graph for which the ratio of derangements to permutations is arbitrarily close to 1/2 but not exactly 1/2. We will also use this construction, but we will randomly remove some edges. By taking a random subraph we can ``interpolate" between $D_{k, \l}$ (a dense digraph whose ratio of derangements to permutations is close to $1/2$) and a sparse random digraph (whose ratio is $0$). In this paper all asymptotics are as $k \rightarrow \infty$. $\l$ is treated as fixed. We use standard big-O, little-o and $\Omega$ notation. We write $x \sim y$ if $x=(1+o(1))y$. All logarithms are base $e$. \subsection{Proof details} Let the random graph $G_{k,\l}(m)$ be chosen uniformly from among all subgraphs of $D_{k,\l}$ with $m$ edges. We will fix some $p, \l$ and let $m = pk^2 \l$ (so $p$ is the probability that any particular edge of $D_{k,\l}$ becomes an edge of $G_{k,\l}$). Let the random variables $X, Y$ be the number of derangements and permutations in $G_{k,\l}(m)$ respectively. Let $\mc{D}, \mc{P}$ be the collection of all possible derangements and permutations on $D_{k,\l}(m)$. \subsubsection{First moments of $X, Y$} We have \begin{align} \mathbb{E}[X] &= \sum_{D \in \mc{D}}\mathbb{P}[D\subseteq G_{k, \l}]=(k!)^\l\frac{\binom{k^2\l-k \l}{m-k \l} }{\binom{k^2\l}{m}}\nonumber\\ &=(k!)^\l \left(\frac{m}{k^2\ell} \right)^{k\ell}\exp\left\{\frac{k^2 \ell^2}{2}\left(\frac{1}{k^2\ell}-\frac{1}{m}\right)+O\left(\frac{k^3}{m^2} + \frac{k}{m} \right) \right\}\nonumber\\ &\sim (k!)^\l p^{k\ell}\exp\left\{\frac{ \ell}{2}\left(1-\frac{1}{p}\right) \right\}\label{eqn:EX} \end{align} where on the second line we have used the following fact: \begin{fact}\label{fact:edgeprob} \[ \frac{\binom{a-x}{b-x} }{\binom{a}{b}} = \frac{(b)_x}{(a)_x} = \rbrac{\frac ba}^x \exp\cbrac{\frac{x^2}{2}\rbrac{\frac1a - \frac1b} + O\rbrac{\frac{x^3}{b^2} + \frac xb}}. \] \end{fact} For completeness we include the proof although it is well-known. \begin{proof} \begin{align*} \frac{(b)_x}{(a)_x} &= \rbrac{\frac ba}^x \cdot \frac{1 \rbrac{1-\frac 1b}\rbrac{1-\frac 2b} \cdots \rbrac{1-\frac {x-1}b} }{1 \rbrac{1-\frac 1a}\rbrac{1-\frac 2a} \cdots \rbrac{1-\frac {x-1}a}}\\ &= \rbrac{\frac ba}^x \cdot \exp\cbrac{\sum_{i=0}^{x-1} \sbrac{ \ln\rbrac{1-\frac{i}{b} }- \ln\rbrac{1-\frac{i}{a}} }}\\ &= \rbrac{\frac ba}^x \cdot \exp\cbrac{\sum_{i=0}^{x-1} \sbrac{ -\frac ib + \frac ia + O\rbrac{\frac{i^2}{a^2} + \frac{i^2}{b^2}} }}\\ &= \rbrac{\frac ba}^x \cdot \exp\cbrac{ \frac{x(x-1)}{2} \rbrac{\frac1a - \frac1b} + O\rbrac{\frac{x^3}{b^2}} }\\ &= \rbrac{\frac ba}^x \exp\cbrac{\frac{x^2}{2}\rbrac{\frac1a - \frac1b} + O\rbrac{\frac{x^3}{b^2} + \frac xb}}. \end{align*} \end{proof} Before we calculate $\mathbb{E}[Y]$ we introduce a function $f_\l(x)$. For any integer $\l \ge 1$, let \begin{equation}\label{eqn:deff} f_\l(x) := \sum_{i=0}^\infty\frac{ x^{i \l}}{(i!)^\l} . \end{equation} Note that the above power series for $f_\l(x)$ converges for all $x$ and therefore in particular each $f_\l$ is continuous in $x$. We have \begin{align} \mathbb{E}[Y] &= \sum_{P \in \mc{P}}\mathbb{P}[P\subseteq G_{k, \l}]=\sum_{i=0}^k\left( \binom{k}{i}(k-i)!\right)^\l \frac{\binom{k^2\l-(k-i) \l}{m-(k-i) \l} }{\binom{k^2\l}{m} }\nonumber\\ &= \sum_{i=0}^k\left( \frac{k!}{i!}\right)^\l \left(\frac{m}{k^2\ell} \right)^{(k-i) \l} \exp\left\{\frac{(k-i)^2 \l^2}{2}\left(\frac{1}{k^2\ell}-\frac{1}{m}\right)+O\left(\frac{k^3 }{m^2} + \frac km \right) \right\}\nonumber\\ &= (k!)^\l p^{k\l} \sum_{i=0}^k\left( \frac{1}{i!}\right)^\l p^{-i \l} \exp\left\{\frac{ \l}{2}\left(1-\frac{1}{p}\right) + O\rbrac{\frac{i+1}{k}} \right\} \label{eqn:EY1}.\end{align} We split the above sum into two ranges of $i $. Note that for $0 \le i \le \sqrt{k}$ we have \newline $\exp\cbrac{O\rbrac{\frac{i+1}{k}}} = 1+O\rbrac{\frac{1}{\sqrt{k}}}$, while for $\sqrt{k} \le i \le k$ we have $\exp\cbrac{O\rbrac{\frac{i+1}{k}}}=O(1)$. Thus line \eqref{eqn:EY1} becomes \begin{equation} (k!)^\l p^{k\l} \sbrac{ \rbrac{1+ O\rbrac{\frac{1}{\sqrt{k}}}}\exp\left\{\frac{ \l}{2}\left(1-\frac{1}{p}\right) \right\}\sum_{0 \le i \le \sqrt{k}} \left( \frac{1}{i!}\right)^\l p^{-i \l} + O(1)\sum_{\sqrt{k}<i \le k} \left( \frac{1}{i!}\right)^\l p^{-i \l} } \label{eqn:EY2}. \end{equation} As $k \rightarrow \infty$ we have $$\sum_{0 \le i \le \sqrt{k}} \left( \frac{1}{i!}\right)^\l p^{-i \l} \rightarrow f_\l(1/p), $$ and $$\sum_{\sqrt{k} < i \le k} \left( \frac{1}{i!}\right)^\l p^{-i \l} \le \sum_{i=\sqrt{k}}^{\infty} \left( \frac{1}{i!}\right)^\l p^{-i \l} =o(1) $$ since the latter is the tail of a convergent series. Thus, returning to our estimate of $\mathbb{E}[Y]$ on line \eqref{eqn:EY2}, we have \begin{equation} \mathbb{E}[Y] \sim (k!)^\l p^{k\l} \exp\left\{\frac{ \l}{2}\left(1-\frac{1}{p}\right) \right\}f_\l(1/p). \label{eqn:EY} \end{equation} \subsubsection{Choosing $p, \l$} Now that we know $\mathbb{E}[X], \mathbb{E}[Y]$ we will choose $p, \l$ to make sure that the ratio of $\mathbb{E}[X]$ to $\mathbb{E}[Y]$ is close to $r$. Using lines \eqref{eqn:EX} and \eqref{eqn:EY} we have \[ \frac{\mathbb{E}[X]}{\mathbb{E}[Y]} \sim \frac{1}{f_\l\rbrac{\frac1p}}, \] so we would like to choose $\l$ and $0<p<1$ so that $f_\l(1/p) = 1/r$. We have \[ \lim_{x \rightarrow \infty} f_\l(x) = \infty , \qquad f_\l(1) = \sum_{i=0}^k\left( \frac{1}{i!}\right)^\l = 1 + 1 + \frac{1}{2^\l} + \frac{1}{6^\l} + \frac{1}{24^\l} + \ldots. \] Note that we can make $f_\l(1)$ arbitrarily close to 2 by taking $\l$ large. Indeed, we have $f_\l(1) \ge 2$ and \begin{align*} f_\l(1) = 2+ \sum_{i\ge 2}\left( \frac{1}{i!}\right)^\l &\leq 2+ \sum_{i\geq 2} \left( \frac{1}{2^{i-1} }\right)^\ell =2+ \frac{1}{2^\ell - 1}. \end{align*} Since $r<1/2$, we can choose $\l$ so that $f_{\l}(1) < 1/r$. Then by the intermediate value theorem there is some $x \in (1, \infty)$ such that $f_\l(x) = 1/r$. We choose $p$ to be the value $1/x$, so $0<p<1$ and $f_\l(1/p) = 1/r$. So we view $\l$ and $p$ as constants determined entirely by $r$. \subsubsection{Second moments of $X, Y$} In this section we show that $\mathbb{E}[X^2] \sim \mathbb{E}[X]^2$ and $\mathbb{E}[Y^2] \sim \mathbb{E}[Y]^2$. This will complete the proof, since then by the second moment method we have that \[ \frac XY \sim \frac{\mathbb{E}[X]}{\mathbb{E}[Y]} \sim \frac{1}{f_\l(1/p)} = r. \] with probability approaching 1 as $k$ goes to infinity. To help us estimate $\mathbb{E}[X^2], \mathbb{E}[Y^2]$ we will find the function $h(a, b)$ (defined below) useful. Suppose we have some fixed matching $B$ of $b$ many edges in the graph $K_{a, a}$. Then by inclusion-exclusion the number of perfect matchings that do not have any edges from $B$ is \[ h(a, b) := \sum_{w=0}^b (-1)^w \binom{b}{w} (a-w)!. \] Note that we always have $h(a, b) \le a!$. We will now observe that, roughly speaking, $h(a,b)\approx \frac{a!}{e}$ whenever $b \approx a \rightarrow \infty$. More formally we have the following \begin{fact}\label{fact:hest} Suppose $a-a^{1/10} \le b \le a$. Then we have \[ h(a, b) = \rbrac{1 + O(a^{-4/5})} \frac{a!}{e} \] as $a \rightarrow \infty$ \end{fact} \begin{proof} We have \begin{align}\label{eqn:h} h(a,b)=\sum_{0 \le w \le b} (-1)^w \binom{b}{w} (a-w)!= a! \sum_{0 \le w \le b} \frac{(-1)^w}{w!}\frac{(b)_w}{(a)_w}. \end{align} Now, for $0 \le w \le a^{1/10}$ we have by Fact \ref{fact:edgeprob} that \begin{align*} \frac{(b)_w}{(a)_w} &= \rbrac{\frac ba}^w \exp\cbrac{ \frac{w^2}{2} \rbrac{\frac1a - \frac1b} + O\rbrac{\frac{w^3}{b^2} + \frac wb}}\\ &= \rbrac{1 + O\rbrac{a^{-9/10}}}^{O\rbrac{a^{1/10}}} \exp \cbrac{O(a^{-4/5}} = 1 + O(a^{-4/5}). \end{align*} Meanwhile for $w \ge a^{1/10}$ we have that the corresponding term in line \eqref{eqn:h} has absolute value \begin{align*} \frac{1}{w!}\frac{(b)_w}{(a)_w} \le \frac{1}{(a^{1/10})!} = \exp\cbrac{ - \Omega\rbrac{a^{1/10} \log a}} \end{align*} by Stirling's approximation. Thus, the sum of all such terms in line \eqref{eqn:h} is at most \[ b \exp\cbrac{ - \Omega\rbrac{a^{1/10} \log a}} = O(a^{-4/5}) \] (this bound is quite comfortable). By the Alternating Series Test we have that \[ \sum_{0 \le w \le a^{1/10}} \frac{(-1)^w}{w!} = \frac 1e + O\rbrac{\frac{1}{(a^{1/10})!}} = \frac 1e + O(a^{-4/5}). \] Breaking up the sum for $h(a, b)$ we have \begin{align*} h(a, b) &= a! \sbrac{\sum_{0 \le w \le a^{1/10}} \frac{(-1)^w}{w!}\frac{(b)_w}{(a)_w} + \sum_{a^{1/10}< w \le b} \frac{(-1)^w}{w!}\frac{(b)_w}{(a)_w}} \\ &= a! \sbrac{\rbrac{1 + O(a^{-4/5})} \sum_{0 \le w \le a^{1/10}} \frac{(-1)^w}{w!} + O(a^{-4/5}) } \\ &= \rbrac{1 + O(a^{-4/5})} \frac{a!}{e}. \end{align*} \end{proof} We find that \begin{align} \mathbb{E}[X^2] &= \sum_{D,D' \in \mc{D}}\mathbb{P}[D,D'\subseteq G_{k \l}]=(k!)^\l\sum_{D' \in \mc{D}}\mathbb{P}[D_0,D'\subseteq M]\nonumber\\ &=(k!)^\l\sum_{b=0}^{k\l}\left[\frac{\binom{k^2\l-(2k \l-b)}{m-(2k \l-b)} }{\binom{k^2\l}{m} }\sum_{\vec{b} \in S_b}\prod_{c=1}^\l\binom{k}{b_c}h(k-b_c, k-b_c) \right].\label{eqn:EX2} \end{align} where in the inner sum, $S_b$ is the set of $\l$-dimensional vectors $\vec{b} = (b_1, \ldots b_\l)$ whose components are nonnegative integers summing to $b$. By \ref{fact:edgeprob}, if $b \le k^{1/10}$ then we have \begin{align*} \frac{\binom{k^2\l-(2k \l-b)}{m-(2k \l-b)} }{\binom{k^2\l}{m} } &= p^{2k\ell-b}\exp\left\{\frac{(2k\l-b)^2}{2}\left(\frac{1}{k^2\ell}-\frac{1}{m}\right)+O\left(\frac{k^3}{m^2}+\frac{k}{m}\right) \right\} \\ & = \rbrac{1 + O(k^{-9/10})} p^{2k\ell-b}\exp\left\{2\ell \left(1-\frac{1}{p}\right) \right\} \end{align*} and by Fact \ref{fact:hest} we have $ h(k-b_c, k-b_c) =\rbrac{1 + O(k^{-4/5})} \frac{(k-b_c)!}{e}$. Therefore the term corresponding to $b$ in \eqref{eqn:EX2} is \begin{align*} \frac{\binom{k^2\l-(2k \l-b)}{m-(2k \l-b)} }{\binom{k^2\l}{m} }\sum_{\vec{b} \in S_b}\prod_{c=1}^\l & \binom{k}{b_c}h(k-b_c, k-b_c)\\ &= \rbrac{1 + O(k^{-4/5})}p^{2k\ell-b}\exp\left\{2\ell \left(1-\frac{1}{p}\right) \right\} \sum_{\vec{b}\in S_b}\prod_{c=1}^\l\binom{k}{b_c} \frac{(k-b_c)!}{e} \\ &= \rbrac{1 + O(k^{-4/5})}(k!)^{\l} p^{2k\ell-b}\exp\left\{2\ell \left(1-\frac{1}{p}\right) - \l \right\} \sum_{\vec{b}\in S_b} \prod_{c=1}^\l \frac{1}{b_c!} \\ &=\rbrac{1 + O(k^{-4/5})}(k!)^{\l} p^{2k\ell-b}\exp\left\{\ell \left(1-\frac{2}{p}\right) \right\} \frac{\l^b}{b!} \end{align*} where on the last line we used the multinomial formula. Meanwhile if $b \ge k^{1/10}$ then the term corresponding to $b$ in \eqref{eqn:EX2} is \begin{align*} \frac{\binom{k^2\l-(2k \l-b)}{m-(2k \l-b)} }{\binom{k^2\l}{m} }\sum_{\vec{b} \in S_b}\prod_{c=1}^\l & \binom{k}{b_c}h(k-b_c, k-b_c) \le \sum_{\vec{b}\in S_b}\prod_{c=1}^\l\binom{k}{b_c} (k-b_c)!\\ & = (k!)^\l \frac{\l^b}{b!} = (k!)^\l \cdot \exp\cbrac{-\Omega\rbrac{k^{1/10} \log k}} \end{align*} and so the sum of all terms in \eqref{eqn:EX2} with $b \ge k^{1/10}$ is at most \[ k\l \cdot (k!)^\l \cdot \exp\cbrac{-\Omega\rbrac{k^{1/10} \log k}} = (k!)^\l \cdot O\rbrac{k^{-4/5}}. \] Therefore \begin{align*} \mathbb{E}[X^2] &=(k!)^\l\sum_{b=0}^{k\l}\left[\frac{\binom{k^2\l-(2k \l-b)}{m-(2k \l-b)} }{\binom{k^2\l}{m} }\sum_{\vec{b} \in S_b}\prod_{c=1}^\l\binom{k}{b_c}h(k-b_c, k-b_c) \right]\\ &= (k!)^\l \sbrac{\sum_{0 \le b \le k^{1/10} } \rbrac{1 + O(k^{-4/5})}(k!)^{\l} p^{2k\ell-b}\exp\left\{\ell \left(1-\frac{2}{p}\right) \right\} \frac{\l^b}{b!} + (k!)^\l \cdot O\rbrac{k^{-4/5}}}\\ & = \rbrac{1 + O(k^{-4/5})} (k!)^{2\l}p^{2k\l}\exp\left\{\ell \left(1-\frac{2}{p}\right) \right\} \sum_{0 \le b \le k^{1/10} } p^{-b} \frac{\l^b}{b!}\\ & = \rbrac{1 + O(k^{-4/5})} (k!)^{2\l}p^{2k\l}\exp\left\{\ell \left(1-\frac{2}{p}\right) \right\}\cdot \rbrac{\exp\left\{\frac{\l}{p} \right\} + O(k^{-4/5}) } \\ & = \rbrac{1 + O(k^{-4/5})} (k!)^{2\l}p^{2k\l}\exp\left\{\ell \left(1-\frac{1}{p}\right) \right\} \\ & \sim \mathbb{E}[X]^2 \end{align*} \vspace{1cm} For $\mathbb{E}[Y^2]$ we find an exact expression to be cumbersome, but the following upper bound will suffice: \begin{align} \mathbb{E}[Y^2] &\le \sum_{\substack{0 \le i, j \le k\\ 0 \le b \le k \l \\ \vec{b} \in S_b}} \frac{\binom{k^2\ell-(2k\ell-(i+j)\ell-b)}{m-(2k\ell-(i+j)\ell-b)} }{\binom{k^2\ell}{m} } \left(\frac{k!}{i!}\right)^\ell \prod_{c=1}^\ell \binom{k-i}{b_c}\binom{k-b_c}{j} h(k-j-b_c, k-i-b_c-2j) \label{eqn:EYsquared} \end{align} The term corresponding to a tuple $(i, j, b, \vec{b})$ above is an upper bound on the contribution to $\mathbb{E}[Y^2]$ due to pairs of permutations $(P, P')$ such that $P$ fixes $i$ vertices per part, $P'$ fixes $j$ vertices per part, and $P$ and $P'$ share a total of $b$ edges where $b_c$ of the shared edges are between $V_c$ and part $V_{c+1}$. The first factor is the edge probability, and the next factor is the number of choices for $P$. The next factor is an upper bound on the number of choices for $P'$. Indeed, we choose the edges of $P'$ from $V_c$ to $V_{c+1}$ by first choosing $b_c$ edges of $P$ to be shared, then we choose $j$ vertices in $V_c$ to be fixed by $P'$, and finally we choose a matching between the remaining vertices (the vertices of $V_c \cup V_{c+1}$ that are not fixed by $P'$ and are not endpoints of the $b_c$ shared edges already chosen). This matching must avoid any edges of $P$, and the vertices to be matched induce at least $k-i-b_c-2j$ edges of $P$, explaining the last factor above. We will now estimate the significant terms in \eqref{eqn:EYsquared}. Assume $i, j, b \le k^{1/10}$. Then by Fact \ref{fact:edgeprob} \begin{align*} \frac{\binom{k^2\ell-(2k\ell-(i+j)\ell-b)}{m-(2k\ell-(i+j)\ell-b)} }{\binom{k^2\ell}{m} } &= p^{2k\ell-(i+j)\ell-b}\exp\left\{\frac{(2k\ell-(i+j)\ell-b)^2}{2k^2\l}\rbrac{1-\frac{1}{p}} +O\rbrac{\frac{1}{k}}\right\} \\ &= p^{2k\ell-(i+j)\ell-b}\exp\left\{2\ell\left(1-\frac{1}{p}\right) +O\rbrac{k^{-9/10}}\right\}. \end{align*} Next we estimate \begin{align*} h(k-j-b_c, k-i-b_c-2j) = \rbrac{1+O\rbrac{k^{-4/5}}} \frac{(k-j-b_c)!}{e} \end{align*} by Fact \ref{fact:hest}. So the product in \eqref{eqn:EYsquared} is \begin{align} \prod_{c=1}^\ell \binom{k-i}{b_c}\binom{k-b_c}{j} h(k-j-b_c, k-i-b_c-2j) & = \rbrac{1+O\rbrac{k^{-4/5}}} \prod_{c=1}^\ell \frac{(k-i)_{b_c}}{b_c!}\frac{(k-b_c)_j}{j!} \frac{(k-j-b_c)!}{e} \nonumber \\ & \le \rbrac{1+O\rbrac{k^{-4/5}}}\rbrac{\frac{k!}{ej!}}^\l \prod_{c=1}^\ell \frac{1}{b_c!}.\label{eqn:prodest1} \end{align} The sum of terms in \eqref{eqn:EYsquared} corresponding to small $i, j, b$ is at most \begin{align} &\rbrac{1+O\rbrac{k^{-4/5}}} \sum_{\substack{0 \le i, j, b \le k^{1/10} \\ \vec{b} \in S_b}} p^{2k\ell-(i+j)\ell-b}\exp\left\{2\ell\left(1-\frac{1}{p}\right) \right\}. \left(\frac{k!}{i!}\right)^\ell \rbrac{\frac{k!}{ej!}}^\l \prod_{c=1}^\ell \frac{1}{b_c!}\nonumber \\ & =\rbrac{1+O\rbrac{k^{-4/5}}} \exp\left\{\ell\left(1-\frac{2}{p}\right) \right\}\sum_{0 \le i, j, b \le k^{1/10}} p^{2k\ell-(i+j)\ell-b} \left(\frac{k!}{i!}\right)^\ell \rbrac{\frac{k!}{j!}}^\l \frac{\l^b}{b!}\nonumber \\ & \le \rbrac{1+O\rbrac{k^{-4/5}}} (k!)^{2\l} p^{2k\l}\exp\left\{\ell\left(1-\frac{1}{p}\right) \right\}\sum_{0 \le i, j , b} \frac{p^{-i\ell} }{(i!)^\l} \cdot \frac{p^{-j\ell} }{(j!)^\l} \cdot \frac{(\l/p)^b}{b!} \nonumber\\ &= \rbrac{1+O\rbrac{k^{-4/5}}} (k!)^{2\l} p^{2k\l}\exp\left\{\ell\left(1-\frac{1}{p}\right) \right\} f_\l\rbrac{\frac1p}^2 \nonumber\\ &\sim \mathbb{E}[Y]^2 \nonumber \end{align} where on the second-to-last line we have used \[ \sum_{0 \le i} \frac{p^{-i\ell} }{(i!)^\l} = f_\l\rbrac{\frac1p}, \qquad \qquad \sum_{0 \le b } \frac{(\l/p)^b}{b!} = \exp\cbrac{\frac {\l}{p}}. \] It remains to show that the sum of all other terms (i.e. terms where $i, j,$ or $b$ is at least $k^{1/10}$) is negligible compared to $\mathbb{E}[Y]^2$, which is of order $(k!)^{2\l} p^{2k\l}$. Note that by Fact \ref{fact:edgeprob} \begin{align*} \frac{\binom{k^2\ell-(2k\ell-(i+j)\ell-b)}{m-(2k\ell-(i+j)\ell-b)} }{\binom{k^2\ell}{m} } &= p^{2k\ell-(i+j)\ell-b}\exp\left\{\frac{(2k\ell-(i+j)\ell-b)^2}{2k^2\l}\rbrac{1-\frac{1}{p}} +O\rbrac{\frac{1}{k}}\right\} \\ &= O\rbrac{p^{2k\ell-(i+j)\ell-b}}. \end{align*} Thus, the sum (over $\vec{b}$) of terms corresponding to a fixed triple $(i, j, b)$ in line \eqref{eqn:EYsquared} big-O of \begin{align*} & p^{2k\ell-(i+j)\ell-b} \sum_{\vec{b} \in S_b} \left(\frac{k!}{i!}\right)^\ell \prod_{c=1}^\ell \binom{k-i}{b_c}\binom{k-b_c}{j} (k-j-b_c)! \nonumber\\ & \le p^{2k\ell-(i+j)\ell-b} \left(\frac{k!}{i!}\right)^\ell \left(\frac{k!}{j!}\right)^\ell \sum_{\vec{b} \in S_b} \prod_{c=1}^\ell \frac{1}{b_c!} \nonumber\\ & =\rbrac{(k!)^{2\l} p^{2k\l} } \cdot \rbrac{ \frac{\l^b}{p^{(i+j)\ell+b}(i!)^\ell(j!)^\l b!} }. \end{align*} It is easy to see that if $i, j$ or $b$ is at least $k^{1/10}$ then the second factor above is $\exp\cbrac{-\Omega\rbrac{k^{1/10} \log k}}$. Since the number of triples $(i, j, b)$ is polynomial in $k$, the sum of all such terms (i.e. where $i, j$ or $b$ is at least $k^{1/10}$) is $o\rbrac{(k!)^{2\l} p^{2k\l} }$ which is a negligible contribution to $\mathbb{E}[Y^2]$. Therefore we have $\mathbb{E}[Y^2]\sim \mathbb{E}[Y]^2.$ \section{Remarks and Open Problems} The reader should note that we did not use a ``binomial" random construction (e.g. keep each edge of $D_{k, \ell}$ with probability $p$ independently) because such a model lacks the concentration we need here. Indeed, for example Janson (\cite{J94}) showed that the number of perfect matchings in $G(n, p)$ is not concentrated even when it is quite large, while the number of perfect matchings of $G(n, m)$ is concentrated. We tried to use a binomial random construction and found that the second moments were too large, which in light of Janson's result makes sense (for example derangements in our graph are just a union of several perfect matchings on bipartite graphs). There are still interesting open problems in \cite{BDHHL}. In particular it is still open whether $S$, the set of possible ratios $(d/p)_G$, is equal to $\mathbb{Q} \cap [0, 1/2]$. Here we would like to pose another open problem that is mostly unrelated to our result. In particular, we ask about stability for digraphs whose ratio $(d/p)_G$ is close to $1/2$: do such digraphs have to resemble the blow-up of a directed graph? \bibliographystyle{abbrv}

https://arxiv.org/abs/1302.2254

Strengthened Cauchy-Schwarz and Hölder inequalities

We present some identities related to the Cauchy-Schwarz inequality in complex inner product spaces. A new proof of the basic result on the subject of Strengthened Cauchy-Schwarz inequalities is derived using these identities. Also, an analogous version of this result is given for Strengthened Hölder inequalities.

\section{Introduction} In \cite{A}, the parallelogram identity in a {\em real} inner product space, is rewritten in Cauchy-Schwarz form (with the deviation from equality given as a function of the angular distance between vectors) thereby providing another proof of the Cauchy-Schwarz inequality in the real case. The first section of this note complements this result by presenting related identities for complex inner product spaces, and thus a proof of the Cauchy-Schwarz inequality in the complex case. Of course, using angular distances is equivalent to using angles. An advantage of the angular distance is that it makes sense in arbitrary normed spaces, in addition to being simpler than the notion of angle. And in some cases it may also be easier to compute. Angular distances are used in Section 2 to give a proof of the basic theorem in the subject of Strengthened Cauchy-Schwarz inequalities (Theorem \ref{scs} below). We also point out that the result is valid not just for vector subspaces, but also for cones. Strengthened Cauchy-Schwarz inequalities are fundamental in the proofs of convergence of iterative, finite element methods in numerical analysis, cf. for instance \cite{EiVa}. They have also been considered in the context of wavelets, cf. for example \cite{DR}, \cite{DRRo1}, \cite{DRRo2}. Finally, Section 3 presents a variant, for cones and in the H\"older case when $1 < p < \infty$, of the basic theorem on Strengthened Cauchy-Schwarz inequalities, cf. Theorem \ref{sh}. \section{Identities related to the Cauchy-Schwarz inequality in complex inner product spaces} It is noted in \cite{A} that in a {\em real} inner product space, the parallelogram identity \begin{equation}\label{par} \|x + y\|^2 + \|x-y\|^2 = 2 \|x\|^2 + 2\|y\|^2 \end{equation} provides the following stability version of the Cauchy-Schwarz inequality, valid for non-zero vectors $x$ and $y$: \begin{equation}\label{bonhilb} (x, y) = \|x\|\|y\|\left(1 - \frac12 \left\|\frac{x}{\|x\|}-\frac{y}{\|y\|}\right\|^2\right). \end{equation} Basically, this identity says that the size of $(x,y)$ is determined by the angular distance $\left\|\frac{x}{\|x\|}-\frac{y}{\|y\|}\right\|$ between $x$ and $y$. In particular, $(x, y) \le \|x\|\|y\|$, with equality precisely when the angular distance is zero. In this section we present some complex variants of this identity, involving $(x,y)$ and also $|(x,y)|$; as a byproduct, the Cauchy-Schwarz inequality in the complex case is obtained. Since different conventions appear in the literature, we point out that in this paper $(x,y)$ is taken to be linear in the first argument and conjugate linear in the second. We systematically replace in the proofs nonzero vectors $x$ and $y$ by unit vectors $u = x/\|x\|$ and $v = y/\|y\|$. \begin{theorem}\label{l1} For all nonzero vectors $x$ and $y$ in a complex inner product space, we have \begin{equation}\label{re} \operatorname{Re} (x, y) = \|x\|\|y\|\left(1 - \frac12 \left\|\frac{x}{\|x\|}-\frac{y}{\|y\|}\right\|^2\right) \end{equation} and \begin{equation}\label{im} \operatorname{Im} (x, y) = \|x\|\|y\|\left(1 - \frac12 \left\|\frac{x}{\|x\|}-\frac{i y}{\|y\|}\right\|^2\right). \end{equation} \end{theorem} \begin{proof} Let $\|u\| =\|v\| = 1$. From (\ref{par}) we obtain \begin{equation*} 4 - \|u-v\|^2 = \|u + v\|^2 = 2 + (u,v) + (v, u) = 2 + (u,v) + \overline{(u, v)} = 2 + 2 \operatorname{Re} (u,v). \end{equation*} Thus, $ \operatorname{Re} (u,v) = 1 - \frac12 \left\|u - v \right\|^2. $ The same argument, applied to $\|u + i v\|^2$, yields $ \operatorname{Im} (u,v) = 1 - \frac12 \left\|u - i v \right\|^2. $ \end{proof} Writing $(x,y) = \operatorname{Re} (x,y)+ i \operatorname{Im} (x,y)$ we obtain the following \begin{corollary}\label{complexCauchy} For all nonzero vectors $x$ and $y$ in a complex inner product space, we have \begin{equation}\label{bonchilb1} (x, y) = \|x\|\|y\|\left(\left(1 - \frac12 \left\|\frac{x}{\|x\|}-\frac{y}{\|y\|}\right\|^2\right) + i \left(1 - \frac12 \left\|\frac{x}{\|x\|}-\frac{i y}{\|y\|}\right\|^2\right)\right). \end{equation} Thus, \begin{equation}\label{bonchilb2} |(x, y)| = \|x\|\|y\|\sqrt{\left(1 - \frac12 \left\|\frac{x}{\|x\|}-\frac{y}{\|y\|}\right\|^2\right)^2 + \left(1 - \frac12 \left\|\frac{x}{\|x\|}-\frac{i y}{\|y\|}\right\|^2\right)^2}. \end{equation} \end{corollary} Next we find some shorter expressions for $|(x,y)|$. Let $\operatorname{Arg} z$ denote the principal argument of $z\in \mathbb{C}$, $z \ne 0$. That is, $0\le \operatorname{Arg} z < 2\pi$, and in polar coordinates, $z= e^{i\operatorname{Arg} z} r$. We choose the principal argument for definiteness; any other argument will do equally well. \begin{theorem} Let $x$ and $y$ be nonzero vectors in a complex inner product space. Then, for every $\alpha\in\mathbb{R}$ we have \begin{equation} \label{stabCauchy} \|x\|\|y\|\left(1 - \frac12 \left\|\frac{e^{i\alpha}x}{\|x\|}-\frac{y}{\|y\|}\right\|^2\right) \le |(x,y)| = \|x\|\|y\|\left(1 - \frac12 \left\|\frac{e^{ - i \operatorname{Arg} (x,y)} x}{\|x\|}-\frac{y}{\|y\|}\right\|^2\right). \end{equation} \end{theorem} \begin{proof} By a normalization, it is enough to consider unit vectors $u$ and $v$. Let $\alpha$ be an arbitrary real number, and set $t = \operatorname{Arg} (u,v)$, so $(u,v) = e^{it} r$ in polar form. Using (\ref{re}) we obtain $$ 1 - \frac12 \left\|e^{i\alpha} u - v \right\|^2 = \operatorname{Re} (e^{i\alpha} u,v) \le |(e^{i\alpha} u,v)| $$ $$ = |(u,v)| = r = (e^{-it} u,v) = \operatorname{Re} (e^{-it} u,v) = 1 - \frac12 \left\|e^{-it} u - v \right\|^2. $$ \end{proof} The preceding result can be regarded as a variational expression for $|(x,y)|$, since it shows that this quantity can be obtained by maximizing the left hand side of (\ref{stabCauchy}) over $\alpha$, or, in other words, by minimizing $\left\|\frac{e^{i\alpha}x}{\|x\|}-\frac{y}{\|y\|}\right\|$ over $\alpha$. \begin{corollary} Cauchy-Schwarz inequality. For all vectors $x$ and $y$ in a complex inner product space, we have $|(x,y)| \le \left\|x\right\| \|y\|$, with equality if and only if the vectors are linearly dependent. \end{corollary} \begin{proof} Of course, if one of the vectors $x$, $y$ is zero, the result is trivial, so suppose otherwise and normalize, writing $u = x/\|x\|$ and $v = y/\|y\|$. From (\ref{stabCauchy}) we obtain, first, $|(u,v)| \le 1$, second, $e^{-i\operatorname{Arg} (u,v)} u = v $ if $|(u,v)| = 1$, so equality implies linear dependency, and third, $|(u,v)| = 1$ if $e^{i\alpha} u = v$ for some $\alpha\in \mathbb{R}$, so linear dependency implies equality. \end{proof} \section{Strengthened Cauchy-Schwarz inequalities} Such inequalities, of the form $|(x,y)| \le \gamma \left\|x\right\| \|y\|$ for some fixed $\gamma \in [0,1)$, are fundamental in the proofs of convergence of iterative, finite element methods in numerical analysis. The basic result in the subject is the following theorem (see Theorem 2.1 and Remark 2.3 of \cite{EiVa}). \begin{theorem}\label{scs} Let $H$ be a Hilbert space, let $F\subset H$ be a closed subspace, and let $V\subset H$ be a finite dimensional subspace. If $F\cap V = \{0\}$, then there exists a constant $\gamma = \gamma (V, F)\in [0,1)$ such that for every $x\in V$ and every $y\in F$, $$ |(x,y)| \le \gamma \left\|x\right\| \|y\|. $$ \end{theorem} There are, at least, two natural notions of angle between subspaces. To see this, consider a pair of 2 dimensional subspaces $V$ and $W$ in $\mathbb{R}^3$, in general position. They intersect in a line $L$, so we may consider that they are parallel in the direction of the subspace $L$, and thus the angle between them is zero. This is the notion of angle relevant to the subject of Strengthened Cauchy-Schwarz inequalities. Alternatively, we may disregard the common subspace $L$, and (in this particular example) determine the angle between subspaces by choosing the minimal angle between their unit normals. Note however that the two notions of angle suggested by the preceding example coincide when the intersection of subspaces is $\{0\}$ (cf. \cite{De} for more information on angles between subspaces). From the perspective of angles, or equivalently, angular distances, what Theorem \ref{scs} states is the intuitively plausible assertion that the angular distance between $V$ and $F$ is strictly positive provided $F$ is closed, $V$ is finite dimensional, and $F\cap V = \{0\}$. Finite dimensionality of one of the subspaces is crucial, though: It is known that if both $V$ and $F$ are infinite dimensional, the angular distance between them can be zero, even if both subspaces are closed. Define the angular distance between $V$ and $F$ as \begin{equation} \label{angdist} \kappa(V, F) := \inf\{\|v - w\|: v\in V, w\in F, \mbox{ and } \|v\| = \|w\|=1\}. \end{equation} The proof (by contradiction) of Theorem \ref{scs} presented in \cite{EiVa} is not difficult, but deals only with the case where both $V$ and $F$ are finite dimensional. And it is certainly not as simple as the following \begin{proof} If either $V=\{0\}$ or $F=\{0\}$ there is nothing to show, so assume otherwise. Let $S(V)$ be the unit sphere of the finite dimensional subspace $V$, and let $v\in S$. Denote by $f(v)$ the distance from $v$ to the unit sphere $S(F)$ of $F$. Then $f(v) > 0$ since $F$ is closed and $v\notin F$. Thus, $f$ achieves a minimum value $\kappa > 0$ over the compact set $S(V)$. By the right hand side of formula (\ref{stabCauchy}), for every $x\in V\setminus \{0\}$ and every $y\in F\setminus \{0\}$ we have $|(x,y)| \le (1 - \kappa^2/2) \left\|x\right\| \|y\|$. \end{proof} In concrete applications of the Strengthened Cauchy-Schwarz inequality, a good deal of effort goes into estimating the size of $\gamma = \cos \theta$, where $\theta$ is the angle between subspaces appearing in the discretization schemes. Since we also have $\gamma = 1 - \kappa^2/2$, this equality can provide an alternative way of estimating $\gamma$, via the angular distance $\kappa$ rather than the angle. Next we state a natural extension of Theorem \ref{scs}, to which the same proof applies (so we will not repeat it). Consider two nonzero vectors $u$, $v$ in a real inner product space $E$, and let $S$ the unit circumference in the plane spanned by these vectors. The angle between them is just the length of the smallest arc of $S$ determined by $u/\|u\|$ and $v/\|v\|$. So to speak about angles, or angular distances, we only need to be able to multiply nonzero vectors $x$ by positive scalars $\lambda = 1/\|x\|$. This suggests that the natural setting for Theorem \ref{scs} is that of cones, rather than vector subspaces. Recall that $C$ is a {\em cone} in a vector space over a field containing the real numbers if for every $x\in C$ and every $\lambda > 0$ we have $\lambda x\in C$. In particular, every vector subspace is a cone. If $C_1$ and $C_2$ are cones in a Hilbert space, the angular distance between them can be defined exactly as before: \begin{equation} \label{angdistc} \kappa(C_1, C_2) := \inf\{\|v - w\|: v\in C_1, w\in C_2, \mbox{ and } \|v\| = \|w\|=1\}. \end{equation} \begin{theorem}\label{scsc} Let $H$ be a Hilbert space with unit sphere $S(H)$, and let $C_1, C_2\subset H$ be (topologically) closed cones, such that $C_1\cap S(H)$ is a norm compact set. If $C_1\cap C_2 = \{0\}$, then there exists a constant $\gamma = \gamma (C_1, C_2)\in [0,1)$ such that for every $x\in C_1$ and every $y\in C_2$, $$ |(x,y)| \le \gamma \left\|x\right\| \|y\|. $$ \end{theorem} \begin{example} Let $H=\mathbb{R}^2$, $C_1 =\{(x,y) \in \mathbb{R}^2: x= -y\}$ and $C_2 =\{(x,y) \in \mathbb{R}^2: x y\ge 0\}$, that is, $C_1$ is the one dimensional subspace with slope $-1$ and $C_2$ the union of the first and third quadrants. Here we can explicitly see that $\gamma (C_1, C_2) = \cos (\pi/4) = 1/\sqrt2$. However, if $C_2$ is extended to a vector space $V$, then the condition $C_1\cap V = \{0\}$ no longer holds and $\gamma (C_1, V) = 1$. So stating the result in terms of cones rather than vector subspaces does cover new, nontrivial cases. \end{example} \section{A Strengthened H\"older inequality} For $1 < p < \infty$, it is possible to give an $L^p-L^q$ version of the Strengthened Cauchy-Schwarz inequality. Here $q:= p/(p - 1)$ denotes the conjugate exponent of $p$. We want to find suitable conditions on $C_1\subset L^p$ and $C_2\subset L^q$ so that there exists a constant $\gamma = \gamma (C_1, C_2)\in [0,1)$ with $ \|fg\|_1 \le \gamma \left\|f\right\|_p \|g\|_q $ for every $f\in C_1$ and every $g\in C_2$. An obvious difference between the H\"older and the Cauchy-Schwarz cases is that in the pairing $(f, g): = \int f\overline{g}$, the functions $f$ and $g$ belong to different spaces (unless $p = q =2$). This means that the hypothesis $C_1\cap C_2 = \{0\}$ needs to be modified. A second obvious difference is that H\"older's inequality deals actually with $|f|$ and $|g|$ rather than with $f$ and $g$. So when finding angular distances we will also deal with $|f|$ and $|g|$. Note that $f\in C_i$ does not necessarily imply that $|f|\in C_i$ (consider, for instance, the second quadrant in $\mathbb{R}^2$). We make standard nontriviality assumptions on measure spaces $(X, \mathcal{A}, \mu)$: $X$ contains at least one point and the (positive) measure $\mu$ is not identically zero. We write $L^p$ rather than $L^p (X, \mathcal{A}, \mu)$. To compare cones in different $L^p$ spaces, we map them into $L^2$ via the Mazur map. Let us write $\operatorname{ sign } z = e^{i\theta}$ when $z = r e^{i\theta} \ne 0$, and $\operatorname{ sign } 0 = 1$ (so $|\operatorname{ sign } z| = 1$ always). The Mazur map $\psi_{r,s} : L^r\to L^s$ is defined first on the unit sphere $S (L^r)$ by $\psi_{r,s} (f) := |f|^{r/s} \operatorname{ sign }f$, and then extended to the rest of $L^r$ by homogeneity (cf. \cite{BeLi}, pp. 197--199 for additional information on the Mazur map). More precisely, $$ \psi_{r,s} (f) := \|f\|_r \psi_{r,s} (f/\|f\|_r) = \|f\|_r^{1 - r/s} |f|^{r/s} \operatorname{ sign }f. $$ By definition, if $\lambda > 0$ then $\psi_{r,s} (\lambda f) = \lambda \psi_{r,s} (f)$. This entails that if $C\subset L^r$ is a cone, then $\psi_{r,s} (C)\subset L^s$ is a cone. Given a subset $A\subset L^r$, we denote by $|A|$ the set $|A| :=\{|f|: f\in A\}$. Observe that if $A$ is a cone then so is $|A|$. \begin{theorem}\label{sh} Let $1 < p < \infty$ and denote by $q:= p/(p - 1)$ its conjugate exponent. Let $C_1\subset L^p$ and $C_2\subset L^q$ be cones, let $S(L^p)$ stand for the unit sphere of $L^p$ and let $\overline{|C_1|}$ and $\overline{|C_2|}$ denote the topological closures of $|C_1|$ and $|C_2|$. If $\overline{|C_1|}\cap S(L^p)$ is norm compact, and $\psi_{p,2} (\overline{|C_1|}) \cap \psi_{q,2} (\overline{|C_2|}) = \{0\}$, then there exists a constant $\gamma = \gamma (C_1, C_2)\in [0,1)$ such that for every $f\in C_1$ and every $g\in C_2$, \begin{equation}\label{sth} \|fg\|_1 \le \gamma \left\|f\right\|_p \|g\|_q. \end{equation} \end{theorem} We use the part of \cite[Theorem 2.2]{A} given next. \begin{theorem}\label{betterhold} Let $1 < p < \infty$, let $q = p/(p-1)$ be its conjugate exponent, and let $M = \max\{p, q\}$. If $f\in L^p $, $g\in L^q$, and $\|f\|_p, \|g\|_q > 0$, then \begin{equation}\label{bonhold} \|fg\|_1 \le \|f\|_p\|g\|_q \left(1 - \frac1M \left\|\frac{|f|^{p/2}}{\|f\|_p^{p/2}}-\frac{|g|^{q/2}}{\|g\|_q^{q/2}}\right\|_2^2\right). \end{equation} \end{theorem} A different proof of inequality (\ref{bonhold}) (with the slightly weaker constant $M = p + q$, but sufficient for the purposes of this note) can be found in \cite{A2}. Next we prove Theorem \ref{sh}. \begin{proof} If either $C_1=\{0\}$ or $C_2=\{0\}$ there is nothing to show, so assume otherwise. Note that since $|C_1|$ and $|C_2|$ are cones, the same happens with their topological closures. The cones $\psi_{p,2} (\overline{|C_1|})$ and $\psi_{q,2} (\overline{|C_2|})$ are also closed, as the following argument shows: The Mazur maps $\psi_{r,s}$ are uniform homeomorphisms between closed balls, and also between spheres, of any fixed (bounded) radius (cf. \cite[Proposition 9.2, p. 198]{BeLi}, and the paragraph before the said proposition). In particular, if $\{f_n\}$ is a Cauchy sequence in $\psi_{q,2} (\overline{|C_2|})$ (for instance) then it is a bounded sequence in $L^2$, so $\psi_{q,2}^{-1} = \psi_{2,q}$ maps it to a Cauchy sequence in $\overline{|C_2|}$, with limit, say, $h$. Then $\lim_n f_n = \psi_{q,2} (h) \in \psi_{q,2} (\overline{|C_2|})$. Likewise, $\psi_{p,2} (\overline{|C_1|})$ is closed. The rest of the proof proceeds as before. Let $v\in \psi_{p,2} (\overline{|C_1|})\cap S(L^2)$ and denote by $F(v)$ the distance from $v$ to $\psi_{q,2} (\overline{|C_2|})\cap S(L^2)$. Then $F(v) > 0$, so $F$ achieves a minimum value $\kappa > 0$ over the compact set $\psi_{p,2} (\overline{|C_1|})\cap S(L^2)$, and now (\ref{sth}) follows from (\ref{bonhold}). \end{proof}

https://arxiv.org/abs/2112.02556

Windmills of the minds: an algorithm for Fermat's Two Squares Theorem

The two squares theorem of Fermat is a gem in number theory, with a spectacular one-sentence "proof from the Book". Here is a formalisation of this proof, with an interpretation using windmill patterns. The theory behind involves involutions on a finite set, especially the parity of the number of fixed points in the involutions. Starting as an existence proof that is non-constructive, there is an ingenious way to turn it into a constructive one. This gives an algorithm to compute the two squares by iterating the two involutions alternatively from a known fixed point.

\section{Introduction} \label{sec:introduction} Fermat's two squares theorem, dated back to 1640, states that a prime $n$ that is one more than a multiple of $4$ can be uniquely expressed as a sum of odd and even squares (Section~\ref{sec:sum-of-two-squares}, Theorem~\ref{thm:fermat-two-squares-thm}). Of the many proofs of this classical number theory result, this one-sentence proof by Zagier~\cite{Zagier-1990-acm} caused a sensation in 1990: \bigskip \begin{mquote} \textit{ \\ The involution on the finite set \[ S = \{(x,y,z) \in \mathbb{N}^{3} \mid \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}} \} \] defined by \begin{equation} \label{eqn:zagier-map} \HOLinline{(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})} \longmapsto \begin{cases} (x + 2 z,z,y - z - x) & \text{if }x < y - z \\ (2 y - x,y,x + z - y) & \text{if }y - z < x < 2y\\ (x - 2 y,x + z - y,y) & \text{if }x > 2y \end{cases} \end{equation} has exactly one fixed point, so \HOLinline{\ensuremath{|}\HOLFreeVar{S}\ensuremath{|}} is odd, and the involution defined by $\HOLinline{(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})} \longmapsto \HOLinline{(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{z}\HOLSymConst{,}\HOLFreeVar{y})}$ also has a fixed point. \\ } \end{mquote} Those who are perplexed by this multi-line sentence are not alone. Even knowing involution, a self-inverse function, and fixed points, those values unchanged by a function, the proof is not obvious at a glance! Listed as number 20 in Formalizing 100 Theorems~\cite{Wiedijk-2020}, there are many formal proofs of this theorem. Some are based on textbook proofs, others follow the ideas in Zagier's proof. All show the existence of the two squares, only a few (Coq~\cite{Thery-2004} and Lean~\cite{Hughes-2019}) include the uniqueness part. Therefore, a formalisation of this one-sentence proof, in a constructive way, is an interesting exercise in theorem-proving. As a bonus, the exercise is a path of discovery due to recent progress in understanding this proof. As Don Zagier remarked after the one sentence, his proof was a condensed version of a 1984 proof by Roger Heath-Brown~\cite{Heath-Brown-1984-acm}, who in turn acknowledged prior work in number theory taken up by Joseph Liouville~\cite{Williams-2010-alt}. This one-sentence proof invokes two involutions: the second one is obvious, but the first one in Equation~\eqref{eqn:zagier-map} has been called ``black magic''~\cite{Trimble-Lama-2008}. The algebraic formulation of this involution has been given a geometric interpretation by Alexander Spivak~\cite{Spivak-2007} in 2007. These are the windmills (Section~\ref{sec:windmills}). They explain why the magic works, and suggest an interplay of the involutions to identify fixed points of each other. Moreover, this provides an algorithm to find the two squares in Fermat's theorem. Thus the one-sentence proof can be made constructive, as elucidated by Zagier~\cite{Zagier-2013-acm} in 2013. \subsection{Contribution} \label{sec:contribution} This paper gives the first formal proof of an algorithm to compute the two squares in Fermat's two squares theorem, by following a constructive version of Zagier's proof in HOL4. As noted before, Zagier's proof has been formalised, in HOL Light~\cite{Harrison-2010}, in NASA PVS~\cite{Narkawicz-2012-acm} and in Coq~\cite{Dubach-Muehlboeck-2021-acm}, although not in this constructive form. All the ideas used in this paper can be found in Shiu~\cite{Shiu-1996} and Zagier~\cite{Zagier-2013-acm}. The novel feature of this work is an elegant and pictorial approach for our formalisation. The emphasis is in providing formal definitions and developing appropriate theories, not only for the present work, but also for supporting further work. \subsection{Overview} \label{sec:overview} Major features in this formalisation are: \begin{itemize}[leftmargin=*] \item the groundwork for Zagier's proof in Section~\ref{sec:sum-of-two-squares}, \item the two involutions for windmills in Section~\ref{sec:windmill-involutions}, \item the existence and uniqueness of two squares in Section~\ref{sec:two-squares-theorem}, \item an algorithm to compute the two squares in Section~\ref{sec:algorithm}, \item theories of involutions and iterations in Section~\ref{sec:orbits}, and \item a correctness proof of our algorithm in Section~\ref{sec:correctness}. \end{itemize} After a review of the work done, we conclude in Section~\ref{sec:conclusion}. \subsection{Notation} \label{sec:notations} Statements starting with a turnstile ($\vdash$) are HOL4 theorems, automatically pretty-printed to \LaTeX{} from the relevant \text{theory} in the HOL4 development. Generally, our notation allows an appealing combination of quantifiers ($\forall, \exists, \exists{!}$), logical connectives (\HOLTokenConj{} for ``and'', \HOLTokenDisj{} for ``or'', \HOLTokenNeg{} for ``not'', also \HOLTokenImp{} for ``implies'' and \HOLTokenEquiv{} for ``if and only if''), set theory ($\in$ for ``element of'', $\times$ for Cartesian product, and comprehensions such as \HOLinline{\HOLTokenLeftbrace{}\HOLBoundVar{x}\;\HOLTokenBar{}\;\HOLBoundVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLNumLit{6}\HOLTokenRightbrace{}}), and functional programming (\HOLTokenLambda{} for abstraction, and juxtaposition for application). Repeated application of a function $f$ is indicated by exponents, \textit{e.g.}, \HOLinline{\HOLFreeVar{f}\;(\HOLFreeVar{f}\;(\HOLFreeVar{f}\;\HOLFreeVar{x}))\;\HOLSymConst{=}\;\HOLFreeVar{f}\ensuremath{\sp{\HOLNumLit{3}}(\HOLFreeVar{x})}}. For a function $f$ from set $S$ to set $T$, we write \HOLinline{\HOLFreeVar{f}\;\ensuremath{:}\;\HOLFreeVar{S}\;\ensuremath{\leftrightarrow}\;\HOLFreeVar{T}} to mean a bijection. The empty set is denoted by \HOLinline{\HOLSymConst{\HOLTokenEmpty{}}}, and a finite set, denoted by \HOLinline{\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}}, has cardinality \HOLinline{\ensuremath{|}\HOLFreeVar{S}\ensuremath{|}}. The set of natural numbers is denoted by $\mathbb{N}$, counting from $0$, and \HOLinline{\HOLConst{count}\;\HOLFreeVar{n}}\;\HOLTokenDefEquality{}\;\HOLinline{\HOLTokenLeftbrace{}\HOLBoundVar{x}\;\HOLTokenBar{}\;\HOLBoundVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{n}\HOLTokenRightbrace{}}, where \HOLTokenDefEquality{} means `equality by definition'. For a natural number $n \in \mathbb{N}$, \HOLinline{\HOLConst{square}\;\HOLFreeVar{n}} means it is a square: \HOLinline{\HOLSymConst{\HOLTokenExists{}}\HOLBoundVar{k}.\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLBoundVar{k}\HOLSymConst{\ensuremath{\sp{2}}}}, \HOLinline{\HOLConst{prime}\;\HOLFreeVar{n}} means it is a prime, and \HOLinline{\HOLConst{\HOLConst{even}}\;\HOLFreeVar{n}} or \HOLinline{\HOLConst{\HOLConst{odd}}\;\HOLFreeVar{n}} denotes its parity. The integer quotient and remainder of $m$ divided by $n$ are written as \HOLinline{\HOLFreeVar{m}\;\HOLConst{\HOLConst{div}}\;\HOLFreeVar{n}} and \HOLinline{\HOLFreeVar{m}\;\HOLConst{\HOLConst{mod}}\;\HOLFreeVar{n}}, respectively. We write \HOLinline{\HOLFreeVar{n}\;\HOLConst{\ensuremath{\mid}}\;\HOLFreeVar{m}} when $n$ divides $m$, which is equivalent to \HOLinline{\HOLFreeVar{m}\;\ensuremath{\equiv}\;\HOLNumLit{0}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLFreeVar{n}\ensuremath{)}} when \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}}. These are basic notations. Others will be introduced as they first appear. \paragraph*{HOL4 Sources} \ifdefined Proof scripts are located in a repository at {\small\url{https://bitbucket.org/jhlchan/project/src/master/fermat/twosq/}}. The scripts are compiled using HOL4, version \texttt{af01322db666}. In this paper, each theorem has \emph{\script{windmill}{197}}, which is hyperlinked to the appropriate line of the corresponding proof script in repository. \else Proof scripts of all theorems are located in a repository (omitted for anonymous review). The scripts are compiled using HOL4, version \texttt{6dcb52a09341}. In this paper, each theorem has \emph{\script{windmill}{197}}, which is hyperlinked to the appropriate line of the corresponding proof script in repository. (For anonymous review, this feature has been removed. See Appendix~\ref{app:cross-reference-theorems} for a cross-reference of theorems in this paper and the proof scripts in supplementary material.) \fi \section{Sum of Two Squares} \label{sec:sum-of-two-squares} The only even prime is \HOLinline{\HOLNumLit{2}\;\HOLSymConst{=}\;\HOLNumLit{1}\ensuremath{{\sp{\HOLNumLit{2}}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}\ensuremath{{\sp{\HOLNumLit{2}}}}}, a sum of two squares. An odd prime, upon division by 4, leaves a remainder of either $1$ or $3$. Only an odd prime of the first type can be expressed as a sum of two squares, as supported by numerical evidence from Table~\ref{tbl:odd-primes-sum}. \begin{table}[h] \caption{Examples of odd primes that can be expressed as a sum of two squares.} \Description{This table provides examples of odd primes that can be expressed as a sum of two squares.} \label{tbl:odd-primes-sum} \[ \begin{array}{r@{\quad\ee\quad}l@{\quad\ee\quad}l} 5 & 4(1) + 1 & 1^{2} + 2^{2}\\ 13 & 4(3) + 1 & 3^{2} + 2^{2}\\ 17 & 4(4) + 1 & 1^{2} + 4^{2}\\ 29 & 4(7) + 1 & 5^{2} + 2^{2}\\ 37 & 4(9) + 1 & 1^{2} + 6^{2}\\ 41 & 4(10) + 1 & 5^{2} + 4^{2}\\ 53 & 4(13) + 1 & 7^{2} + 2^{2}\\ 61 & 4(15) + 1 & 5^{2} + 6^{2}\\ \end{array} \] \end{table} \noindent Pierre de Fermat, in a letter to Marin Mersenne on Christmas day 1640, claimed that he had an ``irrefutable'' proof of this: \begin{theorem}[\textbf{Two Squares Theorem}] \label{thm:fermat-two-squares-thm} \script{twoSquares}{619} A prime $n$ can be expressed uniquely as a sum of odd and even squares if and only if \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}} for some $k$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;(\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenEquiv{}}\\ \;\;\;\;\;\;\;\;\;\;\HOLSymConst{\HOLTokenUnique{}}(\HOLBoundVar{u}\HOLSymConst{,}\HOLBoundVar{v}).\;\HOLConst{\HOLConst{odd}}\;\HOLBoundVar{u}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLConst{\HOLConst{even}}\;\HOLBoundVar{v}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLBoundVar{u}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLBoundVar{v}\HOLSymConst{\ensuremath{\sp{2}}}) \end{HOLmath} \end{theorem} \noindent This paper concentrates on formalising an elementary proof of this result by Roger Heath-Brown, later simplified by Don Zagier. As shown in his one-sentence proof in Section~\ref{sec:introduction}, the idea is this: look at the representations of $n$ not by squares, but in another form. Consider the following set $S_{n}$ of triples $(x,y,z)$: \begin{equation} \label{eqn:mills-set} S_{n} = \{(x,y,z) \in \mathbb{N}\times{\mathbb{N}}\times{\mathbb{N}} \mid \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}} \}. \end{equation} For a prime $n$ of the form $4k + 1$, we have $(1,1,k) \in S_{n}$. Thus the set $S_{n}$ is non-empty, and there are only finitely many triples in $S_{n}$. A triple with \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{z}} will give \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{\sp{2}}}}, \textit{i.e.}, a sum of two squares. If we can show that $S_{n}$ has only one such triple, we have a proof of Fermat's Theorem~\ref{thm:fermat-two-squares-thm}, with both existence and uniqueness. Meanwhile, some general theories will be developed as an exercise in formal proofs, so that they can be applied to similar problems. In addition, we extend the theories to establish not only an algorithm, but also a proof of its correctness, to compute the two unique squares for primes of the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}. \begin{figure*}[h] \begin{center} \begin{tikzpicture}[scale=0.25] \windmill{0}{0}{3}{8}{1} \node at (1.5,2.2) {$x$}; \node at (4.5,4.5) {$y$}; \node at (8.5,3.5) {$z$}; \windmill{14}{0}{3}{4}{2} \node at (15.5,2.2) {$x$}; \node at (16.0,5.5) {$y$}; \node at (18.5,4.2) {$z$}; \windmill{27}{-1}{5}{2}{2} \node at (29.5,3.4) {$x$}; \node at (28.0,6.6) {$y$}; \node at (29.5,5.0) {$z$}; \end{tikzpicture} \caption{Typical windmills, where \HOLinline{\HOLConst{windmill}\;\HOLFreeVar{x}\;\HOLFreeVar{y}\;\HOLFreeVar{z}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}}. The rightmost one has \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{z}}.} \Description{This figure shows typical windmills. The central square is x by x, the four rectangles arranged clockwise around the square are all y by z. The rightmost one has y and z equal.} \label{fig:windmill-sample} \end{center} \end{figure*} \subsection{Windmills} \label{sec:windmills} The following expression will be our main focus: \begin{definition} \label{def:windmill-def} A windmill consists of a central square with four identical rectangular arms. \begin{HOLmath} \;\;\HOLConst{windmill}\;\HOLFreeVar{x}\;\HOLFreeVar{y}\;\HOLFreeVar{z}\;\HOLTokenDefEquality{}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z} \end{HOLmath} \end{definition} \noindent Some typical windmills are shown in Figure~\ref{fig:windmill-sample}. The first term \HOLinline{\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}} is given by a central square of side $x$, and the second term \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}} is given by four arms, each a rectangle of width $y$ and height $z$, arranged clockwise around the square. Therefore each triple in the set $S_{n}$ of Equation~\eqref{eqn:mills-set} can be represented by a windmill, that is, each triple $(x,y,z)$ satisfies \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLConst{windmill}\;\HOLFreeVar{x}\;\HOLFreeVar{y}\;\HOLFreeVar{z}}. Given a prime \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}, we shall look for a windmill with four square arms (the one on the far right in Figure~\ref{fig:windmill-sample}), \textit{i.e.}, \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{z}}, so that \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}} \ee \HOLinline{\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;(\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y})\HOLSymConst{\ensuremath{\sp{2}}}}. First, we collect all triples $(x,y,z)$ which are solutions of \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}}: \begin{definition} \label{def:mills-def} \noindent The mills of a number is its set of windmills. \begin{HOLmath} \;\;\HOLConst{mills}\;\HOLFreeVar{n}\;\HOLTokenDefEquality{}\;\HOLTokenLeftbrace{}(\HOLBoundVar{x}\HOLSymConst{,}\HOLBoundVar{y}\HOLSymConst{,}\HOLBoundVar{z})\;\HOLTokenBar{}\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLConst{windmill}\;\HOLBoundVar{x}\;\HOLBoundVar{y}\;\HOLBoundVar{z}\HOLTokenRightbrace{} \end{HOLmath} \end{definition} \noindent This is the formal definition of the set $S_{n}$ of Equation~\eqref{eqn:mills-set}. The conditions for a proper windmill, with all lengths nonzero, are: \begin{HOLmath} \HOLTokenTurnstile{}\HOLSymConst{\HOLTokenNeg{}}\HOLConst{square}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\ensuremath{\not\equiv}\;\HOLNumLit{0}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{x}\;\HOLBoundVar{y}\;\HOLBoundVar{z}.\;(\HOLBoundVar{x}\HOLSymConst{,}\HOLBoundVar{y}\HOLSymConst{,}\HOLBoundVar{z})\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{mills}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLBoundVar{x}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLBoundVar{y}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLBoundVar{z}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0} \end{HOLmath} \begin{equation} \label{eqn:mills-triple-nonzero} \end{equation} When $n$ is a square, $\HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}}(0)$ for any value of $y$. This would make \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}} infinite. Otherwise: \begin{theorem} \label{thm:mills-finite} \script{windmill}{714} The number of windmills for a number $n$ is finite if and only if $n$ is not a square. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;(\HOLConst{mills}\;\HOLFreeVar{n})\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLSymConst{\HOLTokenNeg{}}\HOLConst{square}\;\HOLFreeVar{n} \end{HOLmath} \end{theorem} \noindent Given an odd $n$ that is not a square, we can determine all its windmill triples $(x,y,z)$ by noting that, since \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}} is even, $x$ must be odd, and $y$ and $z$ form the product \HOLinline{\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}\;\HOLSymConst{=}\;(\HOLFreeVar{n}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}})\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4}}. In Table~\ref{tbl:windmills-29} this is worked out for \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}}, using successive odd $x$ and factors for the product $yz$. The corresponding windmills are shown in Figure~\ref{fig:windmills-29}. \begin{table*}[h] \caption{Determine all the windmill triples of \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}}, by odd $x$ and factors of $yz$.} \Description{This table shows how to detemine all the windmill triples for n = 29, using odd x and factors of the product yz.} \label{tbl:windmills-29} \begin{tabular}{r@{\quad}@{\quad}r@{\quad\ee\quad}r@{\quad}@{\quad}l@{\quad}l} odd $x$ & \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}} & \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}} & triple $(x,y,z)$ & comment\\ \hline $1$ & $29 - 1^{2}$ & $28 \ee 4(7)$ & $(1,1,7), (1,7,1)$ & factors of $7$ are $1, 7$.\\ $3$ & $29 - 3^{2}$ & $20 \ee 4(5)$ & $(3,1,5), (3,5,1)$ & factors of $5$ are $1, 5$.\\ $5$ & $29 - 5^{2}$ & $4 \ee 4(1)$ & $(5,1,1)$ & factor of $1$ is $1$.\\ \end{tabular} \end{table*} \begin{figure*}[h] \begin{center} \begin{tikzpicture}[scale=0.22] \draw[step=1, color=white!60!black] (0,0) grid (65,17); \windmill{8}{8}{1}{1}{7} \windmill{23}{7}{3}{1}{5} \windmill{34}{6}{5}{1}{1} \windmill{44}{7}{3}{5}{1} \windmill{57}{8}{1}{7}{1} \node at (4,1) {$(1,1,7)$}; \node at (22,1) {$(3,1,5)$}; \node at (36,1) {$(5,1,1)$}; \node at (45,1) {$(3,5,1)$}; \node at (55,1) {$(1,7,1)$}; \end{tikzpicture} \caption{All the windmills of \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}}, determined from Table~\ref{tbl:windmills-29}.} \Description{This figure shows all the windmills of n = 29, determined from the previous table.} \label{fig:windmills-29} \end{center} \end{figure*} When a number $n$ has the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}, \[ n \ee 1^{2} + 4(1)k \ee \HOLinline{\HOLConst{windmill}\;\HOLNumLit{1}\;\HOLNumLit{1}\;\HOLFreeVar{k}}, \] showing that its \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLSymConst{\HOLTokenEmpty{}}}: \begin{HOLmath} \HOLTokenTurnstile{}\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenImp{}}\;(\HOLNumLit{1}\HOLSymConst{,}\HOLNumLit{1}\HOLSymConst{,}\HOLFreeVar{n}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4})\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{mills}\;\HOLFreeVar{n} \end{HOLmath} Moreover, when this form corresponds to a prime, this is the only triple $(x,y,z)$ with \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{y}}: \begin{theorem} \label{thm:mills-trivial-prime} \script{windmill}{428} For a prime of the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}, the only windmill with the first and second parameters equal is \HOLinline{\HOLConst{windmill}\;\HOLNumLit{1}\;\HOLNumLit{1}\;\HOLFreeVar{k}}. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{x}\;\HOLBoundVar{z}.\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLConst{windmill}\;\HOLBoundVar{x}\;\HOLBoundVar{x}\;\HOLBoundVar{z}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLBoundVar{x}\;\HOLSymConst{=}\;\HOLNumLit{1}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLBoundVar{z}\;\HOLSymConst{=}\;\HOLFreeVar{n}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4} \end{HOLmath} \end{theorem} \begin{proof} Note that \HOLinline{\HOLFreeVar{k}\;\HOLSymConst{=}\;\HOLFreeVar{n}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4}} for prime \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}. Consider \HOLinline{(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{mills}\;\HOLFreeVar{n}} with \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{y}}. This implies, \[ \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLConst{windmill}\;\HOLFreeVar{x}\;\HOLFreeVar{x}\;\HOLFreeVar{z}} \ee \HOLinline{\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{x}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{}}(\HOLFreeVar{x}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z})}. \] Therefore \HOLinline{\HOLFreeVar{x}\;\HOLConst{\ensuremath{\mid}}\;\HOLFreeVar{n}}. As prime $n$ is not a square, \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{n}}. Hence \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLNumLit{1}}, so \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLNumLit{1}}, and \HOLinline{\HOLFreeVar{z}\;\HOLSymConst{=}\;\HOLFreeVar{k}}. \end{proof} \subsection{Involution} \label{sec:involution} We are going to study involutions on \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}}, the set of windmills for $n$. A function $f$ is an involution on a set $S$, denoted by \HOLinline{\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}}, when it is its own inverse: \begin{HOLmath} \;\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLTokenDefEquality{}\;\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{x}.\;\HOLBoundVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLFreeVar{f}\;\HOLBoundVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;(\HOLFreeVar{f}\;\HOLBoundVar{x})\;\HOLSymConst{=}\;\HOLBoundVar{x} \end{HOLmath} That is, $f$ is a bijection \HOLinline{\HOLFreeVar{f}\;\ensuremath{:}\;\HOLFreeVar{S}\;\ensuremath{\leftrightarrow}\;\HOLFreeVar{S}}, pairing up $x$ and~\HOLinline{\HOLFreeVar{f}\;\HOLFreeVar{x}}, both in $S$. When \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLFreeVar{x}}, the element $x$ is fixed by the involution $f$. We define the following sets: \begin{definition} \label{def:involute-pairs-fixes-def} The pairs and fixes of an involution $f$ on a set $S$. \begin{HOLmath} \;\;\HOLConst{pairs}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLTokenDefEquality{}\;\HOLTokenLeftbrace{}\HOLBoundVar{x}\;\HOLTokenBar{}\;\HOLBoundVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLBoundVar{x}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLBoundVar{x}\HOLTokenRightbrace{}\\ \;\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLTokenDefEquality{}\;\HOLTokenLeftbrace{}\HOLBoundVar{x}\;\HOLTokenBar{}\;\HOLBoundVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLBoundVar{x}\;\HOLSymConst{=}\;\HOLBoundVar{x}\HOLTokenRightbrace{}\\ \end{HOLmath} \end{definition} \noindent Clearly they are disjoint. The subset $\HOLinline{\HOLConst{pairs}\;\HOLFreeVar{f}\;\HOLFreeVar{S}}$ consists of distinct involute pairs, so its cardinality is even: \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLConst{\HOLConst{even}}\;\ensuremath{|}\HOLConst{pairs}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\ensuremath{|} \end{HOLmath} So both \HOLinline{\ensuremath{|}\HOLFreeVar{S}\ensuremath{|}} and \HOLinline{\ensuremath{|}\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\ensuremath{|}} have the same parity. This leads to: \begin{theorem} \label{thm:involute-two-fixes-both-odd} \script{involuteFix}{1182} If two involutions act on the same finite set $S$, their fixes have the same parity. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;(\HOLConst{\HOLConst{odd}}\;\ensuremath{|}\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\ensuremath{|}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLConst{\HOLConst{odd}}\;\ensuremath{|}\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}\ensuremath{|}) \end{HOLmath} \end{theorem} \noindent We shall meet the two involutions on \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}}, a set which is finite for non-square $n$ (by Theorem~\ref{thm:mills-finite}). \section{Windmill Involutions} \label{sec:windmill-involutions} Zagier's one-sentence proof is the interplay of two involutions on the set of windmills (\HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}}) for a prime \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}. \subsection{Flip Map} \label{sec:flip-map} The first involution just swaps the $y$ and $z$ in the triple $(x,y,z)$: \begin{definition} \label{def:flip-def} The flip map for a triple. \begin{HOLmath} \;\;\HOLConst{flip}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})\;\HOLTokenDefEquality{}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{z}\HOLSymConst{,}\HOLFreeVar{y}) \end{HOLmath} \end{definition} \noindent The set \HOLinline{\HOLFreeVar{S}\;\HOLSymConst{=}\;\HOLConst{mills}\;\HOLFreeVar{n}} of windmill triples of a number $n$ can be partitioned by $y, z$ into: \[ \begin{array}{c} S_{y < z} \ee \{(x,y,z) \in S \mid y < z\}\\ S_{y \ee z} \ee \{(x,y,z) \in S \mid y \ee z\}\\ S_{y > z} \ee \{(x,y,z) \in S, \mid y > z\}\\ \end{array} \] An example for \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}} is shown in Figure~\ref{fig:windmills-29-by-flip}. Clearly there is a bijection: $\HOLConst{flip}\colon S_{y < z} \leftrightarrow S_{y > z}$, and $S_{y \ee z} \ee \HOLinline{\HOLConst{fixes}\;\HOLConst{flip}\;\HOLFreeVar{S}}$. Thus the inverse of flip is itself: \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{flip}\;(\HOLConst{flip}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z}))\;\HOLSymConst{=}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z}) \end{HOLmath} showing that: \begin{theorem} \label{thm:flip-involute-mills} \script{windmill}{933} The flip map is an involution on the set of windmills. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{flip}\;\HOLConst{involute}\;\HOLConst{mills}\;\HOLFreeVar{n} \end{HOLmath} \end{theorem} \begin{figure*}[h] \begin{center} \begin{tikzpicture}[scale=0.22] \draw[step=1, color=white!60!black] (0,0) grid (65,17); \windmill{8}{8}{1}{1}{7} \windmill{23}{7}{3}{1}{5} \windmill{34}{6}{5}{1}{1} \windmill{44}{7}{3}{5}{1} \windmill{57}{8}{1}{7}{1} \node at (4,1) {$(1,1,7)$}; \node at (22,1) {$(3,1,5)$}; \node at (36,1) {$(5,1,1)$}; \node at (45,1) {$(3,5,1)$}; \node at (55,1) {$(1,7,1)$}; \coordinate (a) at (0.5,0.5); \coordinate (b) at (31,16.5); \node[draw,thick,color=purple,fit= (a) (b),rounded corners=.55cm,inner sep=2pt] {}; \coordinate (a) at (32.5,0.5); \coordinate (b) at (40,16.5); \node[draw,thick,color=purple,fit= (a) (b),rounded corners=.55cm,inner sep=2pt] {}; \coordinate (a) at (41.5,0.5); \coordinate (b) at (64.5,16.5); \node[draw,thick,color=purple,fit= (a) (b),rounded corners=.55cm,inner sep=2pt] {}; \begin{scope}[scale=10] \coordinate (a) at (1.0,1.1); \coordinate (b) at (6.0,1.1); \draw[<->] (a) to [bend left,looseness=0.8] node[midway,below] {\HOLConst{flip}} (b); \coordinate (a) at (2.5,1.2); \coordinate (b) at (4.6,1.2); \draw[<->] (a) to [bend left,looseness=0.8] node[midway,below] {\HOLConst{flip}} (b); \coordinate (a) at (3.6,0.5); \coordinate (b) at (3.6,0.51); \draw[<->] (a) to [out=-30, in=-150, looseness=200] node[midway,above] {\HOLConst{flip}} (b); \end{scope} \end{tikzpicture} \caption{Partition of windmills of \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}} for \HOLConst{flip}: those with $y < z, y \ee z$, and $y > z$. Note left and right pairing.} \Description{This figure shows a partition of the windmills of n = 29 for the flip map: those with y < z, y = z, and y > z. Note the left and right pairing between y < z and y > z.} \label{fig:windmills-29-by-flip} \end{center} \end{figure*} \subsection{Zagier Map} \label{sec:zagier-map} The other involution is the one devised by Don Zagier, as shown in Equation~\eqref{eqn:zagier-map}: \begin{definition} \label{def:zagier-def} The Zagier map for a triple. \begin{HOLmath} \;\;\HOLConst{zagier}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})\;\HOLTokenDefEquality{}\\ \;\;\;\;\HOLKeyword{if}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{y}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{z}\;\HOLKeyword{then}\;(\HOLFreeVar{x}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}\HOLSymConst{,}\HOLFreeVar{z}\HOLSymConst{,}\HOLFreeVar{y}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{z}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{x})\\ \;\;\;\;\HOLKeyword{else}\;\HOLKeyword{if}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\;\HOLKeyword{then}\;(\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{x}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{z}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{y})\\ \;\;\;\;\HOLKeyword{else}\;(\HOLFreeVar{x}\;\HOLSymConst{\ensuremath{-}}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{x}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{z}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{y}) \end{HOLmath} \end{definition} \noindent Algebraically, this is indeed an involution, as HOL4 can verify without a blink: \begin{HOLmath} \HOLTokenTurnstile{}\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{z}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLConst{zagier}\;(\HOLConst{zagier}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z}))\;\HOLSymConst{=}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z}) \end{HOLmath} \begin{equation} \label{eqn:zagier-involute} \end{equation} That HOL4 can verify this directly from definition is a showcase of its excellent algebraic simplifier, especially for natural numbers. However, we would like to see the magic behind, in terms of the geometry of windmills. Note that this definition differs slightly from Equation~\eqref{eqn:zagier-map} since the else-parts include boundary cases. They actually correspond to improper windmills, and they are irrelevant for the values of $n$ satisfying Equation~\eqref{eqn:mills-triple-nonzero}. \subsection{Mind of a Windmill} \label{sec:windmill-mind} The main purpose of introducing windmills is to read their minds. \begin{figure*}[h] \begin{center} \begin{tikzpicture}[scale=0.28] \windmill{0}{0}{3}{8}{1} \windmill{14}{0}{3}{8}{1} \windmill{27}{-1}{5}{1}{4} \node at (1.5,2.2) {$x$}; \node at (5.0,4.5) {$y$}; \node at (8.5,3.5) {$z$}; \draw[decorate,thick,decoration={brace,amplitude=5pt,mirror,raise=2pt}] (14,2.8) -- (17,2.8) node[midway,below,yshift=-7pt] {$x$}; \draw[decorate,thick,decoration={brace,amplitude=5pt,raise=2pt}] (13,4.2) -- (18,4.2) node[midway,above,yshift=7pt] {$x'$}; \node at (29.5,3.4) {$x'$}; \node at (27.5,8.6) {$y'$}; \node at (28.5,6.2) {$z'$}; \mind{13}{-1}{5} \end{tikzpicture} \caption{A typical \HOLinline{\HOLConst{windmill}\;\HOLFreeVar{x}\;\HOLFreeVar{y}\;\HOLFreeVar{z}\;\HOLSymConst{=}\;\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}}, with a mind (in dashes) and transforms to another windmill.} \Description{This figure shows a typical windmill, with a mind in dashes, on the left. The figure also illustrates how the left windmill transforms to another windmill on the right, with the same mind.} \label{fig:windmill-mind} \end{center} \end{figure*} Referring to Figure~\ref{fig:windmill-mind}, a windmill has a mind (marked in dashes at middle), which is the maximum central square, with side $x'$, that can be fitted with the four arms. When \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLeq{}}\;\ensuremath{\HOLFreeVar{x}\sp{\prime{}}}}, the original square \HOLinline{\HOLFreeVar{x}\ensuremath{{\sp{\HOLNumLit{2}}}}} can grow to the mind \HOLinline{\ensuremath{\HOLFreeVar{x}\sp{\prime{}}}\ensuremath{{\sp{\HOLNumLit{2}}}}}, forming another windmill but keeping the overall shape (on the right). Conversely, going from right to left, we can use the mind as a reference to shrink the square term from \HOLinline{\ensuremath{\HOLFreeVar{x}\sp{\prime{}}}\ensuremath{{\sp{\HOLNumLit{2}}}}} to \HOLinline{\HOLFreeVar{x}\ensuremath{{\sp{\HOLNumLit{2}}}}} by trimming four sides, thereby restoring the arms to original. Transforming a windmill's square term through the mind is the geometric interpretation of Equation~\eqref{eqn:zagier-map}. \begin{table*}[h] \caption{The five cases of Zagier map, transforming a triple $(x,y,z)$ to $(x',y',z')$.} \Description{This figure shows the five cases of Zagier map, transforming a triple (x,y,z) to (x',y',z').} \label{tbl:zagier-map} \begin{tabular}{c@{\quad}l@{\quad}l@{\quad}r@{\quad}l@{\quad}r@{\quad}r@{\quad}r@{\quad}l@{\quad}l} Case & Type & condition & Mind & Picture & $x'$ & $y'$ & $z'$ & condition & Type\\ \hline $1$ & \multirow{ 2}{*}{$x < y$} & $x < y - z$ & $x + 2z$ & Figure~\ref{fig:zagier-map}~(a) & $x + 2z$ & $z$ & $y - x - z$ & $2y' < x'$ & \multirow{ 2}{*}{$y' < x'$}\\ $2$ & & $y - z < x$ & $2y - x$ & Figure~\ref{fig:zagier-map}~(b) & $2y - x$ & $y$ & $x + z - y$ & $x' < 2y'$ & \\ \hline $3$ & $x = y$ & & $x$ & Figure~\ref{fig:zagier-map}~(c) & $x$ & $y$ & $z$ & & $x' = y'$\\ \hline $4$ & \multirow{ 2}{*}{$y < x$} & $x < 2y$ & $x$ & Figure~\ref{fig:zagier-map}~(d) & $2y - x$ & $y$ & $x + z - y$ & $y' - z' < x'$ & \multirow{ 2}{*}{$x' < y'$}\\ $5$ & & $2y < x$ & $x$ & Figure~\ref{fig:zagier-map}~(e) & $x - 2y$ & $x + z - y$ & $y$ & $x' < y' - z'$ & \\ \hline \end{tabular} \end{table*} \begin{figure*}[htbp] \begin{center} \begin{tikzpicture}[scale=0.3] \windmill{0}{0}{3}{8}{2} \windmill{15}{-2}{7}{2}{3} \mind{-2}{-2}{7} \mind{15}{-2}{7} \node at (1.5,2.2) {$x$}; \node at (4.0,5.6) {$y$}; \node at (8.5,4.0) {$z$}; \draw[->,thick] (10,3) -- (12,3) node[midway,above] {\HOLConst{zagier}}; \node at (18.5,4.0) {$x'$}; \node at (26.0,4.0) {$y'$}; \node at (23.5,1.5) {$z'$}; \draw[color=red,ultra thick] (5,3) -- (8,3); \draw[color=red,ultra thick] (22,3) -- (25,3); \draw[decorate,thick,decoration={brace,amplitude=5pt,mirror,raise=2pt}] (5,3) -- (8,3); \draw[decorate,thick,decoration={brace,amplitude=5pt,mirror,raise=2pt}] (22,3) -- (25,3); \node at (0.8,10) {\large{(a) Case $1$: \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{y}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{z}}.}}; \node at (33,3) {$\large{ \begin{array}{r@{\;=\;}l} x' & x + 2z\\ y' & z\\ z' & y - x - z\\ \end{array} }$}; \end{tikzpicture} \begin{tikzpicture}[scale=0.3] \windmill{0}{0}{3}{6}{4} \windmillr{15}{-3}{9}{6}{1} \mind{-3}{-3}{9} \mind{15}{-3}{9} \node at (1.5,2.2) {$x$}; \node at (3.5,7.5) {$y$}; \node at (6.5,4.5) {$z$}; \draw[->,thick] (10,3) -- (12,3) node[midway,above] {\HOLConst{zagier}}; \node at (19.5,5.0) {$x'$}; \node at (21.0,7.6) {$y'$}; \node at (17.0,6.8) {$z'$}; \draw[color=red,ultra thick] (0,6) -- (0,7); \draw[color=red,ultra thick] (18,6) -- (18,7); \draw[decorate,thick,decoration={brace,amplitude=2pt,raise=2pt}] (0,6) -- (0,7); \draw[decorate,thick,decoration={brace,amplitude=2pt,raise=2pt}] (18,6) -- (18,7); \node at (6,10) {}; \node at (6,9) {\large{(b) Case $2$: \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{z}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{x}} and \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{y}}, so \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}}.}}; \node at (33,3) {$\large{ \begin{array}{r@{\;=\;}l} x' & 2y - x\\ y' & y\\ z' & x + z - y\\ \end{array} }$}; \node at (0,-4) {}; \end{tikzpicture} \begin{tikzpicture}[scale=0.3] \windmill{0}{0}{4}{4}{2} \windmill{16}{0}{4}{4}{2} \mind{0}{0}{4} \mind{16}{0}{4} \node at (2.2,3.5) {$x$}; \node at (2.2,6.5) {$y$}; \node at (4.5,5.2) {$z$}; \draw[->,thick] (10,3) -- (12,3) node[midway,above] {\HOLConst{zagier}}; \node at (18.0,3.5) {$x'$}; \node at (18.0,6.5) {$y'$}; \node at (14.8,5.2) {$z'$}; \draw[color=red,ultra thick] (0,4) -- (0,6); \draw[color=red,ultra thick] (16,4) -- (16,6); \draw[decorate,thick,decoration={brace,amplitude=4pt,raise=2pt}] (0,4) -- (0,6); \draw[decorate,thick,decoration={brace,amplitude=4pt,raise=2pt}] (16,4) -- (16,6); \node at (7,9) {}; \node at (6.5,8) {\large{(c) Case $3$: \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{x}}, so \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{z}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{x}} and \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}}.}}; \node at (33,3) {$\large{ \begin{array}{r@{\;=\;}l} x' & 2y - x = x\\ y' & y\\ z' & x + z - y = z\\ \end{array} }$}; \end{tikzpicture} \begin{tikzpicture}[scale=0.3] \windmill{0}{0}{6}{4}{2} \windmillr{18}{2}{2}{4}{4} \mind{0}{0}{6} \mind{16}{0}{6} \node at (3.0,5.5) {$x$}; \node at (2.5,8.5) {$y$}; \node at (4.5,6.8) {$z$}; \draw[->,thick] (10,3) -- (12,3) node[midway,above] {\HOLConst{zagier}}; \node at (19.0,3.6) {$x'$}; \node at (18.0,8.6) {$y'$}; \node at (14.5,6.0) {$z'$}; \draw[color=red,ultra thick] (0,4) -- (0,8); \draw[color=red,ultra thick] (16,4) -- (16,8); \draw[decorate,thick,decoration={brace,amplitude=5pt,raise=2pt}] (0,4) -- (0,8); \draw[decorate,thick,decoration={brace,amplitude=5pt,raise=2pt}] (16,4) -- (16,8); \node at (3,11) {}; \node at (3,10) {\large{(d) Case $4$: \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{x}}, but \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}}.}}; \node at (34,3) {$\large{ \begin{array}{r@{\;=\;}l} x' & 2y - x\\ y' & y\\ z' & x + z - y\\ \end{array} }$}; \end{tikzpicture} \begin{tikzpicture}[scale=0.3] \windmill{0}{0}{7}{3}{2} \windmill{19}{3}{1}{6}{3} \mind{0}{0}{7} \mind{16}{0}{7} \node at (3.5,6.2) {$x$}; \node at (1.5,9.5) {$y$}; \node at (3.5,8.0) {$z$}; \draw[->,thick] (10,3) -- (12,3) node[midway,above] {\HOLConst{zagier}}; \node at (19.5,3.5) {$x'$}; \node at (22.5,7.8) {$y'$}; \node at (25.6,5.8) {$z'$}; \draw[color=yellow,ultra thick] (0,9) -- (3,9); \draw[color=yellow,ultra thick] (4,-2) -- (7,-2); \draw[color=yellow,ultra thick] (16,3) -- (19,3); \draw[color=yellow,ultra thick] (20,4) -- (23,4); \draw[decorate,thick,decoration={brace,amplitude=5pt,mirror,raise=2pt}] (0,9) -- (3,9); \draw[decorate,thick,decoration={brace,amplitude=5pt,mirror,raise=2pt}] (4,-2) -- (7,-2); \draw[decorate,thick,decoration={brace,amplitude=5pt,mirror,raise=2pt}] (16,3) -- (19,3); \draw[decorate,thick,decoration={brace,amplitude=5pt,mirror,raise=2pt}] (20,4) -- (23,4); \node at (3,12) {}; \node at (3,11) {\large{(e) Case $5$: \HOLinline{\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{x}}, so \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{x}}.}}; \node at (34,3) {$\large{ \begin{array}{r@{\;=\;}l} x' & x - 2y\\ y' & x + z - y\\ z' & y\\ \end{array} }$}; \end{tikzpicture} \caption{All five cases of the Zagier map, from $(x,y,z)$ to $(x',y',z')$ through the mind of a windmill.} \Description{This figure shows all the 5 cases of the Zagier map, transform through the mind of a windmill.} \label{fig:zagier-map} \end{center} \end{figure*} The Zagier map transforms $(x,y,z)$ to $(x',y',z')$ via the mind of the windmill, keeping its overall shape. There are three types, depending on whether $x < y$, $x = y$, or $y < x$. Both the first and last types are divided into two cases, as the geometry for the mind is different. Altogether there are five cases, as analysed in Table~\ref{tbl:zagier-map}, and illustrated in Figure~\ref{fig:zagier-map}.\footnote{Dubach and Muehlboeck~\cite{Dubach-Muehlboeck-2021-acm} also identified five types for windmills.} Although five cases of Zagier map have been identified, note that the transformation rule: \[ (x',y',z') \ee (2y - x, y, x + z - y) \] happens to be the same for case $2$ and case $4$. The same rule actually applies to case $3$, which has \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{y}}. Thus the Zagier map can be succinctly expressed as in Definition~\ref{def:zagier-def} with only three branches. Moreover, we can define the mind of a windmill triple as (see Table~\ref{tbl:zagier-map}): \begin{HOLmath} \;\;\HOLConst{mind}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})\;\HOLTokenDefEquality{}\\ \;\;\;\;\HOLKeyword{if}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{y}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{z}\;\HOLKeyword{then}\;\HOLFreeVar{x}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{z}\\ \;\;\;\;\HOLKeyword{else}\;\HOLKeyword{if}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{y}\;\HOLKeyword{then}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\;\HOLSymConst{\ensuremath{-}}\;\HOLFreeVar{x}\\ \;\;\;\;\HOLKeyword{else}\;\HOLFreeVar{x} \end{HOLmath} and verify that the mind is an invariant under the Zagier map for any triple: \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{mind}\;(\HOLConst{zagier}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z}))\;\HOLSymConst{=}\;\HOLConst{mind}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z}) \end{HOLmath} Referring again to Table~\ref{tbl:zagier-map}, the windmills in \HOLinline{\HOLFreeVar{S}\;\HOLSymConst{=}\;\HOLConst{mills}\;\HOLFreeVar{n}} can be partitioned into three triple types: \[ \begin{array}{c@{\quad}l} S_{x < y} \ee \{(x,y,z) \in S \mid x < y\} & \text{covering cases $1$ and $2$}\\ S_{x \ee y} \ee \{(x,y,z) \in S \mid x \ee y\} & \text{covering case $3$}\\ S_{x > y} \ee \{(x,y,z) \in S \mid x > y\} & \text{covering cases $4$ and $5$}\\ \end{array} \] Such a partition for \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}} is shown in Figure~\ref{fig:windmills-29-by-zagier}. Table~\ref{tbl:zagier-map} also shows that, for triples with proper windmills: \begin{itemize}[leftmargin=*] \item a triple of case $1$ maps to case $5$ and vice versa, \item a triple of case $2$ maps to case $4$ and vice versa, and \item a triple of case $3$ maps to itself. \end{itemize} Therefore the Zagier map is its own inverse for proper triples. Combining Equation~\eqref{eqn:zagier-involute} and Equation~\eqref{eqn:mills-triple-nonzero} for the windmills of a prime, we have: \begin{theorem} \label{thm:zagier-involute-mills-prime} \script{windmill}{1475} The Zagier map is an involution on \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}} for a prime $n$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLConst{zagier}\;\HOLConst{involute}\;\HOLConst{mills}\;\HOLFreeVar{n} \end{HOLmath} \end{theorem} \begin{figure*}[h] \begin{center} \begin{tikzpicture}[scale=0.22] \draw[step=1, color=white!60!black] (0,0) grid (65,17); \windmill{8}{8}{1}{1}{7} \windmill{24}{8}{1}{7}{1} \windmill{35}{7}{3}{5}{1} \windmill{47}{7}{3}{1}{5} \windmill{58}{6}{5}{1}{1} \node at (4,1) {$(1,1,7)$}; \node at (22,1) {$(1,7,1)$}; \node at (36,1) {$(3,5,1)$}; \node at (46,1) {$(3,1,5)$}; \node at (60,1) {$(5,1,1)$}; \coordinate (a) at (0.5,0.5); \coordinate (b) at (16.5,16.5); \node[draw,thick,color=purple,fit= (a) (b),rounded corners=.55cm,inner sep=2pt] {}; \coordinate (a) at (18,0.5); \coordinate (b) at (40,16.5); \node[draw,thick,color=purple,fit= (a) (b),rounded corners=.55cm,inner sep=2pt] {}; \coordinate (a) at (41.5,0.5); \coordinate (b) at (64.5,16.5); \node[draw,thick,color=purple,fit= (a) (b),rounded corners=.55cm,inner sep=2pt] {}; \mind[ultra thick]{8}{8}{1} \mind{23}{7}{3} \mind{34}{6}{5} \mind{47}{7}{3} \mind{58}{6}{5} \begin{scope}[scale=10] \coordinate (a) at (2.5,1.2); \coordinate (b) at (4.6,1.2); \draw[<->] (a) to [bend left] node[midway,below] {\HOLConst{zagier}} (b); \coordinate (a) at (3.4,0.5); \coordinate (b) at (6.0,0.5); \draw[<->] (a) to [bend right, looseness=0.8] node[midway,above] {\HOLConst{zagier}} (b); \coordinate (a) at (1.2,0.6); \coordinate (b) at (1.2,0.61); \draw[<->] (a) to [out=-30, in=-150, looseness=250] node[midway,above] {\HOLConst{zagier}} (b); \end{scope} \end{tikzpicture} \caption{Partition of windmills of \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}} for \HOLConst{zagier}: those with $x \ee y, x < y$, and $x > y$. Note pairing by minds.} \Description{This figure shows a partition of windmills of n = 29 for the Zagier map, those with x = y, x < y, and x > y. Note the pairing of minds between those x < y and x > y.} \label{fig:windmills-29-by-zagier} \end{center} \end{figure*} \section{Two Squares Theorem} \label{sec:two-squares-theorem} Now we have enough tools to formalise Fermat's two squares theorem. \subsection{Existence of Two Squares} \label{sec:existence} For the Zagier map, it is straightforward to verify, as indicated in Table~\ref{tbl:zagier-map}, that only a triple of case $3$ can map to itself: \begin{HOLmath} \HOLTokenTurnstile{}\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenImp{}}\;(\HOLConst{zagier}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})\;\HOLSymConst{=}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{y}) \end{HOLmath} Hence $S_{x \ee y} \ee \HOLinline{\HOLConst{fixes}\;\HOLConst{zagier}\;(\HOLConst{mills}\;\HOLFreeVar{n})}$. Applying Theorem~\ref{thm:mills-trivial-prime} which characterises such triples, for certain primes $S_{x \ee y}$ is a singleton: \begin{theorem} \label{thm:zagier-fixes-prime} \script{twoSquares}{162} A prime of the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}} has only $(1,1,k)$ fixed by the Zagier map. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLConst{fixes}\;\HOLConst{zagier}\;(\HOLConst{mills}\;\HOLFreeVar{n})\;\HOLSymConst{=}\;\HOLTokenLeftbrace{}(\HOLNumLit{1}\HOLSymConst{,}\HOLNumLit{1}\HOLSymConst{,}\HOLFreeVar{n}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4})\HOLTokenRightbrace{} \end{HOLmath} \end{theorem} \noindent The fixed points of two involutions play crucial roles in the existence of two squares for Theorem~\ref{thm:fermat-two-squares-thm}: \begin{theorem}[\textbf{Two Squares Existence}] \label{thm:fermat-two-squares-exists} \script{twoSquares}{441} A prime of the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}} is a sum of two squares of different parity. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLSymConst{\HOLTokenExists{}}(\HOLBoundVar{u}\HOLSymConst{,}\HOLBoundVar{v}).\;\HOLConst{\HOLConst{odd}}\;\HOLBoundVar{u}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLConst{\HOLConst{even}}\;\HOLBoundVar{v}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLBoundVar{u}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLBoundVar{v}\HOLSymConst{\ensuremath{\sp{2}}} \end{HOLmath} \end{theorem} \begin{proof} A prime is not a square, so \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}} is finite by Theorem~\ref{thm:mills-finite}. , and both Zagier and flip maps are involutions on \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}}, by Theorem~\ref{thm:zagier-involute-mills-prime} and Theorem~\ref{thm:flip-involute-mills}. Note that Zagier map has a single fixed point by Theorem~\ref{thm:zagier-fixes-prime}. Thus \HOLinline{\ensuremath{|}\HOLConst{fixes}\;\HOLConst{zagier}\;(\HOLConst{mills}\;\HOLFreeVar{n})\ensuremath{|}\;\HOLSymConst{=}\;\HOLNumLit{1}}, so \HOLinline{\ensuremath{|}\HOLConst{fixes}\;\HOLConst{flip}\;(\HOLConst{mills}\;\HOLFreeVar{n})\ensuremath{|}} is odd by Theorem~\ref{thm:involute-two-fixes-both-odd}. Hence \HOLinline{\HOLConst{fixes}\;\HOLConst{flip}\;(\HOLConst{mills}\;\HOLFreeVar{n})\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLSymConst{\HOLTokenEmpty{}}}, containing a triple $(x,y,y)$. Thus \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLConst{windmill}\;\HOLFreeVar{x}\;\HOLFreeVar{y}\;\HOLFreeVar{y}} $\ee$ \HOLinline{\HOLFreeVar{x}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}\HOLSymConst{\ensuremath{\sp{2}}}}. Take \HOLinline{\HOLFreeVar{u}\;\HOLSymConst{=}\;\HOLFreeVar{x}}, and \HOLinline{\HOLFreeVar{v}\;\HOLSymConst{=}\;\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{y}}, then \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{u}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{v}\HOLSymConst{\ensuremath{\sp{2}}}}. Evidently \HOLinline{\HOLFreeVar{v}} is even, and $u$ is odd since $n$ is odd. \end{proof} \noindent Current formalisations of Zagier's proof (HOL Light~\cite{Harrison-2010}, NASA PVS~\cite{Narkawicz-2012-acm} and Coq~\cite{Dubach-Muehlboeck-2021-acm}), or its close relative Heath-Brown's proof (Mizar~\cite{Riccardi-2009} and ProofPower~\cite{Arthan-2016}), stop at just showing the existence of two squares for the primes in \text{Fermat's} Theorem~\ref{thm:fermat-two-squares-thm}, most likely because this already meets the Formalizing 100 Theorems challenge~\cite{Wiedijk-2020}. See also related work in Section~\ref{sec:related-work}. \subsection{Uniqueness of Two Squares} \label{sec:uniqueness} The uniqueness of the two squares in Fermat's Theorem~\ref{thm:fermat-two-squares-thm} is a consequence of the following property of a prime: \begin{theorem}[\textbf{Two Squares Uniquenss}] \label{thm:fermat-two-squares-unique} \script{twoSquares}{205} If a prime $n$ can be expressed as a sum of two squares, the expression is unique up to commutativity. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{a}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{b}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLFreeVar{c}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{d}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLTokenLeftbrace{}\HOLFreeVar{a};\;\HOLFreeVar{b}\HOLTokenRightbrace{}\;\HOLSymConst{=}\;\HOLTokenLeftbrace{}\HOLFreeVar{c};\;\HOLFreeVar{d}\HOLTokenRightbrace{} \end{HOLmath} \end{theorem} \noindent The proof is purely number-theoretic, which has also been formalised by Laurent Th{\'e}ry in Coq~\cite{Thery-2004}. Moreover, we have: \begin{theorem} \label{thm:mod-4-not-squares} \script{helperTwosq}{419} A number of the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{3}} cannot be expressed as a sum of two squares. \begin{HOLmath} \HOLTokenTurnstile{}\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{3}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{u}\;\HOLBoundVar{v}.\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLBoundVar{u}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLBoundVar{v}\HOLSymConst{\ensuremath{\sp{2}}} \end{HOLmath} \end{theorem} \noindent This is an elementary result from possible remainders after division by 4: while a number, such as $u$ or $v$, may have a remainer $0, 1, 2$, or $3$, a square, such as $u^{2}$ or $v^{2}$, can only have a remainder $0$ or $1$. Thus the sum of such remainders can never be $3$. Now we can complete the proof of Fermat's two squares Theorem~\ref{thm:fermat-two-squares-thm}: \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;(\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenEquiv{}}\\ \;\;\;\;\;\;\;\;\;\;\HOLSymConst{\HOLTokenUnique{}}(\HOLBoundVar{u}\HOLSymConst{,}\HOLBoundVar{v}).\;\HOLConst{\HOLConst{odd}}\;\HOLBoundVar{u}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLConst{\HOLConst{even}}\;\HOLBoundVar{v}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLBoundVar{u}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLBoundVar{v}\HOLSymConst{\ensuremath{\sp{2}}}) \end{HOLmath} \begin{proof} For the if part $(\Rightarrow)$, existence is given by Theorem~\ref{thm:fermat-two-squares-exists}, and uniqueness is provided by Theorem~\ref{thm:fermat-two-squares-unique}. For the only-if part $(\Leftarrow)$, an odd prime with \HOLinline{\HOLFreeVar{n}\;\ensuremath{\not\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}} cannot be a sum of two squares by Theorem~\ref{thm:mod-4-not-squares}. \end{proof} \section{Two Squares Algorithm} \label{sec:algorithm} To make Zagier's proof constructive, we need to compute that single triple fixed by flip map. Let $n$ be a prime of the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}. By Theorem~\ref{thm:zagier-fixes-prime}, the only Zagier fixed point is \HOLinline{\HOLFreeVar{u}\;\HOLSymConst{=}\;(\HOLNumLit{1}\HOLSymConst{,}\HOLNumLit{1}\HOLSymConst{,}\HOLFreeVar{k})}, meaning \HOLinline{\HOLConst{zagier}\;\HOLFreeVar{u}\;\HOLSymConst{=}\;\HOLFreeVar{u}}. To change the triple $u$, applying \HOLConst{flip} is the obvious choice. To keep changing the triple, \HOLConst{zagier} should be applied. Thus by applying the composition \HOLinline{\HOLConst{zagier}\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{flip}} repeatedly from the known Zagier fixed point, there is hope that the chain will lead to the only flip fixed point. Figure~\ref{fig:zagier-flip-29} shows that this is indeed the case for \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}}. \begin{figure*}[h] \begin{center} \begin{tikzpicture}[scale=0.22] \draw[step=1, color=white!60!black] (0,0) grid (65,17); \windmill{8}{8}{1}{1}{7} \windmill{24}{8}{1}{7}{1} \windmill{38}{7}{3}{1}{5} \windmill{50}{7}{3}{5}{1} \windmill{58}{6}{5}{1}{1} \node at (4,1) {$(1,1,7)$}; \node at (22,1) {$(1,7,1)$}; \node at (38,1) {$(3,1,5)$}; \node at (51,1) {$(3,5,1)$}; \node at (60,1) {$(5,1,1)$}; \mind[ultra thick]{8}{8}{1} \mind{23}{7}{3} \mind{38}{7}{3} \mind{49}{6}{5} \mind{58}{6}{5} \begin{scope}[scale=10] \coordinate (a) at (1.0,1.1); \coordinate (b) at (2.2,1.1); \draw[->] (a) to [bend left] node[midway,above] {\HOLConst{flip}} (b); \coordinate (a) at (2.8,0.6); \coordinate (b) at (3.8,0.6); \draw[->] (a) to [bend right] node[midway,below] {\HOLConst{zagier}} (b); \coordinate (a) at (4.2,1.2); \coordinate (b) at (5.2,1.2); \draw[->] (a) to [bend left] node[midway,above] {\HOLConst{flip}} (b); \coordinate (a) at (5.2,0.5); \coordinate (b) at (6.0,0.5); \draw[->] (a) to [bend right] node[midway,below] {\HOLConst{zagier}} (b); \end{scope} \end{tikzpicture} \caption{The iteration chain of \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{29}} by the composition \HOLinline{\HOLConst{zagier}\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{flip}}, from Zagier fix to flip fix.} \Description{This figure show the iteration chain of n = 29, by the composition of first flip then Zagier. The chain starts from Zagier fix, ends in flip fix.} \label{fig:zagier-flip-29} \end{center} \end{figure*} In terms of windmills, the flip map keeps the central square, but flips the arms of rectangles from $y$-by-$z$ to $z$-by-$y$. This generally changes the mind of the windmill. The Zagier map keeps the mind, but changes the central square. Similar to the mind being an invariant of the Zagier map, the absolute difference $\left|y - z\right|$ is an invariant of the flip map. If the Zagier map can reduce this difference, the successive iterations of \HOLinline{\HOLConst{zagier}\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{flip}} will be able to locate the flip fixed point. \subsection{Flip Fix Search} \label{sec:flip-fix-search} To find the fixed point of the flip map, we can experiment with this pseudo-code: \bigskip \fbox{\begin{minipage}{0.36\textwidth} \begin{list}{$\circ$}{} \item \emph{Input}: a number \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}. \item \emph{Output}: a triple fixed by the flip map. \item \emph{Method}: \item start with \HOLinline{\HOLFreeVar{u}\;\HOLSymConst{=}\;(\HOLNumLit{1}\HOLSymConst{,}\HOLNumLit{1}\HOLSymConst{,}\HOLFreeVar{k})}, the Zagier fix. \item while ($u$ is not a flip fix) : \item \qquad $u \leftarrow \HOLinline{(\HOLConst{zagier}\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{flip})\;\HOLFreeVar{u}}$ \item end while. \end{list} \end{minipage}} \bigskip \noindent In an HOL4 interactive session, this pseudo-code can be implemented directly as:\footnote{This pseudo-code can be implemented directly in any programming language that supports while-loops and tuples.} \begin{definition} \label{def:two-sq-def} Computing the flip fixed point of \HOLinline{\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}} using a \HOLConst{WHILE} loop. \begin{HOLmath} \;\;\HOLConst{two_sq}\;\HOLFreeVar{n}\;\HOLTokenDefEquality{}\\ \;\;\;\;\HOLConst{WHILE}\;((\HOLSymConst{\HOLTokenNeg{}})\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{found})\;(\HOLConst{zagier}\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{flip})\;(\HOLNumLit{1}\HOLSymConst{,}\HOLNumLit{1}\HOLSymConst{,}\HOLFreeVar{n}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4}),\\ \quad\textrm{where} \;\;\HOLConst{found}\;(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})\;\HOLTokenDefEquality{}\;\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{z} \end{HOLmath} \end{definition} \noindent This simple while-loop may or may not terminate. We shall take up this issue in Section~\ref{sec:termination}. For primes of the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}, it terminates and seems to work. To prove its correctness, we shall develop a theory of permutation iteration, then apply the theory to this algorithm. \section{Permutation Orbits} \label{sec:orbits} In general, the composition of two involutions is no longer an involution, but just a permutation. Let $\varphi \colon S \rightarrow S$ be a permutation, a bijection on the set $S$, denoted by \HOLinline{\HOLFreeVar{\varphi}\;\HOLConst{\HOLConst{permutes}}\;\HOLFreeVar{S}}. For an element \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}} the iteration sequence $\varphi(x)$, \HOLinline{\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLNumLit{2}}(\HOLFreeVar{x})}}, \HOLinline{\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLNumLit{3}}(\HOLFreeVar{x})}}, \textit{etc.}, form its \emph{orbit}. The smallest positive index $n$ such that \HOLinline{\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLFreeVar{n}}(\HOLFreeVar{x})}\;\HOLSymConst{=}\;\HOLFreeVar{x}} is called the \emph{period} of $x$ under $\varphi$. If such a positive index does not exist, the period is defined to be $0$. In HOL4, the definition makes use of \HOLConst{OLEAST}, the optional \HOLConst{LEAST} operator: \begin{definition} \label{def:period-def} The period of function iteration of an element is the least nonzero index for the element iterate to wrap around, otherwise zero. \begin{HOLmath} \;\;\HOLConst{\HOLConst{period}}\;\HOLFreeVar{\varphi}\;\HOLFreeVar{x}\;\HOLTokenDefEquality{}\\ \;\;\;\;\HOLKeyword{case}\;\HOLConst{OLEAST}\;\HOLBoundVar{k}.\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLBoundVar{k}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLBoundVar{k}}(\HOLFreeVar{x})}\;\HOLSymConst{=}\;\HOLFreeVar{x}\;\HOLKeyword{of}\\ \;\;\;\;\HOLTokenBar{}\;\HOLConst{\HOLConst{none}}\;\ensuremath{\triangleright}\;\HOLNumLit{0}\\ \;\;\;\;\HOLTokenBar{}\;\HOLConst{\HOLConst{some}}\;\HOLBoundVar{k}\;\ensuremath{\triangleright}\;\HOLBoundVar{k} \end{HOLmath} \end{definition} \noindent When the set $S$ is finite, the iterates cannot be always distinct. Thus the permutation orbit of any $x \in S$ is finite, with a nonzero period, denoted by \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{period}\;\HOLFreeVar{\varphi}\;\HOLFreeVar{x}}: \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{\varphi}\;\HOLConst{\HOLConst{permutes}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLSymConst{\HOLTokenExists{}}\HOLBoundVar{p}.\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLBoundVar{p}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLBoundVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;\HOLFreeVar{\varphi}\;\HOLFreeVar{x} \end{HOLmath} and by definition the period is minimal, which means that there is no wrap around for element iterates when the index is less than the period: \begin{HOLmath} \HOLTokenTurnstile{}\HOLNumLit{0}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{j}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{j}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLConst{\HOLConst{period}}\;\HOLFreeVar{\varphi}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLFreeVar{j}}(\HOLFreeVar{x})}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLFreeVar{x} \end{HOLmath} This implies a criterion for an exponent index to be divisible by period: \begin{theorem} \label{thm:iterate-period-mod} \script{iteration}{653} For a nonzero period $p$ of $x$, $x$ is fixed by the $k$-th iterate of $\varphi$ if and only if $k$ is a multiple of period~$p$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLNumLit{0}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;\HOLFreeVar{\varphi}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;(\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLFreeVar{k}}(\HOLFreeVar{x})}\;\HOLSymConst{=}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLFreeVar{k}\;\ensuremath{\equiv}\;\HOLNumLit{0}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLFreeVar{p}\ensuremath{)}) \end{HOLmath} \end{theorem} \noindent Moreover, the period is the same for all iterates in the same orbit: \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{\varphi}\;\HOLConst{\HOLConst{permutes}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLFreeVar{j}}(\HOLFreeVar{x})}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLConst{\HOLConst{period}}\;\HOLFreeVar{\varphi}\;\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;\HOLFreeVar{\varphi}\;\HOLFreeVar{x} \end{HOLmath} \subsection{Involution Composition} \label{sec:involution-composition} When the permutation \HOLinline{\HOLFreeVar{\varphi}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}, a composition of two involutions $f$ and \HOLinline{\HOLFreeVar{g}}, we shall investigate whether their fixed points are connected by a chain of composition iterations. Note the following pattern of function application: \begin{equation*} \label{eqn:function-assoc} \begin{split} f\ \circ\ (\HOLinline{\HOLFreeVar{g}}\ \circ\ f)\ \circ\ (\HOLinline{\HOLFreeVar{g}}\ \circ\ f)\ \circ\ (\HOLinline{\HOLFreeVar{g}}\ \circ\ f)\\ \ee (f\ \circ\ \HOLinline{\HOLFreeVar{g}})\ \circ\ (f\ \circ\ \HOLinline{\HOLFreeVar{g}})\ \circ\ (f\ \circ\ \HOLinline{\HOLFreeVar{g}})\ \circ\ f \end{split} \end{equation*} by associativity. Also, $(\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}})^{-1} \ee \HOLinline{\HOLFreeVar{g}}^{-1}\ \circ\ f^{-1} \ee \HOLinline{\HOLFreeVar{g}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{f}}$ for involutions, so inverse is just reversal of application order in this case. Let \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}} for \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}}. With these notations, we can establish some basic results: \begin{theorem} \label{thm:involute-period-1} \script{iterateCompose}{558} When $f$ fixes $x$, the period for $x$ is $1$ if and only if \HOLinline{\HOLFreeVar{g}} also fixes $x$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;(\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLNumLit{1}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}) \end{HOLmath} \end{theorem} \noindent Pick an element $x$ in the set $S$. For involutions, an iterate of (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) can be equal to another iterate of (\HOLinline{\HOLFreeVar{g}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{f}}): \begin{theorem} \label{thm:involute-mod-period} \script{iterateCompose}{401} The $i$-th iterate of (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) equals the $j$-th iterate of (\HOLinline{\HOLFreeVar{g}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{f}}) if and only if (\HOLinline{\HOLFreeVar{i}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{j}}) is a multiple of period $p$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;((\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{i}}(\HOLFreeVar{x})}\;\HOLSymConst{=}\;(\HOLFreeVar{g}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{f})\ensuremath{\sp{\HOLFreeVar{j}}(\HOLFreeVar{x})}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLFreeVar{i}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{j}\;\ensuremath{\equiv}\;\HOLNumLit{0}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLFreeVar{p}\ensuremath{)}) \end{HOLmath} \end{theorem} \noindent When $f$ fixes point $x$, the iterates \HOLinline{(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{i}}(\HOLFreeVar{x})}} and \HOLinline{(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{j}}(\HOLFreeVar{x})}} are related when the sum (\HOLinline{\HOLFreeVar{i}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{j}}) is special: \begin{theorem} \label{thm:involute-two-fix-orbit-1} \script{iterateCompose}{583} When $f$ fixes $x$, the $i$-th and $j$-th iterate of (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) differ by one $f$ application if and only if (\HOLinline{\HOLFreeVar{i}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{j}}) is a multiple of period $p$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;((\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{i}}(\HOLFreeVar{x})}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;((\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{j}}(\HOLFreeVar{x})})\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLFreeVar{i}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{j}\;\ensuremath{\equiv}\;\HOLNumLit{0}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLFreeVar{p}\ensuremath{)}) \end{HOLmath} \end{theorem} \noindent There is a related result, with a similar proof: \begin{theorem} \label{thm:involute-two-fix-orbit-2} \script{iterateCompose}{689} When $f$ fixes $x$, the $i$-th and $j$-th iterate of (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) differ by one \HOLinline{\HOLFreeVar{g}} application if and only if (\HOLinline{\HOLFreeVar{i}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{j}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}) is a multiple of period $p$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;((\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{i}}(\HOLFreeVar{x})}\;\HOLSymConst{=}\;\HOLFreeVar{g}\;((\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{j}}(\HOLFreeVar{x})})\;\HOLSymConst{\HOLTokenEquiv{}}\\ \;\;\;\;\;\;\;\;\;\;\HOLFreeVar{i}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{j}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}\;\ensuremath{\equiv}\;\HOLNumLit{0}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLFreeVar{p}\ensuremath{)}) \end{HOLmath} \end{theorem} \noindent These theorems are useful in the study of iteration orbits starting from fixed points. \subsection{Period Parity} \label{sec:period-parity} Given a finite set $S$, and an element $x \in S$, the iterates \HOLinline{(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{j}}(\HOLFreeVar{x})}} form an orbit, with length equal to the period \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}}. Figure~\ref{fig:orbits-even-odd} shows two orbits, one with an even period, the other with an odd period. \begin{figure*}[h] \centering \begin{tikzpicture}[scale=1.7, ele/.style={fill=black,circle,minimum width=.8pt,inner sep=1pt}, every fit/.style={ellipse,draw,inner sep=5pt}] \node[ele,label=left:{\tiny{$x$}}] (a) at (0,0.5) {}; \node[ele,label=below:{\tiny{$\varphi\ x$}}] (b) at (1,0) {}; \node[ele,label=below:{\tiny{$\varphi^{2}\ x$}}] (c) at (2,0) {}; \node[ele,label=right:{\tiny{$\varphi^{3}\ x$}}] (d) at (3,0.5) {}; \node[ele,label=above:{\tiny{$\varphi^{4}\ x$}}] (e) at (2,1) {}; \node[ele,label=above:{\tiny{$\varphi^{5}\ x$}}] (f) at (1,1) {}; \node[ele,label=above:{\tiny{$\varphi^{6}\ x\qquad$}}] (g) at (0,0.5) {}; \draw[->,dashed,shorten <=2pt,shorten >=2] (a) -- (b); \draw[->,dashed,shorten <=2pt,shorten >=2] (b) -- (c); \draw[->,dashed,shorten <=2pt,shorten >=2] (c) -- (d); \draw[->,dashed,shorten <=2pt,shorten >=2] (d) -- (e); \draw[->,dashed,shorten <=2pt,shorten >=2] (e) -- (f); \draw[->,dashed,shorten <=2pt,shorten >=2] (f) -- (g); \node[ele,draw,fill=white] (ab) at ($(a)!0.5!(b)$) {}; \node[ele,draw,fill=white] (bc) at ($(b)!0.5!(c)$) {}; \node[ele,draw,fill=white] (cd) at ($(c)!0.5!(d)$) {}; \node[ele,draw,fill=white] (de) at ($(d)!0.5!(e)$) {}; \node[ele,draw,fill=white] (ef) at ($(e)!0.5!(f)$) {}; \node[ele,draw,fill=white] (fg) at ($(f)!0.5!(g)$) {}; \draw[thick,color=purple] (a) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (ab); \draw[thick,color=blue] (ab) to [bend right] node[midway,below] {\tiny{$f$}} (b); \draw[thick,color=purple] (b) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (bc); \draw[thick,color=blue] (bc) to [bend right] node[midway,below] {\tiny{$f$}} (c); \draw[thick,color=purple] (c) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (cd); \draw[thick,color=blue] (cd) to [bend right] node[midway,below] {\tiny{$f$}} (d); \draw[thick,color=purple] (d) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (de); \draw[thick,color=blue] (de) to [bend right] node[midway,above] {\tiny{$f$}} (e); \draw[thick,color=purple] (e) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (ef); \draw[thick,color=blue] (ef) to [bend right] node[midway,above] {\tiny{$f$}} (f); \draw[thick,color=purple] (f) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (fg); \draw[thick,color=blue] (fg) to [bend right] node[midway,above] {\tiny{$f$}} (g); \end{tikzpicture} \begin{tikzpicture}[scale=1.7, ele/.style={fill=black,circle,minimum width=.8pt,inner sep=1pt}, every fit/.style={ellipse,draw,inner sep=5pt}] \node[ele,label=left:{\tiny{$x$}}] (a) at (0,0.5) {}; \node[ele,label=below:{\tiny{$\varphi\ x$}}] (b) at (1,0) {}; \node[ele,label=below:{\tiny{$\varphi^{2}\ x$}}] (c) at (2,0) {}; \node[ele,label=right:{\tiny{$\varphi^{3}\ x$}}] (d) at (3,0.5) {}; \node[ele,label=above:{\tiny{$\varphi^{4}\ x$}}] (e) at (2.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{5}\ x$}}] (f) at (1.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{6}\ x$}}] (g) at (0.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{7}\ x\qquad$}}] (h) at (0,0.5) {}; \draw[->,dashed,shorten <=2pt,shorten >=2] (a) -- (b); \draw[->,dashed,shorten <=2pt,shorten >=2] (b) -- (c); \draw[->,dashed,shorten <=2pt,shorten >=2] (c) -- (d); \draw[->,dashed,shorten <=2pt,shorten >=2] (d) -- (e); \draw[->,dashed,shorten <=2pt,shorten >=2] (e) -- (f); \draw[->,dashed,shorten <=2pt,shorten >=2] (f) -- (g); \draw[->,dashed,shorten <=2pt,shorten >=2] (g) -- (h); \node[ele,draw,fill=white] (ab) at ($(a)!0.5!(b)$) {}; \node[ele,draw,fill=white] (bc) at ($(b)!0.5!(c)$) {}; \node[ele,draw,fill=white] (cd) at ($(c)!0.5!(d)$) {}; \node[ele,draw,fill=white] (de) at ($(d)!0.5!(e)$) {}; \node[ele,draw,fill=white] (ef) at ($(e)!0.5!(f)$) {}; \node[ele,draw,fill=white] (fg) at ($(f)!0.5!(g)$) {}; \node[ele,draw,fill=white] (gh) at ($(g)!0.5!(h)$) {}; \draw[thick,color=purple] (a) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (ab); \draw[thick,color=blue] (ab) to [bend right] node[midway,below] {\tiny{$f$}} (b); \draw[thick,color=purple] (b) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (bc); \draw[thick,color=blue] (bc) to [bend right] node[midway,below] {\tiny{$f$}} (c); \draw[thick,color=purple] (c) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (cd); \draw[thick,color=blue] (cd) to [bend right] node[midway,below] {\tiny{$f$}} (d); \draw[thick,color=purple] (d) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (de); \draw[thick,color=blue] (de) to [bend right] node[midway,above] {\tiny{$f$}} (e); \draw[thick,color=purple] (e) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (ef); \draw[thick,color=blue] (ef) to [bend right] node[midway,above] {\tiny{$f$}} (f); \draw[thick,color=purple] (f) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (fg); \draw[thick,color=blue] (fg) to [bend right] node[midway,above] {\tiny{$f$}} (g); \draw[thick,color=purple] (g) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (gh); \draw[thick,color=blue] (gh) to [bend right] node[midway,above] {\tiny{$f$}} (h); \end{tikzpicture} \caption{Orbits of \HOLinline{\HOLFreeVar{\varphi}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}} for point $x$. Left one has even period $6$, right one has odd period $7$.} \Description{This figure shows the orbits for point x of the composition: first g than f. Left one has even period 6, right one has odd period 7.} \label{fig:orbits-even-odd} \end{figure*} In the figure, black dots indicate iterates of \HOLinline{\HOLFreeVar{\varphi}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}, in dashes, and white dots indicate the intermediates, with \HOLinline{\HOLFreeVar{g}} first, then $f$, through the arcs. Since $f$ and \HOLinline{\HOLFreeVar{g}} are involutions, the arcs can go both ways: forward or backward. Let $\alpha$ denote a fixed point of $f$, and $\beta$ denote a fixed point of \HOLinline{\HOLFreeVar{g}}, \textit{i.e.}, \HOLinline{\HOLFreeVar{f}\;\HOLFreeVar{\alpha}\;\HOLSymConst{=}\;\HOLFreeVar{\alpha}}, and \HOLinline{\HOLFreeVar{g}\;\HOLFreeVar{\beta}\;\HOLSymConst{=}\;\HOLFreeVar{\beta}}. We shall look at how these fixed points are related, which is crucial in the correctness proof of our algorithm (see Definition~\ref{def:two-sq-def}). \subsection{Fixed Point Period Even} \label{sec:fixed-point-period-even} Consider an orbit with even period starting with $\alpha$, a fixed point of $f$. Figure~\ref{fig:orbit-fix-even} shows one on the left, and its real picture on the right. \begin{figure*}[h] \centering \begin{tikzpicture}[scale=1.7, ele/.style={fill=black,circle,minimum width=.8pt,inner sep=1pt}, every fit/.style={ellipse,draw,inner sep=5pt}] \node[ele,label=left:{\tiny{$\alpha$}}] (a) at (0,0.5) {}; \node[ele,label=below:{\tiny{$\varphi\ \alpha$}}] (b) at (1,0) {}; \node[ele,label=below:{\tiny{$\varphi^{2}\ \alpha$}}] (c) at (2,0) {}; \node[ele,label=right:{\tiny{$\varphi^{3}\ \alpha$}}] (d) at (3,0.5) {}; \node[ele,label=above:{\tiny{$\varphi^{4}\ \alpha$}}] (e) at (2,1) {}; \node[ele,label=above:{\tiny{$\varphi^{5}\ \alpha$}}] (f) at (1,1) {}; \node[ele,label=above:{\tiny{$\varphi^{6}\ \alpha\qquad$}}] (g) at (0,0.5) {}; \draw[->,dashed,shorten <=2pt,shorten >=2] (a) -- (b); \draw[->,dashed,shorten <=2pt,shorten >=2] (b) -- (c); \draw[->,dashed,shorten <=2pt,shorten >=2] (c) -- (d); \draw[->,dashed,shorten <=2pt,shorten >=2] (d) -- (e); \draw[->,dashed,shorten <=2pt,shorten >=2] (e) -- (f); \draw[->,dashed,shorten <=2pt,shorten >=2] (f) -- (g); \node[ele,draw,fill=white] (ab) at ($(a)!0.5!(b)$) {}; \node[ele,draw,fill=white] (bc) at ($(b)!0.5!(c)$) {}; \node[ele,draw,fill=white] (cd) at ($(c)!0.5!(d)$) {}; \node[ele,draw,fill=white] (de) at ($(d)!0.5!(e)$) {}; \node[ele,draw,fill=white] (ef) at ($(e)!0.5!(f)$) {}; \draw[thick,color=blue] (a) to [out=130, in=-130,looseness=50] node[midway,left] {\tiny{$f$}} (a); \draw[thick,color=purple] (a) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (ab); \draw[thick,color=blue] (ab) to [bend right] node[midway,below] {\tiny{$f$}} (b); \draw[thick,color=purple] (b) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (bc); \draw[thick,color=blue] (bc) to [bend right] node[midway,below] {\tiny{$f$}} (c); \draw[thick,color=purple] (c) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (cd); \draw[thick,color=blue] (cd) to [bend right] node[midway,below] {\tiny{$f$}} (d); \draw[thick,color=purple] (d) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (de); \draw[thick,color=blue] (de) to [bend right] node[midway,above] {\tiny{$f$}} (e); \draw[thick,color=purple] (e) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (ef); \draw[thick,color=blue] (ef) to [bend right] node[midway,above] {\tiny{$f$}} (f); \draw[thick,color=purple] (f) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (g); \DoubleLine[0.3pt]{f}{ab}{red}{red} \DoubleLine[0.3pt]{ef}{b}{red}{red} \DoubleLine[0.3pt]{e}{bc}{red}{red} \DoubleLine[0.3pt]{de}{c}{red}{red} \DoubleLine[0.3pt]{cd}{d}{red}{red} \path[ultra thick, glow=green] (cd) -- (d); \end{tikzpicture} \begin{tikzpicture}[scale=1.7, ele/.style={fill=black,circle,minimum width=.8pt,inner sep=1pt}, every fit/.style={ellipse,draw,inner sep=5pt}] \node[ele,label=left:{\tiny{$\alpha$}}] (a) at (0,0.5) {}; \node[ele,label=below:{\tiny{$\varphi\ \alpha$}}] (b) at (1,0) {}; \node[ele,label=below:{\tiny{$\varphi^{2}\ \alpha$}}] (c) at (2,0) {}; \node[ele,label=right:{\tiny{$\varphi^{3}\ \alpha$}}] (d) at (3,0.5) {}; \node[ele,label=above:{\tiny{$\varphi^{4}\ \alpha$}}] (e) at (2,1) {}; \node[ele,label=above:{\tiny{$\varphi^{5}\ \alpha$}}] (f) at (1,1) {}; \node[ele,label=above:{\tiny{$\varphi^{6}\ \alpha\qquad$}}] (g) at (0,0.5) {}; \draw[->,dashed,shorten <=2pt,shorten >=2] (a) -- (b); \draw[->,dashed,shorten <=2pt,shorten >=2] (b) -- (c); \draw[->,dashed,shorten <=2pt,shorten >=2] (c) -- (d); \draw[->,dashed,shorten <=2pt,shorten >=2] (d) -- (e); \draw[->,dashed,shorten <=2pt,shorten >=2] (e) -- (f); \draw[->,dashed,shorten <=2pt,shorten >=2] (f) -- (g); \draw[thick,color=blue] (a) to [out=130, in=-130,looseness=50] node[midway,left] {\tiny{$f$}} (a); \draw[thick,color=blue] (f) to [bend right] node[midway,left] {\tiny{$f$}} (b); \draw[thick,color=blue] (e) to [bend left] node[midway,right] {\tiny{$f$}} (c); \draw[thick,color=blue] (d) to [out=50, in=-50,looseness=50] node[midway,right] {\tiny{$f$}} (d); \draw[thick,color=purple] (c) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (d); \draw[thick,color=purple] (e) to [bend left] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (b); \draw[thick,color=purple] (f) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (g); \end{tikzpicture} \caption{Orbit from an $f$ fixed point $\alpha$ with even period $6$. Identical points on the left (marked by two parallel lines) are merged on the right (move white dot to black dot). In particular, on the left the two vertices of the shaded line are the same, forming a fixed point of $f$ on the right.} \Description{This figure shows an orbit from an f fixed point alpha, with even period 6. Identical points on the left, marked by two parallel lines, are merged on the right, by moving white dot to black dot. In particular, on the left the two vertices of the shaded line are the same, forming a fixed point of f on the right.} \label{fig:orbit-fix-even} \end{figure*} This orbit is formed by taking the left diagram of Figure~\ref{fig:orbits-even-odd}, but identifying the black dot on $\alpha$ (the leftmost one) with its preceding white dot from $f$, since \HOLinline{\HOLFreeVar{f}\;\HOLFreeVar{\alpha}\;\HOLSymConst{=}\;\HOLFreeVar{\alpha}}, giving the left $f$-loop. This node $\alpha$ is now preceded by two \HOLinline{\HOLFreeVar{g}}-arcs, one from a black dot and one from a white dot. However, \HOLinline{\HOLFreeVar{g}} is an involution, which is injective, so the two dots are identical. The same reasoning shows that all the dots linked by double lines are identical, so that the orbit on the left can be simplified to the one on the right, taking only black dots. Moreover, the rightmost black dot and a preceding white dot from $f$ must be the same, due to \HOLinline{\HOLFreeVar{g}}-arcs from identical dots. This means the half-period iterate, the rightmost black dot, is another fixed point of $f$, say $\alpha'$. Note that $\alpha' \ne \alpha$, for otherwise the period will be affected. This example motivates the following: \begin{theorem} \label{thm:involute-two-fixes-even} \script{iterateCompose}{884} When $f$ fixes $x$, and (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) has an even period $p$ for $x$, then $f$ also fixes \HOLinline{(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{p}\;\HOLConst{div}\;2}(\HOLFreeVar{x})}}, which is not $x$ itself. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{y}\;\HOLSymConst{=}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{p}\;\HOLConst{div}\;2}(\HOLFreeVar{x})}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLConst{\HOLConst{even}}\;\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLFreeVar{x} \end{HOLmath} \end{theorem} \begin{proof} First we show that $f$ fixes $y$. Let \HOLinline{\HOLFreeVar{h}\;\HOLSymConst{=}\;\HOLFreeVar{p}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{2}}. Since period $p$ is even, $p \ee \HOLinline{\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{h}\;\HOLSymConst{=}\;\HOLFreeVar{h}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{h}}$. This implies that \HOLinline{\HOLFreeVar{h}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{h}\;\ensuremath{\equiv}\;\HOLNumLit{0}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLFreeVar{p}\ensuremath{)}}, so \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLFreeVar{y}} by Theorem~\ref{thm:involute-two-fix-orbit-1}. Since (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) is a permutation, \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}}, so \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}}. Next we show that \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLFreeVar{x}}. Suppose \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{x}}. Since for finite $S$ the period \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}}, Theorem~\ref{thm:iterate-period-mod} shows that $p$ divides \HOLinline{\HOLFreeVar{h}\;\HOLSymConst{=}\;\HOLFreeVar{p}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{2}}. Hence \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLNumLit{1}}, which is not even. \end{proof} \noindent Therefore if a fixed point of $f$ has an even period under \HOLinline{\HOLFreeVar{\varphi}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}, it is not alone. This leads directly to: \begin{corollary} \label{cor:involute-fix-singleton-odd} \script{iterateCompute}{1009} If $f$ fixes only a single $x$, then \HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}} has an odd period for $x$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{=}\;\HOLTokenLeftbrace{}\HOLFreeVar{x}\HOLTokenRightbrace{}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLConst{\HOLConst{odd}}\;\HOLFreeVar{p} \end{HOLmath} \end{corollary} \subsection{Fixed Point Period Odd} \label{sec:fixed-point-period-odd} Now consider an orbit with odd period starting with $\alpha$, a fixed point of $f$. Figure~\ref{fig:orbit-fix-odd} shows one on the left, and its real picture on the right. This orbit is formed by taking the right diagram of Figure~\ref{fig:orbits-even-odd}, but identifying the black dot on $\alpha$ (the leftmost one) with its preceding white dot from $f$, since \HOLinline{\HOLFreeVar{f}\;\HOLFreeVar{\alpha}\;\HOLSymConst{=}\;\HOLFreeVar{\alpha}}, giving the left $f$-loop. The same reasoning as the even period orbit of Section~\ref{sec:fixed-point-period-even} shows that all the dots linked by double lines are identical, so that the orbit on the left can be simplified to the one on the right, again taking only black dots. \begin{figure*}[h] \centering \begin{tikzpicture}[scale=1.7, ele/.style={fill=black,circle,minimum width=.8pt,inner sep=1pt}, every fit/.style={ellipse,draw,inner sep=5pt}] \node[ele,label=left:{\tiny{$\alpha$}}] (a) at (0,0.5) {}; \node[ele,label=below:{\tiny{$\varphi\ \alpha$}}] (b) at (1,0) {}; \node[ele,label=below:{\tiny{$\varphi^{2}\ \alpha$}}] (c) at (2,0) {}; \node[ele,label=right:{\tiny{$\varphi^{3}\ \alpha$}}] (d) at (3,0.5) {}; \node[ele,label=above:{\tiny{$\varphi^{4}\ \alpha$}}] (e) at (2.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{5}\ \alpha$}}] (f) at (1.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{6}\ \alpha$}}] (g) at (0.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{7}\ \alpha\qquad$}}] (h) at (0,0.5) {}; \draw[->,dashed,shorten <=2pt,shorten >=2] (a) -- (b); \draw[->,dashed,shorten <=2pt,shorten >=2] (b) -- (c); \draw[->,dashed,shorten <=2pt,shorten >=2] (c) -- (d); \draw[->,dashed,shorten <=2pt,shorten >=2] (d) -- (e); \draw[->,dashed,shorten <=2pt,shorten >=2] (e) -- (f); \draw[->,dashed,shorten <=2pt,shorten >=2] (f) -- (g); \draw[->,dashed,shorten <=2pt,shorten >=2] (g) -- (h); \node[ele,draw,fill=white] (ab) at ($(a)!0.5!(b)$) {}; \node[ele,draw,fill=white] (bc) at ($(b)!0.5!(c)$) {}; \node[ele,draw,fill=white] (cd) at ($(c)!0.5!(d)$) {}; \node[ele,draw,fill=white] (de) at ($(d)!0.5!(e)$) {}; \node[ele,draw,fill=white] (ef) at ($(e)!0.5!(f)$) {}; \node[ele,draw,fill=white] (fg) at ($(f)!0.5!(g)$) {}; \draw[thick,color=blue] (a) to [out=130, in=-130,looseness=50] node[midway,left] {\tiny{$f$}} (a); \draw[thick,color=purple] (a) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (ab); \draw[thick,color=blue] (ab) to [bend right] node[midway,below] {\tiny{$f$}} (b); \draw[thick,color=purple] (b) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (bc); \draw[thick,color=blue] (bc) to [bend right] node[midway,below] {\tiny{$f$}} (c); \draw[thick,color=purple] (c) to [bend right] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (cd); \draw[thick,color=blue] (cd) to [bend right] node[midway,below] {\tiny{$f$}} (d); \draw[thick,color=purple] (d) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (de); \draw[thick,color=blue] (de) to [bend right] node[midway,above] {\tiny{$f$}} (e); \draw[thick,color=purple] (e) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (ef); \draw[thick,color=blue] (ef) to [bend right] node[midway,above] {\tiny{$f$}} (f); \draw[thick,color=purple] (f) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (fg); \draw[thick,color=blue] (fg) to [bend right] node[midway,above] {\tiny{$f$}} (g); \draw[thick,color=purple] (g) to [bend right] node[midway,above] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (h); \DoubleLine[0.3pt]{g}{ab}{red}{red} \DoubleLine[0.3pt]{fg}{b}{red}{red} \DoubleLine[0.3pt]{f}{bc}{red}{red} \DoubleLine[0.3pt]{ef}{c}{red}{red} \DoubleLine[0.3pt]{e}{cd}{red}{red} \DoubleLine[0.3pt]{de}{d}{red}{red} \path[ultra thick, glow=green] (de) -- (d); \end{tikzpicture} \begin{tikzpicture}[scale=1.7, ele/.style={fill=black,circle,minimum width=.8pt,inner sep=1pt}, every fit/.style={ellipse,draw,inner sep=5pt}] \node[ele,label=left:{\tiny{$\alpha$}}] (a) at (0,0.5) {}; \node[ele,label=below:{\tiny{$\varphi\ \alpha$}}] (b) at (1,0) {}; \node[ele,label=below:{\tiny{$\varphi^{2}\ \alpha$}}] (c) at (2,0) {}; \node[ele,label=right:{\tiny{$\varphi^{3}\ \alpha$}}] (d) at (3,0.5) {}; \node[ele,label=above:{\tiny{$\varphi^{4}\ \alpha$}}] (e) at (2.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{5}\ \alpha$}}] (f) at (1.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{6}\ \alpha$}}] (g) at (0.5,1) {}; \node[ele,label=above:{\tiny{$\varphi^{7}\ \alpha\qquad$}}] (h) at (0,0.5) {}; \draw[->,dashed,shorten <=2pt,shorten >=2] (a) -- (b); \draw[->,dashed,shorten <=2pt,shorten >=2] (b) -- (c); \draw[->,dashed,shorten <=2pt,shorten >=2] (c) -- (d); \draw[->,dashed,shorten <=2pt,shorten >=2] (d) -- (e); \draw[->,dashed,shorten <=2pt,shorten >=2] (e) -- (f); \draw[->,dashed,shorten <=2pt,shorten >=2] (f) -- (g); \draw[->,dashed,shorten <=2pt,shorten >=2] (g) -- (h); \draw[thick,color=purple] (f) to [bend left] node[midway,right] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (b); \draw[thick,color=purple] (g) to [bend left] node[midway,below] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (a); \draw[thick,color=purple] (e) to [bend left] node[midway,right] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (c); \draw[thick,color=purple] (d) to [out=50, in=-50,looseness=50] node[midway,right] {\tiny{\HOLinline{\HOLFreeVar{g}}}} (d); \draw[thick,color=blue] (a) to [out=130, in=-130,looseness=50] node[midway,left] {\tiny{$f$}} (a); \draw[thick,color=blue] (d) to [bend right] node[midway,above] {\tiny{$f$}} (e); \draw[thick,color=blue] (c) to [bend right] node[midway,right] {\tiny{$f$}} (f); \draw[thick,color=blue] (b) to [bend right] node[midway,left] {\tiny{$f$}} (g); \end{tikzpicture} \caption{Orbit from an $f$ fixed point $\alpha$ with odd period $7$. Identical points on the left (marked by two parallel lines) are merged~on the right (move white dot to black dot). In particular, on the left the two vertices of the shaded line are the same, forming a fixed point of $g$ on the right.} \Description{This figure shows an orbit from an f fixed point alpha with odd period 7. Identical points on the left, marked by two parallel lines, are merged on the right, by moving white dot to black dot. In particular, on the left the two vertices of the shaded line are the same, forming a fixed point of g on the right.} \label{fig:orbit-fix-odd} \end{figure*} Moreover, the rightmost black dot and a preceding white dot from \HOLinline{\HOLFreeVar{g}} must be the same, due to $f$-arcs from identical dots. This means the half-period iterate, the rightmost black dot, must be a fixed point of \HOLinline{\HOLFreeVar{g}}, say $\beta$. If \HOLinline{\HOLFreeVar{\beta}\;\HOLSymConst{=}\;\HOLFreeVar{\alpha}}, then period \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLNumLit{1}}, in accordance with Theorem~\ref{thm:involute-period-1}. This example motivates the following: \begin{theorem} \label{thm:involute-two-fixes-odd} \script{iterateCompose}{980} When $f$ fixes $x$, and (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) has an odd period $p$ for $x$, then \HOLinline{\HOLFreeVar{g}} fixes \HOLinline{(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{p}\;\HOLConst{div}\;2}(\HOLFreeVar{x})}}, which is not $x$ itself if and only if \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{1}}. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{y}\;\HOLSymConst{=}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLFreeVar{p}\;\HOLConst{div}\;2}(\HOLFreeVar{x})}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLConst{\HOLConst{odd}}\;\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;(\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLNumLit{1}) \end{HOLmath} \end{theorem} \begin{proof} First we show that \HOLinline{\HOLFreeVar{g}} fixes $y$. Let \HOLinline{\HOLFreeVar{h}\;\HOLSymConst{=}\;\HOLFreeVar{p}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{2}}. Since period $p$ is odd, $p \ee \HOLinline{\HOLNumLit{2}\HOLSymConst{\ensuremath{}}\HOLFreeVar{h}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}\;\HOLSymConst{=}\;\HOLFreeVar{h}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{h}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}$. Thus \HOLinline{\HOLFreeVar{h}\;\HOLSymConst{\ensuremath{+}}\;\HOLFreeVar{h}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}\;\ensuremath{\equiv}\;\HOLNumLit{0}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLFreeVar{p}\ensuremath{)}}, so \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{g}\;\HOLFreeVar{y}} by Theorem~\ref{thm:involute-two-fix-orbit-2}. As (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) is a permutation, \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLFreeVar{S}}, so \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}}. Theorem~\ref{thm:involute-period-1} ensures that: \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLNumLit{1}}. \end{proof} \subsection{Fixed Point Orbits} \label{sec:fixed-point-orbits} Let \HOLinline{\HOLFreeVar{\varphi}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}, and $\alpha, \beta$ be fixed points of $f, \HOLinline{\HOLFreeVar{g}}$, respectively. Theorem~\ref{thm:involute-two-fixes-even} and Theorem~\ref{thm:involute-two-fixes-odd} show that: \begin{itemize}[leftmargin=*] \item if the period $p$ of $\alpha$ is even, its orbit has another fixed point of $f$ at the \HOLinline{\HOLFreeVar{h}\;\HOLSymConst{=}\;\HOLFreeVar{p}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{2}} iterate: \HOLinline{\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLFreeVar{h}}(\HOLFreeVar{\alpha})}}. \item if the period $p$ of $\alpha$ is odd, its orbit has another fixed point of \HOLinline{\HOLFreeVar{g}} at the \HOLinline{\HOLFreeVar{h}\;\HOLSymConst{=}\;\HOLFreeVar{p}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{2}} iterate: \HOLinline{\HOLFreeVar{\beta}\;\HOLSymConst{=}\;\HOLFreeVar{\varphi}\ensuremath{\sp{\HOLFreeVar{h}}(\HOLFreeVar{\alpha})}}. \end{itemize} Figure~\ref{fig:orbit-fix-even} and Figure~\ref{fig:orbit-fix-odd} show that these orbits have no more fixed points. The only fixed point, of either $f$ or $g$, occurs at halfway point of the orbit. Thus, fixed point orbits lead directly from one fixed point to another. This is because, assuming one of the intermediate iterate is a fixed point, the iteration path will turn back, due to either $f$ or $g$, both being involutions. This will produce an orbit with a shorter period, but period for an orbit is minimal. Such considerations lead to the following stronger forms of Theorem~\ref{thm:involute-two-fixes-even} and Theorem~\ref{thm:involute-two-fixes-odd}: \begin{theorem} \label{thm:involute-two-fixes-even-odd} \script{iterateCompose}{1100} When $f$ fixes $x$, the $j$-th iterate of (\HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}) from $x$ is a fixed point of either $f$ or \HOLinline{\HOLFreeVar{g}} if and only if $j$ is half of the period $p$. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLConst{\HOLConst{even}}\;\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{j}.\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLBoundVar{j}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLBoundVar{j}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\;\;\;\;\;\;((\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLBoundVar{j}}(\HOLFreeVar{x})}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLBoundVar{j}\;\HOLSymConst{=}\;\HOLFreeVar{p}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{2})\\ \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLConst{\HOLConst{odd}}\;\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{j}.\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLBoundVar{j}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLBoundVar{j}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\;\;\;\;\;\;((\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\ensuremath{\sp{\HOLBoundVar{j}}(\HOLFreeVar{x})}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenEquiv{}}\;\HOLBoundVar{j}\;\HOLSymConst{=}\;\HOLFreeVar{p}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{2}) \end{HOLmath} \end{theorem} This completes our tour of the theory of permutation orbits and fixed points. The results provide the key to formally prove that our two-squares algorithm by iterations is correct. \section{Correctness of Algorithm} \label{sec:correctness} The algorithm to compute the flip fixed point from the known Zagier fixed point, given in Definition~\ref{def:two-sq-def}, makes use of a while-loop. A while-loop consists of a guard $G$ and a body $B$, starting with an element $x$. The body is a function on $x$, producing iterates $B(x)$, \HOLinline{\HOLFreeVar{B}\ensuremath{\sp{\HOLNumLit{2}}(\HOLFreeVar{x})}}, \HOLinline{\HOLFreeVar{B}\ensuremath{\sp{\HOLNumLit{3}}(\HOLFreeVar{a})}}, \textit{etc.}. The guard is a predicate on each iterate: the loop continues only if the test result by the guard stays true. In HOL4, the \HOLConst{WHILE} loop with guard $G$ and body $B$ starting with $x$ is defined as: \begin{HOLmath} \;\;\HOLConst{WHILE}\;\HOLFreeVar{G}\;\HOLFreeVar{B}\;\HOLFreeVar{x}\;\HOLTokenDefEquality{}\;\HOLKeyword{if}\;\HOLFreeVar{G}\;\HOLFreeVar{x}\;\HOLKeyword{then}\;\HOLConst{WHILE}\;\HOLFreeVar{G}\;\HOLFreeVar{B}\;(\HOLFreeVar{B}\;\HOLFreeVar{x})\;\HOLKeyword{else}\;\HOLFreeVar{x} \end{HOLmath} from which one can easily show by induction that: \begin{HOLmath} \HOLTokenTurnstile{}(\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{j}.\;\HOLBoundVar{j}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{k}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLFreeVar{G}\;(\HOLFreeVar{B}\ensuremath{\sp{\HOLBoundVar{j}}(\HOLFreeVar{x})}))\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLConst{WHILE}\;\HOLFreeVar{G}\;\HOLFreeVar{B}\;\HOLFreeVar{x}\;\HOLSymConst{=}\\ \;\;\;\;\;\;\;\;\;\HOLKeyword{if}\;\HOLFreeVar{G}\;(\HOLFreeVar{B}\ensuremath{\sp{\HOLFreeVar{k}}(\HOLFreeVar{x})})\;\HOLKeyword{then}\;\HOLConst{WHILE}\;\HOLFreeVar{G}\;\HOLFreeVar{B}\;(\HOLFreeVar{B}\ensuremath{\sp{\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}(\HOLFreeVar{x})})\;\HOLKeyword{else}\;\HOLFreeVar{B}\ensuremath{\sp{\HOLFreeVar{k}}(\HOLFreeVar{x})} \end{HOLmath} giving this expected result: \begin{theorem} \label{thm:iterate-while-thm} \script{iterateCompute}{922} The \HOLinline{\HOLConst{WHILE}} loop delivers the first body iterate that fails the guard test. \begin{HOLmath} \HOLTokenTurnstile{}(\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{j}.\;\HOLBoundVar{j}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{k}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLFreeVar{G}\;(\HOLFreeVar{B}\ensuremath{\sp{\HOLBoundVar{j}}(\HOLFreeVar{x})}))\;\HOLSymConst{\HOLTokenConj{}}\;\HOLSymConst{\HOLTokenNeg{}}\HOLFreeVar{G}\;(\HOLFreeVar{B}\ensuremath{\sp{\HOLFreeVar{k}}(\HOLFreeVar{x})})\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLConst{WHILE}\;\HOLFreeVar{G}\;\HOLFreeVar{B}\;\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{B}\ensuremath{\sp{\HOLFreeVar{k}}(\HOLFreeVar{x})} \end{HOLmath} \end{theorem} \subsection{Iterate with WHILE} \label{sec:iterate-while} From Section~\ref{sec:orbits}, we learn that for two involutions $f$ and \HOLinline{\HOLFreeVar{g}}, a fixed point $\alpha$ of $f$ is paired up with a fixed point $\beta$ of \HOLinline{\HOLFreeVar{g}} whenever the period of $\alpha$ under the composition \HOLinline{\HOLFreeVar{\varphi}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}} is odd. In fact, $\beta$ lies in the orbit of $\alpha$ at halfway point, the iterate at half period. Since a while-loop also gives an iterate, we have: \begin{theorem} \label{thm:involute-involute-fixes-while} \script{iterateCompose}{1536} For two involutions $f$ and \HOLinline{\HOLFreeVar{g}}, if $f$ fixes $x$ with an odd period, a WHILE loop with \HOLinline{\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}} from $x$ can reach a fixed point of \HOLinline{\HOLFreeVar{g}}. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{f}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{g}\;\HOLConst{involute}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\\ \;\;\;\;\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{f}\;\HOLFreeVar{S}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLConst{\HOLConst{odd}}\;\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLConst{WHILE}\;(\HOLTokenLambda{}\HOLBoundVar{t}.\;\HOLFreeVar{g}\;\HOLBoundVar{t}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLBoundVar{t})\;(\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g})\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S} \end{HOLmath} \end{theorem} \begin{proof} Let guard \HOLinline{\HOLFreeVar{G}\;\HOLSymConst{=}\;(\HOLTokenLambda{}\HOLBoundVar{t}.\;\HOLFreeVar{g}\;\HOLBoundVar{t}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLBoundVar{t})}, and body \HOLinline{\HOLFreeVar{B}\;\HOLSymConst{=}\;\HOLFreeVar{f}\;\HOLSymConst{\HOLTokenCompose}\;\HOLFreeVar{g}}. If period \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLNumLit{1}}, then \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}} by Theorem~\ref{thm:involute-period-1}. So \HOLinline{\HOLSymConst{\HOLTokenNeg{}}\HOLFreeVar{G}\;\HOLFreeVar{x}}, and \HOLinline{\HOLConst{WHILE}\;\HOLFreeVar{G}\;\HOLFreeVar{B}\;\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{x}} since the condition is not met at the start. Therefore \HOLinline{\HOLConst{WHILE}\;\HOLFreeVar{G}\;\HOLFreeVar{B}\;\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{s}}. If period \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{1}}, let \HOLinline{\HOLFreeVar{h}\;\HOLSymConst{=}\;\HOLFreeVar{p}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{2}}, and \HOLinline{\HOLFreeVar{z}\;\HOLSymConst{=}\;\HOLFreeVar{B}\ensuremath{\sp{\HOLFreeVar{h}}(\HOLFreeVar{x})}}. Since \HOLinline{\HOLNumLit{1}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{p}}, $0 < h < p$. Also $f$ and \HOLinline{\HOLFreeVar{g}} are involutions, so \HOLinline{\HOLFreeVar{B}\;\HOLConst{\HOLConst{permutes}}\;\HOLFreeVar{S}}. Hence \HOLinline{\HOLFreeVar{z}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}} by Theorem~\ref{thm:involute-two-fixes-odd}. so \HOLinline{\HOLSymConst{\HOLTokenNeg{}}\HOLFreeVar{G}\;\HOLFreeVar{z}}. We claim \HOLinline{\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{j}.\;\HOLBoundVar{j}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{h}\;\HOLSymConst{\HOLTokenImp{}}\;\HOLFreeVar{G}\;(\HOLFreeVar{B}\ensuremath{\sp{\HOLBoundVar{j}}(\HOLFreeVar{x})})}. To see this, let \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{B}\ensuremath{\sp{\HOLFreeVar{j}}(\HOLFreeVar{x})}}, which is an element of $S$. If \HOLinline{\HOLFreeVar{j}\;\HOLSymConst{=}\;\HOLNumLit{0}}, then \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{x}}. Since period \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{1}}, \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenNotIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}} by Theorem~\ref{thm:involute-period-1}, so \HOLinline{\HOLFreeVar{G}\;\HOLFreeVar{y}}. If \HOLinline{\HOLFreeVar{j}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}}, then $0 < j < h < p$, and \HOLinline{\HOLFreeVar{j}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLFreeVar{h}}. Hence \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{\HOLTokenNotIn{}}\;\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}} by Theorem~\ref{thm:involute-two-fixes-even-odd}, so \HOLinline{\HOLFreeVar{G}\;\HOLFreeVar{y}} again. The claim is proved. By the claim and \HOLinline{\HOLSymConst{\HOLTokenNeg{}}\HOLFreeVar{G}\;\HOLFreeVar{z}}, apply Theorem~\ref{thm:iterate-while-thm} to conclude $\HOLinline{\HOLConst{WHILE}\;\HOLFreeVar{G}\;\HOLFreeVar{B}\;\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{z}} \in \HOLinline{\HOLConst{fixes}\;\HOLFreeVar{g}\;\HOLFreeVar{S}}$. \end{proof} \subsection{Two Squares by WHILE} \label{sec:two-squares-while} We have developed the theory to show that the algorithm in Section~\ref{sec:algorithm} is correct: \begin{theorem} \label{thm:two-sq-thm} \script{twoSquares}{840} For a prime of the form \HOLinline{\HOLNumLit{4}\HOLSymConst{\ensuremath{}}\HOLFreeVar{k}\;\HOLSymConst{\ensuremath{+}}\;\HOLNumLit{1}}, the two squares algorithm of Definiton~\ref{def:two-sq-def} gives a flip fixed point. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;\HOLConst{two_sq}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLConst{flip}\;(\HOLConst{mills}\;\HOLFreeVar{n}) \end{HOLmath} \end{theorem} \begin{proof} Let \HOLinline{\HOLFreeVar{S}\;\HOLSymConst{=}\;\HOLConst{mills}\;\HOLFreeVar{n}}, \HOLinline{\HOLFreeVar{\varphi}\;\HOLSymConst{=}\;\HOLConst{zagier}\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{flip}}, \HOLinline{\HOLFreeVar{u}\;\HOLSymConst{=}\;(\HOLNumLit{1}\HOLSymConst{,}\HOLNumLit{1}\HOLSymConst{,}\HOLFreeVar{n}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4})}, and period \HOLinline{\HOLFreeVar{p}\;\HOLSymConst{=}\;\HOLConst{\HOLConst{period}}\;\HOLFreeVar{\varphi}\;\HOLFreeVar{u}}. By Definition~\ref{def:two-sq-def}, and noting that \HOLinline{(\HOLSymConst{\HOLTokenNeg{}})\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{found}\;\HOLSymConst{=}\;(\HOLTokenLambda{}\HOLBoundVar{t}.\;\HOLConst{flip}\;\HOLBoundVar{t}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLBoundVar{t})}, this is to show: \HOLinline{\HOLConst{WHILE}\;(\HOLTokenLambda{}\HOLBoundVar{t}.\;\HOLConst{flip}\;\HOLBoundVar{t}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLBoundVar{t})\;\HOLFreeVar{\varphi}\;\HOLFreeVar{u}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLConst{flip}\;\HOLFreeVar{S}}. Since a prime is not a square, we have \HOLinline{\HOLConst{\HOLConst{finite}}\;\HOLFreeVar{S}}. Now \HOLinline{\HOLFreeVar{\varphi}\;\HOLConst{\HOLConst{permutes}}\;\HOLFreeVar{S}} as Zagier map and flip map are both involutions, by Theorem~\ref{thm:zagier-involute-mills-prime} and Theorem~\ref{thm:flip-involute-mills}, and \HOLinline{\HOLConst{fixes}\;\HOLConst{zagier}\;\HOLFreeVar{S}\;\HOLSymConst{=}\;\HOLTokenLeftbrace{}\HOLFreeVar{u}\HOLTokenRightbrace{}} by Theorem~\ref{thm:zagier-fixes-prime}. Thus \HOLinline{\HOLFreeVar{u}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLConst{zagier}\;\HOLFreeVar{S}}, and period $p$ is odd by Corollary~\ref{cor:involute-fix-singleton-odd}. So \HOLinline{\HOLConst{WHILE}\;(\HOLTokenLambda{}\HOLBoundVar{t}.\;\HOLConst{flip}\;\HOLBoundVar{t}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLBoundVar{t})\;\HOLFreeVar{\varphi}\;\HOLFreeVar{u}\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{fixes}\;\HOLConst{flip}\;\HOLFreeVar{S}} by Theorem~\ref{thm:involute-involute-fixes-while}. \end{proof} \noindent It is almost trivial to convert \HOLinline{\HOLConst{two_sq}\;\HOLFreeVar{n}} to following algorithm: \begin{definition} \label{def:two-squares-def} Compute the two squares for Fermat's two squares theorem. \begin{HOLmath} \;\;\HOLConst{two_squares}\;\HOLFreeVar{n}\;\HOLTokenDefEquality{}\;(\HOLKeyword{let}\;(\HOLBoundVar{x}\HOLSymConst{,}\HOLBoundVar{y}\HOLSymConst{,}\HOLBoundVar{z})\;=\;\HOLConst{two_sq}\;\HOLFreeVar{n}\;\HOLKeyword{in}\;(\HOLBoundVar{x}\HOLSymConst{,}\HOLBoundVar{y}\;\HOLSymConst{\ensuremath{+}}\;\HOLBoundVar{z})) \end{HOLmath} \end{definition} \noindent giving the two squares in a pair, and its correctness is readily demonstrated: \begin{theorem} \label{thm:two-squares-thm} \script{twoSquares}{1041} The algorithm by Definition~\ref{def:two-squares-def} gives indeed Fermat's two squares. \begin{HOLmath} \HOLTokenTurnstile{}\HOLConst{prime}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}\;\HOLSymConst{\HOLTokenImp{}}\\ \;\;\;\;\;\;\;(\HOLKeyword{let}\;(\HOLBoundVar{u}\HOLSymConst{,}\HOLBoundVar{v})\;=\;\HOLConst{two_squares}\;\HOLFreeVar{n}\;\HOLKeyword{in}\;\HOLFreeVar{n}\;\HOLSymConst{=}\;\HOLBoundVar{u}\HOLSymConst{\ensuremath{\sp{2}}}\;\HOLSymConst{\ensuremath{+}}\;\HOLBoundVar{v}\HOLSymConst{\ensuremath{\sp{2}}}) \end{HOLmath} \end{theorem} \begin{table*}[h] \caption{Running Fermat's two squares algorithm in a HOL4 session, with timing.} \Description{This table shows a sample run of Fermat's two squares algorithm in a HOL4 session, with timing information.} \label{tbl:sample-run} \begin{tabular}{p{0.8\textwidth}} \begin{verbatim} > time EVAL ``two_squares 97``; runtime: 0.00770s, gctime: 0.00086s, systime: 0.00077s. val it = |- two_squares 97 = (9,4): thm > time EVAL ``two_squares 1999999913``; runtime: 2m23s, gctime: 14.7s, systime: 11.3s. val it = |- two_squares 1999999913 = (1093,44708): thm > time EVAL ``two_squares 12345678949``; runtime: 6m02s, gctime: 37.5s, systime: 26.0s. val it = |- two_squares 12345678949 = (110415,12418): thm > EVAL ``9 * 9 + 4 * 4``; val it = |- 9 * 9 + 4 * 4 = 97: thm > EVAL ``1093 * 1093 + 44708 * 44708``; val it = |- 1093 * 1093 + 44708 * 44708 = 1999999913: thm > EVAL ``110415 * 110415 + 12418 * 12418``; val it = |- 110415 * 110415 + 12418 * 12418 = 12345678949: thm \end{verbatim} \end{tabular} \end{table*} Table~\ref{tbl:sample-run} shows a sample run in HOL4 session on a typical laptop, using \HOLConst{EVAL} for evaluation and prefix \HOLConst{time} to obtain timing statistics. Note that these \HOLConst{EVAL} executions are based on optimised symbolic rewriting in HOL4, thus orders of magnitude slower than running native code. \paragraph*{Other algorithms} A prime has a finite set of windmill triples, by Theorem~\ref{thm:mills-finite}. Fermat's two squares for a prime $n$ with \HOLinline{\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}}, which must exist by Theorem~\ref{thm:fermat-two-squares-exists}, can be found by a brute-force search: subtract $n$ by successive odd squares, and check whether the difference is a square. Although there are better ways to test a square than the square-root test, they are not simple to implement. Don Zagier, after his one-sentence proof, referred to an effective algorithm by Wagon~\cite{Wagon-1990-acm} to compute the two squares. The algorithm requires finding a quadratic non-residue of the given prime $n$. The advantage of our algorithm in Definition~\ref{def:two-squares-def} over such alternative methods is that only addition and subtraction are performed. The implementation is rather straightforward. The issue of termination is discussed next. \subsection{Terminating Condition} \label{sec:termination} As mentioned in Section~\ref{sec:flip-fix-search}, for our algorithm the WHILE loop may or may not terminate. To gaurantee termination, convert the WHILE loop to a countdown loop, as follows. First, ensure that the input number $n$ is not a square, so that \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}} is finite (Theorem~\ref{thm:mills-finite}), and check \HOLinline{\HOLFreeVar{n}\;\ensuremath{\equiv}\;\HOLNumLit{1}\;\ensuremath{(}\ensuremath{\bmod}\;\HOLNumLit{4}\ensuremath{)}}, so that \HOLinline{(\HOLNumLit{1}\HOLSymConst{,}\HOLNumLit{1}\HOLSymConst{,}\HOLFreeVar{n}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4})\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{mills}\;\HOLFreeVar{n}}, \textit{i.e.}, \HOLinline{\HOLConst{mills}\;\HOLFreeVar{n}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLSymConst{\HOLTokenEmpty{}}}. Obviously for any triple \HOLinline{(\HOLFreeVar{x}\HOLSymConst{,}\HOLFreeVar{y}\HOLSymConst{,}\HOLFreeVar{z})\;\HOLSymConst{\HOLTokenIn{}}\;\HOLConst{mills}\;\HOLFreeVar{n}}, each $x, y$ or $y$ is less than $n$, hence \HOLinline{\ensuremath{|}\HOLConst{mills}\;\HOLFreeVar{n}\ensuremath{|}\;\HOLSymConst{\HOLTokenLt{}}\;\HOLFreeVar{n}\HOLSymConst{\ensuremath{\sp{3}}}}. Now, use a countdown loop from \HOLinline{\HOLFreeVar{n}\HOLSymConst{\ensuremath{\sp{3}}}} to $0$, start with the triple \HOLinline{(\HOLNumLit{1}\HOLSymConst{,}\HOLNumLit{1}\HOLSymConst{,}\HOLFreeVar{n}\;\HOLConst{\HOLConst{div}}\;\HOLNumLit{4})} for the \HOLinline{\HOLConst{zagier}\;\HOLSymConst{\HOLTokenCompose}\;\HOLConst{flip}} iteration. The iterations trace an orbit. At half-way point, the orbit hits either a flip fixed point, detected by \HOLinline{\HOLFreeVar{y}\;\HOLSymConst{=}\;\HOLFreeVar{z}}, when the period is odd (Theorem~\ref{thm:involute-two-fixes-odd}), or another Zagier fixed point, detected by \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{=}\;\HOLFreeVar{y}}, when the period is even (Theorem~\ref{thm:involute-two-fixes-even}). They provide actual exits from the countdown loop, much earlier than the count drops to zero. \subsection{Lessons Learnt} \label{sec:lessons} This formalisation work can be a self-contained project in a theorem-proving workshop. The ideas are simple, but formulating the theorems properly is not simple. For example, at first the author would like to prove: \begin{equation*} \label{eqn:zagier-inv} \HOLinline{\HOLSymConst{\HOLTokenForall{}}\HOLBoundVar{x}\;\HOLBoundVar{y}\;\HOLBoundVar{z}.\;\HOLConst{zagier}\;(\HOLConst{zagier}\;(\HOLBoundVar{x}\HOLSymConst{,}\HOLBoundVar{y}\HOLSymConst{,}\HOLBoundVar{z}))\;\HOLSymConst{=}\;(\HOLBoundVar{x}\HOLSymConst{,}\HOLBoundVar{y}\HOLSymConst{,}\HOLBoundVar{z})}. \end{equation*} The interactive session produces several subgoals which he cannot resolve immediately. A comparison of Definition~\ref{def:zagier-def} with Equation~\eqref{eqn:zagier-map} shows differences in boundary cases. Finally, some insight from windmills resolves why the boundaries are ignored, and provides the pre-condition \HOLinline{\HOLFreeVar{x}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}\;\HOLSymConst{\HOLTokenConj{}}\;\HOLFreeVar{z}\;\HOLSymConst{\HOLTokenNotEqual{}}\;\HOLNumLit{0}}, see Equation~\eqref{eqn:zagier-involute}. The result is Theorem~\ref{thm:zagier-involute-mills-prime}. Explaining the Zagier map is an involution through the mind of a windmill poses some challenges. Definition~\ref{def:zagier-def} of the Zagier map has $3$ branches, so the initial effort is to treat just $3$ cases. The first case is immediate, but the second case runs into a mess. It is only after drawing a lot of windmills that the author realises these finer points: \begin{itemize}[leftmargin=*] \item there are $3$ types: $x < y, x \ee y$, and $x > y$ for the windmill triple $(x,y,z)$, \item the $3$ types are further subdivided due to geometry of the mind, giving $5$ cases in total, \item the $5$ cases can be condensed into $3$ branches, as the definition shows. \end{itemize} The result is Table~\ref{tbl:zagier-map} in Section~\ref{sec:windmill-mind}. For the permutation orbits in Section~\ref{sec:orbits}, the proofs about relations between iterates start as long-winded arguments treating if-part and only-if part separately. Putting them in this paper prompts the author to rethink the logic. The polished proofs simply employ a chain of logical equivalences. Fixed point orbits have either even or odd period, as treated in Section~\ref{sec:fixed-point-period-even} and Section~\ref{sec:fixed-point-period-odd}. The drawing of the diagrams helps to refine the proofs to be short and sweet, making good use of theorems already proved. About the correctness proof of the algorithm using a while-loop in Section~\ref{sec:correctness}, the author initially applied Hoare logic assertions to derive the desired iterate upon loop exit. \ifdefined This is awkward, as pointed out by Michael Norrish who knows the HOL4 theorem-prover inside out. \else This is awkward, as pointed out by (omitted for anonymous review). \fi The reason is that \HOLinline{\HOLConst{WHILE}} is \emph{defined} as iteration of the body in HOL4. The section had since been rewritten. \paragraph*{Development Effort} The proofs have been streamlined after several revisions. Such refinements result in the script line counts for various theories developed, shown in Table~\ref{tbl:hol4-line-counts}. \begin{table}[h] \caption{Statistics of various theories in this work.} \Description{This table gives the statistics of various theories in this formalisation work.} \label{tbl:hol4-line-counts} \[ \begin{array}{llr} \text{HOL4 Theory} & \text{Description} & \text{\#Lines}\\ \hline \text{involute} & \text{basic involution} & 231\\ \text{iteration} & \text{function iteration and period} & 917\\ \text{iterateCompose} & \text{iteration of involute composition} & 1648\\ \text{iterateCompute} & \text{iteration period computation} & 939\\ \text{windmill} & \text{windmills and their involutions} & 1844\\ \text{twoSquares} & \text{two-squares by windmills} & 1317\\ \end{array} \] \end{table} \noindent The scripts are fully documented, including the traditional proofs as comment before each theorem. Although comments almost double the script size, the line counts are still indicative of the effort to convert ideas into formal proofs. \subsection{Related Work} \label{sec:related-work} As noted in Section~\ref{sec:introduction}, Fermat's two squares theorem has been formalised. However, none of these formal proofs is constructive, in the sense that there is no formal proof of an algorithm to compute the two squares for a prime satisfying the theorem. Fermat's two squares theorem has two parts: existence and uniqueness. All formal proofs include the existence part (see Theorem~\ref{thm:fermat-two-squares-exists}) , using classic and modern existence proofs: the method of infinite descent is used in one system, Gaussian integers are employed in three systems, both Heath-Brown's proof and Zagier's proof are treated in two systems. Only two formal proofs include the uniqueness part (see Theorem~\ref{thm:fermat-two-squares-unique}): Th{\'e}ry~\cite{Thery-2004} proved by algebraic identities and divisibility, and Hughes~\cite{Hughes-2019} proved by unique factorisation of Gaussian integers. Recently, Dubach and Muehlboeck~\cite{Dubach-Muehlboeck-2021-acm} formalised Zagier's proof using involutions in Coq's Mathematical Components Library. They illustrated their proof using the windmills as per this paper, and extended the use of involutions on the same set to formalise also an integer-partition proof of Fermat's two squares theorem by Christopher~\cite{Christopher-2016-acm}. A summary of these formal proofs, in chronological order, is given in Table~\ref{tbl:chronology-formalise-two-squares}. \begin{table*}[h] \caption{Chronology of formalisation of Fermat's two squares theorem.} \Description{This table lists, in chronological order, the formal proofs of Fermat's two squares theorem, by various authors in different theorem provers.} \label{tbl:chronology-formalise-two-squares} \begin{tabular}{llll} Year & Author(s)[reference] & Theorem Prover & Comment\\ \hline 2004 & Laurent Th{\'e}ry~\cite{Thery-2004} & Coq & Gaussian integers, with uniqueness\\ 2007 & Roelof Oosterhuis~\cite{Oosterhuis-2007} & Isabelle & Euler's proof with infinite descent\\ 2009 & Marco Riccardi~\cite{Riccardi-2009} & Mizar & Heath-Brown's proof with involutions\\ 2010 & John Harrison~\cite{Harrison-2010} & HOL Light & Zagier's proof with involutions\\ 2012 & Anthony Narkawicz~\cite{Narkawicz-2012-acm} & NASA PVS & Zagier's proof with involutions\\ 2015 & Mario Carneiro~\cite{Carneiro-2015} & MetaMath & Gaussian integers\\ 2016 & Rob Arthan~\cite{Arthan-2016} & ProofPower & Heath-Brown's proof with involutions\\ 2019 & Chris Hughes~\cite{Hughes-2019} & Lean & Principal Ideal Ring of Gaussian integers, with uniqueness\\ 2021 & Dubach and Muehlboeck~\cite{Dubach-Muehlboeck-2021-acm} & Coq & Zagier's and Christopher's proofs with involutions\\ \end{tabular} \end{table*} \section{Conclusion} \label{sec:conclusion} About Fermat's two squares theorem, G. H. Hardy wrote in his 1940 essay \emph{A Mathematician's Apology}~\cite[Section 13]{Hardy-1940-acm}: \bigskip \begin{mquote}[1em] \textit{This is Fermat's theorem, which is ranked, very justly, as one of the finest of arithmetic. Unfortunately, there is no proof within the comprehension of anybody but a fairly expert mathematician.} \end{mquote} \bigskip \noindent This work has been a rewarding exercise in formalisation, delivering a proof of Fermat's Theorem~\ref{thm:fermat-two-squares-thm} using only natural numbers, involutions, and counting. There is a certain sense of mathematical beauty when a non-trivial result can be shown by elementary means, borrowing elegant ideas by Zagier and Spivak. Moreover, by developing a theory of involution iteration, an algorithm to compute the two squares of the theorem can be formally shown to be correct. \paragraph*{Future Work} The theory in Section~\ref{sec:orbits}, about orbits and fixed points, can be developed using group actions, since the iteration indices form an addition cyclic group under \HOLConst{mod} $p$, where $p$ is the orbit period. One can exploit the symmetry in permutation orbits, especially for permutations arising from two involutions, to improve the algorithm, as shown in the analysis by Shiu~\cite{Shiu-1996}. In HOL4, this direction can start from the algebra of group theory in Chan and Norrish~\cite{Chan-Norrish-cpp-2012}. A formal analysis of the performance of the algorithm for two squares described in Definition~\ref{def:two-sq-def} can be modelled using an approach in Chan~\cite{Chan-ANU-2019-acm}. \ifdefined \section*{Acknowledgements} \label{sec:acknowledgements} \addcontentsline{toc}{section}{Acknowledgments} Many thanks to Michael Norrish for his careful review of the draft, providing useful advice and helpful recommendations to improve this paper. The author is also grateful to the anonymous reviewers who pointed out typographical errors and suggested clarifications. This paper has been revised to incorporate their comments. \else \fi \ifdefined \else \section*{Appendices}

https://arxiv.org/abs/1804.03688

Jensen-type geometric shapes

We present both necessary and sufficient conditions to the convex closed shape $X$ such that the inequality $$ \frac{1}{|X|} \int_X f(x)\:dx \le \frac{1}{|\partial X|} \int_{\partial X} f(x)\:dx$$ is valid for every convex function $f \colon X \to \mathbb{R}$ ($\partial X$ stands for the boundary of $X$).It is proved that this inequality holds if $X$ is (i) an $n$-dimensional parallelotope, (ii) an $n$-dimensional ball, (iii) a convex polytope having an inscribed sphere (tangent to all its facets) with center in the center of mass of $\partial X$.

\section{Introduction} Dragomir and Pearce proved \cite[Theorem 215]{DraPea00} that if $B_n$ is an $n$-dimensional ball then \Eq{*}{ \fint_{B_3} f(x) dx \le \fint_{\partial B_3} f(x) dx } for every convex function $f$; here and below $\fint$ stands for the average integral (more precisely $\fint_X f(x)\:dx:=\tfrac1{|X|}\int_X f(x)\:dx$). During \emph{Conference on Inequalities and Applications 2016} P\'ales stated the problem whether for every convex and closed set $X$ and every convex function $f \colon X \to \R$, the inequality \Eq{E:conv_def}{ \fint_X f(x) dx \le \fint_{\partial X} f(x) dx } is valid. It is however easy to verify that for the triangle $T$ with vertices $(0,-1)$, $(0,1)$, $(1,0)$ and the function $f \colon T \ni (x,y) \mapsto x$, the inequality \eq{E:conv_def} voids (as the inequality $\tfrac13 \le 1-\tfrac{\sqrt{2}}2$ is not valid). Furthermore, this property is invariant under transition, scaling, changing orientation and reflection, whence it is a property of a shape. This motivates us to introduce the following definition. Convex and closed shape $X$ is called \emph{Jensen-type} if for every convex function $f \colon X \to \R$, the inequality \eq{E:conv_def} is satisfied. Using this definition Dragomir--Pearce result can be expressed briefly as {\it $3$-dimensional ball is of Jensen-type} or {\it $B_3$ is of Jensen-type}. The second example can be expressed by {\it 45--45--90 triangle is not of Jensen-type}. Motivated by these preliminaries we are going to prove this property for \mbox{regular polygons}, \mbox{parallelotopes} (in all dimensions), \mbox{balls} (in all dimensions), and \mbox{Platonic solids}. \section{Results} We begin with some necessary condition for $X$ to be Jensen-type. It repeats the argumentation which was already presented in a case of triangle $T$. \begin{lem}\label{lem:necc} If $X$ is of Jensen-type then centers of mass of $X$ and $\partial X$ coincide. \end{lem} \begin{proof} Let $\pi_i \colon \R^n \to \R$ be a projection on $i$-th coordinate ($i \in \{1,\dots,n\}$). Both $\pi_i$ and $-\pi_i$ are convex so, as $X$ is of Jensen-type, we get \Eq{*}{ \fint_X \pi_i(x) dx \le \fint_{\partial X} \pi_i(x) dx \quad\text{ and }\quad \fint_X -\pi_i(x) dx \le \fint_{\partial X} -\pi_i(x) dx. } Thus $\fint_X \pi_i(x) dx = \fint_{\partial X} \pi_i(x) dx$ for $i \in \{1,\dots,n\}$. But centers of mass of $X$ and $\partial X$ equal to $\left(\fint_X \pi_i(x) dx \right)_{i=1}^n$ and $\left(\fint_{\partial X} \pi_i(x) dx \right)_{i=1}^n$, respectively. The above equality states that these points coincide. \end{proof} \begin{xrem} We have presented some necessary condition for a shape to be of Jensen-type. Our conjecture is that every convex shape which satisfies this condition is of Jensen-type. \end{xrem} In the subsequent result we are going to prove that all parallelotopes and $n$-dimensional balls are of Jensen-type. \begin{prop} All parallelotopes are of Jensen-type. \end{prop} \begin{proof} Fix parallelotope $W$ of dimension $n$. Let $\{S_i\}_{i=1}^{2^n}$ be its all facets. Denote by $S_i^*$ the facet opposite to $S_i$. In fact facet $S_i^*$ is a facet $S_i$ shifted by some vector $v_i \in \R^n$. Finally, for $y \in S_i$, let $y^\ast:=y+v_i \in S_i^*$. Now fix a convex function $f \colon W \to \R$. By Hermite-Hadamard inequality we have \Eq{*}{ \int_y^{y^*} f(x) dx &\le \dist(S_i,S_i^*) \cdot \frac{f(y)+f(y^*)}2 \qquad (y \in S_i), } which in view of the equality $|S_i|\dist(S_i,S_i^*)=|W|$ is equivalent to \Eq{*}{ \frac{2|S_i|}{|W|} \int_y^{y^*} f(x) dx &\le f(y)+f(y^*) \qquad (y \in S_i), } We integrate both side over $S_i$ to obtain \Eq{*}{ 2|S_i| \cdot \fint_W f(x)\: dx \le \int_{S_i \cup S_i^*} f(x) \:dx. } Finally, let us sum up the above inequality for $i \in \{1,2,\dots,2^n\}$. Then we obtain \Eq{*}{ 2\abs{\partial W} \cdot \fint_W f(x) dx &\le 2 \int_{\partial W} f(x) dx, } which simplifies to $\fint_W f(x) dx \le \fint_{\partial W} f(x) dx$. \end{proof} In the next proposition we will generalize the Dragomir-Pearce result. \begin{prop} The $n$-dimensional ball is of Jensen-type for every $n \ge 2$. \end{prop} \begin{proof} Fix a convex function $f \colon B_n \to \R$. We have \Eq{*}{ \fint_{B_n} f(x) dx &= \frac{1}{|B_n|} \int_{B_n} f(x) dx = \frac{1}{|B_n|} \int_0^1 r^{n-1} \int_{S_{n-1}} f(rx) dx dr \\ &= \frac{1}{|B_n|} \int_0^1 \frac{r^{n-1}}{2} \int_{S_{n-1}} f(rx)+f(-rx) dx dr } Applying Wright-convexity of $f$ we get \Eq{*}{ \fint_{B_n} f(x) dx &\le \frac{1}{|B_n|} \int_0^1 \frac{r^{n-1}}{2} \int_{S_{n-1}} f(x)+f(-x) dx dr \\ &\le \frac{1}{|B_n|} \int_0^1 r^{n-1} dr \int_{S_{n-1}} f(x) dx = \frac{1}{n |B_n|} \int_{S_{n-1}} f(x) dx. } By the identity $n |B_n|=\abs{S_{n-1}}$, we obtain desired inequality. \end{proof} \subsection{Convex polytopes having an inscribed sphere} We will now struggle with convex polytopes. To avoid misunderstandings the \emph{inscribed sphere} is the sphere which is tangent to all facets. \begin{lem} \label{lem:ostroslup} Let $n \in \N$, $\Delta \subset \R^n$ be a convex $(n-1)$-dimensional shape, $s \in \R^n \setminus \Delta$ and $G=\conv \{\Delta,s\}$. Then for every convex function $f \colon G \to \R$, \Eq{*}{ \fint_G f(x) dx \le \frac{n}{n+1} \fint_{\Delta} f(x) dx + \frac1{n+1} f(s). } \end{lem} \begin{proof} For each $\theta \in (0,1]$ let $T_\theta$ a homothetic transformation of $\Delta$ with center $s$ and scale $\theta$. Denote its image by $\Delta_\theta$. Moreover denote $H:=\dist(p,\Delta)$ and $\pi \colon \Delta \to \Delta_1$ be a projection such that $\pi|_{\Delta_\theta}=T_\theta^{-1}$. We know that \Eq{*}{ x=\theta \cdot \pi(x)+(1-\theta) \cdot s \text{ for all }\theta \in (0,1] \text{ and }x \in \Delta_\theta. } Whence \Eq{*}{ \int_G f(x) dx &= H \cdot \int_0^1 \int_{\Delta_\theta} f(x) dx d\theta \\ &= H \cdot \int_0^1 \int_{\Delta_\theta} f(\theta \cdot \pi(x) +(1-\theta) s) dx d\theta. \\ &= H \cdot \int_0^1 \int_{\Delta_1} \theta^{n-1} f(\theta \cdot x +(1-\theta) s) dx d\theta } Thus, by Jensen's and Fubini's inequalities, \Eq{*}{ \int_G f(x) dx &\le H \cdot \int_{\Delta_1} \int_0^1 \theta^n \: d\theta \cdot f(x) +\int_0^1 \theta^{n-1} (1-\theta) \: d\theta \cdot f(s) \: dx. } By $\int_0^1 \theta^n d\theta =\tfrac{1}{n+1}$ and $\int_0^1 \theta^{n-1}(1-\theta)d\theta=\tfrac{1}{n(n+1)}$ we obtain \Eq{*}{ \int_G f(x) dx &\le H \cdot \left( \tfrac{1}{n+1} \int_{\Delta_1} f(x) dx + \tfrac{1}{n(n+1)} \abs{\Delta_1} \cdot f(s) \right) \\ &= \frac{H \cdot \abs{\Delta_1}}n \cdot \left( \tfrac{n}{n+1} \fint_{\Delta_1} f(x) dx + \tfrac{1}{n+1} \cdot f(s) \right). } To finish the proof we can use the classical equality $|G|=\tfrac1n \cdot H \cdot |\Delta_1|$. \end{proof} \begin{thm} Let $W$ be an $n$-dimensional convex polytope having an inscribed sphere with center $s$. Then \Eq{E:thm1}{ \fint_W f(x) dx \le \frac{n}{n+1} \fint_{\partial W} f(x) dx + \frac1{n+1} f(s) } for every convex function $f \colon W \to \R$. \end{thm} \begin{proof} Let $r$ be a radius of the inscribed sphere. Denote all facets of $W$ by $\{A_1,\dots,A_k\}$, moreover let $G_i=\conv\{A_i,s\}$. We have $\abs{G_i}=\tfrac rn \cdot \abs{A_i}$, in particular $\abs{W}=\tfrac rn \abs{\partial W}$. By Lemma~\ref{lem:ostroslup}, for all $i \in \{1,\dots,k\}$ we have \Eq{*}{ \int_{G_i} f(x) dx &\le \frac{n}{n+1} \cdot \abs{G_i} \cdot \fint_{A_i} f(x) dx + \frac1{n+1} \cdot \abs{G_i} \cdot f(s) \\ &= \frac{n}{n+1} \cdot \frac rn \cdot \abs{A_i} \cdot \fint_{A_i} f(x) dx + \frac1{n+1} \cdot \abs{G_i} \cdot f(s) \\ &= \frac{r}{n+1} \cdot \int_{A_i} f(x) dx + \frac1{n+1} \cdot \abs{G_i} \cdot f(s) } Summing this inequality (side-by-side for $i \in \{1,\cdots,n\}$) we obtain \Eq{*}{ \int_{W} f(x) dx &\le \frac{r}{n+1} \cdot \int_{\partial W} f(x) dx + \frac1{n+1} \cdot \abs{W} \cdot f(s) \\ &= \frac{r\cdot \abs{\partial W}}{n+1} \cdot \fint_{\partial W} f(x) dx + \frac1{n+1} \cdot \abs{W} \cdot f(s) } To finish the proof note that $\frac{r\cdot \abs{\partial W}}{n+1} = \frac{n}{n+1} \cdot \frac{r\cdot \abs{\partial W}}n = \frac{n}{n+1} \cdot \abs{W}$. \end{proof} We can now present some simple corollary. \begin{cor} Let $W$ be a convex $n$-dimensional polytope having an inscribed sphere with center $s$ and $m$ be the center of mass of $\partial W$. Then \Eq{*}{ \fint_W f(x) dx \le \fint_{\partial W} f(x) dx + \frac1{n+1} (f(s)-f(m)) } for every convex function $f \colon W \to \R$. \end{cor} Indeed, by the Jensen's inequality we have $f(m) \le \fint_{\partial W} f(x)\:dx$, thus $0\le \tfrac{1}{n+1} (\fint_{\partial W} f(x)\:dx - f(m) )$. We can now sum this inequality with \eq{E:thm1} side-by-side to obtain desired inequality. As a trivial particular case we obtain some sufficient condition for $W$ to be of Jensen-type. \begin{thm} Let $W$ be a convex polytope having an inscribed sphere. If the center of this sphere coincide with the center of mass $\partial W$, then $W$ is of Jensen-type. \end{thm} Obviously this result implies that all Platonic solids are of Jensen-type. \subsection*{Acknowledgement} I am grateful to Karol Gryszka and Alfred Witkowski for their valuable remarks. \def$'$} \def\R{\mathbb R} \def\Z{\mathbb Z} \def\Q{\mathbb Q{$'$} \def\R{\mathbb R} \def\Z{\mathbb Z} \def\Q{\mathbb Q} \def\mathbb C{\mathbb C}

https://arxiv.org/abs/1509.07716

The width of quadrangulations of the projective plane

We show that every $4$-chromatic graph on $n$ vertices, with no two vertex-disjoint odd cycles, has an odd cycle of length at most $\tfrac12\,(1+\sqrt{8n-7})$. Let $G$ be a non-bipartite quadrangulation of the projective plane on $n$ vertices. Our result immediately implies that $G$ has edge-width at most $\tfrac12\,(1+\sqrt{8n-7})$, which is sharp for infinitely many values of $n$. We also show that $G$ has face-width (equivalently, contains an odd cycle transversal of cardinality) at most $\tfrac14(1+\sqrt{16 n-15})$, which is a constant away from the optimal; we prove a lower bound of $\sqrt{n}$. Finally, we show that $G$ has an odd cycle transversal of size at most $\sqrt{2\Delta n}$ inducing a single edge, where $\Delta$ is the maximum degree. This last result partially answers a question of Nakamoto and Ozeki.

\section{Introduction} \label{sec:introduction} Erd\H os~\cite{Erd74} asked whether there is a constant $c$ such that every $n$-vertex $4$-chromatic graph has an odd cycle of length at most $c\sqrt n$. Kierstead, Szemer\'edi and Trotter~\cite{KST84} proved the conjecture with $c=8$, and the constant was gradually brought down to $c=2$~\cite{Jia01,Nil99}. A natural question, asked by Ngoc and Tuza~\cite{NgoTuz95}, is to determine the infimum of $c$ such that every $4$-chromatic graph on $n$ vertices has an odd cycle of length at most $c\sqrt n$. A construction due to Gallai~\cite{Gal63} shows that $c>1$, and this was subsequently improved to $c>\sqrt 2$ by Ngoc and Tuza~\cite{NgoTuz95}, and independently by Youngs~\cite{You96}. The graphs they used---the so-called \emph{generalized Mycielski graphs}---are a subclass of a rich family of graphs known as \emph{non-bipartite projective quadrangulations}. These are graphs that embed in the projective plane so that all faces are bounded by four edges, but are not bipartite. This family of graphs plays an important role in the study of the chromatic number of graphs on surfaces: it was shown by Youngs~\cite{You96} that all such graphs are $4$-chromatic. Gimbel and Thomassen~\cite{GimTho97} later proved that triangle-free projective-planar graphs are $3$-colorable if and only if they do not contain a non-bipartite projective quadrangulation, and used this to show that the $3$-colorability of triangle-free projective-planar graphs can be decided in polynomial time. Thomassen~\cite{Tho04} also used projective quadrangulations to give negative answers to two questions of Bollob\'as~\cite{Bol78} about $4$-chromatic graphs. Let $G$ be a non-bipartite projective quadrangulation. A key property of such a graph is that any cycle in $G$ is contractible on the surface if and only if it has even length. In particular, the length of a shortest odd cycle in $G$ is precisely the \emph{edge-width} of $G$, the length of a shortest non-contractible cycle. Since any two non-contractible closed curves on the projective plane intersect, it also follows that $G$ does not contain two vertex-disjoint odd cycles. The interest in the study of odd cycles in $4$-colorable graphs also comes from the following question of Erd\H os~\cite{Erd68}: does every $5$-chromatic $K_5$-free graph contain a pair of vertex-disjoint odd cycles? Erd\H os's question may be rephrased as follows: is every $K_5$-free graph without two vertex-disjoint odd cycles $4$-colorable? This was answered in the affirmative by Brown and Jung~\cite{BroJun69}. The non-bipartite projective quadrangulations provide an infinite family of graphs showing that $4$ cannot be replaced by $3$. Note that Erd\H os's question was generalized by Lov\'asz and became known as the Erd\H os--Lov\'asz Tihany Conjecture. So far only a few cases of the conjecture have been proved. \smallskip Our first theorem settles the problem of Ngoc and Tuza for the case of $4$-chromatic graphs with no two vertex-disjoint odd cycles (and in particular, for non-bipartite projective quadrangulations). \begin{theorem}\label{thm:oddcycle2} Let $G$ be a $4$-chromatic graph on $n$ vertices without two vertex-disjoint odd cycles. Then $G$ contains an odd cycle of length at most $\tfrac12(1+\sqrt{8n-7})$. \end{theorem} Note that in the generalized Mycielski graphs found by Ngoc and Tuza~\cite{NgoTuz95}, and independently by Youngs~\cite{You96}, the shortest odd cycles have precisely this number of vertices, so Theorem~\ref{thm:oddcycle2} is sharp for infinitely many values of $n$. \smallskip An \emph{odd cycle transversal} in a graph $G$ is a set of vertices $S$ such that $G-S$ is bipartite. Since any two odd cycles intersect in a non-bipartite projective quadrangulation $G$, it follows that any odd cycle in $G$ is also an odd cycle transversal of $G$. The following slightly more general result holds. If $\gamma$ is a non-contractible closed curve whose intersection with $G$ is a subset $S \subseteq V(G)$ (the minimum size of such a set $S$ is called the \emph{face-width} of $G$), then $G-S$ is bipartite. It follows that the minimum size of an odd cycle transversal of $G$ cannot exceed the face-width of $G$, and it can be proved that the two parameters are indeed equal. \Cref{thm:oddcycle2} immediately implies that a non-bipartite projective quadrangulation on $n$ vertices has an odd cycle transversal with at most $\tfrac12(1+\sqrt{8n-7}) \approx \sqrt{2n}$ vertices. Our next theorem improves the bound to roughly $\sqrt{n}$. \begin{theorem} \label{thm:OCT-upper} Let $G$ be a non-bipartite projective quadrangulation on $n$ vertices. Then $G$ has an odd cycle transversal of size at most $\tfrac14+\sqrt{n-\tfrac{15}{16}}$. \end{theorem} The next result shows that it is almost optimal. \begin{theorem} \label{thm:OCT-lower} There are infinitely many values of $n$ for which there are non-bipartite projective quadrangulations on $n$ vertices containing no odd cycle transversal of size less than $\sqrt n$. \end{theorem} \smallskip Nakamoto and Ozeki (private communication) have asked whether every $n$-vertex non-bipartite projective quadrangulation can be $4$-colored so that one color class has size $1$ and another has size $o(n)$. While we were unable to answer their question in general, our final theorem gives a positive answer when the maximum degree is $o(n)$. \begin{theorem}\label{thm:sqrtD} Let $G$ be a non-bipartite projective quadrangulation on $n$ vertices, with maximum degree $\Delta$. There exists an odd cycle transversal of size at most $\sqrt{2\Delta n}$ inducing a single edge. \end{theorem} The rest of the paper is organized as follows. In \Cref{sec:preliminaries} we introduce the necessary terminology and prove a number of lemmas that will be used later. In \Cref{sec:oddcycle} we prove \Cref{thm:oddcycle2} using a theorem of Lins on graphs embedded in the projective plane (an equivalent form of the Okamura-Seymour theorem in Combinatorial Optimisation), and the Two Disjoint Odd Cycles Theorem of Lov\'asz. In \Cref{sec:OCT} we prove \Cref{thm:OCT-upper} using a theorem of Randby~\cite{Ran97}, and then show that a similar result holds for any $4$-vertex-critical graph in which any two odd cycles intersect. As a consequence of \Cref{thm:OCT-upper}, we also deduce an (almost tight) lower bound on the independence number of non-bipartite projective quadrangulations. In Section~\ref{sec:jap}, we prove Theorem~\ref{thm:sqrtD} and finally, we conclude with some open problems in Section~\ref{sec:conclusion}. \section{preliminaries} \label{sec:preliminaries} Our graph theoretic terminology is standard, and follows Bondy and Murty~\cite{BonMur08}. For the notions from algebraic topology, the reader is referred to~\cite{Mun84}. We denote the (real) projective plane by ${\mathbb P}^2$. We will use the following properties of the projective plane, which can be proved using basic algebraic topology. Namely, every simple closed curve $\gamma:[0,1] \to {\mathbb P}^2$ is either nullhomotopic or non-contractible, and two non-contractible essential simple closed curves intersect transversally an odd number of times. Given a non-bipartite projective quadrangulation $G$, it is not hard to show (see e.g. ~\cite[Lemma 3.1]{KaiSte15}) that all contractible closed walks in $G$ have even length, and all non-contractible closed walks in $G$ have odd length. Given a quadrangulation $G=(V,E)$ of the projective plane ${\mathbb P}^2$, a \emph{support set} $S$ is a circularly ordered subset of $V$ (with repetitions allowed), such that any two consecutive vertices of $S$ are on a common face of $G$. Since $G$ is a quadrangulation, it follows that two distinct consecutive vertices of $S$ are either adjacent or have a common neighbor (in this case, we say that they are \emph{opposite}). The \emph{size} of $S$ is the number of pairs of consecutive vertices of $S$, the \emph{order} of $S$ is the number of pairs of consecutive vertices of $S$ that are adjacent in $G$, and the \emph{parity} of $S$ is the parity of the order of $S$. Note that to any support set $S$ of $G$ we can associate a closed curve of ${\mathbb P}^2$ meeting $G$ in $S$, and encountering the vertices of $S$ in their circular order. We denote such a curve by $\rho(S)$. Conversely, to any curve $\rho$ meeting $G$ in a subset $S\subseteq V$ we can associate a support set $S(\rho)$ whose circular order coincides with the order in which $\rho$ visits the vertices of $S$. Given a support set $S$ of $G$, and two consecutive vertices $v_i,v_{i+1}$ of $S$ that are opposite, a \emph{shift} in $S$ at $(v_i,v_{i+1})$ is the support set obtained from $S$ by adding a common neighbor $u_i$ of $v_i$ and $v_{i+1}$ between $v_i$ and $v_{i+1}$ in the support set ($u_i$ may be chosen arbitrarily among the common neighbors of $v_i$ and $v_{i+1}$ in $G$). Observe that replacing a pair of opposite vertices by two pairs of adjacent vertices does not change the parity of the support set. Moreover, if $S'$ is obtained from $S$ by a shift, then $\rho(S)$ and $\rho(S')$ are homologous. For convenience, we write it as a lemma. \begin{lemma}\label{lem:shift} Let $G$ be a non-bipartite quadrangulation of ${\mathbb P}^2$ and $S$ a support set of $G$. If a support set $S'$ is obtained from $S$ by a sequence of shifts, then $S'$ and $S'$ have the same parity and $\rho(S)$ and $\rho(S')$ are homologous. \end{lemma} Recall that a closed walk in a non-bipartite quadrangulation of ${\mathbb P}^2$ is odd if and only if it is non-contractible. The next lemma shows a similar result for support sets. \begin{lemma}\label{lem:parity} If $G$ is a non-bipartite quadrangulation of ${\mathbb P}^2$ and $S$ a support set of $G$, then $S$ is odd if and only if $\rho(S)$ is non-contractible. In particular, if $S$ is odd, then $G-S$ is bipartite. \end{lemma} \begin{proof} For any two consecutive vertices $v_i$ and $v_{i+1}$ of $S$ that are opposite, we make a shift at $(v_i,v_{i+1})$. Let $S'$ be the support set thus obtained. By Lemma~\ref{lem:shift}, $S$ and $S'$ have the same parity and $\rho(S)$ and $\rho(S')$ are homologous. By the definition of $S'$, any two consecutive vertices are adjacent. It follows that $S'$ is a closed walk in $G$. Since a closed walk in a non-bipartite quadrangulation of ${\mathbb P}^2$ is odd if and only if it is non-contractible, $S$ is odd if and only if $\rho(S)$ is non-contractible. Assume now that $S$ is odd, and so $\rho(S)$ is non-contractible. Then the removal of $S$ yields a graph embedded in the plane, with all inner faces bounded by an even number of edges. Therefore, $G-S$ is bipartite. \end{proof} \begin{lemma}\label{lem:shortest} Let $G$ be a non-bipartite quadrangulation of ${\mathbb P}^2$ and $S$ an odd support set of $G$. Then there is an odd support set $S' \subseteq S$, with order at most the order of $S$, such that the vertices of $S'$ are pairwise distinct, and two vertices of $S'$ are adjacent or opposite if and only if they are consecutive. \end{lemma} \begin{proof} Let $S'\subseteq S$ be an odd support set of order at most the order of $S$, and (with respect to these properties) with minimum size. Assume first that some vertex appears at least twice in $S'$. Then we can divide $S'$ into two support sets of different parities. In particular there is an odd support set $S''\subseteq S'$, of order at most the order of $S'$, and of size less than the size of $S'$. This contradicts the minimality of $S'$. Assume now that two non-consecutive vertices $u$ and $v$ of $S'$ are opposite or adjacent. Again, we can divide $S'$ into two support sets (where $u$ and $v$ are now consecutive), of order at most the order of $S'$ plus one. Since $S'$ is odd, the two support sets have different parities, so one of them is odd (and therefore has order at most the order of $S'$), which contradicts the minimality of $S'$. \end{proof} The same proof also gives the following similar result, which will be needed later. \begin{lemma}\label{lem:shortest2} Let $G$ be a non-bipartite quadrangulation of ${\mathbb P}^2$ and $S$ an odd support set of $G$ corresponding to a non-contractible closed walk in $G$. Then $G$ contains an odd cycle with vertex set $S' \subseteq S$. \end{lemma} \section{Short odd cycles} \label{sec:oddcycle} A key result is the following corollary of a theorem of Lins~\cite{Lin81} (see also~\cite[Corollary 2.4]{ArcBon97}). \begin{lemma}\label{lem:lins} Let $G=(V,E)$ be a projective planar graph with a shortest non-contractible cycle of length $\ell$ and a shortest non-contractible cycle of length $\ell^*$ in its dual graph $G^*$. Then $|E|\ge \ell \cdot \ell^*$. \end{lemma} We now show that the following result follows from Lemma~\ref{lem:lins} as a fairly simple consequence. \begin{theorem}\label{thm:oddcycle} Let $G$ be a non-bipartite projective quadrangulation on $n$ vertices. Then $G$ contains an odd cycle of length at most $\tfrac12(1+\sqrt{8n-7})$. \end{theorem} \begin{proof} A standard application of Euler's formula shows that $G$ has $m=2n-2$ edges. Let $\ell$ be the length of a shortest odd cycle in $G$. By Lemma~\ref{lem:lins}, we know that the dual graph $G^*$ of $G$ contains a non-contractible cycle of length $\ell^* \le (2n-2)/\ell$. Let $C^*$ be such a cycle, and let $(f_1,f_2,\ldots,f_k)$ be the faces of $G$ corresponding to the vertices of $C^*$. Note that any two consecutive faces $f_i,f_{i+1}$ in $C^*$ share an edge, which we call $e_i$. For any $1\le i \le k+1$ (where the indices $1$ and $k+1$ coincide), we will choose a vertex $v_i$ in each edge $e_i$, in a specific way. We start by choosing $v_1$ in $e_1$ arbitrarily, and for any $i>1$ we distinguish two cases. If $v_{i-1} \in e_i$, then we set $v_i=v_{i-1}$, and otherwise we choose for $v_i$ a vertex adjacent to $v_{i-1}$ in $f_i$ (note that such a vertex always exists). Let $S$ be the set of vertices thus chosen (each maximal sequence of consecutive vertices $v_i,v_{i+1},\ldots,v_{j}$ such that $v_i=v_{i+1}=\cdots =v_{j}$ is reduced to a single vertex $v_i$). Any two consecutive vertices are adjacent, so we obtain a closed walk in $G$ of length at most $\ell^*+1$ that is homotopic to $C^*$, and therefore non-contractible. It follows that $S$ is an odd support set where any two consecutive vertices are adjacent. By Lemma~\ref{lem:shortest2}, $G$ contains an odd cycle of length at most $\ell^*+1\le 1+(2n-2)/\ell$. It follows that $\ell^2-\ell \le 2n-2$, so $\ell\le \tfrac12(1+\sqrt{8n-7})$, as desired. \end{proof} A \emph{$k$-separation} in a graph $G$ is a pair $(G_1,G_2)$ of subgraphs of $G$ such that $V(G) = V(G_1) \cup V(G_2)$, $|V(G_1)\cap V(G_2)|=k$, $E(G_1)\cap E(G_2)=\emptyset$, and $E(G_i)\cup V(G_i - G_{3-i})\neq \emptyset$ for $i=1,2$. A graph $G$ is said to be \emph{internally $4$-connected} if $G$ is 3-connected and for every 3-separation $(G_1, G_2)$ in $G$, $|G_1|\le 4$ or $|G_2|\le 4$. The following characterisation of graphs without two vertex-disjoint odd cycles will play a key role in our proofs. It was first proved by Lov\'asz using Seymours's characterisation of regular matroids (see~\cite{Sey95}). A simpler proof was recently given by Kawarabayashi and Ozeki~\cite{KawOze13}. \begin{theorem} \label{thm:disjoint-cycle} Let $G$ be an internally $4$-connected graph. Then $G$ has no two vertex-disjoint odd cycles if and only if $G$ satisfies one of the following conditions: \begin{enumerate} \item $G-v$ is bipartite, for some $v \in V(G)$; \item $G-\{e_1,e_2,e_3\}$ is bipartite for some edges $e_1, e_2, e_3 \in E(G)$ such that $e_1, e_2, e_3$ form a triangle; \item $|V(G)| \leq 5$; \item $G$ can be embedded into the projective plane so that every face boundary has even length. \end{enumerate} \end{theorem} In order to extend Theorem~\ref{thm:oddcycle} to $4$-chromatic graphs without two vertex-disjoint odd cycles, we will need the following technical lemma about precoloring extension in bipartite graphs. \begin{lemma}\label{lem:ext} Let $G$ be a bipartite graph and $X$ be a subset of $V(G)$ of size at most $3$. Then any (proper) precoloring of $X$ extends to a $3$-coloring of $G$, unless \begin{enumerate}[{\normalfont (i)}] \item $X =\{x,y,z\}$ for distinct $x,y,z$, \item $x$, $y$, and $z$ are on the same side of the bipartition of $G$, \item in the precoloring, $x$, $y$ and $z$ have pairwise different colors, and \item any pair of vertices in $\{x,y,z\}$ have a common neighbor in $G$. \end{enumerate} \end{lemma} \begin{proof} Let $A,B$ be the bipartition of $G$. If in the precoloring of $X$, $A$ contains at most two different colors, then we color each vertex of $A-X$ with one of these two colors, and each vertex of $B-X$ with the third color. This yields a (proper) 3-coloring of $G$. Otherwise, by symmetry, $X=\{x,y,z\}\in A$ and $x,y,z$ have distinct colors; this proves conditions (i)--(iii) of the lemma. Assume that $x$ and $y$ have no common neighbor in $G$ (and thus in $B$), and $x,y,z$ are colored $1,2,3$ respectively. Then we color each vertex of $A-X$ with color 3, each neighbor of $x$ with color 2, and the remaining vertices of $B$ with color 1. Since $x$ and $y$ have no common neighbor, this is a proper 3-coloring of $G$ extending the precoloring of $X$. \end{proof} A $k$-chromatic graph is said to be \emph{$k$-vertex-critical} if for any vertex $v$, $G-v$ is $(k-1)$-colorable. A graph is \emph{projective} if it can be embedded in ${\mathbb P}^2$. We will need the following direct consequence of a result of Gimbel and Thomassen~\cite[Theorem 5.4]{GimTho97}: \begin{lemma}\label{lem:GT} Let $G$ be a simple triangle-free projective graph. If $G$ is $4$-vertex-critical, then $G$ is a (non-bipartite) projective quadrangulation. \end{lemma} \begin{proof} Since $G$ is a $4$-chromatic triangle-free projective graph, $G$ contains a non-bipartite projective quadrangulation $H$ as a subgraph by a result of Gimbel and Thomassen~\cite[Theorem 5.4]{GimTho97}. Since $H$ is itself $4$-chromatic and $G$ is vertex-critical, $H$ must be be a spanning subgraph of $G$. If $H$ is a proper subgraph of $G$, then there is an edge $e \in E(G) \setminus E(H)$. Both end vertices of $e$ must lie on the boundary of the same face in some embedding of $H$ in ${\mathbb P}^2$, so adding $e$ to $H$ creates a triangle or a pair of parallel edges, contradicting the hypothesis of the lemma. Therefore $G=H$, so $G$ is a non-bipartite projective quadrangulation. \end{proof} We now extend Theorem~\ref{thm:oddcycle} to $4$-chromatic graphs without two vertex-disjoint odd cycles. \begin{proof}[Proof of Theorem~\ref{thm:oddcycle2}] We prove the result by induction on $n$. We can assume that $G$ is $4$-vertex-critical. In particular, $G$ is connected and does not contain a clique cutset (a clique whose removal disconnects the graph). We may assume that $G$ has at least six vertices (otherwise $G$ has four or five vertices and contains a triangle and the result clearly holds). As a consequence, we may also assume that $G$ is triangle-free, since otherwise $G$ has an odd cycle of length $3\le \tfrac12(1+\sqrt{8n-7})$. \smallskip Assume first that $G$ is internally $4$-connected. In this case we can apply Theorem~\ref{thm:disjoint-cycle}. As $G$ is $4$-chromatic, triangle-free, and has at least six vertices, none of cases (i)--(iii) applies. It follows that $G$ can be embedded into the projective plane. Since $G$ is $4$-vertex-critical and triangle-free, by Lemma~\ref{lem:GT} it is a non-bipartite projective quadrangulation and the result follows directly from Theorem~\ref{thm:oddcycle}. \smallskip Assume now that $G$ is not internally $4$-connected. Since $G$ has no clique-cutset, it is $2$-connected. Hence there exist graphs $G_1=(V_1,E_1)$ and $G_2=(V_2,E_2)$, and sets $X_i \in V_i$ of two or three vertices ($i=1,2$) with $|X_1|=|X_2|$, such that $G_1[X_1]$ and $G_2[X_2]$ are equal (as labelled graphs), and $G$ can be obtained from $G_1,G_2$ by idenfying $X_1$ in $G_1$ and $X_2$ in $G_2$ (call $X$ the corresponding set of vertices in $G$, inducing the same graph as $G_1[X_1]$ and $G_2[X_2]$). Moreover, if $|X_1|=|X_2|=3$, then for $i=1,2$, $V_i - X_i$ contains at least two vertices (this follows from the definition of internal $4$-connectivity). Note that since $G$ is triangle-free, $X$ is bipartite. In what follows, by a slight abuse of notation, we give the same name to a vertex of $X_1$, the vertex of $X_2$ it is indentified with, and the resulting vertex of $X$. Assume that $|X|=2$, say $X=\{x,y\}$. Since $G$ has no clique-cutset, $x$ and $y$ are non-adjacent. If $G$ contains an odd cycle disjoint from $X$ (say in $G_2- X_2$), then since $G$ has no two vertex-disjoint odd cycles, $G_1$ is bipartite. Since $G$ is $4$-vertex-critical, $G_2$ is 3-colorable and by Lemma~\ref{lem:ext} any $3$-coloring of $G_2$ extends to $G_1$, a contradiction. It follows that every odd cycle of $G$ intersects $X$, so $G-X$ is bipartite. As $X$ is a stable set, $G$ is $3$-colorable, which is a contradiction. We can now assume that $|X|=3$, say $X=\{x,y,z\}$. Assume first that $X$ is a stable set. If all odd cycles of $G$ intersect $X$, then $G-X$ is bipartite and since $X$ is stable, $G$ is $3$-colorable, a contradiction. Otherwise $G$ contains an odd cycle disjoint from $X$ (say in $G_2- X_2$). Then $G_1$ is bipartite, and by Lemma~\ref{lem:ext}, $x,y,z$ are on the same side of the bipartition of $X$, and in each $3$-coloring of $G_2$ they have three distinct colors. Moreover, each pair of vertices among $x,y,z$ has a common neighbor in $G_1$. Let $H$ be the graph obtained from $G_2$ by adding a vertex $v$ adjacent to $x,y,z$. Since $G_1-X_1$ contains at least two vertices, $H$ has less vertices than $G$. Note that $H$ is $4$-chromatic (since for any $3$-coloring of $H-v$, the neighbors of $v$ have three distinct colors), and has no two vertex-disjoint odd cycles: each odd cycle $C$ of $H$ is either disjoint from $v$, and is therefore an odd cycle of $G$, or intersects $X$ in two vertices, say $x$ and $y$, and corresponds to an odd cycle of $G$ coinciding with $C$ in $G_2$ and whose intersection with $G_1-X$ is a single common neighbor of $x$ and $y$ in $G_1$ (which is known to exist). By the induction hypothesis, $H$ (and therefore $G$) has an odd cycle of length at most $\tfrac12(1+\sqrt{8n-7})$. Since $G$ is triangle-free, we can assume that $G[X]$ has two non-adjacent vertices, say $y,z$, while $x$ is adjacent to at least one of them, say $y$. By Lemma~\ref{lem:ext}, none of $G_1,G_2$ is bipartite and in particular, every odd cycle intersects $X$. Note that $G-\{y,z\}$ is not bipartite (since otherwise $G$ would be $3$-colorable), so we can assume that $G_2$ has an odd cycle $C_2$ containing $x$ and avoiding $y,z$. Therefore every odd cycle of $G_1$ intersects $x$, so $G_1-x$ is bipartite, say with bipartition $A,B$. Take a $3$-coloring $c$ of $G_2$, and assume without loss of generality that $x$ has color $1$. If $y,z$ are both in $A$ and $x,y,z$ do not have pairwise distinct colors, then $c$ easily extends to a $3$-coloring of $G_1$. Similarly, if $y,z$ have distinct colors and are in distinct partite sets, then $c$ easily extends to a $3$-coloring of $G_1$. So we can assume that either \begin{enumerate} \item $y,z \in A$, $y$ has color $2$, and $z$ has color $3$, or \item $y\in A$, $z\in B$, and $y,z$ are both colored $2$. \end{enumerate} Let $B_x$ be the set of neighbors of $x$ in $B$. We color $A-y$ with color $3$, $B_x$ with color $2$, and $B-(B_x\cup X)$ with color $1$. Since $G$ is triangle-free there are no edges between $y$ and $B_x$, so the resulting $3$-coloring of $G$ is proper, which contradicts the fact that $G$ is $4$-chromatic. This concludes the proof of Theorem~\ref{thm:oddcycle2}. \end{proof} \section{Small odd cycle transversals} \label{sec:OCT} A (multi)graph $G$ embedded in a surface $\Sigma$ is \emph{minimal of face-width $k$} if the face-width of $G$ is $k$, while for any edge $e$ of $G$, the face-width of $G/e$ (the (multi)graph embedded in $\Sigma$ obtained from $G$ be contracting $e$) and the face-width of $G-e$ are less than $k$. We will use the following result of Randby~\cite{Ran97}: \begin{theorem}\label{thm:ran97} For any integer $k$, if a multigraph embedded in ${\mathbb P}^2$ is minimal of face-width $k$, then it contains exactly $2k^2-k$ edges. \end{theorem} \begin{proof}[Proof of Theorem~\ref{thm:OCT-upper}] Let $G$ be a non-bipartite quadrangulation on $n$ vertices of the projective plane ${\mathbb P}^2$, and let $k$ be the face-width of $G$. By Euler's formula, $G$ has $m=2n-2$ edges. If $G$ is not minimal with face-width $k$, we delete or contract edges of $G$ until we obtain a (multi)graph $H$ that is minimal with face-width $k$. Note that $H$ has at most $m=2n-2$ edges by construction, and exactly $2k^2-k$ edges by Theorem~\ref{thm:ran97}. It follows that $2k^2-k\le 2n-2$ and so $k\le \tfrac14+\sqrt{n-\tfrac{15}{16}}$, as desired. \end{proof} We believe that Theorem~\ref{thm:OCT-upper} can be extended to $4$-chromatic graphs with no two vertex-disjoint odd cycles, in the same way Theorem~\ref{thm:oddcycle2} extends Theorem~\ref{thm:oddcycle}. However, we have only been able to prove that $4$-vertex-critical graphs with no two vertex-disjoint odd cycles satisfy the result. \begin{theorem} \label{thm:OCT-upper2} Let $G$ be a $4$-vertex-critical graph on $n$ vertices without two vertex-disjoint odd cycles. Then $G$ has an odd cycle transversal of cardinality at most $\tfrac14+\sqrt{n-\tfrac{15}{16}}$. \end{theorem} \begin{proof} The proof proceeds by induction on $n$. We can assume that $G$ has at least nine vertices (by checking small $4$-vertex-critical graphs, for instance using~\cite{CGSZ15}) and thus no odd cycle transversal on at most three vertices (in particular, $G$ is triangle-free). If $G$ is not internally $4$-connected then there exist graphs $G_1=(V_1,E_1)$ and $G_2=(V_2,E_2)$, and sets $X_i \subseteq V_i$ with at most three vertices ($i=1,2$) with $|X_1|=|X_2|$, such that $G_1[X_1]$ and $G_2[X_2]$ are equal (as labelled graphs), and $G$ can be obtained from $G_1,G_2$ by identifying $X_1$ in $G_1$ and $X_2$ in $G_2$ (let $X$ be the corresponding set of vertices in $G$, inducing the same graph as $G_1[X_1]$ and $G_2[X_2]$). Moreover, $G_1,G_2$ have the property that if $|X_1|=|X_2|=3$, then for $i=1,2$, $V_i - X_i$ contains at least two vertices. If every odd cycle of $G$ intersects $X$, then $X$ is an odd cycle transversal of size at most 3, which is a contradiction. It follows that $G$ contains an odd cycle disjoint from $X$ (say in $G_2-X_2$). Since any two odd cycles intersect, $G_1$ is bipartite. Recall that $G$ is $4$-vertex-critical, so $G_2$ is 3-colorable, and since no 3-coloring of $G_2$ extends to $G_1$ (otherwise $G$ would be 3-colorable), by Lemma~\ref{lem:ext}, we have (i) $X =\{x_1,x_2,x_3\}$ for distinct $x_1,x_2,x_3$, (ii) $x_1$, $x_2$, and $x_3$ are on the same side of the bipartition of $G_1$, (iii) in any 3-coloring of $G_2$, $x_1$, $x_2$ and $x_3$ have pairwise different colors, and (iv) any pair of vertices in $\{x_1,x_2,x_3\}$ have a common neighbor in $G_1$. If there is a vertex $y$ in $G_1$, adjacent to each of $x_1,x_2,x_3$, then $G_2-(V_1-\{y\})$ is $4$-chromatic, which contradicts the fact that $G$ is $4$-vertex-critical (since $G_1-X_1$ contains at least two vertices). It follows that no vertex of $G_1$ is adjacent to each of $x_1,x_2,x_3$. So there is a set $Y=\{y_1,y_2,y_3\}$ of three vertices in $G_1$, such that $y_1$ is adjacent to $x_2$ and $x_3$, $y_2$ is adjacent to $x_1$ and $x_3$, and $y_3$ is adjacent to $x_1$ and $x_2$. Since a $4$-vertex-critical graph has minimum degree at least 3, $G_1$ contains at least seven vertices. If $G_1$ has at least eight vertices, then remove from $G$ all the vertices of $V_1-(X\cup Y)$, and add a vertex $z$ adjacent to $y_1,y_2,y_3$. The resulting graph $H$ is $4$-vertex-critical, has no two vertex-disjoint odd cycles, and is smaller than $G$. By the induction hypothesis, $H$ has an odd cycle transversal $T$ with at most $\tfrac14+\sqrt{n-\tfrac{15}{16}}$ vertices. Note that $z$ does not appear in a minimum odd cycle transversal of $H$, so we can assume that $z\not\in T$. It is easy to check that $T$ is also an odd cycle transversal of $G$. \smallskip \begin{figure}[htbp] \centering \includegraphics[scale=1]{7vg} \caption{A 7-vertex bipartite graph.} \label{fig:7vg} \end{figure} By the above paragraph, we can assume that for any decomposition of $G$ into $G_1$ and $G_2$ as above, on some set $X$ of at most three vertices, $G_1$ induces the graph on seven vertices in Figure~\ref{fig:7vg}. In particular, $X$ induces a stable set of size $3$. Moreover, it is not hard to check that no vertex cutset of $G$ of size at most $3$ intersects $G_1-X$, otherwise $G$ would contain a vertex cutset of size at most $2$. It follows that $G$ can be constructed from some graph $G_0$ and a family $t_1,t_2,\ldots,t_k$ of triples of vertices of $G_0$ by pasting the $X$-part of a copy $G_i$ of the graph of Figure~\ref{fig:7vg} onto each triple $t_i$; see Figure~\ref{fig:expl}, top left. Let $H$ be the graph obtained from $G_0$ by adding, for each $1\le i \le k$, a vertex $z_i$ adjacent to the vertices of $t_i$; see Figure~\ref{fig:expl}, bottom left. Observe that $H$ is $4$-vertex-critical (otherwise $G$ would be $3$-colorable), internally $4$-connected, and any two odd cycles intersect. Note also that $H$ is triangle-free, since otherwise $G$ would also contain a triangle. By Theorem~\ref{thm:disjoint-cycle}, $H$ has an embedding in the projective plane, and by Lemma~\ref{lem:GT}, $H$ is a non-bipartite quadrangulation of the projective plane. We now replace each $z_i$ by $G_i-X$ (see Figure~\ref{fig:expl}, right) and observe that $G$ is itself a quadrangulation of the projective plane. By Theorem~\ref{thm:OCT-upper}, $G$ has an odd cycle transversal with at most $\tfrac14+\sqrt{n-\tfrac{15}{16}}$ vertices, which concludes the proof. \end{proof} \begin{figure}[htbp] \centering \includegraphics[scale=1]{expl} \caption{Graphs $G$, $H$, and their representations as projective quadrangulations.} \label{fig:expl} \end{figure} \medskip We now prove that the bound in Theorem~\ref{thm:OCT-upper} is at most $\tfrac14$ away from the optimum. For a positive integer $k$, let $[k]$ denote the set $\{0,\ldots,k-1\}$. Let $P_k$ be a path with vertex set $[k]$, with vertices in the increasing order along $P_k$. For $k\geq 2$, we define $G_k$ as the graph obtained from the Cartesian product $P_k \Box P_k$ by adding the edges joining $(0,j)$ to $(k-1,k-j-1)$, and those joining $(j,0)$ to $(k-j-1,k-1)$. The graphs $G_k$ embed as quadrangulations in ${\mathbb P}^2$; see \Cref{fig:grids} for embeddings of $G_2$, $G_3$ and $G_4$ in ${\mathbb P}^2$. \begin{figure}[ht] \centering \begin{tikzgraph}[scale=1.2,thin] \def3{1.5} \def1.5{0.5} \draw[dashed] (0,0) circle (3); \foreach\i in {1,...,4} { \path (45+90*\i:3) coordinate (c\i); } \draw (-1.5,1.5)--(-1.5,-1.5) (-1.5,-1.5)--(1.5,-1.5) (1.5,-1.5)--(1.5,1.5) (1.5,1.5)--(-1.5,1.5) (c1)--(-1.5,1.5) (c2)--(-1.5,-1.5) (c3)--(1.5,-1.5) (c4)--(1.5,1.5); \foreach \i in {-1,1} { \foreach \j in {-1,1} { \draw ({\i/2},{\j/2}) node[vertex] {}; } } \end{tikzgraph} \hfil \begin{tikzgraph}[scale=0.75,thin] \def3{2.4} \def1.5{1} \def0{0} \draw[dashed] (0,0) circle (3); \foreach\i in {-1.5,...,1.5} { \path (\i,{sqrt(3^2-abs(\i^2))}) coordinate (n\i) (\i,-{sqrt(3^2-abs(\i^2))}) coordinate (s\i) (-{sqrt(3^2-abs(\i^2))},\i) coordinate (e\i) ({sqrt(3^2-abs(\i^2))},\i) coordinate (w\i); } \foreach\i in {1,...,4} { \path (45+90*\i:3) coordinate (c\i); } \foreach\i in {0} { \draw (n\i)--(s\i) (e\i)--(w\i); } \draw (-1.5,1.5)--(-1.5,-1.5) (-1.5,-1.5)--(1.5,-1.5) (1.5,-1.5)--(1.5,1.5) (1.5,1.5)--(-1.5,1.5) (c1)--(-1.5,1.5) (c2)--(-1.5,-1.5) (c3)--(1.5,-1.5) (c4)--(1.5,1.5); \foreach \i in {-1.5,...,1.5} { \foreach \j in {-1.5,...,1.5} { \draw (\i,\j) node[vertex] {}; } } \end{tikzgraph} \hfil \begin{tikzgraph}[scale=0.6,thin] \def3{3} \def1.5{1.5} \draw[dashed] (0,0) circle (3); \foreach\i in {-3,-1,1,3} { \path ({\i/2},{sqrt(3^2-abs((\i/2)^2))}) coordinate (n\i) ({\i/2},-{sqrt(3^2-abs((\i/2)^2))}) coordinate (s\i) (-{sqrt(3^2-abs((\i/2)^2))},{\i/2}) coordinate (e\i) ({sqrt(3^2-abs((\i/2)^2))},{\i/2}) coordinate (w\i); } \foreach\i in {1,...,4} { \path (45+90*\i:3) coordinate (c\i); } \foreach\i in {-1,1} { \draw (n\i)--(s\i) (e\i)--(w\i); } \draw (-1.5,1.5)--(-1.5,-1.5) (-1.5,-1.5)--(1.5,-1.5) (1.5,-1.5)--(1.5,1.5) (1.5,1.5)--(-1.5,1.5) (c1)--(-1.5,1.5) (c2)--(-1.5,-1.5) (c3)--(1.5,-1.5) (c4)--(1.5,1.5); \foreach \i in {-3,-1,1,3} { \foreach \j in {-3,-1,1,3} { \draw ({\i/2},{\j/2}) node[vertex] {}; } } \end{tikzgraph} \caption{The graphs $G_2$, $G_3$ and $G_4$ embedded in the projective plane ${\mathbb P}^2$.} \label{fig:grids} \end{figure} \begin{proof}[Proof of \Cref{thm:OCT-lower}] The argument is quite similar to the argument given in~\cite{Ree99} for \emph{Escher walls}, a related construction of Lov\'asz and Schrijver. Assume for the sake of contradiction that $G_k$ has an odd cycle transversal $T$ of size at most $k-1$. Then for some $\ell \in [k]$, $T$ is disjoint from the $\ell$-th row $R_\ell$ of $G_k$ (all the vertices $(i,\ell)$, with $i \in [k]$). Consider, for each $i \in [k]$, the set of vertices $L_i=\{(i,j)\,|\,j\in [\ell]\}\cup \{(k-i-1,k-j-1)\,|\,j\in [k-\ell-1]\}$. Note that the sets $L_i$, $i \in [k]$, are vertex-disjoint, so $T$ is disjoint from one of them, say $L_m$. It follows that the set of vertices $C=L_m \cup \{(i,\ell)\,|\, m \le i \le k-m-1\}\subseteq L_m \cup R_\ell$, is disjoint from $T$. Observe that $C$ contains an odd cycle, which contradicts the fact that $T$ was an odd cycle transversal. \end{proof} \Cref{thm:OCT-upper} immediately implies the following almost optimal lower bound on the independence number of projective quadrangulations, which may be new. \begin{corollary} Let $G$ be a non-bipartite projective quadrangulation on $n$ vertices. Then $\alpha(G) \geq \tfrac12\left(n-\tfrac14-\sqrt{n-15/16}\right)$. Moreover, the graph $G_k$ satisfies $\alpha(G_k) = \frac n2-\frac{\sqrt{n}}2$. \end{corollary} \begin{proof} By \Cref{thm:OCT-upper} $G$ has an odd cycle transversal $S$ such that $|S|\le \tfrac14+\sqrt{n-15/16}$. The graph $G-S$ is bipartite on more than $n-\tfrac14-\sqrt{n-15/16}$ vertices, so at least one color class of $G-S$ has more than $\tfrac12\left(n-\tfrac14-\sqrt{n-15/16}\right)$ vertices. We now focus on $G_k$. Since it contains an odd cycle transversal $T$ of $k=\sqrt{n}$ vertices (for instance, take $T=\{(i,i)\,| i \in [k]\}$), it also contains a stable set of size $\tfrac12(n-\sqrt{n})$. We now prove that this bound is tight. For the sake of contradiction, assume that there is a stable set $S$ of size more than $\frac n2-\frac{\sqrt{n}}2=\frac12 (k^2-k)$. Assume first that $k$ is even, and for $0\le i \le k/2$, let $K_i=R_i \cup R_{k-1-i}$. Recall that the $\ell$-th row $R_\ell$ of $G_k$ consists of the vertices $(i,\ell)$, with $i \in [k]$. Observe that each $K_i$ contains a cycle of length $2k$, and since $S$ is a stable set, $|S \cap K_i| \le k$. If $|S\cap K_i| \le k-1$ for every $i\in [k/2]$, then $S$ contains at most $\tfrac{k}2(k-1)=\frac12 (k^2-k)$ vertices, which a contradiction. Therefore $|S \cap K_i| = k$ for some index $i\in [k/2]$. As $(0,i)$ and $(k-1,k-1-i)$ are adjacent it follows that for each $j\in [k]$, $(j,i)\in S$ if and only if $(j,k-1-i) \in S$. Since $k$ is even, we also have that for each $j\in [k]$, $(j,i)\in S$ if and only if $(j-1-j,i)\not\in S$. Let $C_\ell$ be the $\ell$-th column of $G_k$, i.e., the vertices $(\ell,j)$ with $j \in [k]$. By the same argument as above, there exists an index $j$ such that $|S\cap L_j\|=k$, where $L_j=C_j \cup C_{k-1-j}$. As before, we have $(j,i)\in S$ if and only if $(j,k-1-i) \not\in S$, which contradicts the previous paragraph. The proof of the case when $k$ is odd is quite similar and we therefore omit it. The only difference is that in this case the middle row $R_{\lfloor k/2\rfloor}$ and the middle column $C_{\lfloor k/2\rfloor}$ each induce a cycle on $k$ vertices, which therefore contains at most $\tfrac12(k-1)$ vertices of $S$. \end{proof} \section{Almost independent odd cycle transversals}\label{sec:jap} Recall that by Theorem~\ref{thm:oddcycle}, every non-bipartite projective quadrangulation contains an odd cycle of length $O(\sqrt{n})$. This odd cycle is also and odd cycle transversal, and by increasing its size by at most one, we can even make sure that it induces a \emph{proper} subgraph of an odd cycle, i.e., a union of paths (in particular, a bipartite graph). Therefore, every non-bipartite projective quadrangulation can be properly colored with colors $1,2,3,4$ in such a way that only $O(\sqrt{n})$ vertices are colored 1 or 2. Nakamoto and Ozeki (private communication) have asked whether the $4$-coloring might be chosen in such a way that one color class has size $1$ and another has size $o(n)$. We now give an affirmative answer to their question for graphs with sublinear maximum degree. We start with a lemma, whose proof is quite similar to that of Theorem~\ref{thm:oddcycle}. In what follows, the neighborhood of a vertex $v$ is denoted by $N(v)$. \begin{lemma}\label{lem:neighborhood} Let $G=(V,E)$ be a non-bipartite quadrangulation of ${\mathbb P}^2$ and $S$ an odd support set of $G$. Then there is a subset $T \subseteq \bigcup_{v \in S} N(v)$ such that $G[T]$ contains a single edge, and $G-T$ is bipartite. \end{lemma} \begin{proof} Recall that the order of a support set $S$ of $G$ is the number of pairs of consecutive vertices of $S$ that are adjacent in $G$. We prove that there is a support set of order $1$ that is included in $\bigcup_{i=1}^s N(v_i)$. This support set can be obtained as follows. Choose a pair $v_i,v_{i+1}$ of consecutive vertices of $S$ that are adjacent, and let $v_j,v_{j+1}$ be the next pair of consecutive vertices that are adjacent (possibly, $j=i+1$). Then all pairs of consecutive vertices $v_k,v_{k+1}$, with $i<k<j$, are opposite. For each vertex $v_k$, $i<k\le j$, let $s_k$ be a sequence of consecutive neighbors of $v_k$, in their circular order around $v_k$, such that $s_{i+1}$ starts with $v_i$, $s_j$ ends with $v_{j+1}$, and for any $i<k< j$, the last vertex of $s_k$ and the first vertex of $v_{k+1}$ coincide (see Figure~\ref{fig:neighborhood}). We then replace the sequence $v_i,v_{i+1},\ldots,v_{j+1}$ in $S$ by the concatenation of the sequences $s_k$, for $i<k\le j$ (where the last vertex of each sequence $s_k$ is identified with the first vertex of $s_{k+1}$). Note that any two consecutive vertices in this new subsequence are opposite. \smallskip \begin{figure}[htbp] \centering \includegraphics[scale=1]{neighborhood} \caption{Obtaining an odd support set of order 1. Vertices of $S$ are depicted with white dots and vertices of $T$ are depicted with white squares.} \label{fig:neighborhood} \end{figure} We repeat the operation, starting at the next pair (after $v_{j+1}$) of consecutive vertices that are adjacent, until only one such pair remains. The support set thus obtained has order $1$ (and is therefore odd) and is a subset of $\bigcup_{v\in S} N(v)$, as desired. By Lemma~\ref{lem:shortest}, there is a support set $T\subseteq \bigcup_{v\in S} N(v)$ of order $1$ that is not self-intersecting and such that there is a unique pair of adjacent vertices in $T$, and these two vertices are consecutive. It follows that $T$ induces a subgraph with a unique edge, and by Lemma~\ref{lem:parity}, $G-T$ is bipartite. \end{proof} We now turn to the proof of \Cref{thm:sqrtD}. \begin{proof}[Proof of Theorem~\ref{thm:sqrtD}] Let $C$ be a shortest odd cycle in $G$, and let $\ell$ be the length of $C$. Assume first that $\ell\le \sqrt{2n/\Delta}$. Since an odd cycle is an odd support set, it follows from Lemma~\ref{lem:neighborhood} that the union of the neighborhoods of the vertices of $C$ contain a set $T$ of vertices inducing a single edge, and such that $G-T$ is bipartite. Since $G$ has maximum degree at most $\Delta$, $T$ contains at most $\ell \Delta\le \sqrt{2n \Delta}$ vertices, as desired. Assume now that $\ell\ge \sqrt{2n/\Delta}$. Since $G$ has $2n-2$ edges, the dual graph $G^*$ of $G$ has a non-contractible cycle $C^*$ of length less than $2n/\ell$ by Lemma~\ref{lem:lins}. Let $(f_1,f_2,\ldots,f_k)$ be the faces of $G$ corresponding to the vertices of $C^*$. Note that any two consecutive faces $f_i,f_{i+1}$ in $C^*$ share an edge, call it $e_i$. For any edge $e_i$, we choose one of the endpoints $v_i$ of $e_i$ as follows: we start by choosing $v_1$ in $e_1$ arbitrarily, and for any $i>1$ we distinguish two cases. If $v_{i-1} \in e_i$, then we set $v_i=v_{i-1}$, and otherwise we choose for $v_i$ the vertex opposite to $v_{i-1}$ in $f_i$ (note that in this case such vertex is necessarily an endpoint of $e_i$). Note that $S=(v_i\,|\,1\le i \le k)$ is a support set in $G$, and $\rho(S)$ is homologous to $C^*$, and thus non-contractible. By Lemma~\ref{lem:parity}, $S$ is an odd support set (in particular $v_k$ and $v_1$ are adjacent, since all the other pairs of consecutive vertices of $S$ are opposite). By Lemma~\ref{lem:shortest}, $G$ contains a set $S'$ of less than $2n/\ell \le \sqrt{2n \Delta} $ vertices such that $G-S'$ is bipartite and $S'$ induces a subgraph of $G$ with a single edge. This concludes the proof of Theorem~\ref{thm:sqrtD}. \end{proof} Note that a subgraph with a single edge has a proper $2$-coloring such that one of the color classes is a singleton. It follows that $n$-vertex projective quadrangulations with $\Delta=o(n)$ can be $4$-colored in such way that one color class is a singleton and another has size $o(n)$. \section{Conclusion}\label{sec:conclusion} We have seen that Theorem~\ref{thm:oddcycle2} is sharp for infinitely many values of $n$. A natural question is whether the generalized Mycielski graphs are the only extremal graphs. As for the problem of finding a smallest odd cycle transversal, we believe that Theorem~\ref{thm:OCT-upper} is not sharp and that the right bound should be $\sqrt{n}$, which would be tight by Theorem~\ref{thm:OCT-lower}. It was proved by Tardif~\cite{Tar01} that the generalized Mycielki graphs have \emph{fractional} chromatic number $2+o(\tfrac1{n})$, so a natural question is whether the same holds for projective quadrangulations of large edge-width. Note that it was proved by Goddyn~\cite{Goddyn} (see also~\cite{DGMVZ05}), in the same spirit as the result of Youngs~\cite{You96} on the chromatic number, that the \emph{circular} chromatic number of a projective quadrangulation is either $2$ or $4$. \bibliographystyle{plain}

https://arxiv.org/abs/2206.15452

On residues of rounded shifted fractions with a common numerator

For any positive integer $n$ along with parameters $\alpha$ and $\nu$, we define and investigate $\alpha$-shifted, $\nu$-offset, floor sequences of length $n$. We find exact and asymptotic formulas for the number of integers in such a sequence that are in a particular congruence class. As we will see, these quantities are related to certain problems of counting lattice points contained in regions of the plane bounded by conic sections. We give specific examples for the number of lattice points contained in elliptical regions and make connections to a few well-known rings of integers, including the Gaussian integers and Eisenstein integers.

\section{Introduction} For a fixed positive integer $n$, consider the integer sequence \begin{equation}\label{seq:intro} \Floor{\frac{n}{1}}, \Floor{\frac{n}{2}}, \Floor{\frac{n}{3}}, \dots, \Floor{\frac{n}{n}}, \end{equation} where $\Floor{x}$ denotes the floor function. Among the $n$ terms in this sequence, how many are odd? At first glance, it maybe seems reasonable to expect the proportion of odd numbers in the sequence to be roughly half. For $n=10$, the sequence is $10, 5, 3, 2, 2, 1, 1, 1, 1, 1$, of which $7/10=70\%$ are odd. In general, terms in the second half of the sequence are all 1 and thus odd, implying the proportion of odd terms is always at least $50\%$. Looking at more data, this proportion appears to hover around $69\%$ as $n$ grows. This leads to a few natural questions. Why $69\%$? What happens if we replace the floor function with the ceiling function? Or if we round to the nearest integer? In order to answer these questions, we consider three integer sequences defined, for positive integers $n$, in terms of the floor, ceiling, and nearest integer rounding functions: \begin{align*} \Fseq_n &= \#\left\{k\in\Z : 1\le k\le n,\, \Floor{n/k}\text{ is odd}\right\},\\ \Cseq_n &= \#\left\{k\in\Z : 1\le k\le n,\, \Ceil{n/k}\text{ is odd}\right\},\\ \Rseq_n &= \#\left\{k\in\Z : 1\le k\le n,\, \nint{n/k}\text{ is odd}\right\}, \end{align*} where $\Ceil{x}$ denotes the ceiling function and $\nint{x}$ the nearest integer rounding function\footnote{The definition of the nearest integer function can be ambiguous for half-integers. A common convention is to round to the nearest even integer. Since we are interested in parity, we instead choose to always round half-integers up. Hence, $\nint{2.5}=3$, $\nint{3.5}=4$, and so on. As we will see, the asymptotic formula for $\Rseq_n$ will be the same regardless of how we choose to round half-integers.}. \begin{figure} \centering \begin{tabular}{c|rrrrrrrrrrrrrrrrrrrr} $n$ & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & 16 & 17 & 18 & 19 & 20 \\ \hline $\Fseq_n$ & 1 & 1 & 3 & 2 & 4 & 4 & 6 & 4 & 7 & 7 & 9 & 7 & 9 & 9 & 13 & 10 & 12 & 12 & 14 & 12 \\ $\Cseq_n$ & 1 & 1 & 2 & 1 & 3 & 2 & 3 & 2 & 5 & 3 & 4 & 3 & 6 & 5 & 6 & 3 & 7 & 6 & 7 & 6 \\ $\Rseq_n$ & 1 & 1 & 2 & 2 & 4 & 3 & 4 & 4 & 6 & 7 & 6 & 5 & 9 & 8 & 9 & 9 & 10 & 10 & 11 & 12 \end{tabular} \caption{Terms in the sequences $\Fseq_n$, $\Cseq_n$, $\Rseq_n$ for $1\le n\le 20$} \label{fig:first-20-terms} \end{figure} The first 20 terms of each of these three sequences can be found in Figure~\ref{fig:first-20-terms}. The sequences $\Fseq_n$ and $\Cseq_n$ appear in The On-Line Encyclopedia of Integer Sequences (\cite{OEIS}) as, respectively, sequences \href{https://oeis.org/A059851}{A059851} and \href{https://oeis.org/A330926}{A330926}. Plots of $\Fseq_n$, $\Cseq_n$, and $\Rseq_n$, for $1\le n\le 1000$, appear in Figure~\ref{fig:first-three-sequences}. The plots in Figure~\ref{fig:first-three-sequences} suggest that each sequence is roughly linear. As we will show, this is true asymptotically. We will find asymptotic formulas for each of these sequences, and from those we will conclude that \[ \lim\limits_{n\to\infty}\dfrac{\Fseq_n}{n}=\log2, \quad \lim\limits_{n\to\infty}\dfrac{\Cseq_n}{n}=1-\log2, \quad\text{ and } \lim\limits_{n\to\infty}\dfrac{\Rseq_n}{n}=\frac{\pi}{2}-1. \] In particular, the proportion of odd terms in $\Fseq_n$ is asymptotically $\log2\approx 0.693147$. This explains the $69\%$ mentioned in the first paragraph. \begin{figure} \centering \includegraphics[width=.6\textwidth]{first-three-sequences.eps} \caption{Plots of $y=\Fseq_n$ (red, top-most), $y=\Cseq_n$ (black, middle), and $y=\Rseq_n$ (blue, bottom-most), for $1\le n\le 1000$} \label{fig:first-three-sequences} \end{figure} Our results follow from two classical results in analytic number theory: the \emph{Dirichlet divisor problem} (Theorem~\ref{thm:dirichlet}), in which one wants to count the number of lattice points in the first quadrant of the plane beneath a hyperbola; and the \emph{Gauss circle problem} (Theorem~\ref{thm:gauss}), in which one wants to count the number of lattice points contained within a circle in the plane. Each of these classical results leads to a geometric interpretation for one of our sequences. In Proposition~\ref{prop:F_n-Dirichlet}, we will see that $\Fseq_n$ is the difference of the numbers of lattice points in two hyperbolic regions in the plane. In Proposition~\ref{prop:gauss-R_n}, we will see that the number of lattice points contained in a circle of radius $\sqrt{2n}$ is equal to $4\Rseq_n+4n+1$. We then consider more general sequences. Since $\nint{x}=\Floor{x+1/2}$, we can think of the nearest integer function as the usual floor function shifted by $1/2$. If we replace this $1/2$ with an arbitrary real number $\alpha$ and also incorporate a real number $\nu$ to offset the numerator $n$, we obtain an \emph{$\alpha$-shifted, $\nu$-offset, floor sequence of length $n$} \begin{equation}\label{seq:alpha-shift-intro} \Floor{\frac{n-\nu}{1}+\alpha}, \Floor{\frac{n-\nu}{2}+\alpha}, \dots, \Floor{\frac{n-\nu}{n}+\alpha}. \end{equation} We can determine how many of the integers in Sequence~\eqref{seq:alpha-shift-intro} are odd. A natural problem, which we will address, is to find $\alpha$ and $\nu$ for which (asymptotically) half of these integers are odd and half are even. We will see that there is a unique value for $\alpha\in[0,1]$, independent of $\nu$, for which this occurs, and numerically approximate it. Another problem is to count the number of integers in Sequence~\eqref{seq:alpha-shift-intro} that belong to a given congruence class of integers. We will find exact and asymptotic formulas for these counts. In particular, by Corollary~\ref{cor:N_m-slope}, for any $m\in\N$ and $\alpha\in[0,1)$, the proportion of integers in an $\alpha$-shifted, $\nu$-offset, floor sequence of length $n$ that are congruent to 1 modulo $m$ is asymptotically equal to \[ \frac{-\alpha}{1-\alpha}+ \int\limits_0^1 \frac{(1-x)x^{-\alpha}}{1-x^m}\mathrm{d}x \] and, for $2\le r\le m$, the proportion of such integers that are congruent to $r$ modulo $m$ is asymptotically equal to \[ \int\limits_0^1 \frac{(1-x)x^{r-1-\alpha}}{1-x^m}\mathrm{d}x. \] Finally, we will look more closely at connections between the number of integers in a certain congruence class in an $\alpha$-shifted, $\nu$-offset, floor sequence of length $n$ and the number of lattice points in a certain region of the plane. We will demonstrate the connections by obtaining formulas in terms of these counts for the number of lattice points in the following elliptical regions: $x^2+y^2\le n$, $x^2+xy+y^2\le n$, and $x^2+2y^2\le n$, each for any $n\in\N$. Our results here follow from the theory of binary quadratic forms. \subsection{Notation} $\N$ denotes the set of positive integers. All logarithms in this paper use base $\ee$. The cardinality of a finite set $A$ is denoted $\#A$. We use big O notation as follows. For functions $f(x)$ and $g(x)$, if there exist constants $M$ and $a$ such that $|f(x)|\le M g(x)$ for all $x\ge a$, then we write $f(x)=\bigO(g(x))$. \subsection{Organization} This paper is organized as follows. In Section~\ref{sec:F_n and C_n}, we find exact and asymptotic formulas for $\Fseq_n$ and $\Cseq_n$. In Section~\ref{sec:Rseq_n}, we do the same for $\Rseq_n$. In Section~\ref{sec:generalized}, for integers $r,m$ with $1\le r\le m$, and for $\alpha\in[0,1)$, we find exact and asymptotic formulas for $\Num_{n,\alpha,\nu,r,m}$, the number of integers in the $\alpha$-shifted, $\nu$-offset, floor sequence of length $n$ that are congruent to $r$ modulo $m$. Finally, in Section~\ref{sec:applications}, we have two tasks. The first task is to compute a shift $\alpha=\alpha_0$ for which (asymptotically) the $\alpha$-shifted, $\nu$-offset, floor sequence of length $n$ contains as many odd terms as even terms. The second task is to determine the number of lattice points in certain elliptical regions of the plane in terms of the sequences $\Num_{n,\alpha,\nu,r,m}$. \section{The floor and ceiling sequences}\label{sec:F_n and C_n} \subsection{The floor sequence} As was mentioned in the introduction, for $n\in\N$, $\Fseq_n$ is equal to the number of odd integers in the floor sequence \begin{equation}\label{seq:floor} \Floor{\frac{n}{1}}, \Floor{\frac{n}{2}}, \dots, \Floor{\frac{n}{n}}, \end{equation} where the floor function is defined, for $x\in\mathbb{R}$, by $\Floor{x}=\max\{z\in\Z : z\le x\}$. In this subsection, we will find an exact formula for $\Fseq_n$. Then, with that formula and the solution to Dirichlet's divisor problem, we will find an asymptotic formula for $\Fseq_n$. We start by counting the number of integers in Sequence~\eqref{seq:floor} that are greater than or equal to a given integer. \begin{lem}\label{lem:consec_flr} Fix $n\in\N$. For $k\in \mathbb{N}$, Sequence~\eqref{seq:floor} contains $\Floor{n/k}$ terms that are greater than or equal to $k$. \end{lem} \begin{proof} Via the division algorithm, there are $q,r\in \mathbb{Z}$ for which $n=kq+r$ with $0\leq r< k$. Then $\Floor{n/k}=q$. We will show that the first $q$ terms in Sequence~\eqref{seq:floor} are greater than or equal to $k$. Note that \[ \flrf{n}{q} = \flrf{kq+r}{q} = \Floor{k+\frac{r}{q}}\ge k. \] Now consider an arbitrary term $\Floor{n/d}$ in Sequence~\eqref{seq:floor}. Then $1\leq d\leq n$. If $d < q$, then $n/d>n/q$ and hence $\Floor{n/d}\ge\Floor{n/q}\ge k$. On the other hand, if $d > q$, then $d\ge q+1$. This means $kd \ge k(q+1) > n$ and therefore $k>n/d\geq \Floor{n/d}$. We have already shown that $\Floor{n/d}\geq k$ if $d=q$. We conclude that $\Floor{n/d}\geq k$ if and only if $d \in \left\{1,2,\dots ,q\right\}$. \end{proof} We can now find an exact formula for $\Fseq_n$. \begin{prop}\label{prop:floor_seq} For $n\in\N$, \[ \Fseq_n =\sum\limits_{d=1}^n\Floor{\frac{n}{d}}(-1)^{d+1}. \] \end{prop} \begin{proof} We wish to count the number of odd terms in Sequence~\eqref{seq:floor}. Here we can use Lemma~\ref{lem:consec_flr} to note that for a given $k$ there are $\Floor{n/k}$ terms whose value is at least $k$, and thus $\Floor{n/k}-\Floor{n/(k+1)}$ terms whose value is exactly $k$. To compute $\Fseq_n$, we sum for all odd $k$. We find \[ \Fseq_n =\sum\limits_{\substack{1\le k\leq n \\ k \text{ odd}}} \left(\flrf{n}{k}-\flrf{n}{k+1}\right) =\sum\limits_{d=1}^n\Floor{\frac{n}{d}}(-1)^{d+1}, \] as desired. \end{proof} Next, we wish to find an asymptotic formula for $\Fseq_n$. To start, we will manipulate the formula for $\Fseq_n$ obtained in Proposition~\ref{prop:floor_seq} by considering odd and even $d$ separately. \begin{align*} \Fseq_n &=\sum\limits_{d\text{ odd}} \Floor{\frac{n}{d}}-\sum\limits_{d\text{ even}} \Floor{\frac{n}{d}}\\ &=\sum\limits_{d\text{ odd}} \Floor{\frac{n}{d}}-\sum\limits_{d\text{ even}} \Floor{\frac{n}{d}}+\sum\limits_{d\text{ even}} \Floor{\frac{n}{d}}-\sum\limits_{d\text{ even}} \Floor{\frac{n}{d}}\\ &=\sum\limits_{d=1}^n\Floor{\frac{n}{d}} - 2\sum\limits_{b=1}^{\Floor{n/2}}\Floor{\frac{n}{2b}}\\ &=\sum\limits_{d=1}^n\Floor{\frac{n}{d}} - 2\sum\limits_{b=1}^{\Floor{n/2}}\Floor{\frac{n/2}{b}}. \end{align*} We will now use $\tau(n)$, the multiplicative function which counts the number of positive divisors of $n$, and $\mathrm{D}(x)=\sum\limits_{m\le x}\tau(m)$, the \emph{divisor summatory function}, where the summation is taken over positive integers $m$. With this function, we have the following: \[ \mathrm{D}(n) = \sum\limits_{m=1}^n\tau(m) = \sum\limits_{m=1}^n\sum\limits_{d\mid m}1 = \sum\limits_{d=1}^n\Floor{\frac{n}{d}}. \] Geometrically, $\mathrm{D}(n)$ is the number of lattice points in the interior and boundary of the region in the $xy$-plane bounded by the graphs of $x=1$, $y=1$, and the hyperbola $xy=n$. Observe that $\Fseq_n=\mathrm{D}(n)-2\mathrm{D}(n/2)$. For $n\in\N$, we define two regions. Let $H_{1,n}$ be the hyperbolic region in the $xy$-plane with $x,y\ge1$ bounded by the graphs of the hyperbola $xy=n$ and the hyperbola $xy=n/2$. Let $H_{2,n}$ be the region in the $xy$-plane bounded by the graphs of $x=1$, $y=1$, and the hyperbola $xy=n/2$. Then $\mathrm{D}(n)$ counts the number of lattice points in $H_{1,n}\cup H_{2,n}$, and $\mathrm{D}(n/2)$ counts the number of lattice points in $H_{2,n}$. This gives us a geometric interpretation for $\Fseq_n$. \begin{prop}\label{prop:F_n-Dirichlet} For $n\in\N$, $\Fseq_n$ is equal to the number of lattice points in $H_{1,n}$ minus the number of lattice points in $H_{2,n}$. For $H_{1,n}$, we include all points on the boundary except for those on the boundary curve $xy=n/2$. For $H_{2,n}$, we include all points on the boundary. \end{prop} \begin{figure} \centering \includegraphics[scale=.7]{hyperbola-points.eps} \caption{For $n=17$, the graphs of $xy=17$ and $xy=17/2$ along with lattice points in the hyperbolic regions $H_{1,17}$ (circles) and in $H_{2,17}$ (stars).} \label{fig:hyperbola-F_n} \end{figure} \begin{example} To illustrate Proposition~\ref{prop:F_n-Dirichlet}, we have drawn the regions $H_{1,n}$ and $H_{2,n}$ for $n=17$ in Figure~\ref{fig:hyperbola-F_n}. Since $H_{1,17}$ contains 32 points and $H_{2,17}$ contains 20 points, we find $\Fseq_{17}=32-20=12$. \end{example} If we have an asymptotic formula for $\mathrm{D}(x)$, known as the \emph{Dirichlet divisor problem}, then we get an asymptotic formula for $\Fseq_n$. The following theorem of Dirichlet is exactly what we need. (For details, see, e.g., \cite[Theorem 3.3]{Apostol1976}. We will use a slightly modified version of this approach in the proof of Proposition~\ref{prop:sum-of-floors-to-sum-of-fractions} later in this paper.) \begin{thm}\label{thm:dirichlet} For all $x\ge1$, \[ \mathrm{D}(x) =x\log{x}+(2\gamma-1)x+\bigO(\sqrt{x}), \] where $\gamma = \lim\limits_{n\to\infty}\left(-\log{n}+\sum\limits_{k=1}^n 1/k\right)\approx 0.577216$ is the Euler-Mascheroni constant. \end{thm} Next, let $F(x)=\mathrm{D}(x)-2\mathrm{D}(x/2)$. Then $\Fseq_n=F(n)$ and, via Theorem~\ref{thm:dirichlet} we have $F(x)=x\log2+\bigO(\sqrt{x})$. Restricting to integers $n$, we obtain the following result. \begin{prop}\label{prop:floor_seq_asymp} For $n\in\N$, $\Fseq_n = n\log2+\bigO\left(\sqrt{n}\right)$, and hence $\lim\limits_{n\to\infty}\frac{1}{n}\Fseq_n=\log{2}\approx0.693147$. \end{prop} \subsection{The ceiling sequence} As was mentioned in the introduction, for $n\in\N$, $\Cseq_n$ is equal to the number of odd integers in the ceiling sequence \begin{equation}\label{seq:ceiling} \Ceil{\frac{n}{1}}, \Ceil{\frac{n}{2}}, \dots, \Ceil{\frac{n}{n}}, \end{equation} where the ceiling function is defined, for $x\in\mathbb{R}$, by $\Ceil{x}=\min\{z\in\Z : z\ge x\}$. In this subsection, we will find a relation between $\Cseq_n$ and $\Fseq_{n-1}$ which, along with results from the previous subsection, will lead to exact and asymptotic formulas for $\Cseq_n$. \begin{prop}\label{prop:ceiling_seq} For $n\in\N$, \[ \Cseq_n = n-\Fseq_{n-1} = \sum\limits_{d=2}^n \Floor{\frac{n}{d}}(-1)^d. \] \end{prop} \begin{proof} We begin by showing that $\Ceil{n/k}=\Floor{(n-1)/k}+1$ for all $k\in\mathbb{N}$. As in Lemma~\ref{lem:consec_flr}, via the division algorithm, there are $q,r\in \mathbb{Z}$ for which $n=kq+r$ with $0\leq r< k$. If $r=0$, then $n/k=q$ and so $\Ceil{n/k}=\Ceil{q}=q$. We also have \[ \flrf{n-1}{k}=\left\lfloor\frac{n}{k}-\frac{1}{k}\right\rfloor =\left\lfloor q-\frac{1}{k}\right\rfloor =q-1. \] Thus, $\Ceil{n/k}=\Floor{(n-1)/k}+1$. If $r\ne0$, then $1\le r<k$. We have \[ \clf{n}{k} =\clf{kq+r}{k} =\left\lceil q+\frac{r}{k}\right\rceil = q+1. \] and \[ \flrf{n-1}{k} =\flrf{kq+r-1}{k} =\left\lfloor q+\frac{r-1}{k}\right\rfloor =q, \] with the final equality following from the fact that $1\le r<k$. Thus, $\Ceil{n/k}=\Floor{(n-1)/k}+1$. Since we have $\Ceil{n/k}=\Floor{(n-1)/k}+1$ for all $k\in\N$, for each pair of integers $\left(\Ceil{n/k},\Floor{(n-1)/k}\right)$, exactly one integer is odd. Thus, the total number of odd integers in the two sequences \[ \Ceil{\frac{n}{1}},\Ceil{\frac{n}{2}},\dots,\Ceil{\frac{n}{n}} \text{ and } \Floor{\frac{n-1}{1}},\Floor{\frac{n-2}{2}},\dots,\Floor{\frac{n-1}{n}} \] is $n-1$. The number of odd integers in the first sequence is $\Cseq_n$. Since the last term in the second sequence is $\Floor{(n-1)/n}=0$, the number of odd integers in the second sequence is $\Fseq_{n-1}$. Thus, $\Fseq_{n-1}+\Cseq_n=n$. The stated result the follows from Proposition~\ref{prop:floor_seq}. \end{proof} We immediately obtain the following asymptotic formula for $\Cseq_n$. \begin{cor}\label{cor:ceiling_seq_asymp} For $n\in\N$, $\Cseq_n = (1-\log2)n+\bigO\left(\sqrt{n}\right)$, and hence $ \lim\limits_{n\to\infty}\frac{1}{n}\Cseq_n =1-\log2 \approx 0.306853$. \end{cor} \begin{proof} Combining Proposition~\ref{prop:ceiling_seq} with the asymptotic formula for $\Fseq_n$ in Proposition~\ref{prop:floor_seq_asymp}, we find \[ \Cseq_n = n - \Fseq_{n-1} = n - (n-1)\log2 + \bigO\left(\sqrt{n-1}\right) = n-n\log2 + \bigO\left(\sqrt{n}\right). \] The limit result follows. \end{proof} \begin{remark}\label{rmk:an-and-bn} If we just wanted an asymptotic formula for $\Cseq_n$ without finding an exact formula first, we could have used the asymptotic formula for $\Fseq_n$, the observation that \[ \Ceil{n/k} = \begin{dcases*} \Floor{n/k} & if $k\mid n$ \\ \Floor{n/k}+1 & if $k\nmid n$, \end{dcases*} \] and the following lemma. \begin{lem}\label{lem:num_divisors_of_n} For $n\in\N$, the number of positive divisors of $n$ is at most $2\sqrt{n}$. \end{lem} \begin{proof} Suppose $n=de$ for positive integers $d\le e$. Then $1\le d\le \sqrt{n}$. It follows that there are at most $\sqrt{n}$ pairs of divisors $d,e$, and thus at most $2\sqrt{n}$ positive divisors of $n$. \end{proof} Thus, whenever $k$ does not divide $n$, exactly one of $\Floor{n/k}$ and $\Ceil{n/k}$ is odd. When $k$ does divide $n$, then either both $\Floor{n/k}$ and $\Ceil{n/k}$ are odd, or neither is. Since there are at most $2\sqrt{n}$ divisors $d$ of $n$, we have \[ \Fseq_n+\Cseq_n \in [n-2\sqrt{n},n+2\sqrt{n}]. \] Thus $\Fseq_n+\Cseq_n = n + \bigO\left(\sqrt{n}\right)$, from which we conclude that $\Cseq_n=n-\Fseq_n+\bigO\left(\sqrt{n}\right) = (1-\log2)n+\bigO\left(\sqrt{n}\right)$. This gives an alternate proof of Corollary~\ref{cor:ceiling_seq_asymp}. \end{remark} \section{The nearest integer sequence}\label{sec:Rseq_n} As was mentioned in the introduction, for $n\in\N$, $\Rseq_n$ counts the number of odd integers in the nearest integer (rounding) sequence \begin{equation}\label{seq:nint-seq} \nint{\frac{n}{1}}, \nint{\frac{n}{2}}, \dots, \nint{\frac{n}{n}} \end{equation} where the nearest integer function is defined, for $x\in\mathbb{R}$, by \[\nint{x}=\max\{z\in\Z : |z-x|\le |z'-x|\text{ for all } z'\in\Z\}.\] (In other words, we round to the nearest integer, and half-integers are rounded up.) In this subsection, we will find an exact formula for $\Rseq_n$. Then, with that formula and the solution to Gauss's circle problem, we will find an asymptotic formula for $\Rseq_n$. We should first note that $\nint{n/k}=\Floor{n/k+1/2}$, and thus Sequence~\eqref{seq:nint-seq} is equal to the sequence \begin{equation}\label{seq:rounding-seq} \Floor{\frac{n}{1}+\frac{1}{2}}, \Floor{\frac{n}{2}+\frac{1}{2}}, \Floor{\frac{n}{3}+\frac{1}{2}}, \dots, \Floor{\frac{n}{n}+\frac{1}{2}}. \end{equation} Just as we did in obtaining an exact formula for $\Fseq_n$ (Proposition~\ref{prop:floor_seq}), we begin with a lemma. \begin{lem}\label{lem:round_count} Fix $n\in\N$. For $k\in\N$, let $g(k)$ equal the number of terms in Sequence~\eqref{seq:rounding-seq} that are greater than or equal to $k$. Then $g(1)=n$ and, for $k\ge2$, $g(k)=\Floor{2n/(2k-1)}$. \end{lem} \begin{proof} Consider an arbitrary term $\Floor{n/d+1/2}$ in the Sequence~\eqref{seq:rounding-seq}, so that $1\le d\le n$. We first observe that $\Floor{n/d+1/2}\ge\Floor{n/n+1/2} = 1$. Hence, $g(1)=n$. Next, for $k\ge2$, we have either $d\le 2n/(2k-1)$ or $d>2n/(2k-1)$. We'll consider these cases separately. If $d\leq 2n/(2k-1)$, then $k-1/2 \leq n/d$ which implies $k \leq n/d+1/2$. Since $k$ is an integer, we take the floor of both sides to find $k \leq\Floor{n/d+1/2}$. If $d>2n/(2k-1)$, then $k-1/2>n/d$, which implies $k > n/d+1/2 \geq \Floor{n/d+1/2}$. Thus, for $k\ge2$, we have $\Floor{n/d+1/2}\geq k$ for $d \in\left\{1, 2, \dots,\Floor{2n/(2k-1)}\right\}$. Since $\Floor{2n/(2k-1)}\le n$, we conclude that $g(k)=\Floor{2n/(2k-1)}$. \end{proof} With this lemma, we can now find an exact formula for $\Rseq_n$. \begin{prop}\label{prop:rounding_seq} For $n\in\N$, \[ \Rseq_n = -n+\sum\limits_{d=1}^{n}\Floor{\frac{2n}{2d-1}}(-1)^{d+1}. \] \end{prop} \begin{proof} We wish to count the number of odd terms in Sequence~\eqref{seq:nint-seq}. By Lemma~\ref{lem:round_count}, there are $n$ terms that are at least 1, and, for $k\ge2$, there are $\Floor{2n/(2k-1)}$ terms that are at least $k$. We can use this to count the number of terms that are equal to a given value. For $k=1$, there are $n-\Floor{2n/3}$ terms that are equal to 1. For $k\ge2$, there are $\Floor{2n/(2k-1)}-\Floor{2n/(2k+1)}$ terms that are equal to $k$. Summing over odd $k$, we find \begin{align*} \Rseq_n &=n-\Floor{\frac{2n}{3}}+\sum\limits_{\substack{3\le k\le n \\ k\text{ odd}}}\Floor{\frac{2n}{2k-1}}-\Floor{\frac{2n}{2k+1}} \\ &=-n+\sum\limits_{\substack{1\le k\leq n\\ k\text{ odd}}} \flrf{2n}{2k-1}-\flrf{2n}{2k+1}\\ &=-n+\sum\limits_{d=1}^n\Floor{\frac{2n}{2d-1}}(-1)^{d+1}, \end{align*} which completes the proof. \end{proof} Next, we will work toward an asymptotic formula for the summation in $\Rseq_n$. We will use results of Jacobi and Gauss. To start, we need some notation. For $n\in\N$, $r\in\Z$, and $m\in\N$, let $\divs_{r,m}(n)$ denote the number of positive divisors of $n$ that are congruent to $r$ modulo $m$. We will be interested in values of this function for $m=4$ and $r$ equal to 1 or 3. For example, with $n=45$, we have $\divs_{1,4}(45)=4$ and $\divs_{3,4}(45)=2$ because the positive divisors of 45 are 1, 3, 5, 9, 15, and 45. \begin{lem} For $n\in\N$, \[ \Rseq_n = -n + \sum\limits_{k=1}^{2n}\left(\divs_{1,4}(k)-\divs_{3,4}(k)\right). \] \end{lem} \begin{proof} Each integer $d$ divides $\Floor{2n/d}$ integers in the interval $[1,2n]$. If we want to count the number of divisors that are 1 modulo 4 for all of the integers from 1 to $2n$, we see that 1 is a divisor of $\Floor{2n/1}$ terms, 5 is a divisor of $\Floor{2n/5}$ terms, 9 is a divisor of $\Floor{2n/9}$ terms, and so on. In other words, we get \[ \sum\limits_{k=1}^{2n}\divs_{1,4}(k) =\Floor{\frac{2n}{1}}+\Floor{\frac{2n}{5}}+\Floor{\frac{2n}{9}}+\dots. \] Similarly, if we want to add up the number of divisors that are 3 modulo 4 for all of the integers from 1 to $2n$, we get \[ \sum\limits_{k=1}^{2n}\divs_{3,4}(k) =\Floor{\frac{2n}{3}}+\Floor{\frac{2n}{7}}+\Floor{\frac{2n}{11}}+\dots. \] Note that the terms in each of the two above summations are eventually all 0, and thus these are finite sums. Hence, \begin{align*} \sum\limits_{d=1}^n\Floor{\frac{2n}{2d-1}}(-1)^{d+1} &= \Floor{\frac{2n}{1}}-\Floor{\frac{2n}{3}}+\Floor{\frac{2n}{5}}-\Floor{\frac{2n}{7}}+\cdots+(-1)^{n+1}\Floor{\frac{2n}{2n-1}} \\ &= \sum\limits_{k=1}^{2n}\left(\divs_{1,4}(k)-\divs_{3,4}(k)\right). \end{align*} The stated result then follows from the formula for $\Rseq_n$ in Proposition~\ref{prop:rounding_seq}. \end{proof} Next, let $\Jacr_2(n)$ be the number of representations of $n$ as a sum of two integer squares. More precisely, $\Jacr_2(n)=\#\{(a,b)\in\Z^2 : a^2+b^2=n\}$. For example, still with $n=45$, $\Jacr_2(45)=8$ because \[ 45=(\pm3)^2+(\pm6)^2=(\pm6)^2+(\pm3)^2, \] which is a total of 8 combinations. Jacobi's two-square theorem relates $\Jacr_2(n)$ to the number of divisors of $n$ that are 1 modulo 4 and that are 3 modulo 4. \begin{thm}[Jacobi's two-square theorem]\label{thm:jacobi} For $n\in\N$, $\Jacr_2(n)=4(\divs_{1,4}(n)-\divs_{3,4}(n))$. \end{thm} Jacobi proved this theorem, along with theorems about the number of representations of $n$ using four squares, using six squares, and using eight squares, in 1829 with the use of elliptic theta functions. (See \cite{Grosswald1985}.) One may also prove Theorem~\ref{thm:jacobi} in the context of the Gaussian integers, which is the ring \[ \Z[\ii]=\{a+b\ii : a,b\in\Z,\, \ii^2=-1\}. \] The norm of the Gaussian integer $a+b\ii$ is $a^2+b^2$. It follows that $\Jacr_2(n)$ is equal to the number of Gaussian integers with norm equal to a given positive integer $n$. With an understanding of what the prime elements of $\Z[\ii]$ are, along with the fact that $\Z[\ii]$ is a unique factorization domain, one may prove Theorem~\ref{thm:jacobi}. (For details, see, e.g., \cite[Theorem 278]{HardyWright1979} or a wonderful video by 3Blue1Brown on YouTube \cite{3b1b-pi-prime-regularities}.) Visualizing $\Z[\ii]$ as a lattice in the complex plane (with $a+b\ii\in\Z[\ii]$ corresponding to the point $(a,b)$ in the plane), Jacobi's two-square theorem says that the number of lattice points on a circle of radius $\sqrt{n}$ centered at the origin is $4(\divs_{1,4}(n)-\divs_{3,4}(n))$. In general, the point $(a,b)$ is on a circle of radius $\sqrt{a^2+b^2}$ centered at the origin. Thus, adding the numbers of points on circles of radii $\sqrt{1}, \sqrt{2},\dots,\sqrt{2n}$ gives us the total number of lattice points different from the origin in the interior or on the boundary of a circle of radius $\sqrt{2n}$ centered at the origin. Back to our formula for $\Rseq_n$, we now have \begin{equation}\label{eqn:Rseq_n-circle} \Rseq_n = -n + (1/4)\cdot \#\{(a,b)\in\Z^2 : 0<a^2+b^2\le 2n\}. \end{equation} Let $\mathrm{C}(x)$ denote the number of lattice points in the interior or on the boundary of a circle of radius $\sqrt{x}$ centered at the origin. Finding an asymptotic formula for $\mathrm{C}(x)$ is known as the \emph{Gauss circle problem}. Equation~\eqref{eqn:Rseq_n-circle} shows us how to calculate $\mathrm{C}(2n)$ in terms of $\Rseq_n$. \begin{prop}\label{prop:gauss-R_n} For $n\in\N$, the number of lattice points within a circle of radius $\sqrt{2n}$ centered at the origin is $4\Rseq_n+4n+1$. That is, $\mathrm{C}(2n)=4\Rseq_n+4n+1$. \end{prop} \begin{figure} \centering \includegraphics[scale=.5]{gauss-circle.eps} \caption{The number of lattice points within and on a circle of radius $6=\sqrt{2\cdot18}$ is, by Proposition~\ref{prop:gauss-R_n}, $\mathrm{C}(2\cdot18)=4\Rseq_{18}+4\cdot18+1=4\cdot10+4\cdot18+1=113$.} \label{fig:gauss-circle} \end{figure} See Figure~\ref{fig:gauss-circle} for an example with $n=18$, which uses data from Figure~ \ref{fig:first-20-terms}. Geometrically, we see that $\mathrm{C}(n)$ is a non-decreasing function. For any $n\in\N$, \[ 0 \le \mathrm{C}(2n+2)-\mathrm{C}(2n) = 4\Rseq_{n+1}+4(n+1)+1-4\Rseq_{n}-4n-1 = 4\Rseq_{n+1}+4-4\Rseq_n. \] Thus, $\Rseq_{n+1}-\Rseq_n\ge-1$. This proves the following corollary. \begin{cor} The sequence $\Rseq_n$ decreases by at most 1 in any step. \end{cor} A decrease of 1 from $\Rseq_n$ to $\Rseq_{n+1}$ occurs precisely when $2n+2$ and $2n+1$ each cannot be written as a sum of two integer squares. This occurs when the prime factorizations of $2n+2$ and $2n+1$ each contain some prime which is 3 modulo 4 raised to an odd power. From our data in Figure~\ref{fig:first-20-terms}, we see a decrease by 1 for $n=5$. Observe that $2n+2=12=2^2\cdot3^1$ and $2n+1=11^1$. Neither 11 nor 12 is a sum of two integer squares. For comparison, there is no bound for the amount in which the sequences $\Fseq_n$ and $\Cseq_n$ can decrease in any step. Indeed, for $k\in\N$ and $n=2^k$, $\Fseq_n-\Fseq_{n-1}=\Cseq_n-\Cseq_{n-1}=-(k-1)$. We now want an asymptotic formula for $\Rseq_n$. To get there, we use an asymptotic result for $\mathrm{C}(x)$ that is due to Gauss. \begin{thm}\label{thm:gauss} For $x\ge1$, $\mathrm{C}(x) = \pi x + \bigO(\sqrt{x})$. \end{thm} This result appears widely in the literature. See, for instance, \cite[Theorem 41]{Rademacher77} or \cite[Chapter 2, Section 7]{Grosswald1985}. We can now compute an asymptotic formula for $\Rseq_n$. \begin{prop}\label{prop:rounding_seq_asymp} For $n\in\N$, $\Rseq_n = (\pi/2-1)n+\bigO\left(\sqrt{n}\right)$, and hence $\lim\limits_{n\to\infty}\frac{1}{n}\Rseq_n =\pi/2-1 \approx 0.570796$. \end{prop} \begin{proof} Combining Proposition~\ref{prop:gauss-R_n} and Theorem~\ref{thm:gauss}, we find \begin{align*} \Rseq_n &= -n + (1/4)\cdot \left(\mathrm{C}(2n) - 1\right) \\ &= -n + (1/4)\cdot \left(2\pi n - 1 + \bigO(\sqrt{2n})\right) \\ &= -n + (\pi/2) n + \bigO\left(\sqrt{n}\right), \end{align*} which proves the first part. The limit behavior follows. \end{proof} \begin{remark} When we defined $\nint{x}$ in the introduction, we chose to always round up half-integers. Suppose we choose a different convention for rounding half-integers (e.g., always rounding down, or always rounding to the nearest even integer, or something else). Call the new rounding function $\nintp{x}$ and consider the resulting sequence \[ \Rseq_n' =\#\left\{1\le k\le n : \nintp{n/k}\text{ is odd}\right\}. \] To compute $\left|\Rseq_n'-\Rseq_n\right|$, we need only consider those $k$ for which $n/k$ is a half-integer. But $n/k$ is a half-integer when $n/k=l/2$ for odd $l$, which means $k$ is a divisor of $2n$. By Lemma~\ref{lem:num_divisors_of_n}, $2n$ has at most $2\sqrt{2n}$ divisors. Thus, there are at most $2\sqrt{2n}$ such $k$, and so \[ |\Rseq_n'-\Rseq_n| \le 2\sqrt{2n}, \] from which we conclude $\Rseq_n'=\Rseq_n+\bigO\left(\sqrt{n}\right)=n(\pi/2-1)+\bigO\left(\sqrt{n}\right)$. \end{remark} \section{Counting by congruence class with shifted floors}\label{sec:generalized} We can now generalize our results from the preceding sections in three ways. First, we note that we can write $\nint{x}$ in terms of the floor function, by $\nint{x}=\Floor{x+1/2}$. We may therefore think of $\nint{x}$ as a \emph{$1/2$-shifted floor function} and the corresponding sequence \[ \nint{\frac{n}{1}}, \nint{\frac{n}{2}}, \dots, \nint{\frac{n}{n}} = \Floor{\frac{n}{1}+\frac{1}{2}}, \Floor{\frac{n}{2}+\frac{1}{2}}, \dots, \Floor{\frac{n}{n}+\frac{1}{2}} \] as a \emph{$1/2$-shifted floor sequence of length $n$}. We will consider an arbitrary shift $\alpha\in\mathbb{R}$. Second, we now have sequences of length $n$ where the $k$th term (for $k\in\{1,2,\dots,n\}$) is $\Floor{n/k + \alpha}$. We can offset the numerator $n$ by some real number $\nu$, resulting in a general term of the form $\Floor{(n-\nu)/k+\alpha}$. Third, in our earlier work we focused on the number of odd terms in each sequence, which means counting the number of terms in each sequence which are congruent to 1 modulo 2. We will instead count the number of terms in each sequence that are congruent to $r$ modulo $m$ for integers $r$ and $m$ with $m\ge1$. Let's set some notation. For $\alpha,\nu\in\mathbb{R}$, the \emph{$\alpha$-shifted, $\nu$-offset, floor sequence of length $n$} is \begin{equation}\label{seq:alpha-shift} \Floor{\frac{n-\nu}{1}+\alpha}, \Floor{\frac{n-\nu}{2}+\alpha}, \dots, \Floor{\frac{n-\nu}{n}+\alpha}. \end{equation} For $r\in\Z$ and $m\in\N$, let $\Num_{n,\alpha,\nu,r,m}$ equal the number of integers in Sequence~\eqref{seq:alpha-shift} that are congruent to $r$ modulo $m$. In other words, let \[ \Num_{n,\alpha,\nu,r,m} =\#\left\{1\le k\le n : \Floor{\frac{n-\nu}{k}+\alpha}\equiv r\pmod*{m} \right\}. \] Connecting to our earlier work, we see that for $n\in\N$, $\Fseq_n=\Num_{n,0,0,1,2}$ and $\Rseq_n=\Num_{n,1/2,0,1,2}$. Since every integer is in exactly one congruence class modulo $m$, we immediately see that \begin{equation}\label{eqn:sum-is-n} \sum\limits_{r=1}^m \Num_{n,\alpha,\nu,r,m} =n \end{equation} for all $\alpha$, $\nu$, and $m$. If we let $m=1$, we find $\Num_{n,\alpha,\nu,r,1}=n$ for all $\alpha$, $\nu$, and $r$. In what follows, we will assume $m\ge2$. Furthermore, noting that $\Num_{n,\alpha,\nu,r,m} = \Num_{n,\alpha-1,\nu,r-1,m,}$, we may suppose $\alpha\in[0,1)$. Next, since the difference $\Num_{n,\alpha,\nu,r,m}-\Num_{n-1,\alpha,\nu-1,r,m,}$ is 1 or 0 depending on whether $\Floor{(n-\nu)/n+\alpha}$ is congruent to $r$ modulo $m$ or not, we may suppose $\nu\in[0,1)$. Finally, it will be useful to take $r\in[1,m]$, the least positive integer in a given congruence class modulo $m$. \subsection{A sum of differences of floors} Our first task is to write $\Num_{n,\alpha,\nu,r,m}$ as a sum of differences of floors. We will generalize the approaches taken in writing $\Fseq_n$ (Proposition~\ref{prop:floor_seq}) and $\Rseq_n$ (Proposition~\ref{prop:rounding_seq}) as summations involving differences of floors. The following lemma generalizes Lemma~\ref{lem:consec_flr} and Lemma~\ref{lem:round_count}. \begin{lem}\label{lem:consec-general} For $\alpha,\nu\in[0,1)$, $k\in\N$, and $n\in\N$ with $n\alpha\ge\nu$, let $g(k)$ equal the number of integers in an $\alpha$-shifted, $\nu$-offset, floor sequence of length $n$ (Sequence~\eqref{seq:alpha-shift}) that are greater than or equal to $k$. Then \[ g(k) = \begin{dcases*} n & if $k=1$,\\ \Floor{(n-\nu)/(k-\alpha)} & if $k\ge2$. \end{dcases*} \] \end{lem} \begin{proof} To start, note that $\nu<1\le n$ for all $n$ Thus $n-\nu>0$, which implies $(n-\nu)/1+\alpha, (n-\nu)/2+\alpha, \dots, (n-\nu)/n+\alpha$ is a decreasing sequence. Looking at the final term, since $\alpha\ge\nu/n$, we have \[ \frac{n-\nu}{n}+\alpha \ge \frac{n-\nu}{n}+\frac{\nu}{n} =1. \] Thus, every term in this sequence is at least 1, and hence their floors are at least 1. This means $g(1)=n$. Now, suppose $k\ge2$ and let $t=(n-\nu)/(k-\alpha)$. Observe that $0<t<n-\nu\le n$. We will show that the first $\Floor{t}$ terms of Sequence~ \eqref{seq:alpha-shift} are at least $k$ and that the remaining terms are less than $k$. Since Sequence~\eqref{seq:alpha-shift} is a non-increasing sequence, it will suffice to show that \begin{equation}\label{eq:ineq-chain} \Floor{\frac{n-\nu}{t}+\alpha}\ge k > \Floor{\frac{n-\nu}{t+1}+\alpha}. \end{equation} For the first inequality in Equation~\eqref{eq:ineq-chain}, observe that \[ \frac{n-\nu}{t}+\alpha = \frac{n-\nu}{(n-\nu)/(k-\alpha)}+\alpha = k. \] Since $k$ is an integer, we have $\Floor{(n-\nu)/t+\alpha}=\Floor{k}=k\ge k$, as desired. For the second inequality Equation~\eqref{eq:ineq-chain}, we have \[ \Floor{\frac{n-\nu}{t+1}+\alpha} \le \frac{n-\nu}{t+1}+\alpha < \frac{n-\nu}{t}+\alpha = k. \] This completes the proof. \end{proof} Note that in the above proof, we considered the cases $k=1$ and $k\ge2$ separately. Our argument for $k\ge2$ doesn't work for $k=1$ because, for $n\alpha>\nu$, our value of $t$ would be $t=(n-\nu)/(1-\alpha)>n$, and we cannot have more than $n$ terms in a sequence of $n$ terms. This explains why we had to get rid of ``extra'' terms in the summation formula for $\Rseq_n$ (Proposition~\ref{prop:rounding_seq}), which involves $k=1$, $\alpha=1/2$, and $\nu=0$, whereas we had no such adjustment in the summation formula for $\Fseq_n$ (Proposition~\ref{prop:floor_seq}), which has $\alpha=\nu=0$. In what follows, we will count the number of terms in Sequence~\eqref{seq:alpha-shift} that are congruent to $r$ modulo $m$. Since we have slightly different results for $k=1$ and for $k\ge2$ in Lemma~\ref{lem:consec-general}, we will have slightly different results for congruence classes with $r=1$ and with $2\le r\le m$ in the proposition (and subsequent results) below. \begin{prop}\label{prop:generalized-sum-of-floors} Suppose $m\ge2$ and $\alpha,\nu\in[0,1)$. Then for all $n\in\N$ with $n\alpha\ge\nu$, \[ \Num_{n,\alpha,\nu,1,m} = n-\Floor{\frac{(n-\nu)}{1-\alpha}}+\sum\limits_{i\ge0} \left(\Floor{\frac{(n-\nu)}{1+im-\alpha}} - \Floor{\frac{(n-\nu)}{2+im-\alpha}}\right) \] and, for $2\le r\le m$, \[ \Num_{n,\alpha,\nu,r,m} = \sum\limits_{i\ge0} \left(\Floor{\frac{(n-\nu)}{r+im-\alpha}} - \Floor{\frac{(n-\nu)}{r+1+im-\alpha}}\right). \] \end{prop} \begin{proof} By Lemma~\ref{lem:consec-general}, we have a formula for $g(d)$, the number of integers in Sequence~\eqref{seq:alpha-shift} that are greater than or equal to a given value $d$. Now, let $G(d)$ equal the number of integers in Sequence~\eqref{seq:alpha-shift} that are equal to a given value $d$. Then $G(d) = g(d)-g(d+1)$. We first compute $G(1)$ by $G(1) = g(1)-g(2) = n - \Floor{(n-\nu)/2-\alpha}$. Then, for $d\ge2$, \[ G(d) = g(d)-g(d+1) = \Floor{(n-\nu)/(d-\alpha)}-\Floor{(n-\nu)/(d+1-\alpha)}. \] To compute $\Num_{n,\alpha,\nu,r,m}$, we need to count the number of integers in Sequence~\eqref{seq:alpha-shift} that are congruent to $r$ modulo $m$. Since $1\le r\le m$, we have $\Num_{n,\alpha,\nu,r,m} = G(r) + G(r+m) + G(r+2m) + \dots$. For $2\le r\le m$, we have \[ \Num_{n,\alpha,\nu,r,m} = \sum\limits_{i\ge0}G(r+im) = \sum\limits_{i\ge0}\left(\Floor{\frac{(n-\nu)}{r+im-\alpha}} - \Floor{\frac{(n-\nu)}{r+1+im-\alpha}}\right). \] For $r=1$, we have \[ \Num_{n,\alpha,\nu,1,m} = \sum\limits_{i\ge0}G(1+im) = n - \Floor{\frac{(n-\nu)}{2-\alpha}} + \sum\limits_{i\ge1}\left(\Floor{\frac{(n-\nu)}{1+im-\alpha}} - \Floor{\frac{(n-\nu)}{2+im-\alpha}}\right). \] If we include the $i=0$ term in the summation, to make it more closely resemble the formula for $r\ne 1$, we get the stated result. \end{proof} \subsection{An asymptotic formula via summation} Our next task is to evaluate the sum \[ \sum\limits_{i\ge0}\left(\Floor{\frac{(n-\nu)}{r+im-\alpha}}-\Floor{\frac{(n-\nu)}{r+1+im-\alpha}}\right) \] that appears in Proposition~\ref{prop:generalized-sum-of-floors}. We'll start by simplifying $\frac{1}{n-\nu}$ times this sum. For $x\in\mathbb{R}$, let $\{x\}$ denote the fractional part of $x$. (I.e., let $\{x\}=x-\Floor{x}$.) Then \[ \frac{1}{n-\nu}\sum\limits_{i\ge0}\left(\Floor{\frac{n-\nu}{r+im-\alpha}}-\Floor{\frac{n-\nu}{r+1+im-\alpha}}\right) =A_{\alpha,r,m}+\frac{1}{n-\nu}B_{n,\alpha,\nu,r,m} \] for the quantities \[ A_{\alpha,r,m}=\sum\limits_{i\ge0}\left(\frac{1}{r+im-\alpha}-\frac{1}{r+1+im-\alpha}\right) \] and \[ B_{n,\alpha,\nu,r,m}=\sum\limits_{i\ge0}\left(\fpf{n-\nu}{r+im-\alpha}-\fpf{n-\nu}{r+1+im-\alpha}\right). \] In general, $A_{\alpha,r,m}$ is an alternating series in which the absolute values of the terms decrease to zero. By the Alternating Series Test, $A_{\alpha,r,m}$ converges. $B_{n,\alpha,\nu,r,m}$ is also an alternating series. If its terms, in absolute value, were decreasing, then we would have $B_{n,\alpha,\nu,r,m}=\bigO(1)$ and thus $(1/n) B_{n,\alpha,\nu,r,m}=\bigO(1/n)$. Unfortunately, this isn't the case. If we can get an asymptotic formula for $B_{n,\alpha,\nu,r,m}$, then we will have an asymptotic formula for $\Num_{n,\alpha,\nu,r,m}$. To start, we will revisit our results for $\Fseq_n$ and $\Rseq_n$ from earlier in this paper. \begin{example}\label{ex:floor} The floor sequence $\Fseq_n$. For $\alpha=\nu=0$, $r=1$, and $m=2$, we have $\Num_{n,0,0,1,2}=\Fseq_n$ for all $n\in\N$. Then \[ A_{0,1,2} =\frac{1}{1}-\frac{1}{2}+\frac{1}{3}-\frac{1}{4}+\dots =\sum\limits_{k=1}^\infty (-1)^{k+1}\frac{1}{k} \] and \[ B_{n,0,0,1,2} =\fpf{n}{1}-\fpf{n}{2}+\fpf{n}{3}-\fpf{n}{4}+\dots =\sum\limits_{k=1}^\infty (-1)^{k+1}\fpf{n}{k}. \] We see that $A_{0,1,2}=\log2$. (This is the Maclaurin series for $\ln(1+x)$, which converges for $-1<x\le 1$, evaluated at $x=1$.) By Proposition~\ref{prop:floor_seq_asymp}, $\Num_{n,0,0,1,2}=\Fseq_n=n\log2+\bigO\left(\sqrt{n}\right)$. By Proposition~\ref{prop:generalized-sum-of-floors}, $B_{n,0,0,1,2}=\bigO\left(\sqrt{n}\right)$. \end{example} \begin{example}\label{ex:rounding} The rounding sequence $\Rseq_n$. For $\alpha=1/2$, $\nu=0$, $r=1$, and $m=2$, we have $\Num_{n,1/2,0,1,2}=\Rseq_n$ for all $n\in\N$. Then \[ A_{1/2,1,2} =\frac{2}{1}-\frac{2}{3}+\frac{2}{5}-\frac{2}{7}+\dots =\sum\limits_{k=1}^\infty(-1)^{k+1}\frac{2}{2k+1} \] and \[ B_{n,1/2,0,1,2} =\fpf{2n}{1}-\fpf{2n}{3}+\fpf{2n}{5}-\fpf{2n}{7}+\dots =\sum\limits_{k=1}^\infty(-1)^{k+1}\fpf{2n}{2k+1}. \] We see that $A_{1/2,1,2}=\pi/2$. (This is the Maclaurin series for $\arctan(x)$, which converges for $-1\le x\le 1$, evaluated at $x=1$.) By Proposition~\ref{prop:rounding_seq_asymp}, $\Num_{n,1/2,0,1,2}=\Rseq_n=(\pi/2-1)n+\bigO\left(\sqrt{n}\right)$. Applying Proposition~\ref{prop:generalized-sum-of-floors}, we find $B_{n,1/2,0,1,2}=\bigO\left(\sqrt{n}\right)$. \end{example} As we will see with the following proposition, we have $B_{n,\alpha,\nu,r,m}=\bigO\left(\sqrt{n}\right)$ in general. For the proof, we will use Dirichlet's hyperbola method. Our method is adapted from the approach given in an answer on Mathematics Stack Exchange \cite{MSE-floor-summation} which proved that, for an increasing sequence of positive integers $b_1,b_2,b_3,\dots$, \[ \sum\limits_{k\le n}\Floor{\frac{n}{b_k}}(-1)^k = n\sum\limits_{k\le n}\frac{1}{b_k}(-1)^k + \bigO\left(\sqrt{n}\right). \] The terms in our series are not quite in this form, so we will prove a slightly more general result. To do so, we need a generalized notion of the term ``divides'' to work with real numbers. \begin{defn}[Real-ly divides] Let $a\in\mathbb{R}$ and $b\in\mathbb{Z}$. We say $a$ \emph{real-ly divides} $b$ if there is some $d\in\Z$ for which $\Ceil{da}=b$. We denote this by $a\reallydivs b$, and we say $a$ is a \emph{real divisor} of $b$. \end{defn} To see that this definition generalizes the usual definition of ``divides,'' let's suppose that we have $a,b\in\Z$ with $a\mid b$. Then there is some $d\in\Z$ for which $da=b$. Hence, $\Ceil{da}=da=b$, and so $a\reallydivs b$. \begin{remark} We will only consider use this definition of \emph{real-ly divides} with positive numbers. However, if one wants to use this with negative numbers as well, it may be beneficial to modify this definition so that it has some symmetry with positive and negative numbers. One could say a real number $a$ real-ly divides an integer $b$ if there is some integer $d$ for which one of the following holds: either $b\ge0$ and $\Ceil{da}=b$; or $b<0$ and $\Floor{da}=b$. With this, one would additionally have $a\reallydivs b$ if and only if $(-a)\reallydivs b$. \end{remark} The key property that we need is the following lemma. \begin{lem}\label{lem:really-divides-summation} Let $n\in\N$, $\nu\in[0,1)$, and $a\in\mathbb{R}$ with $a\ge1$. Then \[ \Floor{\frac{n-\nu}{a}} = \sum\limits_{\substack{1\le d\le n-\nu \\ a\reallydivs d}} 1. \] \end{lem} \begin{proof} To start, we have $n-\nu>0$ and \[ \Floor{(n-\nu)/a} = \#\left\{ka : k\in\Z,\, k\ge 1,\, ka\le n-\nu\right\}. \] Since $a\ge1$, $\Ceil{k_1a}\ne\Ceil{k_2a}$ for all integers $k_1\ne k_2$. Thus, \[ \Floor{(n-\nu)/a} = \#\left\{\Ceil{ka} : k\in\Z,\, k\ge 1,\, ka\le n-\nu\right\}. \] But this is just the cardinality of the set of numbers that $a$ real-ly divides. We therefore have \[ \Floor{(n-\nu)/a} = \#\left\{d\in\Z : 1\le d\le n-\nu,\, a\reallydivs d\right\} = \sum\limits_{\substack{1\le d\le n-\nu \\ a\reallydivs d}} 1, \] as desired. \end{proof} We can now apply Dirichlet's hyperbola method to show $B_{n,\alpha,\nu,r,m}=\bigO\left(\sqrt{n}\right)$. \begin{prop}\label{prop:sum-of-floors-to-sum-of-fractions} For any increasing sequence $b_0,b_1,b_2,\dots$ of positive real numbers with the property that $b_k\ge1$ and $b_k\ge k$ for all $k$, we have \[ \sum\limits_{k\le n-\nu}\Floor{\frac{n-\nu}{b_k}}(-1)^k = n\sum\limits_{k\le n-\nu}\dfrac{(-1)^k}{b_k}+\bigO\left(\sqrt{n}\right). \] \end{prop} \begin{proof} Let $f(n)=\displaystyle\sum\limits_{k\le n-\nu} \Floor{\frac{n-\nu}{b_k}}(-1)^k$. To start, by Lemma~\ref{lem:really-divides-summation}, we have \[ f(n) = \sum\limits_{k\le n-\nu}(-1)^k\sum\limits_{\substack{d\le n-\nu \\ b_k\reallydivs d}} 1. \] Changing the order of summation, \[ f(n) = \sum\limits_{d\le n-\nu}\sum\limits_{\substack{b_k\reallydivs d \\ k\le n-\nu}} (-1)^k. \] We are summing over $d\le n-\nu$ and real divisors $b_k$ of $d$. Thus, $b_k\le n-\nu$. As we have assumed that $b_k\ge k$, this implies $k\le n-\nu$. Thus, we need not explicitly state that $k\le n-\nu$. % % We have \[ f(n) = \sum\limits_{d\le n-\nu}\sum\limits_{b_k\reallydivs d} (-1)^k. \] If $b_k\reallydivs d$, then $d-1<b_kd'\le d$. Thus, instead of summing over $d\le n-\nu$, we may sum over $d'$ and $k$ such that $b_kd'\le n-\nu$: \[ f(n) = \sum\limits_{b_kd'\le n-\nu}(-1)^k. \] We will now use Dirichlet's hyperbola method. Consider the region $R$ in the first quadrant of the $xy$-plane that is bounded by the hyperbola $xy=n-\nu$, the line $x=1$, and the line $y=1$. Let $A>0$. We split the region $R$ into 3 subregions: $R_1$, the portion of $R$ which lies above the line $y=(n-\nu)/A$; $R_2$, the portion of $R$ which lies to the right of the line $x=A$; and $R_3$, which is the rectangle $[1,A]\times[1,(n-\nu)/A]$. (See Figure~\ref{fig:hyperbola}.) \begin{figure} \centering \hspace{1in} \includegraphics[scale=.7]{hyperbola.eps} \caption{The regions $R_1$, $R_2$, $R_3$} \label{fig:hyperbola} \end{figure} It follows that for each combination of $b_k$ and $d'$ such that $b_kd'\le n-\nu$, there is a point $(x,y)=(d',b_k)$ in $R$. Hence, this point is in exactly one of $R_1$, $R_2$, $R_3$. We can sum over points $(d',b_k)$ in $R_1\cup R_3$ and $R_2\cup R_3$, and then subtract a summation over $R_3$ because we have double counted. We have \begin{equation}\label{eqn:dirichlet-eq-5} f(n) = \sum\limits_{d'\le A}\sum\limits_{b_k\le (n-\nu)/d'}(-1)^k + \sum\limits_{b_k\le (n-\nu)/A}\sum\limits_{d'\le (n-\nu)/b_k}(-1)^k - \sum\limits_{d'\le A}\sum\limits_{b_k\le (n-\nu)/A} (-1)^k. \end{equation} Since $b_0,b_1,b_2,\dots$ is an increasing sequence, $\sum\limits_{b_k\le x}(-1)^k\in\{0,1\}$. This summation is $\bigO(1)$. The first and third double sums in Equation~\eqref{eqn:dirichlet-eq-5} are $\bigO(A)$. Hence, \[ f(n) = \sum\limits_{b_k\le (n-\nu)/A}\sum\limits_{d'\le (n-\nu)/b_k}(-1)^k + \bigO(A) =\sum\limits_{b_k\le (n-\nu)/A}(-1)^k\Floor{\frac{(n-\nu)}{b_k}}+\bigO(A). \] Then, since $\Floor{(n-\nu)/b_k}=(n-\nu)/b_k + \bigO(1)$, \[ f(n) =(n-\nu) \sum\limits_{b_k\le (n-\nu)/A} \frac{(-1)^k}{b_k} + \bigO\left(A+\frac{(n-\nu)}{A}\right). \] Taking $A=\sqrt{(n-\nu)}$, we have \[ f(n) = (n-\nu)\sum\limits_{b_k\le \sqrt{(n-\nu)}}\frac{(-1)^k}{b_k}+\bigO\left(\sqrt{n-\nu}\right) \] All that remains to do is modify the summation so that we sum over $k\le n-\nu$. Let $K_n=\min\{k : b_k>\sqrt{n-\nu}\}$. Then \[ \left|\sum\limits_{K_n\le k\le n}\dfrac{(-1)^k}{b_k}\right| <\left|\frac{(-1)^{K_n}}{b_{K_n}}\right| =\dfrac{1}{b_{K_n}}<\dfrac{1}{\sqrt{n-\nu}} =\bigO\left(\dfrac{1}{\sqrt{n-\nu}}\right), \] where the first inequality holds because we have a finite alternating series. Thus, \begin{align*} \sum\limits_{b_k\le\sqrt{n-\nu}}\dfrac{(-1)^k}{b_k} &= \sum\limits_{k< K_n}\dfrac{(-1)^k}{b_k} \\ &= \sum\limits_{k\le n-\nu}\dfrac{(-1)^k}{b_k} - \sum\limits_{K_n\le k\le n-\nu}\dfrac{(-1)^k}{b_k} \\ &= \sum\limits_{k\le n-\nu}\dfrac{(-1)^k}{b_k}+\bigO\left(1/\sqrt{n-\nu}\right). \end{align*} Therefore, \begin{align*} f(n) &= (n-\nu)\sum\limits_{b_k\le \sqrt{n-\nu}}\frac{(-1)^k}{b_k}+\bigO\left(\sqrt{n-\nu}\right) \\ &= (n-\nu)\sum\limits_{k\le n-\nu}\dfrac{(-1)^k}{b_k}+\bigO\left(\sqrt{n-\nu}\right) \\ &= (n-\nu)\sum\limits_{k\le n-\nu}\dfrac{(-1)^k}{b_k}+\bigO\left(\sqrt{n}\right). \end{align*} Since $\nu$ times the convergent alternating sum is $\bigO(1)$, we get the stated result. \end{proof} Combining Proposition~\ref{prop:generalized-sum-of-floors} and Proposition~\ref{prop:sum-of-floors-to-sum-of-fractions}, we obtain an asymptotic formula for $\Num_{n,\alpha,\nu,r,m}$. We'll first give the result for $2\le r\le m$, followed by a slight modification to get the result for $r=1$. (In the proof of Proposition~\ref{prop:generalized-sum-of-floors}, we saw that counting with $r=1$ is slightly different than counting with $r\ne1$. Fortunately, via Equation~\eqref{eqn:sum-is-n}, if we can count for $r=2,\dots,m$, then we get a count for $r=1$ for free.) \begin{cor}\label{cor:generalized-sum-of-floors} For $2\le r\le m$, and $\alpha,\nu\in[0,1)$, and $n\in\N$ with $n\alpha\ge\nu$, \[ \Num_{n,\alpha,\nu,r,m} = n\sum\limits_{i\ge0}\left(\frac{1}{r+im-\alpha}-\frac{1}{r+1+im-\alpha}\right) + \bigO\left(\sqrt{n}\right). \] \end{cor} \begin{proof} For integers $r,m$ with $2\le r\le m$ and for $\alpha\in[0,1)$, define the sequence $b_0,b_1,b_2,\dots$ as follows. For $i\ge0$, let $b_{2i}=(r+im)-\alpha$ and $b_{2i+1}=(r+im)-\alpha+1$. Then, for $\nu\in[0,1)$, \begin{equation}\label{eqn:summation-floor-difference} \sum\limits_{k\le n-\nu}\Floor{\frac{n-\nu}{b_k}}(-1)^k = \sum\limits_{i\ge0}\left(\Floor{\frac{n-\nu}{r+im-\alpha}}-\Floor{\frac{n-\nu}{r+1+im-\alpha}}\right). \end{equation} (While one series is finite and the other is infinite, note that the terms in the infinite series are zero for all $i>(n-\nu-r+\alpha)/m$.) We wish to show the sequence $b_0,b_1,b_2,\dots$ satisfies the conditions of Proposition~\ref{prop:sum-of-floors-to-sum-of-fractions}. We will show $b_k\ge 1$, $b_k\ge k$, and $b_{k+1}>b_k$ for all $k\ge0$. If $k$ is even, then $k=2i$ for some $i$ and we have $b_k=(r+km/2)-\alpha$. Since $m\ge2$, we have $b_k\ge k + r-\alpha \ge k$. If $k$ is odd, then $k=2i+1$ for some $i$ and we have $b_k=r+(k-1)m/2-\alpha+1$. Since $m\ge2$, we have $b_k\ge (k-1)+r-\alpha+1\ge k$. Thus, $b_k\ge k$ for all $k\ge0$. Additionally, since $b_0=r-\alpha\ge2-\alpha>1$ and $b_k\ge k\ge1$ for all $k\ge1$, we have $b_k\ge1$ for all $k\ge0$. Finally, for $k$ even, $b_{k+1}-b_k=1$, and for $k$ odd, $b_{k+1}-b_k=m-1\ge1$. Hence, $b_{k+1}>b_k$ for all $k\ge0$. Since the sequence $b_0,b_1,b_2,\dots$ satisfies the conditions of Proposition~\ref{prop:sum-of-floors-to-sum-of-fractions}, we conclude that \[ \sum\limits_{k\le n-\nu}\Floor{\frac{n-\nu}{b_k}}(-1)^k = n\sum\limits_{k\le n-\nu}\frac{(-1)^k}{b_k} + \bigO\left(\sqrt{n}\right). \] Next, since we have an alternating series and $b_k\ge k$ for all $k$, we have \[ \left|\sum\limits_{k>n-\nu}\frac{(-1)^k}{b_k} \right| <\left|\frac{(-1)^{n}}{b_{n}}\right| =\frac{1}{b_{n}} <\frac{1}{n} =\bigO\left(\frac{1}{n}\right). \] Thus, \begin{align*} \sum\limits_{k\le n-\nu}\Floor{\frac{n-\nu}{b_k}}(-1)^k &= n\sum\limits_{k\ge0}\frac{(-1)^k}{b_k}-n\sum\limits_{k>n-\nu}\frac{(-1)^k}{b_k}+\bigO\left(\sqrt{n}\right) \\ &= n\sum\limits_{k\ge0}\frac{(-1)^k}{b_k} -n\bigO\left(\frac{1}{n}\right)+\bigO\left(\sqrt{n}\right) \\ &= n\sum\limits_{k\ge0}\frac{(-1)^k}{b_k}+\bigO(1)+ \bigO\left(\sqrt{n}\right). \end{align*} Combined with Equation~\eqref{eqn:summation-floor-difference}, we conclude \[ \sum\limits_{i\ge0}\left(\Floor{\frac{n-\nu}{r+im-\alpha}}-\Floor{\frac{n-\nu}{r+1+im-\alpha}}\right) = n\sum\limits_{i\ge0}\left(\frac{1}{r+im-\alpha}-\frac{1}{r+1+im-\alpha}\right)+\bigO\left(\sqrt{n}\right). \] Together with Proposition~\ref{prop:generalized-sum-of-floors}, this proves the result for $2\le r\le m$. \end{proof} We now use Equation~\eqref{eqn:sum-is-n} and Corollary~\ref{cor:generalized-sum-of-floors} to compute $\Num_{n,\alpha,\nu,r,m}$ for $r=1$. \begin{cor}\label{cor:generalized-sum-of-floors-r=1} For any $m\ge2$, $\alpha,\nu\in[0,1)$, and $n\in\N$ with $n\alpha\ge\nu$, \[ \Num_{n,\alpha,\nu,1,m} = \frac{-\alpha n}{1-\alpha}+n\sum\limits_{i\ge0}\left(\frac{1}{1+im-\alpha}-\frac{1}{2+im-\alpha}\right) + \bigO\left(\sqrt{n}\right). \] \end{cor} \begin{proof} By Equation~\eqref{eqn:sum-is-n}, \[ \Num_{n,\alpha,\nu,1,m} = n-\sum\limits_{r=2}^m \Num_{n,\alpha,\nu,r,m}. \] Then, by Corollary~\ref{cor:generalized-sum-of-floors} \begin{align*} \Num_{n,\alpha,\nu,1,m} & = n - \sum\limits_{r=2}^m \left[n\sum\limits_{i\ge0}\left(\frac{1}{r+im-\alpha}-\frac{1}{r+1+im-\alpha}\right)+\bigO\left(\sqrt{n}\right)\right] \\ & = n - \sum\limits_{r=2}^m n\sum\limits_{i\ge0}\left(\frac{1}{r+im-\alpha}-\frac{1}{r+1+im-\alpha}\right)+\bigO\left(\sqrt{n}\right) \\ & = n - n\sum\limits_{i\ge0}\sum\limits_{r=2}^m\left(\frac{1}{r+im-\alpha}-\frac{1}{r+1+im-\alpha}\right) +\bigO\left(\sqrt{n}\right), \end{align*} where we have changed the order of summation in the last step because we have a finite number of convergent alternating series. We have a telescoping series for each $i\ge0$ which we can manipulate by \begin{align*} \Num_{n,\alpha,\nu,1,m} & = n - n\sum\limits_{i\ge0} \left(\frac{1}{2+im-\alpha}-\frac{1}{m+1+im-\alpha}\right)+\bigO\left(\sqrt{n}\right) \\ & = n -\frac{n}{1-\alpha}+\frac{n}{1-\alpha}- n\sum\limits_{i\ge0} \left(\frac{1}{2+im-\alpha}-\frac{1}{m+1+im-\alpha}\right)+\bigO\left(\sqrt{n}\right) \\ &= n - \frac{n}{1-\alpha} + n\sum\limits_{i\ge0} \left(\frac{1}{1+im-\alpha}-\frac{1}{2+im-\alpha}\right)+\bigO\left(\sqrt{n}\right). \end{align*} Note that we have merely inserted $-n/(1-\alpha)+n/(1-\alpha)$ into our expression, and that we have not changed the order of summation in doing so. Hence, \[ \Num_{n,\alpha,\nu,1,m} = \frac{-\alpha n}{1-\alpha} + n\sum\limits_{i\ge0} \left(\frac{1}{1+im-\alpha}-\frac{1}{2+im-\alpha}\right)+\bigO\left(\sqrt{n}\right), \] as desired. \end{proof} \subsection{An asymptotic formula via integration} From integral calculus, we know how to evaluate sums like $1-1/2+1/3-1/4+\dots$ and $1-1/3+1/5-1/7+\dots$ via integration of Maclaurin series as mentioned in Example~\ref{ex:floor} and Example~\ref{ex:rounding}. We will take the same approach to evaluate the summation that appears in Corollary~\ref{cor:generalized-sum-of-floors} and Corollary~\ref{cor:generalized-sum-of-floors-r=1}. The following proposition shows us how. As with previous work in this section, we will first obtain a result $2\le r\le m$ and then use Equation~\eqref{eqn:sum-is-n} to obtain a result for $r=1$. \begin{prop}\label{prop:alt-sum-integral-formula} For integers $r,m$ with $2\le r\le m$, and for any $\alpha\in [0,1)$, \[ \sum\limits_{i\ge0}\left(\frac{1}{r+im-\alpha}-\frac{1}{r+1+im-\alpha}\right) =\int\limits_0^1\dfrac{(1-x)x^{r-1-\alpha}}{1-x^m}\mathrm{d}x. \] \end{prop} \begin{proof} To start, for $\beta=r-1-\alpha$, a positive real number, and for any non-negative integer $k$, let \[ f_k(x) =x^{\beta}\sum\limits_{i=0}^k\left(x^{im}-x^{im+1}\right). \] Then, for all $k\ge0$ we have \[ |f_k(x)| =\left|x^\beta (1-x)\sum\limits_{i=0}^k x^{im} \right| = \left| x^\beta (1-x) \frac{1-x^{(k+1)m}}{1-x^m} \right| = \left| \frac{x^\beta}{1+x+x^2+\dots+x^{m-1}} \right| \left|\left(1-x^{(k+1)m}\right)\right|, \] \iffalse \begin{align*} |f_k(x)| &=\left|x^\beta (1-x)\sum\limits_{i=0}^k x^{im} \right| ^\\ &= \left| x^\beta (1-x) \frac{1-x^{(k+1)m}}{1-x^m} \right| \\ &= \left| x^\beta (1-x^{(k+1)m}) \frac{1}{(1-x^m)/(1-x)}\right| \\ &= \left| x^\beta \right| \left|(1-x^{(k+1)m})\right| \left| \frac{1}{1+x+x^2+\dots+x^{m-1}}\right|, \end{align*} \fi a product of absolute values of two functions. Each absolute value is at most 1 for all $x\in[0,1]$. (We're using the fact that $\beta>0$ here.) Thus, $|f_k(x)|\le 1\cdot1=1$ for all $x\in[0,1]$. Since 1 is integrable on $[0,1]$, by the Dominated Convergence Theorem we have \[ \lim\limits_{k\to\infty}\int\limits_0^1 f_k(x)\,\mathrm{d}x = \int\limits_0^1 \lim\limits_{k\to\infty} f_k(x)\,\mathrm{d}x. \] Starting with the integral in the statement of this proposition, and applying this limit and integration interchange, we find \begin{align*} \int\limits_0^1\dfrac{(1-x)x^{r-1-\alpha}}{1-x^m}\mathrm{d}x &= \int\limits_0^1 \lim\limits_{k\to\infty} f_k(x)\mathrm{d}x = \lim\limits_{k\to\infty}\int\limits_0^1 f_k(x)\mathrm{d}x = \lim\limits_{k\to\infty} \int\limits_0^1 x^{\beta}\sum\limits_{i=0}^k \left(x^{im}-x^{im+1}\right)\mathrm{d}x \\ &= \lim\limits_{k\to\infty}\sum\limits_{i=0}^k \left(\frac{x^{\beta+1+im}}{\beta+1+im} - \frac{x^{\beta+2+im}}{\beta+2+im}\right)\Bigg|_{x=0}^{x=1} \\ & = \lim\limits_{k\to\infty}\sum\limits_{i=0}^k\left(\frac{1}{\beta+1+im}-\frac{1}{\beta+2+im}\right) \\ & = \sum\limits_{i=0}^\infty\left(\frac{1}{r+im-\alpha}-\frac{1}{r+1+im-\alpha}\right), \end{align*} as desired. \end{proof} Combining Corollary~\ref{cor:generalized-sum-of-floors} and Proposition~\ref{prop:alt-sum-integral-formula}, we obtain the following asymptotic formula for $\Num_{n,\alpha,\nu,r,m}$ with $2\le r\le m$, which we can extend to $r=1$ via Equation~\eqref{eqn:sum-is-n}. This is our main result. \begin{thm}\label{thm:N_m-asymp} Suppose $m\ge1$, $\alpha,\nu\in[0,1)$, and $n\in\N$ with $n\alpha\ge\nu$. Then \[ \Num_{n,\alpha,\nu,1,m} = \frac{-\alpha n}{1-\alpha}+n\int\limits_0^1\dfrac{(1-x)x^{-\alpha}}{1-x^m}\mathrm{d}x + \bigO\left(\sqrt{n}\right), \] and, for $2\le r\le m$, \[ \Num_{n,\alpha,\nu,r,m} = n\int\limits_0^1\dfrac{(1-x)x^{r-1-\alpha}}{1-x^m}\mathrm{d}x + \bigO\left(\sqrt{n}\right). \] \end{thm} \begin{proof} For $2\le r\le m$, the result follows from Corollary~\ref{cor:generalized-sum-of-floors} and Proposition~\ref{prop:alt-sum-integral-formula}. We need to prove the result for $1\le r\le m$, and $\alpha,\nu\in[0,1)$. By Equation~\eqref{eqn:sum-is-n}, \[\Num_{n,\alpha,\nu,1,m} = n - \sum\limits_{r=2}^m\Num_{n,\alpha,\nu,r,m}.\] Thus, \begin{align*} \Num_{n,\alpha,\nu,1,m} = n - \sum\limits_{r=2}^m\Num_{n,\alpha,\nu,r,m} &= n - \sum\limits_{r=2}^m \left(n\int\limits_0^1\dfrac{(1-x)x^{r-1-\alpha}}{1-x^m}\mathrm{d}x+\bigO\left(\sqrt{n}\right)\right) \\ &= n - n\int\limits_0^1\dfrac{x^{-\alpha}(1-x)}{1-x^m}\sum\limits_{r=2}^m x^{r-1}\mathrm{d}x+\bigO\left(\sqrt{n}\right) \\ &= n - n\int\limits_0^1\dfrac{x^{-\alpha}(1-x)}{1-x^m}\sum\limits_{r=2}^m x^{r-1}\mathrm{d}x+\bigO\left(\sqrt{n}\right). \end{align*} The finite series inside the integral will cancel with the denominator nicely if we include one more term. We do so as follows: \begin{align*} \Num_{n,\alpha,\nu,1,m} - n\int\limits_0^1 \dfrac{(1-x)x^{-\alpha}}{1-x^m}\mathrm{d}x &= n - n\int\limits_0^1\dfrac{x^{-\alpha}(1-x)}{1-x^m}\sum\limits_{r=1}^m x^{r-1}\mathrm{d}x+\bigO\left(\sqrt{n}\right) \\ &=n-n\int\limits_0^1 x^{-\alpha}\mathrm{d}x +\bigO\left(\sqrt{n}\right). \end{align*} Since $0\le\alpha<1$, this improper integral converges to $1/(1-\alpha)$. Solving for $\Num_{n,\alpha,\nu,1,m}$, we find \[ \Num_{n,\alpha,\nu,1,m} = n - \frac{n}{1-\alpha} + n\int\limits_0^1\dfrac{(1-x)x^{-\alpha}}{1-x^m}\mathrm{d}x + \bigO\left(\sqrt{n}\right), \] which simplifies to the stated result. \end{proof} Thus, $\Num_{n,\alpha,\nu,r,m}$ is asymptotically linear and our formula is independent of $\nu$. We record the corresponding slope below in the following corollary. \begin{cor}\label{cor:N_m-slope} For $\alpha,\nu\in[0,1)$ and $1\le r\le m$, \[ \lim\limits_{n\to\infty}\frac{1}{n}\Num_{n,\alpha,\nu,r,m} = \begin{dcases*} \frac{-\alpha}{1-\alpha}+\int\limits_0^1\dfrac{(1-x)x^{-\alpha}}{1-x^m}\mathrm{d}x & if $r=1$,\\ \int\limits_0^1\dfrac{(1-x)x^{r-1-\alpha}}{1-x^m}\mathrm{d}x & if $2\le r\le m$. \end{dcases*} \] \end{cor} Via integration, we can compute specific values of $\lim\limits_{n\to\infty}\frac{1}{n}\Num_{n,\alpha,\nu,r,m}$ for $1\le r\le m\le 4$. For $\alpha=\nu=0$, exact and rounded values are in Figure~\ref{fig:shift-0-exact-values} and Figure~\ref{fig:shift-0-decimals}. For $\alpha=1/2$ and $\nu=0$, exact and rounded values are in Figure~\ref{fig:shift-1/2-exact-values} and Figure~\ref{fig:shift-1/2-decimals}. \begin{figure} \centering \begin{tabular}{c||c|c|c|c|} \backslashbox{$r$}{$m$} & 1 & 2 & 3 & 4 \\ \hline \hline 1 & $1$ & $\log\left(2\right)$ & $\frac{1}{9} \, \sqrt{3} \pi$ & $\frac{1}{8} \, \pi + \frac{1}{4} \, \log\left(2\right)$ \\ \hline 2 & & $-\log\left(2\right) + 1$ & $-\frac{1}{18} \, \sqrt{3} \pi + \frac{1}{2} \, \log\left(3\right)$ & $\frac{1}{8} \, \pi - \frac{1}{4} \, \log\left(2\right)$ \\ \hline 3 & & & $-\frac{1}{18} \, \sqrt{3} \pi - \frac{1}{2} \, \log\left(3\right) + 1$ & $-\frac{1}{8} \, \pi + \frac{3}{4} \, \log\left(2\right)$ \\ \hline 4 & & & & $-\frac{1}{8} \, \pi - \frac{3}{4} \, \log\left(2\right) + 1$ \\ \hline \end{tabular} \caption{Values of $\lim\limits_{n\to\infty}\frac{1}{n}\Num_{n,0,0,r,m}$ for $1\le m\le 4$} \label{fig:shift-0-exact-values} \end{figure} \begin{figure} \centering \begin{tabular}{c||c|c|c|c|} \backslashbox{$r$}{$m$} & 1 & 2 & 3 & 4 \\ \hline \hline 1 & $1.000000$ & $0.693147$ & $0.604600$ & $0.565986$ \\ \hline 2 & & $0.306853$ & $0.247006$ & $0.219412$ \\ \hline 3 & & & $0.148394$ & $0.127161$ \\ \hline 4 & & & & $0.087441$ \\ \hline \end{tabular} \caption{$\lim\limits_{n\to\infty}\frac{1}{n}\Num_{n,0,0,r,m}$ for $1\le m\le 4$, rounded to 6 decimal places} \label{fig:shift-0-decimals} \end{figure} \begin{figure} \centering \begin{tabular}{c||c|c|c|c|} \backslashbox{$r$}{$m$} & 1 & 2 & 3 & 4 \\ \hline \hline 1 & $1$ & $\frac{1}{2} \, \pi - 1$ & $\frac{1}{6} \, \sqrt{3} \pi + \frac{1}{2} \, \log\left(3\right) - 1$ & $\frac{1}{4} \, \pi + \frac{1}{4} \, \sqrt{2} \log\left(3+2\sqrt{2}\right) - 1$ \\ \hline 2 & & $-\frac{1}{2} \, \pi + 2$ & $\frac{1}{6} \, \sqrt{3} {\left(\pi - \sqrt{3} \log\left(3\right)\right)}$ & $\frac{1}{4} \, \pi {\left(\sqrt{2} - 1\right)}$ \\ \hline 3 & & & $-\frac{1}{3} \, \sqrt{3} \pi + 2$ & $\frac{1}{4} \, \pi - \frac{1}{4} \, \sqrt{2} \log\left(3+2\sqrt{2}\right)$ \\ \hline 4 & & & & $-\frac{1}{4} \, \pi {\left(\sqrt{2} + 1\right)} + 2$ \\ \hline \end{tabular} \caption{$\lim\limits_{n\to\infty}\frac{1}{n}\Num_{n,1/2,0,r,m}$ for $1\le m\le 4$} \label{fig:shift-1/2-exact-values} \end{figure} \begin{figure} \centering \begin{tabular}{c||c|c|c|c|} \backslashbox{$r$}{$m$} & 1 & 2 & 3 & 4 \\ \hline \hline 1 & $1.000000$ & $0.570796$ & $0.456206$ & $0.408623$ \\ \hline 2 & & $0.429204$ & $0.357594$ & $0.325323$ \\ \hline 3 & & & $0.186201$ & $0.162173$ \\ \hline 4 & & & & $0.103881$ \\ \hline \end{tabular} \caption{$\lim\limits_{n\to\infty}\frac{1}{n}\Num_{n,1/2,0,r,m}$ for $1\le m\le 4$, rounded to 6 decimal places} \label{fig:shift-1/2-decimals} \end{figure} \section{Applications to finding parity and counting lattice points}\label{sec:applications} Now that we have a formula for $\Num_{n,\alpha,\nu,r,m}$, we focus on a few applications: computing a floor shift which results in an asymptotic 50/50 split of even and odd terms; and counting lattice points in a few families of ellipses. \subsection{Shifting for parity} We return to the case where $m=2$ and consider the problem of determining a shift $\alpha$ so that half of the terms are odd and half are even. It amounts to computing $\alpha$ for which \[ \lim\limits_{n\to\infty}\frac{1}{n}\Num_{n,\alpha,\nu,1,2} =\lim\limits_{n\to\infty}\frac{1}{n}\Num_{n,\alpha,\nu,2,2} =1/2. \] By Corollary~\ref{cor:N_m-slope} (with the formula for $r=2$ to avoid the extra term out front), we need $\alpha$ such that \[ \int\limits_0^1\dfrac{(1-x)x^{1-\alpha}}{1-x^2}\mathrm{d}x = \int\limits_0^1\dfrac{x^{1-\alpha}}{1+x}\mathrm{d}x = 1/2. \] (Note that this is independent of $\nu$.) Since we're using $r=2$, for the remainder of this subsection, we will count even entries instead of odd entries in an $\alpha$-shifted floor sequence. \begin{figure} \centering \includegraphics[width=.6\textwidth]{f-alpha-plot.eps} \caption{Plot of $\displaystyle y=f(\alpha) = \int\limits_0^1\dfrac{x^{1-\alpha}}{1+x}\mathrm{d}x$ for $\alpha\in[0,1]$. } \label{fig:f(x)-plot} \end{figure} Let \[ f(\alpha) =\int\limits_0^1\dfrac{x^{1-\alpha}}{1+x}\mathrm{d}x. \] Then $f(\alpha)$ is the (asymptotic) proportion of terms in an $\alpha$-shifted floor sequence of length $n$ that are even. We immediately see that $f$ is continuous, increasing, and concave up for $\alpha\in[0,1]$. Furthermore, $f(0)=1-\log2 < 1/2$ and $f(1)=\log2 > 1/2$. Thus, there is a unique shift $\alpha_0\in(0,1)$ for which $f(\alpha_0)=1/2$. A plot of $f$ appears in Figure~\ref{fig:f(x)-plot}. We see that the value of $\alpha$ for which $f(\alpha)=1/2$ is between $0.6$ and $0.7$. Computing with Sage \cite{sage}, we can shrink the interval down. We compute \[ f(0.682379227335) < 1/2 \text{ and } f(0.682379227345) > 1/2. \] Hence, we have $f(\alpha_0)=1/2$ for \[ \alpha_0 \approx 0.68237922734. \] We can look at some data with this approximate value. A plot of $y=\Num_{n,0.68237922734,0,2,2}$ appears in Figure~\ref{fig:alpha_0_sequence_plot} along with the graph of $y=n/2$. In Figure~\ref{fig:alpha_0_random_table} gives values of $\Num_{n,0.68237922734,0,2,2}$ for random $n$ with $10^5<n<10^6$. \begin{figure} \centering \includegraphics[width=.6\textwidth]{alpha_0_sequence.eps} \caption{Plot of $y=\Num_{n,\alpha,0,2,2}$ for $\alpha=0.68237933734$ and $1\le n\le 1000$ (black) along with the graph of $y=n/2$ (blue).} \label{fig:alpha_0_sequence_plot} \end{figure} \begin{figure} \centering \begin{tabular}{|c|c|c|}\hline & & \\ $n$ & number of even terms & proportion of even terms \\ & & \\ \hline $ 7015 $ & $ 3503 $ & $49.9359\%$ \\ $ 179220 $ & $ 89632 $ & $50.0123\%$ \\ $ 213788 $ & $ 106901 $ & $50.0033\%$ \\ $ 267093 $ & $ 133562 $ & $50.0058\%$ \\ $ 439675 $ & $ 219839 $ & $50.0003\%$ \\ $ 491213 $ & $ 245600 $ & $49.9987\%$ \\ $ 503521 $ & $ 251741 $ & $49.9961\%$ \\ $ 689325 $ & $ 344631 $ & $49.9954\%$ \\ $ 775294 $ & $ 387629 $ & $49.9977\%$ \\ $ 978029 $ & $ 489010 $ & $49.9995\%$ \\ \hline \end{tabular} \caption{The number and proportion of even terms in an $\alpha$-shifted, $\nu$-offset, floor sequence of length $n$ for $\alpha=0.68237933634$ and various random $n$ with $10^5<n<10^6$.} \label{fig:alpha_0_random_table} \end{figure} \subsection{Lattice points in selected ellipses, number rings, and plane tilings} As we saw with Proposition~\ref{prop:gauss-R_n}, the number of lattice points in a circle of radius $\sqrt{2n}$ centered at the origin is $4\Rseq_n+4n+1$, where $\Rseq_n=\Num_{n,1/2,0,1,2}$ is the $1/2$-shifted floor sequence of length $n$. If we think of the plane as being tiled with $1\times1$ squares, then we have a formula for the number of vertices contained in a circle of radius $\sqrt{2n}$ centered at any vertex. In this subsection, we will find formulas involving $\Num_{n,\alpha,\nu,r,m}$ for the number of lattice points contained in the ellipses \[ x^2+y^2=n,\quad x^2+xy+y^2=n,\quad\text{and } x^2+2y^2=n, \] for any $n\in\N$. We will also count vertices contained in a circle of radius equal to the square root of any integer for tilings of the plane by squares or triangles. In Proposition~\ref{prop:generalized-sum-of-floors}, we wrote $\Num_{n,\alpha,\nu,1,m}$ with a summation involving a difference of floors. In what follows, it will be useful to have the formula for $\Num_{n,\alpha,\nu,1,m}$ in the case where $\alpha$ is rational. If we suppose $\alpha=p/q$, then \begin{equation} \label{eqn:num-rational-alpha} \Num_{n,p/q,\nu,1,m}=n - \Floor{\frac{(n-\nu)q}{q-p}} + \sum\limits_{i\ge0}\left(\Floor{\frac{(n-\nu)q}{q+qim-p}}-\Floor{\frac{(n-\nu)q}{2q+qim-p}} \right) \end{equation} for all $n\in\N$ with $n\alpha\ge\nu$. Also, for $n\in\N$, recall the function $\divs_{r,m}(n)$, which counts the number of positive divisors of $n$ that are congruent to $r$ modulo $m$. \subsubsection{Lattice points in the region $x^2+y^2\le n$} In Proposition~\ref{prop:gauss-R_n}, we found a formula for the number of lattice points $(x,y)$ contained in the disc $x^2+y^2\le 2n$ in terms of $\Rseq_n$. This is a result for discs with radius equal to the square root of an even number. We'll extend the result to square roots of odd numbers as well, eventually obtaining a formula for the number of lattice points contained in the disc $x^2+y^2\le n$. If we think of the plane as being tiled with $1\times1$ squares, then our formula will give us the total number of vertices that lie in a disc of radius $\sqrt{n}$ centered at one of the vertices. \begin{prop}\label{prop:circle-radius-n} For $n\in\N$, let $F(n)$ be the number of lattice points in a disc of radius $\sqrt{n}$ centered at the origin. Then \[ F(n) = 4\Num_{\Ceil{n/2},1/2,\{n/2\},1,2}+4\Floor{\frac{n}{2}}+1, \] where $\{n/2\}$ denotes the fractional part of $n/2$. \end{prop} \begin{proof} Let $F(n)=\#\left\{(x,y)\in\Z^2 : x^2+y^2\le n\right\}$. We'll show that the formula works for even $n$ and for odd $n$. If $n$ is even, then $n=2k$ for some $k\in\N$. By Proposition~\ref{prop:gauss-R_n} \[ F(n) = F(2k) = 4\Rseq_k+4k+1=4\Num_{k,1/2,0,1,2}+4k+1=4\Num_{n/2,1/2,0,1,2}+2n+1. \] Note that $\Ceil{n/2}=k=n/2$, $\{n/2\}=\{k\}=0$, and $4\Floor{n/2}=4\Floor{k}=4k=2n$. This proves the formula for $F(n)$ with $n$ even. If $n$ is odd, then $n=2k-1$ for some $k\in\N$. By Jacobi's two-square theorem (Theorem~\ref{thm:jacobi}), \[ F(n) = 1 + 4\sum\limits_{j=1}^n\left(\divs_{1,4}(j)-\divs_{3,4}(j)\right) = 1 + 4\left(\Floor{\frac{n}{1}}-\Floor{\frac{n}{3}}+\Floor{\frac{n}{5}}-\Floor{\frac{n}{7}}+\dots\right). \] In order to get this alternating floor sum, we use $\alpha=\nu=1/2$, $r=1$, and $m=2$ with Equation~\eqref{eqn:num-rational-alpha}. We have \[ \Num_{k,1/2,1/2,1,2} =1-k+\sum\limits_{i\ge0}\left(\Floor{\frac{2k-1}{4i+1}}-\Floor{\frac{2k-1}{4i+3}}\right) = \frac{1-n}{2}+\sum\limits_{i\ge0}\left(\Floor{\frac{n}{4i+1}}-\Floor{\frac{n}{4i+3}}\right) \] for all $k\ge\nu/\alpha=1$. Then, \[ F(n) = 1 + 4\left(\Num_{k,1/2,1/2,1,2}+\frac{n-1}{2}\right) = 1 + 4\Num_{(n+1)/2,1/2,1/2,1,2}+2n-2. \] Note that $\Ceil{n/2}=k=(n+1)/2$, $\{n/2\}=\{k+1/2\}=1/2$, and $4\Floor{n/2}=4\Floor{k-1/2}=4(k-1)=2n-2$. This proves the formula for $F(n)$ with $n$ odd. \end{proof} \begin{example}\label{ex:gaussian-ints} To illustrate Proposition~\ref{prop:circle-radius-n}, we'll compute the number of lattice points in a disc of radius $\sqrt{13}$. Note that $\Ceil{13/2}=7$. The number of lattice points in a disc of radius $\sqrt{13}$ involves the quantity $4\Num_{7,1/2,1/2,1,2}$. To compute this, we examine the length-7 sequence \[ \Floor{\frac{7-1/2}{1}+\frac{1}{2}},\Floor{\frac{7-1/2}{2}+\frac{1}{2}},\dots,\Floor{\frac{7-1/2}{7}+\frac{1}{2}} = 7,3,2,2,1,1,1, \] which contains 5 odd terms. Thus, $\Num_{7,1/2,1/2,1,2}=5$. By Proposition~\ref{prop:circle-radius-n}, the number of lattice points in a disc of radius $\sqrt{13}$ is therefore \[ F(13) =4\Num_{7,1/2,1/2,1,2}+4\Floor{\frac{13}{2}}+1 = 4\cdot5+4\cdot6+1=45. \] A disc of radius $\sqrt{13}$ appears in Fig.~\ref{fig:sqrt-13-disc}, and one confirms that it contains 45 lattice points. \end{example} \begin{figure} \centering \includegraphics[scale=.6]{circle-sqrt-13.eps} \caption{The 45 lattice points in a disc of radius $\sqrt{13}$, which are also the 45 Gaussian integers with norm at most 13.} \label{fig:sqrt-13-disc} \end{figure} \begin{remark} We mentioned the ring of Gaussian integers, $\Z[\ii]$, earlier. For any $z=a+b\ii\in\Z[\ii]$, the norm of $z$ is $N(z)=N(a+b\ii)=a^2+b^2$. Thus, Proposition~\ref{prop:circle-radius-n} gives a formula for the number of Gaussian integers with norm at most $n$. Following from Example~\ref{ex:gaussian-ints}, we know there are 45 Gaussian integers with norm at most 13. We visualize these Gaussian integers in Figure~\ref{fig:sqrt-13-disc}. Viewing $\Z[\ii]$ in the complex plane, the Gaussian integers are the vertices for a tiling of the plane by $1\times1$ square tiles. If we draw a circle of radius $\sqrt{n}$, for $n\in\N$, around any lattice point, then Proposition~\ref{prop:circle-radius-n} gives us a formula for the number of vertices contained in the circle. Any other $1\times1$ square tiling of the plane would involve a rotation and/or shift of this tiling. A circle of radius $\sqrt{n}$ centered at any vertex would contain the same number of lattice points. Thus, Proposition~\ref{prop:circle-radius-n} gives us a formula for the number of vertices contained in a circle of radius $\sqrt{n}$, for $n\in\N$, centered at any vertex of any $1\times1$ square tiling of the plane. \end{remark} \subsubsection{Lattice points in the region $x^2+xy+y^2\le n$} Next, we consider the ellipse $x^2+xy+y^2=n$. In general, the quantity $ax^2+bxy+cy^2$, for constants $a,b,c$, is a \emph{binary quadratic form}. We take results about the number of representations of an integer $n$ by a binary quadratic form from \cite{Dickson1958}. \begin{prop}[{\cite[Exercise XXII.2]{Dickson1958}}]\label{prop:dickson-x^2+xy+y^2} Let $n\in\N$. Then the number of representations of $n=x^2+xy+y^2$, for integers $x,y$, is $6\left(\divs_{1,3}(n)-\divs_{2,3}(n) \right)$. \end{prop} \begin{cor}\label{cor:lattice-points-eisenstein-ellipse} For $n\in\N$, let $F(n)$ be the number of lattice points contained in the elliptical region $x^2+xy+y^2=n$. Then \[ F(n) = 6\Num_{n,0,0,1,3}+1. \] \end{cor} \begin{proof} Let $f(n)=\#\left\{(x,y)\in\Z^2 : x^2+xy+y^2=n\right\}$. Observe that $f(0)=1$. Thus, \begin{equation}\label{eqn:f-step-2} F(n) = \sum\limits_{k=0}^n f(k) = 1 + \sum\limits_{k=1}^n f(k). \end{equation} Next, by Proposition~\ref{prop:dickson-x^2+xy+y^2}, \[ \sum\limits_{k=1}^n f(k) = 6\left(\Floor{\frac{n}{1}}-\Floor{\frac{n}{2}}+\Floor{\frac{n}{4}}-\Floor{\frac{n}{5}}+\Floor{\frac{n}{7}}-\dots\right), \] which is a finite sum since the floors are eventually zero. Using $\alpha=\nu=0$, $r=1$, and $m=3$ with Equation~\eqref{eqn:num-rational-alpha}, we get \[ \Num_{n,0,0,1,3} =\left(\Floor{\frac{n}{1}}-\Floor{\frac{n}{2}}+\Floor{\frac{n}{4}}-\Floor{\frac{n}{5}}+\Floor{\frac{n}{7}}-\dots\right) \] for all $n\in\N$. We conclude that $F(n) = 1 + 6\Num_{n,0,0,1,3}$, as desired. \end{proof} \begin{example}\label{ex:ellipse-30} We'll compute the number of lattice points contained in the ellipse $x^2+xy+y^2=30$. To do so, we consider the sequence \[ \Floor{\frac{30}{1}},\Floor{\frac{30}{2}},\dots,\Floor{\frac{30}{30}} = 30, 15, 10, 7, 6, 5, 4, 3, 3, 3, 2, 2, 2, 2, 2, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1. \] There are 18 terms in this sequence that are congruent to 1 modulo 3. Hence, the ellipse $x^2+xy+y^3=30$ contains $6\Num_{30,0,0,1,3}+1=6\cdot18+1=109$ lattice points. See Figure~\ref{fig:ellipse-30}. \end{example} \begin{figure} \centering \includegraphics[scale=.6]{x2pxypy2e30.eps} \includegraphics[scale=.6]{eisensten-30.eps} \caption{The 109 lattice points within the ellipse $x^2+xy+y^2=30$, and the 109 Eisenstein integers with norm at most 30.} \label{fig:ellipse-30} \end{figure} From Theorem~\ref{thm:N_m-asymp} and Figure~\ref{fig:shift-0-exact-values}, we see that $\Num_{n,0,0,1,3}=\frac{1}{9}\sqrt{3}\pi n+\bigO\left(\sqrt{n}\right)$. We immediately obtain the following corollary. \begin{cor}\label{cor:lattice-points-x2+xy+y2} For $n\in\N$, the number of lattice points in the elliptical region $x^2+xy+y^2\le n$ is $\frac{2}{3}\sqrt{3}\pi n + \bigO\left(\sqrt{n}\right)$. \end{cor} \begin{remark} Consider the ring of Eisenstein integers, $\Z[\omega]$, where $\omega=\frac{-1+\sqrt{-3}}{2}$, a primitive 3rd root of unity. For $z=a-b\omega\in\Z[\omega]$, the norm of $z$ is $N(z)=N(a-b\omega)=a^2+ab+b^2$. Thus, Corollary~\ref{cor:lattice-points-eisenstein-ellipse} gives a formula for the number of Eisenstein integers with norm at most $n$. Following from Example~\ref{ex:ellipse-30}, we know there are 109 Eisenstein integers with norm at most 30. We visualize these Eisenstein integers in Figure~\ref{fig:ellipse-30}. Viewing $\Z[\omega]$ in the complex plane, the Eisenstein integers are the vertices for a plane tiling involving equilateral triangles of side length 1. If we draw a circle of radius $\sqrt{n}$ around any vertex, Corollary~\ref{cor:lattice-points-eisenstein-ellipse} gives us a formula for the number of vertices contained in the circle. Rotating the plane (and hence the tiles) about the center of the circle will leave the number of vertices in the circle unchanged. Thus, Corollary~\ref{cor:lattice-points-eisenstein-ellipse} gives us a formula for the number of vertices contained in a circle of radius $\sqrt{n}$, for $n\in\N$, centered at a vertex of any side-length 1 equilateral triangle tiling of the plane. \end{remark} \subsubsection{Lattice points in the region $x^2+2y^2\le n$} We now consider the ellipse $x^2+2y^2=n$. The result below gives the number of representations of a natural number $n$ in terms of the binary quadratic form $x^2+2y^2$. \begin{prop}[{\cite[Exercise XXII.1]{Dickson1958}}] \label{prop:dickson-x^2+2y^2} Let $n\in\N$. Then the number of representations of $n=x^2+2y^2$, for integers $x,y$, is $2\left(\divs_{1,8}(n)+\divs_{3,8}(n)-\divs_{5,8}(n)-\divs_{7,8}(n) \right)$. \end{prop} \begin{cor}\label{cor:x^2+2y^2} Let $F(n)$ be the number of lattice points contained in the elliptical region $x^2+2y^2=n$. Then $F(0)=1$, $F(1)=3$, $F(2)=5$, $F(5)=11$, and for all $n\in\N$ with $n\ne 1,2,5$, \[ F(n) = 1+2\Num_{\Ceil{n/4},3/4,\{-n/4\},1,2}+2\Num_{\Ceil{n/4},1/4,\{-n/4\},1,2}+2n+2\Floor{\frac{n}{3}}-4\Ceil{n/4}. \] \end{cor} \begin{proof} Let $f(n) = \#\left\{(x,y)\in\Z^2 : x^2+2y^2=n\right\}$. Observe that $f(0)=1$. To start, we have \begin{equation}\label{eqn:f-step-1} F(n) = \sum\limits_{k=0}^n f(k) = 1 + \sum\limits_{k=1}^n f(k). \end{equation} Next, by Proposition~\ref{prop:dickson-x^2+2y^2}, \ \sum\limits_{k=1}^n f(k) = 2\left(\Floor{\frac{n}{1}}+\Floor{\frac{n}{3}}-\Floor{\frac{n}{5}}-\Floor{\frac{n}{7}}+\Floor{\frac{n}{9}}+\dots\right), \ which is a finite sum since the floors are eventually zero. We'll need two alternating floor sums here -- one for $\Floor{n/1}-\Floor{n/5}+\Floor{n/9}-\dots$ and one for $\Floor{n/3}-\Floor{n/7}+\Floor{n/11}-\dots$. Each will need $q=4$ and $m=2$. As we did with Proposition~\ref{prop:circle-radius-n}, we will have $\nu\ne0$ here. For $n\in\N$, let $a=\Ceil{n/4}$ and $b=4a-n$. Then $0\le b<4$. Using $\alpha=p/q=3/4$, $\nu=b/4$, $r=1$, and $m=2$ with Equation~\eqref{eqn:num-rational-alpha}, we get \[ \Num_{a,3/4,b/4,1,2} = a - \Floor{\frac{n}{1}} + \sum\limits_{i\ge0}\left(\Floor{\frac{n}{1+4i}}-\Floor{\frac{n}{5+4i}}\right) \] for all $a\ge\nu/\alpha=b/3$. Using $\alpha=p/q=1/4$, $\nu=b/4$, $r=1$, and $m=2$ with Equation~\eqref{eqn:num-rational-alpha}, we get \[ \Num_{a,1/4,b/4,1,2} = a - \Floor{\frac{n}{3}} + \sum\limits_{i\ge0}\left(\Floor{\frac{n}{3+4i}}-\Floor{\frac{n}{7+4i}}\right) \] for all $a\ge\nu/\alpha=b$. Both equations hold for $a\ge b$. Since $a\ge1$ and $0\le b<4$, this inequality is satisfied for all $a,b$ except for $(a,b)=(1,2), (1,3), (2,3)$, which correspond, respectively, to $n=2$, $n=1$, and $n=5$. Thus, for $n\in\N$ with $n\ne 1,2,5$, \begin{align*} F(n) &=1+2\left(\Num_{a,3/4,b/4,1,2}+\Num_{a,1/4,b/4,1,2}-a+\Floor{\frac{n}{3}}-a+n\right)\\ &=1+2\Num_{\Ceil{n/4},3/4,\{-n/4\},1,2}+2\Num_{\Ceil{n/4},1/4,\{-n/4\},1,2}+2n+2\Floor{\frac{n}{3}}-4\Ceil{\frac{n}{4}} \end{align*} We can fill in the missing values by hand. We compute $f(1)=2$, $f(2)=2$, and $f(5)=0$. And by the formula above with $n=4$, \[ F(4) =1+2\Num_{1,3/4,0,1,2}+2\Num_{1,1/4,0,1,2}+4+2\Floor{4/3}-0 =1+2\cdot1+2\cdot1+4+2 =11. \] Thus, $F(0)=1$, $F(1)=F(0)+f(1)=1+2=3$, $F(2)=F(1)+f(2)=3+2=5$, and $F(5)=F(4)+f(5)=11+0=11$. \end{proof} \begin{figure} \centering \includegraphics[scale=.6]{x2p2y2e29.eps} \includegraphics[scale=.6]{sqrtm2lattice.eps} \caption{The 65 lattice points within the ellipse $x^2+2y^2=29$, and the 65 elements of $\Z[\sqrt{-2}]$ with norm at most 29.} \label{fig:x^2+2y^2=29} \end{figure} \begin{example}\label{ex:x^2+2y^2=29} We'll compute the number of lattice points contained in the ellipse $x^2+2y^2=29$. We have $n=29$, $\Ceil{29/4}=8$, and $\{-29/4\}=3/4$. To compute $\Num_{8,3/4,3/4,1,2}$, we consider the sequence \[ \Floor{\frac{8-3/4}{1}+3/4},\Floor{\frac{8-3/4}{2}+3/4},\dots,\Floor{\frac{8-3/8}{1}+3/4} = 8,4,3,2,2,1,1,1. \] There are 4 odd numbers, so $\Num_{8,3/4,3/4,1,2}=4$. To compute $\Num_{8,1/4,3/4,1,2}$, we consider the sequence \[ \Floor{\frac{8-3/4}{1}+1/4},\Floor{\frac{8-3/4}{2}+1/4},\dots,\Floor{\frac{8-3/8}{1}+1/4} = 7,3,2,2,1,1,1,1. \] There are 6 odd numbers, so $\Num_{8,1/4,3/4,1,2}=6$. Thus, the number of lattice points in this ellipse is \[ F(29) = 1+2\Num_{8,3/4,3/4,1,2}+2\Num_{8,1/4,3/4,1,2}+29+2\Floor{\frac{29}{3}}-3 = 1+2\cdot4+2\cdot6+29+2\cdot9-3 = 65. \] See Figure~\ref{fig:x^2+2y^2=29}. \end{example} \begin{remark} Consider the ring $\Z[\sqrt{-2}]$. For $z=a+b\sqrt{-2}\in\Z[\sqrt{-2}]$, the norm of $z$ is $N(z)=a^2+2b^2$. Thus, Corollary~\ref{cor:x^2+2y^2} gives a formula for the number of elements of $\Z[\sqrt{-2}]$ with norm at most $n$. Following from Example~\ref{ex:x^2+2y^2=29}, we know there are 65 elements of $\Z[\sqrt{-2}]$ with norm at most 29. We visualize these elements in Figure~\ref{fig:x^2+2y^2=29}. \end{remark}

https://arxiv.org/abs/2203.13205

Honeycombs in the Pascal triangle and beyond

In this paper we present a geometric approach to discovering some known and some new identities using triangular arrays. Our main aim is to demonstrate how to use the geometric patterns (by Carlitz), in the Pascal and Hosoya triangles to rediscover some classical identities and integer sequences. Therefore, we use new techniques in classical settings which then provide a new perspective in undergraduate research.

\section{Carlitz's patterns in the Pascal triangle}\label{pascal:triangle} In this section we use Carlitz's patterns seen in Figure \ref{carlitz_patterns:1} to rediscover some well-known binomial identities. The patterns in Figure \ref{carlitz_patterns:1} (a)--(d) can also be seen as Patterns 1, 2, in Figure \ref{hex_patterns:1}(a) and Patterns 3, 4 in Figure \ref{hex_patterns:1}(b). Note that the identity present in Proposition \ref{propo:3} can only be found partially in literature while the identity in Theorem \ref{theorem:4} is new. Here we observe that the hexagonal tiles used in the Pascal triangle is constructed using the entires of the triangle in every other row as shown in Figure \ref{hex_patterns:1}. In this figure, each dot is an entry of the triangle. This setup of the hexagonal tiles can also be seen in Figure \ref{alternatingpatterns:2}. This construction of the honeycomb pattern can also be seen in Figure \ref{alternatingpatterns:2}. Based on this honeycomb pattern we can construct another triangular array with entries of the from each hexagonal cells. These entries are of the form $\binom{2k}{i}$ where $k>0$ and $i$ is odd. The entries of this triangular array are $ 2, 4, 4, 6, 20, 6, 8, 56, 8,\ldots$. We leave it to the readers to conduct \begin{figure}[htbp] \includegraphics[scale=0.30]{alternating_patterns.pdf} \caption{Hexagonal cells in the triangles and patterns.} \label{hex_patterns:1} \end{figure} The first identity in Proposition \ref{propo:1}\eqref{part:a} is obtained by using Pattern 1 in Figure \ref{hex_patterns:1}(a) (or Carlitz's pattern, Figure \ref{carlitz_patterns:1}(b)), in Pascal's triangle. For example, using Figure \ref{hex_patterns:1} (a), Pattern 1, we have that $$\underbrace{\binom{6}{1}}_{\color{blue}{\textbf a}}+ \underbrace{\binom{6}{3}}_{\color{blue}{\textbf b}} +\underbrace{\binom{6}{5}}_{\color{blue}{\textbf c}}= 2^5.$$ This identity is also clear from Figure \ref{alternatingpatterns:2}. The second identity in Proposition \ref{propo:1}\eqref{part:b} is obtained by using Pattern 2 in Figure \ref{hex_patterns:1}(a) (or Carlitz's pattern, Figure \ref{carlitz_patterns:1}(b)), in Pascal's triangle. For example, using in Pattern 2 we have that $$\underbrace{\binom{10}{1}}_{\color{blue}{\textbf d}}+\underbrace{\binom{8}{1}}_{\color{blue}{\textbf e}}+\underbrace{\binom{10}{3}}_{\color{blue}{\textbf f}}+\underbrace{\binom{8}{3}}_{\color{blue}{\textbf g}}+\underbrace{\binom{10}{5}}_{\color{blue}{\textbf h}}+\underbrace{\binom{8}{5}}_{\color{blue}{\textbf i}}+\underbrace{\binom{10}{7}}_{\color{blue}{\textbf j}}+\underbrace{\binom{8}{7}}_{\color{blue}{\textbf k}}+\underbrace{\binom{10}{9}}_{\color{blue}{\textbf l}}=5(2^7) .$$ Therefore, in general, we obtain the following identities. \begin{proposition}\label{propo:1} If $n=2k$ for $k>0$, then the following hold \begin{enumerate}[(a)] \item \label{part:a} $\displaystyle \sum_{i=1}^{k}\binom{2k}{2i-1}=2^{2k-1}$, \item \label{part:b} $\displaystyle \sum_{i=1}^{k}\binom{2k}{2i-1}+\sum_{i=1}^{k-1}\binom{2k-2}{2i-1}=5(2^{2k-3})$. \end{enumerate} \end{proposition} \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.3]{powerof2.pdf} \caption{ Power of 2 in Pascal triangle.} \end{center} \label{alternatingpatterns:2} \end{figure} The identity in Proposition \ref{propo:3} is obtained by using Pattern 3 in Figure \ref{hex_patterns:1}(b). For example, using this pattern we have that $$\underbrace{\binom{12}{1}}_{\color{blue}{\textbf s}}+\underbrace{\binom{12}{3}}_{\color{blue}{\textbf t}}+\underbrace{\binom{10}{3}}_{\color{blue}{\textbf u}}+\underbrace{\binom{12}{5}}_{\color{blue}{\textbf v}}+\underbrace{\binom{12}{7}}_{\color{blue}{\textbf w}}+\underbrace{\binom{10}{7}}_{\color{blue}{\textbf x}}+\underbrace{\binom{12}{9}}_{\color{blue}{\textbf y}}+\underbrace{\binom{12}{11}}_{\color{blue}{\textbf z}}=2^4[9(2^4)-1]=2288.$$ \begin{proposition}\label{propo:3} If $n\ge 2$ is any positive integer, then \[\sum_{i=1}^{2n}\binom{4n}{2i-1}+\sum_{i=1}^{n-1}\binom{4n-2}{4i-1}= \begin{cases} 2^{2n-2}[9 (2^{2n-2})+1], & \text{ if $n$ is even};\\ 2^{2n-2}[9 (2^{2n-2})-1], & \text{ if $n$ is odd}. \end{cases}\] \end{proposition} The identity in Theorem \ref{theorem:4} below is not present in existing literature and is obtained by using the pattern in Figure \ref{carlitz_patterns:1}(d) (or Pattern 4 in Figure \ref{hex_patterns:1}(b)). For example, uisng Figure \ref{hex_patterns:1}, Pattern 4 we have that $$\underbrace{\binom{6}{1}}_{\color{blue}{ \textbf m}}+ \underbrace{\binom{4}{1}}_{\color{blue}{ \textbf n}} +\underbrace{\binom{4}{3}}_{\color{blue}{ \textbf o}}+\underbrace{\binom{6}{5}}_ {\color{blue}{ \textbf p}}= 20.$$ In general, this sum is congruent to $20\bmod 42$. \begin{theorem}\label{theorem:4} If $n\ge 1$, then $$\sum_{i=0}^{n-1}\left[\binom{6n}{6i+1}+\binom{6n}{6i+5}+\binom{6n-2}{6i+1}+\binom{6n-2}{6i+3}\right]\equiv \binom{6}{1}+ \binom{6}{5} +\binom{4}{1}+\binom{4}{3}= 20 \bmod\, 42.$$ \end{theorem} Finally, the Carlitz's pattern see in Figure \ref{carlitz_patterns:1}(e) gives rise to the integer sequence $$2, 4, 4, 6, 20, 6, 8, 56, 56, 8, 10, 120, 252, \ldots=\binom{2n}{2m+1}, \text{ for } n>1 \text{ and }0\le m\le n-1.$$ This sequence is related to the integer sequence {A091044} in \cite{sloane} with terms $1, 2, 2, 3, 10, 3, 4, 28, 28, \ldots$. In fact, the sequence {A091044}, is one half of the odd-numbered entries of even-numbered rows of the Pascal triangle. \section{Carlitz's patterns in the Hosoya triangle}\label{hosoya:triangle} The Hosoya triangle \cite{hosoya}, is a triangular array where the entries are products of Fibonacci numbers. The geometry of the Hosoya triangle presents an interesting method to explore properties of Fibonacci numbers. Several properties and identities have already been discovered and published. Some articles on this topic are by Hosoya, Koshy, and Fl\'orez \emph{et al.} \cite{Blair,florezHiguitaJunesGCD, florezjunes, hosoya, koshy}. \begin{table} [!h] \small \begin{center} \addtolength{\tabcolsep}{-3pt} \scalebox{.9}{% \begin{tabular}{cccccccccccccccccccccccccccc} &&&&&&&&&&&&& 1 &&&&&&&&&&&&&&\\ &&&&&&&&&&&& 1 && 1 &&&&&&&&&&&&&\\ &&&&&&&&&&& 2 && 1 && 2 &&&&&&&&&&&&\\ &&&&&&&&&& 3 && 2 && 2 && 3 &&&&&&&&&&\\ &&&&&&&&& 5 && 3 && 4 && 3 && 5 &&&&&&&&&\\ &&&&&&&& 8 && 5 && 6 && 6 && 5 && 8 &&&&&&&&\\ &&&&&&& 13 && 8 && 10 && 9 && 10 && 8 && 13 &&&&&&&\\ &&&&&& 21 && 13 && 16 && 15 && 15 && 16 && 13 && 21 &&&&&&\\ \end{tabular} \begin{tabular}{cccccccccccccccccccccccccccc} &&&&&&&&&&&& & 1 &&&&&&&&&&&&&&\\ &&&&&&&&&&&& 1 && 2 &&&&&&&&&&&&&\\ &&&&&&&&&&& 1 && 3 && 2 &&&&&&&&&&&\\ &&&&&&&&& & 1 && 4 && 5 && 2 & &&&&&&&&&\\ &&&&&&&& & 1 && 5 && 9 && 7 && 2 &&&&&&&&&\\ &&&&&&&& 1 && 6 && 14 && 16 && 9 && 2 & &&&&&&&\\ &&&&&&& 1 && 7 && 20 && 30 && 25 && 11 && 2 & &&&&&&\\ &&&&&& 1 && 8 && 27 && 50 && 55 && 36 && 13 && 2 &&&&&&\\ \end{tabular}} \end{center} \caption{(a) Hosoya triangle \hspace{4.0cm} (b) Lucas triangle.} \label{tabla1} \end{table} We represent each entry of the Hosoya triangle as $H_{r,k}= F_kF_{r-k+1}$ for positive integers $r$ and $k$ with $r \ge k\ge 1$, (see \cite[p.~188]{koshy}). Recently Blair et al. \cite{Blair1}, published the article titled ``Honeycombs in the Hosoya Triangle". In this article the authors look into the patterns in Figure \ref{carlitz_patterns:1} (a) and (b) (or Patterns 1 and 2 in Figure \ref{hex_patterns:1}) inside the Hosoya triangle and this leads to discovery of some new and geometric interpretation of some known identities. In addition, we obtain a sequence of integers by following the pattern seen in Figure \ref{carlitz_patterns:1}(e). The examples and propositions that follow appeared in \cite{Blair1}. For example, using Pattern 1 of Figure \ref{hex_patterns:1}(a) we have that $$\underbrace{H_{7,2}}_ {\color{blue}{ \textbf a}}+ \underbrace{H_{7,4}}_{\color{blue}{ \textbf b}}+\underbrace{H_{7,6}}_{\color{blue}{ \textbf c}}=F_2F_6+F_4F_4+F_6F_4= 25.$$ In general this gives rise to the integer sequence {A001871} in OEIS (\cite{sloane}). Each term of this sequence is described as the convolution of the even-indexed Fibonacci numbers with itself. This generalizes to the following identity which is a special case of a result by Fl\'{o}rez et al. in \cite{Czabarka}. \begin{proposition}[\cite{Blair1}] \label{prop:6} If $n>0$ is odd, then \[ \sum_{k=1}^{(n-1)/2}H_{(n,2k)}=\sum_{k=1}^{(n-1)/2}F_{2k}F_{n-2k+1}=\left[(2n-6)F_{n-3}+(3n-5)F_{n-2}+[(n-1)/2] \,F_{n-2}\right]/5.\] \end{proposition} Next, we consider Pattern 2 in Figure \ref{hex_patterns:1}(a) embedded in the Hosoya triangle. As an example, using Pattern 2 we have that $$ \underbrace{H_{11,2}}_{\color{blue}{ \textbf d}}+\underbrace{H_{9,2}}_ {\color{blue}{ \textbf e}}+\underbrace{H_{11,4}}_{\color{blue}{ \textbf f}}+\underbrace{H_{9,4}}_{\color{blue}{ \textbf g}}+\underbrace{H_{11,6}}_{\color{blue}{ \textbf h}}+\underbrace{H_{9,6}}_{\color{blue}{ \textbf i}}+\underbrace{H_{11,8}}_{\color{blue}{ \textbf j}}+\underbrace{H_{9,8}}_{\color{blue}{ \textbf k}}+\underbrace{H_{11,10}}_{\color{blue}{ \textbf l}}$$ equals the sum $2(F_2F_{10}+F_4F_8)+F_6F_6+2(F_2F_8+F_4F_6)=390$. Here we observe that the sums of Hosoya triangle entries $H_{n,k}$ following this pattern ---with $n$ an odd integer--- gives rise to the integer sequence {A197649} in OEIS (\cite{sloane}). In general we have the following identity. \begin{proposition}[\cite{Blair1}]\label{prop:7} If $n>0$ is odd, then \[\sum_{k=1}^{(n-1)/2}H_{n,2k}+\sum_{k=1}^{(n+1)/2}H_{n+2,2k}=\sum_{k=1}^{(n-1)/2}F_{2k}F_{n-2k+1}+\sum_{k=1}^{(n+1)/2}F_{2k}F_{n-2k+3}=[(n+1)F_{n+2}-2F_{n+1}]/2.\] \end{proposition} Finally, the pattern seen in Figure \ref{carlitz_patterns:1} (e) gives rise to the integer sequence $1, 3, 3, 8, 9, 8, 21, 24,\ldots$, which is identified as sequence {A141678} in OEIS (\cite{sloane}). This sequence is defined as symmetrical triangle of coefficients based on invert transform of the sequence {A001906}. Note that the sequence {A001906} is the bisection of the Fibonacci sequence given by the recursive sequence $a(n)=3a_{n-1}-a_{n-2}$ for $n\ge 3$, with initial conditions $a_1=0$ and $a_1=1$. \section{Carlitz's patterns in other triangles}\label{other:triangle} Carlitz's patterns can be used in other triangles like the Lucas triangle, Josef's triangle, and Liebniz's harmonic triangle to explore and obtain identities related to several different integer sequences. The Lucas triangle (Table \ref{tabla1}(b)) is constructed using coefficients of polynomials of the form $f_n(x,y)=(x+y)^{(n-1)(x+2y)}$. If this triangle is left-justified, then its rising diagonals give rise to Lucas numbers. On the other hand, Josef's triangles can be constructed like the Hosoya triangle by starting with two outermost columns of Lucas and Fibonacci number sequences (Table \ref{tabla2}(a)). Leibniz's triangle (Table \ref{tabla2}(b)) can be obtained by using the following recursive relation \[L(n, 1) = 1/n; L(n, k) = L(n-1, k-1)-L(n, k-1), \text{for } k>1.\] All three of these triangles are described in detail in \cite{koshy2}. For all three triangles Patterns 1 and 2 in Figure \ref{hex_patterns:1} yields some known identities or integer sequences found in \cite{sloane}. In particular for the Lucas triangle, Pattern 1 (sum of every other entry in the odd-numbered rows), yields the integer sequence $3,12,48,192, \dots$ which is given by the formula $a_n=3(4^{n-1})$ for $n\ge 1$. This is also present in OEIS \cite{sloane} as sequence {A002001}. Pattern 2 in the Lucas triangle yields the sequence $3,15,60,240,\ldots$. Similarly, for Josef's triangle, Pattern 1, yields the integer sequence $2, 10, 40,1 42, 470\ldots$, which is related to the sequence of self-convoluted Fibonacci numbers ( {A001629} in OEIS). In fact, one-half of each term of this sequence, namely $1, 5, 20, 71, 235,\ldots$, is the sequence of even-indexed terms of {A001629}. The sequence $1, 5, 20, 71, 235,\ldots$, is identified in OEIS (\cite{sloane} ) as {A054444}. In Josef's triangle, using Pattern 2 we get the sequence $12,50,182,612,\ldots$ where each term is the sum of consecutive terms of the sequence $2, 10, 40, 142, 470, \ldots$, obtained from Pattern 1 above. Finally, for the Leibniz's harmonic denominator triangle, Pattern 1 gives us the following sequence $$6, 40, 224,1 152, 5632,\ldots.$$ This sequence can be given by the recursive relation $D_n=8D_{n-1}-16D_{n-2}$ for $n>3$ with initial conditions $D_1=1, D_2=6$, and $D_3=40$. This sequence is identified in \cite{sloane} as sequence {A229580}. In the same triangle, Pattern 2 gives us the sequence $46,264,1376,7784,\ldots$. Each term of this sequence is the sum of consecutive terms of the sequence $6,40,224,1152, 5632,\ldots.$ obtained from Pattern 1. \begin{table} [!h] \small \begin{center} \addtolength{\tabcolsep}{-3pt} \scalebox{.8}{% \begin{tabular}{cccccccccccccccccccccccccccc} &&&&&&&&&&&&& 2 &&&&&&&&&&&&&&\\ &&&&&&&&&&&& 1 && 1 &&&&&&&&&&&&&\\ &&&&&&&&&&& 3 && 2 && 3 &&&&&&&&&&&&\\ &&&&&&&&&& 4 && 3 && 3 && 4 &&&&&&&&&&\\ &&&&&&&&& 7 && 5 && 6 && 5 && 7 &&&&&&&&&\\ &&&&&&&& 11 && 8 && 9 && 9 && 8 && 11 &&&&&&&&\\ &&&&&&& 18 && 13 && 15 && 14 && 15 && 13 && 18 &&&&&&&\\ &&&&&& 29 && 21 && 24 && 23 && 23 && 24 && 21 && 29 &&&&&&\\ \end{tabular} \begin{tabular}{cccccccccccccccccccccccccccc} &&&&&&&&&&&& & 1 &&&&&&&&&&&&&&\\ &&&&&&&&&&&& $\frac{1}{2}$ && $\frac{1}{2}$ &&&&&&&&&&&&&\\ &&&&&&&&&&& $\frac{1}{3}$ && $\frac{1}{6}$ && $\frac{1}{3}$ &&&&&&&&&&&\\ &&&&&&&&& & $\frac{1}{4}$ && $\frac{1}{12}$ && $\frac{1}{12}$ && $\frac{1}{4}$ & &&&&&&&&&\\ &&&&&&&& & $\frac{1}{5}$ && $\frac{1}{20}$ && $\frac{1}{30}$ && $\frac{1}{20}$ && $\frac{1}{5}$ &&&&&&&&&\\ &&&&&&&& $\frac{1}{6}$ && $\frac{1}{30}$ && $\frac{1}{60}$ && $\frac{1}{60}$ && $\frac{1}{30}$ && $\frac{1}{6}$ & &&&&&&&\\ &&&&&&& $\frac{1}{7}$ && $\frac{1}{42}$ && $\frac{1}{105}$ && $\frac{1}{140}$ && $\frac{1}{105}$ && $\frac{1}{42}$ && $\frac{1}{7}$ & &&&&&&\\ &&&&&& $\frac{1}{8}$ && $\frac{1}{56}$ && $\frac{1}{168}$ && $\frac{1}{280}$ && $\frac{1}{280}$ && $\frac{1}{168}$ && $\frac{1}{56}$ && $\frac{1}{8}$ &&&&&&\\ \end{tabular}} \end{center} \caption{ (a) Josef's triangle \hspace{3.5cm} (b) Leibniz harmonic triangle} \label{tabla2} \end{table} \section{Appendix --Proofs} In this section we present some proofs of the identities in case the reader is interested in the formal proofs of the results presented earlier. First we present some well-known identities involving binomial coefficients which will be used to prove the propositions in this section. Lemma \ref{lemma:1} Parts \eqref{lemma:parta} and \eqref{lemma:partb} and their proofs can be found in any combinatorics book, for example \cite{gould1, spivey}. Lemma \ref{lemma:1}, part \eqref{lemma:partc} is a special case of an identity found in \cite{gould1}. The proofs of Proposition \ref{propo:1} Parts \eqref{part:a} and \eqref{part:b} follow directly from Lemma \ref{lemma:1} Parts \eqref{lemma:parta} and \eqref{lemma:partb}. \begin{lemma}\label{lemma:1} If $n,k$ are positive integers, then the following hold \begin{enumerate}[(a)] \item $\displaystyle\binom{n+1}{k}=\binom{n}{k}+\binom{n}{k-1}$.\label{lemma:parta} \item $\displaystyle\sum_{k=0}^n \binom{n}{k} =2^n$.\label{lemma:partb} \item $\displaystyle\sum_{k=0}^{2n}(-1)^k\binom{4n}{2k}=(-1)^n2^{2n}$. \label{lemma:partc} \end{enumerate} \end{lemma} \begin{proof}[Proof of Proposition \ref{propo:3}] We first note that the proof of $\displaystyle\sum_{i=1}^{2n}\binom{4n}{2i-1}=2^{4n-1}$ is well-known (see \cite{gould1, spivey}). Therefore, we only prove that \begin{equation}\label{eqn:1} \sum_{i=1}^{n-1}\binom{4n-2}{4i-1}= \begin{cases} 2^{2n-2}[2^{2n-2}+1] & \text{ if } n \text{ is even};\\ 2^{2n-2}[2^{2n-2}-1] & \text{ if } n \text{ is odd}. \end{cases} \end{equation} Observe that Lemma \ref{lemma:1} part (a) implies that \[\displaystyle\sum_{i=1}^{n-1}\binom{4n-2}{4i-1}=\sum_{i=1}^{n-1}\binom{4n-3}{4i-1}+\binom{4n-3}{4i-2}.\] Using the same identity again and regrouping the terms in the sum, we have that \[\sum_{i=1}^{n-1}\binom{4n-2}{4i-1}=\sum_{i=0}^{4n-4}\binom{4n-4}{i}-\sum_{i=0}^{n-1}\binom{4n-4}{4i}+\sum_{i=1}^{n-1}\binom{4n-4}{4i-2}.\] Further simplification gives us, $\displaystyle\sum_{i=1}^{n-1}\binom{4n-2}{4i-1}=\sum_{i=0}^{4n-4}\binom{4n-4}{i}-\sum_{i=0}^{2n-2}(-1)^i\binom{4n-4}{2i}$. Finally, applying Lemma \ref{lemma:1} Parts \eqref{lemma:parta} and \eqref{lemma:partb}, we obtain \[\sum_{i=1}^{n-1}\binom{4n-2}{4i-1}= \begin{cases} 2^{2n-2}[2^{2n-2}+1] & \text{ if } n-1 \text{ is odd};\\ 2^{2n-2}[2^{2n-2}-1] & \text{ if } n-1\text{ is even,} \end{cases} \] which is equivalent to \eqref{eqn:1}. This completes the proof. \end{proof} \begin{proof}[Proof of Theorem \ref{theorem:4}] Note that, $$\sum_{i=0}^{n-1}\left[\binom{6n}{6i+1}+\binom{6n}{6i+5}+\binom{6n-2}{6i+1}+\binom{6n-2}{6i+3}\right]$$ is congruent with $0 \bmod 2$, is congruent with $2 \bmod 3$, \text{ and } is congruent with $6 \bmod 7.$ By the Chinese Remainder Theorem, we can easily solve this system of equation. This proves the theorem. \end{proof} \section{Acknowledgement} The second and third authors were partially supported by The Citadel Foundation.

https://arxiv.org/abs/1203.6668

Making Markov chains less lazy

The mixing time of an ergodic, reversible Markov chain can be bounded in terms of the eigenvalues of the chain: specifically, the second-largest eigenvalue and the smallest eigenvalue. It has become standard to focus only on the second-largest eigenvalue, by making the Markov chain "lazy". (A lazy chain does nothing at each step with probability at least 1/2, and has only nonnegative eigenvalues.)An alternative approach to bounding the smallest eigenvalue was given by Diaconis and Stroock and Diaconis and Saloff-Coste. We give examples to show that using this approach it can be quite easy to obtain a bound on the smallest eigenvalue of a combinatorial Markov chain which is several orders of magnitude below the best-known bound on the second-largest eigenvalue.

\section{Introduction}\label{s:intro} Let $\mathcal{M}$ be an ergodic, reversible Markov chain with finite state space $\Omega$ and transition matrix $P$. It is well known that the eigenvalues of $\mathcal{M}$ satisfy \[ 1 = \lambda_0 > \lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_{N-1} > -1,\] where $N=|\Omega|$. We refer to $\lambda_{N-1}$ as the \emph{smallest eigenvalue} of $\mathcal{M}$. The connection between the mixing time of a Markov chain and its eigenvalues is well-known (see~\cite[Proposition 1]{sinclair}): \begin{equation} \label{mix-time} \tau(\varepsilon) \leq (1-\lambda_\ast)^{-1}\, \ln \frac{1}{\epsilon\, \pi_{\min}} \end{equation} where $\tau(\varepsilon)$ denotes the mixing time of the Markov chain, $\pi_{\min} = \min_{x\in \Omega} \pi(x)$ and \[ \lambda_\ast = \max\{ \lambda_1, \, |\lambda_{N-1}|\}.\] When studying the mixing time of a Markov chain $\mathcal{M}$ using (\ref{mix-time}), the approach which has become standard is to make the chain $\mathcal{M}$ \emph{lazy} by replacing $P$ by $(I+P)/2$, where $I$ denotes the identity matrix. Then all eigenvalues of the lazy chain are nonnegative, and only the second-largest eigenvalue must be investigated. A lazy chain can be implemented so that its expected running time is the same as the mixing time of the original chain. So the problem with lazy chains is not their efficiency. In our opinion, the main problem with lazy Markov chains is conceptual: in order to prove that a Markov chain is fast, we first slow it down. The device of using lazy Markov chains has been called ``crude''~\cite[p.~110]{SJ89} and ``unnatural''~\cite[Chapter 5]{jerrum-book}. In this note, we aim to advertise an approach for bounding the smallest eigenvalue of a Markov chain. This approach was first proposed by Diaconis and Stroock in 1991~\cite[Proposition 2]{DS91}, and a modified version was presented by Diaconis and Saloff-Coste two years later~\cite[p.702]{DSC} (restated as Lemma~\ref{lazy} below). The method of~\cite{DSC} has been applied in~\cite{DSC,DS91,goel}, but in the theoretical computer science community it has become common to work with lazy chains. We urge researchers to first try the approach of~\cite{DSC,DS91} before choosing to work with a lazy version of their chain. Finally we remark that in~\cite{directed} the author wrongly claimed that their~\cite[Lemma 1.3]{directed} was new, when in fact it is precisely the result of~\cite[p.702]{DSC}. We sincerely apologise for this error. \subsection{The method} See~\cite{jerrum-book} for Markov chain definitions not given here. Write $\mathcal{G}$ for the underlying directed graph of the Markov chain $\mathcal{M}$, where $\mathcal{G} = (\Omega, \Gamma)$ and each directed edge $e\in \Gamma$ corresponds to a transition of $\mathcal{M}$. If $P(x,x)>0$ then the edge $xx$ is called a \emph{self-loop} at $x$. Define $Q(e) = Q(x,y) = \pi(x)P(x,y)$ for the edge $e=xy$. A walk in $\mathcal{G}$ is a sequence of states $x_0 x_1 \cdots x_\ell$ such that $P(x_j,x_{j+1})> 0$ for $j=0,\ldots, \ell-1$. The walk is \emph{closed} if $x_\ell=x_0$. If a walk has odd length then we call it an \emph{odd walk}. For each $x\in\Omega$ let $w_x$ be an odd walk from $x$ to $x$ in $\mathcal{G}$. (Such a walk exists for each $x$, since the Markov chain is aperiodic.) Define $\mathcal{W} = \{ w_x : x\in\Omega\}$, a set of ``canonical closed odd walks''. For each transition $e\in\Gamma$ and each $w\in\mathcal{W}$, let $r(e,w)$ denote the number of times that $e$ appears as a directed edge of $w$. We can assume that $r(e,w)\leq 2$ for all transitions $e$ (indeed, if $e$ is a self-loop then we can assume that $r(e,w)\leq 1$.) The \emph{congestion} of $\mathcal{W}$, denoted by $\eta(\mathcal{W})$, is defined by \[ \eta(\mathcal{W}) = \max_{e\in\Gamma} \, Q(e)^{-1}\, \sum_{x\in\Omega,\, e\in w_x}\, r(e,w_x)\, \pi(x)\, |w_x|.\] \begin{lemma} \emph{\cite[p.702]{DSC}}\ Suppose that $\mathcal{M}$ is a reversible, ergodic Markov chain with state space $\Omega$, and let $\mathcal{W}$ be a set of odd walks defined as above. Then \[ (1 + \lambda_{N-1})^{-1} \leq \frac{\eta(\mathcal{W})}{2}.\] \label{lazy} \end{lemma} If $|w_x|=1$ for all $x\in\Omega$ then the bound of Lemma~\ref{lazy} simplifies further to \begin{equation} \label{all-loops} (1+\lambda_{N-1})^{-1} \leq \dfrac{1}{2}\operatorname{max}_{x\in\Omega} P(x,x)^{-1}. \end{equation} \begin{remark} \emph{ Suppose that the graph underlying a Markov chain $\mathcal{M}$ can be obtained from a connected bipartite graph by adding loops to an exponentially small proportion of states. For example, many instances of the \emph{knapsack chain}~\cite{MS99} satisfy this property. Since every closed odd walk must traverse at least one of these self-loop edges, it is very difficult to define a set of canonical closed odd walks with low congestion. So Lemma~\ref{lazy} is unlikely to be easy to apply in this case. } \end{remark} \section{Applications of the method}\label{s:applications} We illustrate the use of Lemma~\ref{lazy} by applying it to three combinatorial Markov chains. Our applications are all ergodic and reversible with uniform stationary distribution, and no edge will be used more than once in any walk $w_x$ that we define. In this case the congestion can be simplified to \begin{equation} \label{simple} \eta(\mathcal{W}) = \operatorname{max}_{e\in\Gamma} \, P(e)^{-1} \, \sum_{x\in\Omega,\,\, e\in w_x} \, |w_x|, \end{equation} where $P(e) = P(x,y) = P(y,x)$ for the transition $e=xy$. \subsection{The switch chain for sampling regular graphs} Our first application is to the Markov chain for sampling regular graphs known as the \emph{switch chain}. A transition of the chain is performed as follows: from the current state $G$ (a $d$-regular graph on vertex set $[n]$) choose an unordered pair of non-incident edges uniformly at random, let $G'$ be the multigraph obtained from $G$ by deleting these edges and inserting a perfect matching of their four endvertices, selected uniformly at random. If $G'$ has no repeated edges then the new state is $G'$, otherwise it is $G$. The lazy version of this chain was analysed by Cooper et al.~\cite{CDG,CDG-corrigendum}. Clearly $P(G,G)\geq \nfrac{1}{3}$ for every state $G$ of this chain, so by (\ref{all-loops}) we immediately conclude that \[ (1 + \lambda_{N-1})^{-1}\leq \nfrac{3}{2}. \] This is several orders of magnitude smaller than the best-known bound on $(1-\lambda_1)^{-1}$, which is $O(d^{23} n^8)$ (see~\cite{CDG-corrigendum}). \subsection{Jerrum and Sinclair's matchings chain } The next application is to the well-known Markov chain for sampling perfect and near-perfect matchings of a fixed graph $G$. A transition of the chain is performed as follows: from the current state $M$ (which is a perfect or near-perfect matching of $G$), choose an edge $e\in E(G)$ uniformly at random. If $M$ is a perfect matching and $e\in M$ then the new state is $M - \{e\}$. If $M$ is a near-perfect matching and both endvertices of $e$ are unmatched in $M$ then the new state is $M \cup \{e\}$. If $M$ is a near-perfect matching, and exactly one endvertex of $e$ is unmatched in $M$ then let $e'$ be the edge of $M$ which matches the other endvertex of $e$: the new state is $(M - \{e'\})\cup\{e\}$. In all other cases the new state is $M$. The lazy version of this chain was analysed by Jerrum and Sinclair~\cite{JS89,JS96}, If $G$ itself is not a perfect matching then $P(M,M)\geq 1/|E|$ for all states $M$ of the chain (that is, for all perfect or near-perfect matchings $M$ of $G$). Therefore (\ref{all-loops}) implies that \[ (1 + \lambda_{N-1})^{-1} \leq \frac{|E|}{2}.\] This bound is at least a factor $n^2$ smaller than the smallest-known bound on $(1-\lambda_1)^{-1}$, which is $O(n|E|q(n))$ for graphs $G$ for which the ratio between the number of near-perfect and perfect matchings is $q(n)$ (see~\cite{JS96}). \subsection{A heat-bath chain for sampling contingency tables} Our final application involves contingency tables. Let $\mathbf{r}=(r_1,\ldots, r_m)$ and $\mathbf{c}=(c_1,\ldots, c_n)$ be two vectors of positive integers with the same sum. A \emph{contingency table} with row sums $r$ and column sums $c$ is an $m\times n$ matrix $X=(x_{i,j})$ with nonnegative integer entries, such that $\sum_{j=1}^n x_{i,j} = r_i$ for $i=1,\ldots, m$ and $\sum_{i=1}^m x_{i,j} = c_j$ for $j=1,\ldots, n$. Let $\Omega_{\mathbf{r},\mathbf{c}}$ denote the set of all contingency tables with row sums $\mathbf{r}$ and column sums $\mathbf{c}$. To avoid trivialities we assume throughout this section that $\min\{ m,n\}\geq 2$. Dyer and Greenhill~\cite{DG-contingency} proposed a Markov chain for sampling contingency tables, which we will call the \emph{contingency chain}. A transition of the chain is performed as follows: choose a $2\times 2$ subsquare of the current table uniformly at random, then replace this $2\times 2$ subsquare by a uniformly chosen $2\times 2$ nonnegative integer matrix with the same row and column sums. The \emph{lazy} contingency chain does nothing at each step with probability $\nfrac{1}{2}$, and otherwise performs a transition as described above. Cryan et al.~\cite{CDGJM} analysed the lazy contingency chain for a constant number of rows. They proved that $(1-\lambda_1)^{-1}\leq n^{f(m)}$ for $m$-rowed contingency tables with $n$ columns, where $m$ is constant and $f(m)$ is an expression satisfying $f(m)\geq 68m^4$. We now analyse the smallest eigenvalue of the (non-lazy) contingency chain. There is always a positive probability that the next state $X'$ of the contingency chain is equal to the current state $X$, since the heat-bath step may simply replace the chosen $2\times 2$ subsquare with its current contents. However, the minimum of $P(X,X)$ over all states $X$ depends on $\mathbf{r}$ and $\mathbf{c}$. (To see this, consider $2\times 2$ squares.) We prefer a bound which depends only on $m$ and $d$, and so we do not simply apply (\ref{all-loops}). \begin{lemma} \label{non-lazy-contingency} Let $\mathbf{r}=(r_1,\ldots, r_m)$ and $\mathbf{c}=(c_1,\ldots, c_n)$ be vectors of positive integers with a common sum which satisfy \[ r_1\geq r_2\geq \cdots \geq r_m\quad \text{ and } \quad c_1\geq c_2 \geq \cdots \geq c_n.\] Suppose that $\min\{r_1,\, c_1\}\geq 2$ and $\max\{ m,n\}\geq 3$. The smallest eigenvalue of the contingency chain on $\Omega_{\mathbf{r},\mathbf{c}}$ satisfies \[ (1+\lambda_{N-1})^{-1} \leq 45\, m^3n^3. \] \end{lemma} \begin{proof} Write $[a] = \{ 1,2,\ldots, a\}$ for $a\in\mathbb{Z}^+$. From $X = (x_{i,j})\in\Omega_{\mathbf{r},\mathbf{c}}$, first suppose that there exists a 5-tuple $(i_1,i_2,i_3,j_1,j_2)$ such that \begin{itemize} \item $i_1,i_2,i_3$ are distinct elements of $[m]$, \item $j_1,j_2$ are distinct elements of $[n]$, \item $x_{i_1,j_1},\, x_{i_2,j_1},\, x_{i_3,j_2}$ are all positive. \end{itemize} Then $(i_1,i_2,i_3,j_1,j_2)$ is called \emph{row-good for $X$}, and $X$ is called \emph{row-good}. If $X$ is row-good, fix the lexicographically least 5-tuple $(i_1,i_2,i_3,j_1,j_2)$ which is row-good for $X$ and consider the following sequence of three transitions on the $3\times 2$ subsquare defined by rows $i_1,i_2,i_3$ and columns $j_1,j_2$: \[ \begin{pmatrix} y_{1,1} & y_{1,2}\\ y_{2,1} & y_{2,2}\\ y_{3,1} & y_{3,2}\end{pmatrix} \,\, \Longrightarrow \,\, \begin{pmatrix} y_{1,1}-1 & y_{1,2}+1\\ y_{2,1} & y_{2,2}\\ y_{3,1}+1 & y_{3,2}-1\end{pmatrix} \,\, \Longrightarrow \,\, \begin{pmatrix} y_{1,1} & y_{1,2}\\ y_{2,1} -1 & y_{2,2}+1\\ y_{3,1}+1 & y_{3,2}-1\end{pmatrix} \,\, \Longrightarrow \begin{pmatrix} y_{1,1} & y_{1,2}\\ y_{2,1} & y_{2,2}\\ y_{3,1} & y_{3,2}\end{pmatrix}. \] (For notational convenience we have written $y_{k,\ell}$ for $x_{i_k,j_\ell}$ in the above.) Note that all intermediate matrices are nonnegative, due to the row-good property. This defines a walk $w_X$ of length 3 from $X$ to $X$ in the graph underlying the contingency chain. We can define 5-tuples $(i_1,i_2,j_1,j_2,j_3)$ which are \emph{column-good for $X$} in the analogous way, and say that $X$ is column-good if there is a 5-tuple which is column-good for $X$. If $X$ is column-good then taking the transpose of each matrix in the sequence of transitions above defines an odd walk $w_X$ of length 3 from $X$ to $X$. Finally, suppose that $X\in\Omega_{\mathbf{r},\mathbf{c}}$ is not row-good and is not column-good. Such an $X$ is said to be \emph{bad}. Then no row or column of $X$ contains more than one positive entry. Since all row and column sums are positive, it follows that $m=n\geq 3$ and that every row and column contains exactly one positive entry. Let $(i_1,i_2,i_3,j_1,j_2,j_3)$ be the lexicographically-least 6-tuple such that \begin{itemize} \item $i_1,i_2,i_3$ are distinct elements of $[m]$, \item $j_1,j_2,j_3$ are distinct elements of $[n]$, \item $x_{i_1,j_1} \geq 2$, while $x_{i_2,j_2}$ and $x_{i_3,j_3}$ are positive. \end{itemize} (The conditions on $\mathbf{r}$ and $\mathbf{c}$ guarantee that such a 6-tuple exists.) Consider the following sequence of 5 transitions, performed on the $3\times 3$ subsquare defined by rows $i_1,i_2,i_3$ and columns $j_1,j_2,j_3$: \begin{align*} \begin{pmatrix} y_{1,1} & 0 & 0\\ 0 & y_{2,2} & 0\\ 0 & 0 & y_{3,3}\end{pmatrix} & \Longrightarrow \,\, \begin{pmatrix} y_{1,1}-1 & 1 & 0\\ 1 & y_{2,2}-1 &0\\ 0 & 0 & y_{3,3}\end{pmatrix} \,\, \Longrightarrow \,\, \begin{pmatrix} y_{1,1}-1 & 1 & 0\\ 0 & y_{2,2}-1 &1\\ 1 & 0 & y_{3,3}-1\end{pmatrix} \\ & \Longrightarrow \,\, \begin{pmatrix} y_{1,1}-2 & 1 & 1\\ 1 & y_{2,2}-1 &0\\ 1 & 0 & y_{3,3}-1\end{pmatrix} \,\, \Longrightarrow \,\, \begin{pmatrix} y_{1,1}-1& 0 & 1\\ 0 & y_{2,2} &0\\ 1 & 0 & y_{3,3}-1\end{pmatrix} \\ & \Longrightarrow \,\, \begin{pmatrix} y_{1,1}& 0 & 0\\ 0 & y_{2,2} &0\\ 0 & 0 & y_{3,3}\end{pmatrix}. \end{align*} This defines a walk $w_X$ of length 5 from $X$ to $X$ in the graph underlying the chain. Now we must analyse the set $\mathcal{W} = \{ w_X : X\in\Omega_{\mathbf{r},\mathbf{c}}\}$ of odd walks defined above. Let $e=(Z,Z')$ be a transition of the contingency chain. Then $Z$ and $Z'$ only differ in a $2\times 2$ subsquare defined by rows $i,i'$ and columns $j,j'$. First we seek row-good $X$ with $e\in w_X$. Let $i''\not\in\{i,i'\}$ be another row index, and fix one of the 6 ways to arrange $(i,i',i'',j,j')$ as $(i_1,i_2,i_3,j_1,j_2)$. This gives enough information to uniquely identify a potential candidate for $X$. For example, if the transition $e$ involves rows $i_1$ and $i_3$ then $X=Z$, while if the transition $e$ involves rows $i_2$ and $i_3$ then $X=Z'$. If $e$ involves rows $i_1$ and $i_2$ then $e$ is the second transition in the sequence, and $X$ can be obtained from $Z$ by reversing the first transition in the sequence: namely, adding 1 to entries $(i_1,j_1)$ and $(i_3,j_2)$ and subtracting 1 from entries $(i_1,j_2)$ and $(i_3,j_1)$. If $X$ is a valid contingency table then $(i_1,i_2,i_3,j_1,j_2)$ is row-good for $X$. If it is the lexicographically least such 5-tuple for $X$ then $e\in w_X$. This identifies at most $12(m-2)$ tables $X$ such that $e\in w_X$. (This is an overcount, but good enough for our purposes.) By choosing a third column index $j''\not\in\{j,j'\}$, an analogous argument shows that there are at most $12(n-2)$ column-good tables $X$ with $e\in w_X$. Finally, we seek bad tables $X$ such that $e\in w_X$. Choose a row index $i''\not\in\{ i,i'\}$ and a column index $j''\not\in\{j,j'\}$, and fix one of the at most 36 ways to arrange $(i,i',i'',j,j',j'')$ as $(i_1,i_2,i_3,j_1,j_2,j_3)$. Now each transition in the sequence alters a different $2\times 2$ subsquare except the first and fourth, which both alter rows $i_1,i_2$ and columns $j_1,j_2$. Hence, arguing as above, there are at most two choices for $X$, for each fixed 6-tuple. This gives at most $72(m-2)(n-2)$ bad tables $X$ such that $e\in w_X$. Combining all this, we find that the congestion parameter $\eta(\mathcal{W})$ satisfies \[ \eta(\mathcal{W}) \leq \binom{m}{2}\,\binom{n}{2}\, \left(36(m-2)+36(n-2)+360(m-2)(n-2)\right) \leq 90\, m^3n^3, \] and applying Lemma~\ref{lazy} completes the proof. \end{proof} Again we observe that this bound on $(1+\lambda_{N-1})^{-1}$ is several orders of magnitude lower than the best-known bound on the second-largest eigenvalue~\cite{CDGJM}. \begin{remark} It has recently been shown~\cite{GU} that the contingency chain described above has no negative eigenvalues. We include Lemma~\ref{non-lazy-contingency} here to illustrate an application of Lemma~\ref{lazy} involving walks of length greater than one. \end{remark}

https://arxiv.org/abs/2302.07629

Automated Reasoning for Physical Quantities, Units, and Measurements in Isabelle/HOL

Formal verification of cyber-physical and robotic systems requires that we can accurately model physical quantities that exist in the real-world. The use of explicit units in such quantities can allow a higher degree of rigour, since we can ensure compatibility of quantities in calculations. At the same time, improper use of units can be a barrier to safety and therefore it is highly desirable to have automated sanity checking in physical calculations. In this paper, we contribute a mechanisation of the International System of Quantities (ISQ) and the associated SI unit system in Isabelle/HOL. We show how Isabelle can be used to provide a type system for physical quantities, and automated proof support. Quantities are parameterised by dimension types, which correspond to base vectors, and thus only quantities of the same dimension can be equated. Since the underlying "algebra of quantities" induces congruences on quantity and SI types, specific tactic support is developed to capture these. Our construction is validated by a test-set of known equivalences between both quantities and SI units. Moreover, the presented theory can be used for type-safe conversions between the SI system and others, like the British Imperial System (BIS).

\section{Introduction} \input{sec/intro} \section{Related Work} \label{sec:related} \input{sec/related} \section{Dimensions} \label{sec:dimensions} \input{sec/dimensions} \section{Physical Quantities and Measurement} \label{sec:quantities} \input{sec/quantities} \section{Unit Systems and the SI} \label{sec:units} \input{sec/units} \section{Unit Conversions and Non-SI Systems} \label{sec:conversions} \input{sec/conversions} \section{Conclusions} \label{sec:conclusions} \input{sec/conclusions} \bibliographystyle{IEEEtran} \subsection{Universe of Dimensions} We begin by defining the dimension universe, the core operators for constructing dimensions, and their properties. If we assume there are $n \in \nat$ base quantities, then a dimension has the form $d_1^{\vec{x}_1} \cdot d_2^{\vec{x}_2} \cdots d_n^{\vec{x}_n}$, a product of dimension symbols ($d_i$) each raised to a power drawn from the vector $\vec{x}$. The encoding of the dimension vector $\vec{x}$ in Isabelle is shown below. \begin{definition}[Dimension Vectors] $$\textbf{typedef}~(N, I)~\textit{dimvec} = (\textit{UNIV} :: (I::enum) \Rightarrow N)$$ \end{definition} \noindent The \textbf{typedef} command in Isabelle/HOL introduces a new type characterised by a non-empty subset of an existing type. In this case, we introduce a new type \textit{dimvec} with two parameters $N$ and $I$. We choose the set $UNIV$ of all total functions from $I$ to $N$ as the characteristic set. In future type definitions, we will use \textbf{typedef} to further restrict the subset. Conceptually, a dimension vector is simply a total function from an enumerable index type $I$, for the possible dimensions, to a numeric type $N$, which should minimally form a ring (e.g. $\num$). The enumerable ($enum$) sort constraint ($I :: enum$) requires that $I$ is isomorphic to a list of values, and is thus also a finite type. In the ISQ we have $I = \{\mathbf{L}, \mathbf{M}, \mathbf{T}, \mathbf{I}, \Theta, \mathbf{N}, \mathbf{J}\}$, for example. Next, we define the core dimension constructors. For simplicity, we present these as functions using $\lambda$-terms, but they are technically defined using the lifting package~\cite{Huffman13Lifting}. \begin{definition}[Dimension Constructors] $$\mathbf{1} \defs (\lambda i.~ 0) \quad \textbf{b}(i) \defs \mathbf{1}(i \mapsto 1)$$ $$\vec{x} \cdot \vec{y} \defs (\lambda i.~ \vec{x}(i) + \vec{y}(i)) \quad \vec{x}^{-1} \defs (\lambda i.~ - \vec{x}(i))$$ \end{definition} \noindent Here, $\mathbf{1}$ denotes a null dimension, which does not map to any physical quantity. It can characterise dimensionless quantities, such as mathematical constants ($\pi$, $e$, etc.) and functions. The function $\textbf{b}(i)$, for $i \in I$, constructs a \emph{base dimension} from the base quantity $i$ by updating the mapping for $i$ in $\mathbf{1}$ to have the power $1$. A base dimension has exactly one entry in the vector mapping to 1, with the others all 0. We also define a predicate $\textit{is-BaseDim} :: (N, I)~\textit{dimvec} \Rightarrow \mathbb{B}$, which determines whether a dimension vector corresponds to a base dimension. A product of two dimensions ($\vec{x} \cdot \vec{y})$ simply pointwise sums together all of the powers, and an inverse ($\vec{x}^{-1}$) negates each of the powers. We can also now obtain division using the usual definition: $\vec{x} / \vec{y} \defs \vec{x} \cdot \vec{y}^{-1}$. With these definitions, we can prove the following group theorem: \begin{theorem} If $(N, +, 0, -)$ forms an abelian group then also $((N, I)\textit{dimvec}, \cdot, \mathbf{1}, {}^{-1})$ forms an abelian group. \end{theorem} \noindent The abelian group laws can therefore be used to equationally rewrite dimension expressions, which is automated using Isabelle's simplifier. Another avenue to efficient proof for dimensions is provided through the Isabelle code generator~\cite{Haftman2010-CodeGen}. Since the set of base quantities $I$ is enumerable, we can always convert a dimension vector to a list of $N$, and vice-versa. We achieve this using a function \textit{mk-dimvec}, which converts a list of $N$ with length $|\!I\!|$ to a dimension vector in $N$. \begin{definition}[Converting Lists to Dimensions] $ $ % \label{def:list-to-dim} \vspace{1ex} \noindent $$ \textit{mk-dvec}(ds) \defs \left(\begin{array}{l} \mathbf{if}~ (length(ds) = ~|\!I\!|) \\ \mathbf{then}~ (\lambda d.~ ds(\textit{enum-ind}(d))) ~\mathbf{else}~ \mathbf{1}\end{array}\right) $$ \end{definition} \noindent Since $I$ is enumerable, every dimension can be assigned a natural number, which also denotes its position in the underlying list. The function $\textit{enum-ind} :: (I\!::\!enum) \Rightarrow \nat$ extracts this positional index of a value in an enumerable type. For ISQ, we have $\textit{enum-ind}(\mathbf{L}) = 0$ and $\textit{enum-ind}(\mathbf{T}) = 2$, for example. We can then construct a dimension from a list $ds$ simply by looking up the value at the enumeration index. Every possible dimension can be constructed using \textit{mk-dvec}, and so we can use it as a so-called ``code datatype'' for the Isabelle code generator. Dimensions are then encoded in SML or Haskell as an algebraic datatype with a single constructor corresponding to \textit{mk-dvec}, for example: \begin{alltt} \textbf{datatype} ('a, 'b) dimvec = Mk_dimvec of 'a list \end{alltt} \noindent We then prove code equation theorems for the group operators, which are homomorphism laws, and enable efficient execution: \begin{theorem} For a dimension vector space $(N, I)~dimvec$, with $|\!xs\!| \,=\, |\!ys\!| \,=\, |\!I\!|$, the following code equations hold: \begin{align*} \mathbf{1} &= \textit{mk-dvec}\,(\textit{replicate}\,|\!I\!|\,0) \\ \textit{mk-dvec}(xs) \cdot \textit{mk-dvec}(ys) &= map~(\lambda (x, y).\, x + y)~(zip~xs~ys) \\ (\textit{mk-dvec}(xs))^n &= \textit{mk-dvec}~(map~(\lambda x.\, n \cdot x)~xs \\ (\textit{mk-dvec}(xs))^{-1} &= \textit{mk-dvec}~(map~(\lambda x.\, - x)~xs \end{align*} \end{theorem} \noindent These theorems give concrete definitional equations for the executable functions on the datatype. The null dimension is a list of $0$ powers of length $|\!I\!|$. Multiplication of two equilength lists $xs$ and $ys$ is pairwise addition of each element. Raising to the $n$th power multiplies each list element by $n$. Taking the inverse power negates each element. With such equations we can perform efficient dimension arithmetic on dimensions constructed from lists. Dimensions in the ISQ are represented using the concrete dimension index type \textit{sdim}: \begin{definition}[ISQ Base Quantities] \label{def:isq-base-quant} \begin{align*} \textbf{datatype}~ sdim &= Length | Mass | Time | Current \\ &| Temperature | Amount | Intensity \\[1ex] \textbf{type-synonym}~ Dimension &= (\num, sdim)~dimvec \end{align*} \end{definition} \noindent It suffices to show that \textit{sdim} is enumerable, using a type class instantiation, and then we can create a specific type synonym \textit{Dimension}, for dimension vectors in the ISQ. For convenience, we then define dimension vectors for each of the base quantities, for example $\mathbf{L} \defs \mathbf{b}(Length)$. \subsection{Dimension Types} \label{sec:dim-types} Having defined our dimension universe, the next step is to characterise the family of dimension types. These dimension types will be used to parameterise our quantities, and ensure only quantities of the type dimension may be compared. We avoid the need for dependent types by first introducing a type class for dimension types. \begin{definition}[Dimension Type Classes] $ $ % \vspace{1ex} $\begin{array}{l} \textbf{class}~\textit{dim-type} = unitary~ + \\ ~~ \textbf{fixes}~\textit{dim-ty-sem} :: \mathcal{D}~itself \Rightarrow Dimension \\[2ex] \textbf{class}~\textit{basedim-type} = \textit{dim-type}~ + \\ ~~ \textbf{assumes}~ \textit{is-BaseDim}\!: \textit{is-BaseDim}(QD(\mathcal{D})) \end{array} $ \end{definition} \noindent A type class characterises a family of types that each implement a given function signature with certain properties, such as algebraic structures like monoids and groups. The \textbf{class} command introduces a type class with a given name, potentially extending existing classes. The \textbf{fixes} subcommand declares a new typed symbol in the signature, and \textbf{assumes} introduces a property of the symbols in the signature. The \textit{dim-type} class characterises a $unitary$ type $\mathcal{D}$ (i.e. a type with cardinality 1) and associates it to a particular dimension. The type $\mathcal{D}~itself$ represents a type as a value in Isabelle/HOL. Thus, \textit{dim-ty-sem} can be seen as a function from types inhabiting the \textit{dim-type} class to particular dimensions, as shown in Figure~\ref{fig:dim-mapping}. We can use the syntactic constructor $TYPE(\alpha)$ to obtain a value of type $\alpha~itself$, for a particular type $\alpha$. This effectively introduces an isomorphism between dimensions at the value level and the type level. For convenience, we introduce the notation $QD(\mathcal{D}) \defs \textit{dim-ty-sem}~\textit{TYPE}(\mathcal{D})$, which obtains the dimension of a given dimension type. The class \textit{basedim-type} further specialises \textit{dim-type} by requiring that the mapped dimension is a base dimension. We use these classes to capture the set of type constructors for dimension types. First we construct types to denote the base dimensions, as unitary types. For example, we define the type length as below: $$\textbf{typedef}~Length = (UNIV :: unit~set)$$ which exploits the fact that a type definition generates a fresh type name from a set (in this case, the set that just contains the only element of the \textit{unit} type). Though there is a seeming clash with the \textit{Length} constructor introduced in \cref{def:isq-base-quant}, these names inhabit different name spaces. \textit{Length} here is a ``tag type'' whose members do not convey information, but represent dimension types syntactically. We define seven such types, one for each of the ISQ base quantities, and also a further special type called \textbf{1}, which corresponds to a dimensionless quantity. Each of the base dimensions instantiates the \textit{basedim-type} class by mapping to the corresponding dimension symbol introduced in the previous section, such that, for example, $QD(Length) = Length$. Next, we introduce the arithmetic operators for dimensions at the type level. The product and inverse type constructors are defined as shown below: \begin{align*} &\textbf{typedef}~(\mathcal{D}_1\!::\!\textit{dim-ty},\mathcal{D}_2\!::\!\textit{dim-ty})~\textit{DimTimes} = (\textit{UNIV} :: unit set) \\ &\textbf{typedef}~(\mathcal{D}\!::\!\textit{dim-ty})~DimInv = (\textit{UNIV} :: unit set) \end{align*} They are similarly tag types, but the parameters must inhabit the \textit{dim-ty} class. This ensures that the dimension types are closed under products and inverse. Using these type constructors, and the base dimension types, we can inductively define algebraic dimensions at the type level. We assign the type constructors the following implementations of \textit{dim-ty-sem}: \begin{definition}[Semantic Interpretation of Dimension Types] \label{def:sem-dim} \begin{align*} \textit{dim-ty-sem}(d :: (\mathcal{D}_1, \mathcal{D}_2)~\textit{DimTimes}) &= QD(\mathcal{D}_1) \cdot QD(\mathcal{D}_2) \\ \textit{dim-ty-sem}(d :: (\mathcal{D})~\textit{DimInv}) &= QD(\mathcal{D})^{-1} \end{align*} \end{definition} These link together the type constructors and the underlying dimension operators. The semantics of a \textit{DimTimes} type calculates the underlying value-level dimension of each parameter $\mathcal{D}_1$ and $\mathcal{D}_2$, and multiplies them together. The \textit{DimInv} type similarly calculates the dimension and then takes the inverse. We give these type constructors the usual mathematical syntax, so that we can write dimension types like $\mathbf{M} \cdot \mathbf{L}$ and $\mathbf{T}^{-1}$. We also define a type synonym for division, namely $(\mathcal{D}_1, \mathcal{D}_2)~\textit{DimDiv} \defs \mathcal{D}_1 \cdot \mathcal{D}_2^{-1}$, and give it the usual syntax. Moreover, we define a fixed number of powers and inverse powers at the type level, such as $\mathcal{D}^{-3} = (\mathcal{D} \cdot \mathcal{D} \cdot \mathcal{D})^{-1}$. We can now also create the set of derived dimensions specified in the ISQ using type synonyms. For example, we define $\textit{Velocity} \defs \mathbf{L} \cdot \mathbf{T}^{-1}$ and $\textit{Pressure} \defs \mathbf{L}^{-1} \cdot \mathbf{M} \cdot \mathbf{T}^{-1}$, which provides a terminology of dimensions for use in formal specifications. We show further example in Figure~\ref{fig:derived-dimensions}, which also demonstrates the mathematical syntax for dimensions implemented in Isabelle/HOL. \begin{figure}[t] \centering \includegraphics[width=.6\linewidth]{figures/Derived_Dimensions.png} \caption{Derived dimension type expressions in Isabelle/HOL} \label{fig:derived-dimensions} \vspace{-1ex} \end{figure} \subsection{Dimension Normalisation} \label{sec:dim-norm} Unlike dimensions at the value level, dimension types with different syntactic forms are incomparable, because they are distinct type expressions. For example, it is intuitively a fact that $\mathbf{L} \cdot \mathbf{T}^{-1} \cdot \mathbf{T} = \mathbf{L}$, which can be proved using the group laws. However, at the type level $\mathbf{L} \cdot \mathbf{T}^{-1} \cdot \mathbf{T}$ and $\mathbf{L}$ are different type expressions, and no built-in normalisation is available in Isabelle. As a result, we need to implement our own normalisation function, $\textit{normalise}(\mathcal{D})$, for ISQ dimensions in Isabelle/ML, so that quantities over dimensions with distinct syntactic forms can be related. Our normalisation function evaluates the dimension vector of a dimension expression, and then uses this to produce a normal form. We implement dimension type evaluation using the ML function $\textit{typ-to-dim} :: typ \Rightarrow int~list$. It converts types formed of the base dimensions and dimension arithmetic operators into a dimension vector list, using the representation given in \cref{def:list-to-dim}. For example, $\textit{typ-to-dim}(\mathbf{L}) = [1, 0, 0, 0, 0, 0, 0]$ and $\textit{typ-to-dim}(\mathcal{D}^{-1}) = map~(\lambda x.\, - x)~(\textit{typ-to-dim}(x))$. Having evaluated the dimension expression, we can use it to construct the normal form. This is an ordered dimension expression of the form $\mathbf{L}^{\vec{x}_1} \cdot \mathbf{M}^{\vec{x}_2} \cdot \mathbf{T}^{\vec{x}_3} \cdots \mathbf{J}^{\vec{x}_7}$, except that we omit terms where $\vec{x}_i = 0$. If every such term is $0$, then the function produces the dimensionless quantity, $\mathbf{1}$. As an example, $\textit{normalise}(\mathbf{T}^4 \cdot \mathbf{L}^{-2} \cdot \mathbf{M}^{-1} \cdot \mathbf{I}^2 \cdot \mathbf{M})$ yields the dimension type $\mathbf{L}^{-2} \cdot \mathbf{T}^4 \cdot \mathbf{I}^2$. This normalisation function is used later in this paper to facilitate coercion between quantities with distinct dimension expressions. \subsection{Quantity Universe and Measurement Systems} \label{sec:quant-univ} We specify our quantity universe as a record with fields for the magnitude and dimension of the quantity. \begin{definition}[Quantity Universe] $$\begin{array}{rl} \textbf{record}~ (N, I\!::\!enum)\, Quantity =&\hspace{-1.5ex} mag\!::\!N\ \\ &\hspace{-1.5ex} dim\!::\!(int, I)\,dimvec \end{array}$$ \end{definition} \noindent The \textit{Quantity} type is parametric over a numeric type $N$ (e.g. $\mathbb{Q}$, $\mathbb{R}$), which should form a field, and the dimension index type $I$. The magnitude is then a number in $N$, and a dimension vector in over $I$. We can now specify the core arithmetic operators on quantities. For convenience of presentation, we use tuple syntax $(x, \mathcal{D})$, though in Isabelle the record fields are used. \begin{definition}[Quantity Arithmetic Operators] \label{def:quant-arith} \begin{align*} 0 &\defs (0, \mathbf{1}) \\ 1 &\defs (1, \mathbf{1}) \\ (x, \mathcal{D}_1) \cdot (y, \mathcal{D}_2) &= (x \cdot y, \mathcal{D}_1 \cdot \mathcal{D}_2) \\ (x, \mathcal{D})^{-1} &= (x^{-1}, \mathcal{D}^{-1}) \\ (x, \mathcal{D}_1) / (y, \mathcal{D}_2) &= (x / y, \mathcal{D}_1 / \mathcal{D}_2) \\ (x, \mathcal{D}) + (y, \mathcal{D}) &= (x + y, \mathcal{D}) \\ (x, \mathcal{D}) - (y, \mathcal{D}) &= (x - y, \mathcal{D}) \\ (x, \mathcal{D}_1) \le (y, \mathcal{D}_2) &\iff (x \le y \land \mathcal{D}_1 = \mathcal{D}_2) \end{align*} \end{definition} \noindent The arithmetic operators are overloaded in Isabelle/HOL, which is why they can validly appear on both sides of these equations. The ``0'' and ``1'' quantities are specified as dimensionless quantities with magnitude 0 and 1, respectively. Multiplication, inverse, and division are total operations that simply distribute through the pair. When multiplying two quantities, we need to multiply both the magnitudes and dimensions. For example, $(7, \mathbf{L}\cdot\mathbf{T}^{-1}) \cdot (2, \mathbf{T}) = (14, \mathbf{L})$. In contrast, addition and subtraction are partial operators that may be applied only when the two quantities have the same dimension. In Isabelle/HOL, the value of an addition or subtraction for different quantities of different dimensions is unspecified. Finally, the order on quantities is simply the order on the magnitudes, but with the requirement that the two dimensions are equal. Quantities as formalised so far specify the form of dimension, but not the system of units being employed. For this, we extend the \textit{Quantity} type to create ``measurement systems'': \begin{definition}[Measurement Systems] $$ \begin{array}{l} \textbf{record} (N, I\!::\!enum, \mathcal{S}\!::\!\textit{unit-system})~\textit{Measurement-System} \\ \quad = (N, I)~Quantity + \textit{unit-sys} :: \mathcal{S} \end{array} $$ \end{definition} \noindent We extend the \textit{Quantity} record with an additional field \textit{unit-sys}. A measurement system is a quantity that specifies the system of units being used via an additional type parameter $\mathcal{S}$, which must inhabit the type class \textit{unit-system}. A unit system type is a unitary type that effectively allows us to tag quantities. This allows us to distinguish quantities using different systems of units, and so prevent improper mixing. For example, the presence of the \textit{SI} tag means that a quantity of length is fundamentally measured in metres, whereas the presence of a tag such as \textit{BIS} may indicate that length is measured in yards. Later, we will use these to facilitate type-safe conversions between different unit systems. All the arithmetic operators can be straightforwardly lifted to measurement systems. Since all such functions are monomorphic (e.g. of type $\alpha \Rightarrow \alpha \Rightarrow \alpha$), mixing of systems is avoided by construction. \subsection{Dimension Typed Quantities} Having defined our universe for quantities, we next enrich this representation with type-level dimensions. For expediency, we assume that all such quantities also have a measurement system attached. Moreover, we focus on quantities with dimensions from the ISQ. \begin{definition}[Quantity Type] $$ \begin{array}{l} \textbf{typedef}~ (N, \mathcal{D}\!::\!\textit{dim-type}, \mathcal{S}::\textit{unit-system})~\textit{QuantT} \\ \quad = \{x :: (N, sdim, S)~\textit{Measurement-System}.~~ dim(x) = QD(\mathcal{D})\} \end{array} $$ \end{definition} \noindent The $(N, \mathcal{D}, \mathcal{S})~\textit{QuantT}$ type represents a quantity with numeric type $N$, dimension type $\mathcal{D}$, and unit system $\mathcal{S}$. The type definition introduces an invariant that requires that the dimension of the underlying quantity $x$ agrees with the one specified in the dimension type. At this level, we use \textit{sdim} as the concrete interpretation of dimensions, as this is required by the \textit{dim-type} class. For convenience, we introduce the type syntax $N[\mathcal{D}, \mathcal{S}]$ to stand for $(N, \mathcal{D}, \mathcal{S})~\textit{QuantT}$. Our \textbf{typedef} also induces two functions for converting between typed and untyped quantitities: $\textit{fromQ} :: N[\mathcal{D}, \mathcal{S}] \Rightarrow (N, \textit{sdim}, \mathcal{S})~\textit{Measurement-System}$ and \textit{toQ} in the opposite direction. Lifting of arithmetic operators $x + y$ and $x - y$ is straightforward for typed quantities, since they are monomorphic and only defined when the dimensions of $x$ and $y$ agree. We can then easily show that typed quantities form an additive abelian group. We also define a scalar multiplication $\textit{scaleQ} :: N \Rightarrow N[\mathcal{D}, \mathcal{S}] \Rightarrow N[\mathcal{D}, \mathcal{S}]$, with notation $n \mathop{*_\qsub} x$, which scales a quantity by a given number without changing the dimension. We can then show that typed quantities form an additive abelian group, and a real vector space, with $(\mathop{*_\qsub})$ as the scalar multiplication operator. Things are more involved when dealing with general multiplication and division, since these need to perform dimension arithmetic at the type level. For example, if we have quantities $x :: \mathbb{R}[\mathbf{I}, SI]$ and $y :: \mathbb{R}[\mathbf{T}, SI]$, then multiplication of $x$ and $y$ is well-defined, and should have the type $\mathbb{R}[\mathbf{I}\cdot\mathbf{T}, SI]$. As a result, we introduce bespoke functions, $qtimes$, $qinverse$, and $qdivide$. We first give the types for these functions: \begin{align*} qtimes &:: N[\mathcal{D}_1, \mathcal{S}] \Rightarrow N[\mathcal{D}_2, \mathcal{S}] \Rightarrow N[\mathcal{D}_1 \cdot \mathcal{D}_2, \mathcal{S}] \\ qinverse &:: N[\mathcal{D}, \mathcal{S}] \Rightarrow N[\mathcal{D}^{-1}, \mathcal{S}] \\ qdivide &:: N[\mathcal{D}_1, \mathcal{S}] \Rightarrow N[\mathcal{D}_2, \mathcal{S}] \Rightarrow N[\mathcal{D}_1 / \mathcal{D}_2, \mathcal{S}] \end{align*} \noindent The first function multiplies two quantities, with the same measurement system, and ``multiplies'' the dimension types using the type constructors introduced in \S\ref{sec:dim-types}. Technically, no multiplication computation takes place, but rather a type constructor denoting multiplication is inserted. Similarly, \textit{qinverse} represents the inverse of the parametrised dimension, and \textit{qdivide} stands for a division. What is achieved here is analogous to dependent types, though we require additional machinery for normalising dimension types (cf. \S\ref{sec:dim-norm}). The definitions of \textit{qtimes} and \textit{qinverse} are obtained simply by lifting of the corresponding functions on quantities in Definition~\ref{def:quant-arith}, which is technically achieved using the \emph{lifting} package~\cite{Huffman13Lifting}. In order to do this we need to prove that the invariant of the \textit{QuantT} type is satisfied, which involves showing that the family of typed quantities is closed under the two functions. For \textit{qmult}, we need to prove that $dim(x \cdot y) = QD(\mathcal{D}_1 \cdot \mathcal{D}_2)$, whenever $dim(x) = QD(\mathcal{D}_1)$ and $dim(y) = QD(\mathcal{D}_2)$, which follows simply through Definitions \ref{def:sem-dim} and \ref{def:quant-arith}. For convenience, we give these functions the usual notation of $x \bullet y$, $x^{-1}$, and $x / y$, but in Isabelle we embolden the operators to syntactically distinguish them. With \textit{qtimes} and \textit{qinverse}, we can also define positive and negative powers, such as $x^{-2} = (x \bullet x)^{-1}$. Equality ($x = y$) in HOL is a homogeneous function of type $\alpha \to \alpha \to \mathbb{B}$; therefore, it cannot be used to compare objects of different types. Consequently, it cannot be used to compare quantities whose dimension types have different syntactic forms (e.g. $\mathbf{L} \cdot \mathbf{T}^{-1} \cdot \mathbf{T}$ and $\mathbf{L}$). This motivates a definition of \emph{heterogeneous (in)equality} for quantities: \begin{align*} \textit{qequiv} &:: N[\mathcal{D}_1, \mathcal{S}] \Rightarrow N[\mathcal{D}_2, \mathcal{S}] \Rightarrow \mathbb{B} \\ \textit{qless-eq} &:: N[\mathcal{D}_1, \mathcal{S}] \Rightarrow N[\mathcal{D}_2, \mathcal{S}] \Rightarrow \mathbb{B} \end{align*} \noindent These functions are defined simply by lifting the functions $(=)$ and $(\le)$ on the underlying quantities. They ignore the dimension types, but the underlying dimensions must nevertheless be the same as per the definitions in \S\ref{sec:quant-univ}. We give these functions the notation $x \cong y$ and $x \lesssim y$, respectively. Relation $(\cong)$ forms an equivalence relation, and $(\lesssim)$ forms a preorder. Moreover, $(\cong)$ is a congruence relation for $(\bullet)$, $({}^{-1})$, and $(\mathop{*_\qsub})$. \subsection{Proof Support} We implement an interpretation-based proof strategy for typed quantity (in)equalities, which allows us to split a conjecture into two parts: (1) equality of the magnitudes; and (2) equivalence of the dimensions. This is supported by a function $\textit{magQ} :: N[\mathcal{D}, \mathcal{S}] \Rightarrow N$, with syntax $\qsem{-}$, which extracts the magnitude from a typed quantity. We can calculate magnitudes using interpretation laws, like the ones below: $$\qsem{x + y} = \qsem{x} + \qsem{y} \quad \qsem{x \bullet y} = \qsem{x} \cdot \qsem{y}$$ \noindent Such laws derive directly from the definition of the quantity operators in Definition~\ref{def:quant-arith}. The equation for addition implicitly makes use of the fact that $x$ and $y$ have the same dimension, and so addition is well-defined in the quantity universe. We then have the following transfer theorems, for the case of two quantitites with the same (syntactic) dimensions: \begin{theorem}[Quantity Transfer Laws] $$x = y \iff (\qsem{x} = \qsem{y}) \qquad x \le y \iff (\qsem{x} \le \qsem{y})$$ \end{theorem} \noindent In both cases, we need not check the equivalence of the dimensions as by construction we know that $x$ and $y$ have the same type, and so also have the same dimensions. It is sufficient simply to check the relation holds of the underlying magnitudes. For our heterogeneous (in)equality relations, we have the following transfer theorems: \begin{theorem}[Heterogeneous Transfer Laws] Given quantities $x :: N[\mathcal{D}_1, \mathcal{S}]$ and $y :: N[\mathcal{D}_2, \mathcal{S}]$, we have $$x \cong y \iff \left(\qsem{x} = \qsem{y} \land QD(\mathcal{D}_1) = QD(\mathcal{D}_2)\right).$$ \end{theorem} \noindent We can then prove heterogeneous equalities by calculation of the underlying magnitudes and dimensions, and use of the numeric and dimensions laws. We supply a proof method called \textit{si-simp}, which uses the simplifier to perform transfer and interpretation, and additionally invokes field simplification laws. An additional method called \textit{si-calc} also compiles dimension vectors (cf. Definition~\ref{def:list-to-dim}) using the code generator, and can thus efficiently prove dimension equalities. We can, for example, prove the following algebraic laws automatically: \begin{theorem}[Quantity Algebraic Laws] \begin{align*} a \mathop{*_\qsub} (x + y) &= (a \mathop{*_\qsub} x) + (a \mathop{*_\qsub} y) \\ x \bullet y &\cong y \bullet x \\ (x \bullet y)^{-1} &\cong x^{-1} \bullet y^{-1} \end{align*} \end{theorem} \subsection{Coercion and Dimension Normalisation} The need for heterogeneous quantity relations ($\cong$, $\lesssim$) can be avoided by the use of coercions to convert between two syntactic representations of the same dimension. Moreover, we can use Isabelle's sophisticated syntax and checking pipeline to normalise dimensions, and so automatically coerce quantities to a normal form. This improves the usability of the library, since the usual relations $(=)$ and $(\le)$ can be used directly. We implement a function $\textit{dnorm} :: N[\mathcal{D}_1, \mathcal{S}] \Rightarrow N[\mathcal{D}_2, \mathcal{S}]$, which can convert between quantities with different dimension forms. In order to use it effectively, it is necessary to know the target dimension $\mathcal{D}_2$ in advance. It is defined below: $$\textit{dnorm}(x) \defs (\mathbf{if}~QD(\mathcal{D}_1) = QD(\mathcal{D}_2)~\mathbf{then}~toQ~(\,fromQ(x))~\mathbf{else}~0)$$ The function checks whether the source and target dimensions ($\mathcal{D}_1$ and $\mathcal{D}_2$) are the same. If they agree, then it performs the coercion by erasing the types with \textit{fromQ} and reinstating the new dimension type with \textit{toQ}. Otherwise, it returns a valid quantity of the target dimension, but with magnitude $0$. For example, if we have $x :: \mathbb{R}[\mathbf{L} \cdot \mathbf{T}^{-1} \cdot \mathbf{T}, SI]$, then we can use $\textit{dnorm}(x) :: \mathbb{R}[\mathbf{L}, SI]$ to obtain a quantity with an equivalent dimension, since $QD(\mathbf{L} \cdot \mathbf{T}^{-1} \cdot \mathbf{T}) = QD(\mathbf{L})$. In general, for two equivalent quantities $x \cong y$, we have it that $\textit{dnorm}(x) = y$. Next, we extend Isabelle's checking pipeline to allow dimension normalisation, so that $\mathcal{D}_2$ can be automatically calculated. We do this by implementing an SML function \textit{check-quant}, which takes a term and enriches it with dimension information. Whenever it encounters an instance of $\textit{dnorm}(t)$, it extracts the type of $t$, which should be $N[\mathcal{D}, \mathcal{S}]$. This being the case, we enrich the instance of \textit{dnorm} to have the type $N[\mathcal{D}, \mathcal{S}] \Rightarrow N[\textit{normalise}(\mathcal{D}), \mathcal{S}]$. We then insert \textit{check-quant} into Isabelle's term checking pipeline. Technically, this is achieved using an Isabelle/ML API function called \texttt{Syntax\_Phases.term\_check}, which allow us to add a new phase into the term checking process. In this case, we add it after type inference has occurred so that we can use the unnormalised dimension type expression as an input to \textit{check-quant}. The soundness of this transformation does not depend on the correctness of \textit{normalise}, since if an incorrect dimension is calculated, \textit{dnorm} will return 0. Nevertheless, the effect is to achieve something akin to dependent types, but in a first-order polymorphic type system.

https://arxiv.org/abs/1212.4551

Condition Numbers of Random Toeplitz and Circulant Matrices

Estimating the condition numbers of random structured matrices is a well known challenge, linked to the design of efficient randomized matrix algorithms. We deduce such estimates for Gaussian random Toeplitz and circulant matrices. The former estimates can be surprising because the condition numbers grow exponentially in n as n grows to infinity for some large and important classes of n-by-n Toeplitz matrices, whereas we prove the opposit for Gaussian random Toeplitz matrices. Our formal estimates are in good accordance with our numerical tests, except that circulant matrices tend to be even better conditioned according to the tests than according to our formal study.

\section{Introduction}\label{sintro} It is well known that random matrices tend to be well conditioned \cite{D88}, \cite{E88}, \cite{ES05}, \cite{CD05}, \cite{SST06}, \cite{B11}, and this property can be exploited for advancing matrix computations (see e.g., \cite{PGMQ}, \cite{PIMR10}, \cite{PQ10}, \cite{PQ12}, \cite{PQa}, \cite{PQZa}, \cite{PQZb}, \cite{PQZC} \cite{PY09}). Exploiting matrix structure in these applications was supported empirically in the latter papers and formally in \cite{T11}. An important step in this direction is the estimation of the condition numbers of structured matrices stated as a challenge in \cite{SST06}. We reply to this challenge by estimating the condition numbers of Gaussian random Toeplitz and circulant matrices both formally (see Sections \ref{scgrtm} and \ref{scgrcm}) and experimentally (see Tables \ref{nonsymtoeplitz}--\ref{tabcondcirc}). Our study shows that Gaussian random Toeplitz circulant matrices do not tend to be ill conditioned and the condition numbers of Gaussian random circulant $n\times n$ matrices tend to grow extremely slow as $n$ grows large. Our numerical tests (the contribution of the second author) are in good accordance with our formal estimates, except that circulant matrices tended to be even better conditioned in the tests than according to our formal study. Our results on Toeplitz matrices are quite surprising because the condition numbers grow exponentially in $n$ as $n\rightarrow \infty$ for some large and important classes of $n\times n$ Toeplitz matrices \cite{BG05}, which is opposit to the behavior of Gaussian random Toeplitz $n\times n$ matrices as we proved and consistently observed in our tests. Clearly, our study of Toeplitz matrices can be equally applied to Hankel matrices. We organize our paper as follows. We recall some definitions and basic results on general matrix computations in the next section and on Toeplitz, Hankel and circulant matrices in Section \ref{stplc}. We define Gaussian random matrices and study their ranks and extremal singular values in Section \ref{srvrm}. In Sections \ref{scgrtm} and \ref{scgrcm} we extend this study to Gaussian random Toeplitz and circulant matrices, respectively. In Section \ref{sexp} we cover numerical tests, which constitute the contribution of the second author. In Section \ref{simprel} we recall some applications of random circulant and Toeplitz matrices, which provide some implicit empirical support for our estimates for their condition numbers. We end with conclusions in Section \ref{srel}. \section{Some definitions and basic results}\label{sdef} Except for Theorem \ref{thcpw} and its application in the proof of Theorem \ref{thcircsing} we work in the field $\mathbb R$ of real numbers. Next we recall some customary definitions of matrix computations \cite{GL96}, \cite{S98}. $A^T$ is the transpose of a matrix $A$. $||A||_h$ is its $h$-norm for $h=1,2,\infty$. We write $||A||$ to denote the 2-norm $||A||_2$. We have \begin{equation}\label{eqnorm12} \frac{1}{\sqrt m}||A||_1\le||A||\le \sqrt n ||A||_1,~~||A||_1=||A^T||_{\infty},~~ ||A||^2\le||A||_1||A||_{\infty}, \end{equation} for an $m\times n$ matrix $A$, \begin{equation}\label{eqnorm12inf} ||AB||_h\le ||A||_h||B||_h~{\rm for}~h=1,2,\infty~{\rm and~any~matrix~product}~AB. \end{equation} Define an {\em SVD} or {\em full SVD} of an $m\times n$ matrix $A$ of a rank $\rho$ as follows, \begin{equation}\label{eqsvd} A=S_A\Sigma_AT_A^T. \end{equation} Here $S_AS_A^T=S_A^TS_A=I_m$, $T_AT_A^T=T_A^TT_A=I_n$, $\Sigma_A=\diag(\widehat \Sigma_A,O_{m-\rho,n-\rho})$, $\widehat \Sigma_A=\diag(\sigma_j(A))_{j=1}^{\rho}$, $\sigma_j=\sigma_j(A)=\sigma_j(A^T)$ is the $j$th largest singular value of a matrix $A$ for $j=1,\dots,\rho$, and we write $\sigma_j=0$ for $j>\rho$. These values have the minimax property \begin{equation}\label{eqminmax} \sigma_j=\max_{{\rm dim} (\mathbb S)=j}~~\min_{{\bf x}\in \mathbb S,~||{\bf x}||=1}~~~||A{\bf x}||,~j=1,\dots,\rho, \end{equation} where $\mathbb S$ denotes linear spaces \cite[Theorem 8.6.1]{GL96}. \begin{fact}\label{faccondsub} If $A_0$ is a submatrix of a matrix $A$, then $\sigma_{j} (A)\ge \sigma_{j} (A_0)$ for all $j$. \end{fact} \begin{proof} \cite[Corollary 8.6.3]{GL96} implies the claimed bound where $A_0$ is any block of columns of the matrix $A$. Transposition of a matrix and permutations of its rows and columns do not change singular values, and thus we can extend the bounds to all submatrices $A_0$. \end{proof} $A^+=T_A\diag(\widehat \Sigma_A^{-1},O_{n-\rho,m-\rho})S_A^T$ is the Moore--Penrose pseudo-inverse of the matrix $A$ of (\ref{eqsvd}), and \begin{equation}\label{eqnrm+} ||A^+||=1/\sigma{_\rho}(A) \end{equation} for a matrix $A$ of a rank $\rho$. $\kappa (A)=\frac{\sigma_1(A)}{\sigma_{\rho}(A)}=||A||~||A^+||$ is the condition number of an $m\times n$ matrix $A$ of a rank $\rho$. Such matrix is {\em ill conditioned} if $\sigma_1(A)\gg\sigma_{\rho}(A)$ and is {\em well conditioned} otherwise. See \cite{D83}, \cite[Sections 2.3.2, 2.3.3, 3.5.4, 12.5]{GL96}, \cite[Chapter 15]{H02}, \cite{KL94}, and \cite[Section 5.3]{S98} on the estimation of the norms and condition numbers of nonsingular matrices. \section{Toeplitz, Hankel and $f$-circulant matrices}\label{stplc} A {\em Toep\-litz} $m\times n$ matrix $T_{m,n}=(t_{i-j})_{i,j=1}^{m,n}$ (resp. Hankel matrices $H=(h_{i+j})_{i,j=1}^{m,n}$) is defined by its first row and first (resp. last) column, that is by the vector $(t_h)_{h=1-n}^{m-1}$ (resp. $(h_g)_{g=2}^{m+n}$) of dimension $m+n-1$. We write $T_n=T_{n,n}=(t_{i-j})_{i,j=1}^{n,n}$ (see equation (\ref{eqtz}) below). ${\bf e}_i$ is the $i$th coordinate vector of dimension $n$ for $i=1,\dots,n$. The reflection matrix $J=J_n({\bf e}_n~|~\dots~|~{\bf e}_1)$ is the Hankel $n\times n$ matrix defined by its first column ${\bf e}_n$ and its last column ${\bf e}_1$. We have $J=J^T=J^{-1}$. A lower {\em triangular Toep\-litz} $n\times n$ matrix $Z({\bf t})=(t_{i-j})_{i,j=1}^n$ (where $t_k=0$ for $k<0$) is defined by its first column ${\bf t}=(t_h)_{h=0}^{n-1}$. We write $Z({\bf t})^T=(Z({\bf t}))^T$. $Z=Z_0=Z({\bf e}_2)$ is the downshift $n\times n$ matrix (see (\ref{eqtz})). We have $Z{\bf v}=(v_i)_{i=0}^{n-1}$ and $Z({\bf v})=Z_0({\bf v})=\sum_{i=1}^{n}v_{i}Z^{i-1}$ for ${\bf v}=(v_i)_{i=1}^n$ and $v_0=0$, \begin{equation}\label{eqtz} T_n=\begin{pmatrix}t_0&t_{-1}&\cdots&t_{1-n}\\ t_1&t_0&\smallddots&\vdots\\ \vdots&\smallddots&\smallddots&t_{-1}\\ t_{n-1}&\cdots&t_1&t_0\end{pmatrix},~Z=\begin{pmatrix} 0 & & \dots & & 0\\ 1 & \ddots & & & \\ \vdots & \ddots & \ddots & & \vdots \\ & & \ddots & 0 & \\ 0 & & \dots & 1 & 0 \end{pmatrix}, ~Z_f=\begin{pmatrix} 0 & & \dots & & f\\ 1 & \ddots & & & \\ \vdots & \ddots & \ddots & & \vdots \\ & & \ddots & 0 & \\ 0 & & \dots & 1 & 0 \end{pmatrix}. \end{equation} Combine the equations $||Z({\bf v})||_1=||Z({\bf v})||_{\infty}=||{\bf v}||_1$ with (\ref{eqnorm12}) to obtain \begin{equation}\label{eqttn} ||Z({\bf v})||\le ||{\bf v}||_1. \end{equation} \begin{theorem}\label{thgs} Write $T_{k}=(t_{i-j})_{i,j=0}^{k-1}$ for $k=n,n+1$. (a) Let the matrix $T_n$ be nonsingular and write ${\bf p}=T_n^{-1}{\bf e}_1$ and ${\bf q}=T_n^{-1}{\bf e}_{n}$. If $p_{1}={\bf e}_1^T{\bf p}\neq 0$, then $p_{1}T_n^{-1}=Z({\bf p})Z(J{\bf q})^T-Z(Z{\bf q})Z(ZJ{\bf p})^T.$ In parts (b) and (c) below let the matrix $T_{n+1}$ be nonsingular and write $\widehat {\bf v}=(v_i)_{i=0}^n=T_{n+1}^{-1}{\bf e}_1$, ${\bf v}=(v_i)_{i=0}^{n-1}$, ${\bf v}'=(v_i)_{i=1}^{n}$, $\widehat {\bf w}=(w_i)_{i=0}^n=T_{n+1}^{-1}{\bf e}_{n+1}$, ${\bf w}=(w_i)_{i=0}^{n-1}$, and ${\bf w}'=(w_i)_{i=1}^{n}$. (b) If $v_0\neq 0$, then the matrix $T_n$ is nonsingular and $v_0T_n^{-1}=Z({\bf v})Z(J{\bf w'})^T-Z({\bf w})Z(J{\bf v}')^T$. (c) If $v_n\neq 0$, then the matrix $T_{1,0}=(t_{i-j})_{i=1,j=0}^{n,n-1}$ is nonsingular and $v_nT_{1,0}^{-1}=Z({\bf w})Z(J{\bf v'})^T-Z({\bf v})Z(J{\bf w}')^T$. \end{theorem} \begin{proof} See \cite{GS72} on parts (a) and (b); see \cite{GK72} on part (c). \end{proof} $Z_f=Z+f{\bf e}_1^T{\bf e}_n$ for a scalar $f\neq 0$ denotes the $n\times n$ matrix of $f$-{\em circular shift} (see (\ref{eqtz})). An $f$-{\em circulant matrix} $Z_f({\bf v})=\sum_{i=1}^{n}v_iZ_f^{i-1}$ is a special Toep\-litz $n\times n$ matrix defined by its first column vector ${\bf v}=(v_i)_{i=1}^{n}$ and a scalar $f$. $f$-circulant matrix is called {\em circulant} if $f=1$ and {\em skew circulant} if $f=-1$. By replacing $f$ with $0$ we arrive at a lower triangular Toep\-litz matrix $Z({\bf v})$. The following theorem implies that the inverses (wherever they are defined) and pairwise products of $f$-circulant $n\times n$ matrices are $f$-circulant and can be computed in $O(n\log n)$ flops. \begin{theorem}\label{thcpw} (See \cite{CPW74}.) We have $Z_1({\bf v})=\Omega^{-1}D(\Omega{\bf v})\Omega.$ More generally, for any $f\ne 0$, we have $Z_{f^n}({\bf v})=U_f^{-1}D(U_f{\bf v})U_f$ where $U_f=\Omega D({\bf f}),~~{\bf f}=(f^i)_{i=0}^{n-1}$, $D({\bf u})=\diag(u_i)_{i=0}^{n-1}$ for a vector ${\bf u}=(u_i)_{i=0}^{n-1}$, $\Omega=(\omega_n^{ij})_{i,j=0}^{n-1}$ is the $n\times n$ matrix of the discrete Fourier transform at $n$ points, $\omega_n={\rm exp}(\frac{2\pi}{n}\sqrt{-1})$ being a primitive $n$-th root of $1$, and $\Omega^{-1}=\frac{1}{n}(\omega_n^{-ij})_{i,j=0}^{n-1}=\frac{1}{n}\Omega^H$. \end{theorem} {\em Hankel} $m\times n$ matrices $H=(h_{i+j})_{i,j=1}^{m,n}$ can be defined equivalently as the products $H=TJ_n$ or $H=J_mT$ of $m\times n$ Toep\-litz matrices $T$ and the Hankel reflection matrices $J=J_m$ or $J_n$. Note that $J=J^{-1}=J^T$ and obtain the following simple fact. \begin{fact}\label{fath} For $m=n$ we have $T=HJ$, $H^{-1}=JT^{-1}$ and $T^{-1}=JH^{-1}$ if $H=TJ$, whereas $T=JH$, $H^{-1}=JT^{-1}$ and $T^{-1}=H^{-1}J$ if $H=JT$. Furthermore in both cases $\kappa (H)=\kappa (T)$. \end{fact} By using the equations above we can readily extend any Toep\-litz matrix inversion algorithm to Hankel matrix inversion and vice versa, preserving the flop count and condition numbers. E.g. $(JT)^{-1}=T^{-1}J$, $(TJ)^{-1}=JT^{-1}$, $(JH)^{-1}=H^{-1}J$ and $(HJ)^{-1}=JH^{-1}$. \section{Gaussian random matrices and their ranks}\label{srvrm} \begin{definition}\label{defcdf} $F_{\gamma}(y)=$ Probability$\{\gamma\le y\}$ (for a real random variable $\gamma$) is the {\em cumulative distribution function (cdf)} of $\gamma$ evaluated at $y$. $F_{g(\mu,\sigma)}(y)=\frac{1}{\sigma\sqrt {2\pi}}\int_{-\infty}^y \exp (-\frac{(x-\mu)^2}{2\sigma^2}) dx$ for a Gaussian random variable $g(\mu,\sigma)$ with a mean $\mu$ and a positive variance $\sigma^2$, and so \begin{equation}\label{eqnormal} \mu-4\sigma\le y \le \mu+4\sigma~{\rm with ~a ~probability ~near ~1}. \end{equation} \end{definition} \begin{definition}\label{defrndm} A matrix (or a vector) is a {\em Gaussian random matrix (or vector)} with a mean $\mu$ and a positive variance $\sigma^2$ if it is filled with independent identically distributed Gaussian random variables, all having the mean $\mu$ and variance $\sigma^2$. $\mathcal G_{\mu,\sigma}^{m\times n}$ is the set of such Gaussian random $m\times n$ matrices (which are {\em standard} for $\mu=0$ and $\sigma^2=1$). By restricting this set to Toeplitz or $f$-circulant matrices we obtain the sets $\mathcal T_{\mu,\sigma}^{m\times n}$ and $\mathcal Z_{f,\mu,\sigma}^{n\times n}$ of {\em Gaussian random Toep\-litz} and {\em Gaussian random $f$-circulant matrices}, respectively. \end{definition} \begin{definition}\label{defchi} $\chi_{\mu,\sigma,n}(y)$ is the cdf of the norm $||{\bf v}||=(\sum_{i=1}^n v_i^2)^{1/2}$ of a Gaussian random vector ${\bf v}=(v_i)_{i=1}^n\in \mathcal G_{\mu,\sigma}^{n\times 1}$. For $y\ge 0$ we have $\chi_{0,1,n}(y)= \frac {2}{2^{n/2}\Gamma(n/2)}\int_{0}^yx^{n-1}\exp(-x^2/2) dx$ where $\Gamma(h)=\int_0^{\infty}x^{h-1}\exp(-x) dx$, $\Gamma (n+1)=n!$ for nonnegative integers $n$. \end{definition} The total degree of a multivariate monomial is the sum of its degrees in all its variables. The total degree of a polynomial is the maximal total degree of its monomials. \begin{lemma}\label{ledl} \cite{DL78}, \cite{S80}, \cite{Z79}. For a set $\Delta$ of a cardinality $|\Delta|$ in any fixed ring let a polynomial in $m$ variables have a total degree $d$ and let it not vanish identically on this set. Then the polynomial vanishes in at most $d|\Delta|^{m-1}$ points. \end{lemma} We assume that Gaussian random variables range over infinite sets $\Delta$, usually over the real line or its interval. Then the lemma implies that a nonzero polynomial vanishes with probability 0. Consequently a square Gaussian random general, Toeplitz or circulant matrix is nonsingular with probability 1 because its determinant is a polynomials in the entries. Likewise rectangular Gaussian random general, Toeplitz and circulant matrices have full rank with probability 1. Hereafter, wherever this causes no confusion, we assume by default that {\em Gaussian random general, Toeplitz and circulant matrices have full rank}. \section{Extremal singular values of Gaussian random matrices}\label{ssvrm} Besides having full rank with probability 1, Gaussian random matrices in Definition \ref{defrndm} are likely to be well conditioned \cite{D88}, \cite{E88}, \cite{ES05}, \cite{CD05}, \cite{B11}, and even the sum $M+A$ for $M\in \mathbb R^{m\times n}$ and $A\in \mathcal G_{\mu,\sigma}^{m\times n}$ is likely to be well conditioned unless the ratio $\sigma/||M||$ is small or large \cite{SST06}. The following theorem states an upper bound proportional to $y$ on the cdf $F_{1/||A^+||}(y)$, that is on the probability that the smallest positive singular value $1/||A^+||=\sigma_l(A)$ of a Gaussian random matrix $A$ is less than a nonnegative scalar $y$ (cf. (\ref{eqnrm+})) and consequently on the probability that the norm $||A^+||$ exceeds a positive scalar $x$. The stated bound still holds if we replace the matrix $A$ by $A-B$ for any fixed matrix $B$, and for $B=O_{m,n}$ the bounds can be strengthened by a factor $y^{|m-n|}$ \cite{ES05}, \cite{CD05}. \begin{theorem}\label{thsiguna} Suppose $A\in \mathcal G_{\mu,\sigma}^{m\times n}$, $B\in \mathbb R^{m\times n}$, $l=\min\{m,n\}$, $x>0$, and $y\ge 0$. Then $F_{\sigma_l(A-B)}(y)\le 2.35~\sqrt l y/\sigma$, that is $Probability \{||(A-B)^+||\ge 2.35x\sqrt {l}/\sigma\}\le 1/x$. \end{theorem} \begin{proof} For $m=n$ this is \cite[Theorem 3.3]{SST06}. Apply Fact \ref{faccondsub} to extend it to any pair $\{m,n\}$. \end{proof} The following two theorems supply lower bounds $F_{||A||}(z)$ and $F_{\kappa (A)}(y)$ on the probabilities that $||A||\le z$ and $\kappa(A)\le y$ for two scalars $y$ and $z$, respectively, and a Gaussian random matrix $A$. We do not use the second theorem, but state it for the sake of completeness and only for square $n\times n$ matrices $A$. The theorems imply that the functions $1-F_{||A||}(z)$ and $1-F_{\kappa (A)}(y)$ decay as $z\rightarrow \infty$ and $y\rightarrow \infty$, respectively, and that the two decays are exponential in $-z^2$ and proportional to $\sqrt{\log y}/y$, respectively. For small values $y\sigma$ and a fixed $n$ the lower bound of Theorem \ref{thmsiguna} becomes negative, in which case the theorem becomes trivial. Unlike Theorem \ref{thsiguna}, in both theorems we assume that $\mu=0$. \begin{theorem}\label{thsignorm} \cite[Theorem II.7]{DS01}. Suppose $A\in \mathcal G_{0,\sigma}^{m\times n}$, $h=\max\{m,n\}$ and $z\ge 2\sigma\sqrt h$. Then $F_{||A||}(z)\ge 1- \exp(-(z-2\sigma\sqrt h)^2/(2\sigma^2))$, and so the norm $||A||$ is likely to have order $\sigma\sqrt h$. \end{theorem} \begin{theorem}\label{thmsiguna} \cite[Theorem 3.1]{SST06}. Suppose $0<\sigma\le 1$, $y\ge 1$, $A\in \mathcal G_{0,\sigma}^{n\times n}$. Then the matrix $A$ has full rank with probability $1$ and $F_{\kappa (A)}(y)\ge 1-(14.1+4.7\sqrt{(2\ln y)/n})n/ (y\sigma)$. \end{theorem} \begin{proof} See \cite[the proof of Lemma 3.2]{SST06}. \end{proof} \section{Extremal singular values of Gaussian random Toeplitz matrices}\label{scgrtm} A matrix $T_n=(t_{i-j})_{i,j=1}^n$ is the sum of two triangular Toeplitz matrices \begin{equation}\label{eqt2tt} T_n= Z({\bf t})+Z({\bf t_-})^T,~{\bf t}=(t_{i})_{i=0}^{n-1},~{\bf t}_-=(t'_{-i})_{i=0}^{n-1},~ t'_0=0. \end{equation} If $T_n\in \mathcal T_{\mu,\sigma}^{n\times n}$, then $T_n$ has $2n-1$ pairwise independent entries in $\mathcal G_{\mu,\sigma}$. Thus (\ref{eqttn}) implies that $$ ||T_n||\le ||Z({\bf t})||+||Z({\bf t_-})^T||\le ||{\bf t}||_1+||{\bf t_-}||_1= ||(t_{i})_{i=1-n}^{n-1}||_1\le \sqrt {2n-1}~||(t_{i})_{i=1-n}^{n-1}||.$$ \noindent Recall Definition \ref{defrndm} and obtain \begin{equation}\label{eqtn} F_{||T_n||}(y)\ge \chi_{\mu,\sigma,2n-1}(y/\sqrt {2n-1}). \end{equation} Next we estimate the norm $||T_n^{-1}||$ for $T_{n}\in \mathcal T_{\mu,\sigma}^{n\times n}$. \begin{lemma}\label{leinp} \cite[Lemma A.2]{SST06}. For a nonnegative scalar $y$, a unit vector ${\bf t}\in \mathbb R^{n\times 1}$, and a vector ${\bf b}\in \mathcal G_{\mu,\sigma}^{n\times 1}$, we have $F_{|{\bf t}^T{\bf b}|}(y)\le \sqrt{\frac{2}{\pi}}\frac{y}{\sigma}$. \end{lemma} \begin{remark}\label{reinp} The latter bound is independent of $\mu$ and $n$; it holds for any $\mu$ even if all coordinates of the vector ${\bf b}$ are fixed except for a single coordinate in $\mathcal G_{\mu,\sigma}$. \end{remark} \begin{theorem}\label{thsigunat1} Given a matrix $T_{n}=(t_{i-j})_{i,j=1}^n\in \mathcal T_{\mu,\sigma}^{n\times n}$, assumed to be nonsingular (cf. Section \ref{srvrm}), write $p_{1}={\bf e}_1^TT_n^{-1}{\bf e}_1$. Then $F_{1/||p_{1}T_n^{-1}||}(y)\le 2n\alpha \beta$ for two random variables $\alpha$ and $\beta$ such that \begin{equation}\label{eqprtinv} F_{\alpha}(y)\le \sqrt{\frac{2n}{\pi}}\frac{y}{\sigma}~{\rm and}~ F_{\beta}(y)\le \sqrt{\frac{2n}{\pi}}\frac{y}{\sigma}~{\rm for}~y\ge 0. \end{equation} \end{theorem} \begin{proof} Recall from part (a) of Theorem \ref{thgs} that $p_{1}T_n^{-1}=Z({\bf p})Z(J{\bf q})^T-Z(Z{\bf q})Z(ZJ{\bf p})^T$. Therefore $||p_{1}T_n^{-1}||\le ||Z({\bf p})||~||Z(J{\bf q})^T||+||Z(Z{\bf q})||~||Z(ZJ{\bf p})^T||$ for ${\bf p}=T_n^{-1}{\bf e}_1$, ${\bf q}=T_n^{-1}{\bf e}_n$, and $p_1={\bf p}^T{\bf e}_1$. It follows that $||p_{1}T_n^{-1}||\le ||Z({\bf p})||~||Z(J{\bf q})||+||Z(Z{\bf q})||~||Z(ZJ{\bf p})||$ since $||A||=||A^T||$ for all matrices $A$. Furthermore $||p_{1}T_n^{-1}||\le||{\bf p}||_1~||J{\bf q}||_1+||Z{\bf q}||_1~||ZJ{\bf p}||_1$ due to (\ref{eqttn}). Clearly $||J{\bf v}||_1=||{\bf v}||_1$ and $||Z{\bf v}||_1\le ||{\bf v}||_1$ for every vector ${\bf v}$, and so (cf. (\ref{eqnorm12})) \begin{equation}\label{eqtpq} ||p_{1}T_n^{-1}||\le 2 ||{\bf p}||_1~||{\bf q}||_1\le 2n ||{\bf p}||~||{\bf q}||. \end{equation} By definition the vector ${\bf p}$ is orthogonal to the vectors $T_n{\bf e}_2,\dots,T_n{\bf e}_n$, whereas ${\bf p}^TT_n{\bf e}_1=1$ (cf. \cite{SST06}). Consequenty the vectors $T_n{\bf e}_2,\dots,T_n{\bf e}_n$ uniquely define the vector ${\bf u}={\bf p}/||{\bf p}||$, whereas $|{\bf u}^TT_n{\bf e}_1|=1/||{\bf p}||$. The last coordinate $t_{n-1}$ of the vector $T_n{\bf e}_1$ is independent of the vectors $T_n{\bf e}_2,\dots,T_n{\bf e}_n$ and consequently of the vector ${\bf u}$. Apply Remark \ref{reinp} to estimate the cdf of the random variable $\alpha=1/||{\bf p}||=|{\bf u}^TT_n{\bf e}_1|$ and obtain that $F_{\alpha}(y)\le \sqrt{\frac{2n}{\pi}}\frac{y}{\sigma}$ for $y\ge 0$. Likewise the $n-1$ column vectors $T{\bf e}_1,\dots,T_{n-1}$ define the vector ${\bf v}=\beta{\bf q}$ for $\beta=1/||{\bf q}||=|{\bf v}^TT_n{\bf e}_n|$. The first coordinate $t_{1-n}$ of the vector $T_n{\bf e}_n$ is independent of the vectors $T{\bf e}_1,\dots,T_{n-1}$ and consequently of the vector ${\bf v}$. Apply Remark \ref{reinp} to estimate the cdf of the random variable $\beta$ and obtain that $F_{\beta}(y)\le \sqrt{\frac{2n}{\pi}}\frac{y}{\sigma}$ for $y\ge 0$. Finally combine these bounds on the cdfs $F_{\alpha}(y)$ and $F_{\beta}(y)$ with (\ref{eqtpq}). \end{proof} By applying parts (b) and (c) of Theorem \ref{thgs} instead of its part (a), we similarly deduce the bounds $||v_0T_{n+1}^{-1}||\le 2\alpha\beta$ and $||v_nT_{n+1}^{-1}||\le 2\alpha\beta$ for two pairs of random variables $\alpha$ and $\beta$ that satisfy (\ref{eqprtinv}) for $n+1$ replacing $n$. We have $p_{1}=\frac{\det T_{n-1}}{\det T_{n}}$, $v_0=\frac{\det T_n}{\det T_{n+1}}$, and $v_n=\frac{\det T_{0,1}}{\det T_{n+1}}$ for $T_{0,1}=(t_{i-j})_{i=0,j=1}^{n-1,n}$. Next we bound the geometric means of the ratios $|\frac{\det T_{h+1}}{\det T_{h}}|$ for $h=1,\dots,k-1$. $1/|p_1|$ and $1/|v_0|$ are such ratios for $k=n-1$ and $k=n$, respectively, whereas the ratio $1/|v_n|$ is similar to $1/|v_0|$, under slightly distinct notation. \begin{theorem}\label{thhdmr} Let $T_h\neq O$ denote $h\times h$ matrices for $h=1,\dots,k$ whose entries have absolute values at most $t$ for a fixed scalar or random variable $t$, e.g. for $t=||T||$. Furthermore let $T_1=(t)$. Then the geometric mean $(\prod_{h=1}^{k-1}|\frac{\det T_{h+1}}{\det T_{h}}|)^{1/(k-1)}=\frac{1}{t}|\det T_{k}|^{1/(k-1)}$ is at most $k^{\frac{1}{2}(1+\frac{1}{k-1})}t$. \end{theorem} \begin{proof} The theorem follows from Hadamard's upper bound $|\det M|\le k^{k/2}t^k$, which holds for any $k\times k$ matrix $M=(m_{i,j})_{i,j=1}^k$ with $\max_{i,j=1}^k|m_{i,j}|\le t$. \end{proof} The theorem says that the geometric mean of the ratios $|\det T_{h+1}/\det T_{h}|$ for $h=1,\dots,k-1$ is not greater than $k^{0.5+\epsilon(k)}t$ where $\epsilon(k)\rightarrow 0$ as $k\rightarrow \infty$. Furthermore if $T_n\in \mathcal T_{\mu,\sigma}^{n\times n}$ we can write $t=||T||$ and apply (\ref{eqtn}) to bound the cdf of $t$. \section{Extremal singular values of Gaussian random circulant matrices}\label{scgrcm} Next we estimate the norms of a random Gaussian $f$-circulant matrix and its inverse. \begin{theorem}\label{thcircsing} Assume $y\ge 0$ and a circulant $n\times n$ matrix $T=Z_1({\bf v})$ for ${\bf v}\in \mathcal G_{\mu,\sigma}^{n\times 1}$. Then (a) $F_{||T||}(y)\ge \chi_{\mu,\sigma,n} (\sqrt {\frac{2}{n}}y)$ for $\chi_{\mu,\sigma,n}(y)$ in Definition \ref{defchi} and (b) $F_{1/||T^{-1}||}(y)\le \sqrt{\frac{2}{\pi}} \frac{ny}{\sigma}$. \end{theorem} \begin{proof} For the matrix $T=Z_1({\bf v})$ we have both equation (\ref{eqt2tt}) and the bound $||{\bf t_-}||_1\le||{\bf t}||_1$, and so $||T||_1\le 2||{\bf t}||_1$. Now part (a) of the theorem follows similarly to (\ref{eqtn}). To prove part (b) recall Theorem \ref{thcpw} and write $B=\Omega T\Omega^{-1}=D({\bf u})$, ${\bf u}=(u_i)_{i=0}^{n-1}=\Omega {\bf v}$. We have $\sigma_j(T)=\sigma_j(B)$ for all $j$ because $\frac{1}{\sqrt n}\Omega$ and $\sqrt n\Omega^{-1}$ are unitary matrices. By combining the equations $u_i={\bf e}_i^T\Omega{\bf v}$, the bounds $||\Re ({\bf e}_i^T\Omega)||\ge 1$ for all $i$, and Lemma \ref{leinp}, deduce that $F_{|\Re (u_i)|}(y)\le \sqrt{\frac{2}{\pi}} \frac{y}{\sigma}$ for $i=1,\dots,n$. We have $F_{\sigma_n(B)}(y)=F_{\min_i|u_i|}(y)$ because $B=\diag(u_i)_{i=0}^{n-1}$, and clearly $|u_i|\ge |\Re (u_i)|$. \end{proof} \begin{remark}\label{retcond} Our extensive experiments suggest that the estimates of Theorem \ref{thcircsing} are overly pessimistic (cf. Table \ref{tabcondcirc}). \end{remark} Combining Theorem \ref{thcpw} with minimax property (\ref{eqminmax}) implies that $$\frac{1}{g(f)}\sigma_j(Z_1({\bf v}))\le \sigma_j(Z_f({\bf v}))\le g(f) \sigma_j(Z_1({\bf v}))$$ for all vectors ${\bf v}$, scalars $f\neq 0$, $g(f)=\max\{|f|^2,{1/|f|^2}\}$, and $j=1,\dots,n$. Thus we can readily extend the estimates of Theorem \ref{thcircsing} to $f$-circulant matrices for $f\neq 0$. In particular Gaussian random $f$-circulant matrices tend to be well conditioned unless $f\approx 0$ or $1/f\approx 0$. \section{Numerical Experiments}\label{sexp} Our numerical experiments with random general, Hankel, Toeplitz and circulant matrices have been performed in the Graduate Center of the City University of New York on a Dell server with a dual core 1.86 GHz Xeon processor and 2G memory running Windows Server 2003 R2. The test Fortran code was compiled with the GNU gfortran compiler within the Cygwin environment. Random numbers were generated with the random\_number intrinsic Fortran function, assuming the uniform probability distribution over the range $\{x:-1 \leq x < 1\}$. The tests have been designed by the first author and performed by his coauthors. We have computed the condition numbers of random general $n\times n$ matrices for $n=2^k$, $k=5,6,\dots,$ with entries sampled in the range $[-1,1)$ as well as complex general, Toeplitz, and circulant matrices whose entries had real and imaginary parts sampled at random in the same range $[-1,1)$. We performed 100 tests for each class of inputs, each dimension $n$, and each nullity $r$. Tables \ref{tab01}--\ref{tabcondcirc} display the test results. The last four columns of each table display the average (mean), minimum, maximum, and standard deviation of the computed condition numbers of the input matrices, respectively. Namely we computed the values $\kappa (A)=||A||~||A^{-1}||$ for general, Toeplitz, and circulant matrices $A$ and the values $\kappa_1 (A)=||A||_1~||A^{-1}||_1$ for Toeplitz matrices $A$. We computed and displayed in Table \ref{tabcondtoep} the 1-norms of Toeplitz matrices and their inverses rather than their 2-norms to facilitate the computations in the case of inputs of large sizes. Relationships (\ref{eqnorm12}) link the 1-norms and 2-norms to one another, but the empirical data in Table \ref{nonsymtoeplitz} consistently show even closer links, in all cases of general, Toeplitz, and circulant $n\times n$ matrices $A$ where $n=32,64,\dots, 1024$. \begin{table}[h] \caption{The norms of random general, Toeplitz and circulant $n\times n$ matrices and of their inverses} \label{nonsymtoeplitz} \begin{center} \begin{tabular}{|c|c|c|c|c|c|c|c|} \hline \textbf{matrix $A$}&\textbf{$n$}&\textbf{$||A||_1$}&\textbf{$||A||_2$}&\textbf{$\frac{||A||_1}{||A||_2}$}&\textbf{$||A^{-1}||_1$}&\textbf{$||A^{-1}||_2$}&\textbf{$\frac{||A^{-1}||_1}{||A^{-1}||_2}$}\\\hline General & $32$ & $1.9\times 10^{1}$ & $1.8\times 10^{1}$ & $1.0\times 10^{0}$ & $4.0\times 10^{2}$ & $2.1\times 10^{2}$ & $1.9\times 10^{0}$ \\ \hline General & $64$ & $3.7\times 10^{1}$ & $3.7\times 10^{1}$ & $1.0\times 10^{0}$ & $1.2\times 10^{2}$ & $6.2\times 10^{1}$ & $2.0\times 10^{0}$ \\ \hline General & $128$ & $7.2\times 10^{1}$ & $7.4\times 10^{1}$ & $9.8\times 10^{-1}$ & $3.7\times 10^{2}$ & $1.8\times 10^{2}$ & $2.1\times 10^{0}$ \\ \hline General & $256$ & $1.4\times 10^{2}$ & $1.5\times 10^{2}$ & $9.5\times 10^{-1}$ & $5.4\times 10^{2}$ & $2.5\times 10^{2}$ & $2.2\times 10^{0}$ \\ \hline General & $512$ & $2.8\times 10^{2}$ & $3.0\times 10^{2}$ & $9.3\times 10^{-1}$ & $1.0\times 10^{3}$ & $4.1\times 10^{2}$ & $2.5\times 10^{0}$ \\ \hline General & $1024$ & $5.4\times 10^{2}$ & $5.9\times 10^{2}$ & $9.2\times 10^{-1}$ & $1.1\times 10^{3}$ & $4.0\times 10^{2}$ & $2.7\times 10^{0}$ \\ \hline Toeplitz & $32$ & $1.8\times 10^{1}$ & $1.9\times 10^{1}$ & $9.5\times 10^{-1}$ & $2.2\times 10^{1}$ & $1.3\times 10^{1}$ & $1.7\times 10^{0}$ \\ \hline Toeplitz & $64$ & $3.4\times 10^{1}$ & $3.7\times 10^{1}$ & $9.3\times 10^{-1}$ & $4.6\times 10^{1}$ & $2.4\times 10^{1}$ & $2.0\times 10^{0}$ \\ \hline Toeplitz & $128$ & $6.8\times 10^{1}$ & $7.4\times 10^{1}$ & $9.1\times 10^{-1}$ & $1.0\times 10^{2}$ & $4.6\times 10^{1}$ & $2.2\times 10^{0}$ \\ \hline Toeplitz & $256$ & $1.3\times 10^{2}$ & $1.5\times 10^{2}$ & $9.0\times 10^{-1}$ & $5.7\times 10^{2}$ & $2.5\times 10^{2}$ & $2.3\times 10^{0}$ \\ \hline Toeplitz & $512$ & $2.6\times 10^{2}$ & $3.0\times 10^{2}$ & $8.9\times 10^{-1}$ & $6.9\times 10^{2}$ & $2.6\times 10^{2}$ & $2.6\times 10^{0}$ \\ \hline Toeplitz & $1024$ & $5.2\times 10^{2}$ & $5.9\times 10^{2}$ & $8.8\times 10^{-1}$ & $3.4\times 10^{2}$ & $1.4\times 10^{2}$ & $2.4\times 10^{0}$ \\ \hline Circulant & $32$ & $1.6\times 10^{1}$ & $1.8\times 10^{1}$ & $8.7\times 10^{-1}$ & $9.3\times 10^{0}$ & $1.0\times 10^{1}$ & $9.2\times 10^{-1}$ \\ \hline Circulant & $64$ & $3.2\times 10^{1}$ & $3.7\times 10^{1}$ & $8.7\times 10^{-1}$ & $5.8\times 10^{0}$ & $6.8\times 10^{0}$ & $8.6\times 10^{-1}$ \\ \hline Circulant & $128$ & $6.4\times 10^{1}$ & $7.4\times 10^{1}$ & $8.6\times 10^{-1}$ & $4.9\times 10^{0}$ & $5.7\times 10^{0}$ & $8.5\times 10^{-1}$ \\ \hline Circulant & $256$ & $1.3\times 10^{2}$ & $1.5\times 10^{2}$ & $8.7\times 10^{-1}$ & $4.7\times 10^{0}$ & $5.6\times 10^{0}$ & $8.4\times 10^{-1}$ \\ \hline Circulant & $512$ & $2.6\times 10^{2}$ & $3.0\times 10^{2}$ & $8.7\times 10^{-1}$ & $4.5\times 10^{0}$ & $5.4\times 10^{0}$ & $8.3\times 10^{-1}$ \\ \hline Circulant & $1024$ & $5.1\times 10^{2}$ & $5.9\times 10^{2}$ & $8.7\times 10^{-1}$ & $5.5\times 10^{0}$ & $6.6\times 10^{0}$ & $8.3\times 10^{-1}$ \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[h] \caption{The condition numbers $\kappa (A)$ of random $n\times n$ matrices $A$} \label{tab01} \begin{center} \begin{tabular}{| c | c | c | c | c | c |} \hline \bf{$n$}&\bf{input} & \bf{min} &\bf{max} &\bf{mean} &\bf{std} \\ \hline $ 32 $ & $ {\rm real} $ & $2.4\times 10^{1}$ & $1.8\times 10^{3}$ & $2.4\times 10^{2}$ & $3.3\times 10^{2}$ \\ \hline $ 64 $ & $ {\rm real} $ & $4.6\times 10^{1}$ & $1.1\times 10^{4}$ & $5.0\times 10^{2}$ & $1.1\times 10^{3}$ \\ \hline $ 128 $ & $ {\rm real} $ & $1.0\times 10^{2}$ & $2.7\times 10^{4}$ & $1.1\times 10^{3}$ & $3.0\times 10^{3}$ \\ \hline $ 256 $ & $ {\rm real} $ & $2.4\times 10^{2}$ & $8.4\times 10^{4}$ & $3.7\times 10^{3}$ & $9.7\times 10^{3}$ \\ \hline $ 512 $ & $ {\rm real} $ & $3.9\times 10^{2}$ & $7.4\times 10^{5}$ & $1.8\times 10^{4}$ & $8.5\times 10^{4}$ \\ \hline $ 1024 $ & $ {\rm real} $ & $8.8\times 10^{2}$ & $2.3\times 10^{5}$ & $8.8\times 10^{3}$ & $2.4\times 10^{4}$ \\ \hline $ 2048 $ & $ {\rm real} $ & $2.1\times 10^{3}$ & $2.0\times 10^{5}$ & $1.8\times 10^{4}$ & $3.2\times 10^{4}$ \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[h] \caption{The condition numbers $\kappa_1 (A)=\frac{||A||_1}{||A^{-1}||_1}$ of random Toeplitz $n\times n$ matrices $A$} \label{tabcondtoep} \begin{center} \begin{tabular}{|c|c|c|c|c|} \hline \textbf{$n$}&\textbf{min}&\textbf{mean}&\textbf{max}&\textbf{std}\\\hline $256$ & $9.1\times 10^{2}$ & $9.2\times 10^{3}$ & $1.3\times 10^{5}$ & $1.8\times 10^{4}$ \\ \hline $512$ & $2.3\times 10^{3}$ & $3.0\times 10^{4}$ & $2.4\times 10^{5}$ & $4.9\times 10^{4}$ \\ \hline $1024$ & $5.6\times 10^{3}$ & $7.0\times 10^{4}$ & $1.8\times 10^{6}$ & $2.0\times 10^{5}$ \\ \hline $2048$ & $1.7\times 10^{4}$ & $1.8\times 10^{5}$ & $4.2\times 10^{6}$ & $5.4\times 10^{5}$ \\ \hline $4096$ & $4.3\times 10^{4}$ & $2.7\times 10^{5}$ & $1.9\times 10^{6}$ & $3.4\times 10^{5}$ \\ \hline $8192$ & $8.8\times 10^{4}$ & $1.2\times 10^{6}$ & $1.3\times 10^{7}$ & $2.2\times 10^{6}$ \\ \hline \end{tabular} \end{center} \end{table} \begin{table}[h] \caption{The condition numbers $\kappa (A)$ of random circulant $n\times n$ matrices $A$} \label{tabcondcirc} \begin{center} \begin{tabular}{|c|c|c|c|c|} \hline \textbf{$n$}&\textbf{min}&\textbf{mean}&\textbf{max}&\textbf{std}\\\hline $256$ & $9.6\times 10^{0}$ & $1.1\times 10^{2}$ & $3.5\times 10^{3}$ & $4.0\times 10^{2}$ \\ \hline $512$ & $1.4\times 10^{1}$ & $8.5\times 10^{1}$ & $1.1\times 10^{3}$ & $1.3\times 10^{2}$ \\ \hline $1024$ & $1.9\times 10^{1}$ & $1.0\times 10^{2}$ & $5.9\times 10^{2}$ & $8.6\times 10^{1}$ \\ \hline $2048$ & $4.2\times 10^{1}$ & $1.4\times 10^{2}$ & $5.7\times 10^{2}$ & $1.0\times 10^{2}$ \\ \hline $4096$ & $6.0\times 10^{1}$ & $2.6\times 10^{2}$ & $3.5\times 10^{3}$ & $4.2\times 10^{2}$ \\ \hline $8192$ & $9.5\times 10^{1}$ & $3.0\times 10^{2}$ & $1.5\times 10^{3}$ & $2.5\times 10^{2}$ \\ \hline $16384$ & $1.2\times 10^{2}$ & $4.2\times 10^{2}$ & $3.6\times 10^{3}$ & $4.5\times 10^{2}$ \\ \hline $32768$ & $2.3\times 10^{2}$ & $7.5\times 10^{2}$ & $5.6\times 10^{3}$ & $7.1\times 10^{2}$ \\ \hline $65536$ & $2.4\times 10^{2}$ & $1.0\times 10^{3}$ & $1.2\times 10^{4}$ & $1.3\times 10^{3}$ \\ \hline $131072$ & $3.9\times 10^{2}$ & $1.4\times 10^{3}$ & $5.5\times 10^{3}$ & $9.0\times 10^{2}$ \\ \hline $262144$ & $6.3\times 10^{2}$ & $3.7\times 10^{3}$ & $1.1\times 10^{5}$ & $1.1\times 10^{4}$ \\ \hline $524288$ & $8.0\times 10^{2}$ & $3.2\times 10^{3}$ & $3.1\times 10^{4}$ & $3.7\times 10^{3}$ \\ \hline $1048576$ & $1.2\times 10^{3}$ & $4.8\times 10^{3}$ & $3.1\times 10^{4}$ & $5.1\times 10^{3}$ \\ \hline \end{tabular} \end{center} \end{table} \section{Implicit empirical support of the estimates of Sections \ref{scgrtm} and \ref{scgrcm}}\label{simprel} The papers \cite{PQa} and \cite{PQZa} describe successful applications of randomized circulant and Toeplitz multipliers to some fundamental matrix computations. These applications were bound to fail if the multipliers were ill conditioned, and so the success gives some implicit empirical support to our probabilistic estimates of Sections \ref{scgrtm} and \ref{scgrcm} and motivates the effort for proving these estimates. Namely it is well known that Gaussian elimination with no pivoting fails numerically where the input matrix has an ill conditioned leading block, even if the matrix itself is nonsingular and well conditioned. In our extensive tests in \cite{PQa} and \cite{PQZa} we consistently fixed this problem by means of multiplication by random circulant matrices. This implies that the random circulant matrices tend to be nonsingular and well conditioned for otherwise the products would be singular or ill conditioned. Likewise in other tests in \cite{PQZa} the column sets of the products $A^TG$ of an $n\times m$ matrix $A^T$ having a numerical rank $\rho$ by random Toeplitz $m\times \rho$ multipliers consisently approximated some bases for the singular spaces associated with the $\rho$ largest singular vaues of the matrix $A$, and this was readily extended to computing a rank-$\rho$ approximation of the matrix $A$, which is a fundamental task of matrix computations \cite{HMT11}. Then again one can immediately observe that these tests would have failed numerically if the multipliers and consequently the products were ill conditioned. \section{Conclusions}\label{srel} Estimating the condition numbers of random structured matrices is a well known challenge (cf. \cite{SST06}). We deduce such estimates for Gaussian random Toeplitz and circulant matrices. The former estimates can be surprising because the condition numbers grow exponentially in $n$ as $n\rightarrow \infty$ for some large and important classes of $n\times n$ Toeplitz matrices \cite{BG05}, whereas we prove the opposit for Gaussian random Toeplitz matrices. Our formal estimates are in good accordance with our numerical tests, except that circulant matrices tended to be even better conditioned in the tests than according to our formal study. The study of the condition number of Hankel matrices is immediately reduced to the study for Toeplitz matrices and vice versa. Can our progress be extended to other important classes of structured matrices? $~$ {\bf Acknowledgements:} Our research has been supported by NSF Grant CCF--1116736 and PSC CUNY Awards 64512--0042 and 65792--0043.

https://arxiv.org/abs/2007.13155

Revisiting a sharpened version of Hadamard's determinant inequality

Hadamard's determinant inequality was refined and generalized by Zhang and Yang in [Acta Math. Appl. Sinica 20 (1997) 269-274]. Some special cases of the result were rediscovered recently by Rozanski, Witula and Hetmaniok in [Linear Algebra Appl. 532 (2017) 500-511]. We revisit the result in the case of positive semidefinite matrices, giving a new proof in terms of majorization and a complete description of the conditions for equality in the positive definite case. We also mention a block extension, which makes use of a result of Thompson in the 1960s.

\section{Introduction} Perhaps the best known determinantal inequality in mathematical sciences is the Hadamard inequality (e.g., \cite[p. 505]{HJ13}) which says that if $A=(a_{ij})$ is an $n\times n$ (Hermitian) positive definite matrix, then \begin{eqnarray}\label{h1} \det A\le a_{11}\cdots a_{nn},\end{eqnarray} and equality holds if and only if $A$ is diagonal. Two decades ago, Zhang and Yang obtained an elegant sharpening of the Hadamard inequality. They proved it for a more general class of matrices and included a term involving off-diagonal entries of the matrix. Let $A$ be an $n\times n$ complex matrix. For a non-empty proper subset $G$ of $\{1,\dots,n\}$, let $A[G]$ denote the principal submatrix of $A$ formed by discarding the $i$th row and column of $A$, for each $i\notin G$. We say $A$ is an $F$-matrix if for all such $G$, $\det(A[G])\ge0$ and $\det(A)\le\det(A[G])\det(A[G^C])$ That is, all principal minors of $A$ are non-negative and they satisfy a Fischer-type inequality. Standard results show that the Hermitian $F$-matrices are exactly the positive semi-definite matrices. \begin{thm}\label{ZYmain} (Zhang-Yang \cite{ZY97}) If $A=(a_{ij})$ is an $F$-matrix, then $a_{ij}a_{ji}\ge0$ for all $i,j$ and, if $\sigma$ is a non-trivial permutation of $\{1,\dots,n\}$, then \begin{equation}\label{maxZY} \det(A)+\Big(\prod_{i=1}^na_{i,\sigma(i)}a_{\sigma(i),i}\Big)^{1/2}\le \prod_{i=1}^na_{ii}. \end{equation} \end{thm} Without noticing the work of Zhang and Yang in \cite{ZY97}, which was written in Chinese, Rozanski, Witula and Hetmaniok recently rediscovered some special cases of (\ref{maxZY}) in \cite{RWH2017}. For this reason, we think it worthwhile to bring the nice result of Theorem \ref{ZYmain} to the attention of the linear algebra community. Besides Theorem \ref{ZYmain}, \cite{ZY97} also contains necessary and sufficient conditions for an $F$-matrix $A$ to satisfy the equation $\det(A)=a_{11}\dots a_{nn}$. The same was done for the equation $\operatorname{permanent}(A)=a_{11}\dots a_{nn}$. However, Zhang and Yang did not give conditions for equality to hold in (\ref{maxZY}). Conditions for equality were included by Rozanski, et al. for the special cases considered in \cite{RWH2017}. There are multiple known ways to prove the Hadamard inequality (\ref{h1}) in the literature (see, e.g., \cite{Hol07, HJ13, MO82}). We recall that one insightful way of seeing the Hadamard inequality is via Schur's majorization inequality (see Lemma \ref{lem1}). For a quick summary of the intimate connection between majorization and determinant inequalities, we refer to \cite{Lin17}. Our initial motivation was that since the original Hadamard inequality (\ref{h1}) is immediate from majorization (see, e.g. \cite[p. 44]{And94}, \cite[p. 67]{Zha13}), it would be nice if Theorem \ref{ZYmain} could also be seen from that perspective. In this note, we give a new proof of Theorem \ref{ZYmain} for positive semi-definite matrices using majorization techniques. Our results include necessary and sufficient conditions for equality to hold when the matrix $A$ is positive definite. Before proceeding, let us fix some notation. For a vector $x\in\mathbb{R}^n$, we denote by $x^{\downarrow}=(x_1^{\downarrow}, \ldots, x_n^{\downarrow})\in\mathbb{R}^n$ the vector with the same components as $x$, but sorted in nonincreasing order. Given $x, y\in \mathbb{R}^n$, we say that $x$ majorizes $y$ (or $y$ is majorized by $x$), written as $x\succ y$, if $$ \sum_{i=1}^k x_i^{\downarrow} \geq \sum_{i=1}^k y_i^{\downarrow} \quad \text{for } k=1,\dots, n-1 $$ and equality holds at $k=n$. Three basic facts about majorization are given below. The first is a matrix characterization, the next is Schur's majorization inequality, and the last is a consequence for the elementary symmetric functions. A {\it doubly stochastic matrix} is square matrix with non-negative entries and all row and column sums equal to 1. \begin{lemma}\label{lem0} \cite[p. 253]{HJ13} If $x$ and $y$ are real row vectors then $x$ majorizes $y$ if and only if there exists a doubly stochastic matrix $S$ such that $y=xS$. \end{lemma} \begin{lemma}\label{lem1} \cite[p. 249]{HJ13} The eigenvalues of a Hermitian matrix majorize its diagonal entries. \end{lemma} Fix a positive integer $n$ and let $e_k(x)$, $k=1, 2, \ldots, n$, denote the $k$th elementary symmetric function in the $n$ variables $x_1, \ldots, x_n$. See \cite[p.114]{MOA11}. By convention, $e_0(x)=1$. \begin{lemma}\label{lem2} \cite[p.115]{MOA11} Let $x, y\in [0,\infty)^n$. If $n\ge2$ and $x\succ y$, then $e_k(x)\le e_k(y)$ for $k=0, 1, \ldots, n$. If $k>1$ and $x, y\in (0,\infty)^n$ then equality holds if and only if $x^\downarrow=y^\downarrow$. \end{lemma} \section{A new proof of Theorem \ref{ZYmain} and more} In this section, $\Lambda$, $V$ and $B$ will be as follows: Let $\Lambda$ be a diagonal matrix with non-negative diagonal entries $\lambda_1,\dots,\lambda _n$. Let $V=(v_{ij})$ be an $n\times n$ matrix whose rows and columns are all unit vectors, that is, all diagonal entries of $V^*V$ and $VV^*$ are equal to $1$. Set $B=(b_{ij})=V^*\Lambda V$. It is important to point out that, in general, $\prod_{i=1}^n\lambda_i\ne \det(B)$, although they are equal when $V$ is unitary. \begin{lemma}\label{2.L} Let $P(t)=\prod_{i=1}^n(\lambda_{i}-t)$ and $Q(t)=\prod_{i=1}^n(b_{ii}-t)$. Fix $s<t\le\min(\lambda_1,\dots,\lambda_n)$. Then $P(t)\le Q(t)$ and $P(s)-P(t)\le Q(s)-Q(t)$. If $n\ge3$ and $P(s)-P(t)= Q(s)-Q(t)$ then $(b_{11},\dots,b_{nn})$ is a permutation of $(\lambda_1,\dots,\lambda_n)$. \end{lemma} \begin{proof} The conditions on $V$ show that the matrix $S=(|v_{ij}|^2)$ is doubly stochastic and a calculation shows that $(b_{11}-t,\dots,b_{nn}-t)=(\lambda_1-t,\dots,\lambda _n-t)S$. By Lemma \ref{lem0} and Lemma \ref{lem2}, $$ e_k(\lambda_1-t,\dots,\lambda _n-t)\le e_k(b_{11}-t,\dots,b_{nn}-t) $$ for $k=0,\dots,n$. In particular, $$ P(t)=e_n(\lambda_1-t,\dots,\lambda _n-t)\le e_n(b_{11}-t,\dots,b_{nn}-t)=Q(t). $$ Also, \begin{align*} P(s)-P(t)&=\prod_{i=1}^n(\lambda_i-t+t-s)-\prod_{i=1}^n(\lambda_i-t)\\ &=\sum_{k=0}^{n-1}e_k(\lambda_1-t,\dots,\lambda _n-t)(t-s)^{n-k}\\ &\le\sum_{k=0}^{n-1}e_k(b_{11}-t,\dots,b_{nn}-t)(t-s)^{n-k} \\ &=\prod_{i=1}^n(b_{ii}-t+t-s)- \prod_{i=1}^n(b_{ii}-t) =Q(s)-Q(t). \end{align*} If $n\ge3$ and $P(s)-P(t)= Q(s)-Q(t)$, then the above estimate reduces to equality throughout, which implies $e_2(\lambda_1-t,\dots,\lambda _n-t)=e_2(b_{11}-t,\dots,b_{nn}-t)$. But $e_2$ is strictly Schur concave on all of $\Bbb R^n$. (See \cite[A.4]{MOA11} to prove concavity and then \cite[A.3.a]{MOA11} to prove strict concavity.) Thus $(b_{11}-t,\dots,b_{nn}-t)$ is a permutation of $(\lambda_1-t,\dots,\lambda_n-t)$ and therefore $(b_{11},\dots,b_{nn})$ is a permutation of $(\lambda_1,\dots,\lambda_n)$. \end{proof} \begin{thm} If $\lambda_1,\dots,\lambda_n$ are non-negative, then \begin{equation}\label{H+} \prod_{i=1}^n\lambda_i\le\prod_{i=1}^n b_{ii}. \end{equation} If $n\ge3$ and $\lambda_1,\dots,\lambda_n$ are strictly positive and distinct then equality holds if and only if $V$ has exactly one non-zero entry in each row and column. \end{thm} \begin{proof} Lemma \ref{2.L} shows that $P(0)\le Q(0)$, which is (\ref{H+}). Note that $V$ has exactly one non-zero entry in each row and column if and only if $S=(|v_{ij}|^2)$ is a permutation matrix. In that case, since $(b_{11},\dots,b_{nn})=(\lambda_1,\dots,\lambda _n)S$, (\ref{H+}) holds with equality. Now suppose $\lambda_1,\dots,\lambda_n$ are strictly positive and distinct. If equality holds in (\ref{H+}), that is, if $e_n(\lambda_1,\dots,\lambda _n)=e_n(b_{11},\dots,b_{nn})$, then $b_{11},\dots,b_{nn}$ are also strictly positive. Therefore Lemma \ref{lem2} shows $(b_{11},\dots,b_{nn})$ is a permutation of $(\lambda_1,\dots,\lambda _n)$. It follows that there are permutation matrices $R$ and $R'$ such that $(\lambda_1^\downarrow,\dots,\lambda _n^\downarrow)=(\lambda_1^\downarrow,\dots,\lambda _n^\downarrow)R'SR$. Clearly, $R'SR=(t_{ij})$ is also a doubly stochastic matrix. Assume that for some $i$, there exists an $m<i$ such that $t_{im}\ne0$. For this $i$ choose the largest such $m$. Then $t_{ij}(\lambda_j^\downarrow-\lambda_i^\downarrow)=0$ for $j>m$, $t_{ij}(\lambda_j^\downarrow-\lambda_i^\downarrow)\ge0$ for $j<m$, and $t_{im}(\lambda_m^\downarrow-\lambda_i^\downarrow)>0$. This shows that $\sum_{j=1}^nt_{ij}(\lambda_j^\downarrow-\lambda_i^\downarrow)>0$, which contradicts $(\lambda_1^\downarrow,\dots,\lambda _n^\downarrow)=(\lambda_1^\downarrow,\dots,\lambda _n^\downarrow)R'SR$. We conclude that $R'PR$ is upper triangular. It is easy to see that the only upper triangular, doubly stochastic matrix is $I$. Thus, $S$ is a permutation matrix. \end{proof} The following examples show that the conditions for equality may change in the non-generic cases where the $\lambda_1,\dots,\lambda_n$ are not distinct or include one or more zeros. \begin{eg} Let $\lambda_1=\lambda_2=1$ and $V=\left(\begin{smallmatrix}i\cos\theta&i\sin\theta&0\\\sin\theta&\cos\theta&0\\0&0&-1\end{smallmatrix}\right)$ for some $\theta$. Then $B=V^*V=\left(\begin{smallmatrix}1&\sin2\theta&0\\\sin2\theta&1&0\\0&0&1\end{smallmatrix}\right)$, but $VV^*=\left(\begin{smallmatrix}1&i\sin2\theta&0\\-i\sin2\theta&1&0\\0&0&1\end{smallmatrix}\right)$. So we have equality in (\ref{H+}) but $V$ does not have only one non-zero entry in each row and column. Significantly, the doubly stochastic matrix $S=\left(\begin{smallmatrix}\cos^2\theta&\sin^2\theta&0\\\sin^2\theta&\cos^2\theta&0\\0&0&1\end{smallmatrix}\right)$ is not uniquely determined; it varies with $\theta$. \end{eg} \begin{eg} With $\Lambda$ and $V$ as defined above, let $\Lambda'=\left(\begin{smallmatrix}\Lambda&0\\0&0\end{smallmatrix}\right)$ and $V'=\left(\begin{smallmatrix}V&0\\0&1\end{smallmatrix}\right)$. Then we have equality in (\ref{H+}) because both sides are zero, but $V'$ does not have one non-zero entry in each row and column unless $V$ does. \end{eg} The conditions imposed on $V$, above, are satisfied by any unitary matrix but the unitaries are only a small subclass of the possible matrices $V$. The next example gives large class of matrices $V$ that are not unitary. \begin{eg} Suppose $T=(t_{ij})$ is an $n\times n$ Hermitian matrix with spectral norm $\|T\|\le 1$ satisfying $t_{ii}=0$ for all $i$. Since $\|T\|\le1$, $I+T$ is positive semi-definite and therefore has a positive semi-definite square root $V=(I+T)^{1/2}$. Note that $V^*V=VV^*=I+T$, a matrix with ones on the diagonal. If $T$ is not zero the matrix $V$ is not unitary. \end{eg} On the other hand, if $V$ is unitary, the conditions on $V$ are satisfied automatically, the product $\lambda_1\cdots\lambda_n$ is the determinant of $B$, and $B$ could be any positive semi-definite matrix. Let $S_n$ be the group of permutations of $\{1,\dots,n\}$ and let $D_n$ denote the collection of permutations that have no fixed point. These are the so-called derangements of $\{1,\dots,n\}$. We also let $e_i$ be the $i$th column of the $n\times n$ identity matrix so that $Ae_i$ extracts the $i$th column of $A$. We remind the reader not to confuse $e_i$ with $e_i(x)$. \begin{thm}\label{refined} Let $n\ge 3$. If $A=(a_{ij})$ is positive semi-definite and $\tau\in D_n$, then \begin{equation}\label{maxtau} \det(A)+\prod_{i=1}^n|a_{i,\tau(i)}|\le \prod_{i=1}^na_{ii}. \end{equation} Equality holds if and only if $A$ is diagonal or the two vectors $Ae_i$ and $Ae_{\tau(i)}$ are collinear for each $i$. \end{thm} \begin{proof} Choose a unitary $V$ such that $A=V^*\Lambda V$, where $\lambda_1,\dots,\lambda_n$ are the (necessarily non-negative) eigenvalues of $A$. This makes $B=A$. Then set $t=\min(\lambda_1,\dots,\lambda_n)$. Lemma \ref{2.L} shows that $P(0)-P(t)\le Q(0)-Q(t)$. The choice of $t$ ensures that $P(t)=0$. Also $P(0)=\det(A)$ and $Q(0)=\prod_{i=1}^na_{ii}$. To prove (\ref{maxtau}) we use the Cauchy-Schwarz inequality to show that $Q(t)\ge\prod_{i=1}^n|a_{i,\tau(i)}|$ for each $\tau\in D_n$. Fix $\tau\in D_n$ and observe that \begin{equation}\label{CS}\begin{aligned} |a_{i,\tau(i)}|^2&=\bigg|\sum_{j=1}^n\bar v_{ji}\lambda_jv_{j,\tau(i)}\bigg|^2\\ &=\bigg|\sum_{j=1}^n\bar v_{ji}(\lambda_j-t)v_{j,\tau(i)}\bigg|^2\\ &\le\sum_{j=1}^n|v_{ji}|^2(\lambda_j-t)\sum_{j=1}^n|v_{j,\tau(i)}|^2(\lambda_j-t)\\ &=(a_{ii}-t)(a_{\tau(i),\tau(i)}-t). \end{aligned}\end{equation} Thus, \begin{equation}\label{aCS} \prod_{i=1}^n|a_{i,\tau(i)}| \le\bigg(\prod_{i=1}^n(a_{ii}-t)\prod_{i=1}^n(a_{\tau(i),\tau(i)}-t)\bigg)^{1/2}=\prod_{i=1}^n(a_{ii}-t)=Q(t). \end{equation} Next we consider conditions for equality. If $A$ is diagonal it is clear that (\ref{maxtau}) holds with equality. Suppose the two vectors $Ae_i$ and $Ae_{\tau(i)}$ are collinear for each $i$. Then $A$ is singular, $\det(A)=0$, and $t=0$. Fix $i$ and choose $a$ and $b$, not both zero, such that $aAe_i=bAe_{\tau(i)}$. Then for each $j$, $e_j^*VA(ae_i-be_{\tau(i)})=0$. But $AV^*e_j=V^*\Lambda e_j=\lambda_jV^*e_j$ so $\lambda_je_j^*V(ae_i-be_{\tau(i)})=0$. Therefore $av_{ji}=bv_{j,\tau(i)}$ for all $j$ such that $\lambda_j\ne0$. This gives equality in (\ref{CS}). Since this holds for all $i$, we have equality in (\ref{aCS}) as well. Since $\det(A)=0$ we also have equality in (\ref{maxtau}). Conversely, suppose equality holds in (\ref{maxtau}). The above proof shows that we must have $P(0)-P(t)= Q(0)-Q(t)$ and equality in (\ref{CS}) for all $i$. If $t>0$, Lemma \ref{2.L} shows that $(a_{11},\dots,a_{nn})$ is a permutation of $(\lambda_1,\dots,\lambda_n)$ giving equality in Hadamard's inequality. Therefore $A$ is diagonal. If $t=0$ then equality in (\ref{CS}) implies that for all $i$ there exist constants $a$ and $b$, not both zero, such that $av_{ji}=bv_{j,\tau(i)}$ for all $j$ such that $\lambda_j\ne 0$. Thus, for all $j$, $\lambda_je_j^*V(ae_i-be_{\tau(i)})=0$ and hence, as above, $e_j^*VA(ae_i-be_{\tau(i)})=0$. This holds for all $j$, and $V$ is invertible, so we conclude that $aAe_i=bAe_{\tau(i)}$, that is, $Ae_i$ and $Ae_{\tau(i)}$ are collinear. \end{proof} \begin{rem} The collinearity condition for equality above may be expressed in terms of matrix rank: For $J\subseteq \{1,\dots,n\}$, let $P_J$ be the orthogonal projection onto $\operatorname{span}\{e_j:j\in J\}$. Let $J^\tau_1,\dots,J^\tau_{n_\tau}$ be the orbits of $\tau\in D_n$. Then $I=\sum_{k=1}^{n_\tau}P_{J_k}$ so $A=\sum_{k=1}^{n_\tau}AP_{J_k}$. The condition that the two vectors $Ae_i$ and $Ae_{\tau(i)}$ are collinear for each $i$ is equivalent to saying that the rank of $AP_{J_k}$ is at most 1 for $k=1,\dots,n_\tau$. \end{rem} The next result follows from Lemma \ref{2.L} and the Cauchy-Schwarz estimates of Theorem \ref{refined}. We state it without proof. \begin{thm} Let $\tau\in D_n$. If for all $i$, the $(i,\tau(i))$ entry of $V^*V$ is zero then $$ \prod_{i=1}^n\lambda_i+\prod_{i=1}^n|b_{i,\tau(i)}|\le\prod_{i=1}^n b_{ii}. $$ \end{thm} Next is our proof of the motivating result: Theorem \ref{ZYmain} for positive semi-definite matrices, including conditions for equality. Observe that if $n=2$, (\ref{maxtau}) reduces to equality for every $A$. \begin{cor}\label{refinedcor} \cite{ZY97} Let $A=(a_{ij})$ be positive semi-definite matrix, and $\sigma\in S_n$. If $\sigma$ is not the identity permutation, then \begin{equation}\label{maxsigma} \det(A)+\prod_{i=1}^n|a_{i,\sigma(i)}|\le \prod_{i=1}^na_{ii}. \end{equation} Equality holds if $A$ is diagonal, or if for each $i$ either: $\sigma(i)=i$ and $Ae_i$, $e_i$ are collinear; $\sigma(i)\ne i$ and $\sigma$ is a transposition; or $\sigma(i)\ne i$ and $Ae_i$, $Ae_{\sigma(i)}$ are collinear. If $A$ is positive definite, these conditions are also necessary for equality. \end{cor} \begin{proof} Let $F$ be the set of fixed points of $\sigma$, a proper subset of $\{1,\dots,n\}$, and let $G$ be its complement. If $F$ is empty, the result follows from Theorem \ref{refined}. Note that both $A[F]$ and $A[G]$ are positive semi-definite. Fischer's inequality (see \cite[Theorem 7.8.5]{HJ13}), followed by Hadamard's inequality, gives \begin{equation}\label{F-H} \det(A)\le\det(A[G])\det(A[F]) \le\det(A[G])\prod_{i\in F}a_{ii}. \end{equation} Note that the restriction of $\sigma$ to $G$ is a permutation of $G$ with no fixed point. By Theorem \ref{refined}, we have \begin{equation}\label{refG} \det(A[G])+\prod_{i\in G}|a_{i,\sigma(i)}|\le\prod_{i\in G}a_{ii}, \end{equation} provided $G$ has at least three elements. We can remove that restriction, however, because (\ref{refG}) becomes equality when $G$ has exactly two elements, it is impossible for $G$ to have exactly one element, and we have excluded the case that $G$ is empty. Combining the last two inequalities gives (\ref{maxsigma}). If $A$ is diagonal then we clearly have equality in (\ref{maxsigma}). Now suppose that for each $i$ either: $\sigma(i)=i$ and $Ae_i$, $e_i$ are collinear; or $\sigma(i)\ne i$ and, if $\sigma$ is not just a transposition, then $Ae_i$, $Ae_{\sigma(i)}$ are collinear. This implies that $A[F]$ is diagonal, and $A$ has (up to reordering of the standard basis) a block diagonal decomposition with blocks $A[F]$ and $A[G]$. Thus we have equality in (\ref{F-H}). The conditions for equality in Theorem \ref{refined} give equality in (\ref{refG}) when $n\ge3$ and equality is trivial when $n=2$ so we have equality in (\ref{maxsigma}). Now suppose that $A$ is positive definite and equality holds in (\ref{maxsigma}). Since $\det(A)>0$, the Fischer inequality shows that $\det(A[F])>0$ and $\det(A[G])>0$. Therefore we have equality in both (\ref{F-H}) and (\ref{refG}). Equality in the Fischer inequality from (\ref{F-H}) implies that $A$ has (up to reordering of the standard basis) a block diagonal decomposition with blocks $A[F]$ and $A[G]$ (see \cite[p. 217]{Zha11}). Equality in the Hadamard inequality from (\ref{F-H}) implies that $A[F]$ is diagonal. Together, these show that if $\sigma(i)=i$, then $Ae_i$ and $e_i$ are collinear. Equality in (\ref{refG}) implies, via Theorem \ref{refined}, that if $G$ has at least three elements and $\sigma(i)\ne i$, then $Ae_i$, $Ae_{\sigma(i)}$ are collinear. If $G$ has fewer than three elements then $\sigma$ can only be a transposition. This completes the proof. \end{proof} Recall that the Hadamard product ``$\circ$" is the entrywise product of matrices. So the Hadamard inequality may be written as $\det A\le \det( A \circ I)$ for a positive semi-definite $A$. Theorem \ref{refined} enables us to state the following result. \begin{thm} \label{t4} Let $\tau\in D_n$ be a derangement and $P=(p_{ij})$ where $p_{ij}$ is $1$ when $j=\tau(i)$ and zero otherwise. For a positive semi-definite matrix $A=(a_{ij})$, \begin{eqnarray} \det (A\circ I)\ge \det A+|\det (A\circ P)|. \end{eqnarray} Equality holds if and only if the matrix $A$ is a diagonal matrix or the two vectors $Ae_i$ and $Ae_{\tau(i)}$ are collinear for each $i$.\end{thm} \begin{rem} We expect that Theorem \ref{t4} will stimulate further investigation of Oppenheim-Schur inequalities (see \cite[p. 509]{HJ13}). \end{rem} In 1961, Thompson \cite{Tho61} published a remarkable determinant inequality. \begin{thm}\label{Th}If $A=(A_{ij})$ is positive definite with each block $A_{ij}$ square, then \begin{eqnarray}\label{tho} \det A\le \det (\det A_{ij}). \end{eqnarray} Equality holds if and only if $A$ is block diagonal. \end{thm} We point out an extension of Theorem \ref{refined} to the block matrix case. \begin{thm}If $A=(A_{ij})$ is an $n\times n$ block positive definite matrix with each block $A_{ij}$ square, then for any derangement $\tau\in D_n$, \begin{eqnarray*} \det A+\prod_{i=1}^{n}|\det A_{i,\tau(i)}|\le \prod_{i=1}^{n}\det A_{ii}. \end{eqnarray*} Equality holds if and only if $A$ is block diagonal. \end{thm} \begin{proof} By (\ref{tho}), it suffices to work with the $n\times n$ positive definite matrix $(\det A_{ij})$. The conclusion then follows by Theorem \ref{t4}. \end{proof} \section*{Acknowledgments} Both authors are grateful to Professor Xingzhi Zhan for pointing to the early work of Xiao-Dong Zhang and Shangjun Yang on this topic. The authors also acknowledge some comments from the referee which help improve the presentation. The work of M. Lin is supported by the National Natural Science Foundation of China (Grant No. 11601314). The work of G. Sinnamon is supported by the Natural Sciences and Engineering Research Council of Canada.

https://arxiv.org/abs/1811.04846

Gaussian Quadrature Rule using ε-Quasiorthogonality

We introduce a new type of quadrature, known as approximate Gaussian quadrature (AGQ) rules using {\epsilon}-quasiorthogonality, for the approximation of integrals of the form \int f(x)d \alpha(x). The measure {\alpha}(\cdot) can be arbitrary as long as it possesses finite moments {\mu}n for sufficiently large n. The weights and nodes associated with the quadrature can be computed in low complexity and their count is inferior to that required by classical quadratures at fixed accuracy on some families of integrands. Furthermore, we show how AGQ can be used to discretize the Fourier transform with few points in order to obtain short exponential representations of functions.

\section{Introduction} \label{Intro} In this paper, we present a new kind of quadrature rule for approximating integrals by sums of the form, \begin{equation} \label{discrete} \int f(x) \, \d \alpha(x) \approx \sum_{i=1}^n w_i f(x_i) \end{equation} having the following characteristics: \begin{enumerate} \item The measure $\alpha( \cdot )$ can be \emph{arbitrary} (positive, signed, complex, ...) as long as it satisfies some weak condition. \item The nodes and weights associated with the quadrature rule can be obtained in low computational complexity through a simple numerical algorithm. \item The quadrature is at least as accurate as the Gaussian quadrature rule, and in many cases is significantly more accurate. \item Low-order rules are able to integrate high-order polynomials with high accuracy. \end{enumerate} The scheme presented in the current work uses a strategy similar to classical Gaussian quadrature rules (of which a few examples can be found in Table~\ref{classicalGaussQuad}). The Gaussian quadrature rule is designed to integrate exactly polynomials of degree at most $2n-1$ using $n$ quadrature points and weights: \[ \int x^k \, \d \alpha(x) \] for various weight functions $\frac{\d \alpha}{\d x}$ (see Table~\ref{classicalGaussQuad}). \begin{table}[htbp] \caption{Examples of classical Gaussian quadratures} \begin{center} \begin{tabular}{ccc} \toprule Name & Interval & Measure ($\d \alpha / \d x$ ) \\ \midrule Gauss-Legendre & $[-1,1]$ & $1$ \\ Gauss-Laguerre & $[0, \infty)$ & $e^{-x}$ \\ Gauss-Hermite & $(-\infty, \infty)$ & $e^{-x^2} $\\ Gauss-Jacobi & $(-1,1)$ & $(1-x)^{\alpha}(1+x)^{\beta} \; , \;\; \alpha, \beta > -1 $ \\ Chebyshev-Gauss (1st kind) & $(-1,1)$ & $1/\sqrt{1-x^2}$ \\ Chebyshev-Gauss (2nd kind) & $[-1,1]$ & $\sqrt{1-x^2}$ \\ \bottomrule \end{tabular} \end{center} \label{classicalGaussQuad} \end{table} The paper is structured as follows. In Section~\ref{GQ}, a brief overview of classical Gaussian quadratures will be presented. In Section~\ref{sec:agq}, the concept of quasiorthogonal polynomial and approximate Gaussian quadrature will be introduced together with an error analysis. This will be followed in Section~\ref{NS} by numerical results. In the same section, we will discuss representations of functions by short sums of exponentials. \section{Gaussian quadrature} \label{GQ} Gaussian quadratures are schemes used to approximate definite integrals of the form, \begin{equation*} \int_a^b f(x) \, \d \alpha(x) \end{equation*} by a finite weighted sum of the form, \begin{equation*} \sum_{n=0}^N w_n \, f(x_n) \end{equation*} where $a<b \in \mathbb{R}$. The coefficients $\{ w_n \}$ are generally referred to as the \emph{weights} of the quadrature, whereas the points $\{ x_n \}$ are referred to as the \emph{nodes}. An $(N+1)$-node Gaussian quadrature can integrate polynomials up to degree $2N+1$ \emph{exactly} and is generally well-suited for the integration of functions that are well-approximated by polynomials. In what follows, we will briefly describe how the nodes and weights of classical Gaussian quadratures can be obtained based on the classical theory of orthogonal polynomials. For this purpose, we shall denote the real and complex numbers by $\mathbb{R}$ and $\mathbb{C}$ respectively. $\alpha( \cdot )$ will represent an arbitrary measure (possibly complex) on $(\mathbb{R}, \mathcal{B})$ or $(\mathbb{C}, \mathcal{B})$ unless otherwise stated. Vectors are represented by lower case letter e.g., $v$. The $i^{th}$ component of a vector $v$ will be written as $v_i$, and we shall use super-indices of the form $v^{(j)}$ when multiple vectors are under consideration. We begin by introducing four key objects: the orthogonal polynomials, the Lagrange interpolants, the moments of a measure $\alpha(\cdot)$ and the Hankel matrix associated with such a measure. \begin{defn}{\bf (Orthogonal polynomial)} A sequence $\{ p^{(k)} (x) \}_{k=0}^{\infty}$ of polynomials of degree $k$ is said to be a sequence of orthogonal polynomials with respect to a positive measure $\alpha(\cdot)$ if, \begin{equation*} \int p^{(k)}(x) p^{(l)} (x) \, \d \alpha(x) = \left\{ \begin{array}{ll} 0 & \mbox{if } k \not = l \\ c_k & \mbox{if } k = l \end{array} \right. \end{equation*} If in addition $c_k = 1 \; \forall k \in \mathbb{N}$, then the sequence is called \emph{orthonormal}. \end{defn} We shall hereafter assume that all such polynomials are monic, i.e., that they can be written as, \begin{equation*} p^{(k)} (x) = x^k + \sum_{n=0}^{k-1} p^{(k)}_n x^n \end{equation*} where $\{ p^{(k)}_n \}_{n=0}^{k-1}$ are some (potentially complex) coefficients. We then introduce Lagrange interpolants, \begin{defn}{\bf (Lagrange interpolant)} Given a set of $(d+1)$ data points $\{ (x_n, y_n) \}_{n=0}^d$, the Lagrange interpolant is the unique polynomial $L (x)$ of degree $d$ such that, \begin{equation*} L (x_n) = y_n , \; n = 0 ... d \end{equation*} It can be written explicitly as, \begin{equation*} L (x) = \sum_{n=0}^d y_n\, \ell_n (x) \end{equation*} where, \begin{equation*} \ell_n (x) = \prod_{\substack{{m=0}\\ {m \not = n}}}^d \frac{x-x_m}{x_n-x_m} \end{equation*} and $\ell_n (x)$ is referred to as the $n^{th}$ Lagrange basis polynomial. \end{defn} Finally we introduce the moments as well as the Hankel matrix associated with a measure $\alpha(\cdot)$, \begin{defn}{\bf(Moment)} Given an arbitrary measure $\alpha(\cdot)$ on $(\mathbb{R}, \mathcal{B})$, its $n^{th}$ moment $\mu_n$ is defined by the following Lebesgue integral, \begin{equation*} \mu_n = \int x^n \, \d \alpha(x) \end{equation*} whenever it exists. \end{defn} \begin{defn}{\bf (Hankel matrix)} An $(N+1) \times (M+1)$ matrix $H$ is called the $(N+1) \times (M+1)$ Hankel matrix associated with the measure $\alpha( \cdot )$ if its entries take the form, \begin{equation} \begin{pmatrix} \mu_0 & \mu_1 & \cdots & \mu_{M} \\ \mu_1 & \mu_2 & \cdots & \mu_{M+1}\\ \vdots &\vdots &\vdots &\vdots \\ \mu_{N} & \mu_{N+1} & \cdots & \mu_{N+M} \end{pmatrix} \label{eq:hankel} \end{equation} i.e., $H_{ij} = \mu_{i+j}$, where $(\mu_0, \mu_1, \ldots, \mu_{N+M} )$ are the first $(N+M)$ moments of $\alpha(\cdot)$ whenever they exist. \end{defn} With these quantities we can now present the main results associated with classical Gaussian quadratures, \begin{thm}{\bf(Gaussian quadrature)} Consider a positive measure $\alpha(\cdot)$ on $([a,b], \mathcal{B})$ (with $a,b \in \mathbb{R}$ potentially infinity) and a sequence of orthonormal polynomials $\{ p^{(k)} (x) \}_{k=0}^{\infty}$ with respect to $\alpha( \cdot)$. Then, the quadrature rule with nodes $\{ x_n\}_{n=0}^k$ consisting in the zeros of $p^{(k+1)}(x)$ and weights $\{ w_n \}_{n=0}^k$ given by, \begin{equation*} w_n = \int \ell_n (x) \, \d \alpha(x) \end{equation*} integrates polynomials of degree $\leq 2k+1$ exactly. \end{thm} This is a classical result which can be found in \cite{Meurant} for instance. Explicit expression for the error incurred in the case of smooth integrand also exist. To close this section, we introduce a further result characterizing the coefficients of the orthogonal polynomials $\{ p^{(k)}(x) \} $. As we shall see in the next section, this characterization lies at the heart of our scheme, \begin{lemma} Consider a positive measure $\alpha(\cdot)$ on $([a,b], \mathcal{B})$ (with $a,b \in \mathbb{R}$ potentially infinity) and a sequence of orthogonal polynomials $\{ p^{(k)} (x) \}_{k=0}^{\infty}$ with respect to $\alpha( \cdot)$. Then, the coefficients $\{ p^{(k+1)}_n \}_{n=0}^k$ of the $(k+1)^{th}$ orthogonal polynomial $p^{(k+1)}(x)$ satisfy the following Hankel system, \[ Hp = \begin{pmatrix} \mu_0 & \mu_1 & \cdots & \mu_{k+1} \\ \mu_1 & \mu_2 & \cdots & \mu_{k+2}\\ \vdots &\vdots &\vdots &\vdots \\ \mu_{k} & \mu_{k+1} & \cdots & \mu_{2k+1} \end{pmatrix} \begin{pmatrix} p_0^{(k+1)} \\ p_1^{(k+1)} \\ \cdots \\ p_{k}^{(k+1)} \end{pmatrix} = 0\] where $\{ \mu_n \}$ are the moments of the measure $\alpha ( \cdot ) $, whenever they exist. \label{AGQ:characterization} \end{lemma} \begin{proof} First write, \begin{equation*} p^{(k+1)} (x) = \sum_{n=0}^{k+1} p^{(k+1)}_n x^n \end{equation*} Let $0 \leq j \leq k $. Then, from orthogonality we have, \begin{equation*} 0 = \int p^{(k+1)} (x) \, x^j \, \d \alpha(x) = \sum_{n=0}^{k+1} p^{(k+1)}_n \int x^{n+j}\, \d \alpha(x) = \sum_{n=0}^{k+1} p^{(k+1)}_n \mu_{n+j} \end{equation*} Putting all these equations in matrix form provides the desired result. \end{proof} The Hankel matrices associated with positive measures commonly encountered with classical Gaussian quadratures have been the subject of extensive study in the past (known as the moment problem). In some cases, they can be proved to be invertible although extremely ill-conditioned (see $\cite{Shohat}$ for details). On the other hand, less is known per regards to more general measures. In any case, in the event where the resulting Hankel matrix would be invertible, it can be expected to be ill-conditioned. Indeed, as an example it can be shown that for a large class of positive measures, the smallest eigenvalue of the $N \times N$ associated Hankel matrix scales like $\O \left (\frac{ \sqrt{N} }{\sigma^{2N}} \right ) $, where $\sigma$ depends only on the interval considered and is equal to $(1+\sqrt{2})$ for the interval $[-1,1]$ (see \cite{Widom:1966}). The question we treat in the next section is whether such Hankel matrices arising from arbitrary measures can be used to derive Gaussian-like quadratures, and what this inherent ill-conditioning entails. \section{Approximate Gaussian quadrature (AGQ)} \label{sec:agq} In this section, we describe the concept of approximate Gaussian quadrature. For this purpose, we will need the concept of $\epsilon$-quasiorthogonal polynomial, which we introduce for the first time below. Before doing so however, we first point to the following key observation. \bigskip \begin{thm} Let $H$ be a $N \times M$ with rank $0 < d < M $. Then, there exists $D \le d+1$ and a vector $a \not = 0$ such that \begin{equation*} Ha = 0, \quad \text{with $a_i = 0$ for all $i > D$} \end{equation*} \label{AGQ:low_rank} \end{thm} \begin{proof} The rank of $H$ is $d$. Therefore if we consider the first $d+1$ columns for $H$ they are linearly dependent. Denote $D$ the smallest integer such that the first $D$ columns of $H$ are linearly dependent. We have $D \le d+1$ and, by definition, there is $a \neq 0$ such that $Ha = 0$ with $a_i = 0$, $i>D$. \end{proof} \bigskip We also have the following corollary, \begin{cor} \label{AGQ:quasiortho_cor} Assume that the $N \times (N+1)$ Hankel matrix $H$ associated with the measure $\alpha(\cdot)$ exists. If $H$ has rank $d<N$ then there exists a nontrivial polynomial $p(x)$ with degree $(D-1)$ where $D \leq d+1$ such that, \begin{equation*} \int p(x) x^j \, \d \alpha(x) = 0 \end{equation*} for all $j = 0,$ \ldots, $N$. \end{cor} \begin{proof} Let $K$ be such that \begin{equation*} D = \inf \{ 0\leq n \leq N : \mathrm{rank}( H(:, 1:n) )= n \} \end{equation*} where $H(:, 1:n)$ is the matrix containing the first $n$ columns of $H$. By theorem \ref{AGQ:low_rank}, there exists a vector $a \not = 0$ such that $Ha = 0$ and $a_i = 0$ for $i>D$.\\ Let $p(x)$ be the polynomial with coefficients given by $a$, i.e. \begin{equation*} p(x) = \sum_{n=0}^D a_n x^n \end{equation*} Then, \begin{align*} \int p(x) x^j \, \d \alpha(x) &= \int \sum_{n=0}^{D} a_n x^{n+j} \, \d \alpha(x)\\ &= \sum_{n=0}^D a_n \mu_{n+j} \\ & = (H a)_j = 0 \end{align*} since $a$ belongs to the null-space of $H$. \end{proof} The consequences of this corollary are far-reaching and constitute the crux of the scheme presented here. Indeed, although we do not generally expect the Hankel matrix $H$ associated with some measure $\alpha( \cdot)$ to be \emph{exactly} low-rank as in the case of Theorem \ref{AGQ:low_rank} (e.g., $H$ has full rank in the case of classical Gaussian quadratures) we can expect that in some cases $H$ will be \emph{approximately} low rank. In other words, given $0 < \epsilon \ll 1 $ we expect, \begin{equation*} D \approx \max \{ 1 \leq i \leq N : \sigma_i > \epsilon \, \sigma_1 \} \end{equation*} where $\{ \sigma_i \}$ are the singular values of $H$, to be much smaller than $N$, i.e., $D \ll N$. We show for instance in Figure \ref{svd} the first $50$ singular values of the Hankel matrix ($N=250$) associated with the Lebesgue measure in $[-1,1]$. The $y$-axis scales as a logarithm in base $10$, and it is seen that the singular values decay faster than exponentially. \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.9] \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Index of singular value}, ymode=log, width=0.6\textwidth, height=200pt, xmin=0, xmax=50, ymin=1e-18, ymax=1e2, xtick={0, 5, 10, 15, 20, 25, 30, 35, 40, 45, 50}, ytick={1e-18, 1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2, 1e0, 1e2}] \input{SingularValues} \end{axis} \end{tikzpicture}% \end{center} \caption{$\log_{10}$ of singular values of Hankel matrix ($N = 250$) associated with the Lebesgue measure on $[-1,1]$.} \label{svd} \end{figure} In light of the above discussion, we might expect in these circumstances the existence of a polynomial $p(x)$ of degree $D \approx \max \{ 1 \leq i \leq N : \sigma_i > \epsilon \sigma_1 \}$ such that \begin{equation*} \left | \int p(x) x^j \, \d \alpha(x) \right | \lesssim \epsilon \end{equation*} for all $0 \leq j \leq N$, and this leads us to the introduction of the concept of $\epsilon$-quasiorthogonal polynomial which we now define, \begin{defn} A polynomial $p(x)$ is called $\epsilon$-quasiorthogonal of order $N$ with respect to the measure $\alpha(\cdot)$ and the basis $\{ L_n (x) \}$ if, \begin{equation*} \left | \int p(x) \, L_n(x) \, \d \alpha(x) \right | \leq \epsilon \end{equation*} for all $n=0,$ \ldots, $N$. \end{defn} Importantly, this definition imposes no restriction per regards to the measure $\alpha(\cdot)$, in opposition with orthogonal polynomials which demand the measure to be positive (\cite{Szego}). In this sense, the relation described is not one of orthogonality for it is not possible to define a nondegenerate inner-product unless $\alpha( \cdot )$ is positive. This is why we chose the name \emph{quasi}-orthogonal. We also note that given $\epsilon \geq \sigma_N(H)$ such polynomial always exists for it suffices to pick $a$ aligned with the right singular vector associated with the smallest singular value $\sigma_N(H)$. From a computational standpoint, there exists an efficient scheme to find such polynomials given a measure $\alpha( \cdot)$ and some $\epsilon>0$. This is the subject of Section \ref{AGQ:comp}. For the remaining of this section, we will focus on demonstrating how such polynomials can be used to obtain efficient quadratures. As will be shown, the construction of the scheme shares a lot with that of classical Gaussian quadrature. This is what constitutes the origin of the denomination. We will need the following technical lemma which proof is provided in appendix, \begin{lemma} \label{AGQ:tech_lemma} Let $\alpha(\cdot)$ be an arbitrary measure on $(\mathbb{R}, \mathcal{B})$, and \begin{equation*} p(x) =x^{d+1} + \sum_{n=0}^{d} p_n x^n \end{equation*} be a monic $\epsilon$-quasiorthogonal polynomial of degree $d+1$ and order $N$ associated with $\alpha(\cdot)$. Further, let, \begin{equation*} f(x) = \sum_{n=0}^{N+d} f_n \, L_n(x) \end{equation*} be some polynomial of degree $N+d$ and $ \tilde{q}(x) = \sum_{n=0}^{d} \tilde{q}_n x^n$ be the Lagrange interpolant of $q(x)$ associated with the zeros of $p(x)$. Finally, let $r(x) = \sum_{n=0}^{N} r_n x^n$ be the unique polynomial such that $ q(x) - \tilde{q}(x)= p(x) r(x) $. Then, \begin{align*} \sum_{n=0}^N |r_n| \leq \lVert\Gamma^{-1} \bar{q} \rVert_{1} \end{align*} where $\bar{q} =\left( q_{d+1}, q_{d+2}, \ldots, q_{N+d}\right )^T $ and $\Gamma$ is the $N \times N$ Toeplitz matrix such that $[ \Gamma ]_{i,j} = p_{j-i}$ if $0\leq j-i \leq d$ and $0$ otherwise. \label{AGQ:errorlemma} \end{lemma} We are now ready to prove our main theorem. \begin{thm}{\bf(Approximate Gaussian quadrature)} \label{AGQ:AGQ_thm} Consider an arbitrary measure $\alpha(\cdot)$ on $(\mathbb{R}, \mathcal{B})$. Let $p(x)$ be a monic $\epsilon$-quasiorthogonal polynomial of degree $d+1$ and order $N$ with respect to $\alpha( \cdot )$, where $0<\epsilon<1$. Then, the quadrature rule with nodes $\{ x_n \}_{n=0}^{d}$ consisting in the zeros of $p(x)$ and weights $\{ w_n \}_{n=0}^{d}$ given by, \begin{equation} \label{AGQ:AGQ_thm:weight} w_n = \int \ell_n (x) \, \d \alpha(x) \end{equation} where $\ell_n (x)$ is the $n^{th}$Lagrange basis polynomial associated with the nodes, integrates polynomials $q(x)$ of degree $\leq N+d$ with an error bounded by, \begin{equation*} \left | \int q(x) \, \d \alpha(x) - \sum_{n=0}^d w_n \, q(x_n) \right | \leq \lVert\Gamma^{-1} \bar{q} \rVert_{1} \, \epsilon \end{equation*} where $\{ q_n \}_{n=0}^{N+d}$ are the coefficients of $q(x)$, $\bar{q} =\left( q_{d+1} , q_{d+2} , \ldots, q_{N+d}\right )^T $ and $\Gamma$ is the $N \times N$ Toeplitz matrix such that $[ \Gamma ]_{i,j} = p_{j-i}$ if $0\leq j-i \leq d$ and $0$ otherwise. \end{thm} \begin{proof} Let $q(x)$ be a polynomial of degree $ N + d$ and consider the Lagrange interpolant at the nodes $\{ x_n \}$, \begin{equation*} \tilde{q} (x) = \sum_{n=0}^d q(x_n) \ell_n (x) \end{equation*} Then consider, \begin{align*} I = \int \left [ q(x) - \tilde{q}(x) \right ] \, \d \alpha(x) \end{align*} The quantity $[q(x) - \tilde{q}(x)]$ is a polynomial of degree at most $(N+d)$ and has zeros located at each of the nodes $\{ x_n \}_{n=0}^d$. Therefore, by the factorization theorem for polynomials we can write, \begin{equation*} q(x) - \tilde{q}(x) = \prod_{n=0}^d (x - x_n) \, r(x) \end{equation*} where $r(x)$ is a polynomial of degree at most $N$. We further note that $\prod_{n=0}^d (x - x_n)$ is a monic polynomial of degree $d+1$ with zeros at $\{ x_n \}_{n=0}^d$ just as $p(x)$. Since monic polynomials are uniquely characterized by their roots we have, \begin{equation*} \prod_{n=0}^d (x - x_n) = p (x) \end{equation*} Therefore, \begin{align*} |I| &= \left | \int p(x) r(x) \, \d \alpha(x) \right | \leq \sum_{n=0}^N | r_n | \,\left | \int p(x) \, x^n \, \d \alpha(x) \right | \leq \sum_{n=0}^N | r_n | \, \epsilon \\ \end{align*} where we used the $\epsilon$-quasiorthogonality of $p(x)$. Finally, thanks to Lemma \ref{AGQ:tech_lemma} we get, \begin{equation*} \left | \int q(x) \, \d \alpha(x) - \sum_{n=0}^d w_n \, q(x_n) \right | \leq \lVert \Gamma^{-1} \bar{q} \rVert_{1} \epsilon \end{equation*} \end{proof} Interestingly, the above analysis reveals that an AGQ of order $d$ is in fact \emph{exact} for polynomials of degree $\leq d$. Some advantages of AGQ is that there is no need for the measure $\alpha( \cdot )$ to have any specific properties beyond the existence of moments of high-enough order. Furthermore, the problem of the existence and uniqueness of the solution to the Hankel system is of no importance; in fact, the larger the null-space of $H$ the better it is. Both characteristics are in sharp contrast with common wisdom regarding classical Gaussian quadratures. First, the positivity of the measure is key in proving the existence of a sequence of orthogonal polynomials necessary to build a classical quadrature (see \cite{Meurant}, Theorem 2.7). Secondly, the notion of orthogonality is at the heart of modern numerical schemes used to obtain nodes and weights for it gives rise to a three-term recurrence relation that is thoroughly exploited computationally (see \cite{Golub:1969, Meurant}). \subsection{Computational considerations} \label{AGQ:comp} The first computational issue we describe here is that of finding an adequate $\epsilon$-quasiorthogonal polynomials of order $N$ given a measure $\alpha(\cdot)$ on $(\mathbb{R}, \mathcal{B})$, some $N \in \mathbb{N}$ and some value $ 0< \epsilon$. For this purpose, we note that a sufficient condition for a monic polynomial $p(x)$ of degree $(d+1)$ to fall within this category is to satisfy the following inequality, \begin{equation*} \lVert H(N,d) \, \bar{p} + h(d) \rVert_{\infty} \leq \epsilon \end{equation*} where $\bar{p} = [p_0, \, p_1 \, , \ldots, p_{d}]^T$, $p(x) = x^{d+1} + \sum_{n=0}^d p_n x^n$, $ h(d) = [\mu_{d+1}, \, \mu_{d+2} \, , \ldots, \mu_{d+N+1}]^T$ and $H(N,d)$ is the $(N+1)\times (d+1)$ Hankel matrix associated with the measure, i.e. \[ H(N,d) = \begin{pmatrix} \mu_0 & \mu_1 & \cdots & \mu_{d} \\ \mu_1 & \mu_2 & \cdots & \mu_{d+1}\\ \vdots &\vdots &\vdots &\vdots \\ \mu_{N} & \mu_{N+1} & \cdots & \mu_{d+N} \end{pmatrix} \] The proof is analogous to that of Corollary \ref{AGQ:quasiortho_cor} and uses the definition of quasiorthogonal polynomials. This inequality provides a constructive way for finding an $\epsilon$-quasiorthogonal polynomial of small degree. This is described in Algorithm \ref{poly_alg}; note that we replace the $\lVert \cdot \rVert_{\infty}$ norm by the more computationally-friendly $\lVert \cdot \rVert_2$ norm which is equivalent. \begin{algorithm2e}[htbp] Let $(d+1) = \mathrm{rank}(H,\delta)$\\ Solve $ \min_p \lVert H(N,d) \, p + h(d) \rVert_2 $ \\ \While{$\lVert H(N,d) \, p + h(d) \rVert_2 > \epsilon $} { $d = d+1$ \\ Solve $ \min_p \lVert H(N,d) \, p + h(d) \rVert_2 $ } \caption{Pseudo-code to determine $\epsilon$-quasiorthogonal polynomial of order $N$ of small degree for desired accuracy $\delta$.} \label{poly_alg} \end{algorithm2e} {\bf Note: } The quadrature obtained from $p(x)$ integrates polynomials of degree $N+d$ with error prescribed by Theorem \ref{AGQ:AGQ_thm}. This error term involves the norm of the inverse of a matrix $\Gamma$ which is upper-triangular, Toeplitz with diagonal entries all equal to $1$ and remaining entries depending on the coefficients of the polynomial $p(x)$. In order to guarantee that an AGQ integrates polynomials of degree $\leq N+d$ with accuracy $\delta$ say, it is sufficient to set $\epsilon \leq \frac{\delta}{C}$ and constrain $p(x)$ to be such that $\lVert \Gamma^{-1} \rVert_{\infty} \leq C$ for some $C>0$. Upon obtaining some characterization of the set $\mathcal{S}_C := \{ p(x) : \lVert \Gamma^{-1} \rVert_{\infty} \leq C \}$, one could potentially carry out the steps described in Algorithm \ref{poly_alg} while restraining the solution to $\mathcal{S}_C$. One would thus guarantee the accuracy of the AGQ \emph{a priori}. Unfortunately, such characterization is not readily available so one is left with the \emph{a posteriori} estimates of Theorem \ref{AGQ:AGQ_thm}. On the other hand, numerical experiments point to the fact that the product $\lVert \Gamma^{-1} \rVert_{\infty} \, \epsilon$ does indeed decay in a fast manner as a function of the degree of $p(x)$, for $\bar{p}$ the solution of the least-squares problem having the smallest norm in Algorithm \ref{poly_alg}. In short, although AGQ in its current state performs well, some improvements are still possible. This constitutes a topic for future research. Once such polynomial has been obtained, its roots constitute the nodes of the approximate Gaussian quadrature as per Theorem \ref{AGQ:AGQ_thm}. The cost of solving a thin $(N+1) \times (d+1)$ least-squares problem is $\O( [N+1) + (d+1)/3] (d+1)^2 ) $ (see \cite{Golub}). Since in general we expect $d \ll N$ the cost is \emph{linear} in $N$. Also, each step of the while loop constitutes a rank-1 update of the system, so $p$ can be recomputed cheaply. Another great computational aspect of the scheme is the availability of a simple analytical formula for the computation of the weights. Indeed, from Theorem \ref{AGQ:AGQ_thm} we have, \begin{equation*} w_n = \int \ell_n (x) \, \d \alpha(x) = \int \sum_{k=0}^d [\ell_n]_k x^k \, \d \alpha(x) = \sum_{k=0}^d [\ell_n]_k \, \mu_k \end{equation*} where $[\ell_n]_k $ is the $k^{th}$ coefficient of the $n^{th}$ Lagrange basis polynomial $ \ell_n (x)$, which can be obtained cheaply from the zeros of $ \ell_n (x)$, i.e., the nodes of the quadrature. We also noticed that it is generally possible to neglect nodes associated with small weights when such are present. This further reduces the cost of the method. As a final comment, the accuracy of the scheme is highly dependent on the accuracy of the nodes. For this reason, we recommend performing the computations in extended arithmetic. In this paper, we used $Maple^\copyright$ in order to compute the nodes and weights of each approximate quadrature with high precision. \section{Numerical simulations} \label{NS} In this section, we demonstrate the efficiency and the versatility of the scheme through a few numerical examples. In section \ref{NS:classical}, we compare fixed-order approximate Gaussian quadratures (AGQ) with two types of classical Gaussian quadratures (Gauss-Legendre and Gauss-Chebyshev) on monomials $x^n$ of increasing degree and show how it quickly becomes advantageous to use an approximate quadrature in those cases. Then in Section \ref{NS:sing}, we give examples related to functions with an integrable singularity at the origin. In section \ref{NS:trig}, we show how the scheme can be applied to monomials on the complex circle, i.e., functions of the form $e^{\i n x}$ where $0 \leq n$. The resulting quadratures are then used in Section \ref{NS:Beylkin} to obtain approximations of functions through short exponential sums which is related to the method of Beylkin \& Monz\'on \cite{Beylkin:2005, Beylkin:2010}. \subsection{Comparison with classical quadratures} \label{NS:classical} In this section, we compare results between the approximate Gaussian quadrature scheme, the Gauss-Legendre $\left ( \d \alpha(x) = \d x \right ) $ and Gauss-Chebyshev $\left ( \d \alpha(x) = \frac{1}{\sqrt{1-x^2} }\d x \right ) $ quadrature. \subsubsection{Integration of monomials} For this benchmark, we fix the order ($N$ in Section \ref{AGQ:comp}) and study the error in approximating integrals of the form, \begin{equation*} \int_{-1}^1 x^n \, \d \alpha (x) \end{equation*} through quadratures involving different number of nodes ($d$ in Section \ref{AGQ:comp}) where $n$ varies between $0$ to $700$. Numerical results are shown in Figure \ref{GLcompare} and \ref{GCcompare}. They were obtained using $N=350$. The results need to be interpreted carefully. The choice of $N$ represents in effect the polynomial order that would be required to approximate a given function $f(x)$ to some accuracy $\epsilon$. A numerical quadrature will then be able to approximate the integral of $f(x)$ if it can integrate all monomials of degree less than $N$ with accuracy $\epsilon$. In Fig.~\ref{GLcompare} for example, we see that the Gauss-Legendre quadrature is exact to machine precision up to $n=39$. However the error increases rapidly to reach $10^{-3}$ near $n=350$. In contrast, although AGQ is not exact for $n \le 39$, the error up to $n \le 350$ remains lower than $10^{-4}$ with only 20 nodes. As we increase the number of nodes (middle and bottom plots) the gain below $n= 350$ is even more significant. \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0, 50,100,150,200,250,300,350, 400,450,500,550,600,650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{GaussLegendre20} \legend{Gauss-Legendre (20 nodes) \\% Approximate Gaussian Quadrature (20 nodes) \\% } \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0, 50,100,150,200,250,300,350, 400,450,500,550,600,650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{GaussLegendre30} \legend{ Gauss-Legendre (30 nodes) \\% Approximate Gaussian Quadrature (30 nodes) \\% } \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0,50,100,150,200,250,300,350, 400,450,500,550,600, 650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{GaussLegendre40} \legend{ Gauss-Legendre (40 nodes) \\% Approximate Gaussian Quadrature (40 nodes) \\% } \end{axis} \end{tikzpicture}% \end{center} \caption{Comparison between the absolute error incurred in the evaluation of the integral $\int_{-1}^1 x^n \, \d x$ through an Approximate Gaussian Quadrature of order $N=350$ (black) and a Gauss-Legendre quadrature (green) for different number of nodes. Top: 20 nodes, Middle: 30 nodes, Bottom: 40 nodes} \label{GLcompare} \end{figure} \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-16, ymax=0, xtick={0, 50,100,150,200,250,300,350, 400,450,500,550,600, 650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{GaussChebyshev20} \legend{ Gauss-Chebyshev (20 nodes) \\% Approximate Gaussian Quadrature (20 nodes) \\% } \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-16, ymax=0, xtick={0, 50,100,150,200,250,300,350, 400,450,500,550,600,650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{GaussChebyshev30} \legend{ Gauss-Chebyshev (30 nodes) \\% Approximate Gaussian Quadrature (30 nodes) \\% } \end{axis} \end{tikzpicture \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-16, ymax=0, xtick={0,50,100,150,200,250,300,350, 400,450,500,550,600,650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{GaussChebyshev40} \legend{ Gauss-Chebyshev (40 nodes) \\% Approximate Gaussian Quadrature (40 nodes) \\% } \end{axis} \end{tikzpicture}% \end{center} \caption{Comparison between the absolute error incurred in the evaluation of the integral $\int_{-1}^1 x^n \, \frac{1}{\sqrt{1-x^2}} \d x$ through an Approximate Gaussian Quadrature of order $N=350$ (black) and a Gauss-Chebyshev quadrature (green) for different number of nodes. Top: 20 nodes, Middle: 30 nodes, Bottom: 40 nodes} \label{GCcompare} \end{figure} The behavior of AGQ in the top plot around $n \approx 40$ where Gauss-Legendre seems to outperform AGQ is not significant. Indeed if a polynomial of order $n \approx 40$ is sufficient to approximate $f$, we would reduce $N$. This would result in an AGQ quadrature much more accurate in the range $n \in [0,40]$. On Figure \ref{bound:Legendre} and \ref{bound:Chebyshev}, we also compare the theoretical bound obtained in Theorem \ref{AGQ:AGQ_thm} with the actual absolute error obtained through a 30-node AGQ for both the Lebesgue and Chebyshev measures respectively. In both cases, it is seen that the bound provides a reasonable estimate for the behavior of the error. \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=350, ymin=1e-17, ymax=0, xtick={0, 50,100,150,200,250,300,350}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{GaussLegendre30_bound} \legend{ Error bound\\% Approximate Gaussian Quadrature error (30 nodes) \\% } \end{axis} \end{tikzpicture}% \end{center} \caption{Comparison between the absolute error incurred in the evaluation of the integral $\int_{-1}^1 x^n \, \d x$ through a $30$-node Approximate Gaussian Quadrature of order $N=350$ and the error bound introduced in Theorem \ref{AGQ:AGQ_thm} (red)} \label{bound:Legendre} \end{figure} \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=350, ymin=1e-17, ymax=0, xtick={0, 50,100,150,200,250,300,350}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{GaussChebyshev30_bound} \legend{ Error bound\\% Approximate Gaussian Quadrature error (30 nodes) \\% } \end{axis} \end{tikzpicture}% \end{center} \caption{Comparison between the absolute error incurred in the evaluation of the integral $\int_{-1}^1 x^n \, \frac{1}{\sqrt{1-x^2}} \d x$ through a $30$-node Approximate Gaussian Quadrature of order $N=350$ and the error bound introduced in Theorem \ref{AGQ:AGQ_thm} (red).} \label{bound:Chebyshev} \end{figure} Finally, an interesting thing to be noted is that in both cases the nodes associated with the approximate Gaussian quadratures were \emph{real} and the weights were \emph{real and positive}; it is a known fact that this should be the case for classical Gaussian quadratures. However, this is by no means obvious for the case of approximate Gaussian quadratures, and we currently have no theory demonstrating that it is always the case for real positive measures. \subsubsection{General integrands} An important difference between AGQ and Gaussian quadratures is that AGQ takes $N$ as a parameter. $N$ represents in effect the order of a polynomial that can approximate $f(x)$ to the desired accuracy. This is function-dependent and therefore may need to be adjusted in AGQ depending on the integrand, if one wishes to have a near optimal quadrature. Generally speaking, AGQ should be able to outperform a classical Gaussian quadrature in all cases since a Gaussian quadrature is a special case of AGQ, by basically choosing $d=N-1$ where $d$ is the degree of the polynomial. Indeed, this is what we observed in our numerical tests. Whenever the classical Gaussian quadrature or CGQ performs well, no gain is obtained with AGQ. We note that, in this case, the usual numerical techniques to evaluate Gaussian quadrature nodes should be more effective than the numerical procedure we are advocating for AGQ (due to ill-conditioning for too stringent a tolerance as mentioned in the introduction). Conversely, when the convergence of CGQ is slow, AGQ provides a significant improvement. This corresponds to situation where expanding $f$ using polynomials requires terms of high degree and then the approximation of AGQ for high order monomials makes a difference. This is illustrated in the examples below. We used the following integrands to investigate the accuracy of AGQ: \begin{align*} \log \left ( 1- \frac{x}{1.05} \right ) &= - \sum_{n=0}^{\infty} \frac{1}{n (1.05)^n} \,x^n ,\;\;\; |x| \leq 1 \\ \frac{1}{ 1- \frac{x}{1.05} } &= - \sum_{n=0}^{\infty} \frac{1}{(1.05)^n} \,x^n ,\;\;\; |x| \leq 1 \\ e^{-10\,x} &= - \sum_{n=0}^{\infty} \frac{(-10)^n}{n!} \,x^n ,\;\;\; 0\leq x \leq 1 \end{align*} The first two integrand have slowly-decaying coefficients and can be approximated in the interval $[-1,1]$ through a sum containing $\O( \log_{1.05}(1/\epsilon) ) $ terms for an accuracy of $\epsilon$. At $\epsilon$-machine ($\epsilon = 10^{-15}$) this implies approximately $700$ terms. The third integrand has very fast decay, and in this case only $50$ terms are sufficient. For each case, we varied the number of nodes in the quadrature. Then for AGQ, we selected the integer $N$ that gave us the most accurate result. In practice, an algorithm would be required to estimate $N$ numerically but we will not address this question here. Results are show in Table \ref{general:log}--\ref{general:exp}. \begin{table}[htbp] \begin{center} \begin{tabular}{cccc} \toprule Number of nodes & Optimal value for $N$ & AGQ & Gauss-Legendre\\ \midrule 10 & 75 & $2.16 \cdot 10^{-8}$ & $1.39 \cdot 10^{-4}$ \\ 15 & 100 & $1.08 \cdot 10^{-8}$ & $3.94 \cdot 10^{-6}$ \\ 20 & 150 & $2.05 \cdot 10^{-11}$ & $1.26 \cdot 10^{-7}$ \\ 25 & 200 & $3.99 \cdot 10^{-14}$ & $4.31 \cdot 10^{-9}$ \\ 30 & 250 & $1.61 \cdot 10^{-15}$ & $1.54 \cdot 10^{-10}$ \\ \bottomrule \end{tabular} \end{center} \caption{Absolute error incurred by an AGQ and a Gauss-Legendre quadrature for the integration of $f(x) = \log \left ( 1- \frac{x}{1.05} \right ) $ over the interval $[-1,1]$ for various number of nodes.} \label{general:log} \end{table}% \begin{table}[htbp] \begin{center} \begin{tabular}{cccc} \toprule Number of nodes & Optimal value for $N$ & AGQ & Gauss-Legendre\\ \midrule 10 & 75 & $5.81 \cdot 10^{-5}$ & $8.15 \cdot 10^{-3}$ \\ 15 & 100 & $2.20 \cdot 10^{-6}$ & $3.60 \cdot 10^{-4}$ \\ 20 & 150 & $4.26 \cdot 10^{-9}$ & $1.56 \cdot 10^{-5}$ \\ 25 & 200 & $1.58 \cdot 10^{-11}$ & $6.76 \cdot 10^{-7}$ \\ 30 & 250 & $4.01 \cdot 10^{-13}$ & $2.92 \cdot 10^{-8}$ \\ 35 & 300 & $1.77 \cdot 10^{-15}$ & $1.25 \cdot 10^{-9}$ \\ \bottomrule \end{tabular} \end{center} \caption{Absolute error incurred by an AGQ and a Gauss-Legendre quadrature for the integration of $f(x) = \frac{1}{ 1- \frac{x}{1.05} }$ over the interval $[-1,1]$ for various number of nodes.} \label{general:geo} \end{table}% \begin{table}[htbp] \begin{center} \begin{tabular}{cccc} \toprule Number of nodes & Optimal value for $N$ & AGQ & Gauss-Legendre\\ \midrule 5 & 15 & $1.09 \cdot 10^{-6}$ & $8.82 \cdot 10^{-5}$ \\ 7 & 7 & $1.29 \cdot 10^{-7}$ & $1.29 \cdot 10^{-7}$ \\ 10 & 10 & $1.02 \cdot 10^{-12}$ & $1.02 \cdot 10^{-12}$ \\ 12 & 12 & $4.44 \cdot 10^{-16}$ & $4.44 \cdot 10^{-16}$ \\ \bottomrule \end{tabular} \end{center} \caption{Absolute error incurred for $e^{-10\,x} $ over the interval $[0,1]$. In that case, Gauss-Legendre converges very fast and AGQ simply provides a quadrature with the same accuracy. The two methods become essentially identical.} \label{general:exp} \end{table} We observe the superior accuracy of AGQ. The first two cases are challenging for CGQ and AGQ does significantly better. For the last case, CGQ converges extremely fast and then AGQ simply finds that the optimal choice is CGQ and provides an estimate with the same accuracy. In summary, $N$ shoud be adjusted depending on the type of integrand. If the integrand is such that expansions in a polynomial basis possess slowy-decaying coefficients, AGQ will provide significantly greater accuracy. If on the contrary, a polynomial expansion converges very rapidly, both AGQ and CGQ will provide essentially identical (and fast) convergence. We also stress that AGQ can be constructed for a wide range of measures whereas CGQ is restricted to positive measures (weight function) only. \subsection{Singular functions} \label{NS:sing} We show how AGQ can be used to integrate functions with integrable singularities. For this purpose, we consider integrand of the form $x^n \log(x)$ for $x \in (0,1]$ and $0 \leq n \leq 700$. In this case, the integral of interest takes the form, \begin{equation*} \int_{0}^1 x^n \, \log(x) \, \d x \end{equation*} This quantity can either be seen as the integration of $x^n \, \log(x)$ with respect to Lebesgue measure or as the integration of the monomial $x^n$ with respect to the measure $\d \alpha(x) = \log(x) \, \d x$. Considering the latter, we build an AGQ of order $N=350$ with different number of nodes and display the absolute error as a function of the degree $n$ and the number of quadrature points. This is shown in Figure \ref{NS:log}. Note that the bound is not plotted beyond $N=350$ for it is no more valid past this point. \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0, 50,100,150,200,250,300,350,400,450,500,550,600,650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{LogMeasure_5} \legend{ Error bound\\% Approximate Gaussian Quadrature error (5 nodes) \\% } \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0, 50,100,150,200,250,300,350,400,450,500,550,600,650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{LogMeasure_10} \legend{ Error bound\\% Approximate Gaussian Quadrature error (10 nodes) \\% } \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.65] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0, 50,100,150,200,250,300,350,400,450,500,550,600,650,700}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{LogMeasure_15} \legend{ Error bound\\% Approximate Gaussian Quadrature error (15 nodes) \\% } \end{axis} \end{tikzpicture}% \end{center} \caption{Comparison between the absolute error incurred in the evaluation of the integral $\int_{-1}^1 x^n \, \log(x) \d x$ through an Approximate Gaussian Quadrature of order $N=350$ (black) and the error bound introduced in Theorem \ref{AGQ:AGQ_thm} (red) for 5, 10 and 15 nodes. \label{NS:log}} \end{figure} We note that in this case we cannot perform a comparison with a classical Gaussian quadrature for no such quadrature exists as is the case with most measures but a few. \subsection{Quadrature for polynomials on the complex circle} \label{NS:trig} In this section, we are interested in integrands that take the form of trigonometric monomials, i.e., functions of the form, \begin{equation*} f(x) = e^{\i n x} \end{equation*} where $0 \leq n$. As their name conveys, such functions are just homogeneous polynomials $z^n$ in the complex plane which have been restricted to the boundary of the unit circle, i.e., $z = e^{ix}$. Thanks to this close relationship with polynomials on the real axis, one can also develop approximate Gaussian quadratures for such functions as well. In fact it suffices to replace the moments $\mu_n$ by the trigonometric moments, \begin{equation*} \tau_n = \int (e^{\i x})^n \, \d \alpha(x) = \int z^n \, \d \alpha(z) \end{equation*} in all that has been presented above and similar results follow. As an example, we built an AGQ of order $N=350$ for trigonometric polynomials with respect to the Lebesgue measure over the interval $[-1,1]$. The absolute error between our approximation and the exact value of the integral, \begin{equation*} \int_{-1}^1 e^{\i n x} \, \d x = \frac{e^{\i n} - e^{-\i n}}{\i n} \end{equation*} are presented in Figure \ref{NS:trigplot}. There, it is seen that as little as $30$ quadrature points are necessary to integrate a complex exponential with frequency $n=500$ with $\approx 10^{-6}$ accuracy. We also plotted the theoretical bound of Theorem \ref{AGQ:AGQ_thm}. Again, it appears to be a good estimate. \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0,50,100,150,200,250,300,350,400,450,500,550,600,650,700,750,800}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{OscillatoryIntegrand_10} \legend{ Error bound\\% Approximate Gaussian Quadrature (10 nodes) \\% } \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0,50,100,150,200,250,300,350,400,450,500,550,600,650,700,750,800}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{OscillatoryIntegrand_20} \legend{ Error bound\\% Approximate Gaussian Quadrature (20 nodes) \\% } \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ylabel={$\log_{10}$ of absolute error}, xlabel={Degree of monomial ($n$)}, ymode=log, width=0.75\textwidth, height=175pt, xmin=0, xmax=700, ymin=1e-17, ymax=0, xtick={0,50,100,150,200,250,300,350,400,450,500,550,600,650,700,750,800}, ytick={1e-16,1e-14, 1e-12, 1e-10,1e-08, 1e-06, 1e-4, 1e-2}] \input{OscillatoryIntegrand_30} \legend{ Error bound\\% Approximate Gaussian Quadrature (30 nodes) \\% } \end{axis} \end{tikzpicture}% \end{center} \caption{Comparison between the absolute error incurred in the evaluation of the integral $\int_{-1}^1 e^{\i n x} \, \d x$ through an Approximate Gaussian Quadrature of order $N=350$ (black) and the error bound introduced in Theorem \ref{AGQ:AGQ_thm} (red) for various number of nodes. Top: 10 nodes, Middle: 20 nodes, Bottom: 30 nodes. \label{NS:trigplot}} \end{figure} It is interesting to look at the location of the nodes for such quadratures. An example is displayed in Figure \ref{NS:trignodes}. The nodes are shown in the complex plane and appear to lie along a curve which rapidly moves upward from $-1$, slowly moves across, and rapidly moves back to $1$ on the real axis. This does not appear to be a coincidence given the fact that functions of the form $e^{\i n x}$ decay exponentially and do not oscillate along the positive imaginary axis. Thus, the underlying curve could be some sort of path of \emph{least oscillation} in an average sense over $0 \leq n \leq N$. At this point, this is a mere qualitative observation, but might be worth investigating in the future. \begin{figure}[htbp] \begin{center} \includegraphics[width=10cm]{trignodes.png} \caption{Location in the complex plane of the nodes of a 20-node AGQ for trigonometric polynomials. The nodes appear to lie on a smooth curve with positive imaginary part.} \label{NS:trignodes} \end{center} \end{figure} \subsection{Approximation of functions through short exponential sums} \label{NS:Beylkin} In this section, we are interested in the approximation of functions by a short sum of exponentials. That is, given a function $f(x)$ defined over an interval $[a,b]$, we seek some approximation in the form, \begin{equation*} f(x) \approx \sum_{m=0}^d \alpha_m e^{ \beta_m x} \end{equation*} for $x \in [a,b]$, and where $d$ should be as small as possible. Such expansions can be viewed as more efficient representations of functions compared to Fourier transforms as they typically require fewer terms. They can form the starting point for various fast algorithms such as the fast multipole method, hierarchical matrices ($\mathcal H$-matrices), etc. Such techniques are particularly desirable when it comes to the solution of integral equations with translation-invariant kernels (see e.g., \cite{Letourneau:2012,Beylkin:2005}). Very powerful techniques based on dynamical systems and recursion ideas were recently introduced by Beylkin \& Monz\'on \cite{Beylkin:2005, Beylkin:2010} in order to approach this problem. As was mentioned earlier, the latter inspired the current work. We will show how AGQ can be used to derive similar approximations through the discretization of the Fourier transform. The final formulation shares some characteristics with the problem of Beylkin \& Monz\'on that can be stated as follows: given the accuracy $\epsilon > 0$, for a smooth function $f(x)$ find the minimal number of complex weights $w_n$ and nodes $e^{t_m}$ such that, \begin{equation*} \left | f(x) - \sum_m w_m e^{t_m x} \right | < \epsilon \end{equation*} for $x \in I$, $I$ being some interval in $\mathbb{R}$. Their scheme is based on an important result regarding Hankel matrices. Consider a Hankel matrix $H$ associated with a sequence $h_k$ where $h_k = f(x_k)$ are uniform samples of $f$. Assume that the null space of $H$ is non-trivial and consider the polynomial whose coefficients are given by a vector in the null space of $H$. The zeros of this polynomial, $\lambda_i$, satisfy the following property (see e.g., \cite{Boley:1998}), \begin{equation*} h_k = \sum_{i=1}^r \lambda_i^k \,d_i \end{equation*} for some $\{ d_i \}$, where $r$ is at most the number of columns of $H$. With our choice for $h_k$, one obtains, \begin{equation*} f(x_k) = \sum_{i=1}^r d_i \, e^{ \log( \lambda_i ) k } \end{equation*} which naturally extends to an interpolation formula for $f(\cdot)$. In \cite{Beylkin:2005, Beylkin:2010}, the authors search for an approximate formula since in general the matrix $H$ is full rank and therefore no efficient representation, that would yield exactly $f(x_k)$, is possible. To achieve this, Beylkin et al.~\cite{Beylkin:2005, Beylkin:2010} show how $\lambda_i$ can be obtained as the roots of a polynomial whose coefficients are given as the entries of a con-eigenvector $u$, i.e., a vector such that, \begin{equation*} H u = \sigma \overline{u} \end{equation*} $\sigma$ being real and nonnegative. The error is then on the order of $\sigma$. They also show that the weights satisfy a well-conditioned Vandermonde system. As will be seen, both our method and theirs involve a Hankel matrix with entries given by the uniform samples of the function to be approximated over the interval considered. However, the current approach avoids the solution of a con-eigenvalue problem altogether and allows for the direct computation of the weights rather than their computation through the solution of a Vandermonde system. Furthermore, since the quasi-orthogonal polynomial obtained through our scheme has small degree, the number of zeros that must be computed is also much smaller. This results in significant computational savings compared to the former method. The resulting error estimates for both methods are different. Indeed, in the case of \cite{Beylkin:2005} one expects the error to be bounded \emph{uniformly} by an expression on the order of the modulus of the small con-eigenvalue $\sigma$ (Theorem 2, \cite{Beylkin:2005}), and such value can be determined \emph{a priori}. In our case however, the error in \emph{not} uniform (as can be seen from the numerical examples). Furthermore, our current error estimate is \emph{a posteriori}. To begin with, consider a function $f(x) \in \mathcal{L}^2 ( \mathbb{R})$ uniformly sampled at $x_n = a + \frac{n (b-a)}{N}$, $n = 0... (N-1)$ for some $N \in \mathbb{N}$ and $a,b \in \mathbb{R}$, and use the Fourier transform to write, \begin{equation*} f(x_n) = \int_{-\infty}^{\infty} e^{2 \pi \i x_n \xi} \hat{f} (\xi) \, \d \lambda( \xi)=\frac{N}{(b-a)}\, \int_{-\infty}^{\infty} e^{2 \pi \i n \zeta} \, e^{2 \pi \i a \zeta} \hat{f} \left (\frac{N}{b-a} \zeta \right ) \, \d \lambda( \zeta) \end{equation*} where $\hat{f} (\xi)$ denotes the Fourier transform of $f(x)$, and $\lambda( \cdot )$ is the Lebesgue measure. We note that $$\frac{N}{b-a} \, e^{2 \pi \i a \zeta} \hat{f} \left (\frac{N}{b-a} \zeta \right )$$ can be seen as a Radon-Nykodym derivative of a certain measure $\alpha( \cdot)$ absolutely continuous with respect to Lebesgue measure (see \cite{Cohn}), i.e., \begin{equation*} \frac{\d \alpha }{ \d \lambda} (\zeta) = \frac{N}{b-a} \, e^{2 \pi \i a \zeta} \hat{f} \left (\frac{N}{b-a} \zeta \right ) \end{equation*} With this measure we have, \begin{equation*} f(x_n) = \int_{-\infty}^{\infty} e^{2 \pi \i n \zeta} \, \d \alpha(\zeta) , \;\; n = 0... N \end{equation*} which is perfectly well-suited for discretization through an approximate Gaussian quadrature as described in the previous section. To find such quadrature, we first need the trigonometric moments of the measure. These moments turn out to have a very simple form. Indeed, a quick look at their definition shows that, \begin{equation*} \tau_n = \int_{\mathbb{T}} e^{\i n \zeta} \, \d \alpha(\zeta) = \int_{\mathbb{T}} e^{\i n \zeta} \, \left [ \frac{N}{b-a} \, e^{2 \pi \i a \zeta} \hat{f} \left (\frac{N}{b-a} \zeta \right ) \right ] \d \lambda(\zeta) = f \left( a + n \frac{(b-a)}{N} \right ) \end{equation*} At this point, we note that the Hankel matrix arising from such moments is exactly the same as the one described in \cite{Beylkin:2005} as previously mentioned. Finally, the nodes $\{ w_n \}$ can be obtained through Eq.\eqref{AGQ:AGQ_thm:weight}. In the end, we obtain \begin{equation*} f(x_n) \approx \sum_{m=0}^d w_m \, e^{ \i n \zeta_m } , \;\;\; n=1,...,N \end{equation*} with error bounded by the expression provided in Theorem \ref{AGQ:AGQ_thm}. To obtain an approximation to $f(x)$ in all of $[a,b]$, we simply allow $\frac{n}{N}$ to vary continuously so that \begin{equation*} \frac{n}{N} = \frac{x-a}{b-a} \end{equation*} for $x \in [a,b]$ and write, \begin{align*} f(x) &\approx \sum_{m=0}^d \alpha_m \, e^{ \i \beta_m x } \\ \alpha_m &= w_m \, e^{-\i \frac{a}{b-a} N \xi_m } \\ \beta_m &= \frac{1}{b-a} N \xi_m \end{align*} When $x$ corresponds to a sample, i.e., $x = x_n$ for some $n$, this reduces to the previous expression. However, when $x$ lies between two samples this last formula should be seen as an interpolation. We do not currently have the complete theory describing the interpolation error. However, it was observed numerically that such error is generally of the same order as that associated with the closest sample whenever the function $f(x)$ is sufficiently oversampled. Numerical examples are provided below. At this point, we describe an algorithm for the construction of such an approximation. The description can be found in pseudo-code in Algorithm \ref{approx_alg}. \begin{algorithm2e}[htbp] \caption{Pseudo-code to for the construction of a short exponential sum approximation of a function $f( \cdot )$ in an interval. \label{approx_alg}} Pick $N \in \mathbb{N}$ sufficiently large (beyond the Nyquist rate)\\ Compute $ \tau_n = f \left( a + n \frac{(b-a)}{N} \right )$\\ Build the Hankel matrix $H_{i,j} = \tau_{i+j}$ for $i,j = 0 .. N$\\ Proceed as described in Algorithm \ref{poly_alg} to find $p(x)$\\ Compute $\{ x_n \}$, the nodes/zeros of $p(x)$ \\ Compute weights $w_n$ following Eq. \eqref{AGQ:AGQ_thm:weight} \\ Build approximation: $ \sum_{n} w_n \, e^{ \i \frac{(x-a)}{(b-a)} N \, \frac{\log(x_n)}{\i} }$ \end{algorithm2e} We now provide a few examples for the representation of some oscillatory functions: the Bessel functions of the first kind $J_{\nu} (100 \pi \, x )$ over the interval $[0,1]$ and for orders $ \nu \in \{ 0, 25 \}$. Such functions are relevant in problems involving the scattering of waves in two dimensions for instance. In both cases, the order of the AGQ is $N=400$ (note that the spectrum of both functions is bounded by about $400 \approx 100 \pi$) and a $40$-terms approximation is obtained using the scheme just introduced. The results are presented in Figure \ref{bessel0} and \ref{bessel25} respectively. Agreement within $10^{-10}$ and $10^{-7}$ absolute error is observed in each cases respectively. It should also be noted that the number of terms lies much below what should be expected with a standard Fourier series given the nature of the oscillations. \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ width=0.75\textwidth, height=0.5\textwidth, xmin=0, xmax=1, ymin=-1., ymax=1, xtick={0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}, ytick={-1,-0.5,0,0.5,1}] \input{Bessel0} \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ymode=log, width=0.75\textwidth, height=0.5\textwidth, xmin=0, xmax=1, ymin=1e-19, ymax=1e-10, xtick={0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}, ytick={1e-19, 1e-18, 1e-17, 1e-16, 1e-15, 1e-14, 1e-13, 1e-12, 1e-11, 1e-10}] \input{Bessel0_error} \end{axis} \end{tikzpicture}% \end{center} \caption{ $40$-term exponential sum approximation of Bessel function of the first kind of order $0$ ($J_{0} (100\pi x)$) in $[0,1]$ (top) and absolute error (bottom) } \label{bessel0} \end{figure} \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ width=0.75\textwidth, height=0.5\textwidth, xmin=0, xmax=1, ymin=-0.25, ymax=0.25, xtick={-1, -0.8, -0.6, -0.4, -0.2, 0, 0.2,0.4, 0.6, 0.8, 1}, ytick={-0.3, -0.2, -0.1, 0., 0.1, 0.2, 0.3}] \input{Bessel25} \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ymode=log, width=0.75\textwidth, height=0.5\textwidth, xmin=0, xmax=1, ymin=1e-17, ymax=1e-7, xtick={-1, -0.8, -0.6, -0.4, -0.2, 0, 0.2,0.4, 0.6, 0.8, 1}, ytick={1e-19, 1e-18, 1e-17, 1e-16, 1e-15, 1e-14, 1e-13, 1e-12, 1e-11, 1e-10, 1e-9, 1e-8, 1e-7, 1e-6}] \input{Bessel25_error} \end{axis} \end{tikzpicture}% \end{center} \caption{ $40$-term exponential sum approximation of Bessel function of the first kind of order $25$ ($J_{25} (100 \pi x)$) in $[0,1]$ (top) and absolute error (bottom) } \label{bessel25} \end{figure} As a final example, we chose to represent the Dirichlet kernel, \begin{equation*} D_N (x) = \sum_{k=-N}^N e^{\i k x} = \frac{\sin \left ( \pi( N + 1/2) \, x \right ) }{\sin \left ( \pi/2 \, x \right )} \end{equation*} over the interval $[-1,1]$. When applied through convolution, the Dirichlet kernel acts as a low-frequency filter. In this sense, a short exponential sum approximation can be used to speed up the filtering process. We picked $N = 200$. To obtain the approximation, we proceeded as described in \cite{Beylkin:2005} and went on to first approximate, \begin{equation*} G_{200} (x) = \sum_{k \geq 0 } \frac{\sin(200 \pi (x+k))}{200 \pi (x+k)} \end{equation*} through a 40-term exponential sum and then built the Dirichlet kernel through the identity, \begin{equation*} D_{200} (x) = G_{200} (x) + G_{200} (1-x) \end{equation*} resulting in a 80-term approximation. It is shown in Figure \ref{dirichlet}. The error is non-uniform as expected from Theorem \ref{AGQ:AGQ_thm} but still remains below $10^{-7}$ for all values in the interval. \begin{figure}[htbp] \begin{center} \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ width=0.75\textwidth, height=0.5\textwidth, xmin=0, xmax=1, ymin=-0.25, ymax=0.25, xtick={0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}, ytick={-0.25, -0.2, -0.15, -0.1, -0.05, 0., 0.05, 0.1, 0.15, 0.2, 0.25}] \input{Dirichlet200} \end{axis} \end{tikzpicture}% \begin{tikzpicture}[scale=0.75] \pgfplotsset{every axis legend/.append style={at={(1.1,0.4)},anchor=north}} \begin{axis}[ ymode=log, width=0.75\textwidth, height=0.5\textwidth, xmin=0, xmax=1, ymin=1e-12, ymax=1e-7, xtick={0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}, ytick={1e-19, 1e-18, 1e-17, 1e-16, 1e-15, 1e-14, 1e-13, 1e-12, 1e-11, 1e-10, 1e-9, 1e-8, 1e-7, 1e-6}] \input{Dirichlet200_error} \end{axis} \end{tikzpicture}% \end{center} \caption{ $80$-term exponential sum approximation of the Dirichlet kernel of order $200$ ($D_{200} ( x)$) in $[-1,1]$ (top) and absolute error (bottom) } \label{dirichlet} \end{figure} \section{Conclusion} We have introduced a new type of quadrature closely related to Gaussian quadratures but which use the concept of $\epsilon$-quasiorthogonality to reduce the number of quadrature nodes and weights. Such quadratures have desirable computational properties and can be applied to a family much broader than that targeted by classical Gaussian quadratures. We have provided the theory for the existence of such quadratures and have provided error estimates together with practical ways of constructing them. We have also carried out various numerical examples displaying the versatility and performance of the method. Finally, we have described how AGQ can be used to approximate functions through short exponential sums and provided further numerical examples in these cases. \section{Acknowledgements} The authors would like to thank Professor Ying Wu from King Abdullah University of Science and Technology (KAUST) for supporting this research through her grant as well as the National Sciences and Engineering Research Council of Canada (NSERC) for their financial support. \newpage

https://arxiv.org/abs/1406.3991

On linear and quadratic Lipschitz bounds for twice continuously differentiable functions

Lower and upper bounds for a given function are important in many mathematical and engineering contexts, where they often serve as a base for both analysis and application. In this short paper, we derive piecewise linear and quadratic bounds that are stated in terms of the Lipschitz constants of the function and the Lipschitz constants of its partial derivatives, and serve to bound the function's evolution over a compact set. While the results follow from basic mathematical principles and are certainly not new, we present them as they are, from our experience, very difficult to find explicitly either in the literature or in most analysis textbooks.

\section{Overview} \label{summary} We consider a function $f : \mathbb{R}^n \rightarrow \mathbb{R}$ of the variables $x \in \mathbb{R}^n$ that is twice continuously differentiable ($C^2$) over an open set containing the compact set $\mathcal{X}$. Because $f$ is $C^2$ over $\mathcal{X}$, its first and second derivatives on this set exist and must be bounded by the Lipschitz constants \vspace{-4mm} \begin{equation}\label{eq:lip1} \underline \kappa_i < \frac{\partial f}{\partial x_i} \Big |_{x} < \overline \kappa_i, \hspace{15mm} i = 1,...,n \end{equation} \vspace{-4mm} \begin{equation}\label{eq:lip2} \underline M_{ij} < \frac{\partial^2 f}{\partial x_j \partial x_i} \Big |_{x} < \overline M_{ij}, \hspace{6mm} i,j = 1,...,n \end{equation} \noindent for all $x \in \mathcal{X}$. The evolution of $f$ between any two points $x_a, x_b \in \mathcal{X}$ may then be bounded as \vspace{-4mm} \begin{equation}\label{eq:bound1L} f(x_b) - f(x_a) \geq \displaystyle \sum_{i=1}^n \mathop {\min} \left[ \begin{array}{l} \underline \kappa_i (x_{b,i} - x_{a,i}), \\ \overline \kappa_i (x_{b,i} - x_{a,i}) \end{array} \right], \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound1U} f(x_b) - f(x_a) \leq \displaystyle \sum_{i=1}^n \mathop {\max} \left[ \begin{array}{l} \underline \kappa_i (x_{b,i} - x_{a,i}), \\ \overline \kappa_i (x_{b,i} - x_{a,i}) \end{array} \right], \end{equation} \noindent where $x_{a,i}$ and $x_{b,i}$ denote the $i^{\rm th}$ elements of the vectors $x_a$ and $x_b$, respectively. The bounds (\ref{eq:bound1L}) and (\ref{eq:bound1U}) are piecewise linear in $x$. Alternatively, one may also use the piecewise quadratic bounds \vspace{-4mm} \begin{equation}\label{eq:bound2L} \begin{array}{l} f(x_b) - f(x_a) \geq \nabla f (x_a)^T (x_b - x_a) + \\ \hspace{10mm}\displaystyle \frac{1}{2} \sum_{i=1}^n \sum_{j=1}^n \mathop {\min} \left[ \begin{array}{l} \underline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}), \\ \overline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \end{array} \right], \end{array} \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound2U} \begin{array}{l} f(x_b) - f(x_a) \leq \nabla f (x_a)^T (x_b - x_a) + \\ \hspace{10mm}\displaystyle \frac{1}{2} \sum_{i=1}^n \sum_{j=1}^n \mathop {\max} \left[ \begin{array}{l} \underline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}), \\ \overline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \end{array} \right], \end{array} \end{equation} \noindent which are locally less conservative but also require more knowledge in the form of both the gradient and the Lipschitz constants on the partial derivatives of $f$. While one may generalize this pattern to even higher orders, we will content ourselves with the linear and quadratic cases as we believe these to be sufficient for most applications -- see, however, \cite{Cartis:13} for a discussion of the cubic case. \section{Derivation of the Linear Bounds} To limit our analysis to a single dimension, we will consider the line segment between $x_a$ and $x_b$. The following one-dimensional parameterization is used: \vspace{-4mm} \begin{equation}\label{eq:Qproof1} \hat f (\gamma) = f(x (\gamma)), \end{equation} \noindent with $x (\gamma) = x_a + \gamma (x_b - x_a), \; \gamma \in [0,1]$. As $f$ is $C^2$, it follows that $\hat f$ is as well, which allows us to use the Taylor series expansion between $\gamma = 0$ and $\gamma = 1$, together with the mean-value theorem \cite{Korn:00}, to state: \vspace{-4mm} \begin{equation}\label{eq:Lproof2} \begin{array}{l} \hat f (1) = \hat f (0) + \displaystyle \frac{d \hat f}{d \gamma} \Big |_{\tilde \gamma} \end{array}, \end{equation} \noindent for some $\tilde \gamma \in (0,1)$. We proceed to define the first-order derivative in terms of the original function $f$. To do this we apply the chain rule: \vspace{-4mm} \begin{equation}\label{eq:Lproof3} \displaystyle \frac{d \hat f}{d \gamma} \Big |_{\gamma} = \displaystyle \mathop {\sum} \limits_{i = 1}^{n} \frac{\partial f}{\partial x_i} \Big |_{x (\gamma)} \frac{d x_i}{d \gamma} \Big |_{\gamma} = \displaystyle \mathop {\sum} \limits_{i = 1}^{n} \frac{\partial f}{\partial x_i} \Big |_{x (\gamma)} (x_{b,i}-x_{a,i}). \end{equation} Noting that $\hat f (0) = f(x_a)$ and $\hat f (1) = f(x_b)$, one may substitute (\ref{eq:Lproof3}) into (\ref{eq:Lproof2}) to obtain \vspace{-4mm} \begin{equation}\label{eq:Lproof4} \begin{array}{l} f (x_b) = f (x_a) + \displaystyle \mathop {\sum} \limits_{i = 1}^{n} \frac{\partial f}{\partial x_i} \Big |_{x (\tilde \gamma)} (x_{b,i}-x_{a,i}) \end{array}. \end{equation} Because $ x(\tilde \gamma) \in \mathcal{X}$, we may use (\ref{eq:lip1}) to bound the individual summation components as \vspace{-4mm} \begin{equation}\label{eq:mvalbound1} \begin{array}{l} x_{b,i} - x_{a,i} \geq 0 \Leftrightarrow \\ \hspace{2mm} \underline \kappa_i (x_{b,i} - x_{a,i}) \leq \displaystyle \frac{\partial f}{\partial x_i} \Big |_{x(\tilde \gamma)} (x_{b,i} - x_{a,i} ) \leq \overline \kappa_i (x_{b,i} - x_{a,i}), \\ x_{b,i} - x_{a,i} \leq 0 \Leftrightarrow \\ \hspace{2mm} \overline \kappa_i (x_{b,i} - x_{a,i}) \leq \displaystyle \frac{\partial f}{\partial x_i} \Big |_{x(\tilde \gamma)} (x_{b,i} - x_{a,i} ) \leq \underline \kappa_i (x_{b,i} - x_{a,i}), \end{array} \end{equation} \noindent or, to account for both cases, as \vspace{-4mm} \begin{equation}\label{eq:mvalbound1pw} \begin{array}{l} \mathop {\min} \left[ \begin{array}{l} \underline \kappa_i (x_{b,i} - x_{a,i}), \\ \overline \kappa_i (x_{b,i} - x_{a,i}) \end{array} \right] \leq \displaystyle \frac{\partial f}{\partial x_i} \Big |_{x (\tilde \gamma)} (x_{b,i} - x_{a,i} ) \\ \hspace{35mm}\leq \mathop {\max} \left[ \begin{array}{l} \underline \kappa_i (x_{b,i} - x_{a,i}), \\ \overline \kappa_i (x_{b,i} - x_{a,i}) \end{array} \right]. \end{array} \end{equation} Substituting this result into (\ref{eq:Lproof4}) then yields (\ref{eq:bound1L}) and (\ref{eq:bound1U}). \section{Derivation of the Quadratic Bounds} The derivation is similar to that of the linear case, and simply involves taking the Taylor series expansion one degree higher, with \vspace{-4mm} \begin{equation}\label{eq:Qproof2} \begin{array}{l} \hat f (1) = \hat f (0) + \displaystyle \frac{d \hat f}{d \gamma} \Big |_{0} + \frac{1}{2} \frac{d^2 \hat f}{d\gamma^2} \Big |_{\tilde \gamma} \end{array} \end{equation} \noindent for some $\tilde \gamma \in (0,1)$. Applying the chain rule \vspace{-4mm} \begin{equation}\label{eq:Qproof3} \displaystyle \frac{d \hat f}{d \gamma} \Big |_{\gamma} = \displaystyle \mathop {\sum} \limits_{i = 1}^{n} \frac{\partial f}{\partial x_i} \Big |_{x (\gamma)} \frac{d x_i}{d \gamma} \Big |_{\gamma} = \nabla f(x(\gamma))^T (x_b - x_a) \end{equation} \noindent and then differentiating once more with respect to $\gamma$ yields \begin{equation}\label{eq:Qproof4} \begin{array}{l} \displaystyle \frac{d^2 \hat f}{d \gamma^2} \Big |_{\gamma} = \displaystyle \mathop {\sum} \limits_{i = 1}^{n} \frac{d}{d \gamma} \left( \frac{\partial f}{\partial x_i} \Big |_{x (\gamma)} \frac{d x_i}{d \gamma} \Big |_{\gamma} \right) \\ \hspace{20mm}= \displaystyle \mathop {\sum} \limits_{i = 1}^{n} \frac{d}{d \gamma} \left( \frac{\partial f}{\partial x_i} \Big |_{x (\gamma)} \right) \frac{d x_i}{d \gamma} \Big |_{\gamma}\;, \end{array} \end{equation} \noindent where we have ignored the terms corresponding to $d^2 x_i / d \gamma^2$ as all such terms are 0. Applying the chain rule again yields \vspace{-4mm} \begin{equation}\label{eq:Qproof5} \begin{array}{l} \displaystyle \frac{d^2 \hat f}{d \gamma^2} \Big |_{\gamma} = \displaystyle \mathop {\sum} \limits_{i = 1}^{n} \mathop {\sum} \limits_{j = 1}^{n} \frac{\partial^2 f}{\partial x_j \partial x_i} \Big |_{x (\gamma)} \frac{d x_j}{d \gamma} \Big |_{\gamma} \frac{d x_i}{d \gamma} \Big |_{\gamma} \\ \hspace{10mm} = \displaystyle \mathop {\sum} \limits_{i = 1}^{n} \mathop {\sum} \limits_{j = 1}^{n} \frac{\partial^2 f}{\partial x_j \partial x_i} \Big |_{x (\gamma)} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}). \end{array} \end{equation} Substituting the results of (\ref{eq:Qproof3}) and (\ref{eq:Qproof5}) into (\ref{eq:Qproof2}), noting that $\hat f (0) = f(x_a)$ and $\hat f (1) = f(x_b)$, and rearranging then leads to \vspace{-4mm} \begin{equation}\label{eq:Qproof6} \begin{array}{l} f (x_b) - f (x_a) = \nabla f(x_a)^T (x_b-x_a) + \\ \hspace{10mm}\displaystyle \frac{1}{2} \mathop {\sum} \limits_{i = 1}^{n} \mathop {\sum} \limits_{j = 1}^{n} \frac{\partial^2 f}{\partial x_j \partial x_i} \Big |_{x (\tilde \gamma)} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}). \end{array} \end{equation} The bounds on the quadratic term are derived in a manner analogous to what was done in the linear case: \vspace{-4mm} \begin{equation}\label{eq:mvalbound2} \begin{array}{l} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \geq 0 \Leftrightarrow \vspace{2mm} \\ \hspace{5mm} \underline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \leq \vspace{2mm} \\ \hspace{10mm}\displaystyle \frac{\partial^2 f}{\partial x_j \partial x_i} \Big |_{x(\tilde \gamma)} (x_{b,i} - x_{a,i} )(x_{b,j} - x_{a,j}) \leq \vspace{2mm}\\ \hspace{35mm}\overline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}), \vspace{2mm}\\ (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \leq 0 \Leftrightarrow \vspace{2mm}\\ \hspace{5mm} \overline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \leq \vspace{2mm}\\ \hspace{10mm}\displaystyle \frac{\partial^2 f}{\partial x_j \partial x_i} \Big |_{x(\tilde \gamma)} (x_{b,i} - x_{a,i} )(x_{b,j} - x_{a,j}) \leq \vspace{2mm}\\ \hspace{35mm}\underline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}), \end{array} \end{equation} \noindent and, taking both cases into account, we obtain \vspace{-4mm} \begin{equation}\label{eq:mvalbound2pw} \begin{array}{l} \mathop {\min} \left[ \begin{array}{l} \underline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}), \\ \overline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \end{array} \right] \leq \vspace{2mm} \\ \hspace{10mm}\displaystyle \frac{\partial^2 f}{\partial x_j \partial x_i} \Big |_{x(\tilde \gamma)} (x_{b,i} - x_{a,i} )(x_{b,j} - x_{a,j}) \leq \vspace{2mm}\\ \hspace{20mm}\mathop {\max} \left[ \begin{array}{l} \underline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}), \\ \overline M_{ij} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \end{array} \right], \end{array} \end{equation} \noindent which may be substituted into (\ref{eq:Qproof6}) to yield (\ref{eq:bound2L}) and (\ref{eq:bound2U}). \section{Other Versions} The bounds (\ref{eq:bound1L})-(\ref{eq:bound2U}) allow a good degree of flexibility by considering both lower and upper bounds on the different partial derivatives. Such flexibility may be useful in certain engineering contexts, where \emph{a priori} knowledge about the system in consideration may be used coherently with the lower and upper bounds on the derivatives \cite{Bunin:13}. However, there are also contexts where these bounds may be needed for purely conceptual reasons and where simpler versions are desired. For example, one might want to suppose \cite{Bunin:13b}: \vspace{-4mm} \begin{equation}\label{eq:symm} \begin{array}{l} \kappa_i = \overline \kappa_i = - \underline \kappa_i \;, \vspace{2mm} \\ M_{ij} = \overline M_{ij} = -\underline M_{ij} \;, \end{array} \end{equation} \noindent which, if we follow the same steps as before, yields \vspace{-4mm} \begin{equation}\label{eq:bound1Ls} f(x_b) - f(x_a) \geq - \displaystyle \sum_{i=1}^n \kappa_i | x_{b,i} - x_{a,i} |, \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound1Us} f(x_b) - f(x_a) \leq \displaystyle \sum_{i=1}^n \kappa_i | x_{b,i} - x_{a,i} |, \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound2Ls} \begin{array}{l} f(x_b) - f(x_a) \geq \nabla f (x_a)^T (x_b - x_a) - \\ \hspace{10mm}\displaystyle \frac{1}{2} \sum_{i=1}^n \sum_{j=1}^n M_{ij} | (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) |, \end{array} \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound2Us} \begin{array}{l} f(x_b) - f(x_a) \leq \nabla f (x_a)^T (x_b - x_a) + \\ \hspace{10mm}\displaystyle \frac{1}{2} \sum_{i=1}^n \sum_{j=1}^n M_{ij} | (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) |. \end{array} \end{equation} One may take this one step further and define the bounds with respect to some standard norms. Defining \vspace{-4mm} \begin{equation}\label{eq:maxkap} \kappa = \mathop {\max} \limits_{i=1,...,n} \kappa_i, \end{equation} \noindent the bounds (\ref{eq:bound1Ls}) and (\ref{eq:bound1Us}) become \vspace{-4mm} \begin{equation}\label{eq:bound1Ls2} f(x_b) - f(x_a) \geq - \kappa \| x_{b} - x_{a} \|_1, \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound1Us2} f(x_b) - f(x_a) \leq \displaystyle \kappa \| x_{b} - x_{a} \|_1. \end{equation} For Bounds (\ref{eq:bound2Ls}) and (\ref{eq:bound2Us}), we may consider the following derivation: \vspace{-4mm} \begin{equation}\label{eq:Mup} \begin{array}{l} \displaystyle \sum_{i=1}^n \sum_{j=1}^n M_{ij} | (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) | \\ \displaystyle \leq \sum_{i=1}^n \sum_{j=1}^n M_{ij} | x_{b,i} - x_{a,i}||x_{b,j} - x_{a,j} | \\ \displaystyle \leq \sum_{i=1}^n \sum_{j=1}^n M_{ij} ( x_{b,i} - x_{a,i})^2, \end{array} \end{equation} \noindent which, with \vspace{-4mm} \begin{equation}\label{eq:maxM} M = \mathop {\max} \limits_{i=1,...,n} \sum_{j=1}^n M_{ij}, \end{equation} \noindent allows for (\ref{eq:bound2Ls}) and (\ref{eq:bound2Us}) to be simplified to: \vspace{-4mm} \begin{equation}\label{eq:bound2Ls2} f(x_b) - f(x_a) \geq \nabla f (x_a)^T (x_b - x_a) - \frac{1}{2} M \| x_{b} - x_{a} \|_2^2, \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound2Us2} f(x_b) - f(x_a) \leq \nabla f (x_a)^T (x_b - x_a) + \frac{1}{2} M \| x_{b} - x_{a} \|_2^2. \end{equation} It may also be shown that the bounds (\ref{eq:bound1L}), (\ref{eq:bound1U}), (\ref{eq:bound2L}), (\ref{eq:bound2U}), (\ref{eq:bound1Ls}), (\ref{eq:bound1Us}), (\ref{eq:bound2Ls}), (\ref{eq:bound2Us}), (\ref{eq:bound1Ls2}), (\ref{eq:bound1Us2}), (\ref{eq:bound2Ls2}), and (\ref{eq:bound2Us2}) all hold with \emph{strict} inequality whenever $x_a \neq x_b$. This follows from (\ref{eq:mvalbound1}) and (\ref{eq:mvalbound2}). We also refer the reader to \cite{Cartis:13} for more alternatives. \section{Local Bounds} As derived, the presented bounds are valid for any arbitrary pair $x_a ,x_b \in \mathcal{X}$, which follows from the validity of the Lipschitz constants over all of $\mathcal{X}$. In certain applications, this globality may, however, add unnecessary conservatism and thus motivate local relaxations \cite{Bunin:13}. Noting that the derivations of the bounds only require them to be valid on the line between $x_a$ and $x_b$, let us define the local Lipschitz constants with respect to these two points in particular as \vspace{-4mm} \begin{equation}\label{eq:lip1loc} \underline \kappa_i^{a,b} < \frac{\partial f}{\partial x_i} \Big |_{x} < \overline \kappa_i^{a,b}, \hspace{10mm} i = 1,...,n, \;\;\; \forall x \in \mathcal{X}_{a,b}, \end{equation} \vspace{-4mm} \begin{equation}\label{eq:lip2loc} \underline M_{ij}^{a,b} < \frac{\partial^2 f}{\partial x_j \partial x_i} \Big |_{x} < \overline M_{ij}^{a,b}, \hspace{3mm} i,j = 1,...,n, \;\;\; \forall x \in \mathcal{X}_{a,b}, \end{equation} \noindent with \vspace{-4mm} \begin{equation}\label{eq:locspace} \mathcal{X}_{a,b} = \{ x_a + \gamma (x_b - x_a) : \gamma \in [0,1] \}. \end{equation} This then yields the corresponding local versions of (\ref{eq:bound1L})-(\ref{eq:bound2U}): \vspace{-4mm} \begin{equation}\label{eq:bound1Lloc} f(x_b) - f(x_a) \geq \displaystyle \sum_{i=1}^n \mathop {\min} \left[ \begin{array}{l} \underline \kappa_i^{a,b} (x_{b,i} - x_{a,i}), \\ \overline \kappa_i^{a,b} (x_{b,i} - x_{a,i}) \end{array} \right], \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound1Uloc} f(x_b) - f(x_a) \leq \displaystyle \sum_{i=1}^n \mathop {\max} \left[ \begin{array}{l} \underline \kappa_i^{a,b} (x_{b,i} - x_{a,i}), \\ \overline \kappa_i^{a,b} (x_{b,i} - x_{a,i}) \end{array} \right], \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound2Lloc} \begin{array}{l} f(x_b) - f(x_a) \geq \nabla f (x_a)^T (x_b - x_a) + \\ \hspace{10mm}\displaystyle \frac{1}{2} \sum_{i=1}^n \sum_{j=1}^n \mathop {\min} \left[ \begin{array}{l} \underline M_{ij}^{a,b} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}), \\ \overline M_{ij}^{a,b} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \end{array} \right], \end{array} \end{equation} \vspace{-4mm} \begin{equation}\label{eq:bound2Uloc} \begin{array}{l} f(x_b) - f(x_a) \leq \nabla f (x_a)^T (x_b - x_a) + \\ \hspace{10mm}\displaystyle \frac{1}{2} \sum_{i=1}^n \sum_{j=1}^n \mathop {\max} \left[ \begin{array}{l} \underline M_{ij}^{a,b} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}), \\ \overline M_{ij}^{a,b} (x_{b,i} - x_{a,i})(x_{b,j} - x_{a,j}) \end{array} \right]. \end{array} \end{equation}

https://arxiv.org/abs/0801.0418

Invariant tensors and graphs

We describe a correspondence between GL_n-invariant tensors and graphs, and show how this correspondence accomodates various types of symmetries and orientations.

\section*{Introduction} Let $V$ be a finite dimensional vector space over a field ${\mathbf k}$ of characteristic zero and ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$ the group of invertible linear endomorphisms of $V$. The classical (Co)Invariant Tensor Theorem recalled in Section~\ref{s1} states that the space of ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-invariant linear maps between tensor products of copies of $V$ is generated by specific `elementary invariant tensors' and that these elementary tensors are linearly independent if the dimension of $V$ is big enough. We will observe that elementary invariant tensors are in one-to-one correspondence with contraction schemes for indices which are, in turn, described by graphs. We then show how this translation between invariant tensors and linear combination of graphs accommodates various types of symmetries and orientations. The above type of description of invariant tensors by graphs was systematically used by M.~Kontsevich in his seminal paper~\cite{kontsevich:93}. Graphs representing tensors appeared also in the work of several other authors, let us mention at least J.~Conant, A.~Hamilton, A.~Lazarev, J.-L.~Loday, S.~Mahajan M.~Mulase, M.~Penkava, K.~Vogtmann, A.~Schwarz and G.~Weingart. We were, however, not able to find a~suitable reference containing all details. The need for such a reference appeared in connection with our paper~\cite{markl:na} that provided a vocabulary between natural differential operators and graph complexes. Indeed, this note was originally designed as an appendix to~\cite{markl:na}, but we believe that it might be of independent interest. It supplies necessary details to~\cite{markl:na} and its future applications, and also puts the `abstract tensor calculus' attributed to R.~Penrose onto a solid footing. \noindent {\bf Acknowledgement.} We would like to express our thanks to J.-L.~Loday and J.~Stasheff for useful comments and remarks concerning the first draft of this note. \vskip 1cm \noindent {\bf Table of content:} \ref{s1}. Invariant Tensor Theorem: A recollection -- page~\pageref{s1} \hfill\break\noindent \hphantom{{\bf Table of content:\hskip .5mm}} \ref{s2}. Graphs appear: An example -- page~\pageref{s2} \hfill\break\noindent \hphantom{{\bf Table of content:\hskip .5mm}} \ref{s3}. The general case -- page~\pageref{s3} \hfill\break\noindent \hphantom{{\bf Table of content:\hskip .5mm}} \ref{s4}. Symmetries occur -- page~\pageref{s4} \hfill\break\noindent \hphantom{{\bf Table of content:\hskip .5mm}} \ref{s5}. A particular case -- page~\pageref{s5} \section{Invariant Tensor Theorem: A recollection} \label{s1} Recall that, for finite-dimensional ${\mathbf k}$-vector spaces $U$ and $W$, one has canonical isomorphisms \begin{equation} \label{conon} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(U,W)^* \cong {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(W,U),\ {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(U,V) \cong U^* \ot V \mbox { and } (U \ot W)^* \cong U^* \ot V^*, \end{equation} where ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(-,-)$ denotes the space of ${\mathbf k}$-linear maps, $(-)^*$ the linear dual and $\ot$ the tensor product over ${\mathbf k}$. The first isomorphism in~(\ref{conon}) is induced by the non-degenerate pairing \[ {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(U,W) \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(W,U) \to {\mathbf k} \] that takes $f \ot g \in {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(U,W) \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(W,U)$ into the trace of the composition $\Tr(f \circ g)$, the remaining two isomorphisms are obvious. In this note, by a {\em canonical isomorphism\/} we will usually mean a composition of isomorphisms of the above types. Einstein's convention assuming summation over repeated (multi)indices is used. We will also assume that the ground field ${\mathbf k}$ is of characteristic zero. In what follows, $V$ will be an $n$-dimensional ${\mathbf k}$-vector space and ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$ the group of linear automorphisms of $V$. We start by considering the vector space ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp Vk,\otexp Vl)$ of ${\mathbf k}$-linear maps $f : \otexp Vk \to \otexp Vl$, $k,l \geq 0$. Since both $\otexp Vk$ and $\otexp Vl$ are natural ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-modules, it makes sense to study the subspace ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}(\otexp Vk,\otexp Vl) \subset {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp Vk,\otexp Vl)$ of ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant maps. As there are no ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant maps in ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp Vk,\otexp Vl)=0$ if $k \not= l$ (see, for instance,~\cite[\S24.3]{kolar-michor-slovak}), the only interesting case is $k=l$. For a permutation $\sigma \in \Sigma_k$, define the {\em elementary invariant tensor \/} $t_\sigma \in {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp Vk,\otexp Vk)$ as the map given by \begin{equation} \label{jdu_k_doktorovi} t_\sigma(v_1 \otimes \cdots \otimes v_k) := v_{\sigma^{-1}(1)}\otimes \cdots \otimes v_{\sigma^{-1}(k)},\ \mbox { for } \Rada v1k \in V. \end{equation} It is simple to verify that $t_\sigma$ is ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant. The following theorem is a celebrated result of H.~Weyl~\cite{weyl}. \begin{itt} \label{itt} The space ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}(\otexp Vk,\otexp Vk)$ is spanned by elementary invariant tensors $t_\sigma$, $\sigma \in \Sigma_k$. If $\dim(V) \geq k$, the tensors $\{t_\sigma\}_{\sigma \in \Sigma_k}$ are linearly independent. \end{itt} This form of the Invariant Tensor Theorem is a straightforward translation of~\cite[Theorem~2.1.4]{fuks} describing invariant tensors in $\otexp{{V^*}}k \ot \otexp Vk$ and remarks following this theorem, see also~\cite[Theorem~24.4]{kolar-michor-slovak}. The Invariant Tensor Theorem can be reformulated into saying that the map \begin{equation} \label{preziji_to?} {\mathcal R}_n : {\mathbf k}[\Sigma_k] \to {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}(\otexp Vk,\otexp Vk) \end{equation} from the group ring of $\Sigma_k$ to the subspace of ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant maps given by ${\mathcal R}_n(\sigma) := t_\sigma$, $\sigma \in \Sigma_k$, is always an epimorphism and is an isomorphism for $n \geq k$ (recall $n$ denoted the dimension of $V$). The tensors $\{t_\sigma\}_{\sigma \in \Sigma_k}$ are not linearly independent if $\dim(V) <k$. For a subset $S \subset\{\rada 1k\}$ such that $\card(S) > \dim(V)$, denote by $\Sigma_S$ the subgroup of $\Sigma_k$ consisting of permutations that leave the complement $\{\rada 1k\}\setminus S$ fixed. It is simple to verify that then \begin{equation} \label{Pozitri_zpet_do_Prahy} \sum_{\sigma \in \Sigma_S}{\rm sgn\/}(\sigma) \cdot t_\sigma = 0 \end{equation} in ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}(\otexp Vk,\otexp Vk)$. By ~\cite[II.1.3]{fuks}, all relations between the elementary invariant tensors are induced by the relations of the above type. In other words, the kernel of the map ${\mathcal R}_n$ in~(\ref{preziji_to?}) is generated by the expressions \[ \sum_{\sigma \in \Sigma_S}{\rm sgn\/}(\sigma) \cdot \sigma \in {\mathbf k}[\Sigma_k], \] where $S$ and $\Sigma_S$ are as above. Observe that, with the convention used in~(\ref{jdu_k_doktorovi}) involving the inverses of $\sigma$ in the right hand side, ${\mathcal R}_n$ is a ring homomorphism. \begin{definition} \label{stab} By the {\em stable range\/} we mean the situation when $\dim(V) \geq k$, that is, when the map ${\mathcal R}_n$ in~(\ref{preziji_to?}) is a monomorphism. \end{definition} \section{Graphs appear: An example} \label{s2} In this section we analyze an example that illustrates how the Invariant Tensor Theorem leads to graphs. We are going to describe invariant tensors in ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}\otexp V2 \sqot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V\right)$. The canonical identifications~(\ref{conon}) determine a ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant isomorphism \[ \Phi : {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}\otexp V2 \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V\right) \cong {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V3,\otexp V3). \] Applying the Invariant Tensor Theorem to ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V3,\otexp V3)$, one concludes that the subspace ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}(\otexp V2 \sqot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V)$ is spanned by $\Phi^{-1}(t_\sigma)$, $\sigma \in \Sigma_3$, and that these generators are linearly independent if $\dim(V) \geq 3$. It is a simple exercise to calculate the tensors $\Phi^{-1}(t_\sigma)$ explicitly. The results are shown in the second column of the table in Figure~\ref{table} in which $X \ot Y \ot F$ is an element of $\otexp V2 \sqot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V)$ and $\Tr(-)$ the trace of a linear map $V \to V$. \begin{figure} \unitlength.9cm \begin{picture}(15,14)(-1,-12) \put(0,-.5){ \put(1,1){\makebox(0,0)[l]{$\Phi^{-1}(t_\sigma)$:}} \put(8,1.45){\makebox(0,0)[l]{coordinate}} \put(8,.9){\makebox(0,0)[l]{form:}} \put(11,1){\makebox(0,0)[l]{graph:}} } \thinlines \put(-2,0){\line(1,0){16}} \put(-2,0.1){\line(1,0){16}} \put(0.6,1){\line(0,-1){12.7}} \put(0.7,1){\line(0,-1){12.7}} \put(7.75,1){\line(0,-1){12.7}} \put(10.6,1){\line(0,-1){12.7}} \put(-2,-1){\makebox(0,0)[l]{$\sigma = {\it identity}$}} \put(1,-1){\makebox(0,0)[l]{$X\ot Y \ot F \mapsto F(X,Y)$}} \put(8,-1){\makebox(0,0)[l]{$X^jY^kF^i_{jk}e_i$}} \put(11.8,-1.5){ \unitlength.5cm \put(0,2){\makebox(0,0)[cc]{\hskip .5mm${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}} \put(0,1.08){\vector(0,1){.85}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0.3,1.2){\makebox(0,0)[l]{\scriptsize$F$}} \put(-1,-1){ \put(0,1){{\vector(1,1){.92}}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0,0.4){\makebox(0,0)[t]{\scriptsize$X$}} } \put(1,-1){ \put(0,1){{\vector(-1,1){.92}}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0,.4){\makebox(0,0)[t]{\scriptsize$Y$}} }} \put(-2,-3){\makebox(0,0)[l]{$\sigma =$}} \put(-1,-3){\sigmadva} \put(1,-3){\makebox(0,0)[l]{$X\ot Y \ot F \mapsto F(Y,X)$}} \put(8,-3){\makebox(0,0)[l]{$X^jY^kF^i_{kj}e_i$}} \put(11.8,-3.5){ \unitlength.5cm \put(0,2){\makebox(0,0)[cc]{\hskip .5mm${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}} \put(0,1.08){\vector(0,1){.85}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0.3,1.2){\makebox(0,0)[l]{\scriptsize$F$}} \put(-1,-1){ \put(0,1){{\vector(1,1){.92}}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0,0.4){\makebox(0,0)[t]{\scriptsize$Y$}} } \put(1,-1){ \put(0,1){{\vector(-1,1){.92}}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0,.4){\makebox(0,0)[t]{\scriptsize$X$}} }} \put(-2,-5){\makebox(0,0)[l]{$\sigma =$}} \put(-1,-5){\sigmatri} \put(1,-5){\makebox(0,0)[l]{$X\ot Y \ot F \mapsto Y \otimes \Tr(F(X,-))$}} \put(8,-5){\makebox(0,0)[l]{$X^jY^iF^k_{jk}e_i$}} \put(11.1,-5.2){\borelioza YX} \put(-2,-7){\makebox(0,0)[l]{$\sigma =$}} \put(-1,-7){\sigmapet} \put(1,-7){\makebox(0,0)[l]{$X\ot Y \ot F \mapsto Y \otimes \Tr(F(-,X))$}} \put(8,-7){\makebox(0,0)[l]{$X^jY^iF^k_{kj}e_i$}} \put(11.1,-7.2){\boreliozaInv YX} \put(-2,-9){\makebox(0,0)[l]{$\sigma =$}} \put(-1,-9){\sigmactyri} \put(1,-9){\makebox(0,0)[l]{$X\ot Y \ot F \mapsto X \otimes \Tr(F(-,Y))$}} \put(8,-9){\makebox(0,0)[l]{$X^iY^jF^k_{kj}e_i$}} \put(11.1,-9.2){\boreliozaInv XY} \put(-2,-11){\makebox(0,0)[l]{$\sigma =$}} \put(-1,-11){\sigmasest} \put(1,-11){\makebox(0,0)[l]{$X\ot Y \ot F \mapsto X \otimes \Tr(F(Y,-))$}} \put(8,-11){\makebox(0,0)[l]{$X^iY^jF^k_{jk}e_i$}} \put(11.1,-11.2){\borelioza XY} \put(14.2,-1){\brace} \put(14.2,-5){\brace} \put(14.2,-9){\brace} \end{picture} \caption{\label{table} Invariant tensors in ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2 \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V)$. The meaning of vertical braces on the right is explained in Example~\protect\ref{456}.} \end{figure} Let us fix a basis $\{\Rada e1n\}$ of $V$ and write $X = X^ae_a$, $Y = Y^ae_a$ and $F(e_a,e_b) = F^c_{ab} e_c$, for some scalars $X^a, Y^a, F^c_{ab} \in {\mathbf k}$, $1\leq a,b,c \leq n$. The corresponding coordinate forms of the elementary tensors are shown in the third column of the table. Observe that the expressions in this column are all possible {\em contractions of indices\/} of the tensors $X$, $Y$ and~$F$. The contraction schemes for indices are encoded by the rightmost column as follows. Given a graph $G$ from this column, decorate its edges by symbols $i,j,k$. For example, for the graph in the bottom right corner of the table, choose the decoration \[ \unitlength.9cm \begin{picture}(5,2)(-1.5,.7) \put(0,2){\put(0.03,0){\makebox(0,0)[cc]{${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}}} \put(0,1){\vector(0,1){.935}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0,.7){\makebox(0,0)[t]{\scriptsize$X$}} \put(.4,.3){ \put(2.09,.85){\makebox(0,0)[cc]{\oval(1.5,1.5)[b]}} \put(2.09,1.15){\makebox(0,0)[cc]{\oval(1.5,1.5)[t]}} \put(2.85,1.22){\line(0,1){.3}} \put(.7,.7){\vector(1,1){.55}} \put(.7,.7){\makebox(0,0)[cc]{\Large$\bullet$}} \put(1.35,1.35){\makebox(0,0)[cc]{\Large$\bullet$}} \put(1.32,1.25){\makebox(0,0)[tc]{\vector(0,1){0}}} \put(1.2,1.45){\makebox(0,0)[r]{\scriptsize $F$}} \put(.7,0.4){\makebox(0,0)[t]{\scriptsize $Y$}} \put(-.5,1.2){\makebox(0,0)[r]{\scriptsize$i$}} \put(1,1){\makebox(0,0)[lt]{\scriptsize$j$}} \put(3,1.4){\makebox(0,0)[l]{\scriptsize$k$}} \put(3.4,1){\makebox(0,0){.}} } \end{picture} \] To each vertex of this edge-decorated graph we assign the coordinates of the corresponding tensors with the names of indices determined by decorations of edges adjacent to this vertex. For example, to the $F$-vertex we assign $F^k_{jk}$, because its left ingoing edge is decorated by $j$ and its right ingoing edge which happens to be the same as its outgoing edge, is decorated by $k$. The vertex \anchor, called {\em the anchor\/}, plays a special role. We assign to it the basis of $V$ indexed by the decoration of its ingoing edge. We get \[ \unitlength.9cm \begin{picture}(5,2)(-2,.7) \put(0,2){\put(0.03,0){\makebox(0,0)[cc]{${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}}} \put(0,2.3){\put(0.03,0){\makebox(0,0)[b]{$e_i$}}} \put(0,1){\vector(0,1){.935}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0,.8){\makebox(0,0)[t]{\scriptsize$X^i$}} \put(.4,.3){ \put(2.09,.85){\makebox(0,0)[cc]{\oval(1.5,1.5)[b]}} \put(2.09,1.15){\makebox(0,0)[cc]{\oval(1.5,1.5)[t]}} \put(2.85,1.22){\line(0,1){.3}} \put(.7,.7){\vector(1,1){.55}} \put(.7,.7){\makebox(0,0)[cc]{\Large$\bullet$}} \put(1.35,1.35){\makebox(0,0)[cc]{\Large$\bullet$}} \put(1.32,1.25){\makebox(0,0)[tc]{\vector(0,1){0}}} \put(1.1,1.45){\makebox(0,0)[r]{\scriptsize $F^k_{jk}$}} \put(.7,0.5){\makebox(0,0)[t]{\scriptsize $Y^j$}} \put(-.5,1.2){\makebox(0,0)[r]{\scriptsize$i$}} \put(1,1){\makebox(0,0)[lt]{\scriptsize$j$}} \put(3,1.4){\makebox(0,0)[l]{\scriptsize$k$}} } \end{picture} \] As the final step we take the product of the factors assigned to vertices and perform the summation over repeated indices. The result is \[ \sum_{1 \leq i,j,k \leq n}X^iY^jF^k_{jk}e_i. \] In this formula we made an exception from Einstein's convention and wrote the summation explicitly to emphasize the idea of the construction. A formal general definition of this process of interpreting graphs as contraction schemes is given below. Let $\wGr_{\rm ex}$ be the vector space spanned by the six graphs in the last column of the table; the hat indicates that the graphs are not oriented. The subscript ``ex'' is an abbreviation of ``example,'' and distinguishes this space from other spaces with similar names used throughout the note. The procedure described above gives an epimorphism \begin{equation} \label{boli_mne_v_krku} \widehat{\mathrm R}}\def\uR{{\underline{R}}_n : \wGr_{\rm ex} \to {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\otexp V2 \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V\right) \end{equation} which is an isomorphism if $n \geq 3$. The map ${\widehat R}} \def\wGr{{\widehat{{\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}}}_n$ defined in this way obviously does not depend on the choice of the basis $\{\Rada e1n\}$ of $V$. The space $\wGr_{\rm ex}$ can also be defined as the span of all directed graphs with three unary vertices \begin{equation} \label{aaa} \unitlength .5cm \begin{picture}(0,1)(0,.2) \put(-.5,0){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0,0){\makebox(0,0)[bl]{\scriptsize $X$}} \put(0.9,0.2){\makebox(0,0){,}} \put(-.5,0){\vector(0,1){1.2}} \end{picture} \hskip 3em \unitlength .5cm \begin{picture}(0,1)(0,.2) \put(-.5,0){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0,0){\makebox(0,0)[bl]{\scriptsize $Y$}} \put(-.5,0){\vector(0,1){1.2}} \hskip 1em \put(0.9,0.2){\makebox(0,0)[lb]{and}} \end{picture} \hskip 4.5em \raisebox{-1em}{\rule{0pt}{0pt}} \unitlength .4cm \begin{picture}(0,1.4)(0,-.3) \put(-.45,.55){\makebox(0,0)[cc]{${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}} \put(-.5,-.8){\vector(0,1){1.2}} \put(0.7,-.25){\makebox(0,0){,}} \end{picture} \end{equation} and one ``planar'' binary vertex \begin{equation} \label{bbb} \unitlength.5cm \begin{picture}(5,1.5)(-2,.4) \put(0,1.08){\vector(0,1){.85}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0.3,1.2){\makebox(0,0)[l]{\scriptsize$F$}} \put(-1,-1){ \put(0,1){{\vector(1,1){.92}}} } \put(1,-1){ \put(0,1){{\vector(-1,1){.92}}} } \end{picture} \end{equation} whose planarity means that its inputs are linearly ordered. In pictures, this order is determined by reading the inputs from left to right. \section{The general case} \label{s3} Let us generalize calculations in Section~\ref{s2} and describe ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-invariant elements in \begin{equation} \label{zabiraji_antibiotika?} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_1},\otexp V{p_1}) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_r},\otexp V{p_r}),{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp Vc,\otexp Vd)\right), \end{equation} where $r,\Rada p1r,\Rada {\hh}1r,c$ and $d$ are non-negative integers. The above space is canonically isomorphic to \[ \otexp{{V^*}}{p_1} \ot \otexp V{{\hh}_1} \ot \cdots \ot \otexp{{V^*}}{p_r} \ot \otexp V{{\hh}_r} \ot \otexp{{V^*}}{c} \ot \otexp V{d}, \] which is in turn isomorphic to\label{888} \begin{equation} \label{budu?} \otexp {{V^*}}{(p_1 + \cdots + p_r + c)} \ot \otexp V{({\hh}_1 + \cdots+ {\hh}_r + d)}, \end{equation} via the isomorphism that moves all $V^*$-factors to the left, without changing their relative order. By the last and first isomorphisms in~(\ref{conon}), the space in~(\ref{budu?}) is isomorphic to \[ {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{(p_1 + \cdots + p_r + c)},\otexp V{({\hh}_1 + \cdots+ {\hh}_r + d)}). \] We will denote the composite isomorphism between~(\ref{zabiraji_antibiotika?}) and the space in the above display by $\Phi$. Since all isomorphisms above are ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant, $\Phi$ is equivariant, too, thus the space~(\ref{zabiraji_antibiotika?}) may contain nontrivial ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant maps only if \begin{equation} \label{vyleci_mne_to?} p_1 + \cdots + p_r + c = {\hh}_1 + \cdots + {\hh}_r + d. \end{equation} Denote by $\wGr$ the space spanned by all directed graphs with $r+1$ planar vertices \[ \raisebox{-3.5em}{\rule{0pt}{0pt}} \unitlength 4mm \linethickness{0.4pt} \begin{picture}(20,5.1)(10.5,19.4) \put(20,20){\vector(1,1){2}} \put(20,20){\vector(-1,1){2}} \put(20,20){\vector(-1,2){1}} \put(18,18){\vector(1,1){1.9}} \put(22,18){\vector(-1,1){1.9}} \put(19,18){\vector(1,2){.935}} \put(20,20){\makebox(0,0)[cc]{\Large$\bullet$}} \put(19,20){\makebox(0,0)[r]{\scriptsize $F_1$}} \put(20.5,18){\makebox(0,0)[cc]{$\ldots$}} \put(20,17){\makebox(0,0)[cc]{% $\underbrace{\rule{16mm}{0mm}}_{\mbox{\scriptsize ${\hh}_1$ inputs}}$}} \put(0,40){ \put(20.5,-18){\makebox(0,0)[cc]{$\ldots$}} \put(20,-17){\makebox(0,0)[cc]{% $\overbrace{\rule{16mm}{0mm}}^{\mbox{\scriptsize $p_1$ outputs}}$}}} \end{picture} \hskip -2.8cm \raisebox{1.2mm}{$\cdots$} \hskip -2.2cm \begin{picture}(20,5.1)(10.5,19.4) \put(20,20){\vector(1,1){2}} \put(20,20){\vector(-1,1){2}} \put(20,20){\vector(-1,2){1}} \put(18,18){\vector(1,1){1.9}} \put(22,18){\vector(-1,1){1.9}} \put(19,18){\vector(1,2){.935}} \put(20,20){\makebox(0,0)[cc]{\Large$\bullet$}} \put(19,20){\makebox(0,0)[r]{\scriptsize $F_r$}} \put(20.5,18){\makebox(0,0)[cc]{$\ldots$}} \put(20,17){\makebox(0,0)[cc]{% $\underbrace{\rule{16mm}{0mm}}_{\mbox{\scriptsize ${\hh}_r$ inputs}}$}} \put(0,40){ \put(20.5,-18){\makebox(0,0)[cc]{$\ldots$}} \put(20,-17){\makebox(0,0)[cc]{% $\overbrace{\rule{16mm}{0mm}}^{\mbox{\scriptsize $p_r$ outputs}}$}}} \end{picture} \hskip -2.8cm \raisebox{1.2mm}{\mbox{and}} \hskip -2.2cm \begin{picture}(20,5.1)(10.5,19.4) \put(20,20){\vector(1,1){2}} \put(20,20){\vector(-1,1){2}} \put(20,20){\vector(-1,2){1}} \put(18,18){\vector(1,1){1.8}} \put(22,18){\vector(-1,1){1.8}} \put(19,18){\vector(1,2){.9}} \put(20,20){\makebox(0,0)[cc]{${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}} \put(20.5,18){\makebox(0,0)[cc]{$\ldots$}} \put(20,17){\makebox(0,0)[cc]{% $\underbrace{\rule{16mm}{0mm}}_{\mbox{\scriptsize $d$ inputs}}$}} \put(0,40){ \put(20.5,-18){\makebox(0,0)[cc]{$\ldots$}} \put(20,-17){\makebox(0,0)[cc]{% $\overbrace{\rule{16mm}{0mm}}^{\mbox{\scriptsize $c$ outputs}}$}}} \end{picture} \hskip -3cm , \hskip 2cm \] where planarity means that linear orders of the sets of input and output edges are specified. Observe that the number of edges of each graph spanning $\wGr$ equals the common value of the sums in~(\ref{vyleci_mne_to?}). For each graph $G \in \wGr$ we define a ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant map $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n(G)$ in the space~(\ref{zabiraji_antibiotika?}) as follows. As in Section~\ref{s2}, choose a basis $(\Rada e1n)$ of $V$ and let $(\rada{e^1}{e^n})$ be the corresponding dual basis of $V^*$. For $F_i \in {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_i},\otexp V{p_i})$, $1 \leq i \leq r$, write \[ F_i = {F_i \hskip .2em}^{a_1^i,\ldots,a^i_{p_i}}_{b_1^i,\ldots,b^i_{{\hh}_i}} \ e_{a_1} \ot \cdots \ot e_{a_{p_i}} \otimes e^{b_1} \ot \cdots \ot e^{b_{{\hh}_i}} \] with some scalars ${F_i \hskip .2em}^{a_1^i,\ldots,a^i_{p_i}}_{b_1^i,\ldots,b^i_{{\hh}_i}} \in {\mathbf k}$ or, more concisely, $F_i = {F_i \hskip .2em}^{A^i}_{B^i}\ e_{A^i} \otimes e^{B^i}$, where $A^i$ abbreviates the multiindex $(a_1^i,\ldots,a^i_{p_i})$, $B^i$ the multiindex $(b_1^i,\ldots,b^i_{{\hh}_i})$, $e_{A^i} := e_{a_1} \ot \cdots \ot e_{a_{p_i}}$, $e^{B^i} := e^{b_1} \ot \cdots \ot e^{b_{{\hh}_i}}$ and, as everywhere in this paper, summations over repeated (multi)indices are assumed. A {\em labelling\/} of a graph $G \in \wGr$ is a function $\ell : {\it Edg}(G) \to \{\rada 1n\}$, where ${\it Edg}(G)$ denotes the set of edges of $G$. Let ${\it Lab}(G)$ be the set of all labellings of $G$. For $\ell \in {\it Lab}(G)$ and $1 \leq i \leq r$, define $A^i(\ell)$ to be the multiindex $(a_1^i,\ldots,a^i_{p_i})$ such that $a^i_s$ equals $\ell(e)$, where $e$ is the edge that starts at the $s$-th output of the vertex $F_i$, $1 \leq s \leq p_i$. Likewise, put $I(\ell) := (\Rada i1c)$ with $i_t := \ell(e)$, where now $e$ is the edge that starts at the $t$-th output of the \raisebox{-.1em}{${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}-vertex, $1 \leq t \leq c$. Let $B^i(\ell)$ and $J(\ell)$ have similar obvious meanings, with `inputs' taken instead of `outputs.' For $F_1 \ot \cdots\ot F_r \in {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_1},\otexp V{p_1}) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_r},\otexp V{p_r})$ define finally \begin{equation} \label{beru_antibiotika} \widehat{\mathrm R}}\def\uR{{\underline{R}}_n(G)(F_1 \ot \cdots\ot F_r) := \sum_{\ell \in {\it Lab}(G)} {F_1 \hskip .2em}^{A^1(\ell)}_{B^1(\ell)} \ot \cdots\ot {F_r \hskip .2em}^{A^r(\ell)}_{B^r(\ell)}\ e_{J(\ell)} \ot e^{I(\ell)} \in {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp Vc,\otexp Vd). \end{equation} It is easy to check that $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n(G)$ is a ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-fixed element of the space~(\ref{zabiraji_antibiotika?}). The nature of the summation in~(\ref{beru_antibiotika}) is close to the {\em state sum model\/} for link invariants, see~\cite[Section~I.8]{kauffman:KnotsandPhysics}, with states being the values of labels of the edges of the graph. \begin{proposition} \label{zabere_to?} Let $r,\Rada p1r,\Rada {\hh}1r,c$ and $d$ be non-negative integers. Then the map \[ \widehat{\mathrm R}}\def\uR{{\underline{R}}_n :\wGr \to {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_1},\otexp V{p_1}) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_r},\otexp V{p_r}),{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp Vc,\otexp Vd)\right) \] defined by~(\ref{beru_antibiotika}) is an epimorphism. If $n \geq e$, where $e$ is the number of edges of graphs spanning $\wGr$ and $n = \dim(V)$, $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$ is also an isomorphism. \end{proposition} Observe that we do not need to assume~(\ref{vyleci_mne_to?}) in Proposition~\ref{zabere_to?}. If~(\ref{vyleci_mne_to?}) is not satisfied, then there are no ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-invariant elements in~(\ref{zabiraji_antibiotika?}) and also the space $\wGr$ is trivial, thus $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$ is an isomorphism of trivial spaces. \begin{proof}[Proof of Proposition~\ref{zabere_to?}] By the above observation, we may assume~(\ref{vyleci_mne_to?}). Consider the diagram \begin{equation} \label{nehoji_se_to} \raisebox{-1.9cm}{\rule{0pt}{4cm}} \unitlength 1cm \linethickness{0.4pt} \begin{picture}(15,1.2)(-.5,1.5) \put(0,0){\makebox(0,0){$\wGr$}} \put(8.6,0){\makebox(0,0){${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_1},\otexp V{p_1}) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_r},\otexp V{p_r}),{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp Vc,\otexp Vd)\right)$}} \put(0,3){\makebox(0,0){${\mathbf k}[\Sigma_k]$}} \put(8.6,3){\makebox(0,0){${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}(\otexp V{(p_1 + \cdots + p_r + c)},\otexp V{({\hh}_1 + \cdots+ {\hh}_r + d)})$}} \put(.6,0){\vector(1,0){1.8}} \put(.75,3){\vector(1,0){4}} \put(0,.5){\vector(0,1){2}} \put(8,.5){\vector(0,1){2}} \put(8.3,1.5){\makebox(0,0)[l]{$\Phi$}} \put(0.3,1.5){\makebox(0,0)[l]{$\Psi$}} \put(7.7,1.5){\makebox(0,0)[r]{$\cong$}} \put(-0.3,1.5){\makebox(0,0)[r]{$\cong$}} \put(1.5,.15){\makebox(0,0)[b]{$\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$}} \put(2.6,3.15){\makebox(0,0)[b]{${\mathcal R}_n$}} \end{picture} \end{equation} in which ${\mathcal R}_n$ is the map~(\ref{preziji_to?}), $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$ is defined in~(\ref{beru_antibiotika}) and $\Phi$ is the composition of canonical isomorphisms and reshufflings of factors described on page~\pageref{888} above. The map $\Psi$ is defined as follows. Let us denote, for the purposes of this proof only, by $\OUT(F_i)$ the linearly ordered set of outputs of the $F_i$-vertex, $1 \leq i \leq r$, and by $\OUT(\ctverecek)$ the linearly ordered set of outputs of \hskip 2pt \raisebox{-1pt}{${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}. The set $\OUT := \OUT(F_1) \cup \cdots \cup \OUT(F_r) \cup \OUT(\ctverecek)$ is linearly ordered by requiring that \[ \OUT(F_1) < \cdots < \OUT(F_r) < \OUT(\ctverecek) \] (we believe that the meaning of this shorthand is obvious). Let $\IN$ be the linearly ordered set of inputs defined in the similar way. The orders define unique isomorphisms \begin{equation} \label{mam_chripku} \OUT \cong (\rada 1k) \ \mbox { and } \IN \cong (\rada 1k) \end{equation} of ordered sets. Since graphs spanning $\wGr$ are determined by specifying how the outputs of vertices are connected to its inputs, there exists a one-to-one correspondence $G \leftrightarrow \varphi_G$ between graphs $G \in \wGr$ and isomorphisms $\varphi_G : \OUT \stackrel{\cong}{\to} \IN$. Given~(\ref{mam_chripku}), such $\varphi_G$ can be interpreted as an element of the symmetric group $\Sigma_k$. The map $\Psi$ is then defined by $\Psi(G) := \varphi_G$. It is simple to verify that the diagram~(\ref{nehoji_se_to}) commutes, so the proposition follows from the Invariant Tensor Theorem. \end{proof} \section{Symmetries occur} \label{s4} In the light of diagram~(\ref{nehoji_se_to}), Proposition~\ref{zabere_to?} may look just as a clumsy reformulation of the Invariant Tensor Theorem. Graphs become relevant when symmetries occur. \begin{example} \label{456} Let $\Sym(\otexp V2,V) \subset {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V)$ be the subspace of symmetric bilinear maps, i.e.~maps satisfying $f(v',v'') = f(v'',v')$ for $v',v'' \in V$. Let us explain how to use calculations of Section~\ref{s2} to describe ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant maps in ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}\otexp V2 \sqot \Sym(\otexp V2,V),V\right)$. The right $\Sigma_2$-action on ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V)$ given by permuting the inputs of bilinear maps is such that the space $\Sym(\otexp V2,V)$ equals the subspace ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V)^{\Sigma_2}$ of $\Sigma_2$-fixed elements. This right $\Sigma_2$-action induces a left $\Sigma_2$-action on ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}\otexp V2 \sqot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V\right)$ which commutes with the ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-action, therefore it restricts to a left $\Sigma_2$-action on the subspace ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\otexp V2 \sqot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V\right)$ of ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant maps. There is also a left $\Sigma_2$-action on the linear space $\wGr_{\rm ex}$ interchanging the inputs of the $F$-vertices of generating graphs. It is simple to check that the map~(\ref{boli_mne_v_krku}) of Section~\ref{s2} is equivariant with respect to these two $\Sigma_2$-actions, hence it induces the map \begin{equation} \label{porad_mi_neni_dobre} \Sigma_2 \backslash \widehat{\mathrm R}}\def\uR{{\underline{R}}_n :\Sigma_2 \backslash \wGr_{\rm ex} \to \Sigma_2 \backslash {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\otexp V2 \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V\right) \end{equation} of left cosets. Observe that, by a standard duality argument, \begin{equation} \label{ttt} \Sigma_2 \backslash{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\otexp V2 \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V2,V),V\right)\cong {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\otexp V2 \ot \Sym(\otexp V2,V),V\right). \end{equation} Let us denote $\wGr_{{\rm ex},\bullet} := \Sigma_2 \backslash\wGr_{\rm ex}$. The bullet $\bullet$ in the subscript signalizes the presence of vertices with fully symmetric inputs. By definition, graphs $G', G'' \in \wGr_{\rm ex}$ are identified in the quotient $\wGr_{{\rm ex},\bullet}$ if they differ only by the order of inputs of the $F$-vertex. In Figure~\ref{table}, this identification is indicated by vertical braces. We see that $\wGr_{{\rm ex},\bullet}$ is again a space {\em spanned by graphs,\/} this time with no linear order on the inputs of the $F$-vertex. So we may {\em define\/} $\wGr_{{\rm ex},\bullet}$ as the space spanned by directed graphs with vertices~(\ref{aaa}) and one binary (ordinary, non-planar) vertex~(\ref{bbb}). We conclude by interpreting~(\ref{porad_mi_neni_dobre}) as the map \begin{equation} \label{Phillips} \widehat{\mathrm R}}\def\uR{{\underline{R}}_n : \wGr_{{\rm ex},\bullet} \to {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\otexp V2 \ot \Sym(\otexp V2,V),V\right). \end{equation} It follows from the properties of the map~(\ref{boli_mne_v_krku}) and the characteristic zero assumption that $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$ is always an epimorphism and is an isomorphism if $n \geq 3$. \end{example} At this point we want to incorporate, by generalizing the pattern used in Example~\ref{456}, symmetries into Proposition~\ref{zabere_to?}. Unfortunately, it turns out that treating the space~(\ref{zabiraji_antibiotika?}) in full generality leads to a notational disaster. To keep the length of formulas within a reasonable limit, we decided to {\em assume from now on\/} that $p_1= \cdots = p_r = 1$, $c=0$ and $d=1$. This means that we will restrict our attention to maps in \begin{equation} \label{red} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_1},V) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_r},V),V\right). \end{equation} For graphs this assumption implies that the vertices $\Rada F1r$ have precisely one output, and that the anchor $\hskip .2em\raisebox{-.1em}{{\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}}\hskip -.2em$ has one input and no outputs. The number of inputs of $F_i$ will be called the {\em arity\/} of $F_i$, $1 \leq i \leq r$. Condition~(\ref{vyleci_mne_to?}) reduces to \[ r = {\hh}_1 + \cdots + {\hh}_r + 1 \] and one also sees that $r$ equals the number of edges of the generating graphs. The above generality is sufficient for all applications we have in mind. A modification to the general case is straightforward but notationally challenging. The space ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{\hh},V)$ admits, for each ${\hh} \geq 0$, a natural right $\Sigma_{\hh}$-action given by permuting inputs of multilinear maps. A {\em symmetry\/} of maps in ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{\hh},V)$ will be specified by a subset ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \subset {\mathbf k}[\Sigma_{\hh}]$. We then denote \[ {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}(\otexp V{\hh},V) := \left\{\adjust {.4} f \in {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{\hh},V) ;\ f {\mathfrak s} = 0 \mbox { for each } {\mathfrak s} \in {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}\right\}. \] For ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}$ as above and a left $\Sigma_{\hh}$-module $U$, we will abbreviate by ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash U$ the left coset ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} U \backslash U$. \begin{example} \label{exxx} Let ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} := I_{\hh} \subset {\mathbf k}[\Sigma_\hh]$ be the augmentation ideal. Then ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{I_{\hh}}(\otexp V{\hh},V)$ is the space of symmetric maps, \[ {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{I_{\hh}}(\otexp V{\hh},V) = \Sym(\otexp V{\hh},V), \] therefore the augmentation ideal describes the symmetry of the local coordinates of vector fields and their derivatives, see~\cite[Example~3.2]{markl:na}. We leave as an exercise to describe in this language the spaces of {\em anti\/}symmetric maps. \end{example} \begin{example} \label{exxy} Let ${\hh} := v+2$, $v \geq 0$, and let $\nabla \subset {\mathbf k}[\Sigma_{\hh}]$ be the image of the augmentation ideal $I_v$ of ${\mathbf k}[\Sigma_v]$ in ${\mathbf k}[\Sigma_{\hh}]$ under the map of group rings induced by the inclusion $\Sigma_v \hookrightarrow \Sigma_v \times \Sigma_2 \hookrightarrow\Sigma_{\hh}$ that interprets permutations of $(\rada 1v)$ as permutations of $(\rada 1v,v+1,v+2)$ keeping the last two elements fixed. Then ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_\nabla(\otexp V{\hh},V)$ consists of multilinear maps $\otexp V{(v+2)} \to V$ that are symmetric in the first $v$ inputs, i.e.~multilinear maps possessing the symmetry of the Christoffel symbols of linear connections and their derivatives, see again~\cite[Example~3.2]{markl:na}. \end{example} \begin{remark} \label{patek_v_IHES} It is clear how to generalize the above notion of symmetry to maps in the left $\Sigma_p$- right $\Sigma_{\hh}$-module ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{\hh},\otexp Vp)$ for general $p,{\hh} \geq 0$. A symmetry of these maps will be specified by subsets ${\mathfrak{{I}}}} \def\So{{\mathfrak{{O}}} \in {\mathbf k}[\Sigma_{{\hh}}]$ and $\So \in {\mathbf k}[\Sigma_{p}]$, the corresponding subspaces will then be \[ {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\So{{\mathfrak{{O}}}}^{\So}(\otexp {V}{{\hh}},\otexp {V}{p}) := \left\{\adjust {.4} f \in {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{\hh},\otexp Vp) ;\ f {\mathfrak s} = 0 = {\mathfrak t} f \mbox { for each } {\mathfrak s} \in {\mathfrak{{I}}}} \def\So{{\mathfrak{{O}}} \mbox { and } {\mathfrak t} \in \So \right\}. \] \end{remark} Suppose we are given subsets ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i \subset {\mathbf k}[\Sigma_{{\hh}_i}]$, $1 \leq i \leq r$. Our aim is to describe ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-invariant elements in the space \begin{equation} \label{reds} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_1}(\otexp V{{\hh}_1},V) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_r}(\otexp V{{\hh}_r},V),V\right). \end{equation} Let \[ {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} :={\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_1 \cup \cdots \cup {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_r \subset {\mathbf k}[\Sigma_{{\hh}_1} \times \cdots \times \Sigma_{{\hh}_r}], \] where ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i$ is, for $1 \leq i \leq r$, identified with its image in ${\mathbf k}[\Sigma_{{\hh}_1} \times \cdots \times \Sigma_{{\hh}_r}]$ under the map induced by the group inclusion $\Sigma_{{\hh}_i} \hookrightarrow \Sigma_{{\hh}_1} \times \cdots \times \Sigma_{{\hh}_r}$. As in Example~\ref{456}, we use the fact that, for $1 \leq i \leq r$, each ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_i},V)$ is a right $\Sigma_{{\hh}_i}$-space, hence the tensor product ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_1},V) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_r},V)$ has a natural right $\Sigma_{{\hh}_1} \times \cdots \times \Sigma_{{\hh}_r}$-action which induces a left $\Sigma_{{\hh}_1} \times \cdots \times \Sigma_{{\hh}_r}$-action on the space~(\ref{red}). This action restricts to the subspace of ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant maps. There is also a left $\Sigma_{{\hh}_1} \times \cdots \times \Sigma_{{\hh}_r}$-action on the space $\wGr$ given by permuting, in the obvious manner, the inputs of the vertices $\Rada F1r$ of generating graphs. The map $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$ of Proposition~\ref{zabere_to?} is equivariant with respect to the above two actions and induces the map \[ {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash \widehat{\mathrm R}}\def\uR{{\underline{R}}_n : {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash \wGr \to {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_1},V) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}(\otexp V{{\hh}_r},V),V\right) \] of left quotients. Denoting $\wGr_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} := {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash \wGr$ and realizing that, by duality, the codomain of ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash \widehat{\mathrm R}}\def\uR{{\underline{R}}_n$ is isomorphic to the subspace of ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-fixed elements in~(\ref{reds}), we obtain the map (denoted again $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$) \begin{equation} \label{jeste_jeden_den} \widehat{\mathrm R}}\def\uR{{\underline{R}}_n : \wGr_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \to {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}{\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_1}(\otexp V{{\hh}_1},V) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_r}(\otexp V{{\hh}_r},V),V\right) \end{equation} which is, by Proposition~\ref{zabere_to?}, an epimorphism and is an isomorphism if $\dim(V) \geq r$. \begin{remark} \label{jaja} As in Example~\ref{456}, it turns out that the quotient $\wGr_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} = {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash \wGr$ is a {\em space of graphs\/} though, for general symmetries, ``space of graphs'' means a free wheeled operad on a certain $\Sigma$-module~\cite{mms}. In the cases relevant for our paper, we however remain in the realm of `classical' graphs, as shown in the following example, see also the proof of Corollary~\ref{boli_mne_za_krkem}. \end{remark} \begin{example} \label{ja} Suppose that, for some $1 \leq i \leq r$, ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i$ equals the augmentation ideal $I_{{\hh}_i}$ of ${\mathbf k}[\Sigma_{{\hh}_i}]$ as in Example~\ref{exxx}. Then, in the quotient ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash \wGr$, one identifies graphs that differ by the order of inputs of the vertex $F_i$. In other words, modding out by ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i \subset {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}$ erases the order of inputs of $F_i$, turning $F_i$ into an ordinary (non-planar) vertex. If ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i = \nabla$ as in Example~\ref{exxy}, one gets a vertex of arity $v+2$, $v \geq 0$, whose first $v$ inputs are symmetric. \end{example} For applications, we still need one more level of generalization that will reflect the antisymmetry of the Chevalley-Eilenberg complex~\cite[Section~2]{markl:na} in the Lie algebra variables. As a motivation for our construction, we offer the following continuation of the calculations in Section~\ref{s2} and Example~\ref{456}. \begin{example} \label{zivotosprava} We will consider the tensor product $V \ot V$ as a left $\Sigma_2$-module, with the action $\tau(v' \ot v'') := - (v'' \ot v')$, for $v',v'' \in V$ and the generator $\tau \in \Sigma_2$. The subspace $(V \ot V)^{\Sigma_2}$ of $\Sigma_2$-fixed elements is then precisely the second exterior power $\mbox{\Large$\land$}}\def\bp{{\mathbf p}^2 V$. This left action induces a ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant right $\Sigma_2$-action on the space ${\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}\otexp V2 \sqot \Sym(\otexp V2,V),V\right)$ such that \[ {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}\otexp V2 \ot \Sym(\otexp V2,V),V\right)/\Sigma_2 \cong {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\adjust {.4}\mbox{\Large$\land$}}\def\bp{{\mathbf p}^2 V \ot \Sym(\otexp V2,V),V\right). \] The above isomorphism restricts to an isomorphism \begin{equation} \label{u} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\otexp V2 \ot \Sym(\otexp V2,V),V\right)/\Sigma_2 \cong {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\mbox{\Large$\land$}}\def\bp{{\mathbf p}^2 V \ot \Sym(\otexp V2,V),V\right). \end{equation} of the subspaces of ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant maps. Likewise, $\wGr_{{\rm ex},\bullet}$ carries a right $\Sigma_2$-action that interchanges the labels $X$ and $Y$ of the \black-vertices of graphs in the last column of Figure~\ref{table} and multiplies the sign of the corresponding generator by $-1$. The map~(\ref{Phillips}) is $\Sigma_2$-equivariant, therefore it induces the map \[ \widehat{\mathrm R}}\def\uR{{\underline{R}}_n/\Sigma_2 : \wGr_{{\rm ex},\bullet} /\Sigma_2 \to {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\otexp V2 \ot \Sym(\otexp V2,V),V\right)/\Sigma_2. \] Let us denote ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^2_{{\rm ex},\bullet} := \wGr_{{\rm ex},\bullet}/\Sigma_2$ and $\rR^2_n := \widehat{\mathrm R}}\def\uR{{\underline{R}}_n/\Sigma_2$. Using~(\ref{u}), one rewrites the above map as an epimorphism \[ \rR^2_n : {\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^2_{{\rm ex},\bullet} \epi {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\adjust {.4}\mbox{\Large$\land$}}\def\bp{{\mathbf p}^2 V \sqot \Sym(\otexp V2,V),V\right) \] which is an isomorphism if $n \geq 3$. The space ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^2_{{\rm ex},\bullet}$ is isomorphic to the span of the set of directed, oriented graphs with one (non-planar) binary vertex $F$, an anchor \anchor, and two `white' vertices \white. By an {\em orientation\/} we mean a linear order of white vertices. A graph with the opposite orientation is identified with the original one taken with the opposite sign. It is clear that, with ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^2_{{\rm ex},\bullet}$ defined in this way, the map ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^2_{{\rm ex},\bullet} \to \wGr_{{\rm ex},\bullet} /\Sigma_2$ that replaces the first (in the linear order given by the orientation) white vertex \white\ by the black vertex \black\ labelled by $X$, and the second white vertex by the black vertex labelled by $Y$, is an isomorphism. The symmetry of the inputs of the vertex $F$ implies the following identities in ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^2_{{\rm ex},\bullet}$: \[ \unitlength.5cm \begin{picture}(10,2)(0,.5) \put(0,2){\makebox(0,0)[cc]{\hskip .5mm${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}} \put(0,1.08){\vector(0,1){.85}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0.3,1.2){\makebox(0,0)[l]{\scriptsize$F$}} \put(-1,-1){ \put(0.15,1.15){{\vector(1,1){.76}}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} \put(1,1){\makebox(0,0)[cc]{$<$}} } \put(1,-1){ \put(-.15,1.15){{\vector(-1,1){.76}}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} } \put(2.5,1){\makebox(0,0){$=-$}} \put(5,0){ \put(0,2){\makebox(0,0)[cc]{\hskip .5mm${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}} \put(0,1.08){\vector(0,1){.85}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0.3,1.2){\makebox(0,0)[l]{\scriptsize$F$}} \put(-1,-1){ \put(0.15,1.15){{\vector(1,1){.76}}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} \put(1,1){\makebox(0,0)[cc]{$>$}} } \put(1,-1){ \put(-.15,1.15){{\vector(-1,1){.76}}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} } \put(2.5,1){\makebox(0,0){$=-$}} } \put(10,0){ \put(0,2){\makebox(0,0)[cc]{\hskip .5mm${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}} \put(0,1.08){\vector(0,1){.85}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0.3,1.2){\makebox(0,0)[l]{\scriptsize$F$}} \put(-1,-1){ \put(0.15,1.15){{\vector(1,1){.76}}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} \put(1,1){\makebox(0,0)[cc]{$<$}} } \put(1,-1){ \put(-.15,1.15){{\vector(-1,1){.76}}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} } } \put(11.5,1){\makebox(0,0){,}} \end{picture} \adjust {1.2} \] from which one concludes that \[ \hskip 5cm \unitlength.5cm \begin{picture}(10,2)(0,.5) \put(0,2){\makebox(0,0)[cc]{\hskip .5mm${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}} \put(0,1.08){\vector(0,1){.85}} \put(0,1){\makebox(0,0)[cc]{\Large$\bullet$}} \put(0.3,1.2){\makebox(0,0)[l]{\scriptsize$F$}} \put(-1,-1){ \put(0.15,1.15){{\vector(1,1){.76}}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} \put(1,1){\makebox(0,0)[cc]{$<$}} } \put(1,-1){ \put(-.15,1.15){{\vector(-1,1){.76}}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} } \put(2.5,1){\makebox(0,0){$=0$.}} \end{picture} \adjust {1.2} \] Therefore ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^2_{{\rm ex},\bullet}$ is in this case one-dimensional, spanned by the equivalence class of the oriented directed graph \[ \begin{picture}(7,5)(1,5) \unitlength.74cm \put(-1,-1.5){ \put(0,-.1){ \put(0,2){\put(0.03,0){\makebox(0,0)[cc]{${\raisebox {.1em}{\rule{.6em}{.6em}} \hskip .1em}$}}} \put(0,1.17){\vector(0,1){.75}} \put(0,1){\makebox(0,0)[cc]{\Large$\circ$}} } \put(.4,.2){ \put(2.09,.85){\makebox(0,0)[cc]{\oval(1.5,1.5)[b]}} \put(2.09,1.15){\makebox(0,0)[cc]{\oval(1.5,1.5)[t]}} \put(2.85,1.22){\line(0,1){.3}} \put(.82,.82){\vector(1,1){.48}} \put(.7,.7){\makebox(0,0)[cc]{\Large$\circ$}} \put(1.35,1.35){\makebox(0,0)[cc]{\Large$\bullet$}} \put(1.32,1.25){\makebox(0,0)[tc]{\vector(0,1){0}}} \put(1,1.45){\makebox(0,0)[r]{\scriptsize $F$}} \put(0.16,0.7){\makebox(0,0)[cc]{$<$}} \put(3.4,1){\makebox(0,0)[b]{.}} }} \end{picture} \adjust {2} \] In the notation of Figure~\ref{table}, the above graph represents the map that sends $(X\land Y) \ot F \in \mbox{\Large$\land$}}\def\bp{{\mathbf p}^2 V \ot \Sym(\otexp V2,V)$ into \[ X \otimes \Tr(F(Y,-)) - Y \otimes \Tr(F(X,-)) \in V. \] \end{example} Let us turn to our final task. We want to describe ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-invariant elements in the space \begin{equation} \label{piano-nad-hlavou} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}\left(\mathop{{\rm \EXT}}\displaylimits_{1 \leq i \leq m} \Sym(\otexp V{{\hh}_i},V) \ot \bigotimes_{m+1 \leq i \leq r} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i}(\otexp V{{\hh}_i},V),V\right) \end{equation} where, as before, $r,\Rada {\hh}1r$ are positive integers, ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i \subset {\mathbf k}[\Sigma_{{\hh}_i}]$ for $m+ 1 \leq i \leq r$, and $m$ is an integer such that $1 \leq m \leq r$. Having in mind the description of the space of symmetric multilinear maps given in Example~\ref{exxx}, we extend the definition of ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i$ also to $1 \leq i \leq m$, by putting ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i: = I_{\hh_i}$. The first step is to identify the exterior power $\mathop{{\LAND}}\displaylimits_{1 \leq i \leq m} \Sym(\otexp V{{\hh}_i},V)$ with the fixed point set of an action of a suitable finite group. This can be done as follows. For $1 \leq w \leq m$, let $A(w) \subset \{\rada 1m\}$ be the subset $A(w) := \{1 \leq i \leq m;\ {\hh}_i = {\hh}_w\}$. Then \[ \{\rada 1m\} = \textstyle\bigcup_{1 \leq w \leq m} A(w) \] is a decomposition of $\{\rada 1m\}$ into not necessarily distinct subsets. Let $\fA \subset \Sigma_m$ be the subgroup of permutations of $\{\rada 1m\}$ preserving this decomposition. The group $\fA$ acts on $\bigotimes _{1 \leq i \leq m} \Sym(\otexp V{{\hh}_i},V)$ by permuting the corresponding factors. If we consider this tensor product as a left $\fA$-module with this permutation action twisted by the signum representation, then \[ \mathop{{\rm \EXT}}\displaylimits_{1 \leq i \leq m} \Sym(\otexp V{{\hh}_i},V) \cong \left(\bigotimes_{1 \leq i \leq m} \Sym(\otexp V{{\hh}_i},V) \right)^\fA. \] The above left $\fA$-action on $\bigotimes _{1 \leq i \leq m} \Sym(\otexp V{{\hh}_i},V)$ induces a dual ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant right $\fA$-action on the space~(\ref{piano-nad-hlavou}). There is a right $\fA$-action on the quotient $\wGr_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} = {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash \wGr$ defined as follows. For a graph $G \in \wGr$ representing an element $[G] \in \wGr_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}$ and for $\sigma \in \fA$, let $G^\sigma$ be the graph obtained from $G$ by permuting the vertices $\Rada F1m$ according to $\sigma$. We then put $[G]\sigma := {\rm sgn\/}(\sigma) [G^\sigma]$. Since, by the definition of $\fA$, $\sigma$ may interchange only vertices with the same number of inputs and the same symmetry, our definition of $G^\sigma$ makes sense. It is simple to see that the map $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$ in~(\ref{jeste_jeden_den}) is $\fA$-equivariant, giving rise to the map \[ \widehat{\mathrm R}}\def\uR{{\underline{R}}_n / \fA : \wGr_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}/\fA \to {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}({\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_1}(\otexp V{{\hh}_1},V) \ot \cdots \ot {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_r}(\otexp V{{\hh}_r},V),V)/\fA \] of right cosets. The codomain of $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n/ \fA$ is easily seen to be isomorphic to the subspace of ${{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}$-equivariant elements in~(\ref{piano-nad-hlavou}). The above calculations are summarized in the following proposition in which ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} := \wGr_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}/\fA$ and $\rR^m_n:= \widehat{\mathrm R}}\def\uR{{\underline{R}}_n / \fA$. \begin{proposition} \label{zabere_to??} Let $r,\Rada \hh 1r$ be non-negative integers, $1 \leq m \leq r$, and ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i \subset {\mathbf k}[\Sigma_{\hh_i}]$ for $m+1 \leq i \leq r$. Then the map \begin{equation} \label{zitra_na_kole} \rR^m_n :{\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \to {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\mathop{{\rm \EXT}}\displaylimits_{1 \leq i \leq m} \Sym(\otexp V{\hh_i},V) \ot \bigotimes_{m+1 \leq i \leq r} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i}(\otexp V{\hh_i},V),V\right) \end{equation} constructed above is an epimorphism. If, moreover, the dimension $n$ of $V$ $\geq$ the number of edges of graphs spanning ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}$, $\rR^m_n$ is also an isomorphism. \end{proposition} The following result says that the presence of vertices with symmetric inputs miraculously {\em extends\/} the stability range (Definition~\ref{stab}). In applications, these vertices will represent the Lie algebra generators in the Chevalley-Eilenberg complex. \begin{proposition} \label{zitra_odletam_z_IHES_do_Prahy} Suppose that $\Rada \hh 1m \geq 2$. If $n \geq e-m$, where $n$ is the dimension of $V$ and $e$ the number of edges of graphs spanning ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}$, then the map $\rR^m_n$ in Proposition~\ref{zabere_to??} is an isomorphism. \end{proposition} \begin{proof} Let $G$ be a graph spanning ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}$ and $S \subset {\it Edg}(G)$ a subset of edges of $G$ such that $\card(S) > n$. For each permutation $\sigma$ of elements of $S$, denote by $G_\sigma$ the graph obtained by cutting the edges belonging to $S$ in the middle and regluing them following the automorphism $\sigma$. The linear combination \begin{equation} \label{eeee} \sum_{\sigma \in \Sigma_S}{\rm sgn\/}(\sigma) \cdot G_\sigma \in {\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \end{equation} is then a graph-ical representation of the expression in~(\ref{Pozitri_zpet_do_Prahy}), thus the kernel of $\rR^m_n$ is generated by expressions of this type. Since, by assumption, $\card(S) \leq n+m$ and $\Rada \hh 1m \geq 2$, the set $S$ must necessarily contain two input edges of the {\em same\/} symmetric vertex of $G$. This implies that the sum~(\ref{eeee}) vanishes, because with each graph $G_\sigma$ it contains the same graph with the opposite sign. This shows that the kernel of $\rR^m_n$ is trivial. \end{proof} \begin{remark} \label{po_navratu_z_polska} By an absolutely straightforward generalization of the above constructions, one can obtain versions of Proposition~\ref{zabere_to??} and Proposition~\ref{zitra_odletam_z_IHES_do_Prahy} describing the space \begin{equation} \label{Eli_chce_prijet_v_prosinci.} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)}\left(\mathop{{\rm \EXT}}\displaylimits_{1 \leq i \leq m} \Sym(\otexp V{\hh_i},V) \ot \bigotimes_{m+1 \leq i \leq r} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i}^{\So_i}(\otexp V{\hh_i},\otexp V{p_i}), {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}}^{\So}(\otexp V{c},\otexp V{d})\right) \end{equation} in terms of a space spanned by graphs. Since the notational aspects of such a generalization are horrendous, we must leave the details as an exercise to the reader. \end{remark} \section{A particular case} \label{s5} We finish this note by a corollary tailored for the needs of~\cite{markl:na}. For non-negative integers $m,b$ and $c$, denote by ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b)\nabla(c)}$ the space spanned by directed, oriented graphs with \begin{itemize} \item[(i)] $m$ unlabeled `white' vertices with fully symmetric inputs and arities $\geq 2$, \item[(ii)] $b$ `black' labelled vertices with fully symmetric inputs and arities $\geq 0$, \item[(iii)] $c$ labelled $\nabla$-vertices, and \item[(iv)] the anchor \anchor. \end{itemize} In item~(iii), a $\nabla$-vertex means a vertex with the symmetry described in Example~\ref{exxy}, see also Example~\ref{ja}. As in Example~\ref{zivotosprava}, an {\em orientation\/} is given by a linear order on the set of white vertices. If $G'$ and $G''$ are graphs in ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b)\nabla(c)}$ whose orientations differ by an odd number of transpositions, then we identify $G' = -G''$ in ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b)\nabla(c)}$. \begin{corollary} \label{boli_mne_za_krkem} For each non-negative integers $m,b$ and $c$ there exists a natural epimorphism \begin{eqnarray*} \lefteqn{ \rR^m_{\bullet(b)\nabla(c),n} :{\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b)\nabla(c)} \epi} \\ && \bigoplus_{\vec h \in \frH} {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_{{\rm GL\/}}\def\GL#1#2{{\GLname\/}^{(#1)}_{#2}(V)} \hskip -.2em \left(\mathop{{\rm \EXT}}\displaylimits_{1 \leq i \leq m} \hskip -.2em \Sym(\otexp V{{\hh}_i},V) \ot \bigotimes_{m+1 \leq i \leq m+b} \hskip -1.2em \Sym(\otexp V{{\hh}_i},V) \bigotimes_{m+b+1 \leq i \leq m+b+c} \hskip -1.2em {\mbox {\it Lin\/}}}\def\Sym{{\mbox {\it Sym\/}}_\Delta(\otexp V{{\hh}_i},V),V\right), \end{eqnarray*} with the direct sum taken over the set $\frH$ of all multiindices $\vec h = (\Rada h1{m+b+c})$ such that \[ \Rada h1m \geq 2,\ \Rada h{m+1}{m+b} \geq 0\ \mbox { and }\ \Rada h{m+b+1}{m+b+c} \geq 2. \] The map $\rR^m_{\bullet(b)\nabla(c),n}$ is an isomorphism if $n = \dim(V) \geq b+c$. \end{corollary} \begin{proof} The map $\rR^m_{\bullet(b)\nabla(c),n}$ is constructed by assembling the maps $\rR^m_n$ from Proposition~\ref{zabere_to??} as follows. For a multiindex $\vec h = (\Rada h1{m+b+c}) \in \frH$ as in the corollary take, in Proposition~\ref{zabere_to??}, $r:= m+b+c$ and \[ {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i = {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}_i(\vec h) := \cases{\adjust {.4} I_{{\hh}_i}}{for $m+1 \leq i \leq m+b$ and}% {\rule{0pt}{1.2em}\nabla}{for $m+b+1 \leq i \leq r$,} \] see Examples~\ref{exxx} and~\ref{exxy} for the notation. Let $\rR^m_n(\vec h)$ be the map~(\ref{zitra_na_kole}) corresponding to the above choices and $\rR^m_{\bullet(b)\nabla(c),n} := \bigoplus_{\vec h \in \frH}\rR^m_n(\vec h)$. We only need to show that the graph space ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b),\nabla(c)}$ is isomorphic to the direct sum of the double quotients ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}(\vec h)} = {\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}(\vec h) \backslash \wGr /\fA$. As we argued in Example~\ref{ja}, the left quotient $\wGr_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}(\vec h)} = {{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}(\vec h)} \backslash \wGr$ is spanned by directed graphs with $r$ labelled vertices $\Rada F1r$ such that the 1st type vertices $\Rada F1m$ (`white' vertices) have fully symmetric inputs and arities ${\hh}_1,\ldots, {\hh}_m$, and the remaining vertices $\Rada F{m+1}r$ are as in items (ii)--(iv) of the definition of ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b)\nabla(c)}$ but with fixed arities $\rada {h_{m+1}}{h_r}$. Modding out $\wGr_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}(\vec h)}$ by $\fA$ identifies graphs that differ by a relabelling of white vertices of the same arity and the sign given by to the signum of this relabelling. This clearly means that the map \[ {\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b),\nabla(c)} \to \bigoplus_{\vec h \in \frH} {\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}(\vec h)} = \bigoplus_{\vec h \in \frH} \wGr_{{\mathfrak{{I}}}} \def\udelta{{\underline{\delta}}(\vec h)} / \fA \] that assigns to the first (in the linear order given by the orientation) white vertex of graphs generating ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b),\nabla(c)}$ label $F_1$, to the second white vertex label $F_2$, etc., is an isomorphism. By simple combinatorics, graphs spanning ${\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b),\nabla(c)}$ have precisely $m+b+c$ edges which completes the proof of the corollary. \end{proof} \begin{remark} \label{ZASE_mne_boli_v_krku} Proposition~\ref{zabere_to??} and its Corollary~\ref{boli_mne_za_krkem} was obtained by applying the double-coset reduction ${\mathfrak{{I}}}} \def\udelta{{\underline{\delta}} \backslash \ {-} /\fA$ and standard duality to the map $\widehat{\mathrm R}}\def\uR{{\underline{R}}_n$ of Proposition~\ref{zabere_to?}. Backtracking all the constructions involved, one can see that, in Corollary~\ref{boli_mne_za_krkem}, the invariant linear map $\rR^m_{\bullet(b)\nabla(c),n}(G)$ corresponding to a graph $G \in {\EuScript {G}\rm r}}\def\plGr{{\rm pl\widehat{\EuScript {G}\rm r}}^m_{\bullet(b)\nabla(c)}$ is given by the `state sum'~(\ref{beru_antibiotika}) {\em antisymmetrized\/} in the white vertices. \end{remark} \def$'${$'$}

https://arxiv.org/abs/1004.2517

Upper and lower bounds for normal derivatives of spectral clusters of Dirichlet Laplacian

In this paper, we prove the upper and lower bounds for normal derivatives of spectral clusters $u=\chi_{\lambda}^s f$ of Dirichlet Laplacian $\Delta_M$, $$c_s \lambda\|u\|_{L^2(M)} \leq \| \partial_{\nu}u \|_{L^2(\partial M)} \leq C_s \lambda \|u\|_{L^2(M)} $$ where the upper bound is true for any Riemannian manifold, and the lower bound is true for some small $0<s<s_M$, where $s_M$ depends on the manifold only, provided that $M$ has no trapped geodesics (see Theorem \ref{Thm3} for a precise statement), which generalizes the early results for single eigenfunctions by Hassell and Tao.

\section{\bf Introduction} Let $M$ be a smooth compact Riemannian manifold with boundary $\partial M = Y$. It is well known that minus the Dirichlet Laplacian $-\Delta_M$ on $M$ has discrete spectrum $0 < \lambda_1^2 < \lambda_2^2 \leq \lambda_3^2 \dots \to \infty$. Let $e_j$ be an $L^2$-normalized eigenfunction corresponding to $\lambda_j^2$, and let $\psi_j$ be the normal derivative of $e_j$ at the boundary. In \cite{Ozawa} Ozawa posed the following question: {\it Do there exist constants $0< c < C < \infty$, depending on $M$ but not on $j$, such that \begin{equation} c \lambda_j \leq \| \psi_j \|_{L^2(Y)} \leq C \lambda_j ? \label{bounds} \end{equation}} Using heat kernel techniques, Ozawa \cite{Ozawa} showed that an averaged version of (\ref{bounds}) holds. More precisely, he showed that $$ \sum_{\lambda_j < \lambda} \psi_j^2(y) = \frac{\lambda^{n+2} }{ (4\pi)^{n/2} \Gamma((n/2)+2)}+ o(\lambda^{n+2}), \quad \forall y \in Y. $$ This asymptotic formula (after integrating over $Y$) would be implied by (\ref{bounds}) in view of Weyl asymptotics for the $\lambda_j$. In \cite{HT} Hassell and Tao proved an upper bound of the form $\|\psi \|_2 \leq C\lambda$ for general manifolds, and a lower bound $c \lambda \leq \|\psi \|_2$ provided that $M$ has no trapped geodesics ( see Theorem \ref{Thm3} for a precise statement). Define the {\bf Spectral Cluster} $\chi_{\lambda}^s f$ with spectral band width $s$, \begin{eqnarray*} \chi_{\lambda}^s f &=&\sum_{\lambda_j\in[\lambda,\lambda+s)}e_j(f) =\int_M[\sum_{\lambda_j\in[\lambda,\lambda+s)}e_j(x)e_j(y)]f(y)dy \label{SpecC} \\ e_j(f)&=& e_j(x)\int_M e_j(y)f(y)dy. \end{eqnarray*} The $L^2\to L^p$ estimates and gradient estimates on spectral clusters have been widely studied (see \cite{G}, \cite{SS1}-\cite{Xu2}). In general, the estimates for single eigenfunctions might be still true for spectral clusters in some sense. It is a natural question: {\it whether the upper and lower bound for normal derivatives of Dirichlet eigenfunctions in \cite{HT} is still true for normal derivatives of spectral clusters $\chi_{\lambda}^s f$.} The key obstacle to answer this question directly from the estimates of Hassell and Tao \cite{HT} for single eigenfunctions is that $\partial_{\nu}e_i$ and $\partial_{\nu}e_j$ are NOT orthogonal in $L^2(\partial M)$ in general when $\lambda_i\neq \lambda_j$. In this paper, based on the ideas in \cite{HT} plus some estimates for extra terms which come up for spectral clusters, we prove the upper bound from (\ref{bounds}) replacing $e_j$ by $\chi_{\lambda}^s f$ on general manifolds, for any $s>0$, \t[{\bf Upper Bound}]\label{Thm1} Let $M$ be a smooth compact Riemannian manifold with boundary, and $u=\chi_{\lambda}^s f$ be the spectral clusters, we have that $\forall s>0$, there exists $C>0$ independent of $\lambda$ and $s$, such that $$ \|\partial_{\nu} u\|_{L^2(\partial M)}\leq C\sqrt{1+s}(\lambda+s) \|u\|_{L^2(M)}.$$ \et Especially for spectral projection $P_{\lambda}(f)=\displaystyle\sum_{\lambda_j\in(0,\lambda]}e_j(f)$, we have the upper bound estimate for its normal derivative: \c\label{Cor1} Let $M$ be a smooth compact Riemannian manifold with boundary, for spectral projection $P_{\lambda}(f)$, we have $$ \|\partial_{\nu}P_{\lambda}(f) \|_{L^2(\partial M)}\leq C\lambda^{3/2} \|P_{\lambda}(f)\|_{L^2(M)}.$$ \ec Next it will be more subtle to study the lower bound from (\ref{bounds}) replacing $e_j$ by $\chi_{\lambda}^s f$. It might only hold when the value of $s$, the width of the spectral cluster, is sufficiently small. This can be seen in the case of the unit disc. If $s>\pi$, we can take a suitable linear combination of two consecutive eigenfunctions with angular dependence $e^{in\theta}$ (these are of the form $e^{in\theta} J_n(\alpha r)$ where $\alpha$ is a zero of the Bessel function $J_n$) and find a function in a ``wide" spectral cluster with zero normal derivative. In order to obtain the lower bound, we first study on bounded Euclidean domains. Following an idea of Rellich \cite{R} for single eigenfunctions on bounded Euclidean domains, we have the lower bound from (\ref{bounds}) replacing $e_j$ by $\chi_{\lambda}^s f$ on bounded Euclidean domains for small $s>0$, \t[{\bf Lower Bound for Euclidean Domains}]\label{Thm2} Let $M \subset \R^n$ be a bounded Euclidean domain, and $R_M=\max_{x,y\in M}|x-y|$ be the diameter of the domain $M$, and $u=\chi_{\lambda}^s f$ be the spectral clusters. Then for $0<s<\frac{1}{2R_M}$, there exists $C_s>0$ independent of $\lambda$, such that $$ \| \partial_{ \nu} u\|_{L^2(\partial M)}\geq C_s\lambda \|u\|_{L^2(M)}.$$ \et Next we turn to study the lower bound on general manifolds. To show a basic picture of our theorem, we refer some simple examples from \cite{HT} for single eigenfunctions, i.e., the cylinder (Example 3 in \cite{HT}), the hemisphere (Example 4 in \cite{HT}), the spherical cylinder (Example 5 in \cite{HT}). In all these examples, the upper bound holds, but the lower bound fails. These examples lead one to expect that the failure of the lower bound is related to the presence of geodesics in $M$ which do not reach the boundary. We obtain the lower bound estimates as in \cite{HT} replacing $e_j$ by $\chi_{\lambda}^s f$ for small $0<s<s_M$, where $s_M$ depends on the manifold only, \t[{\bf Lower Bound for Manifolds}]\label{Thm3} Suppose $M$ has {\bf no trapped geodesics}, i.e., $M$ can be embedded in the interior of a compact manifold with boundary, $N$, of the same dimension, such that every geodesic in $M$ eventually meets the boundary of $N$, and $u=\chi_{\lambda}^s f$ be the spectral clusters. There exists $s_M>0$, which depends on the manifold only, such that for any $0<s<s_M$, there exists $C_s>0$ independent of $\lambda$, such that $$ \|\partial_{\nu} u\|_{L^2(\partial M)}\geq C_s\lambda \|u\|_{L^2(M)}.$$ \et We organize our paper as the following: In section 2, we prove a Rellich-type estimate from Green's formula, and some perturbation estimates to deal with the extra terms in the Rellich-type estimate. In section 3, we prove the the upper bound for general manifolds using the estimates from section 2 following the argument in \cite{HT}. In section 4, we prove the lower bound for Euclidean domains using the fact that the commutator $[-\Delta_M,x\cdot\nabla]=-2\Delta_M$, which gives the idea of the proof of lower bound for general case. In section 5, we show the lower bound for $L^2$ norm of $\partial_{\nu}u$ on an arbitrary compact Riemannian manifold $M$ satisfying the no trapped geodesics condition in Theorem \ref{Thm3} by finding a differential operator $P$ of order $2K-1$ which has a positive commutator with $-\Delta_M$, which depends on a trick due to Morawetz, Ralston and Strauss \cite{MRS}. In Appendix, we study the $L^2$ estimates for spectral clusters near the boundary, which are needed in the proof of Theorem \ref{Thm3}, following the same ideas in section 3 in \cite{HT} for single eigenfunctions. In what follows we shall use the convention that $C$ denotes a constant that is not necessarily the same at each occurrence. \section{\bf Rellich-type estimates and Perturbation estimates} To prove the upper bound, and the lower bound for Euclidean domains, we use the following Lemma which we call a Rellich-type estimate. \la({\bf Rellich-type estimates}) Let $u=\chi_{\lambda}^s f$ be the spectral projection of $f$. Then for any differential operator A, \aa \int_Y \partial_{\nu} u Au d\sigma &= &\int_M <u,[-\Delta,A]u>dg + \int_M <(-\Delta-\lambda^2)u,Au>dg \nn\\ &&-\int_M <u,A(-\Delta-\lambda^2)u>dg.\label{Rellich} \eaa \el {\bf Proof:} The proof is very simple. By Green's Formula, one has \begin{eqnarray*} \int_Y \partial_{\nu} u Au d\sigma - \int_Y u \partial_{\nu} Au d\sigma= \int_M <-\Delta u,Au>dg -\int_M <u,-\Delta Au>dg.\end{eqnarray*} Note that $u\equiv 0$ on $Y$, left side of above equality gives left side of (\ref{Rellich}). Use the fact that $[-\Delta,A]=[-\Delta-\lambda^2,A]$ to write the right side as \begin{eqnarray*} && \int_M <(-\Delta-\lambda^2)u,Au>dg -\int_M <u,(-\Delta-\lambda^2)Au>dg\\ &=& \int_M <u,[-\Delta,A]u>dg + \int_M <(-\Delta-\lambda^2)u,Au>dg \\ &&-\int_M <u,A(-\Delta-\lambda^2)u>dg. \end{eqnarray*}\qe \r If we pick $f=e_j$ the eigenfunction with eigenvalue $\lambda_j^2$, using the fact that $\Delta e_j+\lambda_j^2 e_j=0$, the above Lemma is reduced to Lemma 2.1 in \cite{HT}. \er Since there have two additional terms in (\ref{Rellich}) comparing with Lemma 2.1 in \cite{HT}, we need estimates them by the following Lemma. \la({\bf Perturbation estimates}) Let $u=\chi_{\lambda}^s f$ be the spectral projection of $f$, $A$ is a differential operator with order one, we have \begin{eqnarray*} &&||(-\Delta-\lambda^2)u ||_2\leq 2s(\lambda+s) ||u||_2;\\ && ||A(-\Delta-\lambda^2)u ||_2\leq C_A s(\lambda+s)^2 ||u||_2. \end{eqnarray*} \el {\bf Proof:} For the first inequality, by direct computation, we have: \begin{eqnarray*} ||(-\Delta-\lambda^2)u ||_2^2&=&\int_M<(-\Delta-\lambda^2)u ,(-\Delta-\lambda^2)u >dg\\ &=&\sum_{\lambda_j\in[\lambda,\lambda+s)}(\lambda_j^2-\lambda^2)^2e_j^2(f)\\ &<&\sum_{\lambda_j\in[\lambda,\lambda+s)}(2s\lambda+s^2)^2e_j^2(f)\\ &<&4s^2(\lambda+s)^2||u||_2^2 \end{eqnarray*} For the second inequality, since $A$ is a differential operator with order one and $M$ is compact, we have pointwise estimates $$|A f(x)|\leq C_A|\nabla f(x)|,\qquad \forall x\in M\; and \; \forall f\in C^1(M).$$ With this estimates, by direct computation, we have: \begin{eqnarray*} ||A(-\Delta-\lambda^2)u ||_2^2&\leq& C_A^2||\nabla (-\Delta-\lambda^2)u ||_2^2\\ &=&C_A^2\int_M<\nabla(-\Delta-\lambda^2)u ,\nabla(-\Delta-\lambda^2)u >dg\\ &=&C_A^2\sum_{\lambda_j\in[\lambda,\lambda+s)}(\lambda_j^2-\lambda^2)^2\lambda_j^2e_j^2(f)\\ &<&C_A^2\sum_{\lambda_j\in[\lambda,\lambda+s)}(2s\lambda+s^2)^2\lambda_j^2e_j^2(f)\\ &<&C_A^24s^2(\lambda+s)^4||u||_2^2 \end{eqnarray*} \qe \section{\bf Upper bound for general manifolds} In this section, we shall prove the the upper bound for general manifolds. Here we use the geodesic coordinates with respect to the boundary. We can find a small constant $\delta>0$ so that the map $x=(y,r)\in Y\times [0,\delta) \rightarrow M$, sending $(y,r)$ to the endpoint $x$, of the geodesic of length $r$ which starts a $y\in Y=\partial M$ and is perpendicular to $Y$ is a local diffeomorphism. In this local coordinates $x=(y, r)$, the metric $ g = dr^2 + h_{ij} dy_i dy_j$ and the Riemannian measure \begin{equation} dg = k^2 dr dy, \quad where \quad k^4 = \det h_{ij}.\label{k} \end{equation} and the Laplacian can be written as \begin{eqnarray*} \Delta_g= \sum_{i,j=1}^n g^{ij}(x)\frac{\partial^2}{\partial x_i \partial x_j} + \sum_{i=1}^n b_i(x) \frac{\partial}{\partial x_i}, \end{eqnarray*} where $(g^{ij}(x))_{1\le i,j \le n}$ is the inverse matrix of $(g_{ij}(x))_{1\le i,j \le n}$, and $g^{nn}=1$, and $g^{nk}=g^{kn}=0$ for $k\neq n$. Also the $b_i(x)$ are $C^{\infty}$ and real valued. {\bf Proof of Theorem \ref{Thm1}:} For $u=\chi_{\lambda}^s f$, to prove an upper bound for the $L^2$ norm of $\partial_{\nu} u$, we choose an operator $A$ so that the left hand side of (\ref{Rellich}) in Lemma 2.1 is a positive form in $\partial_{\nu} u$. To do this, we choose $A = \chi(r) \partial_r$, where $\chi \in C_c^\infty(\R)$ is identically $1$ for $r$ close to zero, and vanishes for $r \geq \delta$. The left hand side of (\ref{Rellich}) in Lemma 3.1 is then precisely the square of the $L^2$ norm of $\partial_{\nu} u$. After one integration by parts for the first term of the right hand side of (\ref{Rellich}) in Lemma 3.1, there are first order (vector-valued) differential operators $B_1$, $B_2$ with smooth coefficients, $$\int_M <u,[-\Delta,A]u>dg=\int_M <B_1 u, B_2 u>dg. $$ From Lemma 3.2, each term of the right hand side of (\ref{Rellich}) in Lemma 3.1 is dominated by \begin{eqnarray*} |\int_M <u,[-\Delta,A]u>dg|&=&|\int_M<B_1 u, B_2 u>dg|\leq C_A||\nabla u||_2^2 \leq C_A(\lambda+s)^2||u||_2^2\\ |\int_M <(-\Delta-\lambda^2)u,Au>dg|&\leq& ||(-\Delta-\lambda^2)u||_2||Au||_2\leq C_As(\lambda+s)^2||u||_2^2\\ |\int_M <u,A(-\Delta-\lambda^2)u>dg|&\leq& ||A(-\Delta-\lambda^2)u||_2||u||_2\leq C_As(\lambda+s)^2||u||_2^2 \end{eqnarray*} where $C$ and $C_A$ depend on the domain, but not on $\lambda$. This proves the upper bound for any compact Riemannian manifold with boundary. \qe If we choose $A = Q^*Q\partial_r$ near the boundary in the above proof, with $Q$ an elliptic differential operator of order k in the y variables, one has the $H^k$ estimates for upper bound of the spectral clusters $u=\chi_{\lambda}^s f$: \t\label{Thm-H-k} \aa ||\partial_{\nu} u||_{H^k(Y)}\leq C\sqrt{1+s}(\lambda+s)^k||u||_2\label{deriv-upper-bounds} \eaa for any integer k, and hence (by interpolation) any real k. \et \section{\bf Lower bound for Euclidean domains} In this section, we shall prove Theorem \ref{Thm2}, the lower bound for Euclidean domain $M \subset \R^n$. {\bf Proof of Theorem \ref{Thm2}:} We choose $A$ so that the first term in the right hand side, rather than the left hand side, of (\ref{Rellich}) in Lemma 2.1 is a positive form. Without lose of generality, assume $M\subset \{x\in \R^n| |x|\leq \frac{R_M}{2}\}$. We choose \begin{equation} A = \sum_{i=1}^n x_i \frac{\partial}{\partial x_i}=x\cdot\nabla.\nn \end{equation} As is very well known in scattering theory, the commutator of this with $-\Delta$ (which is minus the Euclidean Laplacian here) is $[-\Delta,A] = -2\Delta$, and for any $g\in C^1(M)$, $|A g(x)|\leq \frac{R_M}{2}|\nabla g(x)|$ for all $x\in M$. Hence, in this case the left side of (\ref{Rellich}) gives us \begin{equation} \int\limits_{Y} \frac{\partial u}{\partial \nu} Au \, d\sigma = \int\limits_{Y} \nu \cdot x \, \big( \frac{\partial u}{\partial \nu} \big)^2 \, d\sigma \leq C \| \partial_{\nu}u \|_2^2, \label{lower-Euc} \end{equation} And the right side of (\ref{Rellich}) gives us \begin{eqnarray} &&RIGHTSIDE \; of\; (\ref{Rellich}) \nn\\ &\geq& \int_M <u,-2\Delta u>dg -\|(-\Delta-\lambda^2)u\|_2\|Au \|_2-\|u\|_2\|A(-\Delta-\lambda^2)u\|_2\nn\\ &\geq& 2\|\nabla u\|_2^2-\frac{R_M}{2}\Big[\|(-\Delta-\lambda^2)u\|_2\|\nabla u\|_2+\|u\|_2\|\nabla(-\Delta-\lambda^2)u\|_2\Big] \nn\\ &\geq& \Big(2\lambda^2-2R_M s(\lambda+s)^2\Big) \|u\|_2^2,\label{lower-Euc-2} \end{eqnarray} which gives the lower bound. The equality in (\ref{lower-Euc}) for single eigenfunctions was proved by Rellich \cite{R}. \qe \section{\bf The lower bound on Riemannian manifolds} To find a lower bound for $L^2$ norm of $\partial_{\nu}\Big(\chi_{\lambda}^s f\Big)$ on an arbitrary compact Riemannian manifold $M$ satisfying the no trapped geodesics conditions of the main Theorem, we need to find a differential operator which has a positive commutator with $-\Delta_M$ as we did for domains in Euclidean spaces. One might wonder whether, on an arbitrary compact Riemannian manifold, with no trapped geodesics, one could choose a first order {\it differential} operator $A$ whose commutator with $-\Delta_M$ had a positive symbol. Example 8 in \cite{HT} shows that this is impossible in general. Firstly, we have a first order pseudo-differential operator $A$ on $N$ which has the required property to leading order, i.e., such that the symbol of $i[-\Delta,A]$ is positive: \begin{lem}\label{Q-lemma}({\bf Lemma 4.1 in \cite{HT}}) Given any geodesic $\gamma$ in $S^*N$, there is a first order, classical, self-adjoint pseudodifferential operator $Q$ satisfying the transmission condition (see \cite{Ho}, section 18.2), and properly supported on $N$, such that the principal symbol $\sigma(i[-\Delta,Q])$ of $i[-\Delta,Q]$ is nonnegative on $T^*M$, and \begin{equation} \sigma(i[-\Delta,Q]) \geq \sigma(-\Delta) = |\xi|^2 \label{comm-cond} \end{equation} on a conic neighborhood $U_\gamma$ of $\gamma \cap T^*M$. \end{lem} We now use Lemma~\ref{Q-lemma} to construct our operator $A$. For each geodesic $\gamma$ in $S^*N$, we have a conic neighborhood $U_\gamma$ as in the Lemma. By compactness of $S^*M$, a finite number of the $U_\gamma$ cover $S^*M$. Let $A$ be the sum of the corresponding $Q_\gamma$. Then Lemma~\ref{Q-lemma} implies that \begin{equation} \sigma(i[-\Delta,A]) \geq |\xi|^2 \; on \; T^*M . \label{A} \end{equation} Secondly, we turn the pseudodifferential operator $A$ into a differential operator $P$ of order $2K-1$ with positive commutator with $-\Delta$ as Hassell and Tao did at Section 5 in \cite{HT} for single eigenfunctions, which depends on a trick due to Morawetz, Ralston and Strauss \cite{MRS}. Recall some facts about spherical harmonics. Let $\Delta_{S^{n-1}}$ denote the Laplacian on the $(n-1)$-sphere, which has eigenvalues $k(n+k-2)$, $k = 0, 1, 2, \dots$, and the corresponding eigenspace be denoted $V_k$. We recall that for every $\phi \in V_k$, the function $r^k \phi$ (thought of as a function on $\R^n$ written in polar coordinates) is a homogeneous polynomial, of degree $k$, on $\R^n$. We summarize the needed results from Section 5 in \cite{HT} as the following proposition: \p[Hassell-Tao \cite{HT}] Since the symbol $a$ of the operator $A$ is odd, there is spherical harmonics expansion of $a$ restricted to the cosphere bundle of $N$ $$ a \restriction_{ S^*N}= \sum_{l=0}^\infty \phi_{2l+1}(x, \frac{\xi}{|\xi|}), \quad \phi_k(x, \cdot) \in V_k(S^*_x N). $$ And there is a nature number $K$ such that the operator $A'$ with symbol $$ a' = \sum_{l=0}^{K-1} \phi_{2l+1}(x, \frac{\xi}{|\xi|}) $$ also has positive commutator with $-\Delta$. Following \cite{MRS}, one can turn $A'$ into a differential operator $P$ of order $2K-1$, by letting $$ p = \sigma(P) = \sum_{l=0}^{K-1} \phi_{2l+1}(x, \frac{\xi}{|\xi|}) |\xi|^{2K- 1}. $$ Moreover, the symbol of $i[-\Delta,P]$ satisfies $$ \sigma(i[-\Delta,P]) = |\xi|^{2K} \big( \sigma(i[-\Delta,P]) |_{|\xi| = 1} \big) \geq c|\xi|^{2K} \quad for\; some\; c > 0. $$ Applying the G$\mathring{a}$rding inequality to $Q = i[-\Delta,P]$, there is \ee \int_M \langle u, Qu \rangle dg \geq c \| u \|_{H^K(M)}^2 - C \Big( \| u \|_{L^2(M)}^2 + \sum_{k=0}^{K-1} \| \pa_r^k u \| _{H^{K-1/2-k}(Y)}^2 \Big), \label{Garding} \eee where $c$ is a positive constant depending on $P$ and $(M,g)$. \ep {\bf Proof of Theorem \ref{Thm3}:} Let $A=P$ in Lemma 2.1, \aa &&RIGHTSIDE \; of\; (\ref{Rellich}) \nn\\ &\geq& \int_M <u,Qu>dg -\|(-\Delta-\lambda^2)u\|_2\|Pu \|_2-\|u\|_2\|P(-\Delta-\lambda^2)u\|_2\nn\\ &\geq& \int_M <u,Qu>dg -C\Big[\|(-\Delta-\lambda^2)u\|_2\|u\|_{H^{2K-1}(M)}\nn\\ && +\|u\|_2\|(-\Delta-\lambda^2)u\|_{H^{2K-1}(M)}\Big] \nn\\ &\geq& \int_M <u,Qu>dg -Cs\lambda^{2K}\|u\|_2^2\nn\\ &\geq& (c-Cs)\lambda^{2K} \| u \|_2^2 - C \Big( \| u \|_{L^2(M)}^2 + \sum_{k=0}^{K-1} \| \pa_r^k u \| _{H^{K-1/2-k}(Y)}^2 \Big), \label{lower-general} \eaa where we use Lemma 2.2 and $\|u\|_{H^k(M)}\leq C\lambda^k\|u\|_2,\; \forall k>0$ to estimate the extra terms, and make use of (\ref{Garding}) to obtain the last inequality. Thus, there exists a constant $s_M>0$, which depends on $M$ only, such that the first term in (\ref{lower-general}) is positive when $0<s<s_M$. Next to consider the left hand side of (\ref{Rellich}). Let us write $u = k^{-1}v$, where $k$ is as in (\ref{k}), so that $v$ satisfies (\ref{v-eqn}) in Appendix. Then $Pu = \tilde P v$, where $\tilde P = P \circ k$ is a differential operator of order $2K-1$. Since $k$ is smooth, we obtain \begin{equation} C_s\lambda^{2K}\|u\|_2^2 \leq \|u\|_2^2 + \sum_{k=0}^{K-1} \| \pa_r^k v \| _{H^{K-1/2-k}(Y)}^2 + \big| \int_{Y} \langle \partial_{\nu}v, \tilde P v \rangle \, d\sigma \big|. \label{eqqq}\end{equation} Since $v=0$ at $Y$, and we are interested in $\tilde P v |_Y$, we may assume that $\tilde P = P' \pa_r$, where $P'$ has order $2K-2$. Using (\ref{v-eqn}) in Appendix, we may replace $\pa_r^2 v$ by $-(\lambda - F)v - \pa_{y_i} (h^{ij} \pa_{y_j} v)+H$ repeatedly, until only $\pa_r \pa_y^\alpha v$ terms remain. Thus we have \begin{eqnarray*} \tilde P v |_Y = \sum_{j=0}^{K-1} \lambda^j P_j (\partial_{\nu}v), \end{eqnarray*} where $P_j$ is a differential operator on $Y$ of order $2(K-1-j)$, independent of $\lambda$. Hence (\ref{eqqq}) becomes \begin{equation} C'_s\lambda^{2K}\|u\|_2^2 \leq \|u\|_2^2 + \sum_{k=0}^{K-1} \lambda^{2k-2} \| \partial_{\nu}u \| _{H^{K-1/2-k}(Y)}^2 + \sum_{j=0}^{K-1} \lambda^{2j}\big| \int_{Y} \langle \partial_{\nu}v, P_j (\partial_{\nu}v) \rangle \, d\sigma \big|.\nn \end{equation} The argument to reduce $Pu$ on $Y$ to $\sum_{j=0}^{K-1} \lambda^j P_j (\partial_{\nu}v)$ on $Y$ is the same as what Hassell and Tao did at Section 5 in \cite{HT} for single eigenfunctions. Using the upper bound estimate (\ref{deriv-upper-bounds}) for $H^k$ norm on the sum over $k$ and for all terms in the sum over $j$ with $j < K-1$, we find \begin{eqnarray*} C''_s\lambda^{2K}\|u\|_2^2 \leq (1 + \lambda^{2K-1})\|u\|_2^2 + \lambda^{2K-2} \| \partial_{\nu}u \|_2^2 + \lambda^{2K-1} \|u\|_2\| \partial_{\nu}u\|_2 . \end{eqnarray*} which gives \begin{equation} \| \partial_{\nu}u \|_2^2+\lambda\|u\|_2\| \partial_{\nu}u\|_2-(C''_s-\lambda^{-1}-\lambda^{-2K} )\lambda^{2}\|u\|_2^2\geq 0.\label{almost} \end{equation} Solve the inequality (\ref{almost}), for $\lambda$ large enough, we have constant $C_s$ independent of $\lambda$, such that \begin{eqnarray*} \| \partial_{\nu}u \|_2\geq C_s\lambda\|u\|_2, \end{eqnarray*} This proves the lower bound. \qe \r One may also prove the lower bound following what Hassell and Tao did at Section 4 in \cite{HT} for single eigenfunctions almost line by line, while one need do some additional estimates on the nonhomogeneous terms like $H$ in (\ref{v-eqn}), which can be looked as small perturbation terms when $s>0$ is small enough. This approach is length and involves many pseudodifferential operator constructions and calculus. \er \section{\bf Appendix: Estimates for spectral clusters near the boundary} Here we study the $L^2$ estimates for spectral clusters near the boundary, which are needed in the proof of Theorem \ref{Thm3}, following some ideas from section 3 in \cite{HT} with its erratum \cite{HT1} for single eigenfunctions and the upper bound for $\partial_{\nu}\Big(\chi_{\lambda}^s f\Big)$ from Theorem \ref{Thm1}. As in section 3, we use the geodesic coordinates system $(y,r)$ near the boundary. Let us denote the boundary of $M$ by $Y$, and write $Y_r$ for the set of points at distance $r$ from the boundary, which is a submanifold for $r \le \delta$. Suppose that $u=\chi_{\lambda}^s f$ is a spectral cluster for Dirichlet Laplacian. Similar as Lemma 3.2 in \cite{HT} with its erratum \cite{HT1} for a single eigenfunction, we derive an estimate on the $L^2$ norm of the spectral cluster $u=\chi_{\lambda}^s f$ on $Y_r$, exploiting the fact that $u$ vanishes on the boundary. \begin{Prop} \label{L7-2} There exists $C > 0$, independent of $\lambda$ and $s$, such that \begin{equation} \int_{Y_r} u^2 d\sigma(y) \leq C\sqrt{1+s}(\lambda+s)^2 r^2\|u\|_2^2 \quad \forall\; r \in [0, \frac{\delta}{3}]. \label{bdy-est}\end{equation} \end{Prop} It will be convenient to change to the function $v = ku$ (this is equivalent to looking at the Laplacian acting on half-densities). Denote $u_l=e_l(f)$, $v_l=ku_l$ for $l\in[\lambda, \lambda+s)$. From the equation (3.2) in \cite{HT}, $v_l$ solves the equation \begin{equation} \pa_r^2 v_l + \pa_i (h^{ij} \pa_j v_l) + \lambda_l^2 v_l + F v_l = 0, \quad h^{ij} = (h_{ij})^{-1}\nn \end{equation} where $$ F = - k^{-1}\pa_r^2 k - k^{-1} \pa_i ( h^{ij} \pa_j k ) $$ is a smooth function on $M$. We have \begin{equation}\label{v-eqn} \pa_r^2 v + \pa_i (h^{ij} \pa_j v) + \lambda^2 v + F v = \sum_{\lambda_l\in[\lambda, \lambda+s)}(\lambda^2-\lambda_l^2)v_l=H \end{equation} As did in section 3 of \cite{HT} for a single eigenfunction, where the nonhomogeneous term $H$ doesn't appear, for the spectral cluster $u$, we define a sort of `energy' $E(r)$ for each value of $r$: \begin{equation} E(r) = \frac1{2} \int_{Y_{r}} \big( v_r^2 + (\lambda^2 + F)v^2 - h^{ij} \pa_i v \pa_j v-Hv \big) dy. \end{equation} This is obtained formally from the energy for hyperbolic operators, with $r$ playing the role of a time variable, by switching the sign of the term involving tangential derivatives. Similar as Lemma 3.1 in \cite{HT}, we have the following estimate for $E(r)$ \begin{lem}\label{L7-1} For $r \in [0, \delta]$, \begin{equation}\label{energy-est} |E(r)| \leq C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2 \end{equation} where $C$ is independent of $\lambda$ and $s$. \end{lem} {\bf Proof of Lemma \ref{L7-1}:} From the upper bound argument in section 3, we know that $E(0) = \frac1{2} \| \partial_{\nu} u \|_2^2 \leq C\lambda^2\| u \|_2^2$. We compute the derivative of $E(r)$: \begin{eqnarray*} \frac{\pa}{\pa r} E(r) = \int_{Y_{r}} \big( \pa_r^2 v \pa_r v +(\lambda^2 + F)v \pa_r v -h^{ij} \pa_i v \pa_r \pa_j v -Hv_r\\ + \frac{\pa h^{ij}}{\pa r} \pa_i v \pa_j v + \frac{\pa F}{\pa r} v^2 -H_rv\big) dy. \end{eqnarray*} Integrating by parts in the third term, using the equation for $v$, and applying Cauchy-Schwarz to last term, we obtain \begin{eqnarray*} \big| \frac{\pa E}{\pa r} \big|(r) &\leq& C \int_{Y_{r}} \big( v^2 + |\nabla v|^2 +\lambda^2|v|^2+\lambda^{-2}|\nabla H|^2\big) dy \\ &\leq& C \int_{Y_{r}} \big( u^2 + |\nabla u|^2+ \lambda^2|u|^2+\lambda^{-2}|\nabla (H/k)|^2\big) k^2 dy. \end{eqnarray*} Thus, for $r_0 \in [0, \delta]$, \begin{eqnarray*} E(r_0) &=& E(0) + \int_0^{r_0} \frac{d}{dr} E(r) dr \\ &\leq & C\lambda^2\| u \|_2^2 + \int_M \big( u^2 + |\nabla u|^2+ \lambda^2|u|^2 +\lambda^{-2}|\nabla (H/k)|^2\big) dg\\ &\leq& C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2, \end{eqnarray*} where we use Lemma 2.2 to estimate them last term. \qe \vspace{4mm} {\bf Proof of Proposition \ref{L7-2}:} Here we follow the main idea of proof of Lemma 3.2 in \cite{HT} with its erratum \cite{HT1} for single eigenfunctions. Consider the $L^2$ norm on $Y_r$, $$L(r) = \int_{Y_r} u^2 k^2 dy = \int_{Y_r} v^2 dy.$$ And we have \begin{equation}\label{u-bd} \int_0^\delta L(r) dr \leq \int_M u^2 \, dg = \|u\|_2^2. \end{equation} By direct computation, we have $$ L'(r_0) = \int_{Y_{r_0}} 2 v v_r\ dy\quad \mathrm{and} \quad L''(r_0) =4 \int_{Y_{r_0}} v_r^2 \ dy \ - 4E(r_0).$$ On the other hand, from Cauchy-Schwarz we have $$ 4 \int_{Y_{r_0}} v_r^2 \ \geq \ \frac{(\int_{Y_{r_0}} 2 v v_r\ dy)^2}{\int_{r=r_0} v^2\ dy} = \frac{L'(r_0)^2}{L(r_0)}.$$ Thus we have the differential inequality for $L(r)$: \begin{equation} L'' \geq \frac{(L')^2}{L} - C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2 , \label{diff-ineq} \end{equation} for some constant $C$ depending only on the manifold $M$. Define the quantity $$ B(r) := \frac{L'(r)^2}{L(r)^2} - \frac{C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2}{L(r)} $$ For any $r\in [0, \delta]$ with $L'(r)>0$, from (\ref{diff-ineq}) we have $$ B'(r) = \frac{2 L' L''}{L^2} - \frac{2 (L')^3}{L^3} + \frac{2 C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2 L'}{L^2} \geq 0. $$ Hence $B(r)$ is non-decreasing for $r\in [0, \delta]$ with $L'(r)>0$. \vspace{4mm} {\bf Claim:} There is $\Lambda>0$ such that for any $\lambda\geq \Lambda$, either $L'(r)\leq 0$ or $B(r)\leq 0$ are true for all $0<r<\delta/3$. \vspace{4mm} Define $O_{\lambda}=\{r | L'(r)>0,\; 0<r<\delta \}=\cup (a_n^{\lambda}, b_n^{\lambda})\subset (0,\delta]$. For any $r\in (a_n^{\lambda}, b_n^{\lambda})$ with $b_n^{\lambda}<\delta$, we have $$ B(r)\leq B(b_n^{\lambda})= - \frac{C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2}{L(b_n^{\lambda})}\leq 0. $$ Hence {\bf Claim} will be true unless there is an unbounded sequence of $\lambda$ such that $B(r_0)> 0$ for some $r_0\in (a_n^{\lambda}, \delta)$ and $0 < r_0 < \delta/3$, where $(a_n^{\lambda}, \delta)$ is one subinterval of $O_{\lambda}$. Then we would have $B(r) > 0$ for all $r \geq r_0$, so $$ L'(r)^2 > 2C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2 L(r) \hbox{ for all } r \geq r_0. $$ In particular $L'(r)$ would be strictly positive for $r \geq r_0$. We rearrange this as $$ (L(r)^{1/2})' = \frac{1}{2} L'(r) L(r)^{-1/2} > \sqrt{C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2} \hbox{ for all } r > r_0. $$ This would give $$ L(r) \geq C'(1+s) \lambda^2(r-r_0)^2\|u\|_2^2 \hbox{ for all } r >\delta/3>r_0. $$ This would contradict the bound (\ref{u-bd}) for large $\lambda$. Hence {\bf Claim} is true. \vspace{4mm} Thus, for $\lambda \geq \Lambda$ (where $\Lambda$ is obtained from {\bf Claim} and depends only on $M$), we must have $B(r) \leq 0$ or $L'(r) \leq 0$ for all $0 < r \leq \delta/3$. In either case $$ (L(r)^{1/2})' = \frac{1}{2} L'(r) L(r)^{-1/2} \leq \sqrt{C\sqrt{1+s}(\lambda+s)^2\|u\|_2^2} \hbox{ for all } r \leq \frac{\delta}{3}. $$ Since $L(0) = 0$, this implies (\ref{bdy-est}) for $\lambda \geq \Lambda$. Next for $\lambda < \Lambda$, since $$ u=\sum_{\lambda_j\in[\lambda,\lambda+s)}e_j(f)\leq \Big(\sum_{\lambda_j\in[\lambda,\lambda+s)}e_j^2\Big)^{1/2}\Big(\sum_{\lambda_j\in[\lambda,\lambda+s)}||e_j(f)||_2^2\Big)^{1/2} =\Big(\sum_{\lambda_j\in[\lambda,\lambda+s)}e_j^2\Big)^{1/2}||u||_2 $$ we have $$ \int_{Y_r} u^2 d\sigma(y)\leq \int_{Y_r} \sum_{\lambda_j\in[\lambda,\lambda+s)}e_j^2 d\sigma(y)\|u\|_2^2 \leq C\Big(\sum_{\lambda_j\in[\lambda,\lambda+s)}\lambda_j^2\Big) r^2\|u\|_2^2\leq C_{\Lambda} r^2\|u\|_2^2 $$ where we make use of the result of Lemma 3.2 in \cite{HT} with its erratum \cite{HT1} for single eigenfunctions: $$ \int_{Y_r} e_j^2 d\sigma(y)\leq C\lambda_j^2, \quad \lambda_j\leq \Lambda+s. $$ \qe \vspace{5mm} {\bf Acknowledgement:} The author would like to thank Professor Andrew Hassell for pointing out a mistake in first version of this paper and some helpful suggestions on this paper. \bibliographystyle{abbrv}

https://arxiv.org/abs/1305.2661

On Improved Bounds on Bounded Degree Spanning Trees for Points in Arbitrary Dimension

Given points in Euclidean space of arbitrary dimension, we prove that there exists a spanning tree having no vertices of degree greater than 3 with weight at most 1.559 times the weight of the minimum spanning tree. We also prove that there is a set of points such that no spanning tree of maximal degree 3 exists that has this ratio be less than 1.447. Our central result is based on the proof of the following claim:Given $n$ points in Euclidean space with one special point $V$, there exists a Hamiltonian path with an endpoint at $V$ that is at most 1.559 times longer than the sum of the distances of the points to $V$.These proofs also lead to a way to find the tree in linear time given the minimal spanning tree.

\section{Abstract} Given points in Euclidean space of arbitrary dimension, we prove that there exists a spanning tree having no vertices of degree greater than 3 with weight at most 1.559 times the weight of the minimum spanning tree. We also prove that there is a set of points such that no spanning tree of maximal degree 3 exists that has this ratio be less than 1.447. Our central result is based on the proof of the following claim: Given $n$ points in Euclidean space with one special point $v$, there exists a Hamiltonian path with an endpoint at $v$ that is at most 1.559 times longer than the sum of the distances of the points to $v$. These proofs also lead to a way to find the tree in linear time given the minimal spanning tree. \section{Introduction} \label{sec:problem} The minimum spanning tree (MST) problem in graphs is perhaps one of the most basic problems in graph algorithms. An MST is a spanning tree with minimal sum of edge weights. Efficient algorithms for finding an MST are well known. One variant on the MST problem is the bounded degree MST problem, which consists of finding a spanning tree satisfying given upper bounds on the degree of each vertex and with minimal sum of edges weights subject to these degree bounds. In general, this problem is NP-hard \cite{NPhard}, so no efficient algorithm exists. However, there are certain achievable results. For undirected graphs, Singh and Lau \cite{SinghLau} found a polynomial time algorithm to generate a spanning tree with total weight no more than that of the bounded degree MST and with each vertex having degree at most one greater than that vertex's bound. If the graph is undirected and satisfies the triangle inequality, Fekete and others \cite{network} bound the ratio of the total weight of the bounded-degree MST to that of any given tree, with a polynomial-time algorithm for generating a spanning tree satisfying the degree constraints and this ratio bound. The Euclidean case, with vertices being points in Euclidean space and edge weights being Euclidean distances, also has a rich history. We denote (following Chan in \cite{1.633}) by $\tau_k^d$ the supremum, over all sets of points in $d$-dimensional Euclidean space, of the ratio of the weight of the bounded degree MST with all degrees at most $k$ to the weight of the MST with no restrictions on degrees ($\tau_k^\infty$ is the supremum of $\tau_k^d$ over all $d$). For $k=2$, the bounded-degree MST problem becomes the Traveling Salesman Problem and $\tau_2^d=2$ \cite{network}, thus making $k=3$ the first unsolved case. Papadimitriou and Vazirani \cite{NPhard} showed that finding the degree-3 MST is NP-hard. Khuller, Raghavachari, and Young \cite{1.67} showed that $1.104\approx (\sqrt{2}+3)/4 \leq \tau_3^2 \leq 1.5$ and $1.035< \tau_4^2 \leq 1.25$. Chan \cite{1.633} improved the upper bounds to 1.402 and 1.143, respectively. Jothi and Raghavachari \cite{degree4} showed that $\tau_4^2 \leq (2+\sqrt{2})/3\approx 1.1381$. $\tau_5^2=1$ since there is always an MST with maximal degree 5 or less \cite{degree5}. These same papers also studied the problem in higher dimensions. Khuller, Raghavachari, and Young \cite{1.67} gave an upper bound on $\tau_3^\infty$ of $5/3\approx 1.667$, which Chan \cite{1.633} improved to $2\sqrt{6}/3\approx 1.633$. These two followed the same approach, proving these bounds on a certain ratio, which we will call $r$. $r$ is the maximum ratio between the shortest path through a collection of points starting at a special point and the size of a star centered at that point. It is conjectured to actually be 1.5. Khuller, Raghavachari, and Young \cite{1.67} showed that $\tau_3^\infty \leq r$. This is achieved in linear time as follows: \begin{enumerate} \item root the original tree \item treating the root as $v$, find a Hamiltonian path with ratio at most $r$ through its children. \item repeat recursively on each child. \end{enumerate} Each vertex then has at most 3 neighbors: two as a child and one as a parent. We improve previous upper bounds on $r$, and thus $\tau_3^\infty$, to 1.559. The proof leads to a linear time algorithm for generating the path and thus the bounded degree tree. Our approach is based on Chan's, but we weigh paths differently and select the number of points to remove when performing the induction based on the distances of points to $v$. We also find, by construction, a non-trivial lower bound of about 1.447 on $\tau_3^\infty$. In Section \ref{sec:lemmas}, we go over $r$ a bit more carefully as well as refering to a useful paper and discuss how we will use it. In Section \ref{sec:upperbound}, we improve the upper bound on $r$ to 1.559, and in Section \ref{sec:lowerbound} we improve the lower bound on $\tau_3^\infty$ to 1.447. \section{Preliminaries} \label{sec:lemmas} $r$ is properly defined as follows: Given point $v$ and $m$ points $a_1,a_2,\ldots,a_m$ in a Euclidean space of arbitrary finite dimension, let $\displaystyle S=\sum_{i=1}^n d(v,a_i)$ and let $L$ be the length of the shortest possible path that starts at $v$ and goes around the other points in some order (it does not go back to $v$). Then $r$ is the supremum of the possible values of $L/S$ over all arrangements of points in any number of dimensions. $r=1.5$ is achieved for $m=2$ in one dimension by the points $v=0,a_1=1,a_2=-1$. We use the results of Young \cite{distsum} multiple times in order to bound certain sums of distances. This paper deals with the maximum of weighted sums (with weights $w_{i,j}$) of lengths between $n$ points in $n-1$ dimensional Euclidean space, given that each point $a_i$ is specified as being no further than some distance $l_i$ from the origin. \begin{equation}\label{eq: Young} \max \left(\sum_{1\leq i<j\leq n} w_{i,j} d(a_i,a_j) \right)=\min \left(\sqrt{\sum_{1\leq i<j\leq n} \frac{w^2_{\displaystyle i,j}}{x_ix_j}}\sqrt{\sum_{i=1}^n l^2_ix_i}\sqrt{\sum_{i=1}^n x_i}\right) \end{equation} where the maximum is taken over all arrangements of points and the minimum is taken over all nonnegative $x_i$. Furthermore, Young specifies a relationship between the optimal arrangement and the values of $x_i$ where equality is achieved. Thus one can iteratively approximate the optimal arrangement using the same method as in \cite{experimental}, and then calculate $x_i$ values from it. Whenever \eqref{eq: Young} is used to give an upper bound on some weighted sum of distances, the values for $x_i$ used are given in Appendix \ref{sec: xis}. \section{Main proof of upper bound on $r$} \label{sec:upperbound} Let $r^*=1.559$. We will prove that $L\leq r^*S$ (as $L$ and $S$ are defined in the introduction), thus showing that $r<1.559$. We will prove this by strong induction on the number of points. Given $m$ vectors $a_1, a_2,\ldots,a_m$ with norms $d_1 \geq d_2 \geq d_3 \geq \ldots \geq d_m > 0$, respectively, we will try to induct by removing $a_1,\ldots,a_n$ for various values of $n$. We will try to traverse the other points, ending at $a_{n+1}$ or $a_{n+2}$. We will then add in the removed points, projected onto a sphere, and look at the average length of a path traversing them and ending at $a_1$ or $a_2$. We will then move them out in stages, seeing how this average path length changes at each stage, in order to bound the final average path length in terms of the values $d_k$. Since the average is an upper bound on the minimum, this gives us a linear inequality on the $d_k$ which is a sufficient condition for the inductive step to work. We then use linear programming to show that one of these inequalities is satisfied and thus that induction is possible. For the algorithm, we will then follow the induction to split the points up into blocks, choose the starting and ending vertex for one block at a time, using brute force to find the shortest path that goes through all the block's points. We start by defining $a_k=0$ and $d_k=0$ for all $k>m$. Introducing these new points does not affect the distance sum or the traversing path length, as the traversing path can go to them first. We will prove the following claim: \begin{claim}\label{claim} There exist two paths $P_1$ and $P_2$ ending at $a_1$ and $a_2$, respectively, such that the average of the lengths of these paths is at most $r^*S$ \end{claim} This clearly implies that $L\leq r^*S$. We will proceed by strong induction on $m$. To induct, remove $a_1$ through $a_n$ (where $n\geq 3$ may vary), use the inductive hypothesis to find two paths $P_1$ and $P_2$ through the other $m-n$ points, ending at $a_{n+1}$ and $a_{n+2}$, respectively. We will then try to find four paths $Q_{11}, Q_{12}, Q_{21}, Q_{22}$ with path $Q_{ij}$ going from $a_{n+i}$ to $a_j$ and going through all points $a_1, \ldots, a_n$, so that the average length of these four paths is at most $r^*(\sum_{i=1}^n d_i)$. We will assume that this is impossible, generate a set of conditions on the values $d_n$, then prove that one of the conditions must be violated. \subsection{Given $n \geq 3$} \label{subsec:given} In this section, we will assume $n>3$ to be a given value. We will select it in Section \ref{sec: recombine}. \begin{figure}[htbp] \centering \includegraphics[scale=.5]{LandSn.png} \label{fig:thickseg} \caption{The thick segments contribute to $L(a_n, \ldots, a_1)$; the dotted segments contribute to $S_n$.} \end{figure} Let $\displaystyle S_n=\sum_{i=1}^n d_i$. Let $L(u_{n+1}, \ldots, u_1)$ be the shortest length of a path $u_{s_{n+1}}, \ldots, u_{s_1}$ where $s$ is a permutation of $1, \ldots, n+1$ so that $s_{n+1}=n+1$ and $\{s_1,s_2\}=\{1,2\}$. Let $\overline{L(u_{n+1}, \ldots, u_1)}$ be the average length over all such paths $u_{s_{n+1}}, \ldots, u_{s_1}$. Then \begin{equation} \label{eq: L()} \begin{split} L(u_{n+1}, \ldots, u_1)&=\frac{1}{n-1}d(u_1,u_2)+\frac{1}{2(n-1)}d(u_1,u_{n+1})+\frac{1}{2(n-1)}d(u_2,u_{n+1})\\ +&\sum_{i=3}^n \frac{3}{2(n-1)}d(u_1,u_i)+\sum_{i=3}^n \frac{3}{2(n-1)}d(u_2,u_i)\\ +&\sum_{i=3}^n \frac{1}{n-1}d(u_i,d_{n+1})+\sum_{3\leq i<j\leq n} \frac{2}{n-1}d(u_i,u_j) \end{split} \end{equation} We wish to find upper bounds on $\overline{L(a_{n+1}, \ldots, a_1)}$ and $\overline{L(a_{n+2},a_n,a_{n-1}, \ldots, a_1)}$. For $1\leq~i\leq~n$, let \begin{equation} \label{eq: defineDnik} \begin{split} D_{n,i,1}&=\overline{L \left( a_{n+1}, \ldots, a_{i+1},a_i, \frac{d_i}{d_{i-1}}a_{i-1}, \ldots, \frac{d_i}{d_1}a_1 \right)} \\ &- \overline{L \left( a_{n+1}, \ldots, a_{i+1}, \frac{d_{i+1}}{d_{i}}a_i, \frac{d_{i+1}}{d_{i-1}}a_{i-1}, \ldots, \frac{d_{i+1}}{d_1}a_1 \right)}. \end{split} \end{equation} and let \begin{equation} \label{eq: defineDnk} D_{n,1}=\overline{L \left( a_{n+1},\frac{d_{n+1}}{d_n}a_n, \ldots, \frac{d_{n+1}}{d_1}a_1 \right)}. \end{equation} $D_{n,i,2}$ and $D_{n,2}$ are defined identically, except $a_{n+1}$ and $d_{n+1}$ are replaced with $a_{n+2}$ and $d_{n+2}$. For $k=1$ or $k=2$, \begin{equation} \label{eq: decompose} \overline{L(a_{n+k},a_n,a_{n-1},\ldots,a_1)}=D_{n,k}+\sum_{i=1}^n D_{n,i,k} \end{equation} Intuitively, we are setting all points at distance $d_{n+k}$, then moving out $n$ points to distance $d_n$, then moving out $n-1$ points, and so on. We will now find values $B_{n,i}$ and $B_n$ independent of the arrangement of $a_1,a_2,\ldots$ satisfying \begin{equation}\label{eq: UgeqD} \begin{split} B_{n,i}(d_i-d_{i+1}) &\geq D_{n,i,1}\\ B_n d_n &\geq D_{n,1} \end{split} \end{equation} The corresponding equations (substituting $a_{n+2}$ for $a_{n+1}$ and $d_{n+2}$ for $d_{n+1}$) will then hold for $D_{n,i,2}$ and $D_{n,2}$. \subsubsection{$B_n$} Define $g(n)$ as the maximum value of \begin{align*} d(u_1,u_2)+\frac{1}{2}d(u_1,u_{n+1})+\frac{1}{2}d(u_2,u_{n+1})+\sum_{j=3}^n\frac{3}{2}d(u_1,u_j)+\\ +\sum_{j=3}^n\frac{3}{2}d(u_2,u_j)+\sum_{j=3}^n d(u_j,u_{n+1})+\sum_{3\leq j<k\leq n} 2d(u_j,u_k) \end{align*} over unit vectors $u_1,\ldots,u_{n+1}$. We use equation \eqref{eq: Young} to obtain upper bounds on $g(n)$, which we then use to find numerical values for $B_n$. Substituting in $\eqref{eq: defineDnk}$ and $\eqref{eq: L()}$, we get that \[D_{n,1}\leq d_n\frac{g(n)}{n-1}\] and similarly for $D_{n,2}$. Thus we can set \begin{equation} \label{eq:Bn} B_n=\frac{g(n)}{n-1} \end{equation} \subsubsection{$B_{n,i}$} For $i<j<k$, \begin{equation} \label{eq: bothstay} d(a_j,a_k)-d(a_j,a_k)=0. \end{equation} For $j\leq i< k$ \begin{equation} \label{eq: onemoves} d \left( \frac{d_i}{d_j}a_j,a_k \right)-d \left(\frac{d_{i+1}}{d_j}a_j,a_k \right)\leq d_i-d_{i+1}. \end{equation} For $k<j\leq i$ \begin{equation} \label{eq: bothmove} d \left( \frac{d_i}{d_j}a_j,\frac{d_i}{d_j}a_k \right)-d \left( \frac{d_{i+1}}{d_j}a_j,\frac{d_{i+1}}{d_j}a_k \right)\leq (d_i-d_{i+1}) d\left( \frac{a_j}{d_j},\frac{a_k}{d_k} \right). \end{equation} Define also $f(i)$ as the maximum value of \begin{align*} d(u_1,u_2)+\sum_{j=3}^i\frac{3}{2}d(u_1,u_j)+\sum_{j=3}^i\frac{3}{2}d(u_2,u_j) +\sum_{3\leq j<k\leq i}2d(u_j,u_k) \end{align*} over unit vectors $u_1,\ldots,u_i$. We use equation \eqref{eq: Young} to obtain upper bounds on $f(i)$, which we then use to find numerical values for $B_{n,i}$. For $i>2$, substituting $\eqref{eq: bothstay}$, $\eqref{eq: onemoves}$, $\eqref{eq: bothmove}$, and $\eqref{eq: L()}$ into $\eqref{eq: defineDnik}$, we get that \begin{align*} D_{n,i,1} &\leq (d_i-d_{i+1})\frac{f(i)}{n-1}+ (d_i-d_{i+1})\left(\frac{1+3(n-i)+(i-2)+2(n-i)(i-2)}{n-1} \right)\\ D_{n,i,1} &\leq (d_i-d_{i+1})\left(\frac{f(i)}{n-1}+2\frac{(n-i)(i-1)}{n-1}+1 \right) \end{align*} and similarly for $D_{n,i,2}$. So we set \begin{equation} \label{eq:Bni} B_{n,i}=\frac{f(i)}{n-1}+2\frac{(n-i)(i-1)}{n-1}+1 \end{equation} For $i=2$, the same substition gives us $B_{n,2}=3$. For $i=1$, the same substition gives us $B_{n,1}=1.5$. If there do not exist four paths $Q_{11}, Q_{12}, Q_{21}, Q_{22}$, then the average length of a path is too great, namely \begin{align*} \frac{1}{2} \left( \overline{L(a_{n+1}, \ldots, a_1)}+\overline{L(a_{n+2},a_n,a_{n-1}, \ldots, a_1)} \right) &> r^*S_n\\ \frac{d_{n+1}+d_{n+2}}{2}B_n + \left(d_n-\frac{d_{n+1}+d_{n+2}}{2}\right)B_{n,n}+\sum_{i=1}^{n-1} \left(d_i-d_{i+1}\right)B_{n,i} &> \sum_{i=1}^{n} r^*d_n \end{align*} \subsection{$n=3$} If $d_4 \leq 0.541d_3$, then, by \eqref{eq: Young}, \[\overline{L \left( a_4, a_3, \frac{d_3}{d_2}a_2, \frac{d_3}{d_1}a_1 \right)}\leq 4.677d_3=3r^*d_3.\] Then, since $B_{3,2}=3<2r^*$ and $B_{3,1}=1.5<r^*$ as in the last section, \[\overline{L(a_4, a_3, a_2, a_1)}\leq r^*(d_1+d_2+d_3).\] Similarly, \[\overline{L(a_5, a_3, a_2, a_1)}\leq r^*(d_1+d_2+d_3)\] so the induction works. Thus for $n=3$ we have the constraint $d_4>0.541d_3$, which is stronger than the one obtained for $n=3$ in the previous section. \section{Choosing $n$} \label{sec: recombine} We obtained linear constraints for various values of $n\leq 10$. These, together with the constraints $d_i\geq d_{i+1}$, make a linear program (given in Appendix \ref{sec: lp}), which is unsatisfiable. Thus one of the constraints must not hold, so the induction works for some $n$. \section{Algorithm} We repeatedly use the inductive step to obtain a sequence of indices $0=n_0<n_1<n_2<~\ldots$. At stage $j$, we remove $n_j-n_{j-1}$ points. The intermediate ending points are then of the form $n_j+k_j$ where each $k_j$ is 1 or 2. Since we are only using $n\leq 10$, we can find all the paths $Q_{11}, Q_{12}, Q_{21}, Q_{22}$ by brute force in linear time. Now, for both possible values of $k_1$, we find which value of $k_0$ gives the shorter path. Then, for both possible values of $k_2$, we find which value of $k_1$ will make the total path after $a_{n_2+k_2}$ shorter. We repeat until we get to some $n_j>m$, at which point we have two paths and choose the shorter one. This whole algorithm is linear. \section{Lower bound on degree-3 tree ratios} \label{sec:lowerbound} Denote by $\sigma$ the sum of edge weights of the minimal spanning tree and by $\sigma_3$ the sum of edge weights of a minimal degree 3 tree. Denote by $(x_1,x_2,\ldots,x_n)$ the coordinates of a point in $n$ dimensions. In six dimensions, let $O$ be the origin and let $v_1,v_2,\ldots,v_7$ be the vertices of a simplex with center at $O$ and radius $\sqrt{6}$. Let the coordinates of $v_i$ be $(v_{i,1},v_{i,2},v_{i,3},v_{i,4},v_{i,5},v_{i,6})$. Note that $d(v_i,v_j)=\sqrt{2*7/6}\sqrt{6}=\sqrt{14}$. Now, given natural $N$ and $0<\alpha < 1$, take the following tree in $7N$ dimensions: \begin{enumerate} \item The origin, $O$, is the root. \item Its $N$ children are $p_1,p_2,\ldots,p_N$. $p_i$ has coordinates 0 except $x_{7i}=1-\alpha$. \item Each $p_i$ has seven children, $q_{i,1},q_{i,2},\ldots,q_{i,7}$ The coordinates of $q_{i,j}$ are all 0 except $x_{7i}=1$ and, for $k$ from 1 to 6, $x_{7i-k}=v_{j,k}$. \end{enumerate} Then $q_{i,1},q_{i,2},\ldots,q_{i,7}$ form a simplex with center distance $\alpha$ from $p_i$ and with each vertex distance $\sqrt{6}$ from the center. It is easy to check that \begin{align*} d(p_i,p_h)&=\sqrt{2}(1-\alpha) \text{ for } i\neq h\\ d(q_{i,j},q_{i,k})&=\sqrt{14}=d(q_{i,j},q_{h,k}) \text{ for } j\neq k,h\neq i\\ d(p_i,q_{i,j})&=\sqrt{6+\alpha^2}\\ \sigma&=N(1-\alpha+7\sqrt{6+\alpha^2}). \end{align*} Then we can pick \[\alpha=-1-\sqrt{7}+\sqrt{4+4\sqrt{7}},\] which gives us $d(q_{i,j},q_{h,k})+d(p_i,p_h)=2d(p_i,q_{i,j})$. Then we can define function $c$ on the vertices so that $c(O)=0, c(p_i)=d(p_i,p_h)/2$ and $c(q_{i,j})=d(q_{i,j},q_{h,k})/2$. In that case, the length of edge $AB$ is at least $c(A)+c(B)$, so $c$ can be thought of a half-edge length. Then, since there are $8N+1$ vertices, there are $8N$ edges, so there is a total of $16N$ edge endpoints. At most 3 of them contribute 0 to $\sigma_3$, at most $3N$ contribute $(1-\alpha)/\sqrt{2}$, and the remainder contribute $\sqrt{14}/2$. Thus \[\sigma_3\geq 3N\left(\frac{1}{2}\sqrt{2}(1-\alpha)\right)+(13N-3)\left(\frac{1}{2}\sqrt{14}\right)\] \[\frac{\sigma_3}{\sigma}=\frac{3N\left(\frac{1}{2}\sqrt{2}(1-\alpha)\right)+(13N-3)\left(\frac{1}{2}\sqrt{14}\right)}{N\left(1-\alpha+7\sqrt{6+\alpha^2}\right)}\] \[\lim_{N \to \infty}\frac{\sigma_3}{\sigma}=\frac{3\left(\frac{1}{2}\sqrt{2}(1-\alpha)\right)+13\left(\frac{1}{2}\sqrt{14}\right)}{1-\alpha+7\sqrt{6+\alpha^2}}\approx 1.4473\] Thus $\tau_3^\infty \geq 1.447$. \section{Acknowledgements} I thank Samir Khuller for suggesting that I work on this problem and Timothy Chan for improved notation and organization.

https://arxiv.org/abs/2205.12266

A simple upper bound for the perimeter of an ellipse

We propose a simple derivation of an upper bound for the perimeter of an ellipse. The procedure, which relies on the use of elliptic integrals, consists in introducing, via inequalities and convexity properties, specific integrals which can be calculated analytically.

\section{Introduction} The perimeter of an ellipse is the length of the continuous line forming its boundary. Computing accurate approximations to the perimeter of an ellipse has been a subject of interest for mathematicians for a long time \cite{Adlaj2012,Borwein1987,Bailey2016}. Many approaches have been used to estimate its value, such as approximation formulas (two formulas derived by Ramanujan are particularly famous \cite{Almkvist1988,Villarino} and we provide a non-exhaustive list in Appendix B) \cite{Hiriart,collect,Barnard2001}, infinite series (Maclaurin, Gauss-Kummer, Euler, etc.), hypergeometric functions \cite{Abbott2009,Chandrupatla2010} (see Appendix C), and of course (elliptic) integrals \cite{Abel1827,Abramowitz1972,Arfken1985,Prudnikov1986,Whittaker1990}. It is well known that the perimeter of an ellipse can be expressed exactly as a complete elliptic integral of the second kind. Approximate formulas can, of course, be obtained by truncating the series representations of exact formulas. The difficulty of calculating exactly the perimeter has aroused a particular interest in the search for bounds. A long time ago, Kepler found the geometric mean of semi-major and semi-minor axis, as a lower bound for the perimeter, and many sophisticated approaches were developed along the years (see for instance \cite{Pfiefer1988,Wang2012,Gusic2015}). The subject is still an active field of research \cite{He2020,Zhao2022}, important for astrophysical applications (trajectories of planets, satellites, \emph{etc.}). An ellipse with semi-major axis $a$ and semi-minor axis $b$ is parametrized by $x=a\cos\phi$ and $y=b\sin\phi$. The infinitesimal variation of curvilinear abscissa $s$ is $ds=\sqrt{dx^2+dy^2}$. The length of a quarter of the ellipse is provided by the elliptic integral \begin{equation}\label{def} L=\int_0^{\pi/2}ds=\int_0^{\pi/2}\sqrt{a^2\sin^2\phi+b^2\cos^2\phi}~d\phi. \end{equation} We will use the latter formula to derive the upper bound. The total perimeter of the ellipse will therefore be \begin{equation} L_e=4L. \end{equation} The well-known simple lower bound $2\pi\sqrt{ab}$ and upper bound $4a+\pi b$ are recalled in sections \ref{sec1} and \ref{sec2} respectively, and the main result of the present work, which is an upper bound more accurate than the one of section \ref{sec2}, but still rather simple, is presented in section \ref{sec3}. \section{The geometric-mean lower bound}\label{sec1} Let us consider two parametrizations of an ellipse \begin{equation} \begin{array}{l} x_1=a\cos\phi,\\ y_1=b\sin\phi \end{array} \end{equation} and \begin{equation} \begin{array}{l} x_2=b\cos\phi,\\ y_2=a\sin\phi. \end{array} \end{equation} The two associated lengths of the quarter of the ellipse are \begin{equation} L_1=\int_0^{\pi/2}\sqrt{a^2\cos^2\phi+b^2\sin^2\phi}~d\phi \end{equation} and \begin{equation} L_2=\int_0^{\pi/2}\sqrt{b^2\cos^2\phi+a^2\sin^2\phi}~d\phi. \end{equation} Since $t\rightarrow\sqrt{t}$ is a concave function, one has \begin{equation} \sqrt{\lambda t_1+(1-\lambda)t_2}\geq\lambda\sqrt{t_1}+(1-\lambda)\sqrt{t_2}, \end{equation} yielding ($\lambda=\cos^2\phi$, $t_1=a^2$ and $t_2=b^2$): \begin{equation} L_1\geq\int_0^{\pi/2}\left(a\cos^2\phi+b\sin^2\phi\right)d\phi \end{equation} and \begin{equation} L_2\geq\int_0^{\pi/2}\left(b\cos^2\phi+a\sin^2\phi\right)d\phi \end{equation} and therefore \begin{equation} L=L_1=L_2=\frac{1}{2}\left(L_1+L_2\right)\geq (a+b)\frac{\pi}{4}. \end{equation} Setting $ab=R^2$, the minimum of $a+b=a+R/a^2=f(a)$ is obtained for $R=\sqrt{ab}$ and is equal to $2R$. In other words, the ellipse which has the minimum perimeter for a given area is the circle of radius $R$. This yields the well-known result \begin{align} \frac{\pi}{2}\sqrt{ab}\leq L \end{align} and for the total perimeter of the ellipse \begin{equation} 2\pi\sqrt{ab}\leq L_e. \end{equation} \begin{figure} \centering \includegraphics[width=10cm]{fig23_ellipse.eps} \caption{quarters of ellipse with semi-major and semi-minor axis $(a=2,b=\sqrt{3})$ and $(a=2,b=\sqrt{3})$ respectively.} \label{fig23} \end{figure} \section{First upper bound}\label{sec2} Equation (\ref{def}) is equivalent to \begin{equation} L=a\int_0^{\pi/2}\sqrt{1-\frac{\left(a^2-b^2\right)}{a^2}\cos^2\phi}~d\phi \end{equation} which can be put in the form \begin{equation} L=a\frac{\pi}{2}\mathscr{F}\left(\frac{b^2-a^2}{a^2}\right)\cos^2\phi~ d\phi=a\frac{\pi}{2}\mathscr{F}\left(-e^2\right) \end{equation} where we have introduced the function \begin{equation} \mathscr{F}(t)=\frac{2}{\pi}\int_0^{\pi/2}\sqrt{1+t\cos^2\phi}~d\phi \end{equation} where $e$ represents the eccentricity of the ellipse defined by \begin{equation} \frac{b^2}{a^2}=1-e^2. \end{equation} The asymptotic behaviour of the function $\mathscr{F}$ is studied in Appendix A. If $-1<t<0$, then $-t/(1+t)>0$ and \begin{equation} \mathscr{F}(t)-\mathscr{F}(-1)=\frac{2}{\pi}\int_0^{\pi/2}\left(\sqrt{1+t\cos^2\phi}-\sqrt{1-\cos^2\phi}\right)d\phi. \end{equation} Since \begin{equation} \sqrt{u}-\sqrt{v}=\frac{u-v}{\sqrt{u}+\sqrt{v}}, \end{equation} we get \begin{equation}\label{inter} \mathscr{F}(t)-\mathscr{F}(-1)=\frac{2}{\pi}\int_0^{\pi/2}\frac{(1+t)\cos^2\phi}{\sqrt{1+t\cos^2\phi}+\sqrt{1-\cos^2\phi}}d\phi. \end{equation} and \begin{equation} \mathscr{F}(t)-\mathscr{F}(-1)\leq\frac{2}{\pi}(1+t)\int_0^{\pi/2}\frac{\cos^2\phi}{\sqrt{1+t}}d\phi, \end{equation} \emph{i.e.} \begin{equation} \mathscr{F}(t)-\mathscr{F}(-1)\leq\frac{2}{\pi}\sqrt{1+t}~\frac{\pi}{4}. \end{equation} because \begin{equation} \int_0^{\pi/2}\cos^2\phi~d\phi=\frac{\pi}{4}. \end{equation} This enables us to write \begin{equation} \mathscr{F}(t)-\mathscr{F}(-1)\leq\frac{\sqrt{1+t}}{2}. \end{equation} Since $-1<-e^2<0$, we get \begin{equation} \mathscr{F}(-e^2)-\mathscr{F}(-1)\leq\frac{\sqrt{1-e^2}}{2}, \end{equation} leading to \begin{equation}\label{firs} L\leq a+\frac{\pi}{4}b. \end{equation} and for the total perimeter of the ellipse \begin{empheq}[box=\fbox]{align}\label{firstup} L_e\leq 4a+\pi b. \end{empheq} \begin{figure} \centering \includegraphics[width=10cm]{fig1_ellipse.eps} \caption{Comparison between the first and second upper bounds of the perimeter of a quarter of an ellipse with semi-major axis $a=2$. On has $f_1(a,b)=a+\frac{\pi}{4}b$ (Eq. (\ref{firs})) and $f_2(a,b)=a+\frac{b^2}{a^2}\left[\sqrt{a^2-b^2}\ln\left(\frac{a+\sqrt{a^2-b^2}}{b}\right)+\frac{\pi}{2}b-a\right]$ (Eq. (\ref{seco})).} \label{fig1} \end{figure} \section{Second upper bound}\label{sec3} Let us go back to Eq. (\ref{inter}): \begin{equation} \mathscr{F}(t)-\mathscr{F}(-1)=\frac{2}{\pi}\int_0^{\pi/2}\frac{(1+t)\cos^2\phi}{\sqrt{1+t\cos^2\phi}+\sqrt{1-\cos^2\phi}}d\phi. \end{equation} We have \begin{equation} \mathscr{F}(t)-\mathscr{F}(-1)\leq\frac{2}{\pi}(1+t)\int_0^{\pi/2}\frac{\cos^2\phi}{\sqrt{1+t}+\sin\phi}d\phi, \end{equation} where we introduce, for $0<\xi<1$, the following integral \begin{equation} \mathscr{K}(\xi)=\int_0^{\pi/2}\frac{\cos^2\phi}{\xi+\sin\phi}d\phi. \end{equation} We obtain \begin{eqnarray} \mathscr{K}(\xi)&=&\int_0^{\pi/2}\frac{1-\sin^2\phi}{\xi+\sin\phi}d\phi=\int_0^{\pi/2}\frac{\left(\xi^2-\sin^2\phi\right)+\left(1-\xi^2\right)}{1+\sin\phi}d\phi\nonumber\\ &=&\int_0^{\pi/2}\left(\xi-\sin\phi\right)d\phi+\left(1-\xi^2\right)\int_0^{\pi/2}\frac{d\phi}{\xi+\sin\phi}\nonumber\\ &=&\xi\frac{\pi}{2}-1+\left(1-\xi^2\right)\mathscr{J}(\xi) \end{eqnarray} with \begin{equation} \mathscr{J}(\xi)=\int_0^{\pi/2}\frac{d\phi}{\xi+\sin\phi}. \end{equation} Setting $t=\tan(\phi/2)$, we have \begin{equation} dt=\frac{1}{2}\left[1+\tan^2\left(\frac{\phi}{2}\right)\right]d\phi, \end{equation} \emph{i.e.} $d\phi=2dt/(1+t^2)$. Then, since \begin{equation} \sin\phi=\frac{2t}{\left(1+t^2\right)}, \end{equation} we get \begin{equation} \frac{d\phi}{\xi+\sin\phi}=\frac{2dt}{\xi t^2+2t+\xi}. \end{equation} Writing \begin{equation} \frac{2}{\xi t^2+2t+\xi}=\frac{2}{\xi(t-\chi_1)(t-\chi_2)} \end{equation} with \begin{equation} \chi_1=\frac{-1-\sqrt{1-\xi^2}}{\lambda} \end{equation} and \begin{equation} \chi_2=\frac{-1+\sqrt{1-\xi^2}}{\lambda}, \end{equation} enables us to write \begin{eqnarray} \mathscr{J}(\xi)&=&\frac{2}{\xi}\int_0^1\frac{dt}{(t-\chi_1)(t-\chi_2)}=\frac{2}{\xi(\chi_1-\chi_2)}\left[\frac{1}{t-\chi_1}-\frac{1}{t-\chi_2}\right]dt\nonumber\\ &=&\frac{2}{\xi(\chi_1-\chi_2)}\left|\ln\left(\frac{t-\chi_1}{t-\chi_2}\right)\right|_0^1=\frac{2}{\xi(\chi_1-\chi_2)}\ln\left[\frac{(1-\chi_1)\chi_2}{(1-\chi_2)\chi_1}\right]. \end{eqnarray} We have also \begin{equation} \chi_1\chi_2=1 \end{equation} and \begin{equation} \chi_1-\chi_2=-\frac{2}{\xi}\sqrt{1-\xi^2} \end{equation} and thus \begin{eqnarray} \mathscr{J}(\xi)&=&-\frac{1}{\sqrt{1-\xi^2}}\ln\left(\frac{\chi_2-1}{\chi_1-1}\right)=\frac{1}{\sqrt{1-\xi^2}}\ln\left(\frac{\xi+1+\sqrt{1-\xi^2}}{\xi+1-\sqrt{1-\xi^2}}\right)\nonumber\\ &=&\frac{1}{\sqrt{1-\xi^2}}\ln\left(\frac{\sqrt{\xi+1}+\sqrt{1-\xi}}{\sqrt{\xi+1}-\sqrt{1-\xi}}\right)\nonumber\\ &=&\frac{1}{\sqrt{1-\xi^2}}\ln\left(\frac{1+\sqrt{1-\xi^2}}{\xi}\right), \end{eqnarray} yielding \begin{equation} \mathscr{K}(\xi)=\frac{\xi\pi}{2}-1+\sqrt{1-\xi^2}~\ln\left(\frac{1+\sqrt{1-\xi^2}}{\xi}\right). \end{equation} Therefore \begin{equation} L=\frac{\pi}{2}a\mathscr{F}\left(\frac{b^2}{a^2}-1\right)\leq\frac{\pi}{2}a\left[\mathscr{F}(-1)+\frac{2}{\pi}\left(\frac{b^2}{a^2}\right)\mathscr{K}\left(\frac{b}{a}\right)\right], \end{equation} or finally \begin{equation}\label{seco} L\leq a+\frac{b^2}{a^2}\left[\sqrt{a^2-b^2}\ln\left(\frac{a+\sqrt{a^2-b^2}}{b}\right)+\frac{\pi}{2}b-a\right] \end{equation} and for the total perimeter of the ellipse \begin{empheq}[box=\fbox]{align} L_e\leq 4a+4\frac{b^2}{a^2}\left[\sqrt{a^2-b^2}~\ln\left(\frac{a+\sqrt{a^2-b^2}}{b}\right)+\frac{\pi}{2}b-a\right]. \end{empheq} This upper bound is better than the first one in Eq. (\ref{firstup}) (see figure \ref{fig1}). \section{Conclusion} In this article, we obtained an upper bound for the perimeter of an ellipse. The calculation consists in simplifying elliptic integrals by applying inequalities and convexity properties. It is hoped that the method may stimulate the derivation of further bounds, improving the compromise between simplicity and accuracy.

https://arxiv.org/abs/1411.7405

A note relating ridge regression and OLS p-values to preconditioned sparse penalized regression

When the design matrix has orthonormal columns, "soft thresholding" the ordinary least squares (OLS) solution produces the Lasso solution [Tibshirani, 1996]. If one uses the Puffer preconditioned Lasso [Jia and Rohe, 2012], then this result generalizes from orthonormal designs to full rank designs (Theorem 1). Theorem 2 refines the Puffer preconditioner to make the Lasso select the same model as removing the elements of the OLS solution with the largest p-values. Using a generalized Puffer preconditioner, Theorem 3 relates ridge regression to the preconditioned Lasso; this result is for the high dimensional setting, p > n. Where the standard Lasso is akin to forward selection [Efron et al., 2004], Theorems 1, 2, and 3 suggest that the preconditioned Lasso is more akin to backward elimination. These results hold for sparse penalties beyond l1; for a broad class of sparse and non-convex techniques (e.g. SCAD and MC+), the results hold for all local minima.

\section{} \begin{abstract} When the design matrix has orthonormal columns, ``soft thresholding" the ordinary least squares (OLS) solution produces the Lasso solution \citep{tibshirani1996regression}. If one uses the Puffer preconditioned Lasso \citep{jia2012preconditioning}, then this result generalizes from orthonormal designs to full rank designs (Theorem 1). Theorem 2 refines the Puffer preconditioner to make the Lasso select the same model as removing the elements of the OLS solution with the largest p-values. Using a generalized Puffer preconditioner, Theorem 3 relates ridge regression to the preconditioned Lasso; this result is for the high dimensional setting, $p>n$. Where the standard Lasso is akin to forward selection \citep{efron2004least}, Theorems 1, 2, and 3 suggest that the preconditioned Lasso is more akin to backward elimination. These results hold for sparse penalties beyond $\ell_1$; for a broad class of sparse and non-convex techniques (e.g. SCAD and MC+), the results hold for all local minima. \end{abstract} \section{Introduction} Preconditioning is a classical computational technique in numerical linear algebra that creates fast algorithms. Several papers have recently proposed and studied the ``preconditioned" Lasso \citep{paul2008preconditioning, huang2011variable, rauhut2011sparse, jia2012preconditioning, qian2012pattern, wauthier2013comparative}. Instead of accelerating standard Lasso algorithms, preconditioning the Lasso creates a new statistical estimator that retains several properties of the Lasso while making the solution less sensitive to the correlation between the columns of the design matrix. This paper demonstrates how penalized least squares estimators with various forms of preconditioning are equivalent to classical quantities in linear regression--the OLS estimator, OLS p-values, and ridge regression. The theorems below do not make any assumptions on the design matrix beyond full rank. Nor do they assume a linear model $Y = \ensuremath{{\mathbf{X}}} \beta + \epsilon$, or assume some conditions on an error term $\epsilon$. Instead of studying the statistical estimation properties of preconditioned penalized least squares problems, the following theorems study the preconditioned Lasso estimators as functions of data $(\ensuremath{{\mathbf{X}}}, Y)$ that return a vector $\hat \beta$. The theorems compare these new functions to classical ``functions" like $ols$ and ridge regression. \subsection{Preliminaries} While the theorems below do not require the linear model, it is the linear model that motivates the estimators (i.e. functions) studied in this paper. The linear model is \begin{equation}\label{linearModel} Y = \ensuremath{{\mathbf{X}}} \beta + \epsilon, \end{equation} where $Y \in R^n$ and $\ensuremath{{\mathbf{X}}} \in R^{n \times p}$ are observed, and $\epsilon \in R^n$ is random noise satisfying $E(\epsilon) = 0$ and $E(\epsilon \epsilon') = \sigma^2 I_p$. The goal is to estimate $\beta \in R^p$ with $\ensuremath{{\mathbf{X}}}$ and $Y$. Throughout the paper, we will assume that $\ensuremath{{\mathbf{X}}}$ is full rank (when $n<p$, it is full row rank). Define $\|x\|_q = (\sum_i x_i^q)^{1/q}$. For $n>p$, define \[\hat \beta^{ols} = \arg \min_{b\in R^p} \| Y - \ensuremath{{\mathbf{X}}} b\|_2^2 = (\ensuremath{{\mathbf{X}}}'\ensuremath{{\mathbf{X}}})^{-1} \ensuremath{{\mathbf{X}}}' Y \in R^p\] as the standard OLS estimator. \begin{definition} Define the function $Lasso_\lambda$, \[Lasso_\lambda( \ensuremath{{\mathbf{X}}}, Y) = \arg \min_{b\in R^p} \| Y - \ensuremath{{\mathbf{X}}} b\|_2^2 + \lambda \|b\|_1.\] \end{definition} For $a\in R$, define $\textrm{sign}(a)$ as $-1, 0,$ or $1$, depending on whether $a$ is negative, zero, or positive. Define $(a)^+$ as equal to $a$ if $a$ is nonnegative and equal to zero if $a$ is negative. The soft-thresholding function is defined as \[t_\lambda(x) = sign(x) (|x| - \lambda)^+.\] To apply $t_\lambda$ to a vector $x$, apply it element-wise, $[t_\lambda(x)]_j = t_\lambda(x_j)$. \begin{lemma}\label{classicalRelation} (Equation 3 in \citep{tibshirani1996regression}) If the design matrix $\ensuremath{{\mathbf{X}}}$ is orthonormal, \begin{equation} \label{classical} Lasso_\lambda(\ensuremath{{\mathbf{X}}}, Y) = t_\lambda(\ensuremath{{\hat \beta^{ols}}}). \end{equation} \end{lemma} \section{Preconditioning the Lasso} Sections 2 and 3 study the low dimensional setting $n > p$. For these sections, let $\ensuremath{{\mathbf{X}}} = UDV'$ be the ``skinny" SVD; $U \in R^{n \times p}$ and $V\in R^{p \times p}$ have orthonormal columns and $D$ is a diagonal matrix. The Puffer transform is defined as $\ensuremath{F} = U D^{-1} U'$ \citep{jia2012preconditioning}. After preconditioning, Equation \eqref{linearModel} becomes \begin{equation} \label{precondModel} \ensuremath{F} Y = (\ensuremath{F} \ensuremath{{\mathbf{X}}}) \beta + \ensuremath{F} \epsilon. \end{equation} While $(\ensuremath{F} \ensuremath{{\mathbf{X}}}) =U V'$ is an orthonormal matrix, it is not orthogonalized by rotating the columns as in a QR decomposition; this would correspond to \textit{right} multiplying $\ensuremath{{\mathbf{X}}}$ by some matrix. Rotating the columns would create a new basis and make the Lasso penalize in the incorrect basis. By \textit{left} multiplying, each row of $\ensuremath{F} \ensuremath{{\mathbf{X}}}$ is a linear combination of the rows in $\ensuremath{{\mathbf{X}}}$. Importantly, the regression estimators that use $(\ensuremath{F} \ensuremath{{\mathbf{X}}}, \ensuremath{F} Y)$ instead of $(\ensuremath{{\mathbf{X}}}, Y)$ still estimate the same vector $\beta$ and the Lasso penalizes in the correct basis; the original regression model in Equation \eqref{linearModel} contains the exact same $\beta$ as Equation \eqref{precondModel}. Define \begin{eqnarray*} \textit{Puffer}(\ensuremath{{\mathbf{X}}}, Y) &=& (\ensuremath{F} \ensuremath{{\mathbf{X}}} , \ensuremath{F} Y) \in \ensuremath{{\mathbb{R}}}^{n \times p} \times \ensuremath{{\mathbb{R}}}^n \ \mbox{ and } \\ Lasso_\lambda(\textit{Puffer}(\ensuremath{{\mathbf{X}}}, Y)) &=& Lasso_\lambda(\ensuremath{F} \ensuremath{{\mathbf{X}}}, \ensuremath{F} Y). \end{eqnarray*} \cite{jia2012preconditioning} showed that if the smallest singular value of $\ensuremath{{\mathbf{X}}}$ is bounded from below, then with $p$ fixed and $n\rightarrow \infty$, the preconditioned Lasso, $Lasso_\lambda(\textit{Puffer}(\ensuremath{{\mathbf{X}}}, Y))$, is sign consistent. Importantly, the Puffer preconditioned Lasso does not require the Irrepresentable Condition from \cite{zhao2006model}. The next theorem shows how this estimator relates to the classical estimator $\hat \beta^{ols} \in R^p$, computed on the full model. This extends the relationship in Lemma \ref{classicalRelation} from the case where $\ensuremath{{\mathbf{X}}}$ is orthonormal, to the case where $\ensuremath{{\mathbf{X}}}$ is full rank. \begin{theorem}\label{theorem1} If $\ensuremath{{\mathbf{X}}}$ is full rank and $n>p$, then \begin{equation}\label{threshold} Lasso_\lambda(\textit{Puffer}(\ensuremath{{\mathbf{X}}}, Y)) = t_\lambda(\hat \beta^{ols} ). \end{equation} \end{theorem} In general, the relationship in Equation \eqref{threshold} does not hold without \textit{Puffer}. All proofs are contained in the Appendix. \section{Correcting for the heterogeneous variability in $\hat \beta^{ols}$} Under the linear model, \begin{equation}\label{olsSE} cov(\ensuremath{{\hat \beta^{ols}}}) = \sigma^2 (\ensuremath{{\mathbf{X}}}'\ensuremath{{\mathbf{X}}})^{-1}. \end{equation} Define \[\Sigma^{ols} = cov(\ensuremath{{\hat \beta^{ols}}}). \ \mbox{ Then, } variance(\ensuremath{{\hat \beta_j^{ols}}}) = \Sigma_{jj}^{ols}.\] If the diagonal elements of $\Sigma^{ols}$ are not all equal, then some elements of $\ensuremath{{\hat \beta^{ols}}}$ will have greater variability than others. Classical confidence intervals and p-values account for this uncertainty. However, the preconditioned Lasso estimator above (i.e. $t_\lambda( \ensuremath{{\hat \beta^{ols}}}) $) does not account for this known heteroskedasticity of $\ensuremath{{\hat \beta^{ols}}}$ by applying a stronger penalty to terms with larger variance. Classical versions of model selection test the following null hypotheses in various ways: \[H_{0,j}: \beta_j = 0, \ \mbox{ for } j \in 1, \dots, p.\] Under the linear model \eqref{linearModel} and $H_{0,j}$, \begin{equation}\label{teststat} Z_j = \sqrt{n} \frac{\hat \beta_j^{ols}}{\sqrt{\sigma^2 (\ensuremath{{\mathbf{X}}}'\ensuremath{{\mathbf{X}}})_{jj}^{-1}}} \stackrel{d}{\Rightarrow}N(0,1) \ \mbox{ as $n \rightarrow \infty$}. \end{equation} The classical ``marginal" p-values are defined as \begin{equation}\label{pvalue} p_j = 2(1 - \Phi(|Z_j|)), \mbox{ for $j = 1, \dots, p$ \end{equation} where $\Phi$ is the cdf of the standard normal distribution. \subsection{Scaling matters} Standard Lasso packages normalize the columns of $\ensuremath{{\mathbf{X}}}$ so that they all have equal $\ell_2$ length. This ensures that each element of the Lasso estimator is equally penalized. However, it does not ensure that the elements of the estimator are equally variable. This section \textit{right} preconditions $\ensuremath{{\mathbf{X}}}$ with a diagonal matrix $N$, making the column lengths heterogeneous. The column lengths are chosen so that the penalty strengths align with the heteroskedasticity of $\ensuremath{{\hat \beta^{ols}}}$, ensuring equal variability among the elements of the estimator $\hat \beta$. After left and right preconditioning with $\ensuremath{F}$ and $N$ respectively, the regression equation becomes \[\ensuremath{F} Y = (\ensuremath{F} \ensuremath{{\mathbf{X}}} N)(N^{-1}\beta) + \ensuremath{F} \epsilon.\] Define $\nu \in R^p$ as the diagonal elements of $(\ensuremath{{\mathbf{X}}}' \ensuremath{{\mathbf{X}}})^{-1}$ and let $N \in R^{ p \times p}$ be a diagonal matrix with $N_{jj} = \sqrt{\nu_j}. $ Because $\ensuremath{{\mathbf{X}}}$ is assumed full column rank and $N_{jj}$ is proportional to the standard error of $\ensuremath{{\hat \beta_j^{ols}}}$, $N_{jj}$ exists and is strictly positive. This ensures that $\textrm{sign}(N^{-1}\beta) = \textrm{sign}(\beta).$ So, inferences in the transformed space carry over to the original model. Take the SVD of $\ensuremath{{\mathbf{X}}} N = U_N D_N V_N'$. Then, define $\ensuremath{F}_N = U_N D_N^{-1} U_N'$ and \[\textit{Puffer}_N(\ensuremath{{\mathbf{X}}}, Y) = (\ensuremath{F}_N \ensuremath{{\mathbf{X}}} N, \ensuremath{F}_N Y) \in \ensuremath{{\mathbb{R}}}^{n \times p} \times \ensuremath{{\mathbb{R}}}^n.\] Theorem \ref{pvalueLasso} shows that if $\lambda = 1.96 \sigma/\sqrt{n}$, then $Lasso_\lambda(\textit{Puffer}_N(\ensuremath{{\mathbf{X}}}, Y))$ selects the same variables as fitting the standard OLS estimator and removing any variables with p-value greater than $.05$. \begin{theorem} \label{pvalueLasso} If $\ensuremath{{\mathbf{X}}}$ is full rank and $n>p$, denote \[\hat \beta^N(\lambda) = Lasso_\lambda(\textit{Puffer}_N(\ensuremath{{\mathbf{X}}}, Y)).\] Let $Z_j$ be the classical test statistic defined in Equation \eqref{teststat}, let $p_j$ be the classical p-value defined in Equation \eqref{pvalue}, and let $\Phi$ represent the cdf of the standard normal distribution. \[\hat \beta_j^N(\lambda) \ne 0 \ \ \ \Leftrightarrow \ \ \ |Z_j| > \lambda \sqrt{n} / \sigma \ \ \ \Leftrightarrow \ \ \ p_j \le 2\left(1-\Phi(\lambda \sqrt{n} /\sigma)\right).\] \end{theorem} Importantly, this is an algebraic equivalence between $Lasso_\lambda(\textit{Puffer}_N(\ensuremath{{\mathbf{X}}}, Y))$ and the classical OLS p-values in Equation \eqref{pvalue}. This requires only a full rank design matrix $\ensuremath{{\mathbf{X}}}$ and does not make any assumption on the error distribution or the elements of $\beta$. Theorem 2 asserts nothing about the statistical reliability of these p-values. If one wishes to use the p-value for statistical inference, then Theorem \ref{pvalueLasso} needs additional assumptions on the error term, $\epsilon$. \subsection{Generalizing to other methods} For simplicity, Theorems 1 and 2 are stated for the Lasso. However, both theorems hold more generally. For $\lambda \ge 0$ and penalty function $\ensuremath{\mbox{pen}}: \ensuremath{{\mathbb{R}}} \rightarrow \ensuremath{{\mathbb{R}}}_+$, define the function $sparse(\ensuremath{{\mathbf{X}}},Y,\lambda, \ensuremath{\mbox{pen}})$ as \begin{equation} sparse(\ensuremath{{\mathbf{X}}},Y,\lambda, \ensuremath{\mbox{pen}}) = \arg \min_b \frac{1}{2}\|Y - \ensuremath{{\mathbf{X}}} b\|_2^2 + \lambda \sum_j \ensuremath{\mbox{pen}}(b_j). \end{equation} Whenever $\ensuremath{\mbox{pen}}$ has a thresholding function $\tilde t_\lambda$ that satisfies a version of Lemma \ref{classicalRelation} with orthonormal designs, i.e. \[sparse(\ensuremath{{\mathbf{X}}}, Y,\lambda, \ensuremath{\mbox{pen}}) = \tilde t_\lambda(\ensuremath{{\hat \beta^{ols}}}) \ \mbox{ for orthonormal } \ensuremath{{\mathbf{X}}},\] then modified versions of Theorems 1 and 2 also apply for this penalty: \begin{eqnarray*} &sparse(\textit{Puffer}(\ensuremath{{\mathbf{X}}}, Y),\lambda, \ensuremath{\mbox{pen}}) &= \tilde t_\lambda(\ensuremath{{\hat \beta^{ols}}}) \\ &\left[ sparse(\textit{Puffer}_N(\ensuremath{{\mathbf{X}}}, Y),\lambda, \ensuremath{\mbox{pen}})\right]_j &= \tilde t_\lambda( \sigma Z_j / \sqrt{n} ). \end{eqnarray*} For example, $\ensuremath{\mbox{pen}}$ could be an $\ell_q$ penalty for $q \in [0,1]$, the elastic net penalty, or a concave penalty such as SCAD or MC+ \citep{fan2001variable, zhang2010nearly}. After preconditioning, a broad class of penalties select the same sequence of models as the Lasso. \section{Relating to ridge regression} For $p > n$, take the ``skinny" SVD of $\ensuremath{{\mathbf{X}}} = UDV'$, where $U \in R^{n \times n}, V \in R^{p \times n}, D \in R^{n \times n}$. Consider a generalized Puffer transformation \[\ensuremath{F}_\tau = U(D^2 + \tau I)^{-1/2}U' \ \mbox{ and } \textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y) = (\ensuremath{F}_\tau \ensuremath{{\mathbf{X}}}, \ensuremath{F}_\tau Y).\] Theorem \ref{mainresult} relates $sparse(\textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y), \lambda, \ensuremath{\mbox{pen}})$ to the ridge estimator \citep{hoerl1970ridge}. For $\tau > 0$, the ridge estimator is \begin{eqnarray} \ensuremath{{\ensuremath{\hat \beta_{ridge}(\tau) }}} &=& \arg \min_b \|Y - \ensuremath{{\mathbf{X}}} b\|_2^2 + \tau \|b\|_2^2 \label{ridge} \\ &=& (\ensuremath{{\mathbf{X}}}'\ensuremath{{\mathbf{X}}} + \tau I)^{-1} \ensuremath{{\mathbf{X}}}' Y. \nonumber \end{eqnarray} Define $\hat \beta_{ridge}(0)$ as the Moore-Penrose estimator \begin{eqnarray*} \hat \beta_{ridge}(0) &=&\arg \min_{b: \ensuremath{{\mathbf{X}}} b = Y} \|b\|_2^2 \\ &=& (\ensuremath{{\mathbf{X}}}'\ensuremath{{\mathbf{X}}})^{+} \ensuremath{{\mathbf{X}}}' Y, \end{eqnarray*} where $(\ensuremath{{\mathbf{X}}}'\ensuremath{{\mathbf{X}}})^{+}$ is the Moore-Penrose pseudoinverse, $VD^{-2}V'$.\footnote{In what follows, replace any matrix inversion with the Moore-Penrose pseudoinverse if $\tau =0$ and the true inverse does not exist.} Define $\ensuremath{\mathscr{P}_\tau}: \ensuremath{{\mathbb{R}}}^p \rightarrow \ensuremath{{\mathbb{R}}}^p$ as \begin{eqnarray} \ensuremath{\mathscr{P}_\tau}(v) &=& \ensuremath{{\mathbf{X}}}'(\ensuremath{{\mathbf{X}}}\X' + \tau I)^{-1} \ensuremath{{\mathbf{X}}} v. \label{proj} \end{eqnarray} For $\tau = 0$, $\ensuremath{\mathscr{P}_0}$ projects onto the row space of $\ensuremath{{\mathbf{X}}}$. Under the linear model (Equation \ref{linearModel}), $\ensuremath{{\ensuremath{\hat \beta_{ridge}(\tau) }}}$ is an unbiased estimator of $\ensuremath{\mathscr{P}_\tau}(\beta)$. The following theorem assumes that $sparse$ uses a penalty function satisfying the following assumptions. \begin{definition} \label{sparsepen} Define a function $\ensuremath{\mbox{pen}}$ as a \textbf{regular sparse penalty} if it is \begin{enumerate} \item non-differentiable at zero and differentiable everywhere else; \item symmetric, $\ensuremath{\mbox{pen}}(a) = \ensuremath{\mbox{pen}}(-a)$; \item monotonically increasing away from zero, $\ensuremath{\mbox{pen}}(a) \ge \ensuremath{\mbox{pen}}(b)$ if $|a| > |b|$; \item Lasso derivative in the neighborhood of zero, \[\lim_{x \rightarrow 0} |\ensuremath{\mbox{pen}}'(x)| = 1,\] where $\ensuremath{\mbox{pen}}'$ is the first derivative of $\ensuremath{\mbox{pen}}$. \end{enumerate} \end{definition} This includes the Lasso, elastic net, SCAD, and MC+. However, $\ell_q$ penalties fail condition (4) when $q < 1$. Such penalties have an unbounded derivative in the neighborhood of zero which creates discontinuities in the solution path; previous research has also excluded such penalties (e.g. \cite{zhang2012general, loh2013regularized}). Theorem \ref{mainresult} says that if $\ensuremath{\mbox{pen}}$ is a regular sparse penalty, then any local minimum for the objective function in $sparse(\textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y), \lambda, \ensuremath{\mbox{pen}} )$, transformed by $\ensuremath{\mathscr{P}_\tau}$, is close to the ridge estimator $\ensuremath{{\ensuremath{\hat \beta_{ridge}(\tau) }}}$. Moreover, $\lambda$ controls the distance between these estimators. \begin{theorem} \label{mainresult} Let $\ensuremath{\mbox{pen}}$ be a regular sparse penalty (Definition \ref{sparsepen}) and let $p\ge n$. Let \[\hat \beta = sparse(\textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y), \lambda, \ensuremath{\mbox{pen}} ).\] If $\hat \beta_j \ne 0$, then \[\ensuremath{{\ensuremath{\hat \beta_{ridge,j}(\tau) }}} - \ensuremath{\mathscr{P}_\tau}(\hat \beta)_j = \lambda \ensuremath{\mbox{pen}}'(\hat \beta_j),\] where $\ensuremath{\mbox{pen}}'$ is the derivative of $\ensuremath{\mbox{pen}}$. If $\hat \beta_j = 0$, then \[|\ensuremath{{\ensuremath{\hat \beta_{ridge,j}(\tau) }}} - \ensuremath{\mathscr{P}_\tau}(\hat \beta)_j| \le \lambda.\] Moreover, these results still hold if $\hat \beta$ is any local minimizer of the objective function for $sparse(\textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y), \lambda, \ensuremath{\mbox{pen}} )$. \end{theorem} Importantly, $sparse$ has been computed with the preconditioned data, $\textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y)$, while $\ensuremath{{\ensuremath{\hat \beta_{ridge}(\tau) }}}$ is the traditional estimator computed with the original data. When $\lambda$ is small, these two estimators are aligned in the row space of $\ensuremath{{\mathbf{X}}}$. Determining the statistically appropriate scale of $\lambda$ requires some care because after preconditioning, the scale of the problem changes; $\|\ensuremath{{\mathbf{X}}}\|_F^2 = O(np)$, but $\|\ensuremath{F}_\tau \ensuremath{{\mathbf{X}}}\|_F^2 = O(n)$. A forthcoming revision to \cite{jia2012preconditioning}, shows that $sparse(\textit{Puffer}_0(\ensuremath{{\mathbf{X}}}, Y), \lambda, \| \cdot \|_1)$ is sign consistent when $\min_j \beta_j$ is larger than $\log n / \sqrt{n}$ and \[\lambda = O\left(\left(\frac{\log n \log p}{p} \right)^{1/2}\right).\] In the high dimensional setting, this $\lambda$ is clearly converging to zero. If Theorem 3 held without $\ensuremath{\mathscr{P}_\tau}$, then it would say that $Lasso_\lambda(\textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y)) = t_\lambda(\ensuremath{{\ensuremath{\hat \beta_{ridge}(\tau) }}})$, which would make a clear analogy to backward elimination. However, the inclusion of $\ensuremath{\mathscr{P}_\tau}$ stains the analogy to backward elimination. In fact, it is not even true that $\ensuremath{\mathscr{P}_\tau}\left(Lasso_\lambda(\textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y))\right) = t_\lambda(\ensuremath{{\ensuremath{\hat \beta_{ridge}(\tau) }}})$; the left side is in the row space of $\ensuremath{{\mathbf{X}}}$, but the right side is not. That said, typical algorithms to compute the Lasso solution path (e.g. \cite{efron2004least}) start at $\lambda =\infty$ with $Lasso_{\lambda= \infty}(\ensuremath{{\mathbf{X}}}, Y) = 0$ and decrease $\lambda$, increasing the number of terms in the model. This resembles forward selection. The results of Theorem 3 are ``backwards" in the sense that for $\lambda \rightarrow 0$, $\ensuremath{\mathscr{P}_\tau}(Lasso_\lambda(\textit{Puffer}_\tau(\ensuremath{{\mathbf{X}}}, Y))) \rightarrow \ensuremath{{\ensuremath{\hat \beta_{ridge}(\tau) }}}$. Suppose $\hat \beta^{(1)}$ and $\hat \beta^{(2)}$ are both local minima of $sparse(\textit{Puffer}_0(\ensuremath{{\mathbf{X}}}, Y),\lambda, \ensuremath{\mbox{pen}})$ for a regular sparse penalty $\ensuremath{\mbox{pen}}(x)$ that is concave on $\{x:x>0\}$. By concavity and the definition of regular sparse penalty, it follows that $|\ensuremath{\mbox{pen}}'(x)| \le 1$ for $x\ne0$. So, the triangle inequality around $\hat \beta_{ridge,j}(0)$ yields \begin{equation}\label{closeMins} \ensuremath{\mathscr{P}_0}( \hat \beta^{(1)}_j - \hat \beta^{(2)}_j ) \le 2 \lambda. \end{equation} So after preconditioning, local minima are exceedingly similar in the row space of $\ensuremath{{\mathbf{X}}}$. Moreover, even if $\hat \beta^{(1)}$ and $\hat \beta^{(2)}$ are (local) minima from different penalty functions, then Equation \ref{closeMins} holds so long as (i) both $\ensuremath{\mbox{pen}}$ functions are regular and concave and (ii) both are computed with the same tuning parameter $\lambda$.\footnote{If they have different tuning parameters $\lambda_1$ and $\lambda_2$, then the bound $2 \lambda$ is replaced by $\lambda_1 + \lambda_2$.} In general, $\beta$ is not identifiable when $p>n$; only the projection of $\beta$ into the row space of $\ensuremath{{\mathbf{X}}}$ is identifiable \citep{shao2012estimation}. This suggests that, after preconditioning, it is difficult statistically distinguish the difference between local minima or the difference between penalty functions. \section{Discussion} Several previous papers have studied different preconditioners for the Lasso \citep{paul2008preconditioning, huang2011variable, rauhut2011sparse, jia2012preconditioning, qian2012pattern, wauthier2013comparative}. This paper connects two types of preconditioning techniques to the classical OLS solution. When performing model selection in the classical setting of $n>>p$, the marginal p-values from OLS are typically considered more informative than the absolute sizes of the elements in $\hat \beta^{ols}$; the p-values account for the potentially heterogeneous standard errors across $\hat \beta^{ols}$. This suggests that $Lasso_\lambda(\textit{Puffer}_N(\ensuremath{{\mathbf{X}}}, Y))$ should be preferred to $Lasso_\lambda(\textit{Puffer}(\ensuremath{{\mathbf{X}}}, Y))$. However, classical intuitions also suggest that two variables might have statistically insignificant p-values because these variables are correlated with each other(see Section 10.1 in \cite{weisberg2014applied}). As such, a backward procedure should have multiple steps. On each step, remove the variable with the largest p-value and refit the OLS with the remaining predictors. Neither of the preconditioning techniques above creates a Lasso solution path that is equivalent to this multi-step approach. A preconditioner that is designed to match this path must be a function of $Y$ and $\lambda$, thus becoming much more complicated. The Lasso (without preconditioning) is akin to forward selection \citep{efron2004least}. Classical methods of forward selection are not model selection consistent, unless the columns of the design matrix are only weakly correlated \citep{tropp2007signal}. Similarly, the Lasso is not sign consistent unless the design matrix satisfies the Irrepresentable Condition \citep{zhao2006model}. After preconditioning with the (generalized) Puffer transformation, Theorems 1, 2, and 3 show that the Lasso is akin to backward elimination. This preconditioner makes the design orthogonal when $n>p$, trivially satisfying all consistency conditions (e.g. Irrepresentable Condition, RIP, etc.). As such, Theorem 1 implies that a one step backward elimination is also sign consistent. This suggests that backward procedures are sign consistent when forward procedures are not. However, backward procedures are not a panacea. They will become unstable whenever the OLS p-values are incorrect (e.g. when $p$ is large compared to $n$ and the CLT does not hold for the test statistic $Z_j$ in Equation \eqref{teststat}). \bibliographystyle{plainnat}

https://arxiv.org/abs/1507.05952

Optimal Testing for Properties of Distributions

Given samples from an unknown distribution $p$, is it possible to distinguish whether $p$ belongs to some class of distributions $\mathcal{C}$ versus $p$ being far from every distribution in $\mathcal{C}$? This fundamental question has received tremendous attention in statistics, focusing primarily on asymptotic analysis, and more recently in information theory and theoretical computer science, where the emphasis has been on small sample size and computational complexity. Nevertheless, even for basic properties of distributions such as monotonicity, log-concavity, unimodality, independence, and monotone-hazard rate, the optimal sample complexity is unknown.We provide a general approach via which we obtain sample-optimal and computationally efficient testers for all these distribution families. At the core of our approach is an algorithm which solves the following problem: Given samples from an unknown distribution $p$, and a known distribution $q$, are $p$ and $q$ close in $\chi^2$-distance, or far in total variation distance?The optimality of our testers is established by providing matching lower bounds with respect to both $n$ and $\varepsilon$. Finally, a necessary building block for our testers and an important byproduct of our work are the first known computationally efficient proper learners for discrete log-concave and monotone hazard rate distributions.

\section{Experimental Evaluation} In this section, we evaluate the efficacy of our proposed monotonicity tester for $d=1$, and compare it to the estimator proposed in~\cite{BatuKR04}, which we will refer to as the BKR tester henceforth. We note that even for $d=1$, our algorithm is seen to outperform the BKR tester in almost all cases. Both our algorithm and BKR require a threshold for the test statistic. It follows from our results that our estimator's threshold should be at least the standard deviation of the statistic when the distribution is monotone -- this is around $\sqrt{2n}$. Also, our threshold should be smaller than the expectation of the statistic when the distribution is not in the class. In particular, we use the value $2m\varepsilon^2+ \sqrt{2n}$ as the threshold for our test statistic. The statistic of BKR imitates an $\ell_2$ test by introducing a discrete statistic that counts the number of collisions within nearby elements. The argument is that monotone distributions will be locally uniform, and therefore the number of such local collisions should be small. They also compare their algorithm with a threshold that is proportional to $O(\log^2 n/\varepsilon)$. However, the constant is not specified in their paper. We considered a few Zipf distributions, which are monotone, and perturbed the probability elements to generate distributions that are $\varepsilon$-far from monotone distributions in total variation distance. We then optimized their threshold for the best possible performance for a range of $n$ and $\varepsilon$. In particular, we came up with the constant of 0.6 for their statistic. Note that one might use uniform distributions, or yet another class of distributions to choose the threshold. As is expected, when we run experiments on the class that was used to determine the threshold, the performance of BKR is comparable with our approach for moderate values of $\varepsilon$. For very small $\varepsilon$, once again their performance degrades, since the number of samples required for concentration of their statistic is proportional to $\varepsilon^{-6}$. However, for experiments on other monotone classes of distributions (i.e., classes which were not used to train the threshold), our algorithm performed significantly better in almost all cases. In particular, we also consider a test between the uniform distribution and perturbations thereof (which are guaranteed to be $\varepsilon$-far from monotone). We give two illustrative plots in the appendix. The number of samples are plotted in logarithmic scale. Both these experiments are run for $n=50,000$. In the uniform case, we take the total variation bound to be $0.05$ and for the Zipf distribution we take it to be $0.07$. As we see for the uniform distribution, our algorithm has accuracy 0.9 with 30000 samples, compared to their accuracy of 0.55. For the Zipf plot, their algorithm is comparable, since we chose the threshold to work the best for this class. \section{Learning product distributions in $\chi^2$ distance} \label{sec:independence-appendix} In this section we prove Lemma~\ref{lem:learning-product}. The proof is analogous to the proof for learning monotone distributions, and hinges on the following result of~\cite{KamathOPS15}. Given $m$ samples from a distribution $q$ over $n$ elements, the add-1 estimator (Laplace estimator) $q$ satisfies: \[ \expectation{\chi^2(p,q) } \le \frac{n}{m+1}. \] Now, suppose $p$ is a product distribution over ${\cal X}=[n_1]\times\cdots\times[n_d]$. We simply perform the add-1 estimation over each coordinate independently, giving a distribution $q^1\times\cdots \times q^d$. Since $p$ is a product distribution the estimates in each coordinate is independent. Therefore, a simple application of the previous result and independence of the coordinates implies \begin{align} \expectation{\chi^2(p,q) } & = \prod_{l=1}^{d}\left(1+ \expectation{\chi^2(p^l,q^l) }\right)-1\nonumber\\ & \le \prod_{l=1}^{d}\left(1+ \frac{n_l}{m+1}\right)-1\nonumber\\ &\le \exp\left(\frac{\sum_l {n_l}}{m+1}\right)-1,\label{eqn:haha} \end{align} where~\eqref{eqn:haha} follows from $e^x\ge1+x$. Using $e^x\le 1+2x$ for $0\le x\le 1$, we have \begin{align} \expectation{\chi^2(p,q) } & \leq 2\frac{\sum_l {n_l}}{m+1}, \end{align} when $m\geq\sum_l n_l$. Therefore, following an application of Markov's inequality, when $m = \Omega((\sum_l n_l)/\varepsilon^2)$, Lemma~\ref{lem:learning-product} is proved. \section{Testing Independence of Random Variables} \label{sec:independence} Let ${\cal X}{\hfill\blksquare}\medskip[n_1]\times\ldots\times[n_d]$, and let $\Pi_d$ be the class of all product distributions over ${\cal X}$. We first bound the $\chi^2$-distance between product distributions in terms of the individual coordinates. \begin{lemma} \label{lem:prod-chi} Let $p = p^1\times p^2\ldots\times p^d$, and $q = q^1\times q^2\ldots\times q^d$ be two distributions in $\Pi_d$. Then \[ \chisqpq{p}{q} = \prod_{\gnote{\ell}=1}^d(1+\chisqpq{p\gnote{^\ell}}{q\gnote{^\ell}})-1. \] \end{lemma} \begin{proof} By the definition of $\chi^2$-distance \begin{align} \chisqpq{p}{q} &= \sum_{{\bf i}\in{\cal X}} \frac{{\dP}_{{\bf i}}^2}{{\dQ}_{{\bf i}}}-1\\ & = \prod_{\gnote{\ell}=1}^d\left[\sum_{i \in[n_\gnote{\ell}]}\frac{\left(p^{\gnote{\ell}}_{i}\right)^2}{q^{\gnote{\ell}}_{i}}\right]-1\\ & = \prod_{\gnote{\ell = 1}}^\gnote{d}\left(1+\chisqpq{p^\gnote{\ell}}{q^\gnote{\ell}}\right)-1. \end{align} \end{proof} Along the lines of learning monotone distributions in $\chi^2$ distance we obtain the following result, proved in Section~\ref{sec:independence-appendix}. \begin{lemma} \label{lem:learning-product} There is an algorithm that takes \[ O\left(\sum_{\ell=1}^d \frac{n_\ell}{\varepsilon^2}\right) \] samples from a distribution $p$ in $\Pi_d$ and outputs a distribution $q\in\Pi_d$ such that with probability at least $5/6$, \[ \chisqpq{p}{q}\le O(\varepsilon^2). \] \end{lemma} This fits precisely in our framework of robust $\chi^2$-$\ell_1$ testing. In particular, applying Theorem~\ref{thm:chisq-test}, we obtain the following result. \begin{theorem} \label{thm:independence} For any $d\ge1$, there exists an algorithm for testing independence of random variables over $[n_1]\times\ldots[n_d]$ with sample and time complexity \[ O\left(\frac{(\prod_{\ell = 1}^dn_\ell)^{1/2} + \sum_{\ell = 1}^d n_\ell}{\varepsilon^2}\right). \] \end{theorem} The following corollaries are immediate. \begin{corollary} Suppose $\prod_{\ell = 1}^d n_\ell^{1/2} \geq \sum_{\ell =1}^d n_\ell$. Then there exists an algorithm for testing independence over $[n_1] \times \dots \times [n_d]$ with sample complexity $\Theta((\prod_{\ell = 1}^d n_\ell)^{1/2}/\varepsilon^2)$. \end{corollary} In particular, \begin{corollary} There exists an algorithm for testing if two distributions over $[n]$ are independent with sample complexity $\Theta(n/\varepsilon^2)$. \end{corollary} \section{Details of the Lower Bounds} \label{sec:lb-appendix} In this section, for the class of distributions $\mathcal{Q}$ described in discussion on lower bounds and a class of interest $\mathcal{C}$, we show that $d_{\mathrm {TV}}(\mathcal{C},\mathcal{Q}) \geq \varepsilon$, thus implying a lower bound of $\Omega(\sqrt{n}/\varepsilon^2)$ for testing $\mathcal{C}$. \subsection{Monotone distributions} We first consider $d=1$ and prove that for appropriately chosen $c$, any monotone distribution over $[n]$ is $\varepsilon$-far from all distributions in $\mathcal{Q}$. Consider any $q \in \mathcal{Q}$. For this distribution, we say that $i\in[n]$ is a \emph{raise-point} if $\dQ_i<\dQ_{i+1}$. Let $R_{q}$ be the set of raise points of $q$. For $q\in\mathcal{Q}$,~\eqref{eqn:fcl} implies at least one in every four consecutive integers in $[n]$ is a raise point, and therefore, $|R_q|\gen/4$. Moreover, note that if $i$ is a raise-point, then $i+1$ is not a raise point. For any monotone (decreasing) distribution $p$, $\dP_i\ge\dP_{i+1}$. For any raise-point $i\in R_q$, by the triangle inequality, \begin{align} \label{eqn:triangle-mon} |\dP_i-\dQ_i|+|\dP_{i+1}-\dQ_{i+1}|\ge |\dP_i-\dP_{i+1}+\dQ_{i+1}-\dQ_i|\ge \dQ_{i+1}-\dQ_i = \frac{2c\varepsilon}{n}. \end{align} Summing over the set $R_q$, we obtain $d_{\mathrm {TV}}(p,q)\ge \frac12|R_q|\cdot \frac{2c\varepsilon}{n}\ge {c\varepsilon}/{4}$. Therefore, if $c\ge4$, then $d_{\mathrm {TV}}(\mathcal{M}_{n},q)\ge\varepsilon$. This proves the lower bound for $d=1$. This argument can be extended to $[n]^d$. Consider the following class of distributions on $[n]^d$. For each point ${\bf i}=(i_1, \ldots, i_d)\in[n]^d$, where $i_1$ is even, generate a random $z\in\{-1,1\}$, and assign to ${\bf i}$ a probability of $(1+z c\varepsilon)/n^{d}$. Let ${\bf e}_1{\hfill\blksquare}\medskip(1,0,\ldots,0)$. Similar to $d=1$, assign a probability $(1-z c\varepsilon)/n^{d}$ to the point ${\bf i}+{\bf e}_1 = (i_1+1,i_2, \ldots, i_d)$. This class consists of $2^{\frac{n^{d/2}}2}$ distributions, and Paninski's arguments extend to give a lower bound of $\Omega(n^{d/2}/\varepsilon^2)$ samples to distinguish this class from the uniform distribution over $[n]^d$. It remains to show that all these distributions are $\varepsilon$ far from $\mathcal{M}_n^d$. Call a point ${\bf i}$ as a raise point if $p_{\bf i}<p_{{\bf i}+{\bf e}_1}$. For any ${\bf i}$, one of the points ${\bf i}$, ${\bf i}+{\bf e}_1$, ${\bf i}+2{\bf e}_1$, ${\bf i}+3{\bf e}_1$ is a raise point, and the number of raise points is at least $n^{d}/4$. Invoking the triangle inequality (identical to~\eqref{eqn:triangle-mon}) over the raise-points, in the first dimension shows that any monotone distribution over $[n]^d$ is at a distance $\frac{c\varepsilon}4$ from any distribution in this class. Choosing $c=4$ yields a bound of $\varepsilon$. \subsection{Testing Product Distributions} Our idea for testing independence is similar to the previous section. We sketch the construction of a class of distributions on ${\cal X}=[n_1]\times\cdots\times[n_d]$. Then $|{\cal X}| = n_1\cdot n_2\ldots\cdot n_d$. For each element in ${\cal X}$ assign a value $(1\pm c\varepsilon)$ and then for each such assignment, normalize the values so that they add to 1, giving rise to a distribution. This gives us a class of $2^{|{\cal X}|}$ distributions. The key argument is to show that a \emph{large} fraction of these distributions are far from being a product distribution. This follows since the degrees of freedom of a product distribution is exponentially smaller than the number of possible distributions. The second step is to simply apply Paninski's argument, now over the larger set of distributions, where we show that distinguishing the collection of distributions we constructed from the uniform distribution over ${\cal X}$ (which is a product distribution) requires $\sqrt{|{\cal X}|}/\varepsilon^2$ samples. \subsection{Log-concave and Unimodal distributions} We will show that any log-concave or unimodal distribution is $\varepsilon$-far from all distributions in $\mathcal{Q}$. Since $\mathcal{LCD}_n \subset \mathcal{U}_n$, it will suffice to show this for every unimodal distribution. Consider any unimodal distribution $p$, with mode $\ell$. Then, $p$ is monotone non-decreasing over the interval $[\ell]$ and non-increasing over $\{\ell+1, \ldots, n\}$. By the argument for monotone distributions, the total variation distance between $p$ and any distribution $q$ over elements greater than $\ell$ is at least $\frac{n-\ell-1}{n}\frac{c\varepsilon}{4}$, and over elements less than $\ell$ is at least $\frac{\ell-1}{n}\frac{c\varepsilon}{4}$. Summing these two gives the desired bound. \subsection{Monotone Hazard distributions} We will show that any monotone hazard rate distribution is $\varepsilon$-far from all distributions in $\mathcal{Q}$. Let $p$ be any monotone-hazard distribution. Any distribution $q\in\mathcal{Q}$ has mass at least $1/2$ over the interval $I=[n/4,3n/4]$. Therefore, by Lemma~\ref{lem:mhr-str}, for any $i\in I$, $\dP_{i+1}\left(1+\frac{\dP_i}{1/4}\right)\ge \dP_i$. As noted before, at least $n/8$ of the raise-points are in $I$. For any $i\in I\cap R_q$, $\dQ_i=(1+c\varepsilon)/n$, $\dQ_{i+1}=(1-c\varepsilon)/n$ \begin{align} d_i = |\dP_i-\dQ_i|+|\dP_{i+1}-\dQ_{i+1}|. \end{align} If $\dP_i\ge(1+2c\varepsilon)/n$ or $\dP_i\le1/n$, then the first term, and therefore $d_i$ is at least $c\varepsilon/n$. If $\dP_i\in(1/n, (1+2c\varepsilon)/n)$, then for $n>5/(c\varepsilon)$ \[ \dP_{i+1}\ge \frac{1}{n}\cdot \frac{1}{1+\frac4{n}}\ge \frac{1-c\varepsilon/2}{n}. \] Therefore the second term of $d_i$ is at $c\varepsilon/2n$. Since there are at least $n/8$ raise points in $I$, \begin{align} d_{\mathrm {TV}}(p,q)\ge \frac12\frac{n}{8}\cdot \frac{c\varepsilon}{2n}\ge \frac{c\varepsilon}{16}. \end{align} Thus any MHR distribution is $\varepsilon$-far from $\mathcal{Q}$ for $c\ge16$. \section{Details on Testing Log-Concavity} \label{sec:LCD} It will suffice to prove Lemma~\ref{lem:lcd-learn}. \begin{prevproof}{Lemma}{lem:lcd-learn} We first draw samples from $p$ and obtain a $O(1/\varepsilon^{3/2})$-piecewise constant distribution $f$ by appropriately flattening the empirical distribution. The proof is now in two parts. In the first part, we show that if $p \in \mathcal{LCD}_n$ then $f$ will be close to $p$ in $\chi^2$ distance over its effective support. The second part involves proper learning of $p$. We will use a linear program on $f$ to find a distribution $q\in\mathcal{LCD}_n$. This distribution is such that if $p\in\mathcal{LCD}_n$, then $\chi^2(p,q)$ is small, and otherwise the algorithm will either output some $q \in \mathcal{LCD}_n$ (with no other relevant guarantees) or \textsc{Reject}\xspace. We first construct $f$. Let $\hat p$ be the empirical distribution obtained by sampling $O(1/\varepsilon^5)$ samples from $p$. By Lemma~\ref{lem:dkw}, with probability at least $5/6$, $d_{\mathrm K}(p,\hat p) \leq \varepsilon^{5/2}/10$. In particular, note that $|p_i - \hat p_i| \leq \varepsilon^{5/2}/10$. Condition on this event in the remainder of the proof. Let $a$ be the minimum $i$ such that $p_i \geq \varepsilon^{3/2}/5$, and let $b$ be the maximum $i$ satisfying the same condition. Let $M = \{a, \dots, b\}$ or $\emptyset$ if $a$ and $b$ are undefined. By the guarantee provided by the DKW inequality, $p_i \geq \varepsilon^{3/2}/10$ for all $i \in M$. Furthermore, $\hat p_i \in p_i \pm \varepsilon^{3/2}/10 \in (1 \pm \varepsilon) \cdot p_i$. For each $i \in M$, let $f_i = \hat p_i$. We note that $|M| = O(1/\varepsilon)$, so this contributes $O(1/\varepsilon)$ constant pieces to $f$. We now divide the rest of the domain into $t$ intervals, all but constantly many of measure $\Theta(\varepsilon^{3/2})$ (under $p$). This is done via the following iterative procedure. As a base case, set $r_0 = 0$. Define $I_j$ as $[l_j, r_j]$, where $l_j = r_{j-1} + 1$ and $r_j$ is the largest $j \in [n]$ such that $\hat p(I_j) \leq 9\varepsilon^{3/2}/10$. The exception is if $I_j$ would intersect $M$ -- in this case, we ``skip'' $M$: set $r_j = a-1$ and $l_{j+1} = b+1$. If such a $j$ exists, denote it by $j^*$. We note that $p(I_j) \leq \hat p(I_j) + \varepsilon^{5/2}/10 \leq \varepsilon^{3/2}$. Furthermore, for all $j$ except $j^*$ and $t$, $r_j + 1 \not \in M$, so $p(I_j) \geq 9\varepsilon^{3/2}/10 - \varepsilon^{3/2}/5 - \varepsilon^{5/2}/10 \geq 3\varepsilon^{3/2}/5$. Observe that this lower bound implies that $t \leq \frac{2}{\varepsilon^{3/2}}$ for $\varepsilon$ sufficiently small. \paragraph{Part 1.} For this part of the algorithm, we only care about the guarantees when $p \in \mathcal{LCD}_n$, so we assume this is the case. For the domain $[n] \setminus M$, we let $f$ be the flattening of $\hat p$ over the intervals $I_1, \dots I_t$. To analyze $f$, we need a structural property of log-concave distributions due to Chan, Diakonikolas, Servedio, and Sun \cite{ChanDSS13a}. This essentially states that a log-concave distribution cannot have a sudden increase in probability. \begin{lemma}[Lemma 4.1 in \cite{ChanDSS13a}] \label{lem:lcd-flat} Let $p$ be a distribution over $[n]$ that is non-decreasing and log-concave on $[1,x] \subseteq [n]$. Let $I = [x,y]$ be an interval of mass $P(I) = \tau$, and suppose that the interval $J = [1,x-1]$ has mass $p(J) = \sigma > 0$. Then $$p(y)/p(x) \leq 1 + \tau/\sigma.$$ \end{lemma} Recall that any log-concave distribution is unimodal, and suppose the mode of $p$ is at $i_0$. We will first focus on the intervals $I_1, \dots, I_{t_L}$ which lie entirely to the left of $i_0$ and $M$. We will refer to $I_j$ as $L_j$ for all $j \leq t_L$. Note that $p$ is non-decreasing over these intervals. The next steps to the analysis are as follows. First we show that the flattening of $p$ over $L_j$ is a multiplicative $(1 + O(1/j))$ estimate for each $p_i \in L_j$. Then, we show that flattening the empirical distribution $\hat p$ over $L_j$ is a multiplicative $(1 + O(1/j))$ estimate of $p(i)$ for each $i \in L_j$. Finally, we exclude a small number of intervals (those corresponding to $O(\varepsilon)$ mass at the left and right side of the domain, as well as $j^*$) in order to get the $\chi^2$ approximation we desire on an effective support. \begin{itemize} \item First, recall that $p(L_j) \leq \varepsilon^{3/2}$ for all $j$. Also, letting $J_j = [1, r_{j-1}]$, we have that $p(J_j) \geq (j-1)\cdot 3\varepsilon^{3/2}/5$. Thus by Lemma \ref{lem:lcd-flat}, $p(r_j) \leq p(l_j) (1 + 2/(j-1))$. Since the distribution is non-decreasing in $L_j$, the flattening $\bar p$ of $p$ is such that $\bar p(i) \in p(i) (1 \pm \frac{2}{j-1})$ for all $i \in L_j$. \item We have that $p(L_j) \geq 3\varepsilon^{3/2}/5$, and $\hat p(L_j) \in p(L_j) \pm \varepsilon^{5/2}/10$, so $\hat p(L_j) \in p(L_j) \cdot (1 \pm \frac{\varepsilon}{6})$, and hence $\hat p(i) \in \bar p(i) \cdot (1 \pm \frac{\varepsilon}{6})$ for all $i \in L_j$. Combining with the previous point, we have that $$\hat p(i) \in p(i) \cdot \left(1 \pm \left(\frac{2\varepsilon}{3(j-1)} + \frac{\varepsilon}{6} + \frac{2}{j-1}\right)\right) \in p(i) \cdot \left(1 \pm \frac{11}{3(j-1)}\right).$$ \end{itemize} A symmetric statement holds for the intervals that lie entirely to the right of $i_0$ and $M$. We will refer to $I_j$ as $R_{t-j}$ for all $j > t_L$. To summarize, we have the following guarantees for the distribution $f$: \begin{itemize} \item For all $i \in M$, $f(i) \in p(i) \cdot (1 \pm \varepsilon)$; \item For all $i \in L_j$ (except $L_1$ and $L_{j^*}$), $f(i) \in p(i) \cdot \left(1 \pm \frac{22}{3j}\right)$; \item For all $i \in R_j$ (except $R_1$), $f(i) \in p(i) \cdot \left(1 \pm \frac{22}{3j}\right)$; \end{itemize} Note that, in particular, we have multiplicative estimates for all intervals, except those in $L_1$, $L_{j^*}$, $R_1$ and the interval containing $i_0$. Let $S$ be the set of all intervals except $L_{j^*}$, $L_j$ and $R_j$ for $j \leq 1/\sqrt{\varepsilon}$, and the one containing $i_0$ Then, since each interval has probability mass at most $O(\varepsilon^{3/2})$ and we are excluding $O(1/\sqrt{\varepsilon})$ intervals, $p(S)>1-O(\varepsilon)$. We now compute the $\chi^2$-distance induced by this approximation for elements in $S$. For an element $i \in L_j \cap S$, we have $$\frac{(f(i) - p(i))^2}{p(i)} \leq \frac{60p(i)}{j^2}.$$ Summing over all $i \in L_j \cap S$ gives $$\frac{60\varepsilon^{3/2}}{j^2}$$ since the probability mass of $L_j$ is at most $\varepsilon^{3/2}$. Summing this over all $L_j$ for $j \geq 1/\sqrt{\varepsilon}$ and $j \neq j^*$ gives \begin{align*} 60\varepsilon^{3/2} \sum_{j = 1/\sqrt{\varepsilon}}^{2/\varepsilon^{3/2}} \frac{1}{j^2} &\leq 60\varepsilon^{3/2} \int_{1/\sqrt{\varepsilon}}^{\infty} \frac{1}{x^2} dx \\ &= 60\varepsilon^{3/2}(\sqrt{\varepsilon}) \\ &= O(\varepsilon^2) \end{align*} as desired. \paragraph{Part 2.} To obtain a distribution $q\in\mathcal{LCD}_n$, we write a linear program. We will work in the log domain, so our variables will be $Q_i$, representing $\log q(i)$ for $i \in [n]$. We will use $F_i = \log f(i)$ as parameters in our LP. There will be no objective function, we simply search for a feasible point. Our constraints will be $$Q_{i-1} + Q_{i+1} \leq 2 Q_{i} \ \ \forall i \in [n-1]$$ $$Q_{i} \leq 0 \ \ \forall i \in [n]$$ $$ \log (1 + \varepsilon) \leq |Q_i - F_i| \leq \log (1 + \varepsilon) \ \text{for}\ i \in M$$ $$ \log \left(1 - \frac{22}{3j}\right) \leq |Q_i - F_i| \leq \log \left(1 + \frac{22}{3j}\right) \ \text{for}\ i \in L_j, j \geq 1/\sqrt{\varepsilon} \ \text{and}\ j \neq j^*$$ $$ \log \left(1 - \frac{22}{3j}\right) \leq |Q_i - F_i| \leq \log \left(1 + \frac{22}{3j}\right) \ \text{for}\ i \in R_j, j \geq 1/\sqrt{\varepsilon}$$ If we run the linear program, then after a rescaling and summing the error over all the intervals in the LP gives us that the distance between $p$ and $q$ to be $O(\varepsilon^2)$ $\chi^2$-distance in a set $S$ which has measure $p(S) \geq 1 - 4\varepsilon$, as desired. If the linear program finds a feasible point, then we obtain a $q \in \mathcal{LCD}_n$. Furthermore, if $p \in \mathcal{LCD}_n$, this also tells us that (after a rescaling of $\varepsilon$), summing the error over all intervals implies that $\chi^2(p_S,q_S) \leq \frac{\varepsilon^2}{500}$ for a known set $S$ with $p(S) \geq 1 - O(\varepsilon)$, as desired. If $M \neq \emptyset$, this algorithm works as described. The issue is if $M = \emptyset$, then we don't know when the $L$ intervals end and the $R$ intervals begin. In this case, we run $O(1/\varepsilon)$ LPs, using each interval as the one containing $i_0$, and thus acting as the barrier between the $L$ intervals (to its left) and the $R$ intervals (to its right). If $p$ truly was log-concave, then one of these guesses will be correct and the corresponding LP will find a feasible point. \end{prevproof} \section{Lower Bounds} \label{sec:lower-bounds} We now prove sharp lower bounds for the classes of distributions we consider. We show that the example studied by Paninski~\cite{Paninski08} to prove lower bounds on testing uniformity can be used to prove lower bounds for the classes we consider. They consider a class $\mathcal{Q}$ consisting of $2^{n/2}$ distributions defined as follows. Without loss of generality assume that $n$ is even. For each of the $2^{n/2}$ vectors $z_0z_1\ldots z_{n/2-1}\in\{-1,1\}^{n/2}$, define a distribution $q\in\mathcal{Q}$ over $[n]$ as follows. \begin{align} \label{eqn:fcl} q_{i} =\begin{cases} \frac{(1+z_\ell c\varepsilon)}{n} & \text{ for } i = 2\ell+1\\ \frac{(1-z_{\ell}c\varepsilon)}{n}& \text{ for } i=2\ell.\\ \end{cases} \end{align} Each distribution in $\mathcal{Q}$ has a total variation distance $c\varepsilon/2$ from $U_n$, the uniform distribution over $[n]$. By choosing $c$ to be an appropriate constant, Paninski~\cite{Paninski08} showed that a distribution picked uniformly at random from $\mathcal{Q}$ cannot be distinguished from $U_n$ with fewer than $\sqrt{n}/\varepsilon^2$ samples with probability at least $2/3$. Suppose $\mathcal{C}$ is a class of distributions such that \begin{itemize} \item The uniform distribution $U_n$ is in $\mathcal{C}$, \item For appropriately chosen $c$, $d_{\mathrm {TV}}(\mathcal{C}, \mathcal{Q})\ge\varepsilon$, \end{itemize} then testing $\mathcal{C}$ is \costasnote{not} easier than distinguishing $U_n$ from $\mathcal{Q}$. Invoking~\cite{Paninski08} immediately implies that testing the class $\mathcal{C}$ requires $\Omega(\sqrt{n}/\varepsilon^2)$ samples. The lower bounds for all the one dimensional distributions will follow directly from this construction, and for testing monotonicity in higher dimensions, we extend this construction to $d\ge1$, appropriately. These arguments are proved in Section~\ref{sec:lb-appendix}, leading to the following lower bounds for testing these classes: \begin{theorem}$ $ \label{thm:lbs} \begin{itemize} \item For any $d\ge1$, any algorithm for testing monotonicity over $[n]^d$ requires $\Omega(n^{d/2}/\varepsilon^2)$ samples. \item For $d\ge1$, any algorithm for testing independence over $[n_1]\times\cdots\times[n_d]$ requires $\Omega\left(\frac{(n_1\cdot n_2\ldots\cdot n_d)^{1/2}}{\varepsilon^2}\right)$ samples. \item Any algorithm for testing unimodality, log-concavity, or monotone hazard rate over $[n]$ requires $\Omega(\sqrt{n}/\varepsilon^2)$ samples. \end{itemize} \end{theorem} \section{Introduction} \label{sec:introduction} The quintessential scientific question is whether an unknown object has some property, i.e. whether a model from a specific class fits the object's observed behavior. If the unknown object is a probability distribution, $p$, to which we have sample access, we are typically asked to distinguish whether $p$ belongs to some class $\mathcal{C}$ or whether it is sufficiently far from it. This question has received tremendous attention in \gnote{the field of} statistics (see, e.g.,~\cite{Fisher25,lehmann2006testing}), where test statistics for important properties such as the ones we consider here have been proposed. Nevertheless, the emphasis has been on asymptotic analysis, characterizing \gnote{the} rates of convergence of test statistics under null hypotheses, as the number of samples tends to infinity. In contrast, \gnote{we wish to study the following problem in the small sample regime:} \vspace{2ex} \begin{center} \smallskip \framebox{ \begin{minipage}{13.5cm} $\Pi(\mathcal{C},\varepsilon)$: Given a family of distributions $\mathcal{C}$, some $\varepsilon>0$, and sample access to an unknown distribution $p$ over \gnote{a} discrete support, how many samples are required to distinguish between $p \in \mathcal{C}$ versus $d_{\mathrm {TV}}(p,\mathcal{C})>\varepsilon$? \end{minipage} } \end{center} \vspace{2ex} The problem has been studied intensely in the literature on \gnote{property testing and sublinear algorithms} \cite{Goldreich98, Fischer01, Rubinfeld06, Ron08,canonne2015survey}, where the emphasis has been on characterizing the optimal tradeoff between $p$'s support size and the accuracy $\varepsilon$ in the number of samples. Several results have been obtained, roughly clustering into three groups, where (i) $\mathcal{C}$ is the class of monotone distributions over $[n]$, or more generally a poset~\cite{BatuKR04,Bhattacharyya11}; (ii) $\mathcal{C}$ is the class of independent, or $k$-wise independent distributions over a hypergrid~\cite{batu2001testing,alon2007testing}; and (iii) $\mathcal{C}$ contains a single-distribution $q$, and the problem becomes that of testing whether $p$ equals $q$ or is far from it~\cite{batu2001testing,Paninski08, valiant2014automatic}. With respect to (iii), \cite{valiant2014automatic} exactly characterizes the number of samples required to test identity to each distribution $q$, providing a single tester matching this bound simultaneously for all $q$. Nevertheless, this tester and its precursors are not applicable to the composite identity testing problem that we consider. If our class $\mathcal{C}$ were finite, we could test against each element in the class, albeit this would not necessarily be sample optimal. If our class $\mathcal{C}$ \gnote{were} a continuum, we \gnote{would} need \emph{tolerant} identity testers, which tend to be more expensive in terms of \gnote{sample complexity} \cite{ValiantV11}, and result in substantially suboptimal testers for the classes we consider. Or we could use approaches related to generalized likelihood ratio test, but their behavior is not well-understood in our regime, and optimizing likelihood over our classes becomes computationally intense. \paragraph{Our Contributions} In this paper, we obtain sample-optimal and computationally efficient testers for $\Pi(\mathcal{C},\varepsilon)$ for the most \gnote{fundamental} shape restrictions to a distribution. Our contributions are the following: \begin{enumerate} \item \jnote{For a known distribution $q$ over $[n]$, and given samples from an unknown distribution $p$, we show that distinguishing the cases: $(a)$ whether the $\chi^2$-distance between $p$ and $q$ is at most $\varepsilon^2/2$, versus $(b)$ the $\ell_1$ distance between $p$ and $q$ is at least $\varepsilon$, requires $\Theta(\sqrt n/\varepsilon^2)$ samples. As a corollary, we provide a simpler argument to show that identity testing requires $\Theta(\sqrt n/\varepsilon^2)$ samples (previously shown in \cite{valiant2014automatic}). } \item For the class $\mathcal{C}=\mathcal{M}_n^d$ of monotone distributions over $[n]^d$ we require an optimal $\Theta\left({n^{d/2} \over \varepsilon^2}\right)$ number of samples, where prior work requires $\Omega\left({\sqrt{n} \log n \over \varepsilon^4}\right)$ samples for $d=1$ and $\tilde{\Omega}\left(n^{d-{1\over 2}} {\rm poly}\left({1 \over \varepsilon}\right)\right)$ for $d>1$~\cite{BatuKR04,Bhattacharyya11}. Our results improve the exponent of $n$ with respect to $d$, shave all logarithmic factors in $n$, and improve the exponent of $\varepsilon$ by at least a factor of $2$. \begin{enumerate} \item A useful building block and interesting byproduct of our analysis is extending Birg\'e's oblivious decomposition for single-dimensional monotone distributions~\cite{Birge87} to monotone distributions in $d\ge1$, and to the stronger notion of $\chi^2$-distance. See Section~\ref{sec:hypergrid}. \item Moreover, we show that $O(\log^d n)$ samples suffice to learn a monotone distribution over $[n]^d$ in $\chi^2$-distance. See Lemma~\ref{lem:fin-learn-mon} for the precise statement. \end{enumerate} \item \gnote{For the class $\mathcal{C} = \Pi_d$ of product distributions over $[n_1]\times\cdots\times[n_d]$, our algorithm requires $O\left(\left((\prod_\ell n_\ell)^{1/2} + \sum_\ell n_\ell\right)/\varepsilon^2\right)$ samples. We note that a product distribution is one where all marginals are independent, so this is equivalent to testing if a collection of random variables are all independent.} \gnote{In the case where $n_\ell$'s are large, then the first term dominates, and the sample complexity is $O(\left(\prod_\ell n_\ell\right)^{1/2}/\varepsilon^2)$. In particular, when $d$ is a constant and all $n_\ell$'s are equal to $n$, we achieve the optimal sample complexity of $\Theta(n^{d/2}/\varepsilon^2)$. To the best of our knowledge, this is the first result for $d \geq 3$, and when $d = 2$, this improves the previously known complexity from $O\left(\frac{n}{\varepsilon^6}{\rm polylog}(n/\varepsilon)\right)$ \cite{batu2001testing,levi2013testing}, significantly improving the dependence on $\varepsilon$ and shaving all logarithmic factors. } \item For the classes $\mathcal{C}=\mathcal{LCD}_n$, $\mathcal{C}=\mathcal{MHR}_n$ and $\mathcal{C}=\mathcal{U}_n$ of log-concave, monotone-hazard-rate and unimodal distributions over $[n]$, we require an optimal $\Theta\left({\sqrt{n} \over \varepsilon^2}\right)$ number of samples. Our testers for $\mathcal{LCD}_n$ and $\mathcal{C}=\mathcal{MHR}_n$ are to our knowledge the first for these classes for the low sample regime we are studying---see~\cite{hall2005testing} and its references for statistics literature on the asymptotic regime. Our tester for $\mathcal{U}_n$ improves the dependence of the sample complexity on $\varepsilon$ by at least a factor of $2$ in the exponent, and shaves all logarithmic factors in $n$, compared to testers based on testing monotonicity. \begin{enumerate} \item A useful building block and important byproduct of our analysis are the first computationally efficient algorithms for properly learning log-concave and monotone-hazard-rate distributions, to within $\varepsilon$ in total variation distance, from ${\rm poly}(1/\varepsilon)$ samples, independent of the domain size $n$. See Corollaries~\ref{cor:learning LCD} and~\ref{cor:learning MHR}. Again, these are the first computationally efficient algorithms to our knowledge in the low sample regime.~\cite{AcharyaDLS15, ChanDSS13b} provide algorithms for density estimation, which are non-proper, i.e. will approximate an unknown distribution from these classes with a distribution that does not belong to these classes. On the other hand, the statistics literature focuses on maximum-likelihood estimation in the asymptotic regime---see e.g. \cite{cule2010theoretical} and its references. \end{enumerate} \item For all the above classes we obtain matching lower bounds, showing that the sample complexity of our testers is optimal with respect to $n$, $\varepsilon$ and when applicable $d$. See Section~\ref{sec:lower-bounds}. Our lower bounds are based on extending Paninski's lower bound for testing uniformity~\cite{Paninski08}. \end{enumerate} At the heart of our tester lies a novel use of the $\chi^2$ statistic. Naturally, the $\chi^2$ and its related $\ell_2$ statistic have been used in several of the afore-cited results. We propose a new use of the $\chi^2$ statistic enabling our optimal sample complexity. The essence of our approach is to first draw a small number of samples (independent of $n$ for log-concave and monotone-hazard-rate distributions and only logarithmic in $n$ for monotone and unimodal distributions) to approximate the unknown distribution $p$ in $\chi^2$ distance. If $p \in \mathcal{C}$, our learner is required to output a distribution $q$ that is $O(\varepsilon)$-close to $\mathcal{C}$ in total variation and $O(\varepsilon^2)$-close to $p$ in $\chi^2$ distance. Then some analysis reduces our testing problem to distinguishing the following cases: \begin{itemize} \item $p$ and $q$ are $O(\varepsilon^2)$-close in $\chi^2$ distance; this case corresponds to $p \in \mathcal{C}$. \item $p$ and $q$ are $\Omega(\varepsilon)$-far in total variation distance; this case corresponds to $d_{\mathrm {TV}}(p,\mathcal{C})>\varepsilon$. \end{itemize} \gnote{ We draw a comparison with \emph{robust identity testing}, in which one must distinguish whether $p$ and $q$ are $c_1\varepsilon$-close or $c_2\varepsilon$-far in total variation distance, for constants $c_2> c_1 > 0$. In \cite{ValiantV11}, Valiant and Valiant show that $\Omega(n/\log n)$ samples are required for this problem -- a nearly-linear sample complexity, which may be prohibitively large in many settings. In comparison, the problem we study tests for $\chi^2$ closeness rather than total variation closeness: a relaxation of the previous problem. However, our tester demonstrates that this relaxation allows us to achieve a substantially sublinear complexity of $O(\sqrt{n}/\varepsilon^2)$. On the other hand, this relaxation is still tight enough to be useful, demonstrated by our application in obtaining sample-optimal testers. } \jnote{We note that while the $\chi^2$ statistic for testing hypothesis is prevalent in statistics providing optimal error exponents in the large-sample regime, to the best of our knowledge, in the small-sample regime, \emph{modified-versions} of the $\chi^2$ statistic have only been recently used for \emph{closeness-testing} in~\cite{AcharyaDJOPS12, ChanDVV13} and for testing uniformity of monotone distributions in~\cite{AcharyaJOS13}. In particular,~\cite{AcharyaDJOPS12} design an unbiased statistic for estimating the $\chi^2$ distance between two \emph{unknown} distributions.} In Section~\ref{sec:testing}, we show that a version of the $\chi^2$ statistic, appropriately excluding certain elements of the support, is sufficiently well-concentrated to distinguish between the above cases. Moreover, the sample complexity of our algorithm is optimal for most classes. Our base tester is combined with the afore-mentioned extension of Birg\'e's decomposition theorem to test monotone distributions in Section~\ref{sec:monotone} (see Theorem~\ref{thm:monotone-final} and Corollary~\ref{cor:high-d}), and is also used to test independence of distributions in Section~\ref{sec:independence} (see Theorem~\ref{thm:independence}). Naturally, there are several bells and whistles that we need to add to the above skeleton to accommodate all classes of distributions that we are considering. For log-concave and monotone-hazard distributions, we are unable to obtain a cheap (in terms of samples) learner that $\chi^2$-approximates the unknown distribution $p$ throughout its support. Still, we can identify a subset of the support where the $\chi^2$-approximation is tight and which captures almost all the probability mass of $p$. We extend our tester to accommodate excluding subsets of the support from the $\chi^2$-approximation. See Theorems~\ref{thm:lcd-main} and~\ref{thm:mhr-main} in Sections~\ref{sec:LCD-main} and \ref{sec:MHR-main}. For unimodal distributions, we are even unable to identify a large enough subset of the support where the $\chi^2$ approximation is guaranteed to be tight. But we can show that there exists a light enough piece of the support (in terms of probability mass under $p$) that we can exclude to make the $\chi^2$ approximation tight. Given that we only use Chebyshev's inequality to prove the concentration of the test statistic, it would seem that our lack of knowledge of the piece to exclude would involve a union bound and a corresponding increase in the required number of samples. We avoid this through a careful application of Kolmogorov's max inequality in our setting. See Theorem~\ref{thm:unimodality} of Section~\ref{sec:unimodal-main}. \paragraph{Related Work.} \jnote{For the problems that we study in thie paper, we have provided the related works in the previous section along with our contributions.} We cannot do justice to the role of shape restrictions of probability distributions in probabilistic modeling and testing. It suffices to say that the classes of distributions that we study are fundamental, motivating extensive literature on their learning and testing~\cite{BBBB:72}. In the recent times, there has been work on shape restricted statistics, pioneered by Jon Wellner, and others.~\cite{JW:09,BW10sn} study estimation of monotone and $k-$ monotone densities, and ~\cite{BJR11,SumardW14} study estimation of log-concave distributions. As we have mentioned, statistics has focused on the asymptotic regime as the number of samples tends to infinity. Instead we are considering the low sample regime and are more stringent about the behavior of our testers, requiring $2$-sided guarantees. We want to accept if the unknown distribution is in our class of interest, and also reject if it is far from the class. For this problem, as discussed above, there are few results when $\mathcal{C}$ is a whole class of distributions. Closer related to our paper is the line of papers~\cite{BatuKR04, ACS10, Bhattacharyya11} for monotonicity testing, albeit these papers have sub-optimal sample complexity as discussed above. \gnote{Testing independence of random variables has a long history in statisics~\cite{rao1981analysis, agresti2011categorical}. The theoretical computer science community has also considered the problem of testing independence of two random variables~\cite{batu2001testing, levi2013testing}. While our results sharpen the case where the variables are over domains of equal size, they demonstrate an interesting asymmetric upper bound when this is not the case.} More recently, Acharya and Daskalakis provide optimal testers for the family of Poisson Binomial Distributions~\cite{AcharyaD15}. \gnote{ Finally, contemporaneous work of Canonne et al~\cite{CanonneDGR15a,CanonneDGR15b} provides a generic algorithm and lower bounds for the single-dimensional families of distributions considered here. We note that their algorithm has a sample complexity which is suboptimal in both $n$ and $\varepsilon$, while our algorithms are optimal. Their algorithm also extends to mixtures of these classes, though some of these extensions are not computationally efficient. They also provide a framework for proving lower bounds, giving the optimal bounds for many classes when $\varepsilon$ is sufficiently large with respect to $1/n$. In comparison, we provide these lower bounds unconditionally by modifying Paninski's construction \cite{Paninski08} to suit the classes we consider. } \input{preliminaries} \input{overview} \input{testing} \input{monotone} \input{unimodal-main} \input{independence} \input{otherclasses} \input{lower-bounds} \section*{Acknowledgements} The authors thank Cl\'ement Canonne and Jerry Li; the former for several useful comments and suggestions on previous drafts of this work, and both for helpful discussions and thoughts regarding independence testing. \bibliographystyle{alpha} \section{Details on MHR testing} \label{sec:MHR} \begin{prevproof}{Lemma}{lem:mhr-learn} As with log-concave distributions, our method for MHR distributions can be split into two parts. In the first step, if $p \in \mathcal{MHR}_n$, we obtain a distribution $q$ which is $O(\varepsilon^2)$-close to $p$ in $\chi^2$ distance on a set $\mathcal{A}$ of intervals such that $p(\mathcal{A}) \geq 1 - O(\varepsilon)$. $q$ will achieve this by being a multiplicative $(1 + O(\varepsilon))$ approximation for each element within these intervals. This step is very similar to the decomposition used for unimodal distributions (described in Section~\ref{sec:unimodal-appendix}), so we sketch the argument and highlight the key differences. The second step will be to find a feasible point in a linear program. If $p \in \mathcal{MHR}_n$, there should always be a feasible point, indicating that $q$ is close to a distribution in $\mathcal{MHR}_n$ (leveraging the particular guarantees for our algorithm for generating $q$). If $d_{\mathrm {TV}}(p,\mathcal{MHR}_n) \geq \varepsilon$, there may or may not be a feasible point, but when there is, it should imply the existence of a distribution $p^* \in \mathcal{MHR}_n$ such that $d_{\mathrm {TV}}(q,p^*) \leq \varepsilon/2$. The analysis will rely on the following lemma from \cite{ChanDSS13a}, which roughly states that an MHR distribution is ``almost'' non-decreasing. \begin{lemma}[Lemma 5.1 in \cite{ChanDSS13a}] \label{lem:mhr-str} Let $p$ be an MHR distribution over $[n]$. Let $I = [a,b] \subset [n]$ be an interval, and $R = [b+1,n]$ be the elements to the right of $I$. Let $\eta = p(I)/p(R)$. Then $p(b+1) \geq \frac{1}{1 + \eta}p(a)$. \end{lemma} \paragraph{Part 1.} As before, with unimodal distributions, we start by taking $O(\frac{b \log b}{\varepsilon^2})$ samples, with the goal of partitioning the domain into intervals of mass approximately $\Theta(1/b)$. First, we will ignore the left and rightmost intervals of mass $\Theta(\varepsilon)$. For all ``heavy'' elements with mass $\geq \Theta(1/b)$, we consider them as singletons. We note that Lemma~\ref{lem:mhr-str} implies that there will be at most $O(1/\varepsilon)$ contiguous intervals of such elements. The rest of the domain is greedily divided (from left to right) into intervals of mass $\Theta(1/b)$, cutting an interval short if we reach one of the heavy elements. This will result in the guarantee that all but potentially $O(1/\varepsilon)$ intervals have $\Theta(1/b)$ mass. Next, similar to unimodal distributions, considering the flattened distribution, we discard all intervals for which the per-element probability is not within a $(1 \pm O(\varepsilon))$ multiplicative factor of the same value for both neighboring intervals. The claim is that all remaining intervals will have the property that the per-element probability is within a $(1 \pm O(\varepsilon))$ multiplicative factor of the true probability. This is implied by Lemma~\ref{lem:mhr-str}. If there were a point in an interval which was above this range, the distribution must decrease slowly, and the next interval would have a much larger per-element weight, thus leading to the removal of this interval. A similar argument forbids us from missing an interval which contains a point that lies outside this range. Relying on the fact that truncating the left and rightmost intervals eliminates elements with low probability mass, similar to the unimodal case, one can show that we will remove at most $\log(n/\varepsilon)/\varepsilon$ intervals, and thus a $\log(n/\varepsilon)/b\varepsilon$ probability mass. Choosing $b = \Omega(\varepsilon^2/\log(n/\varepsilon))$ limits this to be $O(\varepsilon)$, as desired. At this point, if $p$ is indeed MHR, the multiplicative estimates guarantee that the result is $O(\varepsilon^2)$-close in $\chi^2$-distance among the remaining intervals. \paragraph{Part 2.} We note that an equivalent condition for distribution $f$ being MHR is log-concavity of $\log(1 - F)$, where $F$ is the CDF of $f$. Therefore, our approach for this part will be similar to the approach used for log-concave distributions. Given the output distribution $q$ from the previous part of this algorithm, our goal will be check if there exists an MHR distribution $f$ which is $O(\varepsilon)$-close to $q$. We will run a linear program with variables $\mathfrak{f}_i = \log(1 - F_i)$. First, we ensure that $f$ is a distribution. This can be done with the following constraints: \begin{alignat*}{3} &\mathfrak{f}_i &&\leq 0 \hphantom{space}&&\forall i \in [n] \\ &\mathfrak{f}_i &&\geq \mathfrak{f}_{i+1} &&\forall i \in [n-1] \\ &\mathfrak{f}_n &&= -\infty \end{alignat*} To ensure that $f$ is MHR, we use the following constraint: \begin{alignat*}{3} &\mathfrak{f}_{i-1} + \mathfrak{f}_{i+1} &&\leq 2\mathfrak{f}_i \hphantom{space}&&\forall i \in [2,n-1] \end{alignat*} Now, ideally, we would like to ensure $f$ and $q$ are $\varepsilon$-close in total variation distance by ensuring they are pointwise within a multiplicative $(1 \pm \varepsilon)$ factor of each other: $$(1 - \varepsilon) \leq f_i/q_i \leq (1 + \varepsilon)$$ We note that this is a stronger condition than $f$ and $q$ being $\varepsilon$-close, but if $p \in \mathcal{MHR}_n$, the guarantees of the previous step would imply the existence of such an $f$. We have a separate treatment for the identified singletons (i.e., those with probability $\geq 1/b$) and the remainder of the support. For each element $q_i$ identified to have $\geq 1/b$ mass, we add two constraints: $$\log((1-\varepsilon/2b)(1 - Q_i)) \leq \mathfrak{f}_i \leq \log((1+\varepsilon/2b)(1 - Q_i))$$ $$\log((1-\varepsilon/2b)(1 - Q_{i-1})) \leq \mathfrak{f}_{i-1} \leq \log((1+\varepsilon/2b)(1 - Q_{i-1}))$$ If we satisfy these constraints, it implies that $$q_i - \varepsilon/b \leq f_{i} \leq q_i + \varepsilon/b.$$ Since $q_i \geq 1/b$, this implies $$(1 - \varepsilon)q_i \leq f_i \leq (1 + \varepsilon)q_i$$ as desired. Now, the remaining elements each have $\leq 1/b$ mass. For each such element $q_i$, we create a constraint $$(1 - O(\varepsilon))\frac{q_i}{1 - Q_{i-1}} \leq \mathfrak{f}_{i-1} - \mathfrak{f}_i \leq (1 + O(\varepsilon))\frac{q_i}{1 - Q_{i-1}} $$ Note that the middle term is $$-\log\left(\frac{1 - F_i}{1 - F_{i-1}}\right) = -\log\left(1 - \frac{f_i}{1 - F_{i-1}}\right) \in \frac{f_i}{1 - F_{i-1}}\left(1 \pm 2\varepsilon\right),$$ where the second equality uses the Taylor expansion and the facts that $f_i \leq 1/b$ and $1 - F_{i-1} \geq \varepsilon$ (since during the previous part, we ignored the rightmost $O(\varepsilon)$ probability mass). If we satisfy the desired constraints, it implies that $$f_i \in \frac{1}{\left(1 \pm 2\varepsilon\right)}\frac{1 - F_{i-1}}{1 - Q_{i-1}}(1 \mp O(\varepsilon))q_i.$$ Since we are taking $\Omega(1/\varepsilon^4)$ samples and $1 - F_{i-1} \geq \Omega(\varepsilon)$, Lemma~\ref{lem:dkw} implies that $f_i$ is indeed a multiplicative $(1 \pm \varepsilon)$ approximation for these points as well. We note that all points which do not fall into these two cases make up a total of $O(\varepsilon)$ probability mass. Therefore, $f$ may be arbitrary at these points and only incur $O(\varepsilon)$ cost in total variation distance. If we find a feasible point for this linear program, it implies the existence of an MHR distribution within $O(\varepsilon)$ total variation distance. In this case, we continue to the testing portion of the algorithm. Furthermore, if $p \in \mathcal{MHR}_n$, our method for generating $q$ certifies that such a distribution exists, and we continue on to the testing portion of the algorithm. \end{prevproof} \section{Testing for Monotone Hazard Rate} \label{sec:MHR} \gnote{Insert prose} \section{Details on Testing Monotonicity} \label{sec:monotone-appendix} In this section, we prove the lemmas necessary for our monotonicity testing result. \subsection{A Structural Result for Monotone Distributions on the Hypergrid} \label{sec:hypergrid} Birg\'e~\cite{Birge87} showed that any monotone distribution is estimated to a total variation $\varepsilon$ with a $O(\log(n)/\varepsilon)$-piecewise constant distribution. Moreover, the intervals over which the output is constant is independent of the distribution $p$. This result, was strengthened to the Kullback-Leibler divergence by~\cite{AcharyaJOS14a} to study the compression of monotone distributions. They upper bound the KL divergence by $\chi^2$ distance and then bound the $\chi^2$ distance. We extend this result to $[n]^d$. We divide $[n]^d$ into $b^d$ rectangles as follows. Let $\{I_1, \dots, I_b\}$ be a partition of $[n]$ into consecutive intervals defined as: \begin{align*} |I_j| \begin{cases} 1 & \text{ for } 1\le j\le\frac b2, \\ \lfloor 2(1+\gamma)^{j-b/2} \rfloor & \text{ for }\frac b2< j\le b. \end{cases} \end{align*} For ${\bf j}=(j_1,\ldots, j_d)\in[b]^d$, let $I_{\bf j}{\hfill\blksquare}\medskip I_{j_1}\times I_{j_2}\times\ldots\times I_{j_d}$. The $\chi^2$ distance between $p$ and $\bar{\dP}$ can be bounded as \begin{align*} \chi^2(p,\bar{\dP}) =& \left[\sum_{{\bf j}\in[b]^d}\sum_{{\bf i}\in I_{{\bf j}}}\frac{{\dP}_{{\bf i}}^2}{{\dPbar}_{{\bf i}}}\right]-1\\ \le& \left[\sum_{{\bf j}\in[b]^d}{\dP}_{{\bf j}}^+|I_{\bf j}|\right]-1\\ \end{align*} For ${\bf j}=(j_1,\ldots, j_d)\in\mathcal{S}_{\rm large}$, let ${\bf j}^*=(j_1^*,\ldots, j_b^*)$ be \begin{align*} j_i^* =\begin{cases} j_i& \text{ if } j_i\le b/2+1\\ j_i-1& \text{ otherwise}. \end{cases} \end{align*} We bound the expression above as follows. Let $T \subseteq [d]$ be any subset of $d$. Suppose the size of $T$ is $\ell$. Let $\bar T$ be the set of all ${\bf j}$ that satisfy ${\bf j}_i=b/2+1$ for $i\in T$. In other words, over the dimensions determined by $T$, the value of the index is equal to $d/2+1$. The map ${\bf j}\rightarrow{\bf j}^*$ restricted to $T$ is one-to-one, and since at most $d-\ell$ of the coordinates drop, \begin{align*} |I_{{\bf j}}| \le |I_{{\bf j}^*}|\cdot(1+\gamma)^{d-\ell}. \end{align*} Since there are $\ell$ coordinates that do not change, and each of them have $2(1+\gamma)$ coordinates, we obtain \begin{align*} \sum_{{\bf j}\in\bar T}p_{\bf j} \le& \ \sum_{{\bf j}\in\bar T}p_{{\bf j}^*}^-\cdot|I_{{\bf j}}|\cdot (2(1+\gamma))^{\ell}\cdot(1+\gamma)^{d-\ell}\\ =&\ \sum_{{\bf j}\in\bar T}p_{{\bf j}^*}^-\cdot|I_{{\bf j}^*}| \cdot 2^{\ell} (1+\gamma)^{d}. \end{align*} Since the mapping is one-to-one, the probability of observing as element in $\bar T$ is the probability of observing $b/2+1$ in $\ell$ coordinates, which is at most $(2/(b+2))^\ell$ under any monotone distribution. Therefore, \begin{align*} \sum_{{\bf j}\in\bar T}p_{\bf j} \le&\left(\frac2{b+2}\right)^\ell \cdot 2^{\ell} (1+\gamma)^{d}. \end{align*} For any $\ell$ there are ${d\choose \ell}$ choices for $T$. Therefore, \begin{align} \chi^2(p,\bar{\dP})\le\ & \sum_{\ell=0}^d {d\choose \ell} \left(\frac4{b+2}\right)^\ell (1+\gamma)^{d}-1\nonumber\\ =&\ (1+\gamma)^d\left(1+\frac4{b+2}\right)^d-1\nonumber\\ =& \ \left(1+\gamma+\frac4{b+2}+\frac{4\gamma}{b+2}\right)^d-1\nonumber \end{align} Recall that $\gamma=2\log (n)/b>1/b$, implies that the expression above is at most $(1+2\gamma)^d-1$. This implies Lemma~\ref{lem:mon-str}. \subsection{Monotone Learning} Our algorithm requires a distribution $q$ satisfying the properties discussed earlier. We learn a monotone distribution from samples as follows. Before proving this result, we prove a general result for $\chi^2$ learning of arbitrary discrete distributions, adapting the result from~\cite{KamathOPS15}. For a distribution $p$, and a partition of the domain into $b$ intervals $I_1, \ldots, I_b$, let $\bar{\dP}_i= p(I_i)/|I_i|$ be the flattening of $p$ over these intervals. We saw that for monotone distributions there exists a partition of the domain such that $\bar{\dP}$ is \emph{close} to the underlying distribution in $\chi^2$ distance. Suppose we are given $m$ samples from a distribution $p$ and a partition $I_1, \ldots, I_b$. Let $m_j$ be the number of samples that fall in $I_j$. For $i\in I_j$, let \[ \dQ_i {\hfill\blksquare}\medskip \frac{1}{|I_j|}\frac{m_j+1}{m+b}. \] Let $S_j=\sum_{i\in I_j}\dP_i^2$. The expected $\chi^2$ distance between $p$ and $q$ can be bounded as follows. \begin{align} \expectation{\chi^2(p,q) } =& \left[\sum_{j=1}^b \sum_{i\in I_j} \sum_{\ell=0}^{m}{m\choose \ell}(p(I_j))^{\ell}(1-p(I_j))^{m-\ell}\frac{\dP_i^2}{(\ell+1)/(|I_j|(m+b))}\right]-1\nonumber\\ = &\left[\frac{m+b}{m+1}\sum_{j=1}^b \frac{S_j}{\bar{\dP}(I_j)/|I_j|} \left(\sum_{\ell=0}^{m}{m+1\choose \ell+1}(p(I_j))^{\ell+1}(1-p(I_j))^{m+1-\ell+1}\right)\right]-1 \nonumber\\ = & \left[\frac{m+b}{m+1}\sum_{j=1}^b \frac{S_j}{\bar{\dP}(I_j)/|I_j|} \left(1-(1-p(I_j)^{m+1}\right)\right]-1\nonumber\\ \le & \left[\frac{m+b}{m+1}\sum_{j=1}^b \frac{S_j}{\bar{\dP}(I_j)/|I_j|}\right]-1\nonumber\\ = & \left[\frac{m+b}{m+1}\left(\chi^2(p,\bar{\dP})+1\right)\right]-1\nonumber\\ = & \frac{m+b}{m+1}\cdot\chi^2(p,\bar{\dP})+\frac{b}{m+1}\label{eqn:chi-learn}. \end{align} Suppose $\gamma = O(\log (n)/b)$, and $b=O(d\cdot\log (n)/\varepsilon^2)$. Then, by Lemma~\ref{lem:mon-str}, \begin{align} \chi^2(p,\bar{\dP})\le \varepsilon^2. \end{align} Combining this with~\eqref{eqn:chi-learn} gives Lemma~\ref{lem:fin-learn-mon}. \section{Testing Monotonicity} \label{sec:monotone} As an application of our testing framework, we will demonstrate how to test for monotonicity. Let $d\geq1$, and ${\bf i}=(i_1, \ldots, i_d),{\bf j}=(j_1, \ldots, j_d)\in[n]^d$. We say ${\bf i}\succcurlyeq{\bf j}$ if $i_l>j_l$ for $l=1,\ldots,d$. \begin{definition} A distribution $p$ over $[n]^d$ is monotone (decreasing) if for all ${\bf i}\succcurlyeq{\bf j}$, $p_{{\bf i}}\le p_{{\bf j}}$. \end{definition} Our main result of this section is as follows: \begin{theorem} \label{thm:monotone-final} For any $d \geq 1$, there exists an algorithm for testing monotonicity over $[n]^d$ with sample complexity $$O\left(\frac{n^{d/2}}{\varepsilon^2}+\left(\frac{d\log n}{\varepsilon^2}\right)^d\cdot \frac1{\varepsilon^2}\right)$$ and time complexity $O\left(\frac{n^{d/2}}{\varepsilon^2} + \poly(\log n, 1/\varepsilon)^d \right)$. \end{theorem} In particular, this implies the following optimal algorithms for monotonicity testing for all $d \geq 1$: \begin{corollary} \label{cor:high-d} Fix any $d \geq 1$, and suppose $\varepsilon>\frac{\sqrt{d\log n}}{n^{1/4}}$. Then there exists an algorithm for testing monotonicity over $[n]^d$ with sample complexity $O\left(n^{d/2}/\varepsilon^2\right)$. \end{corollary} Our analysis starts with a structural lemma about monotone distributions. In \cite{Birge87}, Birg\'e showed that any monotone distribution $p$ over $[n]$ can be \emph{obliviously} decomposed into $O(\log(n)/\varepsilon)$ intervals, such that the flattening $\bar p$ (recall Definition~\ref{def:flattening}) of $p$ over these intervals is $\varepsilon$-close to $p$ in total variation distance. \cite{AcharyaJOS14a} extend this result, giving a bound between the $\chi^2$-distance of $p$ and $\bar p$. We strengthen these results by extending them to monotone distributions over $[n]^d$. In particular, we partition the domain $[n]^d$ of $p$ into $O( (d\log(n)/\varepsilon^2)^d)$ rectangles, and compare it with $\bar p$, the flattening over these rectangles. \begin{lemma} \label{lem:mon-str} Let $d\ge1$. There is an oblivious decomposition of $[n]^d$ into $O((d\log(n)/\varepsilon^2)^d)$ rectangles such that for any monotone distribution $p$ over $[n]^d$, its flattening $\bar p$ over these rectangles satisfy $\chi^2(p,\bar p) \leq \varepsilon^2$. \end{lemma} This effectively reduces the support size to logarithmic in $n$. At this point, we can apply the Laplace estimator (along the lines of~\cite{KamathOPS15}) and learn a $q$ such that if $p$ was monotone, then $q$ will be $O(\varepsilon^2)$-close in $\chi^2$-distance. \begin{lemma} \label{lem:fin-learn-mon} Let $d\ge1$, and $p$ be a monotone distribution over $[n]^d$. There is an algorithm which outputs a distribution $q$ such that $\expectation{\chi^2(p,q)}\le \frac{\varepsilon^2}{500}$. The time and sample complexity are both $O((d\log (n)/\varepsilon^2)^d/\varepsilon^2)$. \end{lemma} The final step before we apply our $\chi^2$-tester is to compute the distance between $q$ and $\mathcal{M}_n^d$. This subroutine is similar to the one introduced by~\cite{BatuKR04}. The key idea is to write a linear program, which searches for any distribution $f$ which is close to $q$ in total variation distance. We note that the desired properties of $f$ (i.e., monotonicity, normalization, and $\varepsilon$-closeness to $q$) are easy to enforce as linear constraints. If we find that such an $f$ exists, we will apply our $\chi^2$-test to $q$. If not, we output $\textsc{Reject}\xspace$, as this is sufficient evidence to conclude that $p \not \in \mathcal{M}_n^d$. Note that the linear program operates over the oblivious decomposition used in our structural result, so the complexity is polynomial in $(d\log(n)/\varepsilon)^d$, rather than the naive $n^d$. At this point, we have precisely the guarantees needed to apply Theorem~\ref{thm:chisq-test}, directly implying Theorem~\ref{thm:monotone-final}. Proof of the lemmas in this section are provided in Section~\ref{sec:monotone-appendix}. \section{Testing Log-Concavity} \label{sec:LCD-main} In this section we describe our results for testing log-concavity of distributions. Our main result is as follows: \begin{theorem} \label{thm:lcd-main} There exists an algorithm for testing log-concavity over $[n]$ with sample complexity $$O\left(\frac{\sqrt{n}}{\varepsilon^2}+\frac{1}{\varepsilon^5}\right)$$ and time complexity $\poly(n,1/\varepsilon)$. \end{theorem} In particular, this implies the following optimal tester for this class: \begin{corollary} Suppose $\varepsilon > 1/n^{1/5}$. Then there exists an algorithm for testing log-concavity over $[n]$ with sample complexity $O\left(\sqrt{n}/\varepsilon^2\right)$. \end{corollary} Our algorithm will fit into the structure of our general framework. We first perform a very particular type of learning algorithm, whose guarantees are summarized in the following lemma: \begin{lemma} \label{lem:lcd-learn} Given $\varepsilon > 0$ and sample access to a distribution $p$, there exists an algorithm with the following guarantees: \begin{itemize} \item If $p \in \mathcal{LCD}_n$, the algorithm outputs a distribution $q \in \mathcal{LCD}_n$ and an $O(\varepsilon)$-effective support $S$ of $p$ such that $\chi^2(p_S,q_S) \leq \frac{\varepsilon^2}{500}$ with probability at least $5/6$; \item If $d_{\mathrm {TV}}(p,\mathcal{LCD}_n) \geq \varepsilon$, the algorithm either outputs a distribution $q \in \mathcal{LCD}_n$ or \textsc{Reject}\xspace. \end{itemize} The sample complexity is $O(1/\varepsilon^5)$ and the time complexity is $\poly(n,1/\varepsilon)$. \end{lemma} We note that as a corollary, one immediately obtains a $O(1/\varepsilon^5)$ proper learning algorithm for log-concave distributions. The result is immediate from the first item of Lemma~\ref{lem:lcd-learn} and Proposition~\ref{prop:distance-relations}. We can actually do a bit better -- in the proof of Lemma~\ref{lem:lcd-learn}, we partition $[n]$ into intervals of probability mass $\Theta(\varepsilon^{3/2})$. If one instead partitions into intervals of probability mass $\Theta(\varepsilon/\log(1/\varepsilon))$ and works directly with total variation distance instead of $\chi^2$ distance, one can show that $\tilde O(1/\varepsilon^4)$ samples suffice. \begin{corollary} \label{cor:learning LCD} Given $\varepsilon > 0$ and sample access to a distribution $p \in \mathcal{LCD}_n$, there exists an algorithm which outputs a distribution $q \in \mathcal{LCD}_n$ such that $d_{\mathrm {TV}}(p,q) \leq \varepsilon$. The sample complexity is $\tilde O(1/\varepsilon^4)$ and the time complexity is $\poly(n,1/\varepsilon)$. \end{corollary} Then, given the guarantees of Lemma \ref{lem:lcd-learn}, Theorem~\ref{thm:lcd-main} follows from Theorem~\ref{thm:chisq-test}\footnote{To be more precise, we require the modification of Theorem~\ref{thm:lcd-main} which is described in Section~\ref{sec:testing}, in order to handle the case where the $\chi^2$-distance guarantees only hold for a known effective support.}. The details of these results are presented in Section~\ref{sec:LCD}. \section{Testing for Monotone Hazard Rate} \label{sec:MHR-main} In this section, we obtain our main result for testing for monotone hazard rate: \begin{theorem} \label{thm:mhr-main} There exists an algorithm for testing monotone hazard rate over $[n]$ with sample complexity $$O\left(\frac{\sqrt{n}}{\varepsilon^2}+\frac{\log(n/\varepsilon)}{\varepsilon^4}\right)$$ and time complexity $\poly(n,1/\varepsilon)$. \end{theorem} This implies the following optimal tester for the class: \begin{corollary} Suppose $\varepsilon > \sqrt{\log(n/\varepsilon)}/n^{1/4}$. Then there exists an algorithm for testing monotone hazard rate over $[n]$ with sample complexity $O\left(\sqrt{n}/\varepsilon^2\right)$. \end{corollary} We obey the same framework as before, first applying a $\chi^2$-learner with the following guarantees: \begin{lemma} \label{lem:mhr-learn} Given $\varepsilon > 0$ and sample access to a distribution $p$, there exists an algorithm with the following guarantees: \begin{itemize} \item If $p \in \mathcal{MHR}_n$, the algorithm outputs a distribution $q \in \mathcal{MHR}_n$ and an $O(\varepsilon)$-effective support $S$ of $p$ such that $\chi^2(p_S,q_S) \leq \frac{\varepsilon^2}{500}$ with probability at least $5/6$; \item If $d_{\mathrm {TV}}(p,\mathcal{MHR}_n) \geq \varepsilon$, the algorithm either outputs a distribution $q \in \mathcal{MHR}_n$ and a set $S \subseteq [n]$ or \textsc{Reject}\xspace. \end{itemize} The sample complexity is $O(\log(n/\varepsilon)/\varepsilon^4)$ and the time complexity is $\poly(n,1/\varepsilon)$. \end{lemma} As with log-concave distributions, this implies the following proper learning result: \begin{corollary} \label{cor:learning MHR} Given $\varepsilon > 0$ and sample access to a distribution $p \in \mathcal{MHR}_n$, there exists an algorithm which outputs a distribution $q \in \mathcal{MHR}_n$ such that $d_{\mathrm {TV}}(p,q) \leq \varepsilon$. The sample complexity is $O(\log(n/\varepsilon)/\varepsilon^4)$ and the time complexity is $\poly(n,1/\varepsilon)$. \end{corollary} Again, combining the learning guarantees of Lemma \ref{lem:mhr-learn} with the appropriate variant of Theorem~\ref{thm:chisq-test}, we obtain Theorem~\ref{thm:mhr-main}. The details of the argument and proofs are presented in Section~\ref{sec:MHR}. \section{Overview} Our algorithm for testing a distribution $p$ can be decomposed into three steps. \paragraph{Near-proper learning in $\chi^2$-distance.} Our first step requires a learning algorithm with very specific guarantees. In proper learning, we are given sample access to a distribution $p \in \mathcal{C}$, where $\mathcal{C}$ is some class of distributions, and we wish to output $q \in \mathcal{C}$ such that $p$ and $q$ are close in total variation distance. In our setting, given sample access to $p \in \mathcal{C}$, we wish to output $q$ such that $q$ is \emph{close} to $\mathcal{C}$ in total variation distance, and $p$ and $q$ are close in $\chi^2$-distance on an effective support\footnote{We also require the algorithm to output a description of an effective support for which this property holds. This requirement can be slightly relaxed, as we show in our results for testing unimodality.} of $p$. From an information theoretic standpoint, this problem is harder than proper learning, since $\chi^2$-distance is more restrictive than total variation distance. Nonetheless, this problem can be shown to have comparable sample complexity to proper learning for the structured classes we consider in this paper. \paragraph{Computation of distance to class.} The next step is to see if the hypothesis $q$ is close to the class $\mathcal{C}$ or not. Since we have an explicit description of $q$, this step requires no further samples from $p$, i.e. it is purely computational. If we find that $q$ is far from the class $\mathcal{C}$, then it must be that $p \not \in \mathcal{C}$, as otherwise the guarantees from the previous step would imply that $q$ is close to $\mathcal{C}$. Thus, if it is not, we can terminate the algorithm at this point. \paragraph{$\chi^2$-testing.} At this point, the previous two steps guarantee that our distribution $q$ is such that: \begin{itemize} \item If $p \in \mathcal{C}$, then $p$ and $q$ are close in $\chi^2$ distance on a (known) effective support of $p$; \item If $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, then $p$ and $q$ are far in total variation distance. \end{itemize} We can distinguish between these two cases using $O(\sqrt{n}/\varepsilon^2)$ samples with a simple statistical $\chi^2$-test, that we describe in Section~\ref{sec:testing}. \smallskip Using the above three-step approach, our tester, as described in the next section, can directly test monotonicity, log-concavity, and monotone hazard rate. With an extra trick, using Kolmogorov's max inequality, it can also test unimodality. \section{Probability distances} We use the following probability distances in our paper. \begin{definition} The \emph{total variation distance} between distributions $p$ and $q$ is defined as $$d_{\mathrm {TV}}(p,q) {\hfill\blksquare}\medskip \sup_{A} |p(A) - q(A)| = \frac12\|p - q\|_1.$$ \end{definition} For a subset of the domain, the total variation distance is defined as half of the $\ell_1$ distance restricted to the subset. \begin{definition} \label{def:chisq} The \emph{$\chi^2$-distance} between $p$ and $q$ over $[n]$ is defined by $$ \chi^2(p,q) {\hfill\blksquare}\medskip \sum_{i \in [n]}\frac{(p_i-q_i)^2}{q_i} = \left[\sum_{i \in [n]}\frac{p_i^2}{q_i}\right]-1.$$ \end{definition} \begin{definition} The \emph{Kolmogorov distance} between two probability measures $p$ and $q$ over an ordered set ($e.g.$, $\Rho$) with cumulative density functions (CDF) $F_p$ and $F_q$ is defined as $$d_{\mathrm K}(p,q) {\hfill\blksquare}\medskip \sup_{x \in \mathbb{R}} |F_p(x) - F_q(x)|.$$ \end{definition} \section{Preliminaries} We use the following probability distances in our paper. \begin{definition} The \emph{total variation distance} between distributions $p$ and $q$ is defined as $$d_{\mathrm {TV}}(p,q) {\hfill\blksquare}\medskip \sup_{A} |p(A) - q(A)| = \frac12\|p - q\|_1.$$ \end{definition} For a subset of the domain, the total variation distance is defined as half of the $\ell_1$ distance restricted to the subset. \begin{definition} \label{def:chisq} The \emph{$\chi^2$-distance} between $p$ and $q$ over $[n]$ is defined by $$ \chi^2(p,q) {\hfill\blksquare}\medskip \sum_{i \in [n]}\frac{(p_i-q_i)^2}{q_i} = \left[\sum_{i \in [n]}\frac{p_i^2}{q_i}\right]-1.$$ \end{definition} \begin{definition} The \emph{Kolmogorov distance} between two probability measures $p$ and $q$ over an ordered set ($e.g.$, $\Rho$) with cumulative density functions (CDF) $F_p$ and $F_q$ is defined as $$d_{\mathrm K}(p,q) {\hfill\blksquare}\medskip \sup_{x \in \mathbb{R}} |F_p(x) - F_q(x)|.$$ \end{definition} Our paper is primarily concerned with testing against classes of distributions, defined formally as follows: \begin{definition} Given $\varepsilon \in (0,1]$ and sample access to a distribution $p$, an algorithm is said to \emph{test} a class $\mathcal{C}$ if it has the following guarantees: \begin{itemize} \item If $p \in \mathcal{C}$, the algorithm outputs \textsc{Accept}\xspace with probability at least $2/3$; \item If $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, the algorithm outputs \textsc{Reject}\xspace with probability at least $2/3$. \end{itemize} \end{definition} The Dvoretzky-Kiefer-Wolfowitz (DKW) inequality gives a generic algorithm for learning any distribution with respect to the Kolmogorov distance~\cite{DvoretzkyKW56}. \begin{lemma}{(See~\cite{DvoretzkyKW56},\cite{Massart90})} \label{lem:dkw} Suppose we have $n$ \textit{i.i.d.} samples $X_1, \dots X_n$ from a distribution with CDF $F$. Let $F_n(x) {\hfill\blksquare}\medskip \frac{1}{n}\sum_{i=1}^n \mathbf{1}_{\{X_i \leq x\}}$ be the empirical CDF. Then $\Pr[d_{\mathrm K}(F,F_n) \geq \varepsilon] \leq 2e^{-2n\varepsilon^2}$. In particular, if $n = \Omega((1/\varepsilon^2) \cdot \log(1/\delta))$, then $\Pr[d_{\mathrm K}(F,F_n) \geq \varepsilon] \leq \delta$. \end{lemma} We note the following useful relationships between these distances~\cite{GibbsS02}: \begin{proposition} \label{prop:distance-relations} $d_{\mathrm K}(p,q)^2 \leq d_{\mathrm {TV}}(p,q)^2 \leq \frac14 \chi^2(p,q)$. \end{proposition} In this paper, we will consider the following classes of distributions: \begin{itemize} \item Monotone distributions over $[n]^d$ (denoted by $\mathcal{M}_n^d$), for which $i \lesssim j$ implies $f_i \geq f_j$\footnote{This definition describes monotone non-increasing distributions. By symmetry, identical results hold for monotone non-decreasing distributions.}; \item Unimodal distributions over $[n]$ (denoted by $\mathcal{U}_n$), for which there exists an $i^*$ such that $f_i$ is non-decreasing for $i \leq i^*$ and non-increasing for $i \geq i^*$; \item Log-concave distributions over $[n]$ (denoted by $\mathcal{LCD}_n$), the sub-class of unimodal distributions for which $f_{i-1}f_{i+1} \leq f_i^2$; \item Monotone hazard rate (MHR) distributions over $[n]$ (denoted by $\mathcal{MHR}_n$), for which $i < j$ implies $\frac{f_i}{1 - F_i} \leq \frac{f_j}{1 - F_j}$. \end{itemize} \begin{definition} An \emph{$\eta$-effective support} of a distribution $p$ is any set $S$ such that $p(S) \geq 1 - \eta$. \end{definition} The \emph{flattening} of a function $f$ over a subset $S$ is the function $\bar{f}$ such that $\bar{f}_i= p(S)/|S|$. \begin{definition} \label{def:flattening} Let $p$ be a distribution, and support $I_1, \ldots$ is a partition of the domain. The flattening of $p$ with respect to $I_1, \ldots$ is the distribution $\bar p$ which is the flattening of $p$ over the intervals $I_1, \ldots$. \end{definition} \paragraph{Poisson Sampling} Throughout this paper, we use the standard Poissonization approach. Instead of drawing exactly $m$ samples from a distribution $p$, we first draw $m' \sim \rm Poisson(m)$, and then draw $m'$ samples from $p$. As a result, the number of times different elements in the support of $p$ occur in the sample become independent, giving much simpler analyses. In particular, the number of times we will observe domain element $i$ will be distributed as $\rm Poisson(m\dP_i)$, independently for each $i$. Since $\rm Poisson(m)$ is tightly concentrated around $m$, this additional flexibility comes only at a sub-constant cost in the sample complexity with an inversely exponential in $m$, additive increase in the error probability. \section{Introduction} \label{sec:introduction} The quintessential scientific question is whether an unknown object has some property, i.e. whether a model from a specific class fits the object's observed behavior. If the unknown object is a probability distribution, $p$, to which we have sample access, we are typically called to distinguish from samples whether $p$ belongs to some class $\mathcal{C}$ or whether it is sufficiently far from it. This question has received tremendous attention in Statistics, where test statistics for important properties such as the ones we consider here have been proposed. Nevertheless, the emphasis has been on asymptotic analysis, characterizing rates of convergence of test statistics under null hypotheses, as the number of samples tends to infinity. In contrast, we want to study the small sample regime, studying the following problem: \vspace{2ex} \begin{center} \smallskip \framebox{ \begin{minipage}{13.5cm} $\Pi(\mathcal{C},\epsilon)$: Given a family of distributions $\mathcal{C}$, some $\epsilon>0$, and sample access to an unknown distribution $p$ over some discrete support, how many samples are required to distinguish between $p \in \mathcal{C}$ versus $d_{\mathrm {TV}}(p,\mathcal{C})>\epsilon$? \end{minipage} } \end{center} \vspace{2ex} The problem has been studied intensely in the literature on Property Testing and Sublinear Algorithms, where the emphasis has been on characterizing the optimal tradeoff between $p$'s support size and the accuracy $\epsilon$ in the number of samples. Several results have been obtained, roughly clustering into three groups, where (i) $\mathcal{C}$ is the class of monotone distributions over $[n]$, or more generally a poset---see e.g.~\cite{BatuKR04,Bhattacharyya11}; (ii) $\mathcal{C}$ is the class of independent, or $k$-wise independent distributions over a hypergrid---see, e.g., \cite{batu2001testing,alon2007testing}; and (iii) $\mathcal{C}$ contains a single-distribution $q$, and the problem becomes that of testing whether $p$ equals $q$ or is far from it---see, e.g.~\cite{batu2001testing,Paninski08, valiant2014automatic}. With respect to (iii), \cite{valiant2014automatic} characterize exactly the number of samples required to test identity to each distribution $q$, providing a single tester matching this bound simultaneously for all $q$. Nevertheless, this tester and its precursors are not applicable to the composite identity testing problem that we consider. If our class $\mathcal{C}$ were finite, we could test against each element in the class, albeit this would not necessarily be sample optimal. If our class $\mathcal{C}$ is a continuum though, we need tolerant identity testers, which tend to be more expensive in terms of number of samples, and result in substantially suboptimal testers for the classes we consider. Or we could use approaches related to generalized likelihood ratio test, but their behavior is not well-understood in our regime, and optimizing likelihood over our classes becomes computationally intense. \jnote{Our problem falls in the general framework of property testing~\cite{Goldreich98, Fischer01, Rubinfeld06, Ron08,canonne2015survey}, where the objective is to decide if an object has a certain property, or is \emph{far} from having the property. The objects in our case are distributions, and properties are belongingness to \emph{simple} classes of distributions $\mathcal{C}$. The objective of property testing is to solve such problems with as few samples as possible, and as fast (computationally) as possible. This is different from the extensively studied (See \emph{e.g.,}~\cite{Fisher25,lehmann2006testing}) classic problem of (composite) hypothesis testing in statistics where the number of samples are taken to infinity and error exponents and consistency are studied. } In this paper, we obtain sample-optimal and computationally efficient testers for $\Pi(\mathcal{C},\epsilon)$ for the most basic shape restrictions to a distribution. Our contributions are the following: \begin{enumerate} \item \jnote{For a known distribution $q$ over $[n]$, and samples from an unknown distribution $p$, we show that to distinguishing the cases $(a)$ whether the $\chi^2$ distance between $p$ and $q$ is at most $\varepsilon^2/2$, versus $(b)$ the $\ell_1$ distance between $p$ and $q$ is at least $\varepsilon$ requires $O(\sqrt n/\varepsilon^2)$ samples. As a corollary, we provide simpler arguments to show that identity testing requires $\Theta(\sqrt n/\varepsilon^2)$ samples. } \item For the class $\mathcal{C}=\mathcal{M}_n^d$ of monotone distributions over $[n]^d$ we require an optimal $\Theta\left({n^{d/2} \over \varepsilon^2}\right)$ number of samples, where prior work requires $\Omega\left({\sqrt{n} \log n \over \varepsilon^4}\right)$ samples for $d=1$ and $\tilde{\Omega}\left(n^{d-{1\over 2}} {\rm poly}({1 \over \varepsilon})\right)$ for $d>1$~\cite{BatuKR04,Bhattacharyya11}. Our results improve the exponent of $n$ with respect to $d$, shave all logarithmic factors in $n$, and improve the exponent of $\varepsilon$ by at least a factor of $2$. \begin{enumerate} \item A useful building block and interesting byproduct of our analysis is extending Birg\'e's oblivious decomposition for single-dimensional monotone distributions~\cite{Birge87} to monotone distributions in $d\ge1$, and to the stronger notion of $\chi^2$ distance. See Section~\ref{sec:hypergrid}. \item Moreover, we show that $O(\log(n)^{d})$ samples suffice to learn a monotone distribution over $[n]^d$ in $\chi^2$ distance. See Lemma~\ref{lem:fin-learn-mon} for the precise statement. \end{enumerate} \item For the classes $\mathcal{C}=\mathcal{LCD}_n$, $\mathcal{C}=\mathcal{MHR}_n$ and $\mathcal{C}=\mathcal{U}_n$ of log-concave, monotone-hazard-rate and unimodal distributions over $[n]$, we require an optimal $\Theta\left({\sqrt{n} \over \varepsilon^2}\right)$ number of samples. Our testers for $\mathcal{LCD}_n$ and $\mathcal{C}=\mathcal{MHR}_n$ are to our knowledge the first for these classes for the low sample regime we are studying---see~\cite{hall2005testing} and its references for Statistics literature on the asymptotic regime. Our tester for $\mathcal{U}_n$ improves the dependence of the sample complexity on $\varepsilon$ by at least a factor of $2$ in the exponent, and shaves all logarithmic factors in $n$, compared to testers based on testing monotonicity. \begin{enumerate} \item A useful building block and important byproduct of our analysis are the first computationally efficient algorithms for properly learning log-concave and monotone-hazard-rate distributions, to within $\epsilon$ in total variation distance, from ${\rm poly}(1/\epsilon)$ samples, independent of the domain size $n$. See Corollaries~\ref{cor:learning LCD} and~\ref{cor:learning MHR}. Again, these are the first computationally efficient algorithms to our knowledge in the low sample regime. \cite{ChanDSS13b} provide algorithms for density estimation, which are non-proper, i.e. will approximate an unknown distribution from these classes with a distribution that does not belong to these classes. On the other hand, the Statistics literature focuses on maximum-likelihood estimation in the asymptotic regime---see e.g. \cite{cule2010theoretical} and its references. \end{enumerate} \item For all the above classes we obtain matching lower bounds, showing that the sample complexity of our testers is optimal with respect to $n$, $\varepsilon$ and when applicable $d$. See Section~\ref{sec:lower-bounds}. Our lower bounds are based on extending Paninski's lower bound~\cite{Paninski08}. \end{enumerate} In the heart of our tester lies a novel use of the $\chi^2$ statistic. Naturally, the $\chi^2$ and its related $\ell_2$ statistic have been used in several of the afore-cited results. We propose a new use of the $\chi^2$ statistic enabling our optimal sample complexity. The essence of our approach is to first draw a small number of samples (independent of $n$ for log-concave and monotone-hazard-rate distributions and only logarithmic in $n$ for monotone and unimodal distributions) to approximate the unknown distribution $p$ in $\chi^2$ distance. If $p \in \mathcal{C}$, our learner is required to output a distribution $q$ that is $O(\epsilon)$-close to $\mathcal{C}$ in total variation and $O(\epsilon^2)$-close to $p$ in $\chi^2$ distance. Then a small analysis reduces our testing problem to telling apart the following cases: \begin{itemize} \item $p$ and $q$ are $O(\epsilon^2)$-close in $\chi^2$ distance; this case corresponds to $p \in \mathcal{C}$. \item $p$ and $q$ are $\Omega(\epsilon)$-far in total variation distance; this case corresponds to $d_{\mathrm {TV}}(p,\mathcal{C})>\varepsilon$. \end{itemize} In Section~\ref{sec:testing}, we show that a version of the $\chi^2$ statistic, appropriately excluding certain elements of the support, is sufficiently well-concentrated to distinguish between the above cases. Moreover, the sample complexity of our algorithm is optimal for most classes. \explainindetail{Should we put the chi-sq L1 as a separate result, and say why this is a robust identity testing.} Our tester, combined with the afore-mentioned extension of Birg\'e's decomposition theorem is used in Section~\ref{sec:monotone} to test monotone distributions. See Theorem~\ref{thm:monotone-final} and Corollary~\ref{cor:high-d}. Naturally, there are several bells and whistles that we need to add to the above skeleton to accommodate all classes of distributions that we are considering. For log-concave and monotone-hazard distributions, we are unable to obtain a cheap (in terms of samples) learner that $\chi^2$-approximates the unknown distribution $p$ throughout its support. Still, we can identify a subset of the support where the $\chi^2$-approximation is tight and which captures almost all the probability mass of $p$. We extend our tester to accommodate excluding subsets of the support from the $\chi^2$-approximation. See Theorems~\ref{thm:lcd-main} and~\ref{thm:mhr-main} in Sections~\ref{sec:LCD} and \ref{sec:MHR}, which are in the appendix due to lack of space. Some discussion is provided in Section~\ref{sec:blurb}. For unimodal distributions, we are even unable to identify a large enough subset of the support where the $\chi^2$ approximation is guaranteed to be tight. But we can show that there exists a light enough piece of the support (in terms of probability mass under $p$) that we can exclude to make the $\chi^2$ approximation tight. Given that we only use Chebyshev's inequality to prove the concentration of the test statistic, it would seem that our lack of knowledge of the piece to exclude would involve a union bound and a corresponding increase in the required number of samples. We avoid this through a nice application of Kolmogorov's max inequality in our setting. See Theorem~\ref{thm:unimodality} of Section~\ref{sec:unimodal}, which is in the appendix due to lack of space. Some discussion is provided in Section~\ref{sec:blurb}. \paragraph{Related Work.} We cannot do justice to the role of shape restrictions of probability distributions in probabilistic modeling and testing. It suffices to say that the classes of distributions that we study are fundamental, motivating extensive literature on their learning and testing~\cite{BBBB:72}. In the recent times, there has been work on shape restricted statistics, pioneered by Jon Wellner, and others~\cite{JW:09,BW10sn, BJR11,SumardW14}. Due to the sheer volume of literature in statistics in this field, we will restrict ourselves to those already referenced. As we have mentioned, Statistics has focused on the asymptotic regime as the number of samples tends to infinity. Instead we are considering the low sample regime and are more stringent about the behavior of our testers, requiring $2$-sided guarantees. We want to accept if the unknown distribution is in our class of interest, and also reject if it is far from the class. For this problem, as discussed above, there are few results when $\mathcal{C}$ is a whole class of distributions. Closer related to our paper is the line of papers~\cite{BatuKR04, ACS10, Bhattacharyya11} for monotonicity testing, albeit these papers have sub-optimal sample complexity as discussed above. More recently, Acharya and Daskalakis provide optimal testers for the family of Poisson Binomial Distributions~\cite{AcharyaD15}. Finally, contemporaneous work of Canonne et al~\cite{CanonneDGR15} provide testers for the single-dimensional families of distributions considered here, albeit their sample complexity is suboptimal in both $n$ and $\epsilon$. \insertref{Put a bunch of statistics citations.} \input{preliminaries} \input{overview} \input{testing} \input{monotone} \input{unimodal-main} \input{otherclasses} \input{lower-bounds} \input{experiments} \bibliographystyle{IEEEtran} \section{Moments of the Chi-Squared Statistic} \label{sec:chisq-moments} We analyze the mean and variance of the statistic $$ Z = \sum_{i \in \mathcal{A}} \frac{(X_i - mq_i)^2 - X_i}{mq_i},$$ where each $X_i$ is independently distributed according to $\rm Poisson(\text{$m p_i$})$. We start with the mean: \begin{align*} \expectation{Z} &= \sum_{i \in \mathcal{A}} \expectation{\frac{(X_i - mq_i)^2 - X_i}{mq_i}} \nonumber \\ &= \sum_{i \in \mathcal{A}} \frac{\expectation{X_i^2} - 2mq_i\expectation{X_i} + m^2q_i^2 - \expectation{X_i}}{mq_i} \nonumber \\ &= \sum_{i \in \mathcal{A}} \frac{m^2p_i^2 + m p_i - 2m^2q_ip_i + m^2q_i^2 - m p_i}{mq_i} \nonumber \\ &= m \sum_{i \in \mathcal{A}} \frac{(p_i - q_i)^2}{q_i} \nonumber\\ &= m \cdot \chi^2(p_\mathcal{A},q_\mathcal{A}) \nonumber \end{align*} Next, we analyze the variance. Let $\lambda_i = {\bf E}{X_i}=m p_i$ and $\lambda_i' = m q_i$. \begin{align} \Var{Z} &= \sum_{i \in \mathcal{A}}\frac{1}{\lambda_i'^2}\Var{(X_i - \lambda_i)^2 + 2(X_i - \lambda_i)(\lambda_i - \lambda_i') - (X_i - \lambda_i)} \nonumber \\ &= \sum_{i \in \mathcal{A}}\frac{1}{\lambda_i'^2}\Var{(X_i - \lambda_i)^2 + (X_i - \lambda_i)(2\lambda_i -2\lambda_i' - 1) } \nonumber\\ &= \sum_{i \in \mathcal{A}}\frac{1}{\lambda_i'^2}{\bf E}{(X_i - \lambda_i)^4 + 2(X_i - \lambda_i)^3(2\lambda_i -2\lambda_i' - 1) + (X_i - \lambda_i)^2(2\lambda_i -2\lambda_i' - 1)^2 - \lambda_i^2} \nonumber\\ &= \sum_{i \in \mathcal{A}}\frac{1}{\lambda_i'^2}[3\lambda_i^2 + \lambda_i + 2\lambda_i(2\lambda_i - 2\lambda_i' - 1) + \lambda_i(2\lambda_i - 2\lambda_i' - 1)^2 - \lambda_i^2] \nonumber\\ &= \sum_{i \in \mathcal{A}}\frac{1}{\lambda_i'^2}[2\lambda_i^2 + \lambda_i + 4\lambda_i(\lambda_i - \lambda_i') - 2\lambda_i + \lambda_i(4(\lambda_i - \lambda_i')^2 -4(\lambda_i - \lambda_i') + 1)] \nonumber\\ &= \sum_{i \in \mathcal{A}}\frac{1}{\lambda_i'^2}[2\lambda_i^2 + 4\lambda_i(\lambda_i - \lambda_i')^2 ] \nonumber\\ &= \sum_{i \in \mathcal{A}}\left[2 \frac{p_i^2}{q_i^2}+4m\cdot\frac{p_i\cdot(p_i-q_i)^2}{q_i^2}\right] \end{align} The third equality is by noting the random variable has expectation $\lambda_i $ and the fourth equality substitutes the values of centralized moments of the Poisson distribution. \section{Analysis of our $\chi^2$-Test Statistic} \label{sec:chisq-analysis} We first prove the key lemmas in the analysis of our $\chi^2$-test. \begin{prevproof}{Lemma}{lem:means} The former case is straightforward from (\ref{eqn:mean}) and \ref{prp:in-chisq} of $q$. We turn to the latter case. Recall that $\mathcal{A}= \{i:q_i \geq \varepsilon/50n\}$, and thus $q(\bar \mathcal{A}) \leq \varepsilon/50$. We first show that $d_{\mathrm {TV}}(p_{\mathcal{A}}, q_{\mathcal{A}})\geq \frac{6\varepsilon}{25}$, where $p_{\mathcal{A}}, q_{\mathcal{A}}$ are defined as above and in our slight abuse of notation we use $d_{\mathrm {TV}}(p_{\mathcal{A}}, q_{\mathcal{A}})$ for non-probability vectors to denote $\frac12\|p_{\mathcal{A}} - q_{\mathcal{A}}\|_1.$ Partitioning the support into $\mathcal{A}$ and $\compl{\mathcal{A}}$, we have \begin{align} d_{\mathrm {TV}}(p,q)=d_{\mathrm {TV}}(p_{\mathcal{A}}, q_{\mathcal{A}})+d_{\mathrm {TV}}(p_{\compl{\mathcal{A}}}, q_{\compl{\mathcal{A}}}).\label{eqn:tv-decomp} \end{align} We consider the following cases separately: \begin{itemize} \item {\bf $p(\compl{\mathcal{A}})\le \varepsilon/2$:} In this case, \begin{align} d_{\mathrm {TV}}(p_{\compl{\mathcal{A}}}, q_{\compl{\mathcal{A}}}) = \frac12 \sum_{i \in \compl{\mathcal{A}}} |p_i - q_i| \leq \frac12 (p(\compl{\mathcal{A}})+q(\compl{\mathcal{A}})) \le \frac{1}{2}\left(\frac{\varepsilon}{2} + \frac{\varepsilon}{50}\right) = \frac{13\varepsilon}{50}.\nonumber \end{align} Plugging this in~\eqref{eqn:tv-decomp}, and using the fact that $d_{\mathrm {TV}}(p,q) \geq \varepsilon$ shows that $d_{\mathrm {TV}}(p_{\mathcal{A}}, q_{\mathcal{A}}) \geq \frac{6\varepsilon}{25}$. \item {\bf $p(\compl{\mathcal{A}})> \varepsilon/2$:} In this case, by the reverse triangle inequality, \begin{align} d_{\mathrm {TV}}(p_{\mathcal{A}}, q_{\mathcal{A}}) \geq \frac12 (q(\mathcal{A})-p(\mathcal{A})) \geq \frac12 ((1 - \varepsilon/50) - (1 - \varepsilon/2)) = \frac{6\varepsilon}{25} .\nonumber \end{align} \end{itemize} By the Cauchy-Schwarz inequality, \begin{align} \chi^2(p_{\mathcal{A}}, q_{\mathcal{A}}) &\ge 4\frac{d_{\mathrm {TV}}(p_{\mathcal{A}},q_{\mathcal{A}})^2}{q(\mathcal{A})}\nonumber\\ &\geq \frac{\varepsilon^2}{5}. \nonumber \end{align} We conclude by recalling~\eqref{eqn:mean}. \end{prevproof} \begin{prevproof}{Lemma}{lem:vars} We bound the terms of (\ref{eqn:variance}) separately, starting with the first. \begin{align} 2\sum_{i \in \mathcal{A}} \frac{p_i^2}{q_i^2} &= 2\sum_{i \in \mathcal{A}} \left(\frac{(p_i - q_i)^2}{q_i^2} + \frac{2p_iq_i - q_i^2}{q_i^2}\right) \nonumber \\ &= 2\sum_{i \in \mathcal{A}} \left(\frac{(p_i - q_i)^2}{q_i^2} + \frac{2q_i(p_i - q_i) + q_i^2}{q_i^2}\right) \nonumber\\ &\leq 2n + 2\sum_{i \in \mathcal{A}} \left(\frac{(p_i - q_i)^2}{q_i^2} + 2\frac{(p_i - q_i)}{q_i}\right) \nonumber\\ &\leq 4n + 4\sum_{i \in \mathcal{A}} \frac{(p_i - q_i)^2}{q_i^2} \nonumber\\ &\leq 4n + \frac{200n}{\varepsilon} \sum_{i \in \mathcal{A}} \frac{(p_i - q_i)^2}{q_i}\nonumber\\ &= 4n + \frac{200n}{\varepsilon}\frac{E[Z]}{m} \nonumber\\ &\leq 4n + \frac{1}{100}\sqrt{n} E[Z]\label{eq:first-var-term-in} \end{align} The second inequality is the AM-GM inequality, the third inequality uses that $q_i \geq \frac{\varepsilon}{50n}$ for all $i \in \mathcal{A}$, the last equality uses \eqref{eqn:mean}, and the final inequality substitutes a value $m \geq 20000\frac{\sqrt{n}}{\varepsilon^2}$. The second term can be similarly bounded: \begin{align*} 4m \sum_{i \in \mathcal{A}} \frac{p_i(p_i - q_i)^2}{q_i^2} &\leq 4m \left(\sum_{i \in \mathcal{A}} \frac{p_i^2}{q_i^2}\right)^{1/2}\left(\sum_{i \in \mathcal{A}} \frac{(p_i - q_i)^4}{q_i^2}\right)^{1/2} \\ &\leq 4m \left(4n + \frac{1}{100}\sqrt{n} E[Z] \right)^{1/2}\left(\sum_{i \in \mathcal{A}} \frac{(p_i - q_i)^4}{q_i^2}\right)^{1/2} \\ &\leq 4m \left(2\sqrt{n} + \frac{1}{10}n^{1/4} E[Z]^{1/2}\right)\left(\sum_{i \in \mathcal{A}} \frac{(p_i - q_i)^2}{q_i}\right) \\ &= \left(8\sqrt{n} + \frac{2}{5}n^{1/4} E[Z]^{1/2}\right)E[Z] \\ \end{align*} The first inequality is Cauchy-Schwarz, the second inequality uses (\ref{eq:first-var-term-in}), the third inequality uses the monotonicity of the $\ell_p$ norms, and the equality uses~\eqref{eqn:mean}. Combining the two terms, we get $$\Var{Z} \leq 4n + 9\sqrt{n} {\bf E}{Z} + \frac{2}{5}n^{1/4} {\bf E}{Z}^{3/2} .$$ We now consider the two cases in the statement of our lemma. \begin{itemize} \item When $p \in \mathcal{C}$, we know from Lemma~\ref{lem:means} that ${\bf E}{Z} \leq \frac{1}{500} m\varepsilon^2$. Combined with a choice of $m \geq 20000 \frac{\sqrt{n}}{\varepsilon^2}$ and the above expression for the variance, this gives: $$\Var{Z} \leq \frac{4}{20000^2}m^2\varepsilon^4 + \frac{9}{20000 \cdot 500}m^2\varepsilon^4 + \frac{\sqrt{10}}{12500000}m^2\varepsilon^4 \leq \frac{1}{500000}m^2\varepsilon^4.$$ \item When $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, Lemma~\ref{lem:means} and $m \geq 20000\frac{\sqrt{n}}{\varepsilon^2}$ give: $${\bf E}{Z} \geq \frac{1}{5}m\varepsilon^2 \geq 4000\sqrt{n}.$$ Combining this with our expression for variance we get: $$\Var{Z} \leq \frac{4}{4000^2}{\bf E}{Z}^2 + \frac{9}{4000}{\bf E}{Z}^2 + \frac{2}{5\sqrt{4000}}{\bf E}{Z}^2 \leq \frac{1}{100}{\bf E}{Z}^2.$$ \end{itemize} \end{prevproof} \section{A Robust $\chi^2$-$\ell_1$ Identity Test} \label{sec:testing} { Our main result in the Section is Theorem~\ref{thm:chisq-test}. As an immediate corollary, we obtain the following result on testing whether an unknown distribution is close in $\chi^2$ or far in $\ell_1$ distance to a known distribution. In particular, we show the following: \begin{theorem} \label{thm:rob-iden} For a known distribution $q$, there exists an algorithm with sample complexity \[ O(\sqrt n/\varepsilon^2) \] distinguishes between the cases \begin{itemize} \item $\chi^2(p,q)<\varepsilon^2/10$\ \ \ \ \emph{versus} \item $\|p-q\|>\varepsilon^2$. \end{itemize} with probability at least $5/6$. \end{theorem} This theorem follows from our main result of this section, stated next, slightly more generally for classes of distributions. } \begin{theorem} \label{thm:chisq-test} Suppose we are given $\varepsilon \in (0,1]$, a class of probability distributions $\mathcal{C}$, sample access to a distribution $p$ over $[n]$, and an explicit description of a distribution $q$ with the following properties: \begin{enumerate}[label=\textbf{Property \arabic*.},ref=Property \arabic*,align=left] \item $d_{\mathrm {TV}}(q,\mathcal{C}) \leq \frac{\varepsilon}{2}$.\label{prp:q-tv} \item If $p \in \mathcal{C}$, then $\chi^2(p,q) \leq \frac{\varepsilon^2}{500}$. \label{prp:in-chisq} \end{enumerate} Then there exists an algorithm with the following guarantees: \begin{itemize} \item If $p \in \mathcal{C}$, the algorithm outputs \textsc{Accept}\xspace with probability at least $2/3$; \item If $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, the algorithm outputs \textsc{Reject}\xspace with probability at least $2/3$. \end{itemize} The time and sample complexity of this algorithm are $O\left(\frac{\sqrt{n}}{\varepsilon^2}\right)$. \end{theorem} \begin{remark} \label{rmk} As stated in Theorem~\ref{thm:chisq-test}, \ref{prp:in-chisq} requires that $q$ is $O(\varepsilon^2)$-close in $\chi^2$-distance to $p$ over its entire domain. For the class of monotone distributions, we are able to efficiently obtain such a $q$, which immediately implies sample-optimal learning algorithms for this class. However, for some classes, we cannot learn a $q$ with such strong guarantees, and we must consider modifications to our base testing algorithm. For example, for log-concave and monotone hazard rate distributions, we can obtain a distribution $q$ and a set $S$ with the following guarantees: \begin{itemize} \item If $p \in \mathcal{C}$, then $\chi^2(p_S,q_S) \leq O(\varepsilon^2)$ and $p(S) \geq 1 - O(\varepsilon)$; \item If $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, then $d_{\mathrm {TV}}(p,q) \geq \varepsilon/2$. \end{itemize} In this scenario, the tester will simply pretend the support of $p$ and $q$ is $S$, ignoring any samples and support elements in $[n] \setminus S$. Analysis of this tester is extremely similar to what we present below. In particular, we can still show that the statistic $Z$ will be separated in the two cases. When $p \in \mathcal{C}$, excluding $[n] \setminus S$ will only reduce $Z$. On the other hand, when $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, since $p(S) \geq 1 - O(\varepsilon)$, $p$ and $q$ must still be far on the remaining support, and we can show that $Z$ is still sufficiently large. Therefore, a small modification allows us to handle this case with the same sample complexity of $O(\sqrt{n}/\varepsilon^2)$. A further modification can handle even weaker learning guarantees. We could handle the previous case because the tester ``knows what we don't know'' -- it can explicitly ignore the support over which we do not have a $\chi^2$-closeness guarantee. A more difficult case is when there may be a low measure interval hidden in our effective support, over which $p$ and $q$ have a large $\chi^2$-distance. While we may have insufficient samples to reliably identify this interval, it may still have a large effect on our statistic. A naive solution would be to consider a tester which tries all possible ``guesses'' for this ``bad'' interval, but a union bound would incur an extra logarithmic factor in the sample complexity. We manage to avoid this cost through a careful analysis involving Kolmogorov's max inequality, maintaining the $O(\sqrt{n}/\varepsilon^2)$ sample complexity even in this more difficult case. Being more precise, we can handle cases where we can obtain a distribution $q$ and a set of intervals $S = \{I_1,\dots, I_b\}$ with the following guarantees: \begin{itemize} \item If $p \in \mathcal{C}$, then $p(S) \geq 1 - O(\varepsilon)$, $p(I_j) = \Theta(p(S)/b)$ for all $j \in [b]$, and there exists a set $T \subseteq [b]$ such that $|T| \geq b - t$ (for $t = O(1)$) and $\chi^2(p_R,q_R) \leq O(\varepsilon^2)$, where $R = \cup_T I_j$; \item If $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, then $d_{\mathrm {TV}}(p,q) \geq \varepsilon/2$. \end{itemize} This allows us to additionally test against the class of unimodal distributions. The tester requires that an effective support is divided into several intervals of roughly equal measure. It computes our statistic over each of these intervals, and we let our statistic $Z$ be the sum of all but the largest $t$ of these values. In the case when $p \in \mathcal{C}$, $Z$ will only become smaller by performing this operation. We use Kolmogorov's maximal inequality to show that $Z$ remains large when $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$. More details on this tester are provided in Section~\ref{sec:unimodal-appendix}. \end{remark} \begin{algorithm}[h] \caption{Chi-squared testing algorithm}\label{alg:testing} \begin{algorithmic}[1] \State \textbf{Input:} $\varepsilon$; an explicit distribution $q$; (Poisson) $m$ samples from a distribution $p$, where $N_i$ denotes the number of occurrences of the $i$th domain element. \State $\mathcal{A} \leftarrow \{i:q_i \geq \varepsilon/50n\}$ \State $Z \leftarrow \sum_{i \in \mathcal{A}} \frac{(N_i - mq_i)^2 - N_i}{mq_i}$ \If {$Z \leq m\varepsilon^2/10$} \State \Return \textsc{Accept}\xspace \Else \State \Return \textsc{Reject}\xspace \EndIf \end{algorithmic} \end{algorithm} \begin{prevproof}{Theorem}{thm:chisq-test} Theorem \ref{thm:chisq-test} is proven by analyzing Algorithm \ref{alg:testing}. As shown in Section~\ref{sec:chisq-moments}, $Z$ has the following mean and variance: \begin{equation} {\bf E}{Z} = m \cdot \sum_{i \in \mathcal{A}} \frac{(p_i - q_i)^2}{q_i} = m \cdot \chi^2(p_\mathcal{A},q_\mathcal{A}) \label{eqn:mean} \end{equation} \begin{equation} \Var{Z} = \sum_{i \in \mathcal{A}}\left[2 \frac{p_i^2}{q_i^2}+4m\cdot\frac{p_i\cdot(p_i-q_i)^2}{q_i^2}\right] \label{eqn:variance} \end{equation} where by $p_\mathcal{A}$ and $q_\mathcal{A}$ we denote respectively the vectors $p$ and $q$ restricted to the coordinates in $\mathcal{A}$, and we slightly abuse notation when we write $\chi^2(p_\mathcal{A},q_\mathcal{A})$, as these do not then correspond to probability distributions. Lemma~\ref{lem:means} demonstrates the separation in the means of the statistic $Z$ in the two cases of interest, $i.e.,$ $p \in \mathcal{C}$ versus $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, and Lemma~\ref{lem:vars} shows the separation in the variances in the two cases. These two results are proved in Section~\ref{sec:chisq-analysis}. \begin{lemma} \label{lem:means} If $p \in \mathcal{C}$, then ${\bf E}{Z} \leq \frac{1}{500}m\varepsilon^2$. If $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, then ${\bf E}{Z} \geq \frac{1}{5}m\varepsilon^2$. \end{lemma} \begin{lemma} \label{lem:vars} If $p \in \mathcal{C}$, then $\Var{Z} \leq \frac{1}{500000}m^2\varepsilon^4$. If $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$, then $\Var{Z} \leq \frac{1}{100}E[Z]^2$. \end{lemma} Assuming Lemmas~\ref{lem:means} and~\ref{lem:vars}, Theorem~\ref{thm:chisq-test} is now a simple application of Chebyshev's inequality. When $p \in \mathcal{C}$, we have that $${\bf E}{Z} + \sqrt{3}\Var{Z}^{1/2} \leq \left(\frac{1}{500} + \sqrt{3}\left(\frac{1}{500000}\right)^{1/2}\right)m\varepsilon^2 \leq \frac{1}{200}m\varepsilon^2.$$ Thus, Chebyshev's inequality gives $$\Pr\left[Z \geq m\varepsilon^2/10\right] \leq \Pr\left[Z \geq m\varepsilon^2/200\right] \leq \Pr\left[Z - {\bf E}{Z} \geq \sqrt{3}\Var{Z}^{1/2}\right] \leq \frac13.$$ The case for $d_{\mathrm {TV}}(p,\mathcal{C}) \geq \varepsilon$ is similar. Here, $${\bf E}{Z} - \sqrt{3}\Var{Z}^{1/2} \geq \left(1 - \sqrt{3}\left(\frac{1}{100}\right)^{1/2}\right)E[Z] \geq 3m\varepsilon^2/20.$$ Therefore, \[\Pr\left[Z \leq m\varepsilon^2/10\right] \leq \Pr\left[Z \leq 3m\varepsilon^2/20\right] \leq \Pr\left[Z - {\bf E}{Z} \leq - \sqrt{3}\Var{Z}^{1/2}\right] \leq \frac13.\qedhere\qedhere\] \end{prevproof} \section{$t$-modal Distributions} \label{sec:t-modal} For a distribution, $p$ over $[n]$, $i\in[n]$ is said to be a \emph{mode} if $\left(p(i)-p(i-1)\right)\cdot \left(p(i+1)-p(i)\right)$ are of different signs. The family of distributions with at most $t$ modes are called $t$-modal distributions. Monotone distributions are 0-modal, and unimodal and 1-modal distribution classes. These classes of distributions are extremely useful to model mixture distributions, for example it is well known that mixture of $t$ Gaussians is a $t$-modal. PUT SOME RELATED WORK HERE. Recently, in a personal communication~\cite{CanonneDGR15} it was shown that testing $t$-modal distributions is possible with $\tilde{O}(t\sqrt{n}/\varepsilon^4)$ samples. Their time complexity is not mentioned explicitly, however we believe there is an implementation in polynomial time. Indeed, note that any $t$-modal distribution can be written as a mixture of a $t+1$-mixture of unimodal distributions. Indeed, our results will hold for this more general class of distributions. Since, log-concave and monotone hazard distributions are unimodal, our results also apply to them. \subsection{Structural Results} We observed that for monotone distributions, there is an oblivious decomposition of the domain, such that the flattening of the underlying distribution over these intervals has a small $\chi^2$ distance. We now extend this argument to $t$-modal distributions. Consider any distribution with at most $t$ modes. Let $b$ be a number, to be decided later. Consider a partition of $[n]$ into $L\le 1/b+1/t$ intervals such that: \begin{enumerate} \item For each element $i$ with probability at least $1/b$, there is an $I_\ell=\{i\}$. \item There are at most two intervals with $p(I)\le 1/b$. \item Every other interval $I$ satisfies $p(I)\in[\frac1b,\frac2b]$. \end{enumerate} Let $\bar{\dP}$ be the flattening of $p$ over the intervals satisfying these properties. We first prove our decomposition result for unimodal distributions and then describe appropriate extension to mixtures and more modes. Consider any decomposition given as above, and we remove the intervals $I_j$ for which $p(I_j)/|I_j|<\frac{\varepsilon}{50n}$. This operation essentially removes the tiny probability elements. Our key lemma is as follows. \begin{lemma} Let $C>2$. For a unimodal distribution over $[n]$, there are at most most $\frac{4\log(50n/\varepsilon)}{C\varepsilon}$ intervals $I_j$ satisfy $\frac{\dP_j^+}{\dP_j^-}<(1+\varepsilon/C)$. \end{lemma} \begin{proof} To the contrary, if there are more than $\frac{4\log(50n/\varepsilon)}{C\varepsilon}$ intervals, then at least half of them are on one side of the mode, however this implies that the ratio of the largest probability and smallest probability is at least $(1+\varepsilon/C)^j$, and if $j>\frac{2\log(50n/\varepsilon)}{C\varepsilon}$, is at least $50n/\varepsilon$, contradicting that we have removed all such elements. \end{proof} Let $R{\hfill\blksquare}\medskip \frac{4\log(50n/\varepsilon)}{C\varepsilon}$ denote the highest number of intervals with large ratios. For $A\in [L]$, let $I_A$ denote the subset of intervals with indices in $A$. The following result is now immediate from the previous lemma. \begin{lemma} For any unimodal distribution $p$ over $[n]$, there is a subset of intervals $I_A$, such that \begin{itemize} \item $p(I_A)>1-\varepsilon/50-2R/b$, and \item $\chi^2_{{I_A}}(p,\bar{\dP})\le \varepsilon^2/C^2$. \end{itemize} \end{lemma} Consider the following test. \begin{enumerate} \item Take $O(b\log b)$ samples to obtain the intervals satisfying the condition above. \item Use $O(b/\varepsilon^2)$ samples obtain a $q$ that is flat over the intervals of interest. \item If $q$ is at least $\varepsilon/2$ far from the class, output \textsc{Reject}\xspace. \item Remove all singletons and intervals with $q(I)/|I|<\varepsilon/(50n)$. \item \jnote{Remove the intervals with mass at most $1/10b$-- may not be necessary} \item Let $\mathcal{I}_{\rm rem}$ be the remaining intervals \item Let $Z_j$ be the $\chi^2$ statistic over the $i$th interval, computed with $O(\sqrt{n}/\varepsilon^2)$ samples. \item Remove the $R$ largest $Z_j$'s. \item If $\sum_{j}Z_j>m\varepsilon^2/10$, output \textsc{Reject}\xspace. \item Output \textsc{Accept}\xspace. \end{enumerate} We now prove the accuracy of the algorithm. \begin{claim} If the output $q$ is at most $\varepsilon/2$-far from $\mathcal{C}$, and $p$ is at least $\varepsilon-$ far from $\mathcal{C}$, then $p$ is at least $\varepsilon/5$-far from $q$ over the intervals in $\mathcal{I}_{\rm rem}$. \end{claim} \begin{proof} Over the singletons, the largest error introduced is at most $\varepsilon/4$ since we take $O(b/\varepsilon^2)$ samples, \end{proof} \section{Real Todos} \begin{itemize} \item \sout{Everything} \item Test previous item \item Put in the lower bound derivation from Paninski/PBD's. \item $t$-modal distributions -- Think Think Think \end{itemize} \section{Details on testing Unimodality} \label{sec:unimodal} \label{sec:unimodal-appendix} Recall that to circumvent Birg\'e's decomposition, we want to decompose the interval into disjoint intervals such that the probability of each interval is about $O(1/b)$, where $b$ is a parameter, specified later. In particular we consider a decomposition of $[n]$ with the following properties: \begin{enumerate} \item For each element $i$ with probability at least $1/b$, there is an $I_\ell=\{i\}$. \item There are at most two intervals with $p(I)\le 1/{2b}$. \item Every other interval $I$ satisfies $p(I)\in\left[\frac1{2b},\frac2b\right]$. \end{enumerate} Let $I_1, \ldots, I_L$ denote the partition of $[n]$ corresponding to these intervals. Note that $L= O(b)$. \begin{claim} There is an algorithm that takes $O(b\log b)$ samples and outputs $I_1, \ldots, I_L$ satisfying the properties above. \end{claim} The first step in our algorithm is to estimate the \emph{total probability} within each of these intervals. In particular, \begin{lemma} There is an algorithm that takes $m'=O(b\log b/\varepsilon^2)$ samples from a distribution $p$, and with probability at least 9/10 outputs a distribution $\bar{\dQ}$ that is constant on each $I_L$. Moreover, for any $j$ such that $p(I_j)>1/2b$, $\bar{\dQ}(I_j)\in(1\pm\varepsilon)p(I_j)$. \end{lemma} \begin{proof} Consider any interval $I_j$ with $p(I_j)\ge 1/2b$. The number of samples $N_{I_j}$ that fall in that interval is distributed $Binomial(m', p(I_j)$. Then by Chernoff bounds for $m'>12 b\log b/\varepsilon^2$, \begin{align} \probof{|N_{I_j}-m'p(I_j)|>\varepsilon m'p(I_j) }\le& 2\exp\left(\varepsilon^2m'p(I_j)/2\right)\\ \le & \frac1{b^2}, \end{align} where the last inequality uses the fact that $p(I_j)\ge 1/2b$. \end{proof} The next step is estimate the distance of $q$ from $\mathcal{U}_n$. This is possible by a simple dynamic program, similar to the one used for monotonicity. If the estimated distance is more than $\varepsilon/2$, we output \textsc{Reject}\xspace. Our next step is to remove certain intervals. This will be to ensure that when the underlying distribution is unimodal, we are able to estimate the distribution \emph{multiplicatively} over the remaining intervals. In particular, we do the following preprocessing step: \begin{itemize} \item $A= \emptyset$. \item For interval $I_j$, \begin{itemize} \item If \begin{align} q(I_j) &\notin\left((1-\varepsilon)\cdotq(I_{j+1}), (1+\varepsilon)\cdotq(I_{j+1})\right)\ \ \text{ OR }\\ q(I_j) &\notin\left((1-\varepsilon)\cdotq(I_{j-1}), (1+\varepsilon)\cdotq(I_{j-1})\right), \end{align} add $I_j$ to $A$. \end{itemize} \item Add the (at most 2) intervals with mass at most $1/2b$ to $A$. \item Add all intervals $j$ with $q(I_j)/|I_j|<\varepsilon/50n$ to $A$ \end{itemize} If the distribution is unimodal, we can prove the following about the set of intervals ${A^c}$. \begin{lemma} If $p$ is unimodal then, \begin{itemize} \item $p(I_{A^c})\ge 1-\varepsilon/25-1/b - O\left(\log n/(\varepsilon b)\right).$ \item Except \emph{at most one} interval in ${A^c}$ every other interval $I_j$ satisfies, \[ \frac{\dP_j^+}{\dP_j^-}\le (1+\varepsilon). \] \end{itemize} \end{lemma} If this holds, then the $\chi^2$ distance between $p$ and $q$ constrained to $A^c$, is at most $\varepsilon^2$. This lemma follows from the following result. \begin{lemma} Let $C>2$. For a unimodal distribution over $[n]$, there are at most $\frac{4\log(50n/\varepsilon)}{C\varepsilon}$ intervals $I_j$ that satisfy $\frac{\dP_j^+}{\dP_j^-}<(1+\varepsilon/C)$. \end{lemma} \begin{proof} To the contrary, if there are more than $\frac{4\log(50n/\varepsilon)}{C\varepsilon}$ intervals, then at least half of them are on one side of the mode, however this implies that the ratio of the largest probability and smallest probability is at least $(1+\varepsilon/C)^j$, and if $j>\frac{2\log(50n/\varepsilon)}{C\varepsilon}$, is at least $50n/\varepsilon$, contradicting that we have removed all such elements. \end{proof} We have one additional pre-processing step here. We compute $q(A^c)$ and if it is smaller than $1-\varepsilon/25$, we output \textsc{Reject}\xspace. Suppose there are $L'$ intervals in $A^c$. Then, except at most one interval in $L'$ we know that the $\chi^2$ distance between $p$ and $q$ is at most $\varepsilon^2$ when $p$ is unimodal, and the TV distance between $p$ and $q$ is at least $\varepsilon/2$ over $A^c$. We propose the following simple modification to take into account, the one interval that might introduce a high $\chi^2$ distance in spite of having a small total variation. If we knew the interval, we can simply remove it and proceed. Since we do not know where the interval lies, we do the following. \begin{enumerate} \item Let $Z_j$ be the $\chi^2$ statistic over the $i$th interval in $A^c$, computed with $O(\sqrt{n}/\varepsilon^2)$ samples. \item Let $Z_l$ be the largest among all $Z_j$'s. \item If $\sum_{j, j\ne l}Z_j>m\varepsilon^2/10$, output \textsc{Reject}\xspace. \item Output \textsc{Accept}\xspace. \end{enumerate} The objective of removing the largest $\chi^2$ statistic is our substitute for not knowing the largest interval. We now prove the correctness of this algorithm. \paragraph{Case 1 $p\in UM_n$:} We only concentrate on the final step. The $\chi^2$ statistic over all but one interval are at most $c\cdot m\varepsilon^2$, and the variance is bounded as before. Since we remove the largest statistic, the expected value of the new statistic is \emph{strictly dominated} by that of these intervals. Therefore, the algorithm outputs \textsc{Accept}\xspace with at least the same probability as if we removed the spurious interval. \paragraph{Case 2 $p\notin UM_n$:} This is the hard case to prove for unimodal distributions. We know that the $\chi^2$ statistic is large in this case, and we therefore have to prove that it remains large even after removing the largest test statistic $Z_l$. We invoke Kolmogorov's Maximal Inequality to this end. \begin{lemma}[Kolmogorov's Maximal Inequality] For independent zero mean random variables $X_1,\ldots, X_L$ with finite variance, let $S_\ell=X_1+\ldots X_\ell$. Then for any $\lambda>0$, \begin{align} \probof{\max_{1\le\ell\le L}\left|S_\ell\right|\ge\lambda}\le \frac{1}{\lambda^2}\cdot Var\left(S_L\right). \end{align} \end{lemma} As a corollary, it follows that $\probof{\max_{\ell}|X_\ell|>2\lambda}\le \frac{1}{\lambda^2}\cdot Var\left(S_L\right)$. In the case we are interested in, we let $X_i=Z_\ell-\expectation{Z_\ell}$. Then, similar to the computations before, and the fact that each interval has a small mass, it follows that that the variance of the summation is at most $\expectation{Z_\ell}^2/100$. Taking $\lambda = \expectation{S_L-m\varepsilon^2/3}^2/100$, it follows that the statistic does not fall below to $\sqrt n$. This completes the proof of Theorem~\ref{thm:unimodality}. \section{Testing Unimodality} One striking feature of Birge's result is that the decomposition of the domain is oblivious to the samples, and therefore to the unknown distribution. However, such an oblivious decomposition will not work for the unimodal distribution, since the mode is unknown. Suppose we know where the mode of the unknown distribution might be, then the problem can be decomposed into monotone functions over two intervals. Therefore, in theory, one can modify the monotonicity testing algorithm by iterating over all the possible $n$ modes. Indeed, applying a union bound, it then follows that \begin{theorem}(Follows from Monotone) For $\varepsilon>1/n^{1/4}$, there is an algorithm that takes $O(\frac{n^{1/2}}{\varepsilon^2}\log n)$ samples, and can test unimodality. \end{theorem} \jnote{It is then easy to extend this to $t$-modal distributions in the most trivial way by taking $t\sqrt{n}/\varepsilon^2$ samples. We dont like it, and will strive for the separation of testing and learning.}. However, this is not satisfactory since we were only able to prove a lower bound of $\sqrt{n}/\varepsilon^2$. In order to overcome the logarithmic barrier introduced by the union bound, we resort to a non-oblivious decomposition of the domain. We want to decompose the interval into disjoint intervals such that the probability of each interval is about $O(1/b)$, where $b$ is a parameter, specified later. In particular we consider the following decomposition of $[n]$. \begin{enumerate} \item For each element $i$ with probability at least $1/b$, there is an $I_\ell=\{i\}$. \item There are at most two intervals with $p(I)\le 1/{2b}$. \item Every other interval $I$ satisfies $p(I)\in\left[\frac1{2b},\frac2b\right]$. \end{enumerate} Let $I_1, \ldots, I_L$ denote the partition of $[n]$ corresponding to these intervals. Note that $L= O(b)$. \begin{claim} There is an algorithm that takes $O(b\log b)$ samples and outputs $I_1, \ldots, I_L$ satisfying the properties above. \end{claim} The first step in our algorithm is to estimate the \emph{total probability} within each of these intervals. In particular, \begin{lemma} There is an algorithm that takes $m'=O(b\log b/\varepsilon^2)$ samples from a distribution $p$, and with probability at least 9/100 outputs a distribution $\bar{\dQ}$ that is constant on each $I_L$. Moreover, for any $j$ such that $p(I_j)>1/2b$, $q(I_j)\in(1\pm\varepsilon)q(I_j)$. \end{lemma} \begin{proof} Consider any interval $I_j$ with $p(I_j)\ge 1/2b$. The number of samples $N_{I_j}$ that fall in that interval is distributed $Binomial(m', p(I_j)$. Then by Chernoff bounds for $m'>12 b\log b/\varepsilon^2$, \begin{align} \probof{|N_{I_j}-m'p(I_j)|>\varepsilon m'p(I_j) }\le& 2\exp\left(\varepsilon^2m'p(I_j)/2\right)\\ \le & \frac1{b^2}, \end{align} where the last inequality uses the fact that $p(I_j)\ge 1/2b$. \end{proof} The next step is estimate the distance of $q$ from $UM_{n}$. This is possible by a simple dynamic program. If the estimated distance is more than $\varepsilon/2$, we output \textsc{Reject}\xspace. Our next step is to remove certain intervals. This will be to ensure that when the underlying distribution is unimodal, we are able to estimate the distribution \emph{multiplicatively} over the remaining intervals. In particular, we do the following preprocessing step: \begin{itemize} \item $A= \emptyset$. \item For interval $I_j$, \begin{itemize} \item If \begin{align} q(I_j) &\notin\left((1-\varepsilon)\cdotq(I_{j+1}), (1+\varepsilon)\cdotq(I_{j+1})\right)\ \ \text{ OR }\\ q(I_j) &\notin\left((1-\varepsilon)\cdotq(I_{j-1}), (1+\varepsilon)\cdotq(I_{j-1})\right), \end{align} add $I_j$ to $A$. \end{itemize} \item Add the (at most 2) intervals with mass at most $1/2b$ to $A$. \item Add all intervals $j$ with $q(I_j)/|I_j|<\varepsilon/50n$ to $A$ \end{itemize} If the distribution is unimodal, we can prove the following about the set of intervals ${A^c}$. \begin{lemma} If $p$ is unimodal then, \begin{itemize} \item $p(I_{A^c})\ge 1-\varepsilon/25-1/b - O\left(\log n/(\varepsilon b)\right).$ \item Except \emph{at most one} interval in ${A^c}$ every other interval $I_j$ satisfies, \[ \frac{\dP_j^+}{\dP_j^-}\le (1+\varepsilon). \] \end{itemize} \end{lemma} If this holds, then the $\chi^2$ distance between $p$ and $q$ constrained to $A^c$, is at most $\varepsilon^2$. This lemma follows from the following result. \begin{lemma} Let $C>2$. For a unimodal distribution over $[n]$, there are at most most $\frac{4\log(50n/\varepsilon)}{C\varepsilon}$ intervals $I_j$ satisfy $\frac{\dP_j^+}{\dP_j^-}<(1+\varepsilon/C)$. \end{lemma} \begin{proof} To the contrary, if there are more than $\frac{4\log(50n/\varepsilon)}{C\varepsilon}$ intervals, then at least half of them are on one side of the mode, however this implies that the ratio of the largest probability and smallest probability is at least $(1+\varepsilon/C)^j$, and if $j>\frac{2\log(50n/\varepsilon)}{C\varepsilon}$, is at least $50n/\varepsilon$, contradicting that we have removed all such elements. \end{proof} We have one additional pre-processing step here. We compute $q(A^c)$ and if it is smaller than $1-\varepsilon/25$, we output \textsc{Reject}\xspace. Suppose there are $L'$ intervals in $A^c$. Then, except at most one interval in $L'$ we know that the $\chi^2$ distance between $p$ and $q$ is at most $\varepsilon^2$ when $p$ is unimodal, and the TV distance between $p$ and $q$ is at least $\varepsilon/2$ over $A^c$. We propose the following simple modification to take into account, the one interval that might introduce a high $\chi^2$ distance in spite of having a small total variation. If we knew the interval, we can simply remove it and proceed. Since we do not know where the interval lies, we do the following. \begin{enumerate} \item Let $Z_j$ be the $\chi^2$ statistic over the $i$th interval in $A^c$, computed with $O(\sqrt{n}/\varepsilon^2)$ samples. \item Let $Z_l$ be the largest among all $Z_j$'s. \item If $\sum_{j, j\ne l}Z_j>m\varepsilon^2/10$, output \textsc{Reject}\xspace. \item Output \textsc{Accept}\xspace. \end{enumerate} The objective of removing the largest $\chi^2$ statistic is our substitute for not knowing the largest interval. We now prove the correctness of this algorithm. \paragraph{Case 1 $p\in UM_n$:} We only concentrate on the final step. The $\chi^2$ statistic over all but one interval are at most $c\cdot m\varepsilon^2$, and the variance is bounded as before. Since we remove the largest statistic, the expected value of the new statistic is \emph{strictly dominated} by that of these intervals. Therefore, the algorithm outputs \textsc{Accept}\xspace with at least the same probability as if we removed the spurious interval. \paragraph{Case 2 $p\notin UM_n$:} This is the hard case to prove for unimodal distributions. We know that the $\chi^2$ statistic is large in this case, and we therefore have to prove that it remains large even after removing the largest test statistic $Z_l$. We invoke Kolmogorov's Maximal Inequality to this end. \begin{lemma}[Kolmogorov's Maximal Inequality] For independent zero mean random variables $X_1,\ldots, X_L$ with finite variance, let $S_\ell=X_1+\ldots X_\ell$. Then for any $\lambda>0$, \begin{align} \probof{\max_{1\le\ell\le L}\left|S_\ell\right|\ge\lambda}\le \frac{1}{\lambda^2}\cdot Var\left(S_L\right). \end{align} \end{lemma} As a corollary, it follows that $\probof{\max_{\ell}|X_\ell|>2\lambda}\le \frac{1}{\lambda^2}\cdot Var\left(S_L\right)$. In the case we are interested in, we let $X_i=Z_\ell-\expectation{Z_\ell}$. Then, similar to the computations before, and the fact that each interval has a small mass, it follows that that the variance of the summation is at most $\expectation{Z_\ell}^2/100$. Taking $\lambda = \expectation{S_L-m\varepsilon^2/3}^2/100$, it follows that the statistic does not fall below to $\sqrt n$. \begin{theorem} \label{thm:unimodality} For $\varepsilon>n^{-1/4}$ samples, there is an algorithm that takes $O(\sqrt{n}/\varepsilon^2)$ samples to test unimodality. \end{theorem} \section{Testing Unimodality} \label{sec:unimodal-main} One striking feature of Birg\'e's result is that the decomposition of the domain is oblivious to the samples, and therefore to the unknown distribution. However, such an oblivious decomposition will not work for the unimodal distribution, since the mode is unknown. Suppose we know where the mode of the unknown distribution might be, then the problem can be decomposed into monotone functions over two intervals. Therefore, in theory, one can modify the monotonicity testing algorithm by iterating over all the possible $n$ modes. Indeed, by applying a union bound, it then follows that \begin{theorem}(Follows from Monotone) For $\varepsilon>1/n^{1/4}$, there exists an algorithm for testing unimodality over $[n]$ with sample complexity $O\left(\frac{\sqrt{n}}{\varepsilon^2}\log n\right)$. \end{theorem} However, this is unsatisfactory, since our lower bound (and as we will demonstrate, the true complexity of this problem) is $\sqrt{n}/\varepsilon^2$. We overcome the logarithmic barrier introduced by the union bound, by employing a non-oblivious decomposition of the domain, and using Kolmogorov's max-inequality. Our main result for testing unimodality is the following theorem, which is proved in Section~\ref{sec:unimodal-appendix}. \begin{theorem} \label{thm:unimodality} Suppose $\varepsilon>n^{-1/4}$. Then there exists an algorithm for testing unimodality over $[n]$ with sample complexity $O(\sqrt{n}/\varepsilon^2)$. \end{theorem}

https://arxiv.org/abs/1401.7007

A Weighted Ostrowski Type inequality for L$_{1}\left[ a,b\right] $ and applications

The aim of this paper is to obtain some generalized weighted Ostrowski inequalities for differentiable mappings. Some well known inequalities can be derived as special cases of the inequalities obtained here. In addition, perturbed mid-point inequality and perturbed trapezoid inequality are also obtained. The inequalities obtained here have direct applications in Numerical Integration, Probability Theory, Information Theory and Integral Operator Theory. Some of these applications are discussed.

\section{Introduction} Inequalities appear in most of the domains of Mathematics and has applications in numerical integration, probability theory, information theory and integral operator theory. Inequalities as a field came into promince with the publication of a book by Hardy, Littlewood and Polya \cit {6} in 1934. In 1938, Ostrowski \cite{8} discovered a useful inequality, which is known after his name as Ostrowski inequality. In many practical investigation, it is necessary to bound one quantity by another. This classical Ostrowski inequality is very useful for this purpose\textbf{.} Beckenbach and Bellman \cite{2} \ and Mitrinovi\'{c} \cite{12} highlighted the importance of inequalities in their respective publications.\bigskip More recently new inequalities of Ostrowski type were presented Dragomir and Wang \textbf{\cite{5}} in 1997 and Dragomir and Rassias \cite{4} in 2002. The weighted version of Ostrowski inequality was first presented in 1983 by Pecari\'{c} and Savi\'{c} \cite{9}. In 2003, Roumeliotis \cite{11} did some improvement in the weighted version of Ostrowski-Gr\"{u}ss type inequalities. In \cite{10} and \cite{7}, Qayyum and Hussain discussed the weighted version of Ostrowski-Gr\"{u}ss type inequalities. The tools that are used in this paper are weighted Peano kernel approach which is the classical and extensively used approach in developing Ostrowski integral inequalities. The results presented in this paper are very general in nature. \ The inequalities proved by Dragomir et al \textbf{\cite{5}}, Barnett et al \textbf{\cite{1}} and Cerone et al \textbf{\cite{3}} are special cases of the inequalities developed here. Ostrowski \cite{8} proved the classical integral inequality which is stated here without proof. \begin{theorem} Let \ \ $f\ $: $\left[ a,b\right] \rightarrow \mathbb{R} $ \ be\ continuous on $\left[ a,b\right] $ and differentiable on $\left( a,b\right) ,$ whose derivative $f^{\prime }:\left( a,b\right) \rightarrow \mathbb{R} $ is bounded on $\ \left( a,b\right) ,$ i.e. $\left\Vert f^{\prime }\right\Vert _{\infty }=\sup_{t\in \left[ a,b\right] }\left\vert f^{\prime }\left( t\right) \right\vert <\infty $ the \begin{equation} \left\vert f(x)-\frac{1}{b-a}\int_{a}^{b}f(t)dt\right\vert \leq \left[ \frac 1}{4}+\frac{\left( x-\frac{a+b}{2}\right) ^{2}}{\left( b-a\right) ^{2} \right] \left( b-a\right) \left\Vert f^{\prime }\right\Vert _{\infty } \tag{1.1} \end{equation for all $x\in \left[ a,b\right] $. The\ constant$\ \frac{1}{4}\;$is sharp in the sense that it can not be replaced by a smaller one. \end{theorem} Dragomir and Wang \cite{5} proved $\left( 1.1\right) $ for $f^{\text{ \prime }\in L_{1}\left[ a,b\right] $ $,$ as follows: \begin{theorem} Let \ \ $f\ $:$I\subseteq $ \mathbb{R} \rightarrow \mathbb{R} $ \ be\ a differentiable mapping in $I^{\circ }$ and $a,b\in I^{\circ }$with $a<b.$ If $f^{\text{ }\prime }\in L_{1}\left[ a,b\right] $, then the inequality hold \begin{equation} \left\vert \text{ }f(x)-\frac{1}{b-a}\int_{a}^{b}f(t)dt\right\vert \leq \left[ \frac{1}{2}+\frac{\left\vert x-\frac{a+b}{2}\right\vert }{b-a}\right] \left\Vert f^{\prime }\right\Vert _{1} \tag{1.2} \end{equation for all $x\in \left[ a,b\right] $. \end{theorem} They also pointed out some applications of $(1.2)$ in Numerical Integration as well as for special means. Barnett et,al, \cite{1} proved out an inequality of Ostrowski type for twice differentiable mappings which is in terms of the $\left\Vert .\right\Vert _{1}$ norm of the second derivative $f^{\prime \prime }$ and apply it in numerical integration and for some special means. The following inequality of Ostrowski's type for mappings which are twice differentiable, holds \cite{3}. \begin{theorem} Let \ \ $f\ $: $\left[ a,b\right] \rightarrow \mathbb{R} $ \ be\ continuous on $\left[ a,b\right] $ and twice differentiable in \left( a,b\right) $ and $f^{\text{ }\prime \prime }\in L_{1}\left( a,b\right) .$ Then the inequality obtaine \begin{eqnarray} &&\left\vert \text{ }f(x)-\frac{1}{b-a}\int_{a}^{b}f(t)dt-\left( x-\frac{a+ }{2}\right) f^{\text{ }\prime }\left( x\right) \right\vert \label{1.3} \\ &\leq &\frac{1}{2\left( b-a\right) }\left( \left\vert x-\frac{a+b}{2 \right\vert +\frac{1}{2}\left( b-a\right) \right) ^{2}\left\Vert f^{\prime \prime }\right\Vert _{1} \notag \end{eqnarray for all $x\in \left[ a,b\right] $. \end{theorem} J. Roumeliotis \cite{4}, presented product inequalities and weighted quadrature. The weighted inequlity was also obtained in Lebesgue spaces involving first derivative of the function, which is given b \begin{eqnarray} &&\left\vert \text{ }\frac{1}{b-a}\int_{a}^{b}w\left( t\right) f(t)dt-m\left( a,b\right) f\left( x\right) \right\vert \notag \\ &\leq &\frac{1}{2}\left[ m\left( a,b\right) +\left\vert m\left( a,x\right) -m\left( x,b\right) \right\vert \right] \left\Vert f^{\prime \prime }\right\Vert _{1} \label{1.4} \end{eqnarray Motivated and inspired by the work of the above mentioned renowned mathematicians, we will establish a new inequality by using weight function , which will be better and generalized than those developed in $\left[ 1- \right] .$ Some other interesting inequalities are also presented as special cases. In the last, we presented applications for some special means and in numerical integration. \section{Main Results} In order to prove our main result we first give the following essential definition. We assume that the weight function (or density) $w:(a,b)\longrightarrow \lbrack 0,\infty )$ to be non-negative and integrable over its entire domain an \begin{equation*} \int_{a}^{b}w(t)dt<\infty . \end{equation* The domain of $\ w$\ \ \ may be finite or infinite and may vanish at the boundary point. We denote the momen \begin{equation*} m(a,b)=\int_{a}^{b}w(t)dt. \end{equation* We now give our main result. \begin{theorem} \textit{Let \ }$\mathit{\ }f:[a,b]\rightarrow \mathbb{R} $ \textit{be continuous on }$[a,b]$\textit{\ and differentiable on }$(a,b)$ and satisfy the condition $\theta \leq f^{\text{ }\prime }\leq \Phi $ $\ ,\ x\in (a,b).$ Then we have the inequalit \begin{eqnarray} &&\left\vert f(x)-\frac{1}{m(a,b)}w(x)\left( b-a\right) \left( x-\frac{a+b}{ }\right) f^{\text{ }\prime }(x)-\frac{1}{m(a,b)}\int_{a}^{b}f(t)w(t)dt\righ \vert \notag \\ &\leq &\frac{1}{2m^{2}(a,b)}w(x)\left( \frac{1}{2}\left( b-a\right) ^{2}+2\left( x-\frac{a+b}{2}\right) ^{2}\right) \notag \\ && \notag \\ &&\times \left( \frac{1}{2}\left( b-a\right) +\left\vert x-\frac{a+b}{2 \right\vert \right) \left\Vert f^{\text{ }\prime \prime }\right\Vert _{w,1} \label{2.1} \end{eqnarray for all $x\in \left[ a,b\right] $. \end{theorem} \begin{proof} L\textit{et us define the mapping\ }$P(.,.):[a,b]\longrightarrow \mathbb{R} $ given by \begin{equation*} P(x,t)=\left\{ \begin{array}{lll} \int_{a}^{t}w(u)du\text{ \ } & \text{{if}} & t\in \lbrack a,x] \\ \int_{b}^{t}w(u)du\text{ } & \text{{if}} & t\in (x,b] \end{array \right. \end{equation* Integrating by parts, we hav \begin{equation} P(x,t)f^{\text{ }\prime }(t)dt=\int_{a}^{b}f(x)m(a,b)-\int_{a}^{b}f(t)w(t)dt. \tag{2.2} \end{equation Applying the identity $(2.2)$ for $f^{\text{ }\prime }(.),$ we ge \begin{equation*} f^{\text{ }\prime }(t)=\frac{1}{m(a,b)}\int_{a}^{b}P(t,s)f^{\text{ }\prime \prime }(s)ds+\frac{1}{m(a,b)}\int_{a}^{b}f^{\text{ }\prime }(s)w(s)ds. \end{equation* Substituting $f^{\text{ }\prime }(t)$ in the right membership of $\left( 2.2\right) ,$we hav \begin{eqnarray} f(x) &=&\frac{1}{m^{2}(a,b)}\int_{a}^{b}\int_{a}^{b}P(x,t)P(t,s)f^{\text{ \prime \prime }(s)dsdt \notag \\ &&+\frac{1}{m^{2}(a,b)}\int_{a}^{b}P(x,t)dt\int_{a}^{b}f^{\text{ }\prime }(s)w(s)dsdt+\frac{1}{m(a,b)}\int_{a}^{b}f(t)w(t)dt. \label{2.3} \end{eqnarray Sinc \begin{equation*} \int_{a}^{b}P(x,t)dt=w(x)\left( b-a\right) \left( x-\frac{a+b}{2}\right) \end{equation* an \begin{equation*} \int_{a}^{b}f^{\text{ }\prime }(s)w(s)ds=f^{\text{ }\prime }(x)m\left( a,b\right) . \end{equation* From $\left( 2.3\right) $ therefore we obtai \begin{eqnarray} f(x) &=&\frac{1}{m(a,b)}w(x)\left( b-a\right) \left( x-\frac{a+b}{2}\right) f^{\text{ }\prime }(x)+\frac{1}{m(a,b)}\int_{a}^{b}f(t)w(t)dt \notag \\ &&+\frac{1}{m^{2}(a,b)}\int_{a}^{b}\int_{a}^{b}P(x,t)P(t,s)f^{\text{ }\prime \prime }(s)dsdt. \label{2.4} \end{eqnarray No \begin{equation*} \int_{a}^{b}\text{ }\left\vert P(t,s)\right\vert ds=\frac{1}{2}w(t)\left[ \left( t-a\right) ^{2}+\left( t-b\right) ^{2}\right] , \end{equation* \begin{eqnarray*} &&\int_{a}^{b}\left\vert P(x,t)\right\vert \left[ \frac{w(t)}{2}\left( \left( t-a\right) ^{2}+\left( b-t\right) ^{2}\right) \left\vert f^{\text{ \prime \prime }(s)\right\vert ds\right] dt \\ &\leq &\frac{1}{2}w(x)\left( \left( x-a\right) ^{2}+\left( b-x\right) ^{2}\right) \max \left\{ t-a,b-t\right\} \left\Vert f^{\text{ }\prime \prime }\right\Vert _{w,1}. \end{eqnarray* From $\left( 2.4\right) $, we hav \begin{eqnarray} &&\left\vert f(x)-\frac{1}{m(a,b)}w(x)\left( b-a\right) \left( x-\frac{a+b}{ }\right) f^{\text{ }\prime }(x)-\frac{1}{m(a,b)}\int_{a}^{b}f(t)w(t)dt\righ \vert \notag \\ &\leq &\frac{1}{2m^{2}(a,b)}w(x)\left( \left( x-a\right) ^{2}+\left( b-x\right) ^{2}\right) \max \left\{ t-a,b-t\right\} \left\Vert f^{\text{ \prime \prime }\right\Vert _{w,1}. \label{2.5} \end{eqnarray Usin \begin{equation*} \max \left\{ t-a,b-t\right\} =\frac{1}{2}\left( b-a\right) +\left\vert x \frac{a+b}{2}\right\vert \end{equation* in $\left( 2.5\right) $, we get our desired result. \end{proof} \begin{remark} For $w\left( t\right) =1\mathbf{,}\ \ $the inequality $\left( 2.1\right) $ give \begin{eqnarray} &&\left\vert f(x)-\left( x-\frac{a+b}{2}\right) f^{\text{ }\prime }(x)-\frac 1}{\left( b-a\right) }\int\limits_{a}^{b}f(t)dt\right\vert \notag \\ &\leq &\frac{1}{2\left( b-a\right) ^{2}}\left( \frac{1}{2}\left( b-a\right) ^{2}+2\left( x-\frac{a+b}{2}\right) ^{2}\right) \notag \\ && \notag \\ &&\times \left( \frac{1}{2}\left( b-a\right) +\left\vert x-\frac{a+b}{2 \right\vert \right) \left\Vert f^{\text{ }\prime \prime }\right\Vert _{1} \label{2.6} \end{eqnarray which is similar to \bigskip Barnett's result proved in \cite{1}. \end{remark} \begin{corollary} Under the assumptions of Theorem $4$ and choosing $x=\frac{a+b}{2}$ , we have the perturbed midpoint inequalit \begin{eqnarray} &&\left\vert f\left( \frac{a+b}{2}\right) -\frac{1}{m(a,b) \int_{a}^{b}f(t)w(t)dt\right\vert \notag \\ &\leq &\frac{1}{8m^{2}(a,b)}w(x)\left( b-a\right) ^{3}\left\Vert f^{\text{ \prime \prime }\right\Vert _{w,1}. \label{2.7} \end{eqnarray} \end{corollary} \begin{proof} This follows from inequality $\left( 2.1\right) .$ \end{proof} \begin{corollary} Under the assumptions of Theorem $4$, we have the perturbed trapezoidal inequalit \begin{eqnarray} &&\left\vert \frac{f(a)+f(b)}{2}-\frac{1}{m(a,b)}\int_{a}^{b}f(t)w(t)dt \frac{1}{m(a,b)}\frac{\left( b-a\right) ^{2}}{4}\left( w(a)f^{\text{ }\prime }(a)-w(b)f^{\text{ }\prime }(b)\right) \right\vert \notag \\ &\leq &\frac{1}{4m^{2}(a,b)}\left( b-a\right) ^{3}\left[ w(a)+w(b)\right] \left\Vert f^{\text{ }\prime \prime }\right\Vert _{w,1}. \label{2.8} \end{eqnarray} \end{corollary} \begin{proof} Put $x=a$ and $x=b$ in $\left( 2.1\right) $, summing up the obtained inequalities, using the triangle inequality and dividing by 2, we get the required inequality. \end{proof} \begin{remark} The result given in $\left( 2.8\right) $ is different from the comparable results available in \cite{4}. \end{remark} \begin{remark} We can get the best estimation from the inequality $\left( 2.1\right) $ , only when $x=\frac{a+b}{2}$, this yields the inequality $\left( 2.7\right) .$ It shows that mid point estimation is better than the trapezoid estimation.\ \end{remark} \section{\textbf{Application for some special means}} We may now apply inequality $\left( 2.1\right) ,$ to deduce some inequalities for special means by the use of particular mappings as follows: \begin{remark} Consider $f(x)=\sqrt{x}\ \ln x\ ,x\in \left[ a,b\right] \ \subset \left( 0,\infty \right) $ an \begin{equation*} w(x)=\frac{1}{\sqrt{x}}, \end{equation* The inequality $\left( 2.1\right) $ therefore give \begin{eqnarray} &&\left\vert \begin{array}{c} \sqrt{x}\ln x-\frac{1}{2\left( \sqrt{b}-\sqrt{a}\right) x}\left( b-a\right) \left( x-A\right) \left( 1+\frac{1}{2}\ln x\right) \\ -\frac{1}{2\left( \sqrt{b}-\sqrt{a}\right) }\left( b-a\right) \ln I\left( a,b\right \end{array \right\vert \notag \\ &\leq &\frac{1}{8\left( \sqrt{b}-\sqrt{a}\right) ^{2}}\frac{1}{\sqrt{x} \left( \frac{1}{2}\left( b-a\right) ^{2}+2\left( x-A\right) ^{2}\right) \notag \\ &&\times \left( \frac{1}{2}\left( b-a\right) +\left\vert x-A\right\vert \right) \frac{(b-a)}{4ab}\left( 1-\frac{\ln b^{a}-\ln a^{b}}{b-a}\right) . \label{3.1} \end{eqnarray Choosing $x=A$ in $\left( 3.1\right) $, we ge \begin{eqnarray} &&\left\vert \sqrt{A}\ln A-\frac{1}{2\left( \sqrt{b}-\sqrt{a}\right) }\left( b-a\right) \ln I\left( a,b\right) \right\vert \notag \\ &\leq &\frac{1}{128ab\left( \sqrt{b}-\sqrt{a}\right) ^{2}}\frac{1}{\sqrt{x} \left( b-a\right) ^{4}\left( 1-\frac{\ln b^{a}-\ln a^{b}}{b-a}\right) . \label{3.2} \end{eqnarray \bigskip \end{remark} \begin{remark} Consider \ $f(x)=\frac{1}{x}\sqrt{x}\ \ \ ,x\in \lbrack a,b]\subset \lbrack 1,\infty )$ an \begin{equation*} w(x)=\frac{1}{\sqrt{x}} \end{equation* The inequality $\left( 2.1\right) $ therefore give \begin{eqnarray} &&\left\vert \begin{array}{c} \frac{1}{x}\text{ }\sqrt{x}\text{\ \ }+\frac{1}{4\left( \sqrt{b}-\sqrt{a \right) }\frac{1}{x^{2}}\left( b-a\right) \left( x-A\right) \\ -\frac{1}{2\left( \sqrt{b}-\sqrt{a}\right) }\left( b-a\right) L_{-1}^{-1 \end{array \right\vert \notag \\ &\leq &\frac{1}{8\left( \sqrt{b}-\sqrt{a}\right) ^{2}}\frac{1}{\sqrt{x} \left( \frac{1}{2}\left( b-a\right) ^{2}+2\left( x-A\right) ^{2}\right) \notag \\ &&\times \frac{3}{8}\left( \frac{1}{2}\left( b-a\right) +\left\vert x-A\right\vert \right) \left( \frac{b^{2}-a^{2}}{a^{2}b^{2}}\right) . \label{3.3} \end{eqnarray Choosing $x=A$ in $\left( 3.3\right) $, we ge \begin{eqnarray} &&\left\vert \frac{1}{A}\text{ }\sqrt{A}\text{\ \ }-\frac{1}{2\left( \sqrt{b -\sqrt{a}\right) }\left( b-a\right) L_{-1}^{-1}\right\vert \notag \\ &\leq &\frac{3}{256\left( \sqrt{b}-\sqrt{a}\right) ^{2}}\frac{1}{\sqrt{A} \left( b-a\right) ^{3}\left( \frac{b^{2}-a^{2}}{a^{2}b^{2}}\right) . \label{3.4} \end{eqnarray} \end{remark} \begin{remark} Consider\ \ \ $f(x)=x^{p}\sqrt{x}$ $,$\ $x\in \left[ a,b\right] $ $f:\left( 0,\infty \right) \rightarrow R$\ ,$\ $where $p\in R/\left\{ -1,0\right\} $ then for $a<b,$ an \begin{equation*} w(x)=\frac{1}{\sqrt{x}} \end{equation* The inequality $\left( 2.1\right) $ therefore give \begin{eqnarray} &&\left\vert \begin{array}{c} x^{p}\sqrt{x}\text{ }-\frac{1}{2\left( \sqrt{b}-\sqrt{a}\right) }\frac{1}{x \left( b-a\right) \left( x-A\right) \left( p+\frac{1}{2}\right) x^{p} \\ -\frac{1}{2\left( \sqrt{b}-\sqrt{a}\right) }\left( b-a\right) L_{p}^{p} \end{array \right\vert \notag \\ &\leq &\frac{1}{8\left( \sqrt{b}-\sqrt{a}\right) ^{2}}\frac{1}{\sqrt{x} \left( \frac{1}{2}\left( b-a\right) ^{2}+2\left( x-A\right) ^{2}\right) \notag \\ &&\times \left( \frac{1}{2}\left( b-a\right) +\left\vert x-A\right\vert \right) \left( \frac{p^{2}-\frac{1}{4}}{p-1}\right) \left( b^{p-1}-a^{p-1}\right) . \label{3.5} \end{eqnarray Choosing $x=A$\ in $\left( 3.5\right) ,$ we ge \begin{eqnarray} &&\left\vert A^{p}\sqrt{A}\text{ }-\frac{1}{2\left( \sqrt{b}-\sqrt{a}\right) }\left( b-a\right) L_{p}^{p}\right\vert \notag \\ &\leq &\frac{1}{32\left( \sqrt{b}-\sqrt{a}\right) ^{2}}\frac{1}{\sqrt{A} \left( b-a\right) ^{3}\left( \frac{p^{2}-\frac{1}{4}}{p-1}\right) \left( b^{p-1}-a^{p-1}\right) . \label{3.6} \end{eqnarray} \end{remark} \section{\textbf{An Application to Numerical integration}} Let $I_{n}:$ $a=x_{0}<x_{1}<x_{2}<....<x_{n-1}<x_{n}=b$ be a division of the interval $[a,b]$ and $\ \ \xi =\left( \xi _{0},\xi _{1},......,\xi _{n-1}\right) ,$\ a sequence of intermediate points \ $\xi _{i}\in \left[ x_{i},x_{i+1}\right] $ $\left( i=0,1,.....,n-1\right) .$ Consider the perturbed Riemann sum defined b \begin{equation} A=\underset{i=0}{\overset{n-1}{\sum }}m\left( x_{i},x_{i+1}\right) f\left( \xi _{i}\right) -\underset{i=0}{\overset{n-1}{\sum }}w(\xi _{i})h_{i}\left( \xi _{i}-\frac{x_{i}+x_{i+1}}{2}\right) f^ {\acute{} }\left( \xi _{i}\right) \tag{4.1} \end{equation} \begin{theorem} \textit{Let \ }$\mathit{\ }f:[a,b]\rightarrow \mathbb{R} $ \textit{be continuous on }$[a,b]$\textit{\ and differentiable on }$(a,b)$, such that $f^{\text{ {\acute{} }:\left( a,b\right) \rightarrow \mathbb{R} $ is bounded on $\left( a,b\right) $ and assume that\ $\gamma \leq $\ \ $f$ ^ {\acute{} }\leq \Gamma $ for all \ $x\in \left( a,b\right) .$ $f^{\prime \prime }:(a,b)\longrightarrow \mathbb{R} $ belongs \ to $\mathbf{L}_{1}(a,b),$\ \ i.e \begin{equation*} \left\Vert f^{\prime \prime }\right\Vert _{w,1}:=\int\limits_{a}^{b}\left\vert w(t)f(t)\right\vert dt<\infty . \end{equation* we hav \begin{equation} \int\limits_{a}^{b}f(t)w(t)dt=A\left( f,I,w,\xi \right) +R\left( f,I,w,\xi \right) , \tag{4.2} \end{equation where the remainder $R$ satisfies the estimatio \begin{eqnarray} &&\left\vert R\left( f,I,w,\xi \right) \right\vert \notag \\ &\leq &\frac{\left\Vert f^{\text{ }\prime \prime }\right\Vert _{w,1}} 2m\left( x_{i},x_{i+1}\right) }w(\xi _{i})\left( \frac{1}{2}\left( h_{i}\right) ^{2}+2\left( \xi _{i}-\frac{x_{i}+x_{i+1}}{2}\right) ^{2}\right) \notag \\ &&\times \left( \frac{1}{2}\left( h_{i}\right) +\left\vert \xi _{i}-\frac x_{i}+x_{i+1}}{2}\right\vert \right) \label{4.3} \end{eqnarray for any choice $\xi $ of the intermediate points. \end{theorem} \begin{proof} Apply Theorem $4$ on the interval $[x_{i},x_{i+1}]$,\ \ $\xi _{i}\in \lbrack x_{i},x_{i+1}],$ \ where $h_{i}=x_{i+1}-x_{i}$ $\left( i=1,2,3....n-1\right) ,$ to ge \begin{eqnarray*} &&\left\vert m(x_{i},x_{i+1})\text{ }f(\xi _{i})-\overset{x_{i+1}}{\underset x_{i}}{\int }}f(t)w(t)dt-(\xi _{i}-\frac{x_{i}+x_{i+1}}{2})w(\xi _{i})\left( x_{i+1}-x_{i}\right) f^ {\acute{} }\left( \xi _{i}\right) \right\vert \\ &\leq &\frac{\left\Vert f^{\text{ }\prime \prime }\right\Vert _{w,1}} 2m\left( x_{i},x_{i+1}\right) }w(\xi _{i})\left( \frac{1}{2}\left( x_{i+1}-x_{i}\right) ^{2}+2\left( \xi _{i}-\frac{x_{i}+x_{i+1}}{2}\right) ^{2}\right) \\ &&\times \left( \frac{1}{2}\left( x_{i+1}-x_{i}\right) +\left\vert \xi _{i} \frac{x_{i}+x_{i+1}}{2}\right\vert \right) \end{eqnarray*} for all $\xi _{i}\in \lbrack x_{i},x_{i+1}]$ and $i\in \left( 0,1,....n-1\right) .$ Summing the above two inequalities over $i$ from $0$ to $n-1$ and using the generalized triangular inequality, we get the desired estimation. \end{proof} \section{Conclusions} We established weighted Ostrowski type inequality for bounded differentiable mappings which generalizes the previous inequalities developed and discussed in \cite{1},\cite{3},\cite{5} and \cite{8}. Perturbed midpoint and trapezoid inequalities are obtained. Some closely new results are also given. This inequality is extended to account for applications in some special means and numerical integration to show his applicability towards obtaining direct relationship of these means. These generalized inequalities will also be useful for the researchers working in the field of the approximation theory, applied mathematics, probability theory, stochastic and numerical analysis to solve their problems in engineering and in practical life.

https://arxiv.org/abs/2104.08424

Periodicity of quantum walks defined by mixed paths and mixed cycles

In this paper, we determine periodicity of quantum walks defined by mixed paths and mixed cycles. By the spectral mapping theorem of quantum walks, consideration of periodicity is reduced to eigenvalue analysis of $\eta$-Hermitian adjacency matrices. First, we investigate coefficients of the characteristic polynomials of $\eta$-Hermitian adjacency matrices. We show that the characteristic polynomials of mixed trees and their underlying graphs are same. We also define $n+1$ types of mixed cycles and show that every mixed cycle is switching equivalent to one of them. We use these results to discuss periodicity. We show that the mixed paths are periodic for any $\eta$. In addition, we provide a necessary and sufficient condition for a mixed cycle to be periodic and determine their periods.

\section{Introduction} Quantum walks are quantum analogues of classical random walks \cite{AAKV, ADZ, Gu}. In the last two decades, a great deal of research on quantum walks has been carried out, and they have strong connections with various fields. In quantum information, quantum walk models can be seen as a generalization of Grover's search algorithm \cite{Gr, P}. An important fact in mathematics is the spectral mapping theorem of quantum walks \cite{HKSS2014, KSY}. The spectral mapping theorems reduce eigenvalue analysis of time evolution operators to eigenvalue analysis of other self-adjoint operators. They bring quantum walks into close connection with functional analysis \cite{SS} and spectral graph theory \cite{KST}. In the study of periodicity of discrete-time quantum walk, field theory and algebraic number theory have also been leveraged \cite{KKKS, SMA}. Studies of perfect state transfer in continuous-time quantum walks have also been done by algebraic graph theory and algebraic combinatorics. We refer to Godsil's survey \cite{G2012}. Recent studies on state transfer are in \cite{BMW, LW, LLZZ, MDDT, WL, Z}. \subsection{Related works to periodicity} The topic discussed in this paper is periodicity of discrete-time quantum walks. The works of \cite{HKSS2017, KoST} triggered off studies of periodicity on various graphs. For example, the studies done on Grover walks can be summarized in Table~\ref{80}. \begin{table}[h] \centering \begin{tabular}{|c|c|} \hline Graphs & Ref. \\ \hline \hline Complete graphs, complete bipartite graphs, SRGs & \cite{HKSS2017} \\ \hline Generalized Bethe trees & \cite{KSTY2018} \\ \hline Distance regular graphs & \cite{Y2019} \\ \hline Cycle (3-state) & \cite{KKKS} \\ \hline Complete graphs with self loops & \cite{IMT} \\ \hline \end{tabular} \caption{Prior works on periodicity of Grover walks on undirected graphs} \label{80} \end{table} In other models, periodicity of Fourier walks has been considered by Saito \cite{S}, and periodicity of staggered walks has been studied in \cite{KSTY2019}. Recently, periodicity of quantum walks with generalized Grover coins has been considered by Sarkar et al~\cite{SMA}. \subsection{Main Results} In this paper, we study periodicity of mixed graphs. There are three main theorems. See later sections for more detailed terms and definitions. First, we generalize the result in \cite{AGNN} related to classification of mixed cycles to $\eta$-Hermitian adjacency matrices. In \cite{AGNN}, Akbari et al~provided several typical switching functions and classified mixed cycles into three types. We give similar considerations in $\eta$-Hermitian adjacency matrices. Among the four types of switching defined in \cite{AGNN}, one switching cannot be used in $\eta$-Hermitian adjacency matrices. Due to this, we show that there are at most $n+1$ switching equivalence classes of mixed cycles in $\eta$-Hermitian adjacency matrices. The claim is as follows: \begin{thm} \label{Main1} {\it Let $G = (V, \MC{A})$ be a mixed cycle of length $n$. Then, there exists $j \in \{0,1,\dots, n\}$ such that $G$ and $C_n^j$ is $H_{\eta}$-cospectral. Moreover, we have \[ \det H_{\eta}(C_n^j) = (-1)^{n+1} 2 \cos (\eta j) + (-1)^{\lfloor \frac{n}{2} \rfloor} (1 + (-1)^n). \] } \end{thm} The second main theorem is to determine periodicity of mixed paths. Using the model defined in \cite{KST}, we study periodicity of quantum walks defined by mixed graphs. This model is defined by both a mixed graph and a real number $\eta \in [0, 2\pi)$. The claim is as follows: \begin{thm} \label{Main2} {\it Let $G = (V, \MC{A})$ be a mixed path on $n$ vertices equipped with an $\eta$-function $\theta$. Then $G$ is periodic for any $\eta \in [0, 2\pi)$, and the period is $2(n-1)$. } \end{thm} The third main theorem is to determine periodicity of mixed cycles. Since periodicity is determined by eigenvalues, it is sufficient to consider the only $n+1$ types of mixed cycles by the first main theorem. The claim is as follows: \begin{thm} \label{Main3} {\it Let $G = (V, \MC{A})$ be a mixed cycle on $n$ vertices equipped with an $\eta$-function $\theta$. Then, $G$ is periodic if and only if $\eta \in \MB{Q}\pi$. In addition, we suppose that $\eta \in \MB{Q}\pi$ and the mixed cycle $G$ is $H_{\eta}$-cospectral with $C_n^j$. Let $\eta = \frac{p}{q}\pi$, where $p$ and $q$ are coprime. Then, the period $\tau$ of $G$ is the following: \begin{equation} \label{65m} \tau = \begin{cases} \frac{2qn}{(j, 2q)} \quad &\text{if $p$ is odd,} \\ \frac{qn}{(j, q)} \quad &\text{if $p$ is even.} \\ \end{cases} \end{equation} } \end{thm} This paper is organized as follows. In Section~\ref{102}, we prepare terminologies on spectral graph theory. The definitions of mixed graphs and their $\eta$-Hermitian adjacency matrices are provided. In Section~\ref{11}, coefficients of the characteristic polynomials of $\eta$-Hermitian adjacency matrices are discussed. We focus on permutations that contribute to values of determinants. Relationship between the characteristic polynomials of a mixed graph and its underlying graph is clarified. In Section~\ref{104}, we carry out classification of mixed cycles by $\eta$-Hermitian adjacency matrices. We introduce $n+1$ types of mixed cycles and show that every mixed cycle is switching equivalent to one of them. On the other hand, we show that the $n+1$ types of mixed cycles have different eigenvalues except for a finite number of $\eta$. In Section~\ref{105}, we prepare quantum walks defined by mixed graphs. We define periodicity of mixed graphs and provide some characterizations of it. In Section~\ref{106}, we discuss periodicity of mixed paths. We observe action of time evolution matrices on the unit vectors and provide a visual proof. In Section~\ref{107}, we discuss periodicity of mixed cycles. It is easy to provide a necessary and sufficient condition for a mixed cycle to be periodic, but determination of the period is a bit complicated. \section{Preliminaries on spectral graph theory} \label{102} Let $\Gamma =(V, E)$ be a finite simple and connected graph with the vertex set $V$ and the edge set $E$. For $x \in V$, the set of neighbors of $x$ is denoted by $N(x)$. Define $\MC{A} = \MC{A}(\Gamma)=\{ (x, y), (y, x) \mid xy \in E(\Gamma) \}$, which is the set of the symmetric arcs of $\Gamma$. The origin and terminus vertices of $a=(x, y) \in \MC{A}$ are denoted by $o(a), t(a)$, respectively. We express $a^{-1}$ as the inverse arc of $a$. A {\it mixed graph} $G$ consists of a finite set $V(G)$ of vertices together with a subset $\MC{A}(G) \subset V(G) \times V(G) \setminus \{ (x,x) \mid x \in V \}$ of ordered pairs called {\it arcs}. Let $G$ be a mixed graph. Define $\MC{A}^{-1}(G) = \{ a^{-1} \mid a \in \MC{A}(G) \}$ and $\MC{A}^{\pm}(G) = \MC{A}(G) \cup \MC{A}^{-1}(G)$. If there is no danger of confusion, we write $\MC{A}(G)$ as $\MC{A}$ simply. If $(x,y) \in \MC{A} \cap \MC{A}^{-1}$, we say that the unordered pair $\{x,y\}$ is a {\it digon} of $G$. For a vertex $x \in V(G)$, define $\deg_{G} x = \deg_{G^{\pm}} x$. A mixed graph $G$ is $k$-regular if $\deg_{G}x = k$ for any vertex $x \in V(G)$. The graph $G^{\pm} = (V(G), \MC{A}^{\pm})$ is so-called the {\it underlying graph} of a mixed graph $G$, and this is regarded as an undirected graph depending on context. On the other hand, we equate an undirected graph with a mixed graph by considering undirected edges $xy$ as bidirectional arcs $(x,y), (y,x)$. Throughout this paper, we assume that mixed graphs are weakly connected, i.e., we assume that $G^{\pm}$ is connected. Let $G = (V, \MC{A})$ be a mixed graph. For $\eta \in [0, 2\pi)$, the {\it $\eta$-Hermitian adjacency matrix} $H_{\eta} = H_{\eta}(G) \in \MB{C}^{V \times V}$ is defined by \[ (H_{\eta})_{x,y} = \begin{cases} 1 \qquad &\text{if $(x,y) \in \MC{A} \cap \MC{A}^{-1}$,} \\ e^{\eta i} \qquad &\text{if $(x,y) \in \MC{A} \setminus \MC{A}^{-1}$,} \\ e^{-\eta i} \qquad &\text{if $(x,y) \in \MC{A}^{-1} \setminus \MC{A}$,} \\ 0 \qquad &\text{otherwise.} \end{cases} \] When $\eta = \frac{\pi}{2}$, the matrix $H_{\frac{\pi}{2}}$ is nothing but the Hermitian adjacency matrix. This is introduced by Guo--Mohar \cite{GM} and Li--Liu \cite{LL}, independently. When $\eta = \frac{\pi}{3}$, the matrix $H_{\frac{\pi}{3}}$ is called the Hermitian adjacency matrix of the second kind. This is introduced by Mohar \cite{M}. We refer to \cite{AAS, GS, LY} as recent studies on Hermitian adjacency matrices. Note that $H_{\eta}(G^{\pm})$ coincides with the ordinary adjacency matrix of $G^{\pm}$. Define the {\it degree matrix} $D = D(G) \in \MB{C}^{V \times V}$ by $D_{x,y} = (\deg_{G}x)\delta_{x,y}$ for vertices $x,y \in V(G)$, where $\delta_{x,y}$ is the Kronecker delta symbol. For $\eta \in [0, 2\pi)$, the {\it normalized $\eta$-Hermitian adjacency matrix} $\tilde{H}_{\eta}$ is defined by \[ \tilde{H}_{\eta} = D^{-\frac12} H_{\eta} D^{-\frac12}. \] Note that if a mixed graph $G$ is $k$-regular, we have $\tilde{H}_{\eta} = \frac{1}{k} H_{\eta}$. Let $G$ be a mixed graph. The list of the eigenvalues of $H_{\eta}(G)$ together with their multiplicities, denoted by $\Spec(H_{\eta}(G))$, is called {\it $H_{\eta}$-spectrum} of $G$. The same is on $\tilde{H}_{\eta}$. \section{Permutations and characteristic polynomials} \label{11} Let $\Gamma$ be an undirected graph. A mixed graph $G$ is said to be a {\it mixed $\Gamma$} if $G^{\pm}$ is isomorphic to $\Gamma$. Similarly, we say that $G$ is a {\it mixed tree} if $G^{\pm}$ is a tree. Let $G = (V, \MC{A})$ be a mixed graph, and let $x_1,x_2, \dots, x_l \in V$. We say that a sequence $C = (x_1,x_2, \dots, x_l)$ is an {\it $l$-cycle} in $G$ if $C$ is an $l$-cycle in $G^{\pm}$. The {\it girth} of $G$ is defined by the girth of $ G^{\pm}$. Note that if the girth of $G$ is $s+1$, then it has no $l$-cycle for $l \in \{1,2, \dots, s\}$. Let $G = (V,\MC{A})$ be a mixed graph. Put $X = \lambda I - H_{\eta}(G)$. Then \[ \det X = \sum_{\sigma \in \Sym(V)} \sgn(\sigma) \prod_{x \in V} X_{x, \sigma(x)}, \] where $\Sym(V)$ is the set of all permutations of $V$. Let \[ \MC{P}(X) := \left\{ \sigma \in \Sym(V) \, \middle | \, \prod_{x \in V} X_{x, \sigma(x)} \neq 0 \right\}. \] In addition, we define \[ \MC{P}_m(X) := \Big\{ \sigma \in \MC{P}(X) \, \Big | \, |\{ x \in V \mid \sigma(x) \neq x \} | = m \Big\} \] for $m \in \MB{N}$. Any permutation $\sigma \in \Sym(V)$ can be expressed as the product of disjoint cyclic permutations, say $\sigma = \sigma_1 \sigma_2 \cdots \sigma_l$ for some $l \in \MB{N}$. We call each $\sigma_i$ a {\it factor} of $\sigma$. \begin{lem} \label{00} {\it Let $G = (V, \MC{A})$ be a mixed graph with girth $s+1$, and let $X = \lambda I - H_{\eta}(G)$. For $l \in \{1,2, \dots, s\}$, all factors of $\sigma \in \MC{P}_{l}(X)$ are transpositions, where $n = |V|$. } \end{lem} \begin{proof} Let $H_{\eta} = H_{\eta}(G)$. Suppose $\sigma \in \MC{P}_{l}(X)$ has a factor $\sigma_i$ of length $t > 2$. Display as $\sigma_i = (x_1x_2 \cdots x_t)$. Then $(H_{\eta})_{x_1, x_2} (H_{\eta})_{x_2, x_3} \cdots (H_{\eta})_{x_t, x_1} \neq 0$. However, $G$ has no $t$-cycles since $t \leq l \leq s$. Thus, at least one of $(H_{\eta})_{x_1, x_2}, (H_{\eta})_{x_2, x_3}, \dots, (H_{\eta})_{x_t, x_1}$ is $0$. This is a contradiction. \end{proof} On the other hand, for distinct vertices $x,y$ in a mixed graph $G$, we have \begin{equation} \label{01} (H_{\eta})_{x,y} (H_{\eta})_{y,x} = |(H_{\eta})_{x,y}|^2 = \begin{cases} 1 \qquad &\text{if $x$ is adjacent to $y$ in $G^{\pm}$,} \\ 0 \qquad &\text{otherwise} \end{cases} \end{equation} since $H_{\eta}$ is Hermitian. Remarking this, we have the following. \begin{pro} \label{10} {\it Let $G = (V, \MC{A})$ be a mixed graph with girth $s+1$. Let \begin{align*} \det (\lambda I - H_{\eta}(G)) &= \lambda^n + a_1 \lambda^{n-1} + \cdots + a_{n-1} \lambda + a_n, \\ \det (\lambda I - H_{\eta}(G^{\pm})) &= \lambda^n + b_1 \lambda^{n-1} + \cdots + b_{n-1} \lambda + b_n, \end{align*} where $n = |V|$. Then we have $a_l = b_l$ for any $l \in \{1,2, \dots, s\}$. } \end{pro} \begin{proof} Put $X = \lambda I - H_{\eta}(G)$ and $Y = \lambda I - H_{\eta}(G^{\pm})$. For $x, y \in V$, $X_{x,y} \neq 0$ if and only if $Y_{x,y} \neq 0$, so $\MC{P}(X) = \MC{P}(Y)$. In particular, $\MC{P}_{l}(X) = \MC{P}_{l}(Y)$. If $\MC{P}_{l}(X) = \emptyset$, we have $a_l = b_l = 0$. We consider $\MC{P}_{l}(X) \neq \emptyset$. Let $\sigma \in \MC{P}_{l}(X)$. By Lemma~\ref{00}, all factors of $\sigma$ are transpositions. Display as $\sigma = (x_1 y_1)(x_2 y_2) \cdots (x_t y_t)$, where $t = l / 2$. By (\ref{01}), \begin{align*} \sgn(\sigma) \prod_{x \in V} X_{x, \sigma(x)} &= \lambda^{n-2t} (H_{\eta}(G))_{x_1, y_1} (H_{\eta}(G))_{y_1, x_1} \cdots (H_{\eta}(G))_{x_t, y_t} (H_{\eta}(G))_{y_t, x_t} \\ &= \lambda^{n-2t} \cdot 1 \\ &= \lambda^{n-2t} (H_{\eta}(G^{\pm}))_{x_1, y_1} (H_{\eta}(G^{\pm}))_{y_1, x_1} \cdots (H_{\eta}(G^{\pm}))_{x_t, y_t} (H_{\eta}(G^{\pm}))_{y_t, x_t} \\ &= \sgn(\sigma) \prod_{x \in V} Y_{x, \sigma(x)}. \end{align*} Therefore, \[ a_l = \sum_{\sigma \in \MC{P}_{l}(X)} \sgn(\sigma) \prod_{x \in V} X_{x, \sigma(x)} = \sum_{\sigma \in \MC{P}_{l}(Y)} \sgn(\sigma) \prod_{x \in V} Y_{x, \sigma(x)} = b_l. \] \end{proof} \begin{cor} \label{60} {\it Let $G$ be a mixed tree. Then \[ \det (\lambda I - H_{\eta}(G)) = \det (\lambda I - H_{\eta}(G^{\pm})), \] i.e., $G$ and $G^{\pm}$ are $H_{\eta}$-cospectral. } \end{cor} \begin{proof} Since $G$ is a mixed tree, the girth is $\infty$. By Proposition~\ref{10}, all coefficients of both characteristic polynomials are equal. \end{proof} For distinct vertices $x,y$ in a mixed graph $G$, we also have \[ (\tilde{H}_{\eta})_{x,y} (\tilde{H}_{\eta})_{y,x} = |(\tilde{H}_{\eta})_{x,y}|^2 = \begin{cases} \frac{1}{\deg x \deg y} \qquad &\text{if $x$ is adjacent to $y$ in $G^{\pm}$,} \\ 0 \qquad &\text{otherwise.} \end{cases} \] Therefore, the same as Propsition~\ref{10} holds for $\tilde{H}_{\eta}$. \begin{pro} \label{12} {\it Let $G = (V, \MC{A})$ be a mixed graph with girth $s+1$. Let \begin{align*} \det (\lambda I - \tilde{H}_{\eta}(G)) &= \lambda^n + a_1 \lambda^{n-1} + \cdots + a_{n-1} \lambda + a_n, \\ \det (\lambda I - \tilde{H}_{\eta}(G^{\pm})) &= \lambda^n + b_1 \lambda^{n-1} + \cdots + b_{n-1} \lambda + b_n, \end{align*} where $n = |V|$. Then we have $a_l = b_l$ for any $l \in \{1,2, \dots, s\}$. In particular, if $G$ is a mixed tree, then \[ \det (\lambda I - \tilde{H}_{\eta}(G)) = \det (\lambda I - \tilde{H}_{\eta}(G^{\pm})), \] i.e., $G$ and $G^{\pm}$ are $\tilde{H}_{\eta}$-cospectral. } \end{pro} \section{Classification of mixed cycles by $H_{\eta}$-spectra} \label{104} Let $G$ and $G'$ be mixed graphs. We say that $G$ and $G'$ are {\it $H_{\eta}$-cospectral} if they have the same $H_{\eta}$-spectrum. The relation that two mixed graphs are $H_{\eta}$-cospectral is an equivalence relation. We call its equivalence class {\it $H_{\eta}$-cospectral class}. In this section, we determine the equivalence classes in the mixed cycles. Let $G = (V, \MC{A})$ be a mixed graph. A function $\alpha : V \to \{ 1, e^{\pm i \eta} \}$ is called a {\it switching function}. For a switching function $\alpha$, we define the matrix $D(\alpha) \in \MB{C}^{V \times V}$ by $D(\alpha)_{x,y} = \alpha(x) \delta_{x,y}$. Taking a switching function $\alpha$ well, the matrix $D(\alpha) H_{\eta}(G) D(\alpha)^*$ is the $\eta$-Hermitian adjacency matrix of another mixed graph $G'$. Then, we say that {\it $G'$ is obtained by switching with respect to $\alpha$ from $G$}. Clearly, $G$ and $G'$ are $H_{\eta}$-cospectral. Note that if a mixed graph $G'$ is obtained by switching with respect to $\alpha$ from a mixed graph $G$, we also say that $G$ and $G'$ are {\it switching equivalent}. Recent studies related to switching equivalence of mixed graphs are in \cite{KB, WY}. \subsection{Switching functions} In \cite{AGNN}, Akbari et al~defined the four typical switching functions and determined the $H_{\frac{\pi}{2}}$-cospectral classes in the mixed cycles. We generalize their result to general $\eta \in [0, 2\pi)$. Let $G = (V, \MC{A})$ be a mixed cycle. First, we define the three typical switching functions as follows: \begin{itemize} \item[Sw.2$'$.] For a vertex $x \in V$, define the switching function \[ \alpha(v) = \begin{cases} e^{i \eta} \qquad &\text{if $v = x$,} \\ 1 \qquad &\text{otherwise}. \end{cases} \] Let $N_{G^{\pm}}(x) = \{v_1, v_2\}$. If $(v_1, x), (v_2, x) \in \MC{A} \setminus \MC{A}^{-1}$, then we have the mixed graph $(V, \MC{A} \cup \{ (v_1, x)^{-1}, (v_2, x)^{-1} \})$ by switching. \begin{figure}[h] \begin{center} \begin{tikzpicture} [scale = 0.5, fat/.style={circle,fill=black, inner sep = 2mm}, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (s1) at (-2,2) [label = below:$v_1$] {}; \node[slim] (s2) at (0,3) [label = above:$x$] {}; \node[slim] (s3) at (2,2) [label = below:$v_2$] {}; \draw[line width = 1pt][->] (s1) to (s2); \draw[line width = 1pt][->] (s3) to (s2); \end{tikzpicture} \raisebox{10mm}{$\xrightarrow{\text{Sw.2$'$ on $x$}}$} \begin{tikzpicture} [scale = 0.5, fat/.style={circle,fill=black, inner sep = 2mm}, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (s1) at (-2,2) [label = below:$v_1$] {}; \node[slim] (s2) at (0,3) [label = above:$x$] {}; \node[slim] (s3) at (2,2) [label = below:$v_2$] {}; \draw[line width = 1pt] (s1) -- (s2) -- (s3); \end{tikzpicture} \caption{Sw.2$'$ on the vertex $x$} \label{sw2} \end{center} \end{figure} \item[Sw.3$'$.] For a vertex $x \in V$, define the switching function \[ \alpha(v) = \begin{cases} e^{-i \eta} \qquad &\text{if $v = x$,} \\ 1 \qquad &\text{otherwise}. \end{cases} \] Let $N_{G^{\pm}}(x) = \{v_1, v_2\}$. If $(x, v_1), (x, v_2) \in \MC{A} \setminus \MC{A}^{-1}$, then we have the mixed graph $(V, \MC{A} \cup \{ (x, v_1)^{-1}, (x, v_2)^{-1} \})$ by switching. \begin{figure}[h] \begin{center} \begin{tikzpicture} [scale = 0.5, fat/.style={circle,fill=black, inner sep = 2mm}, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (s1) at (-2,1) [label = below:$v_1$] {}; \node[slim] (s2) at (0,2) [label = above:$x$] {}; \node[slim] (s3) at (2,1) [label = below:$v_2$] {}; \draw[line width = 1pt][->] (s2) to (s1); \draw[line width = 1pt][->] (s2) to (s3); \end{tikzpicture} \raisebox{10mm}{$\xrightarrow{\text{Sw.3$'$ on $x$}}$} \begin{tikzpicture} [scale = 0.5, fat/.style={circle,fill=black, inner sep = 2mm}, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (s1) at (-2,1) [label = below:$v_1$] {}; \node[slim] (s2) at (0,2) [label = above:$x$] {}; \node[slim] (s3) at (2,1) [label = below:$v_2$] {}; \draw[line width = 1pt] (s1) -- (s2) -- (s3); \end{tikzpicture} \caption{Sw.3$'$ on the vertex $x$} \label{sw3} \end{center} \end{figure} \item[Sw.4$'$.] For a vertex $x \in V$, define the switching function \[ \alpha(v) = \begin{cases} e^{i \eta} \qquad &\text{if $v = x$,} \\ 1 \qquad &\text{otherwise}. \end{cases} \] Let $N_{G^{\pm}}(x) = \{v_1, v_2\}$. If $(v_1, x) \in \MC{A} \setminus \MC{A}^{-1}$ and $(x, v_2) \in \MC{A} \cap \MC{A}^{-1}$, then we have the mixed graph $(V, (\MC{A} \setminus \{(v_2, x)\}) \cup \{ (v_1, x)^{-1} \})$ by switching. \begin{figure}[h] \begin{center} \begin{tikzpicture} [scale = 0.5, fat/.style={circle,fill=black, inner sep = 2mm}, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (s1) at (-2,1) [label = below:$v_1$] {}; \node[slim] (s2) at (0,2) [label = above:$x$] {}; \node[slim] (s3) at (2,1) [label = below:$v_2$] {}; \draw[line width = 1pt][->] (s1) to (s2); \draw[line width = 1pt] (s2) -- (s3); \end{tikzpicture} \raisebox{10mm}{$\xrightarrow{\text{Sw.4$'$ on $x$}}$} \begin{tikzpicture} [scale = 0.5, fat/.style={circle,fill=black, inner sep = 2mm}, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (s1) at (-2,1) [label = below:$v_1$] {}; \node[slim] (s2) at (0,2) [label = above:$x$] {}; \node[slim] (s3) at (2,1) [label = below:$v_2$] {}; \draw[line width = 1pt] (s1) -- (s2); \draw[line width = 1pt][->] (s2) to (s3); \end{tikzpicture} \caption{Sw.4$'$ on the vertex $x$} \label{sw4} \end{center} \end{figure} \end{itemize} See also Figures~\ref{sw2}, \ref{sw3} and \ref{sw4}. The above switching functions are named after \cite{AGNN}. The lack of ``Sw.1$'$" is due to the generalization of $\eta$. \subsection{The $n+1$ types of mixed cycles} Let $n \in \MB{N}$ and let $j \in \{0,1, \dots, n\}$. We define the {\it mixed cycle $C_n^{j}$ of type $j$} by \begin{align*} V(C_n^{j}) &= \{ x_1, x_2, \dots, x_n \}, \\ \MC{A}(C_n^{j}) &= \{ (x_1, x_2), \dots, (x_j, x_{j+1}) \} \cup \{ (x_{j+1}, x_{j+2}), \dots, (x_n, x_{n+1}) \}^{\pm}, \end{align*} where we set $x_{n+1} = x_1$. We provide the mixed cycles $C_8^3$ and $C_8^8$ in Figure~\ref{31a} as examples. \begin{figure}[h] \begin{center} \begin{tikzpicture} [scale = 0.5, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (v1) at (0,2) [label = above:$x_1$] {}; \node[slim] (v2) at (1.414,1.414) [label =above right:$x_2$] {}; \node[slim] (v3) at (2,0) [label = right:$x_3$] {}; \node[slim] (v4) at (1.414,-1.414) [label =below right:$x_4$] {}; \node[slim] (v5) at (0,-2) [label = below:$x_5$] {}; \node[slim] (v6) at (-1.414,-1.414) [label = below left:$x_6$] {}; \node[slim] (v7) at (-2,0) [label = left:$x_7$] {}; \node[slim] (v8) at (-1.414,1.414) [label = above left:$x_8$] {}; \draw[line width = 1pt][->] (v2) to (v3); \draw[line width = 1pt] (v4) to (v5); \draw[line width = 1pt] (v6) to (v5); \draw[line width = 1pt] (v7) to (v6); \draw[line width = 1pt] (v7) to (v8); \draw[line width = 1pt] (v8) to (v1); \draw[line width = 1pt][->] (v3) to (v4); \draw[line width = 1pt][->] (v1) to (v2); \end{tikzpicture} $\qquad$ \begin{tikzpicture} [scale = 0.5, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (v1) at (0,2) [label = above:$x_1$] {}; \node[slim] (v2) at (1.414,1.414) [label =above right:$x_2$] {}; \node[slim] (v3) at (2,0) [label = right:$x_3$] {}; \node[slim] (v4) at (1.414,-1.414) [label =below right:$x_4$] {}; \node[slim] (v5) at (0,-2) [label = below:$x_5$] {}; \node[slim] (v6) at (-1.414,-1.414) [label = below left:$x_6$] {}; \node[slim] (v7) at (-2,0) [label = left:$x_7$] {}; \node[slim] (v8) at (-1.414,1.414) [label = above left:$x_8$] {}; \draw[line width = 1pt][->] (v2) to (v3); \draw[line width = 1pt][->] (v4) to (v5); \draw[line width = 1pt][->] (v5) to (v6); \draw[line width = 1pt][->] (v6) to (v7); \draw[line width = 1pt][->] (v7) to (v8); \draw[line width = 1pt][->] (v8) to (v1); \draw[line width = 1pt][->] (v3) to (v4); \draw[line width = 1pt][->] (v1) to (v2); \end{tikzpicture} \caption{The mixed cycles $C_8^3$ and $C_8^8$} \label{31a} \end{center} \end{figure} Note that $C_n^0$ is the undirected cycle of length $n$. We will show that any mixed cycle is $H_{\eta}$-cospectral with some type of the mixed cycle, and different types of mixed cycles have different spectra except for a finite number of $\eta$. \begin{lem} \label{02} {\it Let $G = (V, \MC{A})$ be a mixed cycle of length $n$, and let $x, v_1, v_2, \dots, v_l, y \in V$. If $(x,v_1) \in \MC{A} \setminus \MC{A}^{-1}$, $(v_1, v_2), (v_2, v_3), \dots, (v_{l-1}, v_l) \in \MC{A} \cap \MC{A}^{-1}$, and $(v_l, y) \in \MC{A}^{-1} \setminus \MC{A}$. Then, the mixed cycle $(V, \MC{A} \cup \{ (x,v_1)^{-1}, (v_l, y) \} )$ is obtained by switching from $G$. } \end{lem} \begin{proof} We prove by induction on $l$. Consider $l = 1$. Applying Sw.2$'$ with respect to $v_1$, we have the statement. We suppose that the statement follows in the case of $l - 1$. We apply Sw.4$'$ with respect to $v_1$. The switched mixed cycle is in the situation of the case of $l-1$. By the assumption of the induction, we have the statement. \end{proof} \begin{pro} \label{03} {\it Let $G = (V, \MC{A})$ be a mixed cycle of length $n$. Then, there exists $j \in \{0,1,\dots, n\}$ such that $G$ and $C_n^j$ is $H_{\eta}$-cospectral. } \end{pro} \begin{proof} By Lemma~\ref{02}, we have a switched mixed cycle $G'$ such that the directions of all arcs are aligned clockwise or anticlockwise. Applying Sw.4$'$ many times, arcs in the graph are replaced consecutively. This graph is nothing but $C_n^j$ for some $j \in \{0,1,\dots, n\}$. \end{proof} Figure~\ref{30} shows that switching yields the mixed graph of type $3$ from a mixed cycle. Note that Sw.3$'$ is actually unnecessary. However, it can be used to obtain $C_n^j$ with less switching in some cases. \begin{figure}[h] \begin{center} \begin{tikzpicture} [scale = 0.5, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (v1) at (0,2) [label = above:$v_1$] {}; \node[slim] (v2) at (1.414,1.414) [label =above right:$v_2$] {}; \node[slim] (v3) at (2,0) [label = right:$v_3$] {}; \node[slim] (v4) at (1.414,-1.414) [label =below right:$v_4$] {}; \node[slim] (v5) at (0,-2) [label = below:$v_5$] {}; \node[slim] (v6) at (-1.414,-1.414) [label = below left:$v_6$] {}; \node[slim] (v7) at (-2,0) [label = left:$v_7$] {}; \node[slim] (v8) at (-1.414,1.414) [label = above left:$v_8$] {}; \draw[line width = 1pt][->] (v2) to (v3); \draw[line width = 1pt][->] (v4) to (v5); \draw[line width = 1pt] (v6) to (v5); \draw[line width = 1pt][->] (v7) to (v6); \draw[line width = 1pt] (v7) to (v8); \draw[line width = 1pt][->] (v8) to (v1); \draw[line width = 1pt][->] (v3) to (v4); \draw[line width = 1pt] (v1) -- (v2); \end{tikzpicture}} \raisebox{15mm}{$\xrightarrow[\text{on $v_6$}]{\text{using Sw.4$'$}}$} \begin{tikzpicture} [scale = 0.5, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (v1) at (0,2) [label = above:$v_1$] {}; \node[slim] (v2) at (1.414,1.414) [label =above right:$v_2$] {}; \node[slim] (v3) at (2,0) [label = right:$v_3$] {}; \node[slim] (v4) at (1.414,-1.414) [label =below right:$v_4$] {}; \node[slim] (v5) at (0,-2) [label = below:$v_5$] {}; \node[slim] (v6) at (-1.414,-1.414) [label = below left:$v_6$] {}; \node[slim] (v7) at (-2,0) [label = left:$v_7$] {}; \node[slim] (v8) at (-1.414,1.414) [label = above left:$v_8$] {}; \draw[line width = 1pt][->] (v2) to (v3); \draw[line width = 1pt][->] (v3) to (v4); \draw[line width = 1pt] (v7) to (v6); \draw[line width = 1pt] (v7) to (v8); \draw[line width = 1pt][->] (v8) to (v1); \draw[line width = 1pt] (v1) -- (v2); \draw[line width = 1pt][->] (v4) to (v5); \draw[line width = 1pt][->] (v6) to (v5); \end{tikzpicture}}\\ \raisebox{15mm}{$\xrightarrow[\text{on $v_5$}]{\text{using Sw.2$'$}}$} \begin{tikzpicture} [scale = 0.5, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (v1) at (0,2) [label = above:$v_1$] {}; \node[slim] (v2) at (1.414,1.414) [label =above right:$v_2$] {}; \node[slim] (v3) at (2,0) [label = right:$v_3$] {}; \node[slim] (v4) at (1.414,-1.414) [label =below right:$v_4$] {}; \node[slim] (v5) at (0,-2) [label = below:$v_5$] {}; \node[slim] (v6) at (-1.414,-1.414) [label = below left:$v_6$] {}; \node[slim] (v7) at (-2,0) [label = left:$v_7$] {}; \node[slim] (v8) at (-1.414,1.414) [label = above left:$v_8$] {}; \draw[line width = 1pt][->] (v2) to (v3); \draw[line width = 1pt][->] (v3) to (v4); \draw[line width = 1pt][->] (v8) to (v1); \draw[line width = 1pt] (v1) -- (v2); \draw[line width = 1pt] (v4) -- (v5) -- (v6) -- (v7) -- (v8); \end{tikzpicture}} \raisebox{15mm}{$\xrightarrow[\text{on $v_1$}]{\text{using Sw.4$'$}}$} \begin{tikzpicture} [scale = 0.5, slim/.style={circle,fill=black, inner sep = 0.8mm}, ] \node[slim] (v1) at (0,2) [label = above:$v_1$] {}; \node[slim] (v2) at (1.414,1.414) [label =above right:$v_2$] {}; \node[slim] (v3) at (2,0) [label = right:$v_3$] {}; \node[slim] (v4) at (1.414,-1.414) [label =below right:$v_4$] {}; \node[slim] (v5) at (0,-2) [label = below:$v_5$] {}; \node[slim] (v6) at (-1.414,-1.414) [label = below left:$v_6$] {}; \node[slim] (v7) at (-2,0) [label = left:$v_7$] {}; \node[slim] (v8) at (-1.414,1.414) [label = above left:$v_8$] {}; \draw[line width = 1pt][->] (v2) to (v3); \draw[line width = 1pt][->] (v3) to (v4); \draw[line width = 1pt][->] (v1) to (v2); \draw[line width = 1pt] (v1) -- (v2); \draw[line width = 1pt] (v4) -- (v5) -- (v6) -- (v7) -- (v8) -- (v1); \end{tikzpicture}} \caption{Switching a mixed cycle} \label{30} \end{center} \end{figure} \subsection{Determinants of $H_{\eta}(C_n^j)$} Next, we show that different types of mixed cycles have different spectra. We focus on the constant term of the characteristic polynomial of $H_{\eta}(C_n^j)$. Let $P_n$ be the undirected path graph on $n$ vertices. \begin{lem} \label{50} {\it We have \[ \det H_{\eta}(P_n) = (-1)^{\lfloor \frac{n}{2} \rfloor} \frac{1+(-1)^n}{2}. \] } \end{lem} \begin{proof} We will show that \[ \det H_{\eta}(P_n) = \begin{cases} 1 \qquad &\text{if $n \equiv 0 \pmod 4$,} \\ -1 \qquad &\text{if $n \equiv 2 \pmod 4$,} \\ 0 \qquad &\text{if $n \equiv 1,3 \pmod 4$.} \\ \end{cases} \] The determinant is \[ \det(H_{\eta}(P_{n})) = \begin{vmatrix} 0 & 1 & 0 & \cdots & \cdots & \cdots & 0 \\ 1 & 0 & 1 & 0 & \cdots & \cdots & 0 \\ 0 & 1 & 0 & 1 & 0 & \cdots & 0 \\ \vdots & \ddots & \ddots & \ddots & \ddots & & \vdots \\ \vdots & & \ddots & \ddots & \ddots & \ddots & 0 \\ 0 & \cdots & \cdots & 0 & 1 & 0 & 1 \\ 0 & \cdots & \cdots & \cdots & 0 & 1 & 0 \end{vmatrix}. \] We first apply the cofactor expansion along the first row, and we then apply it again along the first column. We have $\det(H_{\eta}(P_{n})) = - \det(H_{\eta}(P_{n-2}))$. Since $\det H_{\eta}(P_1) = 0$ and $\det H_{\eta}(P_2) = -1$, we have the statement. \end{proof} \begin{pro} \label{35} {\it We have \[ \det H_{\eta}(C_n^j) = (-1)^{n+1} 2 \cos (\eta j) + (-1)^{\lfloor \frac{n}{2} \rfloor} (1 + (-1)^n). \] } \end{pro} \begin{proof} We calculate $\det H_{\eta}(C_{n}^{j})$, which is \[ \begin{vmatrix} 0 & e^{\eta i} & 0 & \cdots & \cdots & \cdots & \cdots & 0 & 1 \\ e^{-\eta i} & 0 & e^{\eta i} & 0 & \cdots & \cdots & \cdots & \cdots & 0 \\ 0 & \ddots & \ddots & \ddots & & & & & \vdots \\ \vdots & & e^{-\eta i} & 0 & e^{\eta i} & & & & \vdots \\ \vdots & & & e^{-\eta i} & 0 & 1 & & & \vdots \\ \vdots & & & & 1 & 0 & 1 & & \vdots \\ \vdots & & & & & \ddots & \ddots & \ddots & 0 \\ 0 & \cdots & \cdots & \cdots & \cdots & 0 & 1 & 0 & 1 \\ 1 & 0 & \cdots & \cdots & \cdots & \cdots & 0 & 1 & 0 \end{vmatrix}. \] By the cofactor expansion along the first row, we have $\det H_{\eta}(C_n^j) = -e^{\eta i}D_1 + (-1)^{n+1}D_2$, where \[ D_1 = \begin{vmatrix} e^{-\eta i} & e^{\eta i} & 0 & \cdots & \cdots & \cdots & \cdots & \cdots & 0 & 0 \\ 0 & 0 & e^{\eta i} & 0 & \cdots & \cdots & \cdots & \cdots & \cdots & 0 \\ \vdots & e^{-\eta i} & 0 & e^{\eta i} & & & & & & \vdots \\ \vdots & & \ddots & \ddots & \ddots & & & & & \vdots \\ \vdots & & & e^{-\eta i} & 0 & e^{\eta i} & & & & \vdots \\ \vdots & & & & e^{-\eta i} & 0 & 1 & & & \vdots \\ \vdots & & & & & 1 & 0 & 1 & & \vdots \\ \vdots & & & & & & \ddots & \ddots & \ddots & 0 \\ 0 & \cdots & \cdots & \cdots & \cdots & \cdots & 0 & 1 & 0 & 1 \\ 1 & 0 & \cdots & \cdots & \cdots & \cdots & \cdots & 0 & 1 & 0 \end{vmatrix}, \] and \[ D_2 = \begin{vmatrix} e^{-\eta i} & 0 & e^{\eta i} & 0 & \cdots & \cdots & \cdots & \cdots & 0 & 0 \\ 0 & e^{-\eta i} & 0 & e^{\eta i} & \cdots & \cdots & \cdots & \cdots & \cdots & 0 \\ \vdots & & e^{-\eta i} & 0 & e^{\eta i} & & & & & \vdots \\ \vdots & & & \ddots & \ddots & \ddots & & & & \vdots \\ \vdots & & & & e^{-\eta i} & 0 & e^{\eta i} & & & \vdots \\ \vdots & & & & & e^{-\eta i} & 0 & 1 & & \vdots \\ \vdots & & & & & & 1 & 0 & \ddots & \vdots \\ \vdots & & & & & & & \ddots & \ddots & 1 \\ 0 & \cdots & \cdots & \cdots & \cdots & \cdots & & 0 & 1 & 0 \\ 1 & 0 & \cdots & \cdots & \cdots & \cdots & \cdots & & 0 & 1 \end{vmatrix}. \] We apply the cofactor expansion along the first column to $D_1$. By Corollary~\ref{60}, we have \begin{align*} D_1 &= (-1)^{1+1} e^{-\eta i} \det H_{\eta}(P_{n-2})+ (-1)^{(n-1)+1} e^{(j-1)\eta i} \\ &= e^{-\eta i} \det H_{\eta}(P_{n-2}) + (-1)^{n} e^{(j-1) \eta i}. \end{align*} Applying the cofactor expansion along the first column to $D_2$, we have \begin{align*} D_2 &= (-1)^{1+1}e^{-\eta i} \cdot e^{-(j-1)\eta i} + (-1)^{(n-1) + 1} \det H_{\eta}(P_{n-2}) \\ &= e^{-j \eta i} + (-1)^{n} \det H_{\eta}(P_{n-2}). \end{align*} Therefore, \begin{align*} \det H_{\eta}(C_n^j) &= -e^{\eta i}D_1 + (-1)^{n+1}D_2 \\ &= (-1)^{n+1} 2 \cos (\eta j) + (-2) \det H_{\eta}(P_{n-2}) \\ &= (-1)^{n+1} 2 \cos (\eta j) + (-2) \cdot (-1)^{\lfloor \frac{n-2}{2} \rfloor} \frac{1+(-1)^{n-2}}{2} \tag{by Lemma~\ref{50}} \\ &= (-1)^{n+1} 2 \cos (\eta j) + (-1)^{\lfloor \frac{n}{2} \rfloor} (1 + (-1)^n). \end{align*} \end{proof} Proposition~\ref{03} and Proposition~\ref{35} derive Theorem~\ref{Main1}, which is our first main theorem. In addition, Proposition~\ref{35} yields that, except for a finite number of $\eta$, \[ \det (\lambda I - H_{\eta}(C_n^j)) \neq \det (\lambda I - H_{\eta}(C_n^k)) \] for $j \neq k$. We supplement the phrase ``except for a finite number of $\eta$" here. For example, we consider $\eta = \frac{\pi}{2}$ and the mixed cycles of length 4. We have \begin{align*} \det H_{\frac{\pi}{2}}(C_4^j) &= 2 -2\cos \frac{j \pi}{2} \\ &= \begin{cases} 0 \qquad &\text{if $j \in \{0, 4\}$,} \\ 2 \qquad &\text{if $j \in \{1, 3\}$,} \\ 4 \qquad &\text{if $j=2$.} \end{cases} \end{align*} As this example points out, when $\eta \in \MB{Q}\pi$ and the denominator of $\eta / \pi$ is smaller than a given $n$, the determinants could be equal for different types of mixed cycles. There are only a finite number of such $\eta$ for given $n$. This is the reason why the only three types of mixed cycles appeared in the study of \cite{AGNN}. \subsection{Characteristic polynomials of $H_{\eta}(C_n^j)$} On the other hand, if the determinants are same, the characteristic polynomials of mixed cycles are actually equal. We find the characteristic polynomial of $C_n^j$. An {\it elementary subgraph} of an undirected graph $\Gamma$ is a subgraph of $\Gamma$ such that every component is either $K_2$ or an undirected cycle. Let $\MC{H}(\Gamma)$ denote the set of all the elementary subgraphs of $\Gamma$, and let $\MC{H}_l(\Gamma)$ denote the set of all the elementary subgraphs of $\Gamma$ on $l$ vertices. The {\it rank} and {\it corank} of an undirected graph $\Gamma = (V, E)$ are, respectively, $r(\Gamma) = |V| - c$ and $s(\Gamma) = |E| - |V| + c$, where $c$ is the number of components of $\Gamma$. Let $C_n$ be the undirected cycle graph on $n$ vertices. \begin{pro} {\it We have \[ \det (\lambda I - H_{\eta}(C_n^j)) = \sum_{k = 0}^{\lfloor \frac{n-1}{2} \rfloor} (-1)^k \frac{n}{n-k} \binom{n-k}{k} \lambda^{n-2k} - 2 \cos (\eta j) + (-1)^{\lfloor \frac{n}{2} \rfloor} (1 + (-1)^n). \] } \end{pro} \begin{proof} Let \begin{align*} \det (\lambda I - H_{\eta}(C_n^j)) &= \lambda^n + a_1 \lambda^{n-1} + \cdots + a_{n-1} \lambda + a_n, \\ \det (\lambda I - H_{\eta}(C_n)) &= \lambda^n + b_1 \lambda^{n-1} + \cdots + b_{n-1} \lambda + b_n. \end{align*} Fix $l \in \{1, \dots, n-1 \}$. Since the girth of $C_n^j$ is $n$, we have $a_l = b_l$ by Proposition~\ref{10}. Also, \begin{equation} \label{35a} \MC{H}_l(C_n) = \begin{cases} \{ \Gamma' \in \MC{H}(C_n) \mid \Gamma' \simeq \tfrac{l}{2} K_2 \} \quad &\text{if $l$ is even,} \\ \emptyset \quad &\text{if $l$ is odd,} \end{cases} \end{equation} where $\Gamma' \simeq \tfrac{l}{2} K_2$ denotes that the graph $\Gamma'$ is isomorphic to the disjoint union of the $\frac{l}{2}$ complete graphs $K_2$. Thus, $b_l = 0$ if $l$ is odd. In addition, if $l$ is even, \begin{align*} b_l &= \sum_{\Gamma' \in \MC{H}_l(C_n)} (-1)^{r(\Gamma')} 2^{s(\Gamma')} \tag{by Proposition~7.1 in \cite{B}} \\ &= \sum_{\Gamma' \in \MC{H}_l (C_n)} (-1)^{\frac{l}{2}} \tag{by (\ref{35a})} \\ &= (-1)^{\frac{l}{2}} | \{ \text{$\tfrac{l}{2}$-matching in $C_n$} \} | \\ &= (-1)^{\frac{l}{2}} \binom{n-\frac{l}{2}}{\frac{l}{2}}. \tag{by Exercises in p.14 of \cite{G}} \end{align*} By Proposition~\ref{35}, the characteristic polynomial $\det (\lambda I - H_{\eta}(C_n^j))$ is \begin{align*} & \, \sum_{k = 0}^{\lfloor \frac{n-1}{2} \rfloor} (-1)^k \frac{n}{n-k} \binom{n-k}{k} \lambda^{n-2k} + (-1)^n \{ (-1)^{n+1} 2 \cos (\eta j) + (-1)^{\lfloor \frac{n}{2} \rfloor} (1 + (-1)^n) \} \\ =& \sum_{k = 0}^{\lfloor \frac{n-1}{2} \rfloor} (-1)^k \frac{n}{n-k} \binom{n-k}{k} \lambda^{n-2k} -2 \cos (\eta j) + (-1)^{\lfloor \frac{n}{2} \rfloor + n} (1 + (-1)^n) \\ =& \sum_{k = 0}^{\lfloor \frac{n-1}{2} \rfloor} (-1)^k \frac{n}{n-k} \binom{n-k}{k} \lambda^{n-2k} -2 \cos (\eta j) + (-1)^{\lfloor \frac{n}{2} \rfloor} (1 + (-1)^n). \end{align*} We note that $(-1)^{\lfloor \frac{n}{2} \rfloor + n} \neq (-1)^{\lfloor \frac{n}{2} \rfloor}$, but $(-1)^{\lfloor \frac{n}{2} \rfloor + n} (1 + (-1)^n) = (-1)^{\lfloor \frac{n}{2} \rfloor} (1 + (-1)^n)$ in the last calculation. We have the statement. \end{proof} \subsection{Mixed graphs as $\MB{T}$-gain graphs} In this subsection, we briefly touch gain graphs. $\eta$-Hermitian adjacency matrices are also seen as adjacency matrices of $\MB{T}$-gain graphs. We refer to \cite{MKS, SK} for readers. We use notations and terminologies in \cite{MKS, SK}. The $\eta$-Hermitian adjacency matrix of a mixed graph $G = (V, \MC{A})$ is the adjacency matrix of the $\MB{T}$-gain graph $\Phi = (G^{\pm}, \MB{T}, \varphi)$ defined by the $\MB{T}$-gain $\varphi : \MC{A}^{\pm} \to \MB{T}$ such that \[ \varphi(a) = \begin{cases} 1 \qquad &\text{if $a \in \MC{A} \cap \MC{A}^{-1}$,} \\ e^{\eta i} \qquad &\text{if $a \in \MC{A} \setminus \MC{A}^{-1}$,} \\ e^{-\eta i} \qquad &\text{if $a \in \MC{A}^{-1} \setminus \MC{A}$,} \\ 0 \qquad &\text{otherwise.} \end{cases} \] In \cite{GH, MKS}, the authors mention the determinants of the adjacency matrices of $\MB{T}$-gain graphs. In addition, the coefficients of the characteristic polynomials are also found by using principal minors. The discussions in this section can also be carried out in terms of $\MB{T}$-gain graphs. \section{Preliminaries on quantum walks defined by mixed graphs} \label{105} Let $\eta \in [0, 2\pi)$, and let $G = (V, \MC{A})$ be a mixed graph. The {\it $\eta$-function} $\theta : \MC{A}^{\pm} \to \MB{R}$ of a mixed graph $G$ is defined by \[ \theta(a) = \begin{cases} \eta \qquad &\text{if $a \in \MC{A} \setminus \MC{A}^{-1}$,} \\ -\eta \qquad &\text{if $a \in \MC{A}^{-1} \setminus \MC{A}$,} \\ 0 \qquad &\text{if $a \in \MC{A} \cap \MC{A}^{-1}$.} \end{cases} \] Note that $\theta(a^{-1}) = -\theta(a)$ for any $a \in \MC{A}^{\pm}$. \subsection{Several matrices on quantum walks defined by mixed graphs} In \cite{KST}, the authors provided a quantum walk defined by a mixed graph. Let $G = (V, \MC{A})$ be a mixed graph equipped with an $\eta$-function $\theta$. We define several matrices (operators) on quantum walks. The {\it boundary operator} $K = K(G) \in \MB{C}^{V \times \MC{A}^{\pm}}$ is defined by \[ K_{x,a} = \frac{1}{\sqrt{\deg x}} \delta_{x,t(a)}. \] The {\it coin operator} $C = C(G) \in \MB{C}^{\MC{A}^{\pm} \times \MC{A}^{\pm}}$ is defined by $C = 2K^*K-I$. The {\it shift operator} $S_{\theta} = S_{\theta}(G) \in \MB{C}^{\MC{A}^{\pm} \times \MC{A}^{\pm}}$ is defined by $(S_{\theta})_{ab} = e^{\theta(b)i}\delta_{a,b^{-1}}$. Define the {\it time evolution matrix} $U_{\theta} = U_{\theta}(G) \in \MB{C}^{\MC{A}^{\pm} \times \MC{A}^{\pm}}$ by $U_{\theta} = S_{\theta} C$. \begin{lem} \label{22} {\it Let $G = (V, \MC{A})$ be a mixed graph equipped with an $\eta$-function $\theta$. We have \[ (U_{\theta})_{a,b} = e^{-\theta(a) i} \marukakko{ \frac{2}{\deg_{G} t(b)} \delta_{o(a), t(b)} - \delta_{a, b^{-1}} } \] for any $a,b \in \MC{A}^{\pm}$. } \end{lem} \begin{proof} Indeed, \begin{align*} (U_{\theta})_{a,b} &= (2 S_{\theta} K^* K - S_{\theta})_{a,b} \\ &= 2(S_{\theta} K^* K)_{a,b} - (S_{\theta})_{a,b} \\ &= 2 \sum_{z \in \MC{A}} \sum_{x \in V} (S_{\theta})_{a,z} (K^*)_{z, x} K_{x,b} - e^{\theta(b)i} \delta_{a, b^{-1}} \\ &= 2 \sum_{z \in \MC{A}} \sum_{x \in V} e^{\theta(z)i} \frac{1}{\sqrt{\deg x}} \frac{1}{\sqrt{\deg x}} \delta_{a,z^{-1}} \delta_{x, t(z)} \delta_{x, t(b)} - e^{\theta(a^{-1})i} \delta_{a, b^{-1}} \\ &= 2 \sum_{x \in V} e^{\theta(a^{-1})i} \frac{1}{\deg x} \delta_{x, t(a^{-1})} \delta_{x, t(b)} - e^{\theta(a^{-1})i} \delta_{a, b^{-1}} \\ &= \frac{2 e^{\theta(a^{-1})i}}{\deg_{G} t(b)} \delta_{o(a), t(b)} - e^{\theta(a^{-1})i} \delta_{a, b^{-1}} \\ &= e^{- \theta(a) i} \marukakko{ \frac{2}{\deg_{G} t(b)} \delta_{o(a), t(b)} - \delta_{a, b^{-1}} }. \end{align*} \end{proof} The following is an important theorem that links quantum walks and spectral graph theory. We cite \cite{KST}. In \cite{HKSS2014}, Higuchi et al~proved a similar claim in more general models. \begin{thm}[\cite{KST}] \label{21} {\it Let $G = (V, \MC{A})$ be a mixed graph equipped with an $\eta$-function $\theta$, and let $U_{\theta}$ be the time evolution matrix. Then we have \[ \Spec(U_{\theta}) = \{ e^{\pm i \cos^{-1} (\lambda)} \mid \lambda \in \Spec(\tilde{H}_{\eta}(G)) \} \cup \{ 1 \}^{M_1} \cup \{-1\}^{M_{-1}}, \] where \begin{align*} M_{1} = \frac{1}{2}|\MC{A}^{\pm}| - |V| + \dim \ker ( \tilde{H}_{\eta}(G) - I), \\ M_{-1} = \frac{1}{2}|\MC{A}^{\pm}| - |V| + \dim \ker ( \tilde{H}_{\eta}(G) + I). \end{align*} } \end{thm} The operators (matrices) used in our quantum walks are summarized in Table~\ref{1000}, where $G = (V, \MC{A})$ is a mixed graph equipped with an $\eta$-function $\theta$. \begin{table}[H] \centering \begin{tabular}{|c|c|c|c|} \hline Notation & Name & Indices of rows and columns & Definition \\ \hline \hline $K$ & Boundary & $V \times \MC{A}^{\pm}$ & $K_{x,a} = \frac{1}{ \sqrt{\deg x} } \delta_{x, t(a)}$ \\ \hline $C$ & Coin & $\MC{A}^{\pm} \times \MC{A}^{\pm}$ &$ C = 2K^*K - I$ \\ \hline $S_{\theta}$ & Shift & $\MC{A}^{\pm} \times \MC{A}^{\pm}$ & $(S_{\theta})_{ab} = e^{\theta(b)i}\delta_{a,b^{-1}}$ \\ \hline $U_{\theta}$ & Time evolution & $\MC{A}^{\pm} \times \MC{A}^{\pm}$ & $U_{\theta} = S_{\theta} C$ \\ \hline \end{tabular} \caption{The operators (matrices) used in our quantum walk} \label{1000} \end{table} \subsection{Necessary and sufficient conditions on periodicity} Let $U_{\theta}$ be a time evolution matrix of a mixed graph $G$ equipped with an $\eta$-function $\theta$. We say that $G$ is {\it periodic} if there exists $\tau \in \MB{N}$ such that $U_{\theta}^{\tau} = I$. When the mixed graph $G$ is periodic, the {\it period} is defined by $\min \{ \tau \in \MB{N} \mid U_{\theta}^{\tau} = I \}$. \begin{lem} \label{13} {\it Let $U_{\theta}$ be a time evolution matrix of a mixed graph $G$ equipped with an $\eta$-function $\theta$. Then, we have \begin{equation} \label{70} \{ \tau \in \MB{N} \mid U_{\theta}^{\tau} = I \} = \{ \tau \in \MB{N} \mid \lambda^{\tau} = 1 \text{ \emph{for any} $\lambda \in \Spec(U_{\theta})$} \}. \end{equation} In particular, $G$ is periodic if and only if there exists $\tau \in \MB{N}$ such that $\lambda^{\tau} = 1$ holds for any eigenvalue $\lambda$ of $U_{\theta}$. } \end{lem} \begin{proof} Since $U_{\theta}$ is unitary, there exists a unitary matrix $Q$ such that \[ Q^{*} U_{\theta} Q = \diag (\lambda_1, \cdots, \lambda_{2m}), \] where $m$ is the number of edges of $G^{\pm}$. Thus for $\tau \in \MB{N}$, \[ Q^{*} U_{\theta}^{\tau} Q = \diag (\lambda_1^{\tau}, \cdots, \lambda_{2m}^{\tau}). \] This implies~(\ref{70}). \end{proof} Define $\BM{e}^{(a)} \in \MB{C}^{\MC{A}^{\pm}}$ by $(\BM{e}^{(a)})_z = \delta_{a,z}$. Let $\MC{E}_{\MC{A}^{\pm}} = \{ \BM{e}^{(a)} \mid a \in \MC{A}^{\pm} \}$. This is the canonical basis of $\MB{C}^{\MC{A}^{\pm}}$. \begin{lem} \label{55} {\it Let $U_{\theta}$ be a time evolution matrix of a mixed graph $G$ equipped with an $\eta$-function $\theta$. Then, we have \[ \{ \tau \in \MB{N} \mid U_{\theta}^{\tau} = I \} = \{ \tau \in \MB{N} \mid U_{\theta}^{\tau}\BM{e}^{(a)} = \BM{e}^{(a)} \text{ \emph{for any} $\BM{e}^{(a)} \in \MC{E}_{\MC{A}^{\pm}}$} \}.\] In particular, $G$ is periodic if and only if there exists $\tau \in \MB{N}$ such that $U_{\theta}^{\tau}\BM{e}^{(a)} = \BM{e}^{(a)}$ holds for any vector $\BM{e}^{(a)} \in \MC{E}_{\MC{A}^{\pm}}$. } \end{lem} \begin{proof} It is clear that the left-hand side is included in the right-hand side. We show the reverse inclusion. Suppose $U_{\theta}^{\tau}\BM{e}^{(a)} = \BM{e}^{(a)}$ for any vector $\BM{e}^{(a)} \in \MC{E}_{\MC{A}^{\pm}}$ and for some $\tau \in \MB{N}$. Number the arc set $\MC{A}$ as $a_1, a_2, \dots, a_{2m}$, where $m$ is the number of edges of $G^{\pm}$. Let $P = [ \BM{e}^{(a_1)} \, \BM{e}^{(a_2)} \, \cdots \BM{e}^{(a_{2m})} ]$. Then, we have $U_{\theta}^{\tau} P = P$. Since $P$ is invertible, $U_{\theta}^{\tau} = I$ holds. \end{proof} \section{Periodicity of mixed paths} \label{106} In this section, we discuss periodicity of mixed paths. As a preparation for that, we introduce notations for expressing dynamics of quantum walk. Let $G = (V, \MC{A})$ be a mixed graph equipped with an $\eta$-function $\theta$. We write the components of a vector $\Psi \in \MB{C}^{\MC{A}^{\pm}}$ on the arcs of the graph as in Figure~\ref{48}. If a component of $\Psi$ is $0$, we omit the arc itself corresponding to the component. If a component of $\Psi$ is $1$, we may omit the value on the arc corresponding to the component. \begin{figure}[ht] \begin{center} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[u] (1) at (-1.2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (3, 0) {}; \node[u] (4) at (-1.2, 0.68) {}; \node[u] (5) at (-1.2, -0.68) {}; \node[u] (7) at (4.2, 0) {}; \node[u] (8) at (4.2, 0.68) {}; \node[u] (9) at (4.2, -0.68) {}; \node[u] (12) at (3.3, 0.2) {}; \node[u] (13) at (5.7, 0.2) {}; \draw (0,-0.5) node{$x$}; \draw (3,-0.5) node{$y$}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (2) to (3); \draw[-] (5) to (2); \node[u] (10) at (0.3, 0.2) {}; \node[u] (11) at (2.7, 0.2) {}; \draw[draw= blue,->] (10) to (11); \node[u] (20) at (0.3, -0.2) {}; \node[u] (21) at (2.7, -0.2) {}; \draw[draw= blue,->] (21) to (20); \draw[-] (3) to (7); \draw[-] (3) to (8); \draw[-] (3) to (9); \draw (1.5,0.6) node[blue]{$\Psi_{(x,y)}$}; \draw (1.5,-0.6) node[blue]{$\Psi_{(y,x)}$}; \end{tikzpicture} \caption{The components of a vector written on the arcs of the graph} \label{48} \end{center} \end{figure} We provide an example. Let $G = (V, \MC{A})$ be the mixed graph in Figure~\ref{a01}. We consider a general $\eta \in [0, 2\pi)$ and the $\eta$-function $\theta$. \begin{figure}[ht] \begin{center} \begin{tikzpicture} [scale = 0.8, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm}] \node[v] (1) at (-3, 0) {}; \draw (-3,0.5) node {$v_1$}; \node[v] (2) at (0, 0) {}; \draw (0,0.5) node {$v_2$}; \node[v] (3) at (3, 1.7) {}; \draw (3.5,1.7) node {$v_3$}; \node[v] (4) at (3, -1.7) {}; \draw (3.5,-1.7) node {$v_4$}; \draw[->] (1) to [bend left = 15] (2); \draw[->] (2) to [bend left = 15] (1); \draw[->] (2) to (4); \draw[->] (4) to (3); \draw[->] (3) to (2); \end{tikzpicture} \caption{Mixed graph $G$} \label{a01} \end{center} \end{figure} We focus on the vector $\BM{e}^{((v_3, v_2))} \in \MC{E}_{\MC{A}^{\pm}}$. The actions of the coin operator $C$ and the shift operator $S_{\theta}$ are shown in Figure~\ref{40} and~\ref{41}. Since $U_{\theta} = S_{\theta} C$, the action of the time evolution matrix $U_{\theta}$ is as in Figure~\ref{42}. \begin{figure}[ht] \begin{center} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-3, 0) {}; \draw (-3,0.5) node {$v_1$}; \node[v] (2) at (0, 0) {}; \draw (0,0.5) node {$v_2$}; \node[v] (3) at (3, 1.7) {}; \draw (3.5,1.7) node {$v_3$}; \node[v] (4) at (3, -1.7) {}; \draw (3.5,-1.7) node {$v_4$}; \node[u] (5) at (0.3, 0.37) {}; \node[u] (6) at (2.7, 1.73) {}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (4) to (3); \draw[-] (3) to (2); \draw[draw= blue,->] (6) to (5); \end{tikzpicture} \raisebox{13mm}{$\quad \overset{C}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-3, 0) {}; \draw (-3,0.5) node {$v_1$}; \node[v] (2) at (0, 0) {}; \draw (0,0.5) node {$v_2$}; \node[v] (3) at (3, 1.7) {}; \draw (3.5,1.7) node {$v_3$}; \node[v] (4) at (3, -1.7) {}; \draw (3.5,-1.7) node {$v_4$}; \node[u] (5) at (0.3, 0.37) {}; \node[u] (6) at (2.7, 1.73) {}; \node[u] (7) at (-2.7, 0.2) {}; \node[u] (8) at (-0.3, 0.2) {}; \node[u] (9) at (0.3, -0.37) {}; \node[u] (10) at (2.7, -1.73) {}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (4) to (3); \draw[-] (3) to (2); \draw[draw= blue,->] (6) to (5); \draw[draw= blue,->] (7) to (8); \draw[draw= blue,->] (10) to (9); \draw (-1.5,0.75) node[blue] {$\tfrac{2}{3}$}; \draw (1.1,1.5) node[blue] {$-\tfrac{1}{3}$}; \draw (1.2,-1.5) node[blue] {$\tfrac{2}{3}$}; \end{tikzpicture} \caption{Action of $C$} \label{40} \end{center} \end{figure} \begin{figure}[ht] \begin{center} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-3, 0) {}; \draw (-3,0.5) node {$v_1$}; \node[v] (2) at (0, 0) {}; \draw (0,0.5) node {$v_2$}; \node[v] (3) at (3, 1.7) {}; \draw (3.5,1.7) node {$v_3$}; \node[v] (4) at (3, -1.7) {}; \draw (3.5,-1.7) node {$v_4$}; \node[u] (5) at (0.3, 0.37) {}; \node[u] (6) at (2.7, 1.73) {}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (4) to (3); \draw[-] (3) to (2); \draw[draw= blue,->] (6) to (5); \end{tikzpicture} \raisebox{13mm}{$\quad \overset{S_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-3, 0) {}; \draw (-3,0.5) node {$v_1$}; \node[v] (2) at (0, 0) {}; \draw (0,0.5) node {$v_2$}; \node[v] (3) at (3, 1.7) {}; \draw (3.5,1.7) node {$v_3$}; \node[v] (4) at (3, -1.7) {}; \draw (3.5,-1.7) node {$v_4$}; \node[u] (5) at (0.3, 0.37) {}; \node[u] (6) at (2.7, 1.73) {}; \draw (1.3,1.25) node[blue]{$e^{i\eta }$}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (4) to (3); \draw[-] (3) to (2); \draw[draw= blue,->] (5) to (6); \end{tikzpicture} \caption{Action of $S_{\theta}$} \label{41} \end{center} \end{figure} \begin{figure}[ht] \begin{center} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-3, 0) {}; \draw (-3,0.5) node {$v_1$}; \node[v] (2) at (0, 0) {}; \draw (0,0.5) node {$v_2$}; \node[v] (3) at (3, 1.7) {}; \draw (3.5,1.7) node {$v_3$}; \node[v] (4) at (3, -1.7) {}; \draw (3.5,-1.7) node {$v_4$}; \node[u] (5) at (0.3, 0.37) {}; \node[u] (6) at (2.7, 1.73) {}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (4) to (3); \draw[-] (3) to (2); \draw[draw= blue,->] (6) to (5); \end{tikzpicture} \raisebox{13mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-3, 0) {}; \draw (-3,0.5) node {$v_1$}; \node[v] (2) at (0, 0) {}; \draw (0,0.5) node {$v_2$}; \node[v] (3) at (3, 1.7) {}; \draw (3.5,1.7) node {$v_3$}; \node[v] (4) at (3, -1.7) {}; \draw (3.5,-1.7) node {$v_4$}; \node[u] (5) at (0.3, 0.37) {}; \node[u] (6) at (2.7, 1.73) {}; \node[u] (7) at (-2.7, 0.2) {}; \node[u] (8) at (-0.3, 0.2) {}; \node[u] (9) at (0.3, -0.37) {}; \node[u] (10) at (2.7, -1.73) {}; \draw (1.1,1.4) node[blue]{$-\tfrac{1}{3}e^{i\eta}$}; \draw (-1.5,0.75) node[blue]{$\tfrac{2}{3}$}; \draw (1.1,-1.5) node[blue]{$\tfrac{2}{3} e^{-i\eta}$}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (4) to (3); \draw[-] (3) to (2); \draw[draw= blue,->] (5) to (6); \draw[draw= blue,->] (8) to (7); \draw[draw= blue,->] (9) to (10); \end{tikzpicture} \caption{Action of $U_{\theta}$} \label{42} \end{center} \end{figure} \begin{lem} \label{25} {\it Let $G = (V, \MC{A})$ be a mixed graph equipped with an $\eta$-function $\theta$, and let $a \in \MC{A}^{\pm}$. We have the following: \begin{enumerate}[(1)] \item If $\deg t(a) = 1$, then $U_{\theta} \BM{e}^{(a)} = e^{\theta(a) i}\BM{e}^{(a^{-1})}$; and \item If $\deg t(a) = 2$, then $U_{\theta} \BM{e}^{(a)} = e^{-\theta(b)i}\BM{e}^{(b)}$, where $b$ is the arc in $\{ z \in \MC{A}^{\pm} \mid o(z) = t(a) \} \setminus \{a^{-1}\}$. \end{enumerate} } \end{lem} \begin{proof} Let $U_{\theta} = U_{\theta}(G)$. First, we have \begin{align*} (U_{\theta} \BM{e}^{(a)})_{z} &= \sum_{w \in \MC{A}^{\pm}} (U_{\theta})_{z,w} (\BM{e}^{(a)})_{w} \\ &= (U_{\theta})_{z,a} \\ &= e^{-\theta(z) i} \marukakko{ \frac{2}{\deg_{G} t(a)} \delta_{o(z), t(a)} - \delta_{z, a^{-1}} }. \tag{by Lemma~\ref{22}} \end{align*} Consider the case of $\deg t(a) = 1$. Then $o(z) = t(a)$ if and only if $z = a^{-1}$. Thus, \[ (U_{\theta} \BM{e}^{(a)})_{z} = e^{-\theta(a^{-1}) i} ( 2 \delta_{z, a^{-1}} - \delta_{z, a^{-1}} ) = e^{\theta(a) i} \delta_{z, a^{-1}}. \] We have $U_{\theta} \BM{e}^{(a)} = e^{\theta(a) i}\BM{e}^{(a^{-1})}$. We next consider the case of $\deg t(a) = 2$. Then, \begin{align*} (U_{\theta} e_a)_{z} &= e^{-\theta(z) i} (\delta_{o(z), t(a)} - \delta_{z, a^{-1}}) \\ &= \begin{cases} 0 \quad &\text{if $z=a^{-1}$,} \\ e^{-\theta(b) i} \quad &\text{if $z = b$,} \\ 0 \quad &\text{otherwise.} \end{cases} \end{align*} Therefore, we have $U_{\theta} \BM{e}^{(a)} = e^{-\theta(b)i}\BM{e}^{(b)}$. \end{proof} The above lemma is illustrated as in Figure~\ref{43} and~\ref{44}. Now, we discuss periodicity of mixed paths. The following is our second main theorem. \begin{figure}[ht] \begin{center} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[u] (1) at (-1.2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (3, 0) {}; \node[u] (4) at (-1.2, 0.68) {}; \node[u] (5) at (-1.2, -0.68) {}; \node[u] (6) at (0.3, 0.2) {}; \node[u] (7) at (2.7, 0.2) {}; \draw (0,-0.5) node{$x$}; \draw (3,-0.5) node{$y$}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (2) to (3); \draw[-] (5) to (2); \draw[draw= blue,->] (6) to (7); \end{tikzpicture} \raisebox{6mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[u] (1) at (-1.2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (3, 0) {}; \node[u] (4) at (-1.2, 0.68) {}; \node[u] (5) at (-1.2, -0.68) {}; \node[u] (6) at (0.3, 0.2) {}; \node[u] (7) at (2.7, 0.2) {}; \draw (1.5,0.7) node[blue]{$e^{i\theta((x,y)) }$}; \draw (0,-0.5) node{$x$}; \draw (3,-0.5) node{$y$}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (2) to (3); \draw[-] (5) to (2); \draw[draw= blue,->] (7) to (6); \end{tikzpicture} \caption{Illustration of Lemma~\ref{25} (1)} \label{43} \end{center} \end{figure} \begin{figure}[ht] \begin{center} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[u] (1) at (-1.2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (3, 0) {}; \node[u] (4) at (-1.2, 0.68) {}; \node[u] (5) at (-1.2, -0.68) {}; \node[v] (6) at (6, 0) {}; \node[u] (7) at (7.2, 0) {}; \node[u] (8) at (7.2, 0.68) {}; \node[u] (9) at (7.2, -0.68) {}; \node[u] (10) at (0.3, 0.2) {}; \node[u] (11) at (2.7, 0.2) {}; \node[u] (12) at (3.3, 0.2) {}; \node[u] (13) at (5.7, 0.2) {}; \draw (0,-0.5) node{$x$}; \draw (3,-0.5) node{$y$}; \draw (6,-0.5) node{$z$}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (2) to (3); \draw[-] (5) to (2); \draw[draw= blue,->] (10) to (11); \draw[-] (3) to (6); \draw[-] (6) to (7); \draw[-] (6) to (8); \draw[-] (6) to (9); \end{tikzpicture} \raisebox{6mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.7, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[u] (1) at (-1.2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (3, 0) {}; \node[u] (4) at (-1.2, 0.68) {}; \node[u] (5) at (-1.2, -0.68) {}; \node[v] (6) at (6, 0) {}; \node[u] (7) at (7.2, 0) {}; \node[u] (8) at (7.2, 0.68) {}; \node[u] (9) at (7.2, -0.68) {}; \node[u] (10) at (0.3, 0.2) {}; \node[u] (11) at (2.7, 0.2) {}; \node[u] (12) at (3.3, 0.2) {}; \node[u] (13) at (5.7, 0.2) {}; \draw (4.5,0.7) node[blue]{$e^{-i\theta((y,z)) }$}; \draw (0,-0.5) node{$x$}; \draw (3,-0.5) node{$y$}; \draw (6,-0.5) node{$z$}; \draw (1) to (2); \draw[-] (2) to (4); \draw[-] (2) to (3); \draw[-] (5) to (2); \draw[draw= blue,->] (12) to (13); \draw[-] (3) to (6); \draw[-] (6) to (7); \draw[-] (6) to (8); \draw[-] (6) to (9); \end{tikzpicture} \caption{Illustration of Lemma~\ref{25} (2)} \label{44} \end{center} \end{figure} \begin{thm} {\it Let $G = (V, \MC{A})$ be a mixed path on $n$ vertices equipped with an $\eta$-function $\theta$. Then $G$ is periodic for any $\eta \in [0, 2\pi)$, and the period is $2(n-1)$. } \end{thm} \begin{proof} By Proposition~\ref{12}, $G$ and $G^{\pm}$ are $\tilde{H}_{\eta}$-cospectral since $G$ is a mixed tree. By Theorem~\ref{21}, $U_{\theta}(G)$ and $U_{\theta}(G^{\pm})$ have the same eigenvalues. From Lemma~\ref{13}, periodicity of $G$ and its period are determined by the eigenvalues of the time evolution matrix. Thus, it is sufficient to discuss only periodicity of $G^{\pm}$, which is the undirected path graph $P_n$ on $n$ vertices. By Lemma~\ref{25}, we have $U_{\theta}(P_n)^{2(n-1)} \BM{e}^{(a)} = \BM{e}^{(a)}$ for any vector $\BM{e}^{(a)} \in \MC{E}_{\MC{A}^{\pm}}$. The dynamics is as follows: \begin{align*} \begin{tikzpicture} [scale = 0.8, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (2, 0) {}; \node[u] (4) at (2.5, 0) {}; \draw (2.75,0) node{$\dots$}; \node[u] (5) at (3, 0) {}; \node[v] (6) at (3.5, 0) {}; \node[v] (7) at (5.5, 0) {}; \node[u] (8) at (-1.8, 0.2) {}; \node[u] (9) at (-0.2, 0.2) {}; \draw (1) to (2); \draw[-] (2) to (3); \draw[-] (3) to (4); \draw[-] (5) to (6); \draw[-] (6) to (7); \draw[draw= blue,->] (8) to (9); \end{tikzpicture} & \raisebox{1mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.8, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (2, 0) {}; \node[u] (4) at (2.5, 0) {}; \draw (2.75,0) node{$\dots$}; \node[u] (5) at (3, 0) {}; \node[v] (6) at (3.5, 0) {}; \node[v] (7) at (5.5, 0) {}; \node[u] (8) at (0.2, 0.2) {}; \node[u] (9) at (1.8, 0.2) {}; \draw (1) to (2); \draw[-] (2) to (3); \draw[-] (3) to (4); \draw[-] (5) to (6); \draw[-] (6) to (7); \draw[draw= blue,->] (8) to (9); \end{tikzpicture} \\ & \raisebox{1mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad \cdots$} \\ & \quad \hspace{2mm} \vdots \\ & \raisebox{1mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.8, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (2, 0) {}; \node[u] (4) at (2.5, 0) {}; \draw (2.75,0) node{$\dots$}; \node[u] (5) at (3, 0) {}; \node[v] (6) at (3.5, 0) {}; \node[v] (7) at (5.5, 0) {}; \node[u] (8) at (3.7, 0.2) {}; \node[u] (9) at (5.3, 0.2) {}; \draw (1) to (2); \draw[-] (2) to (3); \draw[-] (3) to (4); \draw[-] (5) to (6); \draw[-] (6) to (7); \draw[draw= blue,->] (8) to (9); \end{tikzpicture} \\ &\raisebox{1mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.8, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (2, 0) {}; \node[u] (4) at (2.5, 0) {}; \draw (2.75,0) node{$\dots$}; \node[u] (5) at (3, 0) {}; \node[v] (6) at (3.5, 0) {}; \node[v] (7) at (5.5, 0) {}; \node[u] (8) at (5.3, -0.2) {}; \node[u] (9) at (3.7, -0.2) {}; \draw (1) to (2); \draw[-] (2) to (3); \draw[-] (3) to (4); \draw[-] (5) to (6); \draw[-] (6) to (7); \draw[draw= blue,->] (8) to (9); \end{tikzpicture} \\ & \raisebox{1mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad \cdots$} \\ & \quad \hspace{2mm} \vdots \\ & \raisebox{1mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.8, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (2, 0) {}; \node[u] (4) at (2.5, 0) {}; \draw (2.75,0) node{$\dots$}; \node[u] (5) at (3, 0) {}; \node[v] (6) at (3.5, 0) {}; \node[v] (7) at (5.5, 0) {}; \node[u] (8) at (-1.8, -0.2) {}; \node[u] (9) at (-0.2, -0.2) {}; \draw (1) to (2); \draw[-] (2) to (3); \draw[-] (3) to (4); \draw[-] (5) to (6); \draw[-] (6) to (7); \draw[draw= blue,->] (9) to (8); \end{tikzpicture} \\ & \raisebox{1mm}{$\quad \overset{U_{\theta}}{\mapsto} \quad$} \begin{tikzpicture} [scale = 0.8, line width = 0.8pt, v/.style = {circle, fill = black, inner sep = 0.8mm},u/.style = {circle, fill = white, inner sep = 0.1mm}] \node[v] (1) at (-2, 0) {}; \node[v] (2) at (0, 0) {}; \node[v] (3) at (2, 0) {}; \node[u] (4) at (2.5, 0) {}; \draw (2.75,0) node{$\dots$}; \node[u] (5) at (3, 0) {}; \node[v] (6) at (3.5, 0) {}; \node[v] (7) at (5.5, 0) {}; \node[u] (8) at (-1.8, 0.2) {}; \node[u] (9) at (-0.2, 0.2) {}; \draw (1) to (2); \draw[-] (2) to (3); \draw[-] (3) to (4); \draw[-] (5) to (6); \draw[-] (6) to (7); \draw[draw= blue,->] (8) to (9); \end{tikzpicture} \end{align*} By Lemma~\ref{55}, we see that $G$ is periodic whose period is $2(n-1)$. \end{proof} Note that the periodicity of the undirected paths has actually studied in \cite{KSTY2018} by eigenvalue analysis. In this paper, we have proven the same fact in a different way. \section{Periodicity of mixed cycles} \label{107} Finally, we discuss periodicity of mixed cycles. Strategy is similar to the mixed paths, but the discussion is more complicated. Recall that a mixed cycle $G$ on $n$ vertices is $H_{\eta}$-cospectral with $C_n^j$ for some $j \in \{0,1,\dots, n\}$ by Proposition~\ref{03}. For $m_1, m_2 \in \MB{Z}$, the greatest common divisor of $m_1$ and $m_2$ is denoted by $(m_1, m_2)$. Note that $(0, m_1) = m_1$. The following is our third main theorem. \begin{thm} {\it Let $G = (V, \MC{A})$ be a mixed cycle on $n$ vertices equipped with an $\eta$-function $\theta$. Then, $G$ is periodic if and only if $\eta \in \MB{Q}\pi$. In addition, we suppose that $\eta \in \MB{Q}\pi$ and the mixed cycle $G$ is $H_{\eta}$-cospectral with $C_n^j$. Let $\eta = \frac{p}{q}\pi$, where $p$ and $q$ are coprime. Then, the period $\tau$ of $G$ is the following: \begin{equation} \label{65} \tau = \begin{cases} \frac{2qn}{(j, 2q)} \quad &\text{if $p$ is odd,} \\ \frac{qn}{(j, q)} \quad &\text{if $p$ is even.} \\ \end{cases} \end{equation} } \end{thm} \begin{proof} By Proposition~\ref{03}, $G$ is $H_{\eta}$-cospectral with $C_n^j$ for some $j \in \{0,1,\dots, n\}$. Since $G$ and $C_n^j$ are 2-regular, they are also $\tilde{H}_{\eta}$-cospectral. By Theorem~\ref{21}, $U_{\theta}(G)$ and $U_{\theta}(C_n^j)$ have the same eigenvalues. From Lemma~\ref{13}, periodicity is determined by the eigenvalues of the time evolution matrices, so it is sufficient to discuss only periodicity of $C_n^j$ for $j \in \{0,1,\dots, n\}$. Let $U_{\theta} = U_{\theta}(C_n^j)$, and let $a$ be an arbitrary arc of $C_n^j$. By Lemma~\ref{25}, we have \begin{equation} \label{26} U_{\theta}^{n} \BM{e^{(a)}} = e^{\pm j \eta i} \BM{e^{(a)}}. \end{equation} If $\eta \not\in \MB{Q}\pi$, then $j l \eta \not\in 2\pi \MB{Z}$ for any $l \in \MB{N}$, so the mixed graph is not periodic. On the other hand, we suppose $\eta \in \MB{Q}\pi$, say $\eta = \frac{p}{q}\pi$. Then, \[ U_{\theta}^{2qn} \BM{e^{(a)}} = e^{\pm 2qj \cdot \frac{p}{q} \pi i} \BM{e^{(a)}} = e^{\pm 2pj \pi i} \BM{e^{(a)}} = \BM{e^{(a)}}. \] Thus, the mixed graph is periodic. Next, we determine the period $\tau$. Let $\eta \in \MB{Q}\pi$ and let $\eta = \frac{p}{q}\pi$, where $p$ and $q$ are coprime. By Lemma~\ref{25}, the vector $U_{\theta}^k \BM{e^{(a)}}$ coincides with a complex multiple of $\BM{e^{(a)}}$ if and only if $k$ is a multiple of $n$. Thus, the period $\tau$ is a multiple of $n$, namely, $\tau = ln$ for some $l \in \MB{N}$. Then, \[ \BM{e^{(a)}} = U_{\theta}^{ln} \BM{e^{(a)}} = e^{\pm l j \cdot \frac{p}{q} \pi i} \BM{e^{(a)}}, \] so $\frac{pjl}{q} \pi \in 2\pi\MB{Z}$, i.e., $pjl \in 2q \MB{Z}$. We would like to find $\min \{ l \in \MB{N} \mid pjl \in 2q \MB{Z} \}$. If $j = 0$, we have $\min \{ l \in \MB{N} \mid pjl \in 2q \MB{Z} \} = 1$, so the period is $n$. This satisfies (\ref{65}). We assume that $j > 0$ in the discussion below. First, we consider the case where $p$ is odd. We will show that \[ \min \{ l \in \MB{N} \mid pjl \in 2q \MB{Z} \} = \frac{2q}{(j,2q)}. \] Since $p$ is odd and $(p,q) = 1$, we have \begin{equation} \label{27} (p,2q) = 1. \end{equation} Let $d = (j,2q)$. There exists $j' \in \MB{N}$ such that \begin{equation} \label{28} j = j' d \end{equation} and \begin{equation} \label{29} (j', 2q) = 1. \end{equation} Therefore, we have \begin{align*} \min \{ l \in \MB{N} \mid pjl \in 2q \MB{Z} \} &= \min \{ l \in \MB{N} \mid jl \in 2q \MB{Z}\} \tag{by (\ref{27})} \\ &= \min \{ l \in \MB{N} \mid j' d l \in 2q \MB{Z}\} \tag{by (\ref{28})} \\ &= \min \{ l \in \MB{N} \mid d l \in 2q \MB{Z} \} \tag{by (\ref{29})} \\ &= \frac{2q}{d}. \tag{since $d$ is a divisor of $2q$} \end{align*} Next, we consider the case where $p$ is even. We will show that \begin{equation} \min \{ l \in \MB{N} \mid pjl \in 2q \MB{Z} \} = \frac{q}{(j,q)}. \label{66} \end{equation} If $p=0$, then $q=1$ since $p$ and $q$ are coprime. We have $\min \{ l \in \MB{N} \mid pjl \in 2q \MB{Z} \} = 1$ and $\frac{q}{(j,q)} = 1$. Equality~(\ref{66}) is satisfied in this case. We assume that $p > 0$. Since $p$ is even, we have \begin{equation} \label{31} p = 2p' \end{equation} for some $p' \in \MB{N}$. $p$ and $q$ are coprime, so \begin{equation} \label{32} (p',q) = 1. \end{equation} Let $d = (j,q)$. There exists $j' \in \MB{N}$ such that \begin{equation} \label{33} j = j' d \end{equation} and \begin{equation} \label{34} (j', q) = 1. \end{equation} Therefore, we have \begin{align*} \min \{ l \in \MB{N} \mid pjl \in 2q \MB{Z} \} &= \min \{ l \in \MB{N} \mid p' j l \in q \MB{Z}\} \tag{by (\ref{31})} \\ &= \min \{ l \in \MB{N} \mid jl \in q \MB{Z}\} \tag{by (\ref{32})} \\ &= \min \{ l \in \MB{N} \mid j' d l \in q \MB{Z}\} \tag{by (\ref{33})} \\ &= \min \{ l \in \MB{N} \mid d l \in q \MB{Z} \} \tag{by (\ref{34})} \\ &= \frac{q}{d}. \tag{since $d$ is a divisor of $q$} \end{align*} We have the statement. \end{proof} \section*{Acknowledgements} This paper is based on the graduation theses of the second and third authors. We would like to thank their advisor, Professor Norio Konno, for his fruitful comments and helpful advice.

https://arxiv.org/abs/0907.3772

The Maximum Wiener Index of Trees with Given Degree Sequences

The Wiener index of a connected graph is the sum of topological distances between all pairs of vertices. Since Wang gave a mistake result on the maximum Wiener index for given tree degree sequence, in this paper, we investigate the maximum Wiener index of trees with given degree sequences and extremal trees which attain the maximum value.

\section{Introduction} The Wiener index of a molecular graph, introduced by Wiener \cite{wiener1947} in 1947, is one of the oldest and most widely used topological indices in the quantitative structure property relationships. In the mathematical literature, the Wiener index seems to be the first studied by Entringer et al. \cite{entringer1976}. For more information and background, the readers may refer to a recent and very comprehensive survey \cite{dobrynin2001} and a book \cite{rouvray2002} which is dedicated to Harry Wiener on the Wiener index and the references therein. Through this paper, all graphs are finite, simple and undirected. Let $G= (V,~E)$ be a simple connected graph with vertex set $V(G)=\{v_1,\cdots, v_n\}$ and edge set $E(G)$. Denote by $d_G(v_i)$ (or for short $ d(v_i)$) the {\it degree} of vertex $v_i$. The {\it distance} between vertices $v_i$ and $v_j$ is the minimum number of edges between $v_i$ and $v_j$ and denoted by $d_G(v_i, v_j)$ (or for short $d(v_i,v_j)$). The {\it Wiener index} of a connected graph $G$ is defined as \begin{equation}\label{weiner-def} W(G)=\sum_{\{v_i, v_j\}\subseteq V(G)}d(v_i, v_j). \end{equation} A {\it tree} is a connected and acyclic graph. A {\it caterpillar} is a tree in which a single path (called {\it Spine}) is incident to (or contains) every edge. For other terminology and notions, we follow from \cite{bondy1976}. Entringer et al. \cite{entringer1976} proved that the path $P_n$ and the star $K_{1, n-1}$ have the maximum and minimum Wiener indices, respectively, in the set consisting of all trees of order $n$. Dankelmann \cite{dankelmann1994} obtained the all extremal graphs in the set of all connected graphs with given the order and the matching number which attained the maximum Wiener value. Moreover, Fischermann et al. \cite{fischermann2002} and Jelen et al. \cite{jelen2003} independently determined all trees which have the minimum Wiener indices among all trees of order $n$ and maximum degree $\Delta$. A nonincreasing sequence of nonnegative integers $\pi=(d_1,d_2,\cdots, d_{n})$ is called {\it graphic} if there exists a simple graph having $\pi$ as its vertex degree sequence. Hence it is natural to consider the following problem. \begin{problem}\label{problem} Let $\pi=(d_1, \cdots, d_n)$ be graphic degree sequence and $${\mathcal{G}}_{\pi}=\{G: {\rm \ the\ degree \ sequence\ of} \ G\ {\rm is} \ \pi\}.$$ Find the upper (lower) bounds for the Wiener index of all graphs $G$ in ${\mathcal G}_{ \pi}$ and characterize all extremal graphs which attain the upper (lower) bounds. \end{problem} Moreover, we call a graph {\it maximum (minimum) optimal} if it maximizes (minimizes) the Wiener index in $\mathcal{G_{\pi}}$. Recently, by the different techniques, Wang \cite{wang2008} and Zhang et al.\cite{zhang2008} independently characterized the tree that minimizes the Wiener index among trees of given degree sequences. Moreover, they proved that the minimum optimal trees for a given tree degree sequence $\pi$ are unique. On the other hand, Wang in \cite{wang2008} also "{\it proved}" the only maximum optimal tree that maximizes the Wiener index among trees of given degree sequences. The result can be stated as follows: \begin{theorem}\cite{wang2008}\label{wang} Given the degree sequence and the number of vertices, the greedy caterpillar maximizes the Wiener index, where the greedy caterpillar with degree sequence $(d_1,\cdots, d_n)$ ($ d_1\ge d_2\ge \cdots \ge d_k\ge 2>d_{k+1}=1$) is formed by attaching pending edges to a path $v_1, v_2, \cdots, v_k$ of length $k-1$ such that $$d(v_1)\ge d(v_k)\ge d(v_2)\ge d(v_{k-1})\ge \cdots\ge d(v_{\lceil\frac{k+1}{2}\rceil}). $$ \end{theorem} Unfortunately, this result is not correct. For example: \begin{example}\label{example} Let $\pi=(13, 5, 5, 5, 4, 3, 1, \cdots, 1)$ be a degree sequence of tree with $31$ vertices. Let $T_1$ and $T_2$ be two trees with degree sequences $\pi$ (see Fig.1). \setlength{\unitlength}{0.1in} \begin{picture}(60,15) \put(3,5){\circle{0.5}} \put(3.25,5){\line(1,0){10}} \put(13.5,5){\circle{0.5}}\put(13.75,5){\line(1,0){10}} \put(24,5){\circle{0.5}} \put(24.25,5){\line(1,0){10}} \put(34.5,5){\circle{0.5}} \put(34.75,5){\line(1,0){10}} \put(45,5){\circle{0.5}} \put(45.25,5){\line(1,0){10}} \put(55.5,5){\circle{0.5}} \put(2.8,5.2){\line(-1,2){3.5}} \put(3.2,5.2){\line(1,2){3.5}} \put(13.5,5.25){\line(0,1){7}} \put(23.8,5.2){\line(-1,2){3.5}} \put(24.2,5.2){\line(1,2){3.5}} \put(34.3,5.2){\line(-1,2){3.5}} \put(34.7,5.2){\line(1,2){3.5}} \put(44.8,5.2){\line(-1,2){3.5}} \put(45.2,5.2){\line(1,2){3.5}} \put(55.3,5.2){\line(-1,2){3.5}} \put(55.7,5.2){\line(1,2){3.5}} \put(-0.8,12.5){\circle{0.5}} \put(6.87,12.5){\circle{0.5}} \put(2.2,12){$\cdots$}\put(2.2,13.3){$12$} \put(2.5,3.3){$v_{1}$} \put(13.5,12.5){\circle{0.5}} \put(13,3.3){$v_{2}$} \put(20.2,12.5){\circle{0.5}} \put(27.8,12.5){\circle{0.5}} \put(23.2,13.3){$2$} \put(23.2, 3.3){$v_{3}$} \put(30.61,12.5){\circle{0.5}} \put(38.3,12.5){\circle{0.5}} \put(33.7,12){$\cdots$} \put(34.8,13.3){$3$} \put(34, 3.3){$v_{4}$} \put(41.3,12.5){\circle{0.5}} \put(48.8,12.5){\circle{0.5}} \put(44,12){$\cdots$} \put(44.8,13.3){$3$}\put(44, 3.3){$v_{5}$} \put(51.7,12.5){\circle{0.5}}\put(59.2,12.5){\circle{0.5}} \put(54.5,12){$\cdots$}\put(55,13.3){$4$}\put(55, 3.3){$v_{6}$} \put(28, 1){$T_1$} \end{picture} \setlength{\unitlength}{0.1in} \begin{picture}(60,15) \put(3,5){\circle{0.5}} \put(3.25,5){\line(1,0){10}} \put(13.5,5){\circle{0.5}}\put(13.75,5){\line(1,0){10}} \put(24,5){\circle{0.5}} \put(24.25,5){\line(1,0){10}} \put(34.5,5){\circle{0.5}} \put(34.75,5){\line(1,0){10}} \put(45,5){\circle{0.5}} \put(45.25,5){\line(1,0){10}} \put(55.5,5){\circle{0.5}} \put(2.8,5.2){\line(-1,2){3.5}} \put(3.2,5.2){\line(1,2){3.5}} \put(13.3,5.2){\line(-1,2){3.5}} \put(13.7,5.2){\line(1,2){3.5}} \put(23.8,5.2){\line(-1,2){3.5}} \put(24.2,5.2){\line(1,2){3.5}} \put(34.5,5.25){\line(0,1){7}} \put(44.8,5.2){\line(-1,2){3.5}} \put(45.2,5.2){\line(1,2){3.5}} \put(55.3,5.2){\line(-1,2){3.5}} \put(55.7,5.2){\line(1,2){3.5}} \put(-0.8,12.5){\circle{0.5}} \put(6.87,12.5){\circle{0.5}} \put(2.2,12){$\cdots$}\put(2.2,13.3){$12$} \put(2.5,3.3){$v_{1}$} \put(9.7,12.5){\circle{0.5}} \put(17.2,12.5){\circle{0.5}} \put(12.6,12){$\cdots$} \put(13.5,13.3){$3$} \put(13,3.3){$v_{2}$} \put(20.2,12.5){\circle{0.5}} \put(27.8,12.5){\circle{0.5}} \put(23.2,13.3){$2$} \put(23.2, 3.3){$v_{3}$} \put(34.5,12.5){\circle{0.5}} \put(34, 3.3){$v_{4}$} \put(41.3,12.5){\circle{0.5}} \put(48.8,12.5){\circle{0.5}} \put(44,12){$\cdots$} \put(44.8,13.3){$3$}\put(44, 3.3){$v_{5}$} \put(51.7,12.5){\circle{0.5}}\put(59.2,12.5){\circle{0.5}} \put(54.5,12){$\cdots$}\put(55,13.3){$4$}\put(55, 3.3){$v_{6}$} \put(20, 1){$T_2$} \put(25,-1){\bf Figure 1 $T_1$ and $T_2$} \end{picture} \end{example} Clearly, $T_2$ is a greedy caterpillar and $T_1$ is not a greedy caterpillar. Moreover, they have the same degree sequences $\pi$. By calculation, it is easy to see that $$W(T_2)=9870< W(T_1)=9886.$$ Hence this example illustrates that Theorem~\ref{wang} in \cite{wang2008} is not correct. Motivated by Problem \ref{problem} and Example \ref{example}, we try to investigate the extremal trees which attain the maximum Wiener index among all trees with given degree sequences. The problem seems to be difficult. Because we find that the extremal tree depends on the values of components of degree sequences. The rest of the paper is organized as follows. In Section 2, we discuss some properties of the extremal tree with the maximum Wiener index and give an upper bound in terms of degree sequences. In Section 3, the extremal trees with the maximum Wiener index among given degree sequences $(d_1, \cdots, d_n)$, where $d_1\ge \cdots \ge d_k\ge 2>d_{k+1}=1$ and $k\le 6$ are characterized. Moreover, the extremal maximal trees are not unique. \section{Properties of extremal trees with the maximum Wiener index } Let $\mathcal{T_{\pi}}$ be the set of all trees with degree sequences $\pi=(d_1, d_2, \cdots, d_n)$ with $d_1\ge d_2\ge\cdots\ge d_n$. Shi in \cite{shi1993} proved that a maximum optimal tree must be a caterpillar. \begin{lemma}\cite{shi1993}\label{shi} Let $T^*$ be a maximum optimal tree in $\mathcal{T_{\pi}}$. Then $T^*$ is a caterpillar. \end{lemma} From Lemma~\ref{shi}, we only need to consider all caterpillars with a degree sequence $\pi$. In order to study the structure of the maximum optimal trees, we present a formula for Wiener index of any caterpillar. \begin{lemma}\label{formula} Let $T$ be a caterpillar of order $n$ with the degree sequence $\pi=(d(v_1), \cdots,$ $ d(v_k), d(v_{k+1}),\cdots,d(v_n))$(see Figure 2). \setlength{\unitlength}{0.1in} \begin{picture}(60,20) \put(10,5){\circle{0.5}} \put(10.25,5){\line(1,0){10}} \put(20.5,5){\circle{0.5}}\put(20.75,5){\line(1,0){5}} \put(26.5,4.55){$\cdot$}\put(27.5,4.55){$\cdot$}\put(28.5,4.55){$\cdot$} \put(30,5){\circle{0.5}} \put(31.5,4.55){$\cdot$}\put(32.5,4.55){$\cdot$}\put(33.5,4.55){$\cdot$} \put(35.25,5){\line(1,0){5}} \put(40.5,5){\circle{0.5}} \put(40.75,5){\line(1,0){10}} \put(51,5){\circle{0.5}} \put(9.8,5.2){\line(-1,2){3.5}} \put(10.2,5.2){\line(1,2){3.5}} \put(20.3,5.2){\line(-1,2){3.5}} \put(20.7,5.2){\line(1,2){3.5}} \put(29.8,5.2){\line(-1,2){3.5}} \put(30.2,5.2){\line(1,2){3.5}} \put(40.3,5.2){\line(-1,2){3.5}} \put(40.7,5.2){\line(1,2){3.5}} \put(50.8,5.2){\line(-1,2){3.5}} \put(51.2,5.2){\line(1,2){3.5}} \put(6.2,12.5){\circle{0.5}} \put(13.8,12.5){\circle{0.5}} \put(16.65,12.5){\circle{0.5}} \put(24.25,12.5){\circle{0.5}} \put(26.23,12.5){\circle{0.5}} \put(33.7,12.5){\circle{0.5}} \put(36.7,12.5){\circle{0.5}} \put(44.23,12.5){\circle{0.5}} \put(47.2,12.5){\circle{0.5}} \put(54.7,12.5){\circle{0.5}} \put(9,12){$\cdots$}\put(19,12){$\cdots$}\put(29,12){$\cdots$} \put(39,12){$\cdots$}\put(50,12){$\cdots$} \put(9,13.3){$y_1$}\put(18.2,13.3){$y_{2}-1$} \put(28.5,13.3){$y_{i}-1$} \put(37.5,13.3){$y_{k-1}-1$} \put(50,13.3){$y_{k}$} \put(9.5,3.3){$v_{1}$} \put(20,3.3){$v_{2}$} \put(30, 3.3){$v_{i}$} \put(40,3.3){$v_{k-1}$} \put(50.2,3.3){$v_{k}$} \put(25,1){\bf Figure 2}\put(34, 1){$T$} \end{picture} If $d(v_i)=y_i+1\ge 2$ for $i=1, \cdots, k$ and $d(v_{k+1})=\cdots=d(v_n)=1$, then \begin{equation}\label{weiner-f} W(T)=(n-1)^2+F(y_1, \cdots, y_k), \end{equation} where \begin{equation}\label{fx} F(y_1, \cdots, y_k)=\sum_{i=1}^{k-1}(\sum_{j=1}^iy_j)(\sum_{j=i+1}^ky_j). \end{equation} \end{lemma} \begin{proof} It is well known \cite{hosoya1971} that the formula (\ref{weiner-def}) is equal to $$W(T)=\sum_{e}n_1(e)n_2(e),$$ where $e=(u, v)$ is an edge of $T$, and $n_1(e)$ (resp. $n_2(e)$) is the number of vertices of the component of $T-e$ containing $u$ (resp. $v$). For $e_i=(v_i, v_{i+1})\in E(T),$ the numbers of vertices of the two components of $T-e_i$ are $\sum_{j=1}^id(v_j)-(i-1)$ and $\sum_{j=i+1}^kd(v_j)-(k-i-1)$ for $i=1, \cdots, k-1,$ respectively. Hence \begin{eqnarray*} W(T)&=&\sum_{e\in E(T)}n_1(e)n_2(e)\\ &=&\sum_{e {\rm {\ is \ pendent\ edge}}}n_1(e)n_2(e)+ \sum_{ e {\rm \ is \ not \ pendent\ edge}}n_1(e)n_2(e)\\ &=& (n-1)(n-k)+\sum_{i=1}^{k-1}(\sum_{j=1}^id(v_j)-(i-1))(\sum_{j=i+1}^kd(v_j)-(k-i-1)) \\ &=&(n-1)(n-k)+\sum_{i=1}^{k-1}(1+\sum_{j=1}^iy_j)(1+\sum_{j=i+1}^ky_j)\\ &=&(n-1)(n-k)+(k-1)(1+\sum_{j=1}^ky_j)+\sum_{i=1}^{k-1}(\sum_{j=1}^iy_j)(\sum_{j=i+1}^ky_j)\\ &=&(n-1)^2+F(y_1, \cdots, y_k), \end{eqnarray*} where last equality is due to $\sum_{j=1}^ky_j=\sum_{j=1}^kd(v_j)-k=2(n-1)-(n-k)-k=n-2$. This completes the proof. \end{proof} {\bf Remark} In this sequel, the caterpillar $T$ in Lemma~\ref{formula} is denoted by $T(y_1,\cdots, y_k)$. Then degree sequence of $T(y_1,\cdots, y_k)$ is $(y_1+1, \cdots, y_k+1, 1, \cdots, 1)$. The following theorem give a characterization of a maximum optimal tree. \begin{theorem}\label{optimal-equal} Let $\pi=(d_1, \cdots, d_n)$ with $d_1\ge\cdots \ge d_k\ge 2\ge d_{k+1}=\cdots=d_n=1$. Then $T$ is a maximum optimal tree in ${\mathcal{T}}_{\pi}$ if and only if $T$ is a caterpillar $T(x_1, \cdots, x_k)$ and $(x_1, \cdots, x_k)$ satisfies \begin{equation} F(x_1, \cdots , x_k)=\max\{F(y_1, \cdots, y_k)=\sum_{i=1}^{k-1}(\sum_{j=1}^iy_j)(\sum_{j=i+1}^ky_j):\ y_1\ge y_k \}, \end{equation} where $(y_1, \cdots, y_k)$ is any permutation of $(d_1-1, \cdots, d_k-1)$. \end{theorem} \begin{proof} Necessity. Since $T$ is a maximum optimal tree in ${\mathcal{T}}_{\pi}$, by Lemmas~\ref{shi}, $T$ must be a caterpillar and can be denoted by $T(z_1, \cdots, z_k)$ with $(z_1, \cdots, z_k)$ is the permutation of $(d_1-1, \cdots, d_k-1)$. Moreover, by Lemma~\ref{formula}, we have $$W(T(z_1, \cdots, z_k))=(n-1)^2+F(z_1, \cdots, z_k).$$ For any permutation $(y_1, \cdots, y_k)$ of $(d_1-1, \cdots, d_k-1)$ with $y_1\ge y_k$, there exists a caterpillar $T_1$ with the degree sequence $\pi$ such that $$W(T_1)=(n-1)^2+F(y_1, \cdots, y_k).$$ Because $T(z_1, \cdots, z_k)$ is a maximum optimal tree in ${\mathcal{T}}_{\pi}$, we have $$F(y_1, \cdots, y_k)=W(T_1)-(n-1)^2\le W(T(z_1, \cdots, z_k))-(n-1)^2=F(z_1, \cdots, z_k).$$ Sufficiency. If $T$ is a caterpillar $T(x_1, \cdots, x_k)$ and $(x_1, \cdots, x_k)$ satisfies \begin{equation} F(x_1, \cdots , x_k)=\max\{F(y_1, \cdots, y_k)=\sum_{i=1}^{k-1}(\sum_{j=1}^iy_j)(\sum_{j=i+1}^ky_j): y_1\ge y_k \}, \end{equation} where the maximum is taken over all permutations $(y_1, \cdots, y_k)$ of $(d_1-1, \cdots, d_k-1)$. Let $T_1$ be any tree with the degree sequence $\pi$. By Lemma~\ref{shi}, there exists a caterpillar $T_2$ with the degree sequence $\pi$ such that $W(T_1)\le W(T_2)$. Then $T_2$ must be $T(y_1, \cdots, y_k)$, where $(y_1, \cdots, y_k)$ is the permutation of $(d_1-1, \cdots, d_k-1)$. Hence $$W(T_1)\le W(T_2)= (n-1)^2+F(y_1, \cdots, y_k)\le (n-1)^2+F(x_1, \cdots, x_k)=W(T(x_1, \cdots, x_k)).$$ Therefore $T(x_1, \cdots, x_k)$ is a maximum optimal tree. This completes the proof. \end{proof} Now we can present an upper bound for the Wiener index of any tree with given degree sequence $\pi$ in terms of degree sequences. \begin{theorem}\label{upperbound} Let $T$ be a tree with a given degree sequence $\pi=(d_1,\cdots, d_n)$, where $d_1\ge \cdots\ge d_k>d_{k+1}=\cdots=d_n=1$. Then \begin{equation} W(T)\le (n-1)^2+\frac{k(k-1)}{4}\sum_{i=1}^k(d_i-1)^2 \end{equation} with equality if and only if $k=2$ and $d_1=d_2$. \end{theorem} \begin{proof} Let $T(x_1, \cdots, x_k)$ be a caterpillar and $(x_1, \cdots, x_k)$ satisfy \begin{equation} F(x_1, \cdots , x_k)=\max\{F(y_1, \cdots, y_k)=\sum_{i=1}^{k-1}(\sum_{j=1}^iy_j)(\sum_{j=i+1}^ky_j): y_1\ge y_k \}, \end{equation} where $(y_1, \cdots, y_k)$ is any permutation of $(d_1-1, \cdots, d_k-1)$. By Theorem~\ref{optimal-equal}, $W(T)\le W(T(x_1, \cdots, x_k))$. Clearly, \begin{eqnarray*} F(x_1, \cdots, x_k)=\sum_{i=1}^{k-1}(\sum_{j=1}^ix_j)(\sum_{j=i+1}^kx_j) =\frac{1}{2}(x_1, \cdots, x_k)C(x_1,\cdots, x_k)^T, \end{eqnarray*} where $$ C=\left(\begin{array}{cccccc} 0 &1& 2 &\cdots & k-2& k-1\\ 1 & 0& 1& \cdots & k-3& k-2\\ \cdots & \cdots & \cdots & \cdots & \cdots & \cdots \\ k-1& k-2 & k-3&\cdots & 1 &0 \end{array}\right).$$ By Perron-Frobenius theorem (for example, see \cite{horn1985}), the largest eigenvalue $\lambda_1(C)$ of $C$ is at most $\frac{k(k-1)}{2}$ with equality if and only if $k=2$. Hence by Rayleigh quotient, $$(x_1, \cdots, x_k)C(x_1,\cdots, x_k)^T\le \lambda_1(C)\sum_{i=1}^kx_i^2$$ with equality if and only if $(x_1, \cdots, x_k)^T$ is an eigenvector of $C$ corresponding to the eigenvalue $\lambda_1(C)$. Therefore, $$F(x_1, \cdots, x_k)\le \frac{k(k-1)}{4}\sum_{i=1}^kx_i^2$$ with equality if and only if $k=2$ and $x_1=x_2$. Hence $$ W(T)\le (n-1)^2+\frac{k(k-1)}{4}\sum_{i=1}^k{x_i}^2\le (n-1)^2+\frac{k(k-1)}{4}\sum_{i=1}^k(d_i-1)^2$$ with equality if and only if $k=2$ and $d_1=d_2$, since $(d(v_1), \cdots, d(v_k))$ is a permutation of $(d_1, \cdots, d_k)$. This completes the proof. \end{proof} \begin{lemma}\label{function} Let $w_1\ge w_2\ge\cdots\ge w_k\ge 1$ be the positive integers with $k\ge 5$. Let $$F(z_1, \cdots, z_k)=\max\{F(y_1, \cdots, y_k)=\sum_{i=1}^{k-1}(\sum_{j=1}^iy_j)(\sum_{j=i+1}^ky_j): y_1\ge y_k\},$$ where $(y_1, \cdots, y_k)$ is any permutation of $(w_1, \cdots, w_k)$. Then there exists a $2\le t\le k-2$ such that the following holds: \begin{equation}\label{z1-zt} z_1+\cdots+ z_{t-2}\le z_{t+1}+\cdots+z_k \end{equation} and \begin{equation}\label{zt-zk} z_1+\cdots+z_{t-1}>z_{t+2}+\cdots+z_k. \end{equation} Further, if equations (\ref{z1-zt}) is strict, then \begin{equation}\label{z1-zt-ztrict} z_1\ge z_2\ge\cdots\ge z_t, \quad \quad z_t\le z_{t+1}\le \cdots\le z_k. \end{equation} If equations (\ref{z1-zt}) becomes equality, then \begin{equation}\label{lemma-z1-zt-1} z_1\ge z_2\ge\cdots\ge z_t, \quad \quad z_t\le z_{t+1}\le \cdots\le z_k \end{equation} or \begin{equation}\label{lemma-z1-zt-2} z_1\ge z_2\ge\cdots\ge z_{t-1}, \quad \quad z_{t-1}\le z_{t}\le \cdots\le z_k. \end{equation} \end{lemma} \begin{proof} Let $$f(p)=\sum_{i=1}^{p-2}z_i-\sum_{i=p+1}^kz_i,\ \ \ 2\le p\le k-2.$$ Clearly $f(2)<0$, $f(k-1)>0$ and $$f(2)\le f(3)\le\cdots\le f(k-1).$$ Hence there exists a $ 2\le t\le k-2$ such that $f(t)\le 0$ and $f(t+1)>0$. In other words, equations (\ref{z1-zt}) and (\ref{zt-zk}) hold. By the definition of $F(z_1, \cdots, z_k),$ we have for $1\le i\le k-1$, \begin{eqnarray*} 0&\le& F(z_1, \cdots,z_{i-1}, z_i, z_{i+1}, \cdots, z_k)-F(z_1, \cdots, z_{i-1}, z_{i+1}, z_i, \cdots, z_k)\\ &=&(z_{i+1}-z_i)(\sum_{j=1}^{i-1}z_j-\sum_{j=i+2}^kz_j) . \end{eqnarray*} But for $1\le i\le t-2$, by (\ref{z1-zt}), we have $\sum_{j=1}^{i-1}z_j<\sum_{j=i+2}^kz_j$. Hence $z_1\ge \cdots\ge z_{t-1}$. On the other hand, for $t\le i\le k-1$, by (\ref{zt-zk}), we have $\sum_{j=1}^{i-1}z_j>\sum_{j=i+2}^kz_j$. Therefore $z_t\le z_{t+1}\cdots\le z_k$. If (\ref{z1-zt}) is strict, then $(z_1+\cdots+z_{t-2})-(z_{t+1}+\cdots+z_k)<0$, which implies $z_{t-1}\ge z_t$. So (\ref{z1-zt-ztrict}) holds. If (\ref{z1-zt}) becomes equality, i.e., $z_1+\cdots+z_{t-2}=z_{t+1}+\cdots+z_k$, then it is easy to see that (\ref{lemma-z1-zt-1}) or (\ref{lemma-z1-zt-2}) holds. This completes the proof. \end{proof} \begin{corollary}\label{k=6fun} Let $w_1\ge w_2\ge\cdots\ge w_6\ge 1$ be the positive integers. Let $$F(z_1, \cdots, z_6)=\max\{F(y_1, \cdots, y_6)= \sum_{i=1}^{5}(\sum_{j=1}^iy_j)(\sum_{j=i+1}^6y_j):\ \ y_1\ge y_6\},$$ where $(y_1, \cdots, y_6)$ is any permutation of $(w_1, \cdots, w_6)$. Then $(z_1, \cdots, z_6)$ is equal to one of the following five $(w_1, w_6, w_5, w_4, w_3, w_2)$, $(w_1, w_5, w_6, w_4, w_3, w_2)$, $(w_1, w_4, w_6, w_5, w_3, w_2)$, $(w_1, w_4, w_5, w_6, w_3, w_2)$ and $(w_1, w_3, w_6, w_5, w_4, w_2)$. \end{corollary} \begin{proof} By Lemma~\ref{function}, there are just three cases: {\bf Case 1} $t=2$. Then by Lemma~\ref{function}, $z_1\ge z_2$ and $z_2\le z_3\le z_4\le z_5\le z_6$. Hence $(z_1, \cdots, z_6)$ must be $(w_1, w_6, w_5, w_4, w_3, w_2)$. {\bf Case 2} $t=3$. Then $z_1\le z_4+ z_5+z_6$ and $z_1+z_2>z_5+z_6.$ Moreover, $z_1\ge z_2\ge z_3$ and $ z_3\le z_4\le z_5\le z_6$; or $z_1\ge z_2$ and $z_2\le z_3\le z_4\le z_5\le z_6$. Therefore $(z_1, \cdots, z_6)$ must be one of $(w_1, w_6, w_5, w_4, w_3, w_2)$, $(w_1, w_5, w_6, w_4, w_3, w_2)$, $(w_1, w_4, w_6, w_5, w_3, w_2)$ and $(w_1, w_3, w_6, w_5, w_4, w_2)$. {\bf Case 3} $t=4$. Then $z_1+z_2\le z_5+z_6$. Moreover, $z_1\ge z_2\ge z_3$ and $ z_3\le z_4\le z_5\le z_6$; or $z_1\ge z_2\ge z_3\ge z_4$ and $z_4\le z_5\le z_6$. Therefore, $(z_1, \cdots, z_6)$ must be one of $(w_1, w_4, w_6, w_5, w_3, w_2)$, $(w_1, w_5, w_6, w_4, w_3, w_2)$ and $(w_1, w_4, w_5, w_6, w_3, w_2)$. This completes the proof. \end{proof} \begin{theorem}\label{char} Let $\pi=(d_1, \cdots, d_n)$ be a tree degree sequence with $d_1\ge d_2\ge\cdots\ge d_k\ge 2$, $d_{k+1}=\cdots=d_n=1$ and $k\ge 5$. If a caterpillar $T(x_1, \cdots, x_k)$ is a maximum optimal tree in ${\mathcal{T}}_{\pi}$ with $F(x_1, \cdots, x_k)$ in equation (\ref{weiner-f}). Then there exists a $2\le t\le k-2$ such that either $$\sum_{i=1}^{t-2}x_i\le\sum_{i=t+1}^kx_i,\quad \sum_{i=1}^{t-1}x_i>\sum_{t+2}^kx_i,\quad x_1\ge x_2\ge\cdots\ge x_{t-1}\ge x_t, \quad x_{t}\le x_{t+1}\le \cdots\le x_k;$$ or $$\sum_{i=1}^{t-2}x_i=\sum_{i=t+1}^kx_i,\quad \sum_{i=1}^{t-1}x_i>\sum_{t+2}^kx_i, \quad x_1\ge x_2\ge\cdots\ge x_{t-1}\ge x_t, \quad x_{t}\le x_{t+1}\le \cdots\le x_k;$$ or $$\sum_{i=1}^{t-2}x_i=\sum_{i=t+1}^kx_i,\quad \sum_{i=1}^{t-1}x_i>\sum_{t+2}^kx_i, \quad x_1\ge x_2\ge\cdots\ge x_{t-1}, \quad x_{t-1}\le x_{t}\le \cdots\le x_k.$$ \end{theorem} \begin{proof} It follows from Theorem~\ref{optimal-equal} and Lemma~\ref{function} that the assertion holds. \end{proof} \section{The maximum optimal tree with many leaves} In this section, for a given degree sequence $\pi=(d_1, \cdots, d_n)$ with at least $n-6$ leaves, we give the maximum optimal trees with the maximum Wiener index in ${\mathcal{T}}_{\pi}$. Moreover, the maximum optimal tree may be not unique. \begin{theorem}\label{k=2--4} Let $\pi=(d_1, \cdots,d_k, \cdots, d_n)$ be tree degree sequence with $n-k$ leaves for $2\le k\le 4.$ Then the maximum optimal tree in ${\mathcal T}_{ \pi}$ is the greedy caterpillar. In other words, if $k=2$, then $W(T)=(n-1)^2+(d_1-1)(d_2-1)$, for $T\in { \mathcal{T}}_{\pi}.$ If $k=3$, then for any $T\in {\mathcal{T}}_{\pi},$ $$W(T)\le (n-1)^2+(d_1-1)(d_2+d_3-2)+(d_1+d_2-2)(d_3-1)$$ with equality if and only if $T$ is the caterpillar $T(d_1-1, d_3-1, d_2-1).$ If $k=4,$ then for any $T\in {\mathcal{T}}_{\pi},$ $$W(T)\le (n-1)^2+(d_1-1)(d_2+d_3+d_4-3)+(d_1+d_2-2)(d_3+d_4-2)+(d_1+d_2+d_3-3)(d_4-1)$$ with equality if and only if $T$ is the caterpillar $T(d_1-1, d_4-1, d_3-1, d_2-1)$. \end{theorem} \begin{proof} If $k=2$, it is obvious. If $k=3$, it is easy to see that $F(d_1-1, d_2-1, d_3-1)\le F(d_1-1, d_3-1, d_2-1). $ By Theorem~\ref{optimal-equal}, the assertion holds. If $k=4$, then by Theorem~\ref{optimal-equal}, let $T$ be a caterpillar $T(x_1, x_2, x_3, x_4)$ and $$F(x_1, x_2, x_3, x_4)=\max\{F(y_1, y_2, y_3, y_4): y_1\ge y_4\},$$ where $(y_1, y_2, y_3, y_4)$ is any permutation of $(d_1-1, d_2-1, d_3-1, d_4-1)$. Because $$F(x_1, x_2, x_3, x_4)-F(x_2, x_1, x_3, x_4)=(x_1-x_2)(x_3+x_4)\ge 0$$ and $$ F(x_1, x_2, x_3, x_4)-F(x_1, x_2, x_4, x_3)=(x_4-x_3)(x_1+x_2)\ge 0,$$ we have $x_1\ge x_2$ and $x_4\ge x_3$. So $(x_1, x_2, x_3, x_4)=(d_1-1, d_4-1, d_3-1, d_2-1)$. This completes the proof. \end{proof} \begin{theorem}\label{k=5} Let $\pi=(d_1, \cdots,d_k, \cdots, d_n)$ be tree degree sequence with $n-5$ leaves. (1). If $d_1> d_2+d_3$, then the maximum optimal tree in ${\mathcal T}_{ \pi}$ is the only caterpillar $T(d_1-1, d_5-1, d_4-1, d_3-1, d_2-1)$. (2). If $d_1=d_2+d_3$, then there are the exactly two maximum optimal trees in ${\mathcal T}_{ \pi}$: one tree is the caterpillar $T(d_1-1, d_5-1, d_4-1, d_3-1, d_2-1)$; the other tree is the caterpillar $T(d_1-1, d_4-1, d_5-1, d_3-1, d_2-1)$. (3). If $d_1< d_2+d_3$, then the maximum optimal tree in ${\mathcal T}_{ \pi}$ is the only caterpillar $T(d_1-1, d_4-1, d_5-1, d_3-1, d_2-1)$. \end{theorem} \begin{proof} By Theorem\ref{optimal-equal}, let $T(x_1, x_2, x_3, x_4, x_5)$ be a maximum optimal tree in ${\mathcal{T}}_{\pi}$. If $d_1>d_2+d_3$, then by Theorem~\ref{char}, it is easy to see that $t=2$, and $x_1\ge x_2$ and $x_2\le x_3\le x_4\le x_5$. Hence $(x_1, x_2, x_3, x_4, x_5)=(d_1-1, d_5-1, d_4-1, d_3-1, d_2-1)$. If $d_1<d_2+d_3$, then by Theorem~\ref{char}, it is easy to see that $x_1\ge x_2\ge x_3$ and $x_3\le x_4\le x_5$. Hence $(x_1, x_2, x_3, x_4, x_5)=(d_1-1, d_4-1, d_5-1, d_3-1, d_2-1)$ or $(d_1-1, d_3-1, d_5-1, d_4-1, d_2-1)$. But $W(T(d_1-1, d_4-1, d_5-1, d_3-1, d_2-1))-W(T(d_1-1, d_3-1, d_5-1, d_4-1, d_2-1))=2(d_1-d_2)(d_3-d_4)\ge 0$ with equality if and only if $d_1=d_2$ or $d_3=d_4$. Hence the assertion (3) holds. If $d_1=d_2+d_3$, then by Theorem ~\ref{char}, it is easy to see that either $x_1\ge x_2$ and $x_2\le x_3\le x_4\le x_5$; or $x_1\ge x_2\ge x_3$ and $x_3\le x_4\le x_5$. Hence $(x_1, x_2, x_3, x_4, x_5)=(d_1-1, d_5-1, d_4-1, d_3-1, d_2-1)$ or $(d_1-1, d_4-1, d_5-1, d_3-1, d_2-1)$. Moreover, $F(d_1-1, d_5-1, d_4-1, d_3-1, d_2-1)=F (d_1-1, d_4-1, d_5-1, d_3-1, d_2-1)$. Hence (2) holds. \end{proof} \begin{lemma}\label{6diff} Let $w_1\ge w_2\ge\cdots\ge w_6\ge 1$ be positive integers and $$ F(y_1, \cdots, y_k)=\sum_{i=1}^{k-1}(\sum_{j=1}^iy_j)(\sum_{j=i+1}^ky_j).$$ Then \begin{equation}\label{k61} F(w_1, w_6, w_5, w_4, w_3, w_2)-F(w_1, w_5, w_6, w_4, w_3, w_2)=(w_1-w_2-w_3-w_4)(w_5-w_6), \end{equation} \begin{equation}\label{k62} F(w_1, w_5, w_6, w_4, w_3, w_2)-F(w_1, w_4, w_6, w_5, w_3, w_2)=2(w_1-w_2-w_3)(w_4-w_5), \end{equation} \begin{equation}\label{k63} F(w_1, w_4, w_6, w_5, w_3, w_2)-F(w_1, w_4, w_5, w_6, w_3, w_2)=(w_1+w_4-w_2-w_3)(w_5-w_6), \end{equation} \begin{equation}\label{k64} F(w_1, w_4, w_5, w_6, w_3, w_2)-F(w_1, w_3, w_6, w_5, w_4, w_2)=(3w_3-3w_4-w_5+w_6)(w_1-w_2). \end{equation} \end{lemma} \begin{proof} By a simple calculation, it is easy to see that the assertion holds. \end{proof} \begin{theorem}\label{k=6} Let $\pi=(d_1, \cdots,d_6, \cdots, d_n)$ be tree degree sequence with $n-6$ leaves, i.e., $d_1\ge\cdots\ge d_6\ge 2$ and $d_7=\cdots=d_n=1$. (1). If $d_1>d_2+d_3+d_4-2$, then there is only one maximum optimal tree $T(d_1-1, d_6-1, d_5-1, d_4-1, d_3-1, d_2-1)$ in ${\mathcal T}_{ \pi}$. (2). If $d_1=d_2+d_3+d_4-2$, then there are exactly two maximum optimal trees in ${\mathcal T}_{ \pi}$: one maximum optimal tree is $T(d_1-1, d_6-1, d_5-1, d_4-1, d_3-1, d_2-1)$; the other maximum optimal tree is $T(d_1-1, d_5-1, d_6-1, d_4-1, d_3-1, d_2-1)$. (3). $d_2+d_3-1<d_1<d_2+d_3+d_4-2$, then there is only one maximum optimal tree $T(d_1-1, d_5-1, d_6-1, d_4-1, d_3-1, d_2-1)$ in ${\mathcal T}_{ \pi}$. (4). If $d_2+d_3-1=d_1$, then there are exactly two maximum optimal trees in ${\mathcal T}_{ \pi}$: one maximum optimal tree is $T(d_1-1, d_5-1, d_6-1, d_4-1, d_3-1, d_2-1)$; the other maximum optimal tree is $T(d_1-1, d_4-1, d_6-1, d_5-1, d_3-1, d_2-1)$. (5). If $\ \ \max\{d_2+d_3-d_4,\ d_2+\frac{1}{3}(d_5-d_6)\}<d_1<d_2+d_3-1$, then there is only one maximum optimal tree $T(d_1-1, d_4-1, d_6-1, d_5-1, d_3-1, d_2-1)$ in ${\mathcal T}_{ \pi}$. (6). If $d_1=d_2+d_3-w_4> d_2+\frac{1}{3}(d_5-d_6)$, then there are exactly two maximum optimal trees in ${\mathcal T}_{ \pi}$: one maximum optimal tree is $T(d_1-1, d_4-1, d_6-1, d_5-1, d_3-1, d_2-1)$; the other maximum optimal tree is $T(d_1-1, d_4-1, d_5-1, d_6-1, d_3-1, d_2-1)$. (7). If $d_1= d_2+\frac{1}{3}(d_5-d_6)>d_2+d_3-d_4$, then there are exactly two maximum optimal trees in ${\mathcal T}_{ \pi}$: one maximum optimal tree is $T(d_1-1, d_4-1, d_6-1, d_5-1, d_3-1, d_2-1)$; the other maximum optimal tree is $T(d_1-1, d_3-1, d_6-1, d_5-1, d_4-1, d_2-1)$. (8). If $d_1=d_2+d_3-d_4= d_2+\frac{1}{3}(d_5-d_6)$, then there are exactly three maximum optimal trees in ${\mathcal T}_{ \pi}$: they are $T(d_1-1, d_4-1, d_6-1, d_5-1, d_3-1, d_2-1)$; $T(d_1-1, d_4-1, d_5-1, d_6-1, d_3-1, d_2-1)$ and $T(d_1-1, d_3-1, d_6-1, d_5-1, d_4-1, d_2-1)$. (9). If $d_2+\frac{1}{3}(d_5-d_6)\le d_1<d_2+d_3-d_4$, or $d_1\le d_2+\frac{1}{3}(d_5-d_6)< d_2+d_3-d_4$, then there is only one maximum optimal tree $T(d_1-1, d_4-1, d_5-1, d_6-1, d_3-1, d_2-1)$ in ${\mathcal T}_{ \pi}$. (10). If $d_2+d_3-d_4\le d_1<d_2+\frac{1}{3}(d_5-d_6)$; or $d_1\le d_2+d_3-d_4< d_2+\frac{1}{3}(d_5-d_6)$, then there is only one maximum optimal tree $T(d_1-1, d_3-1, d_6-1, d_5-1, d_4-1, d_2-1)$ in ${\mathcal T}_{ \pi}$. (11). If $d_1< d_2+\frac{1}{3}(d_5-d_6)= d_2+d_3-d_4$, then there are exactly two maximum optimal trees in ${\mathcal T}_{ \pi}$: one maximum optimal tree is $T(d_1-1, d_3-1, d_6-1, d_5-1, d_4-1, d_2-1)$; the other maximum optimal tree is $T(d_1-1, d_4-1, d_5-1, d_6-1, d_3-1, d_2-1)$. \end{theorem} \begin{proof} The proof is referred to appendix since it is technique. \end{proof} {\bf Remark}. From Theorem~\ref{k=6}, we can see that the maximum optimal trees depend on the values of all components of the tree degree sequences and not unique, while the minimum optimal tree is unique for a given tree degree sequence. Moreover, Theorem~\ref{k=6} explains that it seems to be difficult for characterize all the maximum optimal trees for a given tree degree sequence. \frenchspacing

https://arxiv.org/abs/0910.5456

Neighborhoods of univalent functions

The main result shows a small perturbation of a univalent function is again a univalent function, hence a univalent function has a neighborhood consisting entirely of univalent functions.For the particular choice of a linear function in the hypothesis of the main theorem, we obtain a corollary which is equivalent to the classical Noshiro-Warschawski-Wolff univalence criterion.We also present an application of the main result in terms of Taylor series, and we show that the hypothesis of our main result is sharp.

\section{Introduction} We denote by $U_{r}=\left\{ z\in \mathbb{C}:\left\vert z\right\vert <r\right\} $ the open disk of radius $r>0$ centered at the origin and we let $U=U_{1}$. The class of functions $f:D\rightarrow \mathbb{C}$ analytic in the domain $D$ will be denoted by $\mathcal{A}\left( D\right) $. It is known that if $f:D\rightarrow \mathbb{C}$ is a univalent map in a domain $D$, then $f^{\prime }\neq 0$ in $D$. The non-vanishing of the derivative of an analytic function (local univalence) is not in general sufficient to insure the univalence of the function, as it can be seen by considering for example the exponential function $f\left( z\right) =e^{z}$ defined in the upper half-plane. The classical Noshiro-Warschawski-Wolff univalence criterion gives a partial converse of the above result, as follows: \begin{theorem} If $f:D\rightarrow \mathbb{C}$ is analytic in the convex domain $D$ and \begin{equation*} \func{Re}f^{\prime }\left( z\right) >0,\qquad z\in D, \end{equation*}% then $f$ is univalent in $D$. \end{theorem} In the present paper we introduce the constant $K\left( f,D\right) $ associated with a function $f:D\rightarrow \mathbb{C}$ analytic in a domain $D$, which is a measure of the "degree of univalence" of $f$ (see Proposition \ref{characterization of univalence} and the remark following it). Using the constant $K\left( f,D\right) $ thus introduced, in Theorem \ref{main theorem} we obtain a sufficient condition for univalence, which shows that a small perturbation of a univalent function is again univalent. As a theoretical consequence of this result, it follows that a univalent function has a neighborhood consisting entirely of univalent functions (see Remark \ref{Nbds of univalent functions}). The Theorem \ref{main theorem} is sharp, in the sense that we cannot replace the upper bound appearing in the hypothesis of this theorem by a larger one, as shown in Example \ref{Exemplul 1}. For the particular choice of a linear function in Theorem \ref{main theorem}, we obtain a simple sufficient condition for univalence (Corollary \ref{corollary of main theorem 2}), which is shown to be equivalent to the Noshiro-Warschawski-Wolff univalence criterion. The main result in Theorem \ref{main theorem} can be viewed therefore as a generalization of this classical result, in which the linear function is replaced by a general univalent function. The paper concludes with another application of the main result in the case of analytic functions defined in the unit disk. Thus, in Theorem \ref{Application to Taylor series} and the corollary following it, we obtain sufficient conditions for the univalence of an analytic function defined in the unit disk in terms of the coefficients of its Taylor series representation, which might be of independent interest. \section{Main results} Given a function $f:D\rightarrow \mathbb{C}$ analytic in the domain $D$ we introduce the constant $K\left( f,D\right) $ defined as follows: \begin{equation} K\left( f,D\right) =\inf_{\substack{ a,b\in D \\ a\neq b}}\left\vert \frac{% f\left( a\right) -f\left( b\right) }{a-b}\right\vert \end{equation} Note that from the definition follows immediately that if the function $f$ is not univalent in $D$ then $K\left( f,D\right) =0$ The constant $K\left( f,D\right) $ characterizes the univalence of the function $f$ in $D$ in the following sense: \begin{proposition} \label{characterization of univalence}Let $f:D\rightarrow \mathbb{C}$ be an analytic function in the domain $D$. If $K\left( f,D\right) >0$ then $f$ is univalent in $D$. Conversely, if $f$ is univalent in $D$ and $\Omega \subset \overline{\Omega }% \subset D$ is a domain strictly contained in $D$, then $K\left( f,\Omega \right) >0$. \end{proposition} \begin{proof} The first statement follows from the inequality \begin{equation*} \left\vert f\left( a\right) -f\left( b\right) \right\vert \geq \left\vert a-b\right\vert K\left( f,D\right) >0, \end{equation*}% for any distinct points $a,b\in D$. To prove the converse, note that% \begin{equation*} K\left( f,\Omega \right) \geq \inf_{a,b\in \overline{\Omega }}\left\vert F\left( a,b\right) \right\vert , \end{equation*}% where $F:\overline{\Omega }\times \overline{\Omega }\rightarrow \mathbb{C}$ is the function defined by% \begin{equation*} F\left( a,b\right) =\left\{ \begin{array}{l} \frac{f\left( a\right) -f\left( b\right) }{a-b},\qquad a\neq b \\ f^{\prime }\left( a\right) ,\qquad \quad a=b% \end{array}% \right. . \end{equation*} Note that since $f:D\rightarrow \mathbb{C}$ is analytic, $F$ is continuous on the closed set $\overline{\Omega }\times \overline{\Omega }$, and therefore $F$ attains its minimum modulus on this set:% \begin{equation*} K\left( f,\Omega \right) \geq \inf_{a,b\in \overline{\Omega }}\left\vert F\left( a,b\right) \right\vert =\left\vert F\left( \alpha ,\beta \right) \right\vert , \end{equation*}% for some $\alpha ,\beta \in \overline{\Omega }$. If $\alpha \neq \beta $, then $\left\vert F\left( \alpha ,\beta \right) \right\vert =\left\vert \frac{f\left( \alpha \right) -f\left( \beta \right) }{\alpha -\beta }\right\vert >0$ since $\alpha ,\beta \in \overline{\Omega }% \subset D$ and $f$ is univalent in $D$, and if $\alpha =\beta $ then $% \left\vert F\left( \alpha ,\alpha \right) \right\vert =\left\vert f^{\prime }\left( \alpha \right) \right\vert >0$, again by the univalence of $f$ in $D$% . It follows that in all cases we have% \begin{equation*} K\left( f,\Omega \right) \geq \left\vert F\left( \alpha ,\beta \right) \right\vert >0, \end{equation*}% concluding the proof. \end{proof} \begin{remark} \label{K might be 0}Note that the converse in the above proposition may not hold for $\Omega =D$ without the additional hypothesis, as shown in the example below. In order to have the equivalence% \begin{equation*} f\text{ univalent in }D\Longleftrightarrow \text{ }K\left( f,D\right) >0, \end{equation*}% one needs additional hypotheses, which guarantee the existence of a continuous extension of $f,f^{\prime }$ to $\overline{D}$, such that $f$ is injective on $\overline{D}$ and $f^{\prime }\neq 0$ in $\overline{D}$. For example, in the case $D=U$, if the boundary of the image domain $f\left( U\right) $ is a Jordan curve of class $C^{1,\alpha }$ $\left( 0<\alpha <1\right) $, by Carath\'{e}odory theorem the function $f$ has a continuous injective extension to $\overline{D}$, and also, by Kelogg-Warschawski theorem, the function $f^{\prime }$ has continuous extension to $\overline{D} $, with $f^{\prime }\neq 0$ in $\overline{D}$ (see for example \cite% {Pommerenke}, p. 24 and pp. 48 -- 49). Following the proof above with $% \Omega $ replaced by $U$, we obtain $K\left( f,U\right) >0$, and therefore in this case we have \begin{equation*} f\text{ univalent in }U\Longleftrightarrow \text{ }K\left( f,U\right) >0. \end{equation*} \end{remark} \begin{example} Let $D=U-\left[ 0,1\right] $ be the unit disk with a slit along the positive real axis. Since $D$ is simply connected, there exists a conformal map $% f:U\rightarrow D$ between the unit disk $U$ and $D$ (see Figure \ref{Figura 1} below). The map $f$ has a continuous extension to $\overline{U}$, and without loss of generality we may assume that there exists $\theta \in (0,2\pi )$ such that $f\left( e^{i\theta }\right) =f\left( e^{-i\theta }\right) \in \left( 0,1\right) $. The function $f$ is univalent in $U$, but $K\left( f,U\right) =0$ since \begin{equation*} K\left( f,U\right) \leq \lim_{\substack{ a\rightarrow e^{i\theta } \\ % b\rightarrow e^{-i\theta }}}\left\vert \frac{f\left( a\right) -f\left( b\right) }{a-b}\right\vert =\left\vert \frac{f\left( e^{i\theta }\right) -f\left( e^{-i\theta }\right) }{e^{i\theta }-e^{-i\theta }}\right\vert =0. \end{equation*} \end{example} \begin{figure}[thb] \begin{center} \includegraphics[scale=0.6]{fig1} \caption{An example of a univalent function in $U$ for which $K\left( f,U\right) =0$.} \label{Figura 1} \end{center} \end{figure} The main result is contained in the following: \begin{theorem} \label{main theorem}Let $f:D\rightarrow \mathbb{C}$ be a non-constant analytic function in the convex domain $D$. If there exists an analytic function $g:D\rightarrow \mathbb{C}$ univalent in $D$ such that% \begin{equation} \left\vert f^{\prime }\left( z\right) -g^{\prime }\left( z\right) \right\vert \leq K\left( g,D\right) ,\qquad z\in D, \label{sufficient condition for univalency} \end{equation}% then the function $f$ is also univalent in $D$. \end{theorem} \begin{proof} Assuming that $f$ is not univalent in $D$, there exists distinct points $% z_{1,2}\in D$ such that $f\left( z_{1}\right) =f\left( z_{2}\right) $. Integrating the derivative of $f-g$ along the line segment $\left[ z_{1},z_{2}\right] \subset D$ and using the hypothesis (\ref{sufficient condition for univalency}) we obtain% \begin{eqnarray*} \left\vert g\left( z_{2}\right) -g\left( z_{1}\right) \right\vert &=&\left\vert \left( f\left( z_{2}\right) -g\left( z_{2}\right) \right) -\left( f\left( z_{1}\right) -g\left( z_{1}\right) \right) \right\vert \\ &=&\left\vert \int_{\left[ z_{1},z_{2}\right] }f^{\prime }\left( z\right) -g^{\prime }\left( z\right) dz\right\vert \\ &\leq &\int_{\left[ z_{1},z_{2}\right] }\left\vert f^{\prime }\left( z\right) -g^{\prime }\left( z\right) \right\vert \left\vert dz\right\vert \\ &\leq &\int_{\left[ z_{1},z_{2}\right] }K\left( g,D\right) \left\vert dz\right\vert \\ &=&K\left( g,D\right) \left\vert z_{1}-z_{2}\right\vert . \end{eqnarray*} Since the points $z_{1,2}$ are assumed to be distinct, from the definition of the constant $K\left( g,D\right) $ we obtain equivalently% \begin{equation} \left\vert \frac{g\left( z_{2}\right) -g\left( z_{1}\right) }{z_{2}-z_{1}}% \right\vert \leq K\left( g,D\right) =\inf_{\substack{ a,b\in D \\ a\neq b}}% \left\vert \frac{g\left( a\right) -g\left( b\right) }{a-b}\right\vert \leq \left\vert \frac{g\left( z_{2}\right) -g\left( z_{1}\right) }{z_{2}-z_{1}}% \right\vert , \end{equation}% and therefore% \begin{equation} K\left( g,D\right) =\inf_{\substack{ a,b\in D \\ a\neq b}}\left\vert \frac{% g\left( a\right) -g\left( b\right) }{a-b}\right\vert =\left\vert \frac{% g\left( z_{2}\right) -g\left( z_{1}\right) }{z_{2}-z_{1}}\right\vert . \label{minimum attained} \end{equation} Consider now the auxiliary function $G:D-\left\{ z_{2}\right\} \rightarrow \mathbb{C}$ defined by \begin{equation} G\left( z\right) =\frac{g\left( z\right) -g\left( z_{2}\right) }{z-z_{2}}% ,\qquad z\in D-\left\{ z_{2}\right\} , \end{equation}% and note that since $g$ is analytic in $D$, $G$ is also analytic in $% D-\left\{ z_{2}\right\} $ and moreover the limit \begin{equation} \lim_{z\rightarrow z_{2}}G\left( z\right) =\lim_{z\rightarrow z_{2}}\frac{% g\left( z\right) -g\left( z_{2}\right) }{z-z_{2}}=g^{\prime }\left( z_{2}\right) \end{equation}% exists and it is finite. The function $G$ can be therefore extended by continuity to an analytic function in $D$, denoted also by $G$. Since% \begin{equation*} \inf_{z\in D}\left\vert G\left( z\right) \right\vert =\inf_{\substack{ z\in D \\ z\neq z_{2}}}\left\vert G\left( z\right) \right\vert =\inf_{\substack{ % z\in D \\ z\neq z_{2}}}\left\vert \frac{g\left( z\right) -g\left( z_{2}\right) }{z-z_{2}}\right\vert \geq \inf_{\substack{ a,b\in D \\ a\neq b }}\left\vert \frac{g\left( a\right) -g\left( b\right) }{a-b}\right\vert =K\left( g,D\right) , \end{equation*}% combining with (\ref{minimum attained}) we obtain that \begin{equation*} \inf_{z\in D}\left\vert G\left( z\right) \right\vert \geq K\left( g,D\right) =\left\vert \frac{g\left( z_{2}\right) -g\left( z_{1}\right) }{z_{2}-z_{1}}% \right\vert =\left\vert G\left( z_{1}\right) \right\vert \geq \inf_{z\in D}\left\vert G\left( z\right) \right\vert , \end{equation*}% which shows that minimum value of the modulus of $G$ in $D$ is attained at $% z_{1}$:% \begin{equation*} \inf_{z\in D}\left\vert G\left( z\right) \right\vert =\left\vert G\left( z_{1}\right) \right\vert . \end{equation*} However, since the function $g$ is univalent in $D$, from the definition of $% G$ it follows that $G\left( z\right) \neq 0$ for any $z\in D-\left\{ z_{2}\right\} $, and also $G\left( z_{2}\right) =g^{\prime }\left( z_{2}\right) \neq 0$, and therefore the function $G$ does not vanish in $D$. Applying the maximum modulus principle to the analytic function $1/G$ it follows that $\left\vert G\right\vert $ must be constant in $D$, and therefore $G$ is constant in $D$. It follows that \begin{equation} g\left( z\right) =g\left( z_{2}\right) +c\left( z-z_{2}\right) ,\qquad z\in D, \label{g must be linear} \end{equation}% for a certain constant $c\in \mathbb{C}$ (from the definition of $G$ it can be seen that the constant $c$ can be written in the form $c=g^{\prime }\left( z_{2}\right) e^{i\theta }$, for some $\theta \in \mathbb{R}$). The relation (\ref{g must be linear}) shows that $g$ is a linear function, and therefore the constant $K\left( g,D\right) $ becomes in this case% \begin{eqnarray*} K\left( g,D\right) &=&\inf_{\substack{ a,b\in D \\ a\neq b}}\left\vert \frac{g\left( a\right) -g\left( b\right) }{a-b}\right\vert \\ &=&\inf_{\substack{ a,b\in D \\ a\neq b}}\left\vert \frac{\left( g\left( z_{2}\right) +c\left( a-z_{2}\right) \right) -\left( g\left( z_{2}\right) +c\left( b-z_{2}\right) \right) }{a-b}\right\vert \\ &=&\inf_{\substack{ a,b\in D \\ a\neq b}}\left\vert \frac{c\left( a-b\right) }{a-b}\right\vert \\ &=&\left\vert c\right\vert . \end{eqnarray*} The hypothesis (\ref{sufficient condition for univalency}) of the theorem can be written therefore as follows% \begin{equation*} \left\vert f^{\prime }\left( z\right) -c\right\vert \leq \left\vert c\right\vert ,\qquad z\in D, \end{equation*}% which shows that either $f$ is linear in $D$ (and thus univalent, since $f$ is assumed to be non-constant in $D$), or the following strict inequality holds% \begin{equation*} \left\vert f^{\prime }\left( z\right) -c\right\vert <\left\vert c\right\vert ,\qquad z\in D. \end{equation*} Repeating the proof above with $g\left( z\right) \equiv cz$ we obtain% \begin{eqnarray*} \left\vert cz_{2}-cz_{1}\right\vert &=&\left\vert \left( f\left( z_{2}\right) -cz_{2}\right) -\left( f\left( z_{1}\right) -cz_{1}\right) \right\vert \\ &=&\left\vert \int_{\left[ z_{1},z_{2}\right] }f^{\prime }\left( z\right) -cdz\right\vert \\ & \leq & \int_{\left[ z_{1},z_{2}\right] }\left\vert f^{\prime }\left( z\right) -c\right\vert \left\vert dz\right\vert \\ &<& \left\vert c\right\vert \left\vert z_{2}-z_{1}\right\vert , \end{eqnarray*}% a contradiction. The contradiction obtained shows that the function $f$ is univalent in $D$, concluding the proof of the theorem. \end{proof} In the particular case $D=U$, from the previous thereom we obtain immediately the following sufficient criterion for univalence in the unit disk: \begin{theorem} \label{main theorem 2}Let $f:U\rightarrow \mathbb{C}$ be a non-constant analytic function in the unit disk. If there exists an analytic function $% g:U\rightarrow \mathbb{C}$ univalent in $U$ such that% \begin{equation} \left\vert f^{\prime }\left( z\right) -g^{\prime }\left( z\right) \right\vert \leq K\left( g,U\right) ,\qquad z\in U, \label{sufficient condition for univalency 2} \end{equation}% then the function $f$ is also univalent in $U$. \end{theorem} As a corollary of Theorem \ref{main theorem} we obtain immediately the following: \begin{corollary} \label{corollary of main theorem 2}If $f:D\rightarrow \mathbb{C}$ is non-constant and analytic in the convex domain $D$ and there exists $c>0$ such that% \begin{equation} \left\vert f^{\prime }\left( z\right) -c\right\vert \leq c,\qquad z\in D, \label{Noshiro-Warschawski-Wolff type condition} \end{equation}% then $f$ is univalent in $D$. \end{corollary} \begin{proof} Considering the univalent function $g:D\rightarrow \mathbb{C}$ defined by $g\left( z\right) =cz$, we have $g^{\prime }\left( z\right) =c$ for $z\in D$ and% \begin{equation*} K\left( g,D\right) =\inf_{\substack{ a,b\in D \\ a\neq b}}\left\vert \frac{g\left( a\right) -g\left( b\right) }{a-b}\right\vert =\inf_{\substack{ % a,b\in D \\ a\neq b}}\left\vert \frac{ca-cb}{a-b}\right\vert =c, \end{equation*}% and therefore the claim follows from Theorem \ref{main theorem} above. \end{proof} \begin{remark} Let us note that the previous corollary can also be obtained as a direct consequence of the classical Noshiro-Warschawski-Wolff univalence criterion, since the hypothesis (\ref{Noshiro-Warschawski-Wolff type condition}) implies the hypothesis% \begin{equation} \func{Re}f^{\prime }\left( z\right) >0,\qquad z\in D. \label{Noshiro-Warschawski-Wolff hypothesis} \end{equation}% of this theorem (the fact that the above inequality is a strict inequality follows from the maximum principle, the function $f$ being assumed to be non-constant in $D$). Conversely, the Noshiro-Warschawski-Wolff univalence criterion follows from the previous corollary. To see this, note that in order to prove the univalence of $f$, it suffices to prove the univalence of $f$ in $D_{r}=r D$, for an arbitrarily fixed $r\in (0,1)$. If the condition (\ref{Noshiro-Warschawski-Wolff hypothesis}) holds, there exists $c>0$ such that $$f^{\prime }\left( D_{r}\right) \subset \left\{ w\in \mathbb{C}:\left\vert w-c\right\vert <c\right\} ,$$ or equivalent% \begin{equation*} \left\vert f^{\prime }\left( z\right) -c \right\vert <c,\qquad z\in D_{r}. \end{equation*}% Applying Corollary \ref{corollary of main theorem 2} to the restriction of of $f$ to $D_r$, it follows that the function $f$ is univalent in $D_{r}.$ Since $r\in \left( 0,1\right) $ was arbitrarily fixed, it follows that $f$ is univalent in $U$, concluding the proof of the claim. \end{remark} The remark above shows that Corollary \ref{corollary of main theorem 2} and the Noshiro-Warschawski-Wolff univalence criterion are equivalent, and therefore Theorem \ref{main theorem} is a generalization of it. The Noshiro-Warschawski-Wolff univalence criterion can be viewed as a particular case of the main Theorem \ref{main theorem}, corresponding to the choice of a linear function $g$. \begin{remark} \label{Nbds of univalent functions}Fixing an arbitrarily univalent function $% g:U\rightarrow \mathbb{C}$ for which $K\left( g,U\right) \neq 0$ (see Remark % \ref{K might be 0} above), Theorem \ref{main theorem 2} shows that a whole neighborhood $V\left( g\right) =\left\{ f\in \mathcal{A}:\left\vert \left\vert f^{\prime }-g^{\prime }\right\vert \right\vert \leq K\left( g,U\right) \right\} $ of $g$ consists entirely of univalent functions in $U$ ($\left\vert \left\vert \cdot \right\vert \right\vert $ denotes here the supremum norm in the space $\mathcal{A}_{0}=\left\{ f\in \mathcal{A}:f\left( 0\right) =0\right\} $ of normalized analytic functions). Loosely stated, Theorem \ref{main theorem 2} shows that an univalent function has a neighborhood consisting entirely of univalent functions. \end{remark} The hypotheses of Theorem \ref{main theorem} and Theorem \ref{main theorem 2} are sharp, in the sense that we cannot replace the right side of the inequalities (\ref{sufficient condition for univalency}), respectively (\ref% {sufficient condition for univalency 2}), by larger constants, as can be seen from the following example. \begin{example} \label{Exemplul 1}Consider the function $f:U\rightarrow \mathbb{C}$ defined by $f\left( z\right) =z+az^{2}$, $z\in U$, where $a\in \mathbb{C}$ is a parameter. Using Theorem \ref{main theorem 2} above with $g\left( z\right) \equiv z$, for which $K\left( g,U\right) =1$, we obtain that the function $f$ is univalent in $U$ if \begin{equation*} \left\vert 2az\right\vert \le 1,\qquad z\in U, \end{equation*}% that is if $\left\vert 2a\right\vert \leq 1$. This result is sharp, since the function $f$ is univalent iff $\left\vert a\right\vert \leq \frac{1}{2}$, as it can be checked by direct computation. \end{example} The univalence of the function $f$ in the previous example can also be obtained by using the Noshiro-Warschawski-Wolff univalence criterion (for $% \left\vert a\right\vert \leq 1/2$ we have $\func{Re}f^{\prime }\left( z\right) >0$ for any $z\in U$). The next example shows that we may still use Theorem \ref{main theorem 2} also in situations when the Noshiro-Warschawski-Wolff univalence criterion cannot be applied: \begin{example} \label{Exemplul 2}Consider the linear map $g:U\rightarrow \mathbb{C}$ defined by $g\left( z\right) =\frac{z}{1-z}$. The function $g$ is univalent in $U$ and we have% \begin{equation*} K\left( g,U\right) =\inf_{\substack{ a,b\in U \\ a\neq b}}\left\vert \frac{% g\left( a\right) -g\left( b\right) }{a-b}\right\vert =\inf_{\substack{ % a,b\in U \\ a\neq b}}\left\vert \frac{\frac{a}{1-a}-\frac{b}{1-b}}{a-b}% \right\vert =\inf_{\substack{ a,b\in U \\ a\neq b}}\frac{1}{\left\vert 1-a\right\vert \left\vert 1-b\right\vert }=1. \end{equation*} The function $f:U\rightarrow \mathbb{C}$ defined by $f\left( z\right) =\frac{% z^{2}}{1-z}$ is analytic in $U$ and satisfies% \begin{equation*} \left\vert f^{\prime }\left( z\right) -g^{\prime }\left( z\right) \right\vert =1\leq K\left( g,U\right) ,\qquad z\in U, \end{equation*}% and therefore by Theorem \ref{main theorem 2} it follows that $f$ is univalent in the unit disk. The univalence of $f$ does not follow however by the Noshiro-Warschawski-Wolff univalence criterion since $\func{Re}f^{\prime }\left( z\right) $ takes (arbitrarily small) negative values for $z\in U$ sufficiently close to $1$. \end{example} As another application of Theorem \ref{main theorem 2}, in the next result we show that by perturbing the coefficients of the Taylor series of an univalent function, the resulting function is also univalent. More precisely, we have the following: \begin{theorem} \label{Application to Taylor series} Let $g:U\rightarrow \mathbb{C}$ be an analytic univalent function with Taylor series representation% \begin{equation} g\left( z\right) =\sum_{n=0}^{\infty }b_{n}z^{n},\qquad z\in U\text{.} \label{Taylor series for g} \end{equation} If the coefficients $a_{0},a_{1},\ldots \in \mathbb{C}$ satisfy the inequality \begin{equation} \sum_{n=1}^{\infty }n\left\vert a_{n}-b_{n}\right\vert <K\left( g,U\right) \label{hypothesis on coefficients of Taylor series} \end{equation}% then the function $f:U\rightarrow \mathbb{C}$ defined by \begin{equation} f\left( z\right) =\sum_{n=0}^{\infty }a_{n}z^{n},\qquad z\in U, \label{Taylor series for f} \end{equation}% is analytic and univalent in $U$. \end{theorem} \begin{proof} Since $g$ is univalent in $U$, the radius of convergence of the Taylor series (\ref{Taylor series for g}) is at least $1$, hence \begin{equation*} \lim \sup \sqrt[n]{\left\vert b_{n}\right\vert }\leq 1, \end{equation*}% and therefore $\left\vert b_{n}\right\vert \leq 1$ for all $n$ sufficiently large. Using the hypothesis (\ref{hypothesis on coefficients of Taylor series}) we obtain% \begin{equation*} \lim \sup \sqrt[n]{\left\vert a_{n}\right\vert }\leq \lim \sup \sqrt[n]{% \left\vert b_{n}\right\vert +\left\vert a_{n}-b_{n}\right\vert }\leq \lim \sup \sqrt[n]{1+\frac{K\left( g,U\right) }{n}}=1, \end{equation*}% and therefore the radius of convergence of the series in (\ref{Taylor series for f}) is at least $1$, thus the function $f$ is well defined by (\ref% {Taylor series for f}) and it is analytic in $U$. Since \begin{eqnarray*} \left\vert f^{\prime }\left( z\right) -g^{\prime }\left( z\right) \right\vert &=&\left\vert \sum_{n=0}^{\infty }na_{n}z^{n-1}-\sum_{n=0}^{\infty }nb_{n}z^{n-1}\right\vert \\ &\leq &\sum_{n=1}^{\infty }n\left\vert a_{n}-b_{n}\right\vert \left\vert z\right\vert ^{n-1} \\ &\leq &\sum_{n=1}^{\infty }n\left\vert a_{n}-b_{n}\right\vert \\ &<&K\left( g,U\right) , \end{eqnarray*}% for any $z\in U$, by Theorem \ref{main theorem 2} follows that $f$ is univalent in $U$, concluding the proof. \end{proof} Using a comparison with the generalized harmonic series, from the above we can obtain the following: \begin{corollary} Let $g:U\rightarrow \mathbb{C}$ be an analytic univalent function with Taylor series representation% \begin{equation} g\left( z\right) =\sum_{n=0}^{\infty }b_{n}z^{n},\qquad z\in U\text{.} \end{equation} If the coefficients $a_{0},a_{1},\ldots \in \mathbb{C}$ satisfy the inequality \begin{equation} \left\vert a_{n}-b_{n}\right\vert <K\left( g,U\right) \frac{\zeta \left( p\right) }{n^{p+1}},\qquad n=1,2,\ldots , \end{equation}% for some $p>1$ ($\zeta $ denotes the Riemann zeta function), then the function $f:U\rightarrow \mathbb{C}$ defined by \begin{equation} f\left( z\right) =\sum_{n=0}^{\infty }a_{n}z^{n},\qquad z\in U, \end{equation}% is analytic and univalent in $U$. \end{corollary} \begin{example} Considering the function $g\left( z\right) =\frac{z}{1-z}=\sum_{n=1}^{\infty }z^{n}$ defined in Example \ref{Exemplul 2}, which is analytic and univalent in $U$ and has $K\left( g,U\right) =1$, from the previous theorem it follows that the function $f:U\rightarrow \mathbb{C}$ defined by $f\left( z\right) =\sum_{n=0}^{\infty }a_{n}z^{n}$ is analytic and univalent in $U$ if the coefficients $a_{n}$ satisfy the inequality% \begin{equation*} \sum_{n=1}^{\infty }n\left\vert a_{n}-1\right\vert <1. \end{equation*} Using for example the fact that $\zeta \left( 2\right) =\frac{\pi ^{2}}{6}% \approx 1.645$, from the previous corollary it follows that the function $f$ is also analytic and univalent in $U$ if the coefficients $a_{n}$ satisfy the inequality% \begin{equation*} \left\vert a_{n}-1\right\vert \leq \frac{\pi ^{2}}{6n^{3}}\approx \frac{1.645% }{n^{3}},\qquad n=1,2,\ldots \end{equation*} \end{example}

https://arxiv.org/abs/1812.04874

Exponential convexifying of polynomials

Let $X\subset\mathbb{R}^n$ be a convex closed and semialgebraic set and let $f$ be a polynomial positive on $X$. We prove that there exists an exponent $N\geq 1$, such that for any $\xi\in\mathbb{R}^n$ the function $\varphi_N(x)=e^{N|x-\xi|^2}f(x)$ is strongly convex on $X$. When $X$ is unbounded we have to assume also that the leading form of $f$ is positive in $\mathbb{R}^n\setminus\{0\}$. We obtain strong convexity of $\varPhi_N(x)=e^{e^{N|x|^2}}f(x)$ on possibly unbounded $X$, provided $N$ is sufficiently large, assuming only that $f$ is positive on $X$. We apply these results for searching critical points of polynomials on convex closed semialgebraic sets.

\section{Introduction} In \cite{KS} we considered several questions concerning convexification of a polynomial $f$ which is positive on a closed convex set $X\subset \mathbb R^n$. One of the main results in \cite{KS}, is the following [Theorem 5.1]: \emph{if $X$ is a compact set than there exists a positive integer $N$ such that the function \begin{equation}\label{conv1} \phi_N(x)=(1+|x|^2)^Nf(x) \end{equation} is strongly convex on $X$}. Moreover, explicit estimates for the exponent $N$ were given in \cite{KS}. They depend on the diameter of $X$, the size of coefficients of the polynomial $f$ and on the minimum of $f$ on $X$. In fact a stronger version of \eqref{conv1} was given in \cite{KS}; \emph{there exists an integer $N$, which can be explicitly estimated, such that the polynomials \begin{equation* \phi_{N,\xi}=(1+|x-\xi|^2)^Nf(x),\quad \xi\in X, \end{equation*} are strongly convex on $X$}. The fact that $N$ can be chosen independent of $\xi$ was crucial for a construction of an algorithm which for a given polynomial $f$, positive in the convex compact semialgebraic set $X$, produces a sequence $a_\nu\in X$ starting from an arbitrary point $a_0\in X$, defined by induction: $a_\nu=\operatorname{argmin}_{X}\phi_{N,a_{\nu-1}}$, i.e., $a_{\nu}\in X$ is the unique point of $X$ at which $\phi_{N,a_{\nu-1}}$ has a global minimum on $X$. The sequence $a_\nu$ converges to a lower critical point of $f$ on $X$ (see \cite[Theorem 7.5]{KS}). In the case of non-compact closed convex set $X$ the results mentioned above require an additional assumption, that the \emph{leading form} $f_d$ of $f$, satisfy \begin{equation}\label{eqpositivityfd} f_d(x)>0\quad\hbox{for }x\in\mathbb R^n\setminus\{0\}. \end{equation} Under this assumption we have that: \emph{if a polynomial $f$ is positive on $X$ then for any $R>0$ there exists $N_0$ such that for each $\xi\in X$, $|\xi|\leq R$, $N>N_0$ the polynomial $\phi_{N,\xi}$ is strongly convex on $X$}. The assumption \eqref{eqpositivityfd} is necessary for local convexity of $\phi_{N,\xi}$ in a neighborhood of infinity, see \cite[Proposition 6.3]{KS}. However, this assumption is not sufficient to obtain convexity of the polynomial $\phi_{N,\xi}$ for some fixed $N>0$ independent of $\xi\in X$. For instance the polynomial $f(x)=1+x^2$, has this property, cf. \cite[Example 4.5]{KS}. The main goal of this paper is to study convexification of polynomials functions by exponential factors of the form $e^{N|x-\xi|^2}$ or by double exponential of the form $e^{e^{N|x-\xi|^{2}}}$. Surprisingly they play distinct roles. We set $$ \varphi _{N,\xi}(x):=e^{N|x-\xi|^2}f(x). $$ and prove the following (see Theorem \ref{uwypuklanie na zwartym} and Corollary \ref{uwypuklanie na zwartymcor2}): \emph{if a polynomial $f$ is positive on a compact and convex set $X\subset \mathbb R^n$, than there exists effectively computed number $N_0$ such that for any $N> N_0$ and $\xi\in \mathbb R^n$ the function $\varphi _{N,\xi}(x)$ is strongly convex on $X$.} If $X$ is not compact, we obtain the above assertions under the assumption \eqref{eqpositivityfd}, see Theorem \ref{convexonconvexsetnewvar}. In general the assumption \eqref{eqpositivityfd} can not be ommited as we show in Example \ref{exacounter}. Surprisingly convexification in the noncompact case without assumption \eqref{eqpositivityfd} is possible using double exponential factors. Namely in Theorems \ref{tedoubleexp} and \ref{convexonconvexsetnewxdoublevarsemi} we prove that: \emph{if $X\subset \mathbb R^n$ is a convex and closed semialgebraic set and $f$ is a polynomial positive on $X$, then for any $R>0$ there exists effectively computed number $N_0$ such that for any $N> N_0$ and any $\xi\in \mathbb R^n$, $|\xi|\leq R$, the function $$ \varPhi_{N,\xi}(x):=e^{e^{N|x-\xi|^{2}}}f(x) $$ is strongly convex on $X$.} In the case when $X$ is a convex and closed set, but non necessary semialgebraic, the result still holds (Theorems \ref{tedoubleexp2} and \ref{convexonconvexsetnewxdouble}) under an additional assumption \begin{equation}\label{eqestmonX} \inf\{f(x): x\in X\}\ge m>0. \end{equation} In the above theorems one can replace $\varPhi_{N,\xi}$ by the function $$ \Phi_{N,\xi}(x):=e^{Ne^{|x-\xi|^{2}}}f(x). $$ It turns out that convexification of polynomials using exponential function is somehow more natural and powerful than the convexification by the factors of the form $(1+|x-\xi|^2)^N$ done in \cite{KS}. In particular it applies also to the noncompact case and the explicit formulae for the exponent $N$ are nicer. We believe that the results mentioned above could be of interest, also to study o-minimal structures expanded by the exponent function. It fits particularly to the structure $\mathbb R_{\exp}$ semialgebraic sets expanded by the exponent function. The remarkable fact that $\mathbb R_{\exp}$ is indeed an o-minimal structure was established by A. Wilkie \cite{W}. It would be interesting to explain a different power of exponential and double exponential for convexification. The main difficulty when determining explicitely the number $N$ such that the function $\varphi_N$ is strongly convex on a convex compact set $X$, comes from an effective estimation of the number $m$ in \eqref{eqestmonX} and the number $R=\max\{|x|:x\in X\}$. Using results of G.~Jeronimo, D. Perrucci, E.~Tsigaridas \cite{JPT} we show in Theorem \ref{uwypuklaniecalkowitychna zwartym}) and Theorem \ref{uwypuklanie na zwartym} how it is feasible when $X$ is a compact semialgebraic set described by polynomial inequalities with integer coefficients and $f$ is also a polynomial with integer coefficients (see Theorem \ref{uwypuklanie na zwartym calk}). As an application to optimization we propose an algorithm which produces, starting from an arbitrary point $a_0\in X$, a sequence $a_\nu \in X$ which tends to a lower critical point of a polynomial $f$ restricted to $X$ or to infinity. We assume that $X\subset \mathbb R^n$ is a closed convex semialgebraic set and $f$ a polynomial which is bounded from below on $X$. Then by adding to $f$ an appropriate constant we may assume that $f\ge m> 0$ on $X$. If $X$ is unbounded we assume also condition \eqref{eqpositivityfd}. Hence by the above mentioned theorems we obtain strong convexity of $\varphi_\xi(x)=e^{|x-\xi|^2}f(x)$ for $\xi\in X$. Let us choose any $a_0\in X$ and set by induction: $a_\nu=\operatorname{argmin}_{X}\varphi_{N,a_{\nu-1}}$. Then we prove that the sequence $a_\nu $ tends to a lower critical point of a polynomial $f$ restricted to $X$ or to infinity. Note that computing $a_\nu$, that is minimizing $\varphi_{N,a_{\nu-1}}$ on $X$, is usually easier since the function is convex. This type of algorithm, based on convexification, is called sometimes {\it proximal}, see for instance \cite{Bolte}. Observe that computing the critical point of $\varphi_{N,a_{\nu-1}}$ involves only algebraic equations. The paper is organized as follows. In Section \ref{one variable} we prove that the function $\varphi _N$ in one variable is strongly convex on a closed interval $I\subset \mathbb R$, provided $f(x)>m$ for $x\in I$ and some $m>0$ and $N\in\mathbb{R}$ is sufficiently large. We also estimate from above the number $N$. In Section \ref{Sect3} we consider this problem in the several variables case on a compact convex set $X$ (see Theorem \ref{uwypuklanie na zwartym}). In Sections \ref{Sect44} and \ref{secdoubleexp} we consider the case when the set $X$ is not compact. \section{Convexifying polynomials}\label{one variable} \subsection{Convexifying $C^2$-functions in one variable} In this section we prove that if $f$ is a function of class $C^2$ positive on a closed interval $I\subset \mathbb R$ (not necessary compact), then for $N$ large enough the function $t\mapsto e^{Nt^2}f(t)$ is strongly convex on $I$. Let $f: \mathbb R\rightarrow \mathbb R$ be $C^2$ function. For any $N\in \mathbb R $ and $p,q\in\mathbb R$ we define the following function: $$ \varphi _{N,p,q}(t):=e^{N(t^2+pt+q)}f(t), \ \ t\in \mathbb R. $$ For positive numbers $m,D$ we put $$ \mathcal{N}(m,D):=\frac{D}{2m}+\frac{D^2}{2m^2}. $$ \begin{lemat}\label{lemat dla funkcji2} Let $f$ be a function of class $C^2$, which is positive on a closed interval $I\subset \mathbb R$. Let $m,D\in\mathbb{R}$ be such that \begin{equation}\label{minimum f} 0<m\leq \inf\{f(t): t \in I\}, \end{equation} and \begin{equation}\label{ograniczenie pochodnych} |f'(t)|\leq D, \quad |f''(t)|\leq D \quad for \quad t \in I. \end{equation} Assume that $p^2\leq 4q$ and \begin{equation}\label{nierownosc dla N2} N>\mathcal{N}(m,D) \end{equation} then $$\varphi _{N,p,q}'' (t)\geq -\frac{D^2}{m}-D+2Nm>0$$ for $t \in I$, thus $\varphi _{N,p,q}$ is strongly convex on $I$. \end{lemat} \begin{proof} By definition of $\varphi_{N,p,q}$ we have \begin{equation*}\label{eqwaznewprzykladzie} \varphi_{N,p,q}''(t)=e^{N(t^2+pt+q)}[N^2(2t+p)^2f(t)+2N(2t+p)f'(t)+2Nf(t)+f''(t)]. \end{equation*} Hence, from the assumptions we obtain \begin{equation}\label{oszacowanie} \varphi_{N,p,q}''(t)\geq e^{N(t^2+pt+q)}[N^2(2t+p)^2m-2N|2t+p|D+2Nm-D] \end{equation} for $t\in I$.Note that the function $$ \mathbb{R} \ni \lambda \mapsto N^2m\lambda^2-2N D\lambda+2Nm-D$$ attains its minimum, equal $-\frac{D^2}{m}-D+2Nm$, at the point $\lambda =\frac{D}{Nm}$. Thus for $ N>\mathcal{N}(m,D) $ we have $$N^2m\lambda^2-2ND|\lambda|+2Nm-D>0$$ for any $\lambda\in \mathbb R$. Therefore $$ \varphi_{N,p,q}''(t)\geq -\frac{D^2}{m}-D+2Nm>0\quad \hbox{for }t\in I, $$ which implies that $\varphi_{N,p,q}$ is strongly convex on $I$. \end{proof} From Lemma \ref{lemat dla funkcji2} we immediately obtain \begin{cor}\label{cor lemat dla funkcji2} Let $f$ be a function of class $C^2$ on a closed interval $I\subset \mathbb R$. Let $m,D\in\mathbb{R}$ be such that \begin{equation} m< \inf\{f(t): t \in I\}, \end{equation} and \begin{equation}\label{ograniczenie pochodnychdwa} |f'(t)|\leq D, \quad |f''(t)|\leq D, \quad \hbox{for} \quad t \in I. \end{equation} Than for any $\xi\in\mathbb R$ and any $N\geq 1$ the function $$ \psi_{N,\xi}(t)=e^{N(t-\xi)^2}[f(t)-m+D],\quad t\in I $$ is strongly convex on $I$. In particular the function $$ \varphi(t)=e^{(t-\xi)^2}[f(t)-m+D],\quad t\in I, $$ is strongly convex on $I$. \end{cor} \subsection{Convexifying polynomials in several variables}\label{Sect3} We will show that the function $\varphi _N$ in $n$ variables is strongly convex on a compact convex set $X\subset\mathbb{R}^n$, provided $f$ is a polynomial positive on $X$ and $N$ is suficiently large. Let $f\in \mathbb R [x]$ be a real polynomial in $x=(x_1, \ldots , x_n)$ of the form \begin{equation}\label{funkcja f} f=\sum_{j=0}^d\sum_{|\nu |=j}a_{\nu }x^\nu , \end{equation} where $a_\nu \in \mathbb R,$ $x^\nu =x^{\nu_{1}}_1\cdots x^{\nu_{n}}_n$ and $|\nu |=\nu _1 + \cdots + \nu _n$ for $\nu =(\nu _1, \cdots , \nu _n)\in \mathbb N^n$ (we assume that $0\in \mathbb N$). For $R>0$ we denote $$ D_n (f,R):=\max \bigg\{1,\sum_{j=1}^d\sum_{|\nu |=j}j|a_ \nu |R^{j-1};\sum_{j=1}^d\sum_{|\nu |=j}j(j-1)|a_ \nu |R^{j-2}\bigg\}. $$ \begin{tw}\label{uwypuklanie na zwartym} Let $f\in \mathbb R [x]$ be a polynomial which is positive on a compact and convex set $X\subset \mathbb R^n$. Let $R=\max\{|x|: x \in X\}$ and $$0<m\leq \min\{f(x): x\in X\}.$$ Than for any $\xi\in \mathbb R^n$, any $D \geq D_n (f,R)$ and any real $N> \mathcal{N}(m,D)$ the function $\varphi _{N,\xi}(x):=e^{N|x-\xi|^{2}}f(x)$ is strongly convex on $X$. \end{tw} \begin{proof} Let \begin{equation}\label{eqdefA} A:=\{(\alpha,\beta) \in \mathbb R^n \times \mathbb R^n: \langle \alpha, \beta \rangle=0,\; |\beta|=1\}, \end{equation} where $\langle \cdot,\cdot \rangle$ denotes the standard scalar product on $\mathbb R^n.$ Set $$\gamma _{\alpha, \beta}(t):=\beta t + \alpha, \ \ t\in \mathbb R.$$ Clearly, the family of all curves $\gamma _{\alpha, \beta}$, where $(\alpha,\beta)\in A$ describes all affine lines in $\mathbb R^n.$ Denote by $B\subset A$ the set of all $(\alpha,\beta)\in A$ such that the line parametrized by $\gamma _{\alpha, \beta}$ intersects the set X. Then $B$ is a compact set and $$B\subset \{(\alpha ,\beta )\in A: |\alpha |\leq R\}.$$ We will prove that for any $(\alpha ,\beta )\in B$ and $N > \mathcal{N}(m,D)$ the function $\varphi_{N,\xi} \circ \gamma _{ \alpha , \beta }$ is strongly convex on $$I_{\alpha ,\beta }:=\{t\in \mathbb R: \gamma_{\alpha ,\beta }(t) \in X\}.$$ Because $X$ is a compact and convex set, so $I_{\alpha ,\beta }$ is a closed interval or only one point. It is obvious that for $(\alpha;\beta) \in B$ the set $\{t\in \mathbb R: |\gamma _{\alpha ,\beta }(t)|\leq R\}$ is an interval, which contains the point 0 or it is equal to $\{0\}$. Denote this interval by $[-R_{\alpha ,\beta },R_{\alpha ,\beta }]$ (under convention $[0,0]=\{0\}$). Then $$I_{\alpha ,\beta }\subset [-R_{\alpha ,\beta },R_{\alpha ,\beta }]\subset [-R;R].$$ Let $f$ be of the form (\ref{funkcja f}). Than for $t\in I_{\alpha, \beta}$ we have $|\gamma _{\alpha ,\beta } (t)| \leq R $. Let us fix $(\alpha, \beta) \in B$. Then $$ |(f\circ \gamma _{\alpha ,\beta })'(t)|\leq \sum_{j=1}^d\sum_{|\nu |=j}j|a_ \nu |R^{j-1} $$ and $$ |(f\circ \gamma _{\alpha ,\beta })''(t)|\leq \sum_{j=1}^d\sum_{|\nu |\leq j}j(j-1)|a_ \nu |R^{j-2}. $$ Consequently, $$ |(f\circ \gamma _{\alpha ,\beta })'(t)|\leq D, \ \ \ |(f\circ \gamma _{\alpha ,\beta })''(t)|\leq D \ \ \ \hbox{for} \ \ \ t\in I_{\alpha, \beta}.$$ Take any $\xi\in\mathbb R^n$, then \begin{multline* |\gamma _{\alpha ,\beta }(t)-\xi|^2=\langle \beta t+\alpha-\xi ,\beta t+\alpha-\xi \rangle \\ = t^2-2\langle \beta,\xi\rangle t+|\alpha| ^2-2\langle \alpha,\xi\rangle+|\xi|^2 \end{multline* then for $p=-2\langle \beta,\xi\rangle$ and $q=|\alpha| ^2-2\langle \alpha,\xi\rangle+|\xi|^2$, we have $p^2\leq 4q$ and $$ \varphi _N \circ \gamma _{\alpha, \beta}(t)=e^{N(t^2+pt+q)}f(\gamma _{\alpha, \beta}(t)). $$ So, by Lemma {\ref{lemat dla funkcji2}} we get that $\left(\varphi _N \circ \gamma _{\alpha, \beta}\right)''(t)\geq -\frac{D^2}{m}-D+2Nm>0$ for $t\in I_{\alpha,\beta}$ and $\varphi_{N,\xi}$ is strongly convex on $X$, provided $N>\mathcal{N}(m,D)$. \end{proof} From Theorem \ref{uwypuklanie na zwartym} we obtain the following corollary. \begin{cor}\label{uwypuklanie na zwartymcor2} Let $f\in \mathbb R [x]$ and let $X\subset \mathbb R^n$ be a compact and convex set. Let $R=\max\{|x|: x \in X\}$ and let $m\in \mathbb{R}$ be a constant such that $$ m\leq \min\{f(x): x\in X\}. $$ Than for any $D> D_n (f,R)$ and any $\xi\in\mathbb R^n$, the function $$ \varphi_\xi(x):=e^{|x-\xi|^2}[f(x)-m+D], \quad x\in \mathbb{R}^n $$ is strongly convex on $X$. \end{cor} By a similar argument as in the proof of Theorem \ref{uwypuklanie na zwartym}, we obtain the following fact. \begin{rem}\label{uwypuklanie na zwartymrem} Let $f:\mathbb{R}^n\to \mathbb{R}$ be a function of class $C^2$ and let $X\subset \mathbb R^n$ be a compact and convex set. Assume that $m,D\in \mathbb{R}$ are numbers satisfying $$ m < \min\{f(x): x\in X\} $$ and the first and second directional derivatives of $f$ in directions of vectors of length $1$, are bounded by $D$ on $X$. Then the function $$ \varphi_\xi(x)=e^{|x-\xi|^2}[f(x)-m+D], \quad x\in \mathbb{R}^n, \quad \xi\in \mathbb{R}^n $$ is strongly convex on $X$. \end{rem} \subsection{Convexifying polynomials with integer coefficients}\label{SectINTEGER} For actual applications of Theorem \ref{uwypuklanie na zwartym} it is important to compute the number $\mathcal{N}(m,D)$ for a given convex semialgebraic set $X $ and a polynomial $f$ which is positive on $X$. Hence the main difficulty is to compute (or rather estimate) $m=\min\{f(x):x\in X\}$ and $R=\max\{|x|:x\in X\}$. This actually possible if we suppose that $f$ has integer coefficients and $X$ is described by equations and inequalities with integer coefficients. More precisely, let $X\subset \mathbb R^n$, $n\ge 2$, be a compact semialgebraic set of the form \begin{multline}\label{formXc} X=\{x\in\mathbb{R}^n:g_1(x)=0,\ldots,g_l(x)=0,g_{l+1}(x)\geq 0, \ldots,\\ g_k(x)\geq 0\}, \end{multline} where $g_1,\ldots,g_k\in \mathbb{Z}[x]$. Under the above notations G. Jeronimo, D.~Perrucci, E.~Tsigaridas in \cite{JPT} proved that \begin{tw}\label{uwypuklaniecalkowitychna zwartym} Let $f, g_1,\ldots,g_k\in \mathbb{Z}[x]$ be polynomials with degrees bound by an even integer $d$ and coefficients of absolute values at most $H$, and let $\tilde H=\max\{H,2n+2k\}$. If $f(x)>0$ for $x\in X$ and $X$ of the form \eqref{formXc} is compact, then $$ \min\{f(x):x\in X\}\geq \left(2^{4-\frac{n}{2}}\tilde H d^n\right)^{-n2^nd^n}. $$ \end{tw} For a positive real number $H$ and positive integers $d,n,k$ we put $$ \frak{b}(n,d,H,k)=\left(2^{4-\frac{n}{2}}\max\{H,2n+2k\}d^n\right)^{-n2^nd^n} $$ From Theorems \ref{uwypuklaniecalkowitychna zwartym} and \ref{uwypuklanie na zwartym} we immediately obtain \begin{tw}\label{uwypuklanie na zwartym calk} Let $X\subset \mathbb R^n$ be a compact and convex semialgebraic set of the form \eqref{formXc} and let $f, g_1,\ldots,g_k\in \mathbb{Z}[x]$ be polynomials with degrees bound by an even integer $d $ and coefficients of absolute values at most $H$. Set $$ R=\sqrt{\big[\frak{b}(n+1,\max\{d,4\},H,k+2)\big]^{-1}-1},\quad m=\frak{b}(n.d,H,k). $$ Then \begin{equation}\label{boundR} \max\{|x|:x\in X\}\leq R. \end{equation} Moreover, if $f(x)>0$ for $x\in X$, then for any $D \geq D_n \left(f,R\right)$, $N> \mathcal{N}\left(m,D\right)$ and for any $\xi\in\mathbb R^n$ the function $$\varphi _{N,\xi}(x):=e^{N|x-\xi|^{2}}f(x)$$ is strongly convex on $X$. \end{tw} \begin{proof} By Theorem \ref{uwypuklaniecalkowitychna zwartym} we have $0<m\leq \min\{f(x):x\in X\}$. Let $$ Y=\{(x,y)\in\mathbb R^n\times\mathbb R:x\in X,\,(1+|x|^2)y^2-1=0,\,y\ge 0\}, $$ and let $h(x,y)=y^2$. Then $Y\subset \mathbb R^{n+1}$ is a compact semialgebraic set defined by $k+2$ polynomial equations and inequalities of degrees bounded by $\max\{d,4\}$. Moreover, the absolute values of coefficients of those polynomials and $h$ are bounded by $H$. Then, by Theorem \ref{uwypuklanie na zwartym calk}, $$ \min\{h(x,y):(x,y)\in Y\}\geq \frak{b}(n+1,\max\{d,4\},H,k+2) $$ and consequently we obtain \eqref{boundR}. Summing up, Theorem \ref{uwypuklanie na zwartym} gives the assertion. \end{proof} \section{Convexifying polynomials on non-compact sets}\label{Sect44} In this section we will show that the function $\varphi _N(x)=e^{N|x|^2}f(x)$ in $n$ variables is strongly convex on a closed and convex set $X\subset\mathbb{R}^n$ (not necessary compact), provided the polynomial $f$ takes values larger than a certain number $m>0$, the leading form of a polynomial $f$ has only positive values and $N$ is suficiently large. \subsection Convexifying polynomials in one variable} For a polynomial $f\in\mathbb R[t]$ of the form $f(t)=a_0t^d+a_1t^{d-1}+\cdots+a_d$, $a_0,\ldots,a_d\in\mathbb R$, $a_0\ne 0$, we put $$ K(f):=2\max_{1\leq i\leq d}\left|\frac{a_i}{a_0}\right|^{1/i}. $$ \begin{lemat}\label{K(g)<K(f)var} Let $f\in \mathbb{R}[t]$ be a polynomial of degree $d>0$ which is positive on a closed interval $I\subset \mathbb R$ (not necessary compact). Let $m\in\mathbb R$ be a positive number such that $$ \inf\{f(t):t\in I\}\geq m. $$ Let $g_N\in\mathbb R[t]$ be a polynomial of the form $$ g_N=2Nf^2-(f')^2+ff'', $$ and let $\Theta_N\in\mathbb{R}[t,\xi]$ be a polynomial of the form \begin{equation}\label{thetaNformvar} \Theta _N(t,\xi):=4N^2(t-\xi)^2f(t)+4N(t-\xi)f'(t)+2Nf(t)+f''(t) \end{equation} for $N\in \mathbb{R}$ and $N \geq 1$. Then for $N\geq \mathcal{N}(m,D)$, where $D\geq D_1(f,R)$ and $R\geq \max\{K(f),K(g_1)\}$, we have $$ \Theta_N(t,\xi)>0\quad\hbox{for }(t,\xi)\in I\times \mathbb R. $$ \end{lemat} \begin{proof} Consider the following quadratic function in $x$ $$ 4N^2x^2f(t)+4Nxf'(t)+2Nf(t)+f''(t). $$ Then its discrirminant is of the form $\Delta(t)=-16N^2g_N(t)$. Take $R\geq \max\{K(f),K(g_1)\}$, $D\geq D_1(f,R)$ and $N>\mathcal{N}(m,D)$. Then we have $$ g_N(t)\geq 2Nf^2(t)-D^2-f(t)D\geq f^2(t)\left(2N-\tfrac{D^2}{m^2}-\tfrac{D}{m}\right)>0 $$ for $t\in I$, $|t|\leq R$. On the other hand $g_N(t)\geq g_1(t)>0$ for $t\in I$, $|t|\geq R$. So $\Delta(t)<0$ for $t\in I$ and we deduce the assertion. \end{proof} \begin{tw}\label{thm3unboundedvar} Let $f \in \mathbb R[t] $ be a polynomial of degree $d>0$ and let $I\subset \mathbb{R}$ be a closed interval (not necessary compact). Assume that there exists $m\in \mathbb R$ such that $$ 0<m\leq \inf\{f(t): t\in I\}. $$ Let $R>\max\{K(f),K(2f^2-(f')^2+ff'')$ and $D\geq D_1(f,R)$. Then for any $N\in \mathbb{R}$, $N\geq \mathcal{N}(m,D)$, and any $\xi\in\mathbb R$ the function $$ \varphi_{N,\xi}(t)=e^{N(t-\xi)^2}f(t) $$ is strongly convex on $I$. \end{tw} \begin{proof} It suffices to observe that $\varphi''_{N,\xi}(t)=e^{N|t-\xi|^2}\Theta_N(t,\xi)$ and apply Lemma \ref{K(g)<K(f)var}. \end{proof} \subsection Convexifying polynomials in several variables} \begin{tw}\label{convexonconvexsetnewvar} Let $X\subset \mathbb R^n$ be a convex closed set. Assume that $f$ is a polynomial of degree $d>0$ which is positive on~$X$, \begin{equation}\label{assumptioncompactvar} f_d^{-1}(0) = \{0\} \end{equation} and there exists $m>0$ such that \begin{equation}\label{estfcompact1var} \inf \{f(x):x\in X\} \ge m. \end{equation} Then there exists $N_0\in\mathbb N$ such that for any integer $N\geq N_0$ and any $\xi\in\mathbb R^n$ the function $\varphi_{N,\xi}(x)=e^{N|x-\xi|^2}f(x)$ is strongly convex on $X$. \end{tw} \begin{proof} Take any line of the form $\gamma_{\alpha,\beta}(t)=\beta t +\alpha$, where $\alpha,\beta\in\mathbb R^n$, $|\beta|=1$ and $\langle \alpha,\beta\rangle=0$. Then $$ (\varphi_{N,\xi}\circ \gamma_{\alpha,\beta})(t)=e^{N(t^2+|\alpha|^2-2\langle \beta,\xi\rangle t-2\langle\alpha,\xi\rangle)}f(\gamma_{\alpha,\beta}(t)). $$ Then \begin{equation*}\label{eqbis} \begin{split} (\varphi_{N,\xi}\circ \gamma_{\alpha,\beta})''(t)=&e^{N(t^2+|\alpha|^2-2\langle \beta,\xi\rangle t-2\langle\alpha,\xi\rangle)}[4N^2(f\circ\gamma_{\alpha,\beta})(t)y^2\\ &+4N(f\circ\gamma_{\alpha,\beta})'(t)y+ 2N(f\circ\gamma_{\alpha,\beta})(t)\\ &+(f\circ\gamma_{\alpha,\beta})''(t)], \end{split} \end{equation*} where $y=t+\langle\beta,\xi\rangle$. Consider the function in the square bracket as a quadratic function in $y$. Then its discriminant is of the form $$ \Delta=-16N^2[2N(f\circ\gamma_{\alpha,\beta})^2(t)+(f\circ\gamma_{\alpha,\beta})(t)(f\circ\gamma_{\alpha,\beta})''(t)-((f\circ\gamma_{\alpha,\beta})'(t))^2]. $$ Note that $(f\circ\gamma_{\alpha,\beta})'(t)$ and $(f\circ\gamma_{\alpha,\beta})''(t)$ are the first and the second directional derivatives of $f$ at $\gamma_{\alpha,\beta}(t)$ in the direction $\beta$ and $|\beta|=1$. Observe that there exists $N_0$ such that for any $N\geq N_0$ we have $\Delta<0$. Indead, it suffices to prove that for any $x\in X$ and any $\beta \in\mathbb R^n$, $|\beta|=1$ we have \begin{equation}\label{eqvar1} 2Nf(x)^2+f(x)\partial^2_\beta f(x)-(\partial_\beta f(x))^2>0. \end{equation} If $f_d(x)<0$ for $x\in\mathbb R^n\setminus \{0\}$ then the set $X$ is compact and the inequality follows from the assumption that $f(x)\geq m$ for $x\in X$. Indead, let $D\geq \max \{|\partial _\beta f(x)|,|\partial^2_\beta f(x)|\}$ for $x\in X$, $|\beta|=1$. Since $X$ is compact, then $$ 2Nf(x)^2+f(x)\partial^2_\beta f(x)-(\partial_\beta f(x))^2\ge 2Nm^2-mD-D^2>0 $$ for $N>\mathcal{N}(m,D)$. This gives \eqref{eqvar1}. Consider the case when $f_d(x)>0$ for $x\in\mathbb R^n\setminus\{0\}$, and let $$ f_{d*}=\inf \{f_d(x):x\in S_n\}, $$ where $S_n$ is the unit sphere in $\mathbb{R}^n$, i.e., $S_n=\{x\in\mathbb{R}^n:|x|=1\}$. Let $f$ be a polynomial of the form \eqref{funkcja f}. We set $$ \|f\|: \sum_{|\nu|\leq d}|a_\nu| $$ Then $\|f\|\geq \|f_d\|\geq f_{d*}$. If $f_{d*}>0$ then we set $$ \mathbb{K}(f):= \frac{2\|f\|}{f_{d*}} $$ and $$ m(f):=f_{d*}-\sum_{j=0}^{d-1}\mathbb{K}(f)^{j-d}\sum_{|\nu|=j}|a_\nu|. $$ In the further part of the proof we will need the following lemma. \begin{lemat}\label{lemmamestvar} If $d=\deg f>0$ and $f_{d*}>0$, then $m(f)>0$ and $f(x)\geq m(f)|x|^d$ for any $x\in \mathbb{R}^n$ such that $|x|\geq \mathbb{K}(f)$. \end{lemat} \begin{proof} Put $$ h(t):=f_{d*}t^d-\sum_{j=0}^{d-1}\left(\sum_{|\nu|=j}|a_\nu|\right)t^{j}. $$ Since $\frac{\|f\|}{f_{d*}}\geq 1$, then $$ K(h)=2\max_{1\leq i\leq d}\left|\frac{\sum_{|\nu|=d-i}|a_\nu|}{f_{d*}}\right|^{{1}/{i}}< 2\max_{1\leq i\leq d}\left|\frac{\|f\|}{f_{d*}}\right|^{{1}/{i}}=\mathbb{K}(f), $$ and since $h'(t)>0$ for $t >K(h)$, then $h(|x|)\geq h(\mathbb{K}(f))>0$ for $|x|\geq \mathbb{K}(f)$. Moreover, $m(f)\mathbb{K}(f)^d=h(\mathbb{K}(f))$, so $m(f)>0$. On the other hand $$ m(f)|x|^d\leq \left(f_{d*}-\sum_{j=0}^{d-1}|x|^{j-d}\sum_{|\nu|=j}|a_\nu|\right)|x|^d=h(|x|)\leq f(x) $$ for $|x|\geq \mathbb{K}(f)$. This gives the assertion of Lemma \ref{lemmamestvar}. \end{proof} Take $R\geq \mathbb{K}(f)$, and $D\geq D_n(f,R)$ then for $N\geq \mathcal{N}(m,D)$ we have $$ 2Nf(x)^2+f(x)\partial^2_\beta f(x)-(\partial_\beta f(x))^2\geq 2Nm^2-mD-D^2>0 $$ for $|x|\leq R$. For $|x|\geq R$, we have \begin{multline*} 2Nf(x)^2+f(x)\partial^2_\beta f(x)-(\partial_\beta f(x))^2\\ \geq 2Nm^2(f)|x|^{2d}-m(f)|x|^dD_n(f,|x|)-D_n^2(f,|x|). \end{multline*} Since for $|x|\geq 1$, $$ D_n(f,|x|)\leq D_n(f,1)|x|^{d-1}, $$ then for $|x|\geq R$ and $|\beta|=1$ we have \begin{multline*} 2Nf(x)^2+f(x)\partial^2_\beta f(x)-(\partial_\beta f(x))^2\\ \geq 2Nm^2(f)|x|^{2d}-m(f)D_n(f,1)|x|^{2d-1}-D_n^2(f,1)|x|^{2d-2}\\ \geq |x|^{2d}[2Nm^2(f)-m(f)D_n(f,1)-D_n^2(f,1)]>0. \end{multline*} for $N>\mathcal{N}(m(f),D_n(f,1))$. This gives \eqref{eqvar1}. Moreover, there exists $\epsilon>0$ such that $$ 2Nf(x)^2+f(x)\partial^2_\beta f(x)-(\partial_\beta f(x))^2>\epsilon. $$ for any $x\in X$ and $|\beta|=1$. From \eqref{eqvar1} and the above there follows that $\Delta<-16N^2\epsilon$ for any $\alpha,\beta$ and $t$ such that $\gamma_{\alpha,\beta}(t)\in X$. Since $f(x)\geq m$ for $x\in X$, then $(\varphi_{N,\xi}\circ \gamma_{\alpha,\beta})''(t)\geq \tfrac{\epsilon}{m}$ if $\gamma_{\alpha,\beta}(t)\in X$. This gives the assertion. \end{proof} By analogous argument as for Theorem \ref{convexonconvexsetnewvar} and under notations of the proof we obtain the following corollary. \begin{cor}\label{uwypuklanie na wypuklymcor2} Let $f\in \mathbb R [x]$ be a polynomial of degree $d$ and let $X\subset \mathbb R^n$ be a convex and closed set. Under assumptions of Theorem \ref{convexonconvexsetnewvar} and notations of the proof, for any $R\geq \mathbb{K}(f)$, $D\geq D_n(f,R)$ and $N>\max\{\mathcal{N}(m,D),\mathcal{N}(m(f),D_n(f,1))\}$, and any $\xi\in\mathbb R^n$ the function $\varphi_{N,\xi}(x)=e^{N|x-\xi|^2}f(x)$ is strongly convex on $X$. In particular the function $$ \varphi_\xi(x):=e^{|x-\xi|^2}[f(x)-m+D $$ is strongly convex on $X$. \end{cor} The assumption \eqref{assumptioncompactvar} that $f_d(x)\ne 0$ for $x\ne 0$, in Theorem \ref{convexonconvexsetnewvar}, can not be omited as the following example shows. \begin{exa}\label{exacounter} Let $f\in \mathbb R[x,y,z]$ be a polynomial of the form $$ f(x,y,z)=(y^2+z^2+1)\left[(x-1)^2(x+1)^2 +(yz+1)^2+y^2\right]. $$ Since $(y^2+z^2+1)\left[(yz+1)^2+y^2\right]\ge \frac{1}{2}$ for $(y,z)\in\mathbb R^2$ then we easily see that $$ f(x,y,z)\ge \tfrac{1}{2}\quad \hbox{for }(x,y,z)\in\mathbb R^3. $$ Note that $\deg f=6$, and the leading form $f_6(x,y,z)=(y^2+z^2)(x^4+y^2z^2)$ has nontrivial zeroes. Now take any $N\in \mathbb R$ and $\varphi_N(x,y,z)=e^{N(x^2+y^2+z^2)}f(x,y,z)$. Then for $\xi\ne 0$ we have $$ \varphi_N(0,\xi^{-1},-\xi)=e^{N(\xi^{-2}+\xi^2)}(\xi^{-2}+\xi^2+1)(1+\xi^{-2}) $$ and $$ \varphi_N(-1,\xi^{-1},-\xi)=\varphi_N(1,\xi^{-1},-\xi)=e^{N(\xi^{-2}+\xi^2)}(\xi^{-2}+\xi^2+1)e^N\xi^{-2}. $$ Hence for sufficiently large $\xi$, $$ \varphi_N(-1,\xi^{-1},-\xi)=\varphi_N(1,\xi^{-1},-\xi)<\varphi_N(0,\xi^{-1},-\xi), $$ therefore $\varphi_N$ can not be a convex function. \end{exa} The assumption \eqref{assumptioncompactvar} in Theorem \ref{convexonconvexsetnewvar}, cannot be replaced by a condition $\lim_{|x|\to\infty} f(x)=\infty$. Indead, consider a modification of the previous example of the form $$ f_k(x,y,z)=(y^2+z^2+1)^k\left[(x-1)^2(x+1)^2 +(yz+1)^2+y^2\right], $$ where $k\geq 2$. Then $\lim_{|(x,y,z)|\to\infty} f(x,y,z)=\infty$ and the function $\varphi_N(x,y,z)=e^{N(x^2+y^2+z^2)}f_k(x,y,z)$ is not convex for any $N\in\mathbb R$ by the previous argument. It turns out that the use of a double exponential function leads to a convexity of an appropriate function on $X$. We show it in the next section. \section{Double exponential convexifying polynomials}\label{secdoubleexp} In this section, without the assumption that the leading form of a polynomial $f\in\mathbb R[x]$ in $n$ variables has only positive values, we will show that the function $\varPhi _N(x)=e^{e^{N|x|^2}}f(x)$ is strongly convex on a closed and convex semialgebraic set $X\subset\mathbb{R}^n$ (not necessary compact), provided the polynomial $f$ takes positive values on $X$ and $N$ is suficiently large. \begin{tw}\label{tedoubleexp} Let $X\subset \mathbb R^n$ be a closed and convex semialgebraic set, and let $f\in\mathbb R[x]$ be a polynomial which has only positive values on $X$. Then there exists $N_0\in\mathbb R$ such that for any $N\geq N_0$ the function $\varPhi _N(x)=e^{e^{N|x|^2}}f(x)$ is strongly convex on $X$. \end{tw} \begin{proof} Let $f$ be of the form \eqref{funkcja f}, $d=\deg f$. Then $f(x)$, , the first and second directional derivatives of $f$ in directions of vectors of length $1$ at $x\in X$, are bounded by $\tilde D(1+|x|^d)$, where $$ \tilde D:=|a_0|+\sum_{|\nu|=1}|a_\nu|+\sum_{j=2}^d\sum_{|\nu |=j}j(j-1)|a_ \nu |. $$ Take an affine line in $\mathbb R^n$ of the for $$ \gamma _{\alpha, \beta}(t):=\beta t + \alpha, \ \ t\in \mathbb{R}, $$ where $(\alpha,\beta)\in A$ and the set $A$ is defined in \eqref{eqdefA}. Then $|\beta|=1$, $\langle \alpha,\beta\rangle=0$, and $|\gamma_{\alpha,\beta}(t)|^2=t^2+|\alpha|^2$. Let write the second derivative of $\varPhi_N\circ \gamma_{\alpha,\beta}$ in the form $$ (\varPhi_N\circ\gamma_{\alpha,\beta})''(t)=e^{e^{N(t^2+|\alpha|^2)}}(a(t) t^2 +b(t)t +c(t)), $$ where \begin{equation*} \begin{split} a(t) & = 4N^2(e^{N(t^2+|\alpha|^2)}+e^{2N(t^2+|\alpha|^2)})f\circ\gamma_{\alpha,\beta}(t), \\ b(t) & = 4Ne^{N(t^2+|\alpha|^2)}(f\circ\gamma_{\alpha,\beta})'(t), \\ c(t) & = 2Ne^{N(t^2+|\alpha|^2)}f\circ\gamma_{\alpha,\beta}(t) +(f\circ\gamma_{\alpha,\beta})''(t) \end{split} \end{equation*} The discriminant of the polynomial $ P_t(\lambda) = a(t)\lambda^2 + b(t)\lambda+c(t) $ is of the form \begin{equation*} \begin{split} \Delta=16N^2e^{2N(t^2+|\alpha|^2)} &\left[\left((f\circ\gamma_{\alpha,\beta})'(t)\right)^2\right. \\ &-f\circ\gamma_{\alpha,\beta}(t)(f\circ\gamma_{\alpha,\beta})''(t)\left(1- e^{-N(t^2+|\alpha|^2)}\right)\\ &\left.-2N(f\circ\gamma_{\alpha,\beta})^2(t)\left(1+ e^{N(t^2+|\alpha|^2)}\right)\right]. \end{split} \end{equation*} So, by the choice of the number $\tilde D$, we have \begin{equation*} \begin{split} \Delta\leq 32N^2e^{2N(t^2+|\alpha|^2)}&\left[ \tilde D^2\left(1+|\gamma_{\alpha,\beta}(t)|^d\right)^2\right.\\ &\;\; \left.-N(f\circ\gamma_{\alpha,\beta})^2(t)\left(1+ e^{N|\gamma_{\alpha,\beta}(t)|^2}\right) \right]. \end{split} \end{equation*} Since the set $X$ is semialgebraic and $f^{-1}(0)\cap X=\emptyset$, then by H\"ormander-{\L}ojasiewicz inequality, see eg. \cite[Corollary 2.4]{KS0}, there exist $C,K,\mathcal{L}>0$, where $K,\mathcal{L}\in \mathbb{Z}$, $K\geq d$, depend on $d$ and the {\it complexity} of $X$, (i.e., degrees and the number of polynomials describing $X$) such that $$ f(x)\geq C\left({1+|x|^K}\right)^{-\mathcal{L}}\quad \hbox{for }x\in X. $$ Moreover, the numbers $K,\mathcal{L}$ are effectively computable. By the above, \begin{equation*} \begin{split} \Delta\leq 32N^2e^{2N(t^2+|\alpha|^2)}\left({1+|\gamma_{\alpha,\beta}(t)|^K}\right)^{-\mathcal{L}}&\left[ \tilde D^2\left(2+|\gamma_{\alpha,\beta}(t)|^K\right)^{\mathcal{L}+2}\right. \\ &\left.\;\; -NC^ \left(1+ e^{N|\gamma_{\alpha,\beta}(t)|^2}\right)\right]. \end{split} \end{equation*} If $N$ is large enough, then for any $x\in\mathbb R^n$ we have $$\tilde D^2(2+|x|^K)^{\mathcal{L}+2}<NC^2(1+e^{N|x|^2}).$$ Therefore $\Delta<0$, so $P_t(\lambda) >0$ for any $\lambda \in \mathbb R$. Consequently $$ (\varPhi_N\circ\gamma_{\alpha,\beta})''(t)=e^{e^{N(t^2+|\alpha|^2)}}P_t(t) >0, $$ for $t\in \mathbb R$. Note that $\lim_{|t|\to \infty}(\varPhi_N\circ \gamma_{\alpha,\beta})''(t)=+\infty$, hence there exists $\mu>0$ such that $(\varPhi_N\circ\gamma_{\alpha,\beta})''(t)\geq \mu $ for $t\in\mathbb R$. Moreover, the number $\mu$ can be chosen independet of $\gamma_{\alpha,\beta}$. This gives the assertion. \end{proof} \begin{rem}\label{remeffectcomputed} The number $N_0$ in Theorem \ref{tedoubleexp} can be effectively computed, provided we can estimate the constant $C$. More precisely, under notations in the proof, if $k > (\mathcal{L}+2)K$, then for $|x|\geq 1$ we have $$ NC^2\left(1+e^{N|x|^2}\right)\geq NC^2+\sum_{j=0}^k\frac{C^2N^{j+1}}{j!}|x|^{2j}>\tilde D^2\left(2+|x|^K\right)^{\mathcal{L}+2} $$ for $$ N>k!\max_{i=0,\ldots,\mathcal{L}+2}\left({\tilde D^2C^{-2}2^{\mathcal{L}+2-i}\binom{\mathcal{L}+2}{i}}\right) $$ If additionally $N\geq \tilde D^2 C^{-2}3^{\mathcal{L}+2}$, then the above inequality holds for any $x\in\mathbb R^n$. \end{rem} \begin{rem}\label{remdoubleexp} We cannot omit the assumption in Theorem \ref{tedoubleexp} that the set $X$ is semialgebraic. For instance if $f(x,y)=-y^2+y$ and $X=\{(x,y)\in\mathbb R^2:e^{-e^x}\leq y\leq \tfrac{1}{2},\; x\geq 0\}$, then $f(x,y)>0$ on $X$, but the function $\varPhi_N(x,y)$ is not convex on $X$ for any $N\in \mathbb R$. \end{rem} Assuming that $f(x)\geq m$ on $X$, for some $m>0$, we can omit the assumption in Theorem \ref{tedoubleexp} on semialgebraicity of $X$. More precisely, by a similar argument as in the proof of Theorem \ref{tedoubleexp} we obtain \begin{tw}\label{tedoubleexp2} Let $X\subset \mathbb R^n$ be a closed and convex set, and let $f\in\mathbb R[x]$ be a polynomial such that $f(x)\geq m$ for $x\in X$ and some $m>0$. Then there exists $N_0\in\mathbb R$ such that for any $N\geq N_0$ the function $\varPhi _N(x)=e^{e^{N|x|^2}}f(x)$ is strongly convex on $X$. \end{tw} \begin{rem}\label{tedoubleexp2bis} By a similar argument as for the proof of Theorem \ref{tedoubleexp2} we obtain that the assertion of this Theorem occurs not only for the function $\varPhi_{N}$ but also for the function $\Phi_{N}(x)=e^{Ne^{|x|^2}}f(x)$. More precisely, we have: Let $X\subset \mathbb R^n$ be a closed and convex set, and let $f\in\mathbb R[x]$ be a polynomial such that $f(x)\geq m$ for $x\in X$ and some $m>0$. Then there exists $N_0\in\mathbb R$ such that for any $N\geq N_0$ the function $\Phi _{N}$ is strongly convex on $X$. \end{rem} A similar argument as for Theorem \ref{tedoubleexp} gives the following theorems \begin{tw}\label{convexonconvexsetnewxdoublevarsemi} Let $X\subset \mathbb R^n$ be a convex closed semialgebraic set and let $r>0$. If $f$ is a polynomial such that \begin{equation}\label{estfcompact1xdoublevarsemi} f(x)>0\quad\hbox{for }x\in X, \end{equation} then there exists $N_0\in\mathbb N$ such that for any integer $N\geq N_0$ and any $\xi\in\mathbb R^n$, $|\xi|\leq r$, the function $\varPhi_{N,\xi}(x)=e^{e^{N|x-\xi|^2}}f(x)$ is strongly convex on $X$. Moreover, there exists $\alpha\in\mathbb R$ such that the function $$ \varPhi_\xi(x)=e^{e^{|x-\xi|^2}}[f(x)+\alpha], \quad x\in \mathbb{R}^n $$ is strongly convex on $X$, provided $ \xi\in \mathbb{R}^n$, $|\xi|\leq r$. \end{tw} \begin{tw}\label{convexonconvexsetnewxdouble} Let $X\subset \mathbb R^n$ be a convex closed set and let $r>0$. Assume that $f$ is a polynomial of degre $d>0$ such that there exists $m\in\mathbb R$ such that \begin{equation}\label{estfcompact1xdouble} 0<m < \inf \{f(x):x\in X\}. \end{equation} Then there exists $N_0\in\mathbb N$ such that for any integer $N\geq N_0$ and any $\xi\in\mathbb R^n$, $|\xi|\leq r$ the function $\varPhi_{N,\xi}(x)=e^{e^{N|x-\xi|^2}}f(x)$ is strongly convex on $X$. Moreover, there exists $\alpha\in\mathbb R$ such that the function $$ \varPhi_\xi(x)=e^{e^{|x-\xi|^2}}[f(x)+\alpha], \quad x\in \mathbb{R}^n $$ is strongly convex on $X$, provided $ \xi\in \mathbb{R}^n$, $|\xi|\leq r$. \end{tw} It is impossible to obtain $N$ in the above theorem, such that the function $\varPhi_{N,\xi}$ is convex for any $\xi\in X$ as the following example shows. \begin{exa}\label{exacounter2} Let $f\in \mathbb R[x,y,z]$ be a polynomial of the form $$ f(x,y,z)=\left[(yz+1)^2+y^2\right]\left[\left(xz^2-1\right)^2\left(xz^2+1\right)^2 +y^2+z^2+1\right]. $$ Analogously as in Example \ref{exacounter} we see that $$ f(x,y,z)\ge \tfrac{1}{2}\quad \hbox{for }(x,y,z)\in\mathbb R^3, $$ and the leading form $f_{16}(x,y,z)=y^2z^{10}x^4$ has nontrivial zeroes. Now take any $N\in \mathbb R$ and $\varPhi_N(x,y,z)=e^{e^{N(x^2+y^2+z^2)}}f(x,y,z)$. Then for $\xi=(0,t^{-1},-t)$, $t>0$ we have $$ \frac{\partial^2 \varPhi_{N,\xi}}{\partial x^2}(\xi)=e\left(2Nf(\xi)+\frac{\partial^2 f}{\partial x^2}(\xi)\right). $$ Since $$ f(\xi)=2t^{-2}+t^{-4}+1 $$ and $$ \frac{\partial^2 f}{\partial x^2}(\xi)=-4t^2, $$ then we easily see that $\frac{\partial^2 \varPhi_{N,\xi}}{\partial x^2}(\xi)<0$ for sufficiently large $t$. So, $\varPhi_{N,\xi}$ can not be a convex function. \end{exa} \begin{rem}\label{remtripleexpmotnotimproves} It is worth noting that the use of triple exponential convexifying $\phi_N(x)=e^{e^{e^{N|x|^2}}}f(x)$ of a polynomial $f$ does not improve convexity of the function $\phi_{N,\xi}(x)=e^{e^{e^{N|x-\xi|^2}}}f(x)$ regardless of $\xi\in X$. \end{rem} \section{Algorithm for searching lower critical points}\label{Algorithm0} \subsection{Searching lower critical points in a compact set} In this part we give an algorithm which produces, starting from an arbitrary point, a sequence of points converging to a lower critical point of a polynomial on a convex compact semialgebraic set. A similar algorithm was proposed in \cite{KS}. Let $X\subset \mathbb R^n$ be a closed set and let $f$ be a function of class $C^1$ in a neighborhood $U\subset \mathbb R^n$ of $X$. We denote the set of lower critical points of the function $f$ on the set $X$ by $\Sigma _X f.$ It is obvious that the set of ordinary critical points $\Sigma f$ of the function $f$ is contained in the set $\Sigma _X f.$ Our algorithm for approximation of lower critical points of $f$ is based on the iteration of computation of the smallest value of the strongly convex function $\varphi _{\xi }$ on the convex and compact set $X$. More precisely, let $$ R\ge \max\{|x|:x\in X\} $$ Take any polynomial $f\in\mathbb{R}[x]$ of the form \eqref{funkcja f}. Let $$ m=-\sum_{j=0}^dR^j\sum_{|\nu |=j}|a_{\nu }|, $$ and let $$ D>D_n (f,2R). $$ Then we have $$ f(x)-m+D\geq D \quad\hbox{for }x\in X, $$ and from Corollary \ref{uwypuklanie na zwartymcor2}, we have that for any $\xi\in X$, the function $$ \varphi_\xi(x)=e^{|x-\xi|^2}[f(x)-m+D],\quad x\in \mathbb{R}^n $$ is $\mu$-strongly convex on $X$ for some $\mu>0$. Since we are looking for lower critical points of $ f $, so without loss of generality, we may assume that $-m+D=0$, therefore $$ \varphi_\xi(x)=e^{|x-\xi|^2}f(x),\quad x\in \mathbb{R}^n $$ is $\mu$-strongly convex function in $X$ for any $\xi\in X$. Any strictly convex function $\varphi$ defined on a compact and convex set $X$ has the unique point, denoted by $\operatorname{argmin}_X \varphi$, in which the function $\varphi$ has the minimal value on the set $X$. Therefore, chosing an arbitrary point $a_0 \in X$, we can determine by induction a sequence $a_\nu \in X$, $\nu\in\mathbb{N}$, in the following way \begin{equation}\label{eqanu} a_\nu :=\operatorname{argmin}_X \varphi_{a_{\nu -1}}\quad \hbox{for }\nu \geq 1. \end{equation} \begin{tw}\label{approxi} Let $X\subset \mathbb R^n$ be a compact convex semialgebraic set and $f:\mathbb R ^n \rightarrow \mathbb R$ a positive polynomial on $X$. Let $a_\nu $ be a sequence defined as $a_ \nu :=\operatorname{argmin}_X \varphi_{a_{\nu -1}}$ with $a_0 \in X.$ Then the limit $$ a_*=\lim_{\nu \rightarrow \infty } a_ \nu $$ exist and $a_* \in \Sigma _X f.$ \end{tw} The proof of Theorem \ref{approxi} follows word by word the proof of Theorem 6.5 in \cite{KS}, where we should use the following three lemmas instead of the corresponding lemmas in \cite{KS}. \begin{lemat} For any $\nu \in \mathbb N$, we have $$|a_{\nu +1}-a_{\nu }|=\operatorname{dist}(a_\nu , f^{-1}(f(a_{\nu +1})) \cap X).$$ \end{lemat} \begin{lemat}\label{lemma44x} For any $\nu \in \mathbb N$ we have $$f(a_{\nu +1})\leq \frac{f(a_\nu )-\frac{\mu }{2}|a_{\nu +1}-a_\nu |^2}{e^{|a_{\nu +1}-a_\nu |^2}}.$$ In particular the sequence $f(a_\nu )$ is decreasing. \end{lemat} \begin{proof} Since $\varphi _\xi $ is strongly convex, the definition of $a_{\nu +1}$ implies that the function $$ [0,1] \ni t\mapsto \varphi _{a_\nu }(a_\nu +t(a_{\nu +1}-a_\nu )) $$ decrease, so $\langle a_{\nu +1}-a_{\nu }, \nabla \varphi _{a_{\nu }}(a_{\nu +1 })\rangle \leq 0$. Again by the fact that $\varphi _{a_\nu }$ is $\mu$-strictly convex, we get $$ f({a_\nu })\geq f(a_{\nu +1})e^{|a_{\nu }-a_{\nu +1}|^2}+\frac{\mu}{2}|a_{\nu }-a_{\nu +1} |.$$ This gives the assertion. \end{proof} We can also addapt the following lemma (\cite[Lemma 6.3]{KS}). \begin{lemat} Let $f:[0,\eta ]\rightarrow \mathbb R$ be a $C^1$ function such that $0<f\leq C$ and $f'\leq -\eta $ on $[0,\eta ]$ for some $C\geq \frac{1}{2}$ and $\eta >0.$ Assume that $\varphi (x)=e^{x^2}f(x)$ is strictly convex on $[0,\eta ].$ Then $b_1:=\operatorname{argmin}_{[0,\eta ]}\varphi \geq \frac{\eta }{2C}.$ Hence $f(0)-f(b_1)\geq \frac{\eta ^2}{2C}.$ \end{lemat} \begin{rem}\label{effectivemin} The function $\varphi_{a_{\nu-1}}$ is defined by using the function $\exp$. However, to determine the minimum value of this function on a compact convex semialgebraic set $X$ it is enough to solve only polynomial equations and inequalities. More precisely, the set $X$ is the union of a finite collection of basic semialgebraic sets, so we may assume that $$ a_\nu\in X=\{x\in\mathbb{R}^n:g_1(x)\geq 0,\ldots,g_k(x)\geq 0\}, $$ where $g_1,\ldots,g_k\in\mathbb{R}[x]$. Then $$ X=\{x\in\mathbb{R}^n:g_1(x)e^{|x-a_{\nu-1}|^2}\geq 0,\ldots,g_k(x)e^{|x-a_{\nu-1}|^2}\geq 0\}. $$ Therefore, when applying Lagrange Multipliers or Karush-Kuhn-Tucker Theorem to compute the point $a_\nu$ it is enought to solve a system of polynomial equations and inequalities. \end{rem} \subsection{Searching lower critical points in an unbounded set} Let $X\subset \mathbb R^n$ be a convex and closed semialgebraic set. Let $f\in \mathbb R[x]$ be a polynomial of degree $d>0$ of the form \eqref{funkcja f} and let $f_d$ be the leading form of $f$. Assume that $f_{d*} >0$. Then by Theorem \ref{convexonconvexsetnewvar}, we may effectively compute a real number $N\geq 1$ such that the function $\varphi_{N,\xi}(x)=e^{N|x-\xi|^2}f(x)$ for $\xi\in\mathbb R^n$ is strongly convex on $X$. Moreover, $\varphi_{N,\xi}(x)\geq f(x)\geq m(f)|x|^d$ for $x\in X$, $|x|\geq \mathbb{K}(f)$, so we have \begin{equation*}\label{factlim} \lim_{x\in X,\,|x|\to\infty} \varphi_{N,\xi}(x)=+\infty. \end{equation*} Then we may uniquely determine the sequence \begin{equation}\label{eqanu2} a_\nu :=\operatorname{argmin}_{X} \varphi_{a_{\nu -1}}\quad \hbox{for }\nu \geq 1. \end{equation} Analogous argument as for Theorem \ref{approxi} gives the following theorem. \begin{tw}\label{approxi2} Let $a_\nu $ be a sequence defined by \eqref{eqanu2}, Then the limit $$ a_*=\lim_{\nu \rightarrow \infty } a_ \nu $$ exist and $a_* \in \Sigma _X f.$ \end{tw} \begin{rem}\label{generalcase} If $X\subset \mathbb R^n$ is a closed and convex semialgebraic set and a polynomial $f\in\mathbb R[x]$ is positive on $X$ and it is proper on $X$ (i.e., $\lim_{x\in X,\,|x|\to \infty}f(x)=+\infty$), then by Theorem \ref{convexonconvexsetnewxdoublevarsemi} one can repeat the argument from Theorem \ref{approxi} and obtain a sequence $a_\nu\in X$ such that $\lim_{\nu\to\infty}a_\nu=a_*\in\Sigma_Xf$. If we assume only that $f(x)>0$ on $X$, then the sequence $a_\nu$ can tend to infinity. Moreover, in the construction of $a_\nu$ we have to change $N$ step by step \end{rem}

https://arxiv.org/abs/1710.05342

Three problems on exponential bases

We consider three special and significant cases of the following problem. Let D be a (possibly unbounded) set of finite Lebesgue measure in R^d. Find conditions on D for which the standard exponential basis on the unit cube of R^d is a frame, a Riesz sequence, or a Riesz basis on L^2(D).

\section{Introduction} \label{sec:intro} We are interested in the following problem. Let $D\subset\mathbb R^d$ be a set of Lebesgue measure $|D|<\infty$. Let $E( \Z^d)=\{e^{2\pi i n\cdot x}\}_{n\in\Z^d}$ be the standard exponential basis for the unit cube $ \cal Q_d =[-\frac 12, \frac 12]^d$. Can $E(\Z^d)$ be frame, or a Riesz sequence, or a Riesz basis for $L^2(D)$? We have recalled definitions and general facts about frames, Riesz sequences and Riesz bases in Section 2. Our investigation was motivated by the following problems: \begin{itemize} \item[{\bf P1.}] {\it (The broken interval)}. Let $ J= [0,\alpha)\cup [\alpha+r, L+r)$, with $0<\alpha<\, L$ and $r>0$. For which values of the parameters is the set $E(\Z)$ a Riesz basis, a Riesz sequence or a frame in $L^2(J)?$ \end{itemize} It is easy to verify that $E(\Z)$ is a frame on $J$ when $L+r\leq 1$ and it is a Riesz sequence when either $\alpha\ge 1$ of $L-\alpha\ge 1$ (see also Lemma \ref{L-dil-basis}). It is proved in \cite{Laba} that $E(\Z)$ is an orthonormal basis for $J$ if and only if the measure of $J$ is $L=1$ and the "gap" $ r$ is a non-negative integer. \begin{itemize} \item[{\bf P2.}] {\it (The rotated square)}. Let $Q_h= [-\frac h2, \frac h2]\times [-\frac h2, \frac h2]$ be a square with side $h>0$. For $\theta\in [0, 2\pi)$, we let $\rho_\theta:\mathbb R^2\to\mathbb R^2$ be the rotation $\rho_\theta(x,y)= (x\cos\theta-y\sin\theta, \ x\sin\theta+y\cos\theta)$. For which values of $\theta $ is $E(\Z^2)$ a Riesz basis, a Riesz sequence or a frame on $ \rho_\theta( Q_h)$? \end{itemize} % Also the solution to this problem is trivial only for certain values of the parameters (for example, when $\theta$ is an integer multiple of $\frac \pi 2$). \medskip The next problem was kindly suggested by Chun Kit Lai. \begin{itemize} \item[{\bf P3.}] {\it (The translated parallelepiped)}. Let $P \subset \mathbb R^d$ be a parallelepiped with sides parallel to the vectors $ v_1$, ...,\, $ v_d\in\mathbb R^d$. Find conditions on these vectors for which that the set $E(\Z^d)$ is a Riesz basis, a Riesz sequence, or a frame in $L^2(P)$ . \end{itemize} We recall that a {\it lattice} is the image of $\Z^d$ by a linear invertible transformation \newline $B:\mathbb R^d\to\mathbb R^d$ and we observe that Problem 3 is equivalent to the following: {\it for which lattices $\Lambda=B\Z^d$ is the set $E(B\Z^d)=\{e^{2\pi i Bn\cdot x}\}_{n\in\Z^d}$ a Riesz basis, or a Riesz sequence, or a frame in $L^2(\cal Q_d)$? } Problem 3 is related to certain optimization problems on lattices that have deep applications in computer sciences and in cryptography. See Section 7.1 for details and references. \medskip We first prove necessary and sufficient conditions for which $E(\Z^d)$ is a Riesz sequence or a frame on a given domain $D\subset \mathbb R^d$ and then we completely solve Problems 1, 2 and 3. We let \begin{equation}\label{e-Phi} \Phi(x)=\sum_{m\in\Z^d} \chi_D(x+m), \end{equation} where $\chi_D$ denotes the characteristic function of $D$. Note that $\Phi(x)$ only takes non-negative integer values. Our first result is the following \begin{theorem} \label{T-Riesz } $E(\Z^d)$ is a Riesz sequence in $L^2(D)$ if and only if there exist constants $0<A\leq B <\infty$ for which $A\leq \Phi(x) \leq B$ for a.e. $x\in \cal Q_d $. \end{theorem} That is, we prove that $E(\Z^d)$ is a Riesz sequence in $L^2(D)$ if and only if the integer translates of $D$ (i.e., the sets $D+n=\{x+n,\, x\in D\}$, with $n\in\Z^d$) cover $\mathbb R^d $ with the possible exception of a set of measure zero. \medskip It is interesting to compare Theorem \ref{T-Riesz } with results in \cite{AAC, K, GL}. In these papers the authors consider domains that {\it multi-tile } $\mathbb R^d$, i.e. bounded measurable sets $S \subset\mathbb R^d$ for which there exist a set of translations $\Lambda $ and an integer $h>0$ such that $\sum_{\lambda\in\Lambda} \chi_{S+\lambda}(x)\equiv h$ a.e.; if $h=1$, we say that $S$ {\it tiles} $\mathbb R^d$. It is proved in \cite[Theorem 1]{GL} and in \cite[Theorem 1]{K} that bounded domains that multi-tile $\mathbb R^d$ with a lattice of translation have an exponential basis; in the recent \cite[Theorem 4.4] {AAC} the converse of \cite[Theorem 1]{K} is proved. If $\Phi$ is as in \eqref{e-Phi} and $\Phi(x)\equiv k$ a.e., then $D$ multi-tiles $\mathbb R^d$ with lattice of translations $\Z^d$. By Theorem \ref{T-Riesz }, $E(\Z^d)$ is a Riesz sequence on $L^2(D)$; when $D$ is bounded, it is shown in \cite[Theorem 1]{K} that $E(\Z^d)$ can be completed to an exponential basis for $L^2(D)$, but when $D$ is not bounded an example in \cite{AAC} shows that that may not be possible. \medskip Next, we investigate conditions for which $E(\Z^d)$ is a frame on $D $. The following result is proved in \cite[Lemma 2.10]{GLi}. See also the recent \cite[Theorem 2]{BHM}. \begin{theorem} \label{T-frame} $E(\Z^d)$ is a frame on $L^2( D)$ if and only if for every $ m, s\in\Z^d$, with $ m\ne s$, we have that \begin{equation}\label{e-ass-fr} |(D+ m)\cap (D+ s)|=0. \end{equation} \end{theorem} In other words, $E(\Z^d)$ is a frame in $L^2(D)$ if and only if the integer translates of $D$ only overlap on sets of measure zero. Equivalently, $E(\Z^d)$ is a frame on $L^2( D)$ if and only if $ \Phi(x) \leq 1$ for a.e. $x\in\mathbb R^d$. \medskip We finally prove the following \begin{theorem}\label{T-basis} Assume that $|D|=1$. The following are equivalent in $L^2(D)$: \begin{itemize}\item [a)] $E(\Z^d)$ is a frame \item [b)] $E(\Z^d)$ is a complete \item [c)] The integer translates of $D$ tile $\mathbb R^d$. \item[d)] $E(\Z^d)$ is an orthonormal Riesz basis \item[e)] $E(\Z^d)$ is a Riesz sequence \end{itemize} \end{theorem} We recall that a set $\{w_i\}_{i\in I}$ is {\it complete} in a Hilbert space $ (H ,\ \l \ , \ \r_H)$ if and only if $\l u, w_i\r_H=0$ for every $i\in I$ implies $u=0$. A frame is complete but the converse is not necessarily true. B. Fuglede proved in \cite{F} that if $\Lambda=A\Z^d$ is a lattice in $\mathbb R^d$, the set $E(A\Z^d)$ is an orthogonal exponential basis in $L^2(D)$ if and only if $\{D +\mu\}_{\mu\in( A^t)^{-1} \Z^d }$ tiles $\mathbb R^d$. Here, $(A^t)^{-1}$ denotes the inverse of the transpose of $A$. Thus, the equivalence of c) and d) in Theorem \ref{T-basis} is a special case of Fuglede's theorem. The connections between tiling and exponential bases are deep and interesting and have been intensely investigated. We refer the reader to the introduction and to the references cited in \cite{BHM}. See also \cite{K2}. \medskip This paper is organized as follows: in Section 2 we present preliminary definitions and known results. We prove Theorems \ref{T-Riesz }, \ref{T-frame} and \ref{T-basis} in Sections 3 and 4. We solve Problems 1, 2 and 3 in Sections 5, 6 and 7. \medskip \noindent {\it Acknowledgments.} The first author wishes to thank C.K. Lai for bringing Problem 3 to our attention and E. Hernandez for stimulating discussions that helped us improve the quality of this paper. We also wish thank the anonymous referees for their thorough reading of our manuscript and for their valuable suggestions. \section{ Preliminary and notation} We denote with $x\cdot y= x_1y_1+\,...\,+ x_dy_d$ the inner product of $x=(x_1, ...,\, x_d)$, $y=(y_1, ...,y_d)\in\mathbb R^d$. We let $\|f\|_2=\left(\int_{\mathbb R^d} |f(x)|^2 dx\right)^{\frac 12}$ be the standard norm in $L^2(\mathbb R^d)$; we let ${\bf c}=\{c_j\}_{j\in\Z^d}$ and we denote with $\|{\bf c}\|_{\ell^2}=(\sum_{j\in\Z^d } |c_j|^2)^{\frac 12}$ the standard norm in $\ell^2(\Z^d)$. We denote with $\l f ,\, g \r_{2}=\int_{\mathbb R^d} f(x)\bar g(x)dx$ the inner product in $L^2(\mathbb R^d)$. When there is no ambiguity we will use the same notation also for the inner product in $\ell^2(\Z^d)$. The Fourier transform of a function $f\in L^2(\mathbb R^d)\cap L^1(\mathbb R^d)$ is $\hat f(x)=\int_{\mathbb R^d} f(t) e^{-2\pi i x\cdot t} dt. $ We will often say that a family of sets $\{D_\lambda\}_{\lambda\in\Lambda}$ { covers} $\mathbb R^d$ with the understanding that $\mathbb R^d-\cup_{\lambda\in\Lambda} D_\lambda $ may be a nonempty set of measure zero. We use the notation $\tau_w$ to denote the translation operator $g \to g (\cdot+w)$. \subsection{Frames and Riesz bases} We have used the excellent textbooks \cite{Heil} and \cite{Cr} for most of the definitions and preliminary results presented in this section. Let $H$ be a separable Hilbert space with inner product $\langle\ ,\ \rangle $ and norm $||\ ||=\sqrt{\l \ , \ \r} $. % A sequence of vectors ${\mathcal V}= \{v_j\}_{j\in\Z} \subset H $ is a {\it frame} if there exist constants $0< A, \ B<\infty$ such that the following inequality holds for every $w\in H$. \begin{equation}\label{e2-frame} A||w||^2\leq \sum_{j\in\Z} |\l w, v_j\r |^2\leq B ||w||^2. \end{equation} % We say that ${\cal V}$ is a {\it tight frame } if $A=B$ and is a {\it Parseval frame} if $A=B=1$. The left inequality in \eqref{e2-frame} implies that ${\cal V}$ is complete in $H$ but it may not be linearly independent. A {\it Riesz basis} is a linearly independent frame. An equivalent definition of Riesz basis is the following: the set ${\mathcal V}$ is a {\it Riesz sequence} if there exists constants $0<A\leq B <\infty$ such that, for every finite set of coefficients $ \{a_j\}_{j\in J}\subset\mathbb C $, we have that \begin{equation}\label{e2- Riesz-sequence} A \sum_{j\in J} |a_j|^2 \leq \left\Vert \sum_{j\in J} a_j v_j \right\Vert^2 \leq B \sum_{j\in J} |a_j|^2, \end{equation} and it is {\it Riesz basis} if it also satisfies \eqref{e2-frame}. If ${\mathcal V}$ is a Riesz basis, the constants $A$ and $B$ in \eqref{e2-frame} and \eqref{e2- Riesz-sequence} are the same (see \cite[Proposition 3.5.5]{Cr}). An orthonormal basis is a Riesz basis; we can write $w=\sum_{j\in\Z} \l v_j, \, w\r v_j$ for every $v\in H$ and this representation formula yields the following important identities: for every $ w,\ z\in H$, \begin{equation}\label{e-Planch} ||w||^2= \sum_{n\in\Z } |\l v_n,\, w\r|^2, \quad \l w, z\r= \sum_{n\in\Z } \l v_n,\, w\r\overline{\l v_n, z\r}. \end{equation} The following useful proposition can be found in \cite[Prop. 3.2.8]{Cr}. \begin{proposition}\label{prop-C} A sequence of unit vectors in $H$ is a Parseval frame if and only if it is an orthonormal Riesz basis. \end{proposition} \medskip Let $D\subset \mathbb R^d$ be a measurable set, with $|D|<\infty$. An {\it exponential basis} of $L^2(D)$ is a Riesz basis made of functions in the form of $ e^{2\pi i x\cdot \lambda }$, where $ \lambda \in\mathbb R^d$. Exponential bases are important in the applications because they allow one to represent functions in $L^2(D)$ in a stable manner, with coefficients that are easy to calculate. The following lemma is easy to prove (see e.g \cite[Prop. 2.1]{DK}). \begin{lemma}\label{L-dil-basis} Let $D_1\subset D\subset D_2$ be measurable sets of $\mathbb R^d$ , with $|D_2|<\infty$. % Let ${\cal V}=\{ e^{2\pi i x\cdot \lambda_n }\}_{n\in\Z}$ be Riesz basis of $L^2(D)$ with frame constants $0<A\leq B<\infty$; then, ${\cal V}$ is a Riesz sequence on $L^2(D_2)$ and a frame on $L^2(D_1)$ with the same frame constants. \end{lemma} \subsection{The Beurling density} In \cite{B, B1} A. Beurling characterized sampling sets by means of their density. For $h > 0$ and $x \in\mathbb R^d$, we let $Q _h(x)$ denote the closed cube centered at $x$ with side length $h$. Let $\Lambda=\{\lambda_j\}_{j\in\Z}\subset \mathbb R^d$ be {\it uniformly discrete}, i.e., we assume that $|\lambda_j-\lambda_k|\ge \delta>0$ whenever $\lambda_j\ne\lambda_k$. Following \cite{CDH} we denote with \begin{align*} {\mathcal D}^+(\Lambda) &= \limsup_{h\to \infty} \frac{\sup_{x\in\mathbb R^d} |\Lambda\cap Q_h(x)|}{h^d}\\ {\mathcal D}^-(\Lambda) &= \liminf_{h\to \infty} \frac{\inf_{x\in\mathbb R^d} |\Lambda\cap Q_h(x)|}{h^d} \end{align*} the upper and lower density of $\Lambda$. If ${\mathcal D}^-(\Lambda)={\mathcal D}^+(\Lambda)$ we say that $\Lambda$ has {\it uniform Beurling density} ${\mathcal D}(\Lambda)$. Theorem \ref{T-density-exp} below is a generalization of theorems of Landau and Beurling \cite{B, Landau} in dimension $d\ge 1$. See also \cite{NO} and \cite[Sect. 2]{S}. \begin{theorem}\label{T-density-exp} If $E(\Lambda)=\{e^{2\pi i \lambda_j\cdot x}\}_{j\in\Z}$ is a frame in $L^2(D)$, then ${\mathcal D}^-(\Lambda)\ge |D|$. If $E(\Lambda)$ is a Riesz sequence in $L^2(D)$, then ${\mathcal D}^+(\Lambda)\leq |D|$. \end{theorem} Thus, a necessary condition for $E(\Lambda)$ to be a Riesz basis in $L^2(D)$ is that ${\mathcal D} (\Lambda)= |D|$. In the special case of $\Lambda=\Z^d$ we have the following \begin{corollary}\label{C-density} If $E(\Z^d)$ is a frame in $L^2(D)$ then $|D|\leq 1$; if $E(\Z^d)$ is a Riesz sequence in $L^2(D)$ then $|D|\ge 1$. \end{corollary} \subsection{Shift invariant spaces} We let $$ V^2(\varphi):=\overline{\Span \{\tau_k\varphi \}_{k\in\Z^d } } $$ where $\varphi\in L^{2}(\mathbb R^d)$ and ``bar'' denotes the closure in $L^2(\mathbb R^d)$. The space $ V^2(\varphi)$ is {\it shift-invariant}, i.e. if $f\in V^2(\varphi)$ then also $\tau_m f\in V^2(\varphi)$ for every $m\in\Z^d$. Shift-invariant spaces of functions appear naturally in signal theory and in other branches of applied sciences. Following \cite{Aldroubi}, \cite{Christiansen}, \cite{CCS} we say that the translates $\{ \tau_k\varphi \}_{k\in\Z^d}$ form a Riesz basis in $V^2 ( \varphi ) $ if there exist constants $0<A,\ B<\infty$ such that, for every finite set of coefficients ${\bf d}= \{d_j\} \subset\mathbb C $, we have that \begin{equation}\label{E-p-basis-2} A \|{\bf d}\|_{\ell^2} \leq \| \sum_j d_j \tau_j\varphi \|_{2} \leq B\|{\bf d}\|_{\ell^2}. \end{equation} If \eqref{E-p-basis-2} holds, then $ V^2(\varphi )= \left\{ f= \sum_{k\in\Z^d}d_k \tau_k\varphi, \ {\bf d} \in \ell^2\ \right\}, $ and the sequence $\{d_k \}_{k\in\Z^d}$ is uniquely determined by $f$. \medskip The following theorem is well known: see e.g \cite{JM} or \cite[Prop. 1.1]{AS}. \begin{theorem} \label{T-basis2} The set $\{ \tau_m\varphi \} _{m\in\Z^d}$ is a Riesz basis in $V^2(\varphi)$ with frame constants $0<A,\ B<\infty$ if and only if, \begin{equation}\label{e1} A= \inf_{y\in \cal Q_d}\sum_{m \in\Z^d} |\hat \varphi(y+m)|^2 \leq \sup_{y\in \cal Q_d}\sum_{m \in\Z^d} |\hat \varphi(y+m)|^2 = B. \end{equation} \end{theorem} \section{ Proof of Theorem \ref{T-Riesz }} Let $\ell^2_0(\Z^d)\subset \ell^2(\Z^d)$ be the set of sequences ${\bf a}=(a_n)_{n\in\Z^d}$ such that $a_n=0 $ whenever $|n|\ge N$, with $N=N({\bf a})\ge 0$. Let $ S({\bf a}) = \sum_{n\in\Z^d} a_n e^{2\pi i n\cdot x}$. Recall that $E(\Z^d)$ is a Riesz sequence in $L^2(D)$ if and only if there exists constants $0<A,\ B<\infty$ such that \begin{equation}\label{ineq-S} A\|{\bf a}\|_2^2\leq \|S({\bf a}) \|_{L^2(D)} ^2\leq B\|{\bf a}\|_2 ^2 \end{equation} for every ${\bf a}\in \ell^2_0(\Z^d).$ We gather: \begin{align}\nonumber \|S({\bf a}) \|_{L^2(D)} ^2= & \int_D \left|\sum_{n\in\Z^d} a_n e^{2\pi i n\cdot x}\right|^2 dx= \int_D\left(\sum_{n, m\in\Z^d} a_n \overline{a_m}\, e^{2\pi i (n-m)\cdot x}\right)dx \\\label{Sa} =& \sum_{n, m\in\Z^d} a_n \overline{a_m}\, \int_D e^{2\pi i (n-m)\cdot x} dx = \sum_{n, m\in\Z^d} a_n \overline{a_m} \widehat{\chi_D}(n-m). \end{align} Let $T_D$ be the operator, initially defined in $\ell^2_0(\Z^d)$, as: \begin{equation}\label{e-TD} T_D({\bf a})_m= \sum_{n \in\Z^d} a_n \widehat{\chi_D}(n-m), \ m\in\Z^d. \end{equation} The calculation above shows that $ \|S({\bf a}) \|_{L^2(D)} ^2=\l T_D({\bf a}), \ {\bf a}\r_2, $ where $\l\, , \,\r_2 $ denotes the inner product in $\ell^2(\Z^d)$. We can easily verify that $T_D({\bf a})$ is self-adjoint and, in view of \eqref{Sa}, that $ \l T_D({\bf a}),\ {\bf a}\r_2\ge 0$ for every ${\bf a}\in\ell^2_0(\Z^d)$; thus, \eqref{ineq-S} holds if and only if \begin{equation}\label{e-1} A\|{\bf a}\|_2\leq \l T_D({\bf a}), \ {\bf a}\r_2 \leq B\|{\bf a}\|_2,\quad {\bf a}\in \ell^2_0(\Z^d). \end{equation} % To prove \eqref{e-1} we need the following \begin{lemma}\label{L-Haase} Assume that $\dsize ||T_D||_{\ell^2\to\ell^2}= \sup_{||{\bf a}||_2=1}|| T_D ({\bf a})||_2 <\infty$. The inequality below holds for every ${\bf a}\in \ell^2_0(\Z^d)$ such that $||{\bf a} ||_2=1$. \begin{equation}\label{e2} \displaystyle \frac{ ||T_D({\bf a})||_2^2}{ ||T_D||_{\ell^2\to\ell^2}}\leq \l T_D({\bf a}),\, {\bf a}\r_2 \leq ||T_D||_{\ell^2\to\ell^2} . \end{equation} \end{lemma} \begin{proof}[Proof of Theorem \ref{T-Riesz }] Let $\Phi(x) $ be as in \eqref{e-Phi}. We show that if there exist constants $0<A' \leq B' <\infty$ such that $A' \leq \Phi(x)\leq B' $ a.e. in $\cal Q_d$, then $E(\Z^d)$ is a Riesz sequence in $L^2(D)$. Since $\Phi(x)=\sum_{m \in\Z^d} \chi_D(x+m) =\sum_{m \in\Z^d} |\chi_D(x+m)|^2 $, by Theorem \ref{T-basis2} the set $\{ \tau_m\hat \chi_D \} _{m\in\Z^d}$ is a Riesz basis of $V^2(\hat \chi_D )$ with frame constants $0<A' ,\ B' <\infty$. In view of \eqref{E-p-basis-2} and \eqref{e-TD}, the inequality \begin{equation}\label{e-prime} A' \|{\bf a}\|_2\leq \|T_D{\bf a}\|_2^2\leq B' \|{\bf a}\|_2 \end{equation} holds for every ${\bf a}\in \ell^2_0(\Z^d)$. By Lemma \ref{L-Haase}, we have \eqref{e-1}, as required. If $E(Z^d)$ is a Riesz sequence on $D$, we argue as in the proof of \cite[Theorem 3.1]{selvan}. Using Plancherel's identity and the Poisson summation formula, from \eqref{Sa} we obtain \begin{align}\nonumber ||S({\bf a}) ||_{L^2(D)} ^2 &= \sum_m {\Big \vert} \sum_{n\in\Z^d} a_n\widehat{\chi_D}(n-m) {\Big \vert} ^2 = \int_{\cal Q_d} {\Big\vert} \sum_{m\in\Z^d} \sum_{n\in\Z^d} a_n\widehat{\chi_D}(n-m) e^{2\pi i x\cdot m} {\Big\vert}^2dx \\ \nonumber & =\int_{\cal Q_d} {\Big\vert} {\Big( } \sum_{n\in\Z^d} a_n e^{2\pi i x\cdot n} {\Big )}\sum_{m\in\Z^d} \widehat{\chi_D}(n-m) e^{2\pi i x\cdot (m-n)} {\Big\vert}^2dx % \\ \label{new e}& = \int_{\cal Q_d} \, {\Big\vert}\sum_{n\in\Z^d} a_n e^{2\pi i n\cdot x}{\Big\vert}^2\, |\Phi(x)|^2 dx. % \end{align} By assumption, the integral in \eqref{new e} is finite and so so $\Phi(x) <\infty$ a.e.. To show that $\Phi(x) >0$ a.e. we argue by contradiction: suppose that there exists $\Omega\subset D$, with $|\Omega|>0$, where $\Phi(x)\equiv 0$. We can assume that $\Omega\subset \cal Q_d$. Since $E(\Z^d)$ is a Riesz basis in $L^2(\cal Q_d)$, we can write $\chi_{\cal Q_d-\Omega}(x)=\sum_{n\in\Z^d} b_ne^{2\pi i n\cdot x}$, with $\vec b\in \ell^2(\Z^d)$. Thus, $\int_{\cal Q_d} |\Phi(x)|^2 \, \left|\sum_{n\in\Z^d} b_n e^{2\pi i n\cdot x}\right|^2dx =0$ which, together with \eqref{new e}, contradicts \eqref{ineq-S}. \end{proof} \begin{proof}[Proof of Lemma \ref{L-Haase}] The right inequality in \eqref{e2} is \cite[Theorem 13.8]{Haase} so we only need to prove the left inequality. Let $\alpha= \sup_{||{\bf a}||_2=1}|\l T_D({\bf a}),\, {\bf a}\r|$ and $U= \alpha I-T_D$, where $I$ is the identity operator in $\ell^2(\Z^d)$. It is easy to verify that $U$ is positive and that \begin{equation}\label{e-id3} T_D\,U\,T_D+\,U\,T_D\,U\,=\alpha^2T_D-\alpha T_D^2. \end{equation} The operators $T_D\,U\,T_D$ and $\,U\,T_D\,U\,$ are positive too; indeed, for every ${\bf a}\in \ell^2$, we have that $\l T_D\,U\,T_D {\bf a}, \, {\bf a}\r_2= \l \,U\,(T_D{\bf a}), \ T_D{\bf a}\r_2\ge 0$ and $\l \,U\,T_D\,U\,{\bf a} , \, {\bf a}\r_2= \l T_D(\,U\,{\bf a}), \ \,U\, {\bf a}\r_2 \ge 0$ because $T_D$ and $\,U\,$ are both positive. By \eqref{e-id3}, also the operator $\alpha T_D- T_D^2$ is positive. For every ${\bf a}\in \ell^2$ with $||{\bf a}||_2=1$, we have that $$ \l (\alpha T_D- T_D^2) {\bf a}, \ {\bf a}\r_2= \alpha \l T_D{\bf a}, \ {\bf a}\r_2- \l T_D^2{\bf a}, \ {\bf a}\r_2 = \alpha \l T_D{\bf a}, \ {\bf a}\r_2-||T_D{\bf a}||_2^2\ge 0 $$ and the left inequality in \eqref{e2} is proved. \end{proof} \medskip \noindent {\it Remark.} From the identity \eqref{new e} it follows that the constants $A$ and $B$ in \eqref{ineq-S} are the minimum and maximum of $\Phi(x)$ on the unit square $\cal Q_d$. Thus, $A $ and $B $ are integers. When $|D|=1$, Theorem \ref{T-basis} shows that $E(\Z^d)$ is a Riesz sequence if and only the integer translates of $D$ tile $\mathbb R^d$, and so $A =B =1$. In general, if $k \leq |D|< k+1$ for some positive integer $k$, we can easily verify that the integer translates of $D $ cover $\mathbb R^d$ $k$ times but not $k+1$ times. Thus, $A \leq |D|$ and $B \ge |D|$. \section{Proof of Theorem \ref{T-frame} } \medskip Let $D\subset \mathbb R^d$ be measurable, with $|D|\leq 1$. By Lemma \ref{L-dil-basis}, the theorem is trivial when $D\subset \cal Q_d $, so we assume that $ D- \cal Q_d $ has positive measure. Let $D_1$, ...,\, $D_N ,\,...$ be a (possibly infinite) family of disjoint sets of positive measure such that $D-\cal Q_d= \cup_{j} D_j$. We can choose the $D_j$ in such way that, for certain vectors $v_1$, ... $v_N,\, ... \in \Z^d$, we have that $D_j+v_j\subset \cal Q_d. $ Let $D_0= D\cap \cal Q_d $ and $v_0=0$. % We prove the following \begin{lemma} \label{L-frame2} $E(\Z^d)$ is frame for $L^2( D)$ if and only, for every $ v \in\Z^d$ and every $k\ne j$, \begin{equation}\label{e-assumptions-frame}|( v +D_j)\cap D_k |= 0. \end{equation} \end{lemma} It is easy to verify that \eqref{e-assumptions-frame} is equivalent to \eqref{e-ass-fr}, and so Theorem \ref{T-frame} is equivalent to Lemma \ref{L-frame2}. \begin{proof} Assume that $|(D_1 + v) \cap D_0|>0 $ for some $v\in\Z^d$ (the proof is similar in the other cases). We can assume without loss of generality that $D_1 + v \subset \cal Q_d $ (see Figure 1); otherwise we let $D_1= D_1' \cup D_1''$, with $D'_1+v\subset \cal Q_d$ and we replace $D_1$ with $D_1'$. We show that $E(\Z^d)$ is not a frame on $L^2(D)$. \begin{figure}[h!] \begin{center} \begin{tikzpicture} % \fill[gray!20] (-1.5,0)-- (-1.5, 3)-- ( -.5,3)--(0, 4)--(.5,3)--(1.5,3)--(1.5, 0)--( .7, 0)-- (.7, .7)--(-.7, .7)-- (-.7, 0)--(-1.5,0); \fill[gray!20] ( -.5,3)--(0, 4)--(.5,3)-- (-.5,3); \draw[black]( -.5,0)--(0, 1)--(.5,0)-- (-.5,0); \fill[gray!60] ( -.5,0)--(0, 1)--(.5,0)-- (-.5,0); \draw[ black] (-1.5,0)-- (-1.5, 3)-- ( -.5,3)--(0, 4)--(.5,3)--(1.5,3)--(1.5, 0)--( .7, 0)-- (.7, .7)--(-.7, .7)-- (-.7, 0)--(-1.5,0);; \draw[thick, black, dashed] (-1.5,0)--(-1.5,3)--(1.5, 3)--(1.5,0)--(-1.5, 0); \draw [-> ] (0,3)-- (0,1.5); \draw (.2 ,2.2) node [black ] {$ v$}; \draw (-.5 ,1.5) node [black ] {$D_0$}; \draw (0 ,3.3) node [black ] { \small{$D_1$}}; \draw (0 , .2) node [black ] { \tiny{$ D_1\!+\!v$}}; \draw (2 , 1.5) node [black ] {$\cal Q_d $}; \end{tikzpicture} \caption{ } \end{center} \end{figure} Every $f\in L^2(D)$ can be written as $f = f_0 + f_1$ where $f_0 = f \chi_{D-D_1}$ and $f_1 = f \chi_{D_1}$. % Recall that $\tau_w g(x)= g (x+w)$. It follows that \begin{align}\nonumber |\inner{e^{2\pi i n\cdot x}}{f}_{L^2(D)}|^2 &= |\inner{e^{2\pi i n\cdot x}}{f_0}_{L^2(D-D_1)} + \inner{e^{2\pi i n\cdot x}}{f_1}_{L^2(D_1)}|^2 \\\label{1} &= |\inner{e^{2\pi i n\cdot x }}{f_0}_{L^2(D-D_1)} + \inner{e^{2\pi i n\cdot (x-v)}}{\tau_{- v}f_1 }_{L^2( D_1+v))}|^2 \\\nonumber &= |\inner{e^{2\pi i n\cdot x}}{f_0}_{L^2(\cal Q_d )} + \inner{e^{2\pi i n\cdot x}}{\tau_{-v}{f_1}}_{L^2(\cal Q_d )}|^2 \\\nonumber &= |\inner{e^{2\pi i n\cdot x}}{f_0}_{L^2(\cal Q_d )}|^2 + |\inner{e^{2\pi i n\cdot x}}{\tau_{-v}{f_1}}_{\cal Q_d )}|^2 \\\nonumber & \quad + 2\Real{\left(\inner{e^{2\pi i n\cdot x}}{f_0}_{L^2(\cal Q_d )} \conjugate{\inner{e^{2\pi i n\cdot x}}{\tau_{-v}{f_1}}_{L^2(\cal Q_d )}}\right)}. \end{align} We have used the change of variables $x\to x-v$ in the second inner product in \eqref{1} and the fact that $e^{2\pi i n\cdot v}=1$. Since $E(\Z^d)$ is an orthonormal basis in $\cal Q_d $, the identities \eqref{e-Planch} in Section 2 and the calculation above yield % \begin{equation}\label{e-id1} % \begin{split} \sum_{n \in \Z^d} |\inner{e^{2\pi i n\cdot x}}{f}_{L^2(D)}|^2 % &= \norm{f_0}_{L^2(\cal Q_d )}^2 + \norm{ \tau_{-v} f_1 }_{L^2(\cal Q_d )}^2 \\ & + 2\Real{ \inner{f_0}{\tau_{-v}{f_1}}_{L^2(\cal Q_d )} }. \end{split} \end{equation} If we let $B= (D-D_1)\cap (D_1 +v)$ we can choose $ f=f_1+f_0$, with $ f_1(x)= \chi_{B}(x+v)$ and $f_0(x) =- \chi_B(x) $; from \eqref{e-id1} it readily follows that % $\sum_{n \in \Z^d} |\inner{e^{2\pi i n\cdot x}}{f}_{L^2(D)}|^2 =0$, which contradicts \eqref{e2-frame}. \medskip We now assume that $|(w+D_j)\cap D_k|=0$ for every $k\ne j$ and every $w\in\Z^d$; we prove that $E(\Z^d)$ is a tight frame in $L^2(D)$. We assume for simplicity that $ D_1+v \subset \cal Q_d $ for some $v \in\Z^d$. Let $f=f_0+f_1$ be as in the first part of the proof. By assumption, $|(D_1+v) \cap (D-D_1)| =0$, and so \eqref{e-id1} yields \begin{align*}\sum_{n \in \Z^d} |\inner{e^{2\pi i n\cdot x}}{f }_{L^2(D)}|^2& = \norm{f_0}_{L^2(\cal Q_d )}^2 + \norm{ \tau_{-v} f_1 }_{L^2(\cal Q_d )}^2 \\ &=\norm{f \chi_{D-D_1}}_{L^2(\cal Q_d )}^2 + \norm{ f \chi_{D_1 +v}}_{L^2(\cal Q_d )}^2 = ||f||_{L^2(D)}. \end{align*} Thus, $E(\Z^d)$ is a tight frame in $L^2(D)$ as required. \end{proof} \medskip The proof of Theorem \ref{T-frame} shows that if \eqref{e-assumptions-frame} is not satisfied, we can produce a function $f\in L^2(D)$ for which $\l f,\, e^{2\pi i x\cdot n}\r_{L^2(D)}=0$ for every $n\in\Z^d$, and so $E(\Z^d)$ is not complete. This observation proves the following: \begin{corollary}\label{C-complete} $E(\Z^d)$ is complete in $L^2(D)$ if and only if the integer translates of $D$ intersect on sets of measure $0$. \end{corollary} \begin{proof}[Proof of Theorem \ref{T-basis}] We have proved that a) $\iff$ b); we show that $b)\iff c)$. By Corollary \ref{C-complete}, $E(\Z^d)$ is complete in $L^2(D)$ if and only if the integer translates of $D$ overlap only on sets of measure zero. Thus, c) $\Rightarrow$ b). Let us prove that b) $\Rightarrow$ c); let $D_0,\, D_1$, ...,\, $D_N ,\,...$ and $v_0, \,v_1, \, ...,\, \, v_N,\, ...$ be as in the proof of Lemma \ref{L-frame2}. Since $|(D_j+v_j)\cap (D_k+v_k)|=0 $ when $k\ne j$, and \begin{equation}\label{e-Q-D} 1= |\cal Q_d|=|D|= |D_0| + \sum_j |D_j+v_j| \end{equation} necessarily $ \cup_j (D_j+v_j) =\cal Q_d$ and the integer translates of $D$ tiles $\mathbb R^d$. By Fuglede's theorem, c) $\iff$ d). Clearly d) $\Rightarrow$ e); to finish the proof of the theorem we show that e) $\Rightarrow$ c). By Theorem \ref{T-Riesz }, the integer translates of $D$ cover $\mathbb R^d$; thus, $ \cup_j (D_j+v_j) =\cal Q_d$ and from \eqref{e-Q-D} follows that the $D_j+v_j$'s can only intersect on set of measure zero. Thus, the integer translated of $D$ can only intersect on sets of measure zero and c) is proved. \end{proof} \section{The broken interval} In this section we solve the first problem stated in the introduction. We let $J= [0,\alpha)\cup [\alpha+r, L+r)\subset \mathbb R$, with $0<\alpha<L$ and $r>0$. By Lemma \ref{L-frame2} and Theorem \ref{T-Riesz }, $E(\Z)$ is a frame on $L^2(J)$ if and only if the integer translates of $J$ do not overlap in $[0,1]$ and it is a Riesz sequence if and only if the integer translates of $J$ cover $\mathbb R$. \medskip Let $[r]$ be the integer part of $r$, i.e., the largest integer $n \leq r$; let $\{r\}= r - [r]$ be the fractional part of $r$. We prove the following \begin{theorem}\label{T-riesz-interval} a) $E(\Z)$ is a frame on $J$ if and only if $L + \{r\} \leq 1$. \\ b) $E(\Z)$ is a Riesz sequence on $J $ if and only if one of the following is true: \begin{itemize} \item[i)] $\alpha \geq 1$ or $L - \alpha \geq 1$ \item[ ii)] $\{r\} = 0$ and $L \geq 1$ \item[ iii)] $1 \leq L < 2$ and $L + \{r\} \geq 2$; \end{itemize} \medskip \end{theorem} % To prove b) we will need the following \begin{lemma}\label{L-frac-part} The integer translates of $J$ cover $\mathbb R$ if and only if the integer translates of $J' = [0,\alpha) \cup [\alpha + \{r\}, L + \{r\})$ cover $\mathbb R$. \end{lemma} \begin{proof} If the integer translates of $J$ cover $\mathbb R$ then for $ x\in\mathbb R$, there is an integer $m$ such that either $x \in (m, \alpha +m)$ or $x \in (\alpha + r + m,\ L + r + m)$. If $x \in (m, \alpha + m)$ then clearly $x\in J'+m$ as well. If $x \in (\alpha + r + m,\ L + r + m)$, then $x \in (\alpha + \{r\} + [r] + m,\ L + \{r\} + [r] + m)$, i.e. $x$ is in the translation of $J'$ by $[r]+m$. The converse is similar. \end{proof} \begin{proof}[Proof of Theorem \ref{T-riesz-interval}] By Theorem \ref{T-frame} and Lemma \ref{L-frac-part}, $E(\Z)$ is a frame on $J$ if and only if the integer translates of $[0,\alpha)\cup[\alpha + \{r\}, L + \{r\})$ do not intersect in $[0,1]$, or if and only if $(0,\alpha) \cap (\alpha + \{r\} - 1, L + \{r\} - 1) = \varnothing$. This is equivalent to having either $\alpha \leq \alpha + \{r\} -1$, which is impossible, or $L + \{r\} \leq 1.$ That proves part a). \medskip Let us prove part b). By Lemma \ref{L-frac-part} we can assume that $r = \{r\}$, i.e. that $0 \leq r < 1$. By Theorem \ref{T-Riesz }, $E(\Z)$ is a Riesz sequence on J if and only if the integer translates of J cover $\mathbb R$. If one of the connected components $[0,\alpha)$ or $[\alpha + r, L + r)$ covers $\mathbb R$ by integer translations, we have that either $\alpha \geq 1$ or $L - \alpha \geq 1$, and i) is proved. If neither component covers $\mathbb R$ by integer translations, i.e. if both $\alpha < 1$ and $L - \alpha < 1$, we can consider 2 sub-cases: \begin{itemize}\item If $r = 0$, we have that $J=[0, L)$, and the integer translates of $J$ cover $\mathbb R$ if and only if $L \geq 1$; that proves ii). \item Suppose next that $r > 0$. The integer translates of $J$ cover $\mathbb R$ if and only if the "gap" $(\alpha, \alpha + r)$ is covered by integer translates of $J$. This is possible if and only if $ (1, \alpha +1) \cup (\alpha + r - 1, L + r -1)\supset (\alpha, \alpha + r) $ (see Figure2). \begin{figure}[h!] \begin{center} \begin{tikzpicture} % \draw[ black, ->] (-.5 ,0 )-- ( 5 , 0 ) ; \draw[ black, thick] (0,0)-- ( 1.5, 0); \draw[ black, thick] ( 2.5,0)-- ( 3.5, 0 ); \draw[ black, thick] (1,1)-- ( 2.5, 1); \draw[ black, thick] ( 3.5,1)-- ( 4.5, 1); \draw ( 0,-.2 ) node [black ] {{\tiny $ 0 $}}; \draw ( .2, .2 ) node [black ] {{\tiny $ J $}}; \draw ( 1.5, -.2) node [black ] {{\tiny $ \alpha $}}; \draw (2.5, -.2 ) node [black ] {{\tiny $ r+\alpha $}}; \draw (3.5,-.2 ) node [black ] {{\tiny $ r+L $}}; \draw (5,1 ) node [black ] {{\tiny{\bf $J+1 $}}}; \draw [black,fill] (1,1 ) circle [radius=0.06]; \draw [black,fill] (2.5,1 ) circle [radius=0.06]; \draw [black,fill] (3.5,1 ) circle [radius=0.06]; \draw [black,fill] (4.5,1 ) circle [radius=0.06]; \draw[black, dashed] (1,1)--(1,0); \draw [black,fill] (0, 0) circle [radius=0.06]; \draw [black,fill] (1.5, 0) circle [radius=0.06]; \draw [black,fill] (2.5, 0) circle [radius=0.06]; \draw [black,fill] (3.5, 0) circle [radius=0.06]; \draw [black,fill] (1, 0) circle [radius=0.06]; \draw (1,-.2 ) node [black ] {{\tiny $ 1 $}}; \end{tikzpicture} \caption{ } \end{center} \end{figure} We have $(1, \alpha + 1) \cap (\alpha, \alpha + r) = (1, \alpha + r)$, because $\alpha , r < 1$. Thus, J covers $\mathbb R$ if and only if $(\alpha + r -1, L + r - 1) \cap (\alpha, \alpha + r) \supset (\alpha, 1)$. This is equivalent to the conditions % \begin{center} $ \begin{cases} r - 1 \leq 0 \\ L + r -1 \geq 1 \\ \alpha + r \geq 1 \end{cases} \Longleftrightarrow \begin{cases} r \leq 1 \\ L + r \geq 2 \\ \alpha + r \geq 1 \end{cases}.$ \end{center} % Since $\alpha < 1$ and $L - \alpha < 1$ by assumption, and recalling that $r < 1$, we can see at once that the condition $L + r \geq 2$ implies $\alpha + r \geq 1$. Indeed, if $\alpha + r < 1$, then $$L + r = L - \alpha + \alpha + r < L - \alpha + 1 < 2.$$ Thus, the integer translates of $J$ covers $\mathbb R$ if and only if $L + r \geq 2$, and we have iii). The theorem is proved. \end{itemize} % \end{proof} \section{The rotated square} Let $ Q_h= Q_h (0)=[-\frac{h}{2}, \frac{h}{2}] \times [-\frac{h}{2}, \frac{h}{2}]$ be the square in $\mathbb R^2$ centered at the origin with sides of length $h$. Let $A_\theta=\begin{bmatrix} \cos{(\theta)} & \sin{(\theta)} \\ -\sin{(\theta)} & \cos{(\theta)} \end{bmatrix}$ be the matrix of a rotation by an angle $\theta$, and let $ Q_{h,\theta} =A_{\theta}Q_h(0) $ be the square obtained from the rotation of $Q_h(0)$. The following theorem offers a complete solution to Problem 2: \begin{theorem} a) $E(\Z^2)$ is a Riesz sequence on $L^2(Q_{h,\theta})$ if and only if $ h\ge 1-\sin(2\theta) $. b) $E(\Z^2)$ is a frame on $L^2(Q_{h,\theta})$ if and only if $h\leq \frac{1}{\sin\theta+\cos\theta}$. \end{theorem} \begin{proof} We first prove Part a). Let $P_1= (\frac 12, \frac 12), \ P_2= (-\frac 12, \frac 12)\ P_3= (-\frac 12, -\frac 12)\ P_4= (\frac 12, -\frac 12)$ be the vertices of $\cal Q_2$. We first find conditions on $h$ and $\theta$ for which the points $P_1$, ..., $P_4$ lie on the sides of $Q_{h,\theta}$. \begin{figure}[h!] \begin{center} \begin{tikzpicture} % \draw[ black, dashed] (-1.5 ,-1.5 )-- ( 1.5 , -1.5 )-- (1.5 , 1.5 )--(-1.5 , 1.5 )--(-1.5 , -1.5 ); \fill[gray!10](-2.4,0)-- ( 0, -2.4)-- ( 2.4,0)-- (0, 2.4 )--(-2.4,0); \draw[ black] (-2.4,0)-- ( 0, -2.4)-- ( 2.4,0)-- (0, 2.4 )--(-2.4,0); \draw[ black, dashed] (-1.5 ,-1.5 )-- ( 1.5 , -1.5 )-- (1.5 , 1.5 )--(-1.5 , 1.5 )--(-1.5 , -1.5 ); \draw[ black] (-3 ,0)-- ( 0, -3 )-- ( 3 ,0)-- (0, 3 )--(-3 ,0); \draw (0 ,0) node [black ] { $ Q_{h,\theta}$}; % \draw (1.8, -1.8 ) node [black ] {{\small $ P_4 $}}; \draw (1.8, 1.8 ) node [black ] {{\small $ P_1 $}}; \draw (-1.8, 1.8 ) node [black ] {{\small $ P_2 $}}; \draw (-1.8, -1.8 ) node [black ] {{\small $ P_3 $}}; \draw (-3.3,0 ) node [black ] {{\small $ Q_3 $}}; \draw (3.3,0 ) node [black ] {{\small $ Q_1 $}}; \draw (0, -3.3 ) node [black ] {{\small $ Q_4 $}}; \draw (0,3.3) node [black ] {{\small $ Q_2 $}}; \draw (2.5,2 ) node [black ] {{ $Q_{l(\theta), \theta}$}}; \fill[gray!60] ( 1.5 , 1.5 )-- ( 1.1 , 1.5 )-- (1.1 , 1.3 )--( 1.5 , 1.3)--( 1.5 , 1.5 ); \draw[ black, thick] ( 1.5 , 1.5 )-- ( 1.1 , 1.5 )-- (1.1 , 1.3 )--( 1.5 , 1.3)--( 1.5 , 1.5 ); \fill[gray!60] ( -1.5 , -1.5 )-- (- 1.1 , -1.5 )-- (-1.1 , -1.3 )--(- 1.5 , -1.3)--( -1.5 , -1.5 ); \draw[ black, thick] (- 1.5 ,- 1.5 )-- ( -1.1 , - 1.5 )-- (-1.1 , -1.3 )--( -1.5 , -1.3)--( -1.5 , -1.5 ); \fill[gray!60] ( 1.5 , -1.5 )-- ( 1.1 , -1.5 )-- ( 1.1 , -1.3 )--( 1.5 , -1.3)--( 1.5 , -1.5 ); \draw[ black, thick] ( 1.5 ,- 1.5 )-- ( 1.1 , - 1.5 )-- ( 1.1 , -1.3 )--( 1.5 , -1.3)--( 1.5 , -1.5 ); \fill[gray!60] ( -1.5 , 1.5 )-- (- 1.1 , 1.5 )-- (-1.1 , 1.3 )--(- 1.5 , 1.3)--( -1.5 , 1.5 ); \draw[ black, thick] (- 1.5 , 1.5 )-- ( -1.1 , 1.5 )-- (-1.1 , 1.3 )--( -1.5 , 1.3)--( -1.5 , 1.5 ); \draw [black,fill] (1.5, 1.5) circle [radius=0.06]; \draw [black,fill] (1.5, -1.5) circle [radius=0.06]; \draw [black,fill] (-1.5, 1.5) circle [radius=0.06]; \draw [black,fill] (-1.5,- 1.5) circle [radius=0.06]; \end{tikzpicture} \caption{ } \end{center} \end{figure} Let $\ell_{1}: y-\frac 12=\tan(\theta)(x-\frac 12)$, $ \ell_{2}: y-\frac 12=-\frac 1{\tan(\theta) } (x+\frac 12)$ and $\ell_{3}: y+\frac 12=\tan(\theta)(x+\frac 12) $ be the equations of the sides of $Q_{h,\theta}$ that contain the points $P_1$, $P_2$ and $P_3$, resp. It is easy to verify that $\ell_2$ intersects $\ell_1$ and $\ell_3$ at the points $ Q_2= \left( -\frac 12\cos(2\theta), \ \frac 12(1-\sin(2\theta))\right)$ and $ Q_3= \left( -\frac 12(1-\sin(2\theta)) , \ -\frac 12\cos(2\theta)\right) $, and that the length of the segment that join $Q_2$ and $Q_3$ equals to $ l(\theta) =1-\sin(2\theta)$. Thus, when $h\ge l(\theta)$, the set $E(\Z^2)$ is a Riesz sequence on $L^2(Q_{h, \theta})$. We show that when $h < l(\theta)$ the integer translates of $Q_{h, \theta}$ do not cover the plane anymore. Indeed, if $h < l(\theta)$, the four vertices of $\cal Q_2$ are outside the square $Q_{l(\theta), \theta}$ and have positive distance from the boundary of $Q_{h,\theta}$. We can find a small rectangle $R$ with sides parallel to the sides of $\cal Q_2$ for which $ R+P_j \subset \cal Q_2-Q_{h, \theta}$ for every $j$ (see Figure 3). The integer translates of $Q_{h,\theta}$ cannot cover the rectangles $R+P_j$ and so the condition of Theorem \ref{T-Riesz } is not verified. \medskip \begin{figure}[h!] \begin{center} \begin{tikzpicture} % \draw[ black, dashed] (-2, 1)--(3,1)--(3,6)--(-2, 6)--(-2, 1); \draw (3.5,5.2) node [black ] {{\small $ \cal Q_2 $}}; \draw[black] (1,7)--(4,3)--(0,0)--(-3,4)--(1,7); \fill[gray!60](.5,6)--(.5, 6.3)--(.2, 6.3)--(.2, 6)--(.5, 6); \draw[black ] (.5,6)--(.5, 6.3)--(.2, 6.3)--(.2, 6)--(.5, 6); % \draw (.1,6.5) node [black ] {{\small ${ R } $}}; \fill[gray!30](1,6)--(-2,4)--(0,1)--( 3,3)--(1,6); \draw[black] (1,6)--(-2,4)--(0,1)--( 3,3)--(1,6); \draw (1, 6.4) node [black ] {{\small $ Q_1 $}}; \draw (-2.4, 4) node [black ] {{\small $ Q_2 $}}; \draw (0, .6 ) node [black ] {{\small $ Q_3 $}}; \draw (3.4, 3) node [black ] {{\small $ Q_4 $}}; \draw (0,3) node [black ] {{ $Q_{s(\theta), \theta}$}}; \draw [black,fill] (1,6) circle [radius=0.06]; \draw [black,fill] (-2,4) circle [radius=0.06]; \draw [black,fill] (0,1)circle [radius=0.06]; \draw [black,fill] (3,3)circle [radius=0.06]; \end{tikzpicture} \caption{ } \end{center} \end{figure} Let us prove Part b): the vertices $Q_1$, ..,\, $Q_4$ of $Q_{h,\theta}$ lie on the sides of $\cal Q_2$ if and only if there exists $0<t\leq 1$ for which $Q_1=(t-\frac 12,\, \frac 12)$, $Q_2= (-\frac 12,\, t-\frac 12)$, $Q_3= ( \frac 12 -t,\, -\frac 12)$ and $Q_4= ( \frac 12, \, \frac 12 -t)$. If we let $\tan(\theta)$ be the slope of the line that joins $Q_3$ and $Q_4$, we can see at once that $ \tan(\theta)= \frac {1-t}{t}$; thus, $ t=\frac{1}{1+\tan(\theta)}= \frac{\cos(\theta)}{\sin(\theta)+\cos(\theta)}.$ The length of the segment $[Q_3Q_4]$ is then: $$ s(\theta)= \sqrt{ t^2+(1-t )^2}= \frac{1}{\sin(\theta)+\cos(\theta)} $$ and $E(\Z^2)$ is a frame on $ Q_{h,\theta}$ whenever $h\leq s(\theta)$. Let us show that $E(\Z^2)$ is not a frame on $Q_{h,\theta}$ whenever $h>s(\theta)$. Indeed, if $h>s(\theta)$, the set $Q_{h,\theta}-\cal Q_2$ has positive measure; we can find a small rectangle $R$ with sides parallel to the sides of $\cal Q_2$ and vectors $n_1$, ..., $n_4\in\Z^2$ such that $R+n_j\subset Q_{h,\theta}-\cal Q_2$ (see Figure 3). Thus, the integers translates of $Q_{h,\theta}$ overlap on the rectangles $R+n_j$ and by Theorem \ref{T-frame}, $E(\Z^2)$ is not a frame on $L^2(Q_{h,\theta})$. \end{proof} \section{The translated parallelepiped} In this section we solve Problem 3. Let $P\subset \mathbb R^d$ be a parallelepiped with sides parallel to vectors $ v_1$, ...,\, $ v_d$. We let $A=\{a_{i,j}\}_{1\leq i,j\leq d}$ be the matrix whose columns are $ v_1$, ...,\, $ v_d$; we let $A^{-1}=\{b_{i,j}\}_{1\leq i,j\leq d}$. We prove the following \begin{theorem}\label{T-frame-parall} a) The set $E(\Z^d)$ is a frame on $L^2(P)$ if and only if $\det(A) \leq 1$ and $\max_{{1\leq i,k, j\leq d}\atop{j\ne k}}|a_{i,j}|+ |a_{i,k}|\leq 1$ b) The set $E(\Z^d)$ is a Riesz sequence on $L^2(P)$ if and only if $\det(A ) \ge1$ and $\max_{1\leq i, j\leq d} |b_{i,j}| \leq 1$. \end{theorem} \begin{figure}[h!] \begin{center} \begin{tikzpicture} % \draw[ black, ->] (-1, 0)--(7,0); \draw[ black, ->] (0,-1)--(0, 5); \draw[black] (0,0)--(2,2)--(3.5,2)--(1.5,0)--(0,0); \draw[black] (1.8 ,1.5)--(3.8 ,3.5)--(5.3 ,3.5)--(3.3 ,1.5)--(1.8 ,1.5); \draw(1, -.3) node[black] {$v_1$}; \draw(1.8, -.3 ) node[black] {$1$}; \draw (1.1 ,.5) node [black ] {{\small $ P $}}; \draw (6 ,2.8) node [black ] {{\small $ P +(1,1) $}}; \draw[black, dashed] (1.8, 1.5)-- (1.8,0); \draw[black, dashed] (1.8, 1.5)-- (0,1.5); \draw(-.2, 1.5 ) node[black] {$1$}; \draw(1.7, 2 ) node[black] {$v_2$}; \end{tikzpicture} \caption{ } \end{center} \end{figure} \begin{proof} Observe that if $|P|= \det(A) > 1$, the set $E(\Z^d)$ cannot be a frame on $L^2(P)$ so in part a) we assume $\det(A) \leq 1$. Similarly, for part b) we assume that $\det(A) \ge 1$. \medskip We prove part a) by induction on the dimension $d$. By Theorem \ref{T-frame}, the set $E(\Z^d)$ is a frame on $L^2(P)$ if and only if the integer translates of $P$ overlap only on sets of zero measure. In dimension $d=2$, we let $ v_1=(a_{1,1}, a_{2,1})$ and $ v_ 2=(a_{1,2}, a_{2,2})$ be the vectors that are parallel to the sides of $P$. When the components of $v_1$ and $v_2$ are non-negative, we can easily verify that $P$ overlaps with $P+(1,1)$ if and only if the sum of the projections of $v_1$ and $v_2$ on the $x_1$ and $x_2$ axes has measure $\ge 1$ (see Figure 4). Thus, $P$ overlaps with $P+(1,1)$ if and only if $ a_{1,1} + a_{1,2} \ge 1$ and $a_{2,1} + a_{2,2} \ge 1$. These conditions imply that no pair of integer translates of $P$ intersect. For general $v_1$ and $v_2$ we can similarly verify that the integer translates of $P$ do not intersect if and only if $ | a_{1,1} |+ |a_{1,2}| \ge 1$ and $|a_{2,1}| + |a_{2,2} |\ge 1$. We now assume that part a) of the theorem is valid in dimension $d\ge 2$. We prove that is is valid also in dimension $d+1$. Let $P$ be a parallelepiped in $\mathbb R^{d+1}$. The integer translates of $P$ overlap on sets of positive measure in $\mathbb R^{d+1}$ if and only if the integer translates of the faces of $P$ overlap on sets of positive measure in $\mathbb R^{d}$. Let $P_h$ be the face of $P$ spanned by the vectors $ v_1$, ..., $ v_{h-1}, \ v_{h+1}$, ...,\, $ v_{d+1}$. Let $ e_1=(1, 0, ...,\,0)$, ..., $ e_{d+1} =(0, ...0,\,1)$ be the standard orthonormal basis in $\mathbb R^{d+1}$ and let $H_j$ be the orthogonal complement of $e_j$. Clearly, the integer translates of $P_h$ overlap if and only if the integer translates of the orthogonal projections of $P_h$ on the $H_j$'s overlap. The projection of $P_h$ on $H_k$ is a parallelepiped in $\mathbb R^d$ spanned by the vectors $ w_1$, ..., $ w_{h-1}, \ w_{h+1}$, ...,\, $ w_{d+1}$ where $ w_j$ is the projection of $v_j$ on $H_k$, i.e., it is the vector $ v_j$ with the $k$-th component removed. By assumptions, $\max_{{1 \leq i,k, j\leq d+1\atop{i\ne k}}\atop{k\ne j\ne h}}\{|a_{i,j }| + |a_{i,k }| \}\leq 1$. This inequality is valid for every face of $P$ and for every projection, and so we have that $\max_{{1 \leq i,j, k\leq d+1 }\atop{k\ne j}}\{|a_{i,k}|+ |a_{i,j}|\}\leq 1$ as required. \medskip We now prove part b). By Theorem \ref{T-Riesz }, the integer translates of $P$ must cover $\mathbb R^d$. Since $P$ is the image of the unit cube $[0,1]^d$ via the linear transformation $A(x)= A x$, we can write $P=A([0,1]^d)$. Thus, $E(\Z^d)$ is a Riesz sequence in $L^2(P)$ if and only if $ \bigcup_{n\in\Z^d} (A([0,1]^d) + n) =\mathbb R^d $, or: $$ A^{-1}\left(\bigcup_{n\in\Z^d} (A([0,1]^d) + n)\right)= \bigcup_{n\in\Z^d} ( [0,1]^d + A^{-1} n)=\mathbb R^d.$$ The translates of the unit cube $[0,1]^d$ cover $\mathbb R^d$ if and only if the components of the vectors $A^{-1}e_k$ are all $\leq 1$, i.e., if and only if $\max_{1\leq i,j\leq d}|b_{i,j}|\leq 1$. \end{proof} \subsection{The shortest vector problem} Let $A:\mathbb R^d\to\mathbb R^d$ be linear and invertible; consider the parallelepiped $P=A(Q)$, where $Q = [0,1]^d.$ The sides of $P$ are parallel to the columns of the matrix that represents $A$. By Corollary \ref{C-complete}, the set $E(\Z^d)$ is complete in $L^2(A(Q))$ if and only if the integer translates of $A(Q)$ do not intersect. % The integer translates of $A(Q)$ intersect if and only if there are $x,\ y \in Q $ such that $Ax = Ay + n$, for some nonzero $n \in \Z^d$. We can also say that the translates of $A(Q_d)$ intersect if and only if there exist $x, y \in Q$ and $n \in \Z^d$ such that $A^{-1}n = x-y$, i.e., if and only if there exists $n \in \Z^d$ such that $A^{-1}n \in D = \{w \mid \| w\|_{\infty} < 1 \}$. \medskip These considerations show that Problem 3 is related to the so-called {\it shortest vector problems} (SVP): Given a lattice $\mathcal{L}$ and a norm $||\ ||$ on $\mathbb{R}^d$, find the minimum length $\lambda = \min_{0 \neq v \in \mathcal{L}}\| v\|$ of a nonzero lattice point. The SVP is known to be NP-hard (see \cite{Aj}). The conjectured intractability of the SVP and of other optimization problems on lattices is central in the construction of secure lattice-based cryptosystems. For more information on this problem see e.g. \cite{AD}, \cite{Kn} and the references cited in these papers. \medskip

https://arxiv.org/abs/1801.01724

A Lipschitz condition along a transversal foliation implies local uniqueness for ODEs

We prove the following result: if a continuous vector field $F$ is Lipschitz when restricted to the hypersurfaces determined by a suitable foliation and a transversal condition is satisfied at the initial condition, then $F$ determines a locally unique integral curve. We also present some illustrative examples and sufficient conditions in order to apply our main result.

\section{Introduction} Uniqueness for ODEs is an important and quite old subject, but still an active field of research \cite{cons,DiNoSi14,ferreira}, being Lipschitz uniqueness theorem the cornerstone on the topic. Besides the existence of many generalizations of that theorem, see \cite{agalak,cl,hartman}, one recent and fruitful line of research has been the searching for alternative or weaker forms of the Lipschitz condition. For instance, let $U\subset{\mathbb R}^2$ be an open neighborhood of $(t_0,x_0)$ and $f:U\subset{\mathbb R}^2\to{\mathbb R}$ be continuous and consider the scalar initial value problem \begin{equation}\label{eq-ivp-scalar} x'(t)=f(t,x(t)),\, x(t_0)=x_0. \end{equation} It was proved, independently by Mortici, \cite{mortici}, and Cid and Pouso \cite{cidpouso1,cp2}, that local uniqueness holds provided that the following conditions are satisfied: \begin{itemize} \item$f(t,x)$ is Lipschitz with respect to $t$, \item$f(t_0,x_0)\not=0$. \end{itemize} A more general result had been proved before by Stettner and Nowak, \cite{sn}, but in a paper restricted to German readers. They proved that if $U\subset {\mathbb R}^2$ is an open neighborhood of $(t_0,x_0)$, $f:U\subset {\mathbb R}^2\to{\mathbb R}$ is continuous and $(u_1,u_2)\in{\mathbb R}^2$ such that \begin{itemize} \item $|f(t,x)-f(t+ku_1,x+ku_2)|\le L |k|\quad \mbox{on $D$}$, \item $u_2\not= f(t_0,x_0) u_1$, \end{itemize} then the scalar problem \eqref{eq-ivp-scalar} has a unique local solution. By taking either $(u_1,u_2)=(0,1)$ or $(u_1,u_2)=(1,0)$ this result covers both the classical Lipschitz uniqueness theorem and the previous alternative version. Moreover this result has been remarkably generalized in \cite{DiNoSi14} by Dibl{\'\i}k, Nowak and Siegmund by allowing the vector $(u_1,u_2)$ to depend on $t$. Let us now consider the autonomous initial value problem for a system of differential equations \begin{equation}\label{peq-aut} z'(t)=F(z(t)),\ z(t_0)=p_0, \end{equation} where $n\in{\mathbb N}$, $F:U\subset{\mathbb R}^{n+1}\to{\mathbb R}^{n+1}$ and $p_0\in U$. Trough the paper we shall need the following definition: if $g:D\subset{\mathbb R}^{n+1}\to E$, where $E$ is a normed space, we will say that $g$ is {\it Lipschitz in $D$ when fixing the first variable} if there exists $L>0$ such that for all $(s,x_1,x_2,\ldots,x_n),(s,y_1,y_2,\ldots,y_n)\in D$ we have that $$\|g(s,x_1,x_2,\ldots,x_n)-g(s,y_1,y_2,\ldots,y_n)\|_{E} \le L \|(x_1,x_2,\ldots,x_n)-(y_1,y_2,\ldots,y_n)\|,$$ and where $\|\cdot\|$ stands for any norm in ${\mathbb R}^{n+1}$. Moreover, for any function $g$ with values in ${\mathbb R}^{n+1}$ we denote $g=(g_1,g_2,\dots,g_{n+1})$. The following alternative version of Lipschiz uniqueness theorem for systems has been proved by Cid in \cite{Cid03}. \begin{thm} \label{cor-cid} Let $U\subset{\mathbb R}^{n+1}$ an open neighborhood of $p_0$ and $F:U\to{\mathbb R}^{n+1}$ continuous. If moreover \begin{itemize} \item $F$ is Lipschitz in $U$ when fixing the first variable, \item $F_{1}(p_0)\ne0$, \end{itemize} then there exists $\a>0$ such that problem \eqref{peq-aut}has a unique solution in $[t_0-\a,t_0+\a]$. \end{thm} \begin{rem} The classical Lipschitz theorem is included in the previous one. In order to see this, let $n\in{\mathbb N}$, $U\subset{\mathbb R}^{n+1}$ be an open set, $f:U\to{\mathbb R}^n$ and $(t_0,x_0)\in U$ and consider the non-autonomous problem \begin{equation}\label{peq} x'(t)=f(t,x(t)),\ x(t_0)=x_0. \end{equation} As it is well known, problem \eqref{peq} is equivalent to the autonomous one \eqref{peq-aut}, where \begin{displaymath}F(z_1,z_2,\ldots,z_{n+1}):=(1,f(z_1,z_2,\ldots,z_{n+1})),\end{displaymath} and $p_0:=(t_0,x_0)$. Now, if $f(t,x)$ is Lipschitz with respect to $x$ then $F(z_1,z_2,\ldots,z_{n+1})$ is Lipschitz when fixing the first variable and moreover $F_{1}(p_0)=1\ne0$, so Theorem \ref{cor-cid} applies. \end{rem} Recently, Dibl{\'\i}k, Nowak and Siegmund have obtained in \cite{SiNoDi16} a generalization of both \cite{Cid03} and \cite{sn}. Their result reads as follows: \begin{thm}\label{thDNS} Let $U\subset{\mathbb R}^{n+1}$ an open neighborhood of $p_0$, $F:U\to{\mathbb R}^{n+1}$ be continuous and $\cal{V}$ a linear hyperplane in ${\mathbb R}^{n+1}$ such that \begin{itemize} \item $F$ is Lipschitz continuous along $\cal{V}$, that is, there exists $L>0$ such that if $x,y\in U$ and $x-y\in \cal{V}$ then $$\|F(x)-F(y)\|\le L \|x-y\| $$ \item Transversality condition: $F(p_0)\not\in \cal{V}$. \end{itemize} then there exists $\a>0$ such that problem~\eqref{peq-aut} has a unique solution in $[t_0-\a,t_0+\a]$. \end{thm} The previous theorem has the following geometric meaning: uniqueness for the autonomous system \eqref{peq-aut} follows provided that the continuous vector field $F$ is Lipschitz when restricted to a family of parallel hyperplanes to $\cal{V}$ that covers $U$ and that the vector field at the initial condition $F(p_0)$ is transversal to $\cal{V}$. Our main goal in this paper is to extend Theorem \ref{thDNS} from the linear foliation generated by the hyperplane $\cal{V}$ to a general $n$-foliation. The paper is organized as follows: in Section 2 we present our main result which relies on an appropriate change of coordinates and Theorem \ref{cor-cid}. We will show by examples that our result is in fact a meaningful generalization of Theorem \ref{thDNS}. In Section 3 we present some useful results about Lipschitz functions, including the definition of a modulus of Lipschitz continuity along a hyperplane that will be used in Section 4 for obtaining explicit sufficient conditions on $F$ for the existence of a suitable $n$-foliation. Another key ingredient for that result shall be a general rotation formula proved too at Section 4. Through the paper $\<\cdot,\cdot \>$ shall denote the usual scalar product in the Euclidean space. \section{The main result: a general uniqueness theorem} \begin{dfn}Let $p_0\in {\mathbb R}^{n+1}$. Assume there exist open subsets $V\subset{\mathbb R}^n$, $U\subset{\mathbb R}^{n+1}$, an open interval $J\subset {\mathbb R}$ with $0\in J$ and a family of differentiable functions $\{g_s:V\to U\}_{s\in J}$ such that $g_0(0)=p_0\in U$ and $\Phi:(s,y)\in J\times V \to g_s(y)\in U$ is a diffeomorphism. Then we say $\{g_s\}_{s\in J}$ is a \emph{local $n$-foliation of U at $p_0$}. \end{dfn} \begin{rem}An observation regarding notation. If $\Phi:{\mathbb R}^{n+1}\to{\mathbb R}^{n+1}$ is a diffeomorphism, we denote by $\Phi'$ its derivative and by $\Phi^{-1}$ its inverse. Also, we write $(\Phi^{-1})'$ for the derivative of the inverse. Observe that $\Phi'$ takes values in ${\mathcal M}_{n+1}({\mathbb R})$ so, although we cannot consider the functional inverse of $\Phi'$, we can consider the inverse matrix, whenever it exists, of every $\Phi'(x)$ for $x\in{\mathbb R}^{n+1}$. We denote this function by $(\Phi')^{-1}$. Clearly, the chain rule implies that \begin{displaymath}(\Phi')^{-1}(x)=(\Phi^{-1})'(\Phi(x)).\end{displaymath} \end{rem} The following is our main result. \begin{thm}\label{thmgen} Let $U\subset{\mathbb R}^{n+1}$, $V\subset{\mathbb R}^{n}$ be open sets, $p_0\in U$, $F:U\subset{\mathbb R}^{n+1}\to{\mathbb R}^{n+1}$ a continuous function and $\{g_s:V\to U\}_{s\in J}$ a local $n$-foliation of $U$ at $p_0$ which defines the diffeomorphism $\Phi:J\times V\to U$. If the following assumptions hold, \begin{itemize} \item{(C1)} Transversality condition: \begin{equation}\label{transcon}\<\(\frac{\partial\Phi_1^{-1}}{\partial z_1}(p_0),\dots,\frac{\partial\Phi_1^{-1}}{\partial z_{n+1}}(p_0)\),F(p_0)\>\ne0,\end{equation} \item{(C2)} Lipschitz condition along the foliation: $F\circ \Phi$ and $(\Phi')^{-1}$ are Lipschitz in a neighborhood of zero when fixing the first variable, \end{itemize} then there exists $\a>0$ such that problem~\eqref{peq-aut} has a unique solution in $[t_0-\a,t_0+\a]$. \end{thm} \begin{proof} Consider the change of coordinates \begin{align}\label{coc}z=(z_1,\dots,z_{n+1})=\Phi(s,y_1,\dots,y_n):=g_s(y_1,\dots,y_n).\end{align} Since $\{g_s\}_{s\in J}$ is a foliation, $\Phi$ is a diffeomorphism. Then, considering $y=(s,y_1,\dots,y_n)$, differentiating \eqref{coc} with respect to $t$ and taking into account equation \eqref{peq-aut}, \begin{equation}\label{eqderiv}\frac{\dif z}{\dif t}=\Phi'(y)\frac{\dif y}{\dif t}=F(z)=(F\circ \Phi)(y).\end{equation} Since $\Phi$ is a diffeomorphism, $\Phi'(y)$ is an invertible matrix for every $y$, so \[\frac{\dif y}{\dif t}=\Phi'(y)^{-1}(F\circ \Phi)(y).\] By definition of $g_s$, $\Phi(0)=p_0$, so we can consider the problem \begin{equation}\label{redeq}\frac{\dif y}{\dif t}(t)=h(y),\ y(t_0)=0,\end{equation} where \begin{equation*}h(y)=\Phi'(y)^{-1}F( \Phi(y)).\end{equation*} Now, by (C2) we have that $h$ is the product of locally Lipschitz functions when fixing the first variable. Furthermore, if $e_1=(1,0,\dots,0)\in{\mathbb R}^n$ and taking into account (C1), \[h_1(0)=e_1^T\Phi'(0)^{-1}F(p_0)=e_1^T(\Phi^{-1})'(p_0)F(p_0)=\<\(\frac{\partial\Phi_1^{-1}}{\partial z_1}(p_0),\dots,\frac{\partial\Phi_1^{-1}}{\partial z_{n+1}}(p_0)\),F(p_0)\>\ne0.\] Hence, we can apply Theorem \ref{cor-cid} to problem \eqref{redeq} and conclude that problem~\eqref{peq-aut} has, locally, a unique solution. \end{proof} \begin{rem} 1) Condition \eqref{transcon} can be easily interpreted geometrically: the vector \begin{displaymath}\(\frac{\partial\Phi_1^{-1}}{\partial z_1}(p_0),\dots,\frac{\partial\Phi_1^{-1}}{\partial z_{n+1}}(p_0)\),\end{displaymath} is normal to the hypersurface given by $g_0(V)$ at $p_0$. So, condition \eqref{transcon} means that the vector $F(p_0)$ is not tangent to that hypersurface, and therefore it is called the \emph{'transversality condition'}. \medbreak \noindent 2) Notice that from \cite[Example 3.1]{Cid03} we know that if the transversality condition \eqref{transcon} does not hold then the Lipschitz condition along the foliation, that is (C2), is not enough to ensure uniqueness. On the other hand, by \cite[Example 3.4]{Cid03} we also know that (C1) and a Lipschitz condition along a local (n-1)-foliation do not imply uniqueness. So, in some sense, conditions (C1) and (C2) are sharp. \end{rem} Theorem \ref{thmgen} generalizes the main result in \cite{SiNoDi16}, where only foliations consisting of hyperplanes are considered. In the next example we show the limitations of linear (or affine) coordinate changes which are used in \cite{SiNoDi16}. \begin{exa}\label{exa-change} Let $F(x,y):=1+(y-x^2)^{\frac{2}{3}}$. Is there a linear change of coordinates $\Phi$ such that $F\circ\Phi$ is Lipschitz in a neighborhood of zero when fixing the first variable? The answer is no. Any linear change of variables $\Phi$ will be given by two linearly independent vectors $v, w\in{\mathbb R}^2$ as $\Phi(z,t)=z w+t v$. If $F\circ\Phi$ is Lipschitz in a neighborhood of zero when fixing the first variable, that is, $z$, that implies that the directional derivative of $F$ at any point of the neighborhood in the direction of $v$, whenever it exists, is a lower bound for any Lipschitz constant. To see that this cannot happen, take $S=\{(x,y)\in{\mathbb R}^2\ : y=x^2\}$ and realize that $F$ is differentiable in ${\mathbb R}^2\backslash S$, with \[\nabla F(x,y)=\frac{2}{3}(y-x^2)^{-\frac{1}{3}}(-2x,1),\quad \mbox{for} \, \, (x,y)\in{\mathbb R}^2\backslash S.\] Let $v=(v_1,v_2)\in{\mathbb R}^2$. The directional derivative of $F$ at $(x,y)$ in the direction of $v$ is \[D_vF(x,y)=\<\nabla F(x,y),v\>=\frac{2}{3}(y-x^2)^{-\frac{1}{3}}(v_2-2v_1x),\quad \mbox{for} \, \, (x,y)\in{\mathbb R}^2\backslash S.\] Now consider a neighborhood $N$ of $0$. In particular, we can consider the points of the form $(x,y)=(\l,\l^2+\mu)\in N\backslash S$ for $\mu\ne0$ and $\l\in(-\epsilon,\epsilon)$, so \[D_vF(x,y)=\frac{2}{3}\frac{v_2-2 \lambda v_1}{ \mu ^{1/3}}.\] This quantity is unbounded in $ N\backslash S$ unless the numerator is $0$ for every $\l\in(-\epsilon,\epsilon)$, but that means that $v=0$, so $v$ and $w$ cannot be linearly independent. Hence, no linear change of coordinates $\Phi$ makes $F\circ\Phi$ Lipschitz in a neighborhood of zero when fixing the first variable.\par \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.5]{Example} \end{center} \end{figure} Nevertheless, take $(x,y)=\Phi(z,t)=g_z(t)=(t,z+t^2)$. We have $\Phi^{-1}(x,y)=(y-x^2,x)$ and both are differentiable, so $\Phi$ is a diffeomorphism. Now, $(F\circ\Phi)(z,t)=1+z^\frac{2}{3}$, which is clearly Lipschitz when fixing the first variable. In the figure you can see the parabolas $g_z(t)$ foliating the plane, where $g_0(t)$ is the thicker one. \end{exa} \begin{exa} With what we learned from Example \ref{exa-change}, it is easy to see that uniqueness for the scalar initial value problem \begin{equation}\label{eqex1}x'(t)=1+(x(t)-t^2)^{\frac{2}{3}}, \quad x(0)=0,\end{equation} can not be dealt with \cite[Theorem 2]{SiNoDi16} neither with \cite[Theorem 1]{DiNoSi14}. However, by using the local 1-foliation associated to diffeomorphism $\Phi$ given in Example \ref{exa-change}, it is easy to show that conditions (C1) and (C2) of Theorem \ref{thmgen} are satisfied. Therefore, we have the local uniqueness of solution. \end{exa} \section{Some results about Lipschitz functions} We will now establish some properties of Lipschitz functions that will be useful for checking condition (C2) in Theorem \ref{thmgen}. Before that, consider the following Lemma. \begin{lem}\label{leminv} Let $A,B,C\in{\mathcal M}_n({\mathbb R})$, $A$ and $C$ invertible. Then \[ \|ABC\| \ge \frac{\|B\|}{\|A^{-1}\| \|C^{-1}\|},\] where $\|\cdot\|$ is the usual matrix norm. \end{lem} \begin{proof}It is enough to observe that \[\|B\|=\|A^{-1}ABCC^{-1}\|\le\|A^{-1}\|\|ABC\|\|C^{-1}\|.\] \end{proof} \begin{lem}\label{lemeq}Let $U$ be an open subset of ${\mathbb R}^n$ and $g:U\to GL_n({\mathbb R})$. \begin{enumerate} \item If $g$ is locally Lipschitz and $g^{-1}$ (the inverse matrix function) is locally bounded, then $g^{-1}$ is locally Lipschitz. \item If $g$ is locally Lipschitz when fixing the first variable and $g^{-1}$ is locally bounded, then $g^{-1}$ is locally Lipschitz when fixing the first variable. \end{enumerate} \end{lem} \begin{proof} 1. Let $K$ be a compact subset of $U$, $k_1$ be a Lipschitz constant for $g$ in $K$ and $k_2$ a bound for $g^{-1}$ in $K$. Then, for $x,y\in K$, using Lemma \ref{leminv}, \begin{align*}k_1\|x-y\| & \ge\|g(x)-g(y)\|=\|g(x)(g(y)^{-1}-g(x)^{-1})g(y)\|\ge\frac{\|g(y)^{-1}-g(x)^{-1}\|}{k_2^2}. \end{align*} Hence, $\|g(x)^{-1}-g(y)^{-1}\|\le k_1k_2^2\|x-y\|$ in $K$ and $g^{-1}$ is locally Lipschitz.\par 2. We proceed as in 2. Let $K$ be a compact subset of $U$, $(t,x),(t,y)\in K$, $k_1$ be a Lipschitz constant for $g$ in $K$ when fixing $t$ and $k_2$ a bound for $g^{-1}$ in $K$. Then, \begin{align*}k_1\|x-y\| & \ge\|g(t,x)-g(t,y)\|=\|g(t,x)(g(t,y)^{-1}-g(t,x)^{-1})g(t,y)\|\ge\frac{\|g(t,y)^{-1}-g(t,x)^{-1}\|}{k_2^2}. \end{align*} Hence, $\|g(t,x)^{-1}-g(t,y)^{-1}\|\le k_1k_2^2\|x-y\|$ and $g^{-1}$ is locally Lipschitz when fixing the first variable.\par \end{proof} \begin{cor}\label{coreq}Let $U$ be an open subset of ${\mathbb R}^n$, $f:U\to f(U)\subset{\mathbb R}^n$ be a diffeomorphism (notice that, in that case, $f':U\to GL_n({\mathbb R})$). \begin{enumerate} \item If $f'$ is locally Lipschitz and $(f')^{-1}$ is locally bounded, then $(f')^{-1}$ is locally Lipschitz. \item If $f'$ is locally Lipschitz and $(f')^{-1}$ is locally bounded, then $(f^{-1})'$ is locally Lipschitz. \item If $f'$ is locally Lipschitz when fixing the first variable and $(f')^{-1}$ is locally bounded, then $(f')^{-1}$ is locally Lipschitz when fixing the first variable. \end{enumerate} \end{cor} \begin{proof} 1. Just apply Lemma \ref{lemeq}.1 to $g=f'$.\par 2. Notice that \begin{displaymath}(f^{-1})'(x)=(f')^{-1}(f^{-1}(x)),\end{displaymath} and that $(f')^{-1}$ is locally Lipschitz by the previous claim. On the other hand, since $f'$ is locally continuous we have that $f$ is locally a ${\cal C}^1$-diffeomorphism, and thus $f^{-1}$ is locally Lipschitz. Therefore $(f^{-1})'$ is locally Lipschitz since it is the composition of two locally Lipschitz functions. \par 3. Just apply Lemma \ref{lemeq}.2 to $g=f'$. \end{proof} \subsection{A modulus of continuity for Lipschitz functions along an hyperplane} Let $U$ be an open subset of ${\mathbb R}^{n+1}$, $p_0\in U$ and consider the tangent space of $U$ at $p$, which can be identified with ${\mathbb R}^{n+1}$. Consider now the real Grassmannian $\Gr(n,n+1)$, that is, the manifold of hyperplanes of ${\mathbb R}^{n+1}$. We know that $\Gr(n,n+1)\cong \Gr(1,n+1)={\mathbb P}^n$, that is, we can identify unequivocally each hyperplane with their perpendicular lines, which are elements of the projective space ${\mathbb P}^n$.\par \begin{dfn} Consider $B_{n+1}(p,\d)\subset{\mathbb R}^{n+1}$ to be the open ball of center $p$ and radius $\d$. Then, for a function $F:U\to{\mathbb R}^{n+1}$ and every $p\in U$, $v\in{\mathbb P}^n$ and $\d\in{\mathbb R}^+$ we define the \textit{modulus of continuity} \[\omega_F(p,v,\d):=\sup_{\substack{x,y\in B_{n+1}(p,\d)\\ x-p,y-p\perp v\\x\ne y}}\frac{\|F(x)-F(y)\|}{\|x-y\|}\in[0,+\infty].\] We also define \[\omega_F(p,v):=\lim_{\d\to0} \omega_F(p,v,\d)=\lim_{\d\to0}\sup_{\substack{x,y\in B_{n+1}(p,\d)\\ x-p,y-p\perp v\\x\ne y}}\frac{\|F(x)-F(y)\|}{\|x-y\|}=\lils{\substack{(x,y)\to(p,p)\\x-p,y-p\perp v\\x\ne y}}\frac{\|F(x)-F(y)\|}{\|x-y\|}\in[0,+\infty].\] \end{dfn} \begin{rem} If $\omega_F(p,v)<+\infty$, then there exist $\d,\epsilon\in{\mathbb R}^+$ such that \[\|F(x)-F(y)\|\le(\omega_F(p,v)+\epsilon)\|x-y\|,\ x,y\in B_{n+1}(p,\d),\ x-p,y-p\perp v.\] Equivalently, \[\|F(x+p)-F(y+p)\|\le(\omega_F(p,v)+\epsilon)\|x-y\|,\ x,y\in B_{n+1}(0,\d),\ x,y\perp v.\] Let $A$ be a orthonormal matrix such that its first column is parallel to $v$. In that case, since $A$ is orthogonal, $x\perp e_1$ implies that $Ax \perp v$.Then, \begin{equation*}\|F(Ax+p)-F(Ay+p)\|\le(\omega_F(p,v)+\epsilon)\|A(x-y)\|,\ x,y\in B_{n+1}(0,\d),\ x,y\perp e_1.\end{equation*} That is, taking into account that $\|A\|=1$, \begin{equation*}\|F(A(0,x)+p)-F(A(0,y)+p)\|\le(\omega_F(p,v)+\epsilon)\|x-y\|,\ x,y\in B_{n}(0,\d).\end{equation*} Hence, if $\phi(x)=Ax+p$ then $F\circ \phi$ is locally Lipschitz in an neighborhood of the origin \textit{when the first variable is equal to zero}. \end{rem} The following lemma illustrates the relation between the modulus of continuity $\omega_F$ and the partial derivatives of $F$. \begin{lem}\label{lemreg}Assume $F$ is continuously differentiable in a neighborhood $N$ of $p$. Then \[\omega_F(p,v)=\sup_{\substack{w\perp v\\\|w\|=1}}\|D_wF(p)\|.\] \end{lem} \begin{proof} Since $F'(z)$ is continuous at $p$, for $\{\epsilon_n\}\to 0$ there exists $\{\d_n\}\to 0$ such that if $z\in B_{n+1}(p,\d_n)$ and $\|w\|=1$ then $\|F'(z)(w)\|\le \|F'(p)(w)\|+\epsilon_n$. Hence, using the Mean Value Theorem, \begin{align*} \sup_{\substack{x,y\in B_{n+1}(p,\d_n)\\ x-p,y-p\perp v\\x\ne y}}\frac{\|F(x)-F(y)\|}{\|x-y\|}\le & \sup_{\substack{x,y,z \in B_{n+1}(p,\d_n)\\ x-p,y-p \perp v\\x\ne y}}\frac{\|F'(z)(x-y)\|}{\|x-y\|} \le \sup_{\substack{z \in B_{n+1}(p,\d_n)\\ u \in B_{n+1}(0,2 \d_n) \\ u \perp v\\u\ne 0}}\frac{\|F'(z)(u)\|}{\|u\|} \\ = & \sup_{\substack{z \in B_{n+1}(p,\d_n)\\ d \in (0,2 \d_n) \\ w \perp v\\ \|w\|=1}}\frac{\|F'(z)(d w)\|}{\|d w\|}= \sup_{\substack{z \in B_{n+1}(p,\d_n) \\ w \perp v\\ \|w\|=1}} \|F'(z)(w)\| \\ \le & \sup_{\substack{w \perp v\\ \|w\|=1}} \|F'(p)( w)\|+\epsilon_n = \sup_{\substack{w \perp v\\ \|w\|=1}} \|D_w F(p)\|+\epsilon_n .\end{align*} Then, taking the limit when $n\to\infty$, we obtain \[\omega_F(p,v)\le \sup_{\substack{w\perp v\\\|w\|=1}}\|D_wF(p)\|.\] On the other hand, assume $w\in{\mathbb S}^n$ and $w\perp v$. Then $F(p+tw)=F(p)+t(D_wF(p)+g(t))$ where $g$ is continuous and $\lim_{t\to0}g(t)=0$. Therefore, \[\|D_wF(p)\|=\left\|\frac{F(p+tw)-F(p)}{t}-g(t)\right\|\le \sup_{\substack{x,y\in B_{n+1}(p,t)\\ x-p,y-p\perp v\\x\ne y}}\left[\frac{\|F(x)-F(y)\|}{\|x-y\|}+|g(t)|\right].\] Taking the limit when $t$ tends to zero, $\|D_wF(p)\|\le\omega_F(p,v)$, which ends the proof. \end{proof} \begin{rem} This definition of the modulus of continuity $\omega_F(\cdot,\cdot)$ is somewhat similar to the definition of strong absolute differentiation which appears in \cite[expression (1)]{ChIn}: \noindent Let $(X,d_X)$ and $(Y,d_Y)$ be two metric spaces and consider $F:X\to Y$ and $p\in X$. We say $F$ is \emph{strongly absolutely differentiable at $p$} if and only if the following limit exists: \[F^{|\prime|}(p):=\lim_{\substack{(x,y)\to(p,p)\\ x\ne y}}\frac{d_Y(F(x),F(y))}{d_X(x,y)}.\] However, notice that there some important differences between $\omega_F(\cdot,\cdot)$ and $F^{|\prime|}$ when $X={\mathbb R}^n$ and $Y={\mathbb R}^m$. First, since $\omega(\cdot,\cdot)$ is defined with a supremmum, $\omega(\cdot,\cdot)$ is well defined in more cases than $F^{|\prime|}$. Also, in the definition of $\omega_F(\cdot,v)$, we are avoiding the direction of a certain vector $v$. This means that, while strong absolute differentiation implies continuity at the point (see \cite[Theorem 3.1]{ChIn}), $\omega(\cdot,\cdot)$ does not.\par Regarding the similarities, when the partial derivatives of $F$ exist, $F^{|\prime|}=\|\sum_{k=1}^n\frac{\partial F}{\partial x_k}\|$ (see \cite[Theorem 3.6]{ChIn}). \end{rem} \begin{exa}\label{examc}Consider again $F(x,y):=1+(y-x^2)^{\frac{2}{3}}$ and $S=\{(x,y)\in{\mathbb R}^2\ : y=x^2\}$. As was stated in Example \ref{exa-change}, we have that $F|_{{\mathbb R}^2\backslash S}\in{\mathcal C}^\infty({\mathbb R}^2\backslash S)$ and \[\nabla F(x,y)=\frac{2}{3}(y-x^2)^{-\frac{1}{3}}(-2x,1),\quad \mbox{for}\ (x,y)\in{\mathbb R}^2\backslash S.\] Therefore, $\omega(p,v)<+\infty$ for every $(p,v)\in ({\mathbb R}^2\backslash S)\times{\mathbb P}^1$.\par On the other hand, for $p=(x_0,x_0^2)\in S$ and $v=(v_1:v_2)\in{\mathbb P}^1$, if $x=(x_1,y_1)-p\perp v$ then $x=\l(-v_2,v_1)+p$ for some $\l\in {\mathbb R}$. Analogously, we take $y=\mu(-v_2,v_1)+p$ for some $\mu\in {\mathbb R}$. Hence, \begin{align*}\omega_F(p,v)= & \lils{\substack{(x,y)\to(p,p)\\x-p,y-p\perp v\\x\ne y}}\frac{\|F(x)-F(y)\|}{\|x-y\|}=\lils{\substack{(\l,\mu)\to(0,0)\\\l\ne \mu}}\frac{|F(\l(-v_2,v_1)+p)-F(\mu(-v_2,v_1)+p)|}{\|(\l-\mu)(-v_2,v_1)\|} \\= & \lils{\substack{(\l,\mu)\to(0,0)\\ \l\ne \mu}}\frac{|[\l(2x_0v_2+v_1)-\l^2v_2^2]^\frac{2}{3}-[\mu(2x_0v_2+v_1)-\mu^2v_2^2]^\frac{2}{3}|}{|\l-\mu|} \end{align*} We now can consider two cases: $(v_1:v_2)=(-2x_0:1)$ and $(v_1:v_2)\ne(-2x_0:1)$. In the first case, taking into account that $z^2+z+1\ge 3/4$ for every $z\in{\mathbb R}$, \begin{align*}\omega_F(p,v)= & \lils{\substack{(\l,\mu)\to(0,0)\\\l\ne \mu}} \frac{|(-\l^2v_2^2)^\frac{2}{3}-(-\mu^2v_2^2)^\frac{2}{3}|}{|\l-\mu|}=\lils{\substack{(\l,\mu)\to(0,0)\\\l\ne \mu}} \frac{|\mu^\frac{4}{3}-\l^\frac{4}{3}||v_2|^\frac{2}{3}}{|\l-\mu|} \\= & |v_2|^\frac{2}{3}\lils{\substack{(\l,\mu)\to(0,0)\\\l\ne \mu}} \left|\mu^\frac{1}{3}+\frac{\l}{\mu^\frac{2}{3}+\mu^\frac{1}{3}\l^\frac{1}{3}+\l^\frac{2}{3}}\right|=|v_2|^\frac{2}{3}\lils{\substack{(\l,\mu)\to(0,0)\\\l\ne \mu}} \left|\mu^\frac{1}{3}+\l^\frac{1}{3}\frac{1}{\(\frac{\mu}{\l}\)^\frac{2}{3}+\(\frac{\mu}{\l}\)^\frac{1}{3}+1}\right| \\ \le & |v_2|^\frac{2}{3}\lils{\substack{(\l,\mu)\to(0,0)\\\l\ne \mu}} \left|\mu^\frac{1}{3}+\frac{4}{3}\l^\frac{1}{3}\right|=0. \end{align*} Observe that in this deduction we have assumed $\l\ne0$. It is clear that, when $\l=0$, the limit is zero as well.\par In the case $(v_1:v_2)\ne(-2x_0:1)$ the quotient inside the limit is not bounded and $\omega_F(p,v)=+\infty$. Therefore, \[\omega_F^{-1}([0,+\infty))=({\mathbb R}^2\backslash S)\times{\mathbb P}^1\cup\{((x,x^2),(-2x:1))\in{\mathbb R}^2\times{\mathbb P}^1\ :\ x\in{\mathbb R}\}.\] \end{exa} \section{Sufficient conditions ensuring a Lipschitz condition along a foliation} The next Lemma is a key ingredient in the main result of this section. It gives an alternative expression to the rotation matrix provided by the Rodrigues' Rotation Formula and generalizes it for $n$-dimensional vector spaces. \begin{lem}[Codesido's Rotation Formula]\label{crf} Let $x,y\in{\mathbb R}^{n+1}$ and define $K_x^y\in{\mathcal M}_{n+1}({\mathbb R})$ as \[K_x^y :=yx^T-xy^T.\] Now, let $u,v\in {\mathbb S}^n$, $v\not= -u$, and define $R_u^v\in{\mathcal M}_{n+1}({\mathbb R})$ as \begin{equation}\label{Codfor}R_u^v:=\Id+K_u^v+\frac{1}{1+\<u,v\>}(K_u^v)^2,\end{equation} where $\Id$ is the identity matrix of order $n+1$. Then, $R_u^v\in \operatorname{SO}(n+1)$ and $R_u^vu=v$, that is, $R_u^v$ is a rotation in ${\mathbb R}^{n+1}$ that sends the unitary vector $u$ to $v$. Furthermore, the function $R:\{(u,v)\in{\mathbb S}^n\times{\mathbb S}^n\ :\ u\ne-v\}\to \operatorname{SO}(n+1)$, defined by $R(u,v):=R_u^v$, is analytic. \end{lem} \begin{proof} First, we show that $R_u^v\in\operatorname{O}(n+1)$, that is, $(R_u^v)^T=(R_u^v)^{-1}$. Observe that $(K_u^v)^T=-K_u^v$ and so $[(K_u^v)^2]^T=(K_u^v)^2$. That is, $(R_u^v)^T= \Id-K_u^v+\frac{1}{1+\<u,v\>}(K_u^v)^2$. Therefore, \begin{align*}(R_u^v)^TR_u^v= & \left[\Id-K_u^v+\frac{1}{1+\<u,v\>}(K_u^v)^2\right]\left[\Id+K_u^v+\frac{1}{1+\<u,v\>}(K_u^v)^2\right] \\ = & \Id+\frac{1-\<u,v\>}{1+\<u,v\>}(K_u^v)^2+\frac{1}{(1+\<u,v\>)^2}(K_u^v)^4.\end{align*} Now, \begin{align*}(K_u^v)^2= & (vu^T-uv^T)^2=vu^Tvu^T+uv^Tuv^T-vu^Tuv^T-uv^Tvu^T=\<u,v\>(vu^T+uv^T)-(vv^T+uu^T),\\ (K_u^v)^4= & \left[\<u,v\>(vu^T+uv^T)-(vv^T+uu^T)\right]^2=\(\<u,v\>^2-1\)(K_u^v)^2.\end{align*} Therefore, \begin{align*}(R_u^v)^TR_u^v= & \Id+\frac{1-\<u,v\>}{1+\<u,v\>}(K_u^v)^2-\frac{1-\<u,v\>^2}{(1+\<u,v\>)^2}(K_u^v)^2=\Id.\end{align*} Clearly, $R_u^v$ is analytic on $S=\{(u,v)\in{\mathbb S}^n\times{\mathbb S}^n\ :\ u\ne-v\}$ and so is the determinant function. Now, we are going to prove that $S$ is a connected set: firstly, define the linear subspaces \begin{displaymath}V_1:=\{z \in {\mathbb R}^{2n+2} : \ z_i=-z_{n+1+i}, \quad i=1,2,\ldots n+1\},\end{displaymath} \begin{displaymath}V_2:=\{z \in {\mathbb R}^{2n+2} : \ z_i=0, \quad i=1,2,\ldots n+1\},\end{displaymath} \begin{displaymath}V_3:=\{z \in {\mathbb R}^{2n+2} : \ z_{n+1+i}=0, \quad i=1,2,\ldots n+1\},\end{displaymath} and note that $\codim(V_i)=n+1\ge 2$ for all $i\in \{1,2,3\}$. Then, it is know that $X:={\mathbb R}^{n+1}\setminus (V_1 \cup V_2\cup V_3)$ is connected, see \cite[Chapter V, Problem 5]{Dieu69}, and since the projection $\pi: X \to S$ defined as \begin{displaymath}\pi(z)=\left(\frac{(z_1,z_2,\ldots,z_{n+1})}{\|(z_1z_2,\ldots,z_{n+1})\|},\frac{(z_{n+2},z_{n+3},\ldots,z_{2n+2})}{\|(z_{n+2},z_{n+3},\ldots,z_{2n+2})\|} \right),\end{displaymath} is continuous and onto, we have that $S$ is connected too. Therefore, $|R_u^v|$ is continuous on the connected set $S$ and takes values in $\{-1,1\}$, so $|R_u^v|$ is constant. Since $|R_u^u|=|\Id|=1$ we have that $|R_u^v|=1$ on $S$, that is, $R_u^v\in\operatorname{SO}(n+1)$.\par Last, observe that \begin{align*}R_u^vu & =u+(vu^T-uv^T)u+\frac{\<u,v\>(vu^T+uv^T)u-(vv^T+uu^T)u}{1+\<u,v\>} \\ &=u+v-uv^Tu+\frac{\<u,v\>(v+uv^Tu)-(vv^Tu+u)}{1+\<u,v\>} \\ & =v+\frac{\<u,v\>(v+uv^Tu+u-uv^Tu)-(vv^Tu+u)+u-uv^Tu}{1+\<u,v\>}=v+\frac{\<u,v\>(v+u)-vv^Tu-uv^Tu}{1+\<u,v\>}\\ &=v+\frac{\<u,v\>(v+u)-\<u,v\>v-\<u,v\>u}{1+\<u,v\>}=v.\end{align*} \end{proof} \begin{rem} For $n=1$ the function $R$ admits a continuous extension to ${\mathbb S}^1\times{\mathbb S}^1$. Indeed, let us consider $u,v\in {\mathbb S}^1$, $v\not=-u$. Then $u=(\cos(\a),\sin(\a))$ and $v=(\cos(\b),\sin(\b))$ for some $\a,\b \in {\mathbb R}$, with $\b\not= \a+(2k+1)\pi$, $k\in {\mathbb Z}$. Now, a direct computation shows that \begin{displaymath}R_{u}^v=\left(\begin{array}{ccc} \cos(\a-\b) & \sin(\a-\b) \\ -\sin(\a-\b) & \cos(\a-\b) \end{array}\right).\end{displaymath} Therefore, \begin{displaymath}\lim_{v\to -u} R_{u}^v =\lim_{\beta \to \alpha+\pi} \left(\begin{array}{ccc} \cos(\a-\b) & \sin(\a-\b) \\ -\sin(\a-\b) & \cos(\a-\b) \end{array}\right)= \left(\begin{array}{cc}-1 & 0 \\0 & -1\end{array}\right).\end{displaymath} However, for $n\ge 2$ the function $R$ does not admit a continuous extension to ${\mathbb S}^n\times{\mathbb S}^n$. To see this, consider $u\in{\mathbb S}^n$, $w\in{\mathbb R}^{n+1}$, $w\perp u$, $w\not=0$ and define $v(w)=(w-u)/\|w-u\|$. Observe that $v(w)\in {\mathbb S}^n$, $v(w)\not= -u$, $\displaystyle\lim_{\|w\|\to 0} v(w)=-u$ and \[K_u^{v(w)}=\frac{1}{\|w-u\|}K_u^w.\] Hence, \begin{align*}R_u^{v(w)} & =\Id+\frac{1}{\|w-u\|}K_u^w+\frac{\| w-u\|}{\| w-u\|+\<u,w\>-1}\frac{1}{\| w-u\|^2}(K_u^w)^2 \\ & =\Id+\frac{1}{\|w-u\|}K_u^w+\frac{-ww^T-\|w\|^2uu^T}{\| w-u\|(\| w-u\|-1)}. \end{align*} Now, consider $\bar{w}\perp u$ with $\|\bar{w}\|=1$.Therefore, if it exists, \[\lim_{v \to -u}R_u^{v}=\lim_{t \to 0}R_u^{ v(t \bar{w})}=\Id+\lim_{t \to0}\frac{-t^2 (\bar{w}\bar{w}^T-uu^T)}{\sqrt{t^2+1}(\sqrt{t^2+1}-1)}=\Id-2 (\bar{w}\bar{w}^T-uu^T).\] But in ${\mathbb R}^{n+1}$, with $n\ge 2$, there exist at least two independent unitary vectors $\bar{w}_1$ and $\bar{w}_2$ in $\<u\>^{\perp}$, each of them leading to a different value of the right-hand side of the previous expression. Hence, the $\displaystyle \lim_{v \to -u}R_u^{v}$ does not exist and thus $R$ can not be continuously extended to ${\mathbb S}^n\times{\mathbb S}^n$. \end{rem} The following is the main result in this section and gives sufficient conditions for the existence of a $n$-foliation which allows $F$ to satisfy condition (C2) in Theorem \ref{thmgen}. \begin{thm}\label{mainthm}Let $U$ be an open subset of ${\mathbb R}^{n+1}$, $p_0\in U$ and $F:U\to{\mathbb R}^{n+1}$ continuous. Assume there exists an open interval $J$ with $0\in J$ and a simple path $\c=(\c_1,\c_2)\in{\mathcal C}^1(J, U\times{\mathbb P}^n)$ such that the following conditions hold: \begin{itemize} \item[(i)] $\c_1(0)=p_0.$ \item[(ii)] There exist $\d,M\in{\mathbb R}^+$, such that $\omega_F(\c_1(t),\c_2(t),\d)<M$ for all $t\in J$. \item[(iii)] $\c_1'(0) \not\perp \c_2(0)$. \end{itemize} Then, there exists an open neighborhood of zero $\hat{U}\subset U\subset {\mathbb R}^{n+1}$ such that $\Phi(s,y)$ is a local $n$-foliation of $\hat{U}$. Moreover, $F\circ \Phi$ and $(\Phi')^{-1}$ are Lipschitz in a neighborhood of zero when fixing the first variable. \end{thm} \begin{proof} Assume, without loss of generality, that $\c_1$ is parameterized by arc length, that is, $\|\c_1'(t)\|=1$ for all $t\in J$. Consider ${\mathbb S}^n$ as covering space of ${\mathbb P}^n$ with the usual projection $\pi:{\mathbb S}^n\to{\mathbb P}^n$. Take $v_0\in\pi^{-1}(\c_2(0))$, such that $v_0\not= -e_1$ where $e_1=(1,0,\dots,0)\in{\mathbb R}^{n+1}$, and consider the lift $\tilde\c=(\c_1,\tilde\c_2):J\to V\times{\mathbb S}^n$ of $\c$ such that $\tilde \c(0)=(p_0,v_0)$. Now, $\tilde\c_2$ is continuous, and $\<e_1,\tilde \c_2(0)\>=\<e_1,v_0\>\not=-1$ so we can consider an open interval $\tilde J \subset J$ where $\<e_1,\tilde \c_2(s)\>\ne-1$ (that is, $\tilde \c_2(s)\ne-e_1$) for $s\in\tilde J$. Since $\tilde \c$ is differentiable and $\|\tilde \c_2(s)\|=1$ for every $s\in \tilde J$, we can consider the continuously differentiable function \begin{center}\begin{tikzcd}[row sep=tiny] \tilde J \arrow{r}{A} & \operatorname{SO}(n+1)\\ s \arrow[mapsto]{r} & A(s):=R_{e_1}^{\tilde\c_2(s)} \end{tikzcd} \end{center} where $R_u^v$ is defined as in Lemma \ref{crf}. Observe that denoting by $a_j(s)$ the columns of $A(s)$, that is, \begin{displaymath}A(s)=\left(\begin{array}{c|c|c|c}a_1(s) & a_2(s) & \ldots & a_{n+1}(s)\end{array}\right),\end{displaymath} we have that $a_1(s)=\tilde\c_2(s)$ and $\{ a_2(s),a_3(s),\ldots, a_{n+1}(s)\}$ is an orthonormal basis of $\tilde\c_2(s)^{\perp}$, (remember that $A(s) e_1=\tilde\c_2(s)$ and that $A(s)$ is an orthogonal matrix). Now, we can define the differentiable function $\Phi: \tilde J\times {\mathbb R}^{n}\to {\mathbb R}^{n+1}$ given by \begin{displaymath}\Phi(s,y):=\c_1(s)+A(s)(0,y).\end{displaymath} \noindent {\it Claim 1. $g_s(y):=\Phi(s,y)$ is a local $n$-foliation.} We easily compute \begin{displaymath}\frac{\partial \Phi}{\partial s}(s,y)=\c_1'(s)+A'(s) (0,y),\end{displaymath} \begin{displaymath}\frac{\partial \Phi}{\partial y}(s,y)=\left(\begin{array}{c|c|c|c}a_2(s) & a_3(s) & \ldots & a_{n+1}(s)\end{array}\right).\end{displaymath} So \begin{displaymath}\Phi'(0,0)=\left(\begin{array}{c|c|c|c|c}\c_1'(0) & a_2(0) & a_3(0) & \ldots & a_{n+1}(0)\end{array}\right),\end{displaymath} and since, by (iii), $\c_1'(0)\not\perp \tilde \c_2(0)=a_1(0)$ we have \begin{displaymath}J_{\Phi}(0,0)=|\Phi'(0,0)|\not=0.\end{displaymath} Then, by the inverse function theorem there exist open sets $\hat{J}\subset \tilde J ,\hat{V}\subset V$ and $\hat{U}\subset U$ such that $\hat{J}\times\hat{V}$ contains the origin and $\Phi:\hat{J}\times\hat{V}\to\hat{U}$ is a diffeomorphism. Moreover, by (i), $ \Phi(0,0)=p_0$, so $\Phi(s,y)$, a local $n$-foliation of $\hat{U}$. \medbreak \noindent {\it Claim 2. $F\circ \Phi$ is Lipschitz continuous in a neighborhood of zero when fixing the first variable.} Notice that, by construction, $\Phi(s,y)- \c_1(s)\in\<\tilde\c_2(s)\>^{\perp}$. Now, condition (ii) implies that \begin{align*} & \|F\circ\Phi(s,y_1)-F\circ\Phi(s,y_2)\| = \|F(\c_1(s)+A(s)(0,y_1))-F(\c_1(s)+A(s)(0, y_2))\| \\ \le & \omega_F(\c_1(s),\c_2(s),\d)\|\c_1(s)+A(s)(0,y_1)-\c_1(s)+A(s)(0, y_2)\|\le M \sup_{s\in \hat{J}}\|A(s)\| \|y_1-y_2\|.\end{align*} for every $s\in \hat{J}$ and $y_1,y_2\in B_{n}\left(0,\displaystyle\frac{\d}{\sup_{s\in \hat{J}}\|A(s)\|}\right)$.\par \noindent {\it Claim 3. $(\Phi')^{-1}$ is Lipschitz continuous in a neighborhood of zero when fixing the first variable.}\par Fix $s\in \hat{J}$. We have that \[\Phi'(s,y)=\left(\begin{array}{c|c|c|c|c}\c_1'(s)+A'(s)(0,y) & a_2(s) & a_3(s) & \ldots & a_{n+1}(s)\end{array}\right).\] Then, \[\|\Phi'(s,x)-\Phi'(s, y)\|\le \sup_{s\in\hat{J}}\|A(s)\| \|x- y\|,\] so $\Phi'$ is Lipschitz continuous in a neighborhood of zero when fixing $s$. On the other hand, $(\Phi')^{-1}$ is a continuous function, therefore locally bounded. Hence, by Corollary \ref{coreq}.3, $(\Phi')^{-1}$ is Lipschitz continuous in a neighborhood of zero when fixing the first variable. \end{proof} \renewcommand{\abstractname}{Acknowledgments} \begin{abstract}The authors want to express their gratitude towards Prof. Santiago Codesido (Universit\'e de Gen\`eve, Switzerland) for suggesting the rotation matrix expression \eqref{Codfor} and several useful discussions. \end{abstract}

https://arxiv.org/abs/1707.00857

Existence results for a linear equation with reflection, non-constant coefficient and periodic boundary conditions

This work is devoted to the study of first order linear problems with involution and periodic boundary value conditions. We first prove a correspondence between a large set of such problems with different involutions to later focus our attention to the case of the reflection. We study then different cases for which a Green's function can be obtained explicitly and derive several results in order to obtain information about its sign. Once the sign is known, maximum and anti-maximum principles follow. We end this work with more general existence and uniqueness of solution results.

\section{Introduction} In a previous paper by the authors \cite{Cab4}, a Green's function for the following linear problem with reflection was found. \begin{equation}\label{eqoriginal}x'(t)+\omega x(-t)=h(t), t\in I;\quad x(-T)=x(T),\end{equation} where $T\in{\mathbb R}^+$, $\omega\in{\mathbb R}\backslash\{0\}$ and $h\in L^1(I)$, with $I=[-T,T]$. The precise form of this Green's function was given by the following theorem. \begin{thm}[\cite{Cab4}, Proposition 3.2]\label{Greenf} Suppose that $\omega \neq k \, \pi/T$, $k \in {\mathbb Z}$. Then problem (\ref{eqoriginal}) has a unique solution given by the expression \begin{equation} \label{e-u} u(t):=\int_{-T}^T\overline G(t,s)h(s)\dif s, \end{equation} where $$\overline{G}(t,s):=\omega\,G(t,-s)-\frac{\partial G}{\partial s}(t,s)$$ and $G$ is the Green's function for the harmonic oscillator $$x''(t)+\omega^2x(t)=0;\ x(T)=x(-T),\ x'(T)=x'(-T).$$ \end{thm} The sign properties of this Green's function were further studied in \cite{Cab5}, where the methods similar to those found in \cite{gijwjiea, gijwjmaa, gijwems} are used in order to derive existence and multiplicity results.\par Still, obtaining the Green's function of problem \eqref{eqoriginal} with a non-constant coefficient has not been accomplished yet. In this article we will study this case and further generalize the existence of Green's functions. Through a correspondence theorem, we will also be able to extend these results to problems with other involutions. We will also obtain new maximum and anti-maximum principles and existence and uniqueness results. \section{Order one linear problems with involutions} Assume $\phi$ is a differentiable involution on $[\phi(T),T]$. Let $a,b,c,d\in L^1([\phi(T),T])$ and consider the following problem \begin{equation}\label{proinv1} d(t)x'(t)+c(t)x'(\phi(t))+b(t)x(t)+a(t)x(\phi(t))=h(t),\ x(\phi(T))=x(T). \end{equation} It would be interesting to know under what circumstances problem \eqref{proinv1} is equivalent to another problem of the same kind but with a different involution, in particular the reflection. The following two results will help us to clarify this situation. \begin{lem}[\textsc{Correspondence of Involutions}]\label{corofinv} Let $\phi$ and $\psi$ be two differentiable involutions\footnote{Every differentiable involution is a diffeomorphism.} on the intervals $[\phi(T),T]$ and $[\psi(S),S]$ respectively. Let $t_0$ and $s_0$ be the unique fixed points of $\phi$ and $\psi$ respectively. Then, there exists an orientation preserving diffeomorphism $f:[\psi(S),S]\to[\phi(T),T]$ such that $f(\psi(s))=\phi(f(s))\nkp\fa s\in[\psi(S),S]$. \end{lem} \begin{proof} Let $g:[\psi(S),s_0]\to[\phi(T),t_0]$ be an orientation preserving diffeomorphism, that is, $g(s_0)=t_0$. Let us define $$f(s):=\begin{cases} g(s) & \text{ if } s\in[\psi(S),s_0], \\ (\phi\circ g\circ\psi)(s) & \text{ if } s\in(s_0,S].\end{cases}$$ \par Clearly, $f(\psi(s))=\phi(f(s))\nkp\fa s\in[\psi(S),S]$. Since $s_0$ is a fixed point for $\psi$, $f$ is continuous. Furthermore, because $\phi$ and $\psi$ are involutions, $\phi'(t_0)=\psi'(s_0)=-1$, so $f$ is differentiable. $f$ is invertible with inverse $$f^{-1}(t):=\begin{cases} g^{-1}(t) & \text{ if } t\in[\phi(T),t_0], \\ (\psi\circ g^{-1}\circ\phi)(t) & \text{ if } t\in(t_0,T].\end{cases}$$ $f^{-1}$ is also differentiable for the same reasons. \end{proof} \begin{rem}A similar argument could be done in the case of involutions defined on open, possibly not bounded, intervals. \end{rem} \begin{rem} The expression obtained for $f$ reminds us of the characterization of involutions given in \cite[Property 6]{Wie}. \end{rem} \begin{rem} It is easy to check that if $\phi$ is an involution defined on ${\mathbb R}$ with fixed point $t_0$ then $\psi(t):=\phi(t+t_0-s_0)-t_0+s_0$ is an involution defined on ${\mathbb R}$ with fixed point $s_0$ (cf. \cite[Property 2]{Wie}). For this particular choice of $\phi$ and $\psi$, we can take $g(s)=s-s_0+t_0$ in Lemma \ref{corofinv} and, in such a case, $f(s)=s-s_0+t_0$ for all $s\in{\mathbb R}$. \end{rem} \begin{cor}[\textsc{Change of Involution}] Under the hypothesis of Lemma \ref{corofinv}, problem \eqref{proinv1} is equivalent to \begin{equation}\label{proinv2} \frac{d(f(s))}{f'(s)}y'(s)+\frac{c(f(s))}{f'(\psi(s))}y'(\psi(s))+b(f(s))y(s)+a(f(s))y(\psi(s))=h(f(s)),\ y(\psi(S))=y(S). \end{equation} \end{cor} \begin{proof}Consider the change of variable $t=f(s)$ and $y(s):=x(t)=x(f(s))$. Then, using Lemma \ref{corofinv}, it is clear that $$\frac{\dif y}{\dif s}(s)=\frac{\dif x}{\dif t}(f(s))\frac{\dif f}{\dif s}(s)\quad\text{and}\quad \frac{\dif y}{\dif s}(\psi(s))=\frac{\dif x}{\dif t}(\phi(f(s)))\frac{\dif f}{\dif s}(\psi(s)).$$ Making the proper substitutions in problem \eqref{proinv1} we get problem \eqref{proinv2} and vice-versa. \end{proof} This last results allows us to restrict our study of problem \eqref{proinv1} to the case where $\phi$ is the reflection $\phi(t)=-t$. In the following section we will further restrict our assumptions to the case where $c\equiv0$ in problem \eqref{proinv1}. A comment on how to proceed without this assumption will be done in the Appendix at the end of this work. \section{Study of the homogeneous equation} In this section we will study some different cases for the homogeneous equation \begin{equation}\label{gen-eq} x'(t)+a(t)x(-t)+b(t)x(t)=0,\ t\in I,\end{equation} where $a,b\in L^1(I)$. In order to solve it, we can consider the decomposition of equation \eqref{gen-eq} used in \cite{Cab4}. For any given function $f$, let $f_e(x):=\frac{f(x)+f(-x)}{2}$ be its even part and $f_o(x):=\frac{f(x)-f(-x)}{2}$ its odd part. Then, the solutions of equation \eqref{gen-eq} satisfy \begin{align}\label{eq3.1}\begin{pmatrix}x_o' \\ x_e'\end{pmatrix} & =\begin{pmatrix}a_o-b_o & -a_e-b_e \\ a_e-b_e & -a_o-b_o\end{pmatrix}\begin{pmatrix}x_o \\ x_e\end{pmatrix}. \end{align} Realize that, a priori, solutions of system \eqref{eq3.1} need not to be pairs of even and odd functions, nor provide solutions of \eqref{gen-eq}.\par In order to solve this system, we will restrict problem \eqref{eq3.1} to those cases where the matrix $$M(t)=\begin{pmatrix}a_o-b_o & -a_e-b_e \\ a_e-b_e & -a_o-b_o\end{pmatrix}(t)$$ satisfies that $[M(t),M(s)]:=M(t)M(s)-M(s)M(t)=0\nkp\fa t,s\in I$, for in that case the solution of the system \eqref{eq3.1} is given by the exponential of the integral of $M$\footnote{See the Appendix for more details on this matter.}. Clearly, $$[M(t),M(s)]=2 \begin{pmatrix} a_e(t)b_e(s)-a_e(s)b_e(t) & a_o(s)[a_e(t)+b_e(t)]-a_o(t)[a_e(s)+b_e(s)]\\a_o(t)[a_e(s)+b_e(s)]-a_o(s)[a_e(t)+b_e(t)] & a_e(s)b_e(t)-a_e(t)b_e(s)\end{pmatrix}.$$ Let $A(t):=\int_0^t a(s)\dif s$, $B(t):=\int_0^t b(s)\dif s$. Let $\overline M$ be a primitive (save possibly a constant matrix) of $M$. We study now the different cases where $[M(t),M(s)]=0\nkp\fa t,s\in I$. We will always assume $a\not\equiv0$, since the case $a\equiv0$ is the well-known case of an ODE.\par \textbf{(C1). $b_e=k\,a,\ k\in{\mathbb R},\ |k|<1$.} \par In this case, $a_o=0$ and $\overline M$ has the form $$\overline M=\begin{pmatrix}B_e & -(1+k)A_o\\ (1-k)A_o & -B_e\end{pmatrix}.$$ If we compute the exponential (see note in the Appendix for more information) we get $$e^{\overline M(t)}=e^{-B_e(t)}\begin{pmatrix}\cos\(\sqrt{1-k^2}A(t)\) & -\frac{1+k}{\sqrt{1-k^2}}\sin\(\sqrt{1-k^2}A(t)\)\\ \frac{\sqrt{1-k^2}}{1+k}\sin\(\sqrt{1-k^2}A(t)\) & \cos\(\sqrt{1-k^2}A(t)\)\end{pmatrix}.$$ Therefore, if a solution to equation \eqref{gen-eq} exists, it has to be of the form $$u(t)=\a e^{-B_e(t)}\cos\(\sqrt{1-k^2}A(t)\)+\b e^{-B_e(t)}\frac{1+k}{\sqrt{1-k^2}}\sin\(\sqrt{1-k^2}A(t)\).$$ with $\a$, $\b\in{\mathbb R}$. It is easy to check that all the solutions of equation \eqref{gen-eq} are of this form with $\b=-\a$. \par \textbf{(C2). $b_e=k\,a,\ k\in{\mathbb R},\ |k|>1$.} This case is much similar to (C1) and it yields solutions of system \eqref{eq3.1} of the form $$u(t)=\a e^{-B_e(t)}\cosh\(\sqrt{k^2-1}A(t)\)+\b e^{-B_e(t)}\frac{1+k}{\sqrt{k^2-1}}\sinh\(\sqrt{k^2-1}A(t)\),$$ which are solutions of equation \eqref{gen-eq} when $\b=-\a$.\par \textbf{(C3). $b_e=a$.} In this case the solutions of system \eqref{eq3.1} are of the form \begin{equation}\label{eqc3}u(t)=\a e^{-B_e(t)}+2\b e^{-B_e(t)}A(t)\end{equation} which are solutions of equation \eqref{gen-eq} when $\b=-\a$.\par \textbf{(C4). $b_e=-a$.} In this case the solutions of system \eqref{eq3.1} are the same as in case (C3), but they are solutions of equation \eqref{gen-eq} when $\b=0$.\par \textbf{(C5). $b_e=a_e=0$.} In this case the solutions of system \eqref{eq3.1} are of the form $$u(t)=\a e^{A(t)-B(t)}+\b e^{-A(t)-B(t)},$$ which are solutions of equation \eqref{gen-eq} when $\a=0$.\par \section{The cases (C1)--(C3) for the complete problem} In the more complicated setting of the following nonhomogeneous problem \begin{equation}\label{eq2cp} x'(t)+a(t)\,x(-t)+b(t)\,x(t)=h(t),\enskip a.\,e. t\in I,\quad x(-T)=x(T), \end{equation} we have still that, in the cases (C1)--(C3), it can be sorted out very easily. In fact, we get the expression of the Green's function for the operator. We remark that in the three considered cases along this section the function $a$ must be even on $I$. We note also that $a$ is allowed to change its sign on $I$. \par First, we are going to prove a generalization of Theorem \ref{Greenf}.\par Consider problem \eqref{eq2cp} with $a$ and $b$ constants. \begin{equation}\label{eq2cp2} x'(t)+a\,x(-t)+b\,x(t)=h(t),\enskip t\in I,\quad x(-T)=x(T). \end{equation} Considering the homogeneous case ($h=0$), differentiating and making proper substitutions, we arrive to the problem. \begin{equation}\label{eqhog} x''(t)+(a^2-b^2)x(t)=0,\enskip t\in I,\quad x(-T)=x(T),\quad x'(-T)=x'(T). \end{equation} Which, for $b^2<a^2$, is the problem of the harmonic oscillator. It was shown in \cite[Proposition 3.1]{Cab4} that, under uniqueness conditions, the Green's function $G$ for problem \eqref{eqhog} satisfies the following properties in the case $b^2<a^2$, but they can be extended almost automatically to the case $b^2>a^2$. \begin{lem} The Green's function $G$ satisfies the following properties. \begin{enumerate} \item $G\in{\mathcal C}(I^2,{\mathbb R})$, \item $\frac{\partial G}{\partial t}$ and $\frac{\partial^2 G}{\partial t^2}$ exist and are continuous in $\{(t,s)\in I^2\ |\ s\ne t\}$, \item $\frac{\partial G}{\partial t}(t,t^-)$ and $\frac{\partial G}{\partial t}(t,t^+)$ exist for all $t\in I$ and satisfy $$\frac{\partial G}{\partial t}(t,t^-)-\frac{\partial G}{\partial t}(t,t^+)=1\nkp\fa t\in I,$$ \item $\frac{\partial^2 G}{\partial t^2}+(a^2-b^2)G=0\text{ in }\{(t,s)\in I^2\ |\ s\ne t\},$ \item \begin{enumerate} \item $G(T,s)=G(-T,s)\nkp\fa s\in I$, \item $\frac{\partial G}{\partial t}(T,s)=\frac{\partial G}{\partial t}(-T,s)\nkp\fa s\in(-T,T)$. \end{enumerate} \item $G(t,s)=G(s,t)$, \item $G(t,s)=G(-t,-s)$, \item $\frac{\partial G}{\partial t}(t,s)=\frac{\partial G}{\partial s}(s,t)$, \item $\frac{\partial G}{\partial t}(t,s)=-\frac{\partial G}{\partial t}(-t,-s)$, \item $\frac{\partial G}{\partial t}(t,s)=-\frac{\partial G}{\partial s}(t,s)$. \end{enumerate} \end{lem} With these properties, we can prove the following Theorem (cf. \cite[Proposition 3.2]{Cab4}). \begin{thm}\label{Greenf2} Suppose that $a^2-b^2 \neq n^2 \, (\pi/T)^2$, $n=0,1,\dots$ Then problem \eqref{eq2cp2} has a unique solution given by the expression \begin{equation*} \label{e-u2} u(t):=\int_{-T}^T\overline G(t,s)h(s)\dif s, \end{equation*} where \begin{equation}\label{e-a-b} \overline{G}(t,s):=a\,G(t,-s)-b\,G(t,s)+\frac{\partial G}{\partial t}(t,s)\end{equation} is called the \textbf{Green's function} related to problem \eqref{eq2cp2}. \end{thm} \begin{proof} Since problem \eqref{eq2cp2}, in the homogeneous case, can be reduced to a problem with the equation of problem \eqref{eqhog}, the classical theory of ODE tells us that problem \eqref{eq2cp2} has at most one solution for all $a^2-b^2 \neq n^2 \, (\pi/T)^2$, $n=0,1,\dots$ Let us see that function $u$ defined in (\ref{e-u}), with $\overline G$ given by \eqref{e-a-b}, fulfills (\ref{eq2cp2}): \begin{eqnarray*} & & u'(t)+a\, u(-t)+b\, u(t)=\frac{\dif}{\dif t}\int_{-T}^{-t}\overline G(t,s)h(s)\dif s+\frac{\dif}{\dif t}\int_{-t}^t\overline G(t,s)h(s)\dif s+\frac{\dif}{\dif t}\int_{t}^T\overline G(t,s)h(s)\dif s\\ &+&a\int_{-T}^T\overline G(-t,s)h(s)\dif s+b\int_{-T}^T\overline G(t,s)h(s)\dif s\\ &=&(\overline G(t,t^-)-\overline G(t,t^+))h(t)+\int_{-T}^T\left[a\frac{\partial G}{\partial t}(t,-s)-b\frac{\partial G}{\partial t}(t,s)+\frac{\partial^2 G}{\partial t^2}(t,s)\right]h(s)\dif s\\ &+&a\int_{-T}^T\left[a\,G(-t,-s)-b\,G(-t,s)+\frac{\partial G}{\partial t}(-t,s)\]h(s)\dif s +b\int_{-T}^T\left[a\,G(t,-s)-b\,G(t,s)+\frac{\partial G}{\partial t}(t,s)\]h(s)\dif s. \end{eqnarray*} Using properties $(I)-(X)$, we deduce that this last expression is equal to $h(t)$, so the equation in problem \eqref{eq2cp2} is satisfied. Property $(V)$ allows us to verify the boundary conditions. $$ u(T)-u(-T)=$$ $$\int_{-T}^T\left[a\,G(T,-s)-b\,G(T,s)+\frac{\partial G}{\partial t}(T,s)-a\,G(-T,-s)+b\,G(-T,s)-\frac{\partial G}{\partial t}(-T,s)\right]h(s)\dif s=0.$$ \end{proof} This last theorem leads us to the question ``Which is the Green's function for the case (C3) with $a,b$ constants?". The following Lemma answers that question. \begin{lem}\label{lemGc3}Let $a\ne 0$ be a constant and let $G_{C3}$ be a real function defined as $$G_{C3}(t,s):=\frac{t-s}{2}-a\,s\,t+\begin{cases} -\frac{1}{2}+a\,s & \text{ if } |s|<t, \\ \frac{1}{2}-a\,s & \text{ if } |s|<-t, \\ \frac{1}{2}+a\,t & \text{ if } |t|<s, \\ -\frac{1}{2}-a\,t & \text{ if } |t|<-s.\end{cases}$$ Then the following properties hold. \begin{itemize} \item $\frac{\partial G_{C3}}{\partial t}(t,s)+a(G_{C3}(t,s)+G_{C3}(-t,s))=0$ for a.\,e. $t,s\in (-1,1)$. \item $\frac{\partial G_{C3}}{\partial t}(t,t^+)-\frac{\partial G_{C3}}{\partial t}(t,t^-)=1\nkp\fa t\in(-1,1)$. \item $G_{C3}(-1,s)=G_{C3}(1,s)\nkp\fa s\in(-1,1)$. \end{itemize} \end{lem} These properties are straightforward to check.\ Clearly, $G_{C3}$ is the Green's function for the problem $$x'(t)+a[x(t)+x(-t)]=h(t), t\in[-1,1];\quad x(1)=x(-1),$$ that is, the Green's function for the case (C3) with $a,b$ constants and $T=1$. For other values of $T$, it is enough to make a change of variables. \begin{rem} The function $G_{C3}$ can be obtained from the Green's functions for the case $(C1)$ with $a$ constant, $b_o\equiv0$ and $T=1$ taking the limit $k\to 1^-$ for $T=1$. \end{rem} The following theorem shows how to obtain a Green's function for non constant coefficients of the equation using the Green's function for constant coefficients. We can find the same principle, that is, to compose a Green's function with some other function in order to obtain a new Green's function, in \cite[Theorem 5.1, Remark 5.1]{Cab6} and also in \cite[Section 2]{Gau}.\par But first, we need to now hot the Green's function should be defined in such a case. Theorem \ref{Greenf2} gives us the expression of the Green's function for problem \eqref{eq2cp2}, $\overline{G}(t,s):=a\,G(t,-s)-b\,G(t,s)+\frac{\partial G}{\partial t}(t,s)$. For instance, in the case (C1), if $\omega=\sqrt{a^2-b^2}$, $$2\omega\sin(\omega T)\overline{G}(t,s):=\begin{cases} a \cos[\omega (s + t - T)]+ b \cos[\omega (s - t + T)] + \omega \sin[\omega (s - t + T)],& t>|s|,\\ a\cos[\omega (s + t - T)] +b \cos[\omega (-s + t + T)] - \omega \sin[\omega (-s + t + T)], & s>|t|,\\ a \cos[\omega (s + t + T)] +b \cos[\omega (-s + t + T)] - \omega \sin[\omega (-s + t + T)], & -t>|s|,\\ a \cos[\omega (s + t + T)] +b \cos[\omega (s - t + T)] + \omega \sin[\omega (s - t + T)], & -s>|t|. \end{cases}$$ Also, observe that $\overline G$ is continuous except at the diagonal, where $\overline G(t,t^-)-\overline G(t,t^+)=1$. Similarly, we can obtain the explicit expression of the Green's function $\overline G$ for the cases (C2) and (C3) (see Lemma \ref{lemGc3}). In any case, we have that the Green's function for problem \eqref{eq2cp2} can be expressed as $$2\omega\sin(\omega T)\overline{G}(t,s):=\begin{cases} \overline G_1(t,s),& t>|s|,\\ \overline G_2(t,s), & s>|t|,\\ \overline G_3(t,s), & -t>|s|,\\ \overline G_4(t,s), & -s>|t|, \end{cases}$$ were the $\overline G_j$, $j=1,\dots,4$ are analytic functions defined on ${\mathbb R}^2$. In order to simplify the statement of the following Theorem, consider the following conditions.\par $\mathbf{(C1^*)}$. (C1) is satisfied, $(1-k^2)A(T)^2\neq (n \, \pi)^2$ for all $n=0,1,\dots$ and $\cos\(\sqrt{1-k^2}A(T)\)\ne0$.\par $\mathbf{(C2^*)}$. (C2) is satisfied and $(1-k^2)A(T)^2\neq (n \, \pi)^2$ for all $n=0,1,\dots$\par $\mathbf{(C3^*)}$. (C3) is satisfied and $A(T)\ne0$.\par Assume one of $(C1^*)$--$(C3^*)$. In that case, by Theorem \ref{Greenf2} and Lemma \ref{lemGc3}, we are under uniqueness conditions for the solution for the following problem \cite{Cab4}. \begin{equation}\label{eq2} x'(t)+x(-t)+k\,x(t)=h(t),\enskip t\in [-|A(T)|,|A(T)|],\quad x(A(T))=x(-A(T)). \end{equation} The Green's function $G_2$ for problem \eqref{eq2} is just an specific case of $\overline G$ and can be expressed as $$\overline{G_2}(t,s):=\begin{cases} k_1(t,s),& t>|s|,\\ k_2(t,s), & s>|t|,\\ k_3(t,s), & -t>|s|,\\ k_4(t,s), & -s>|t|. \end{cases}$$ Define now \begin{equation}\label{Ggenral}G_1(t,s):=e^{B_e(s)-B_e(t)}H(t,s)=e^{B_e(s)-B_e(t)}\begin{cases} k_1(A(t),A(s)),& t>|s|,\\ k_2(A(t),A(s)), & s>|t|,\\ k_3(A(t),A(s)), & -t>|s|,\\ k_4(A(t),A(s)), & -s>|t|. \end{cases}\end{equation} Defined this way, $G_1$ is continuous except at the diagonal, where $G_1(t,t^-)-\overline G_1(t,t^+)=1$. Now we can state the following Theorem. \begin{thm}\label{thmcases123} Assume one of $(C1^*)$--$(C2^*)$. Let $G_1$ be defined as in \eqref{Ggenral}. Assume $G_1(t,\cdot)h(\cdot)\in L^1(I)$ for every $t\in I$. Then problem \eqref{eq2cp} has a unique solution given by $$u(t)=\int_{-T}^TG_1(t,s)h(s)\dif s.$$ \end{thm} \begin{proof} First realize that, since $a$ is even, $A$ is odd, so $A(-t)=-A(t)$. It is important to note that if $a$ has not constant sign in $I$, then $A$ may be not injective on $I$. From the properties of $\bar G_2$ as a Green's function, it is clear that $$\frac{\partial \bar G_2}{\partial t}(t,s)+\bar G_2(-t,s)+k\,\bar G_2(t,s)=0\quad\text{for a.\,e. }t,s\in A(I),$$ and so, $$\frac{\partial H}{\partial t}(t,s)+a(t)H(-t,s)+ka(t)\,H(t,s)=0\quad\text{for a.\,e. }t,s\in I,$$ Hence \begin{align*} & u' (t)+a(t)\, u(-t)+(b_o(t)+k\,a(t))\, u(t)= \frac{\dif}{\dif t}\int_{-T}^TG_1(t,s)h(s)\dif s+a(t)\int_{-T}^TG_1(-t,s)h(s)\dif s\\ &\quad+(b_o(t)+k\,a(t))\int_{-T}^TG_1(t,s)h(s)\dif s\\ =\ &\frac{\dif}{\dif t}\int_{-T}^{t} e^{B_e(s)-B_e(t)}H(t,s)h(s)\dif s+\frac{\dif}{\dif t}\int_{t}^T e^{B_e(s)-B_e(t)} H(t,s)h(s)\dif s\\&\quad+ a(t)\int_{-T}^T e^{B_e(s)-B_e(t)}H(-t,s)h(s)\dif s+(b_o(t)+k\,a(t))\int_{-T}^T e^{B_e(s)-B_e(t)}H(t,s)h(s)\dif s\\ =\ &[H(t,t^-)-H(t,t^+)]h(t)+a(t)\, e^{-B_e(t)}\int_{-T}^Te^{B_e(s)}\frac{\partial H}{\partial t}(t,s)h(s)\dif s\\&\quad- b_o(t)e^{-B_e(t)}\int_{-T}^Te^{B_e(s)}H(t,s)h(s)\dif s+ a(t)e^{-B_e(t)}\int_{-T}^Te^{B_e(s)} H(-t,s)h(s)\dif s\\ &\quad+ (b_o(t)+k\,a(t))e^{-B_e(t)}\int_{-T}^Te^{B_e(s)}H(t,s)h(s)\dif s\\ =\ & h(t)+\, a(t)e^{-B_e(t)}\int_{-T}^Te^{B_e(s)}\[\frac{\partial H}{\partial t}(t,s)+a(t)H(-t,s)+ka(t)\,H(t,s)\]h(s)\dif s=h(t). \end{align*} The boundary conditions are also satisfied. $$u(T)-u(-T)= e^{-B_e(T)}\int_{-T}^Te^{B_e(s)} [H(T,s)-H(-T,s)]h(s)\dif s=0.$$ In order to check the uniqueness of solution, let $u$ and $v$ be solutions of problem \eqref{eq2}. Then $u-v$ satisfies equation \eqref{gen-eq} and so is of the form given when we first studied the cases $(C1^*)$--$(C3^*)$ (see Section 3). Also, $(u-v)(T)-(u-v)(-T)=2(u-v)_o(T)=0$, but this can only happen, by what has been imposed by conditions $(C1^*)$--$(C3^*)$, if $u-v\equiv0$, thus proving the uniqueness of solution. \end{proof} \begin{exa} Consider the problem $$x'(t)=\cos(\pi t)x(-t)+\sinh(t)x(t)=\cos(\pi t)+\sinh(t), \;x(3/2)=x(-3/2). $$ Clearly we are in the case (C1). If we compute the Green's function according to Theorem \ref{thmcases123} we obtain $$2\sin(\sin (\pi T))G_1(t,s)=e^{\cosh (s)-\cosh (t)}\begin{cases} \sin \left(\frac{\sin (\pi s)}{\pi }-\frac{\sin (\pi t)}{\pi }-\frac{\sin (\pi T)}{\pi }\right)+ \cos \left(\frac{\sin (\pi s)}{\pi }+\frac{\sin (\pi t)}{\pi }-\frac{\sin (\pi T)}{\pi }\right), |t|<s,\\ \sin \left(\frac{\sin (\pi s)}{\pi}-\frac{\sin (\pi t)}{\pi }+\frac{\sin (\pi T)}{\pi }\right)+\cos \left(\frac{\sin (\pi s)}{\pi }+\frac{\sin (\pi t)}{\pi }+\frac{\sin (\pi T)}{\pi }\right), |t|<-s,\\ \sin \left(\frac{\sin (\pi s)}{\pi}-\frac{\sin (\pi t)}{\pi }+\frac{\sin (\pi T)}{\pi }\right)+ \cos \left(\frac{\sin (\pi s)}{\pi }+\frac{\sin (\pi t)}{\pi }-\frac{\sin (\pi T)}{\pi }\right), |s|<t,\\ \sin \left(\frac{\sin (\pi s)}{\pi }-\frac{\sin (\pi t)}{\pi }-\frac{\sin (\pi T)}{\pi }\right) + \cos \left(\frac{\sin (\pi s)}{\pi }+\frac{\sin (\pi t)}{\pi }+\frac{\sin (\pi T)}{\pi }\right), |s|<-t.\end{cases}$$ \begin{figure}[hhht]\label{figure2g} \center{\includegraphics[width=.5\textwidth]{grafico1.png}\includegraphics[width=.5\textwidth]{grafico2.png}}\caption{Graphs of the kernel \textit{(left)} and of the functions involved in the problem \textit{(right)}.} \end{figure} \end{exa} One of the most important direct consequences of Theorem \ref{thmcases123} is the existence of maximum and antimaximum principles in the case $b\equiv0$\footnote{Note that this discards the case (C3), for which $b\equiv0$ implies $a\equiv 0$, because we are assuming $a\not\equiv0$.}. To show this and what happens in the case $b$ constant, $b\ne0$, we recall here a couple results from \cite{Cab4}. \begin{thm}[{\cite[Theorem 4.3]{Cab4}}]\label{alphasign}Let $b=0$, $\a=aT$. \par \begin{itemize} \item If $\a\in(0,\frac{\pi}{4})$ then $\overline G$ is strictly positive on $I^2$. \item If $\a\in(-\frac{\pi}{4},0)$ then $\overline G$ is strictly negative on $I^2$. \item If $\a=\frac{\pi}{4}$ then $\overline G$ vanishes on $P:=\{(-T,-T),(0,0),(T,T),(T,-T)\}$ and is strictly positive on $(I^2)\backslash P$. \item If $\a=-\frac{\pi}{4}$ then $\overline G$ vanishes on $P$ and is strictly negative on $(I^2)\backslash P$. \item If $\a\in{\mathbb R}\backslash[-\frac{\pi}{4},\frac{\pi}{4}]$ then $\overline G$ is not positive nor negative on $I^2$. \end{itemize} \end{thm} \begin{cor}[{\cite[Corollary 4.4]{Cab4}}]\label{coralphasign} Let ${\mathcal F}_\l(I)$ be the set of real differentiable functions $f$ defined on $I$ such that $f(-T)-f(T)=\l$. The operator $R_a:{\mathcal F}_\l(I)\to L^1(I)$ defined as $R_a(x(t))=x'(t)+a\, x(-t)$, with $a\in{\mathbb R}\backslash\{0\}$, satisfies \begin{itemize} \item $R_m$ is strongly inverse positive if and only if $a\in(0,\frac{\pi}{4T}]$ and $\l\ge0$, \item $R_m$ is strongly inverse negative if and only if $a\in[-\frac{\pi}{4T},0)$ and $\l\ge0$. \end{itemize} \end{cor} With these results we get the following corollary to Theorem \ref{thmcases123}. \begin{cor}\label{corsigng}Under the conditions of Theorem \ref{thmcases123}, if $a$ is nonnegative on $I$ and $b=0$,\par \begin{itemize} \item If $A(T)\in(0,\frac{\pi}{4})$ then $G_1$ is strictly positive on $I^2$. \item If $A(T)\in(-\frac{\pi}{4},0)$ then $G_1$ is strictly negative on $I^2$. \item If $A(T)=\frac{\pi}{4}$ then $G_1$ vanishes on $P:=\{(-A(T),-A(T)),(0,0),(A(T),A(T)),( A(T),- A(T))\}$ and is strictly positive on $(I^2)\backslash P$. \item If $A(T)=-\frac{\pi}{4}$ then $G_1$ vanishes on $P$ and is strictly negative on $(I^2)\backslash P$. \item If $A(T)\in{\mathbb R}\backslash[-\frac{\pi}{4},\frac{\pi}{4}]$ then $G_1$ is not positive nor negative on $I^2$. \end{itemize} Furthermore, the operator $R_a:{\mathcal F}_\l(I)\to L^1(I)$ defined as $R_a(x(t))=x'(t)+a(t)\, x(-t)$ satisfies \begin{itemize} \item $R_a$ is strongly inverse positive if and only if $A(T)\in(0,\frac{\pi}{4T}]$ and $\l\ge0$, \item $R_a$ is strongly inverse negative if and only if $A(T)\in[-\frac{\pi}{4T},0)$ and $\l\ge0$. \end{itemize} \end{cor} The second part of this last corollary, drawn from positivity (or negativity) of the Green's function could have been obtained, as we show below, without having so much knowledge about the Green's function. In order to show this, consider the following proposition in the line of the work of Torres \cite[Theorem 2.1]{Tor}.\par \begin{pro}\label{proredpro} Consider the homogeneous initial value problem \begin{equation}\label{eqhomivp} x'(t)+a(t)\,x(-t)+b(t)\,x(t)=0,\ t\in I;\ x(t_0)=0.\end{equation} If problem \eqref{eqhomivp} has a unique solution ($x\equiv 0$) on $I$ for all $t_0\in I$ then, if the Green function for \eqref{eq2cp} exists, it has constant sign.\par What is more, if we further assume $a+b$ has constant sign, the Green's function has the same sign as $a+b$. \end{pro} \begin{proof} Without lost of generality, consider $a$ to be a $2T$-periodic $L^1$ function defined on ${\mathbb R}$ (the solution of \eqref{eq2cp} will be considered in $I$). Let $G_1$ be the Green's function for problem \eqref{eq2cp}. Since $G_1(T,s)=G_1(-T,s)$ for all $s\in I$, and $G_1$ is continuous except at the diagonal, it is enough to prove that $G_1(t,s)\ne0\nkp\fa t,s\in I$.\par Assume, on the contrary, that there exists $t_1,s_1\in I$ such that $G_1(t_1,s_1)=0$. Let $g$ be the $2T$-periodic extension of $G_1(\cdot,s_1)$. Let us assume $t_1>s_1$ (the other case would be analogous). Let $f$ be the restriction of $g$ to $(s_1,s_1+2T)$. $f$ is absolutely continuous and satisfies \eqref{eqhomivp} a.e. for $t_0=t_1$, hence, $f\equiv0$. This contradicts the fact of $G_1$ being a Green's function, therefore $G_1$ has constant sign.\par Realize now that $x\equiv1$ satisfies $$x'(t)+a(t)x(-t)+b(t)x(t)=a(t)+b(t),\ x(-T)=x(T).$$ Hence, $\int_{-T}^TG_1(t,s)(a(s)+b(s))\dif s=1$ for all $t\in I$. Since both $G_1$ and $a+b$ have constant sign, they have the same sign. \end{proof} The following corollaries are an straightforward application of this result to the cases (C1)--(C3) respectively. \begin{cor}\label{cor1sig} Assume $a$ has constant sign. Under the assumptions of (C1) and Theorem \ref{thmcases123}, $G_1$ has constant sign if $$|A(T)|< \frac{\arccos(k)}{2\sqrt{1-k^2}}.$$ Furthermore, $\sign(G_1)=\sign(a)$. \end{cor} \begin{proof} The solutions of \eqref{gen-eq} for the case (C1), as seen before, are given by $$u(t) =\a e^{-B_e(t)}\[ \cos\(\sqrt{1-k^2}A(t)\)-\frac{1+k}{\sqrt{1-k^2}}\sin\(\sqrt{1-k^2}A(t)\)\].$$ Using a particular case of the phasor addition formula\footnote{$\a\cos \c+\b\sin \c=\sqrt{\a^2+\b^2}\sin(\c+\theta)$, where $\theta\in[-\pi,\pi)$ is the angle such that $\cos\theta=\frac{\b}{\sqrt{\a^2+\b^2}}$, $\sin\theta=\frac{\a}{\sqrt{\a^2+\b^2}}$.}, $$u(t)=\a e^{-B_e(t)}\sqrt{\frac{2}{1-k}}\sin\(\sqrt{1-k^2}A(t)+\theta\),$$ where $\theta\in[-\pi,\pi)$ is the angle such that \begin{equation}\label{sincos}\sin\theta=\sqrt{\frac{1-k}{2}}\quad\text{and}\quad\cos\theta=-\frac{1+k}{\sqrt{1-k^2}}\sqrt{\frac{1-k}{2}}=-\sqrt{\frac{1+k}{2}}.\end{equation} Observe that this implies that $\theta\in\(\frac{\pi}{2},\pi\)$.\par In order for the hypothesis of Proposition \ref{proredpro} to be satisfied, it is enough and sufficient to ask for $0\not\in u(I)$ for some $\a\ne0$. Equivalently, that $$\sqrt{1-k^2}A(t)+\theta\neq \pi n\nkp\fa n\in{\mathbb Z}\nkp\fa t\in I,$$ That is, $$A(t)\neq \frac{\pi n-\theta}{\sqrt{1-k^2}}\nkp\fa n\in{\mathbb Z}\nkp\fa t\in I.$$ Since $A$ is odd and injective and $\theta\in\(\frac{\pi}{2},\pi\)$, this is equivalent to \begin{equation}\label{firststimate}|A(T)|< \frac{\pi-\theta}{\sqrt{1-k^2}}.\end{equation} Now, using the double angle formula for the sine and \eqref{sincos}, $$\frac{1-k}{2}=\sin^2\theta=\frac{1-\cos(2\theta)}{2}\text{,\quad this is,\quad} k=\cos(2\theta),$$ which implies, since $2\theta\in(\pi,2\pi)$, $$\theta=\pi-\frac{\arccos(k)}{2},$$ where $\arccos$ is defined such that it's image is $[0,\pi)$. Plugging this into inequality \eqref{firststimate} yields $$|A(T)|<\sigma(k):= \frac{\arccos(k)}{2\sqrt{1-k^2}},\quad k\in(-1,1).$$ \par Using $|k|<1$, $a+b=(k+1)a+b_o$ and the continuity of $G_1$ with respect to $a$ and $b$, we can prove that the sign of the Green's function is given by Proposition \ref{proredpro}. \end{proof} \begin{rem} In the case $a$ is a constant $\omega$ and $k=0$, $A(I)=[-|\omega|T,|\omega|T]$, and the condition can be written as $|\omega|T<\frac{\pi}{4}$, which is consistent with the results found in \cite{Cab4}.\end{rem} \begin{rem} Observe that $\sigma$ is strictly decreasing on $(-1,1)$ and $$\lim_{k\to-1^+}\sigma(k)=+\infty,\quad\lim_{k\to1^-}\sigma(k)=\frac{1}{2}.$$ \end{rem} \begin{cor}\label{corC3sign} Under the conditions of (C3) and Theorem \ref{thmcases123}, $G_{1}$ has constant sign if $|A(T)|<\frac{1}{2}$. \end{cor} \begin{proof} This corollary is a direct consequence of equation \eqref{eqc3}, Proposition \ref{proredpro} and Corollary \ref{thmcases123}. Observe that the result is consistent with $\sigma(1^-)=\frac{1}{2}$. \end{proof} In order to prove the next corollary, we need the following 'hyperbolic version' of the phasor addition formula. It's proof can be done without difficulty. \begin{lem}\label{hyppha} Let $\a$, $\b,\c\in{\mathbb R}$, then $$\a\cosh \c + \b\sinh\c= \sqrt{|\a^2-\b^2|}\begin{cases}\cosh\(\frac{1}{2}\ln\left|\frac{\a+\b}{\a-\b}\right|+\c\) & \text{if}\quad \a>|\b|, \\ -\cosh\(\frac{1}{2}\ln\left|\frac{\a+\b}{\a-\b}\right|+\c\) & \text{if}\quad- \a>|\b|, \\ \sinh\(\frac{1}{2}\ln\left|\frac{\a+\b}{\a-\b}\right|+\c\) & \text{if}\quad \b>|\a|, \\ -\sinh\(\frac{1}{2}\ln\left|\frac{\a+\b}{\a-\b}\right|+\c\) & \text{if }\quad -\b>|\a|, \\ \a\,e^\c & \text{if }\quad\a=\b,\\ \a\,e^{-\c} & \text{if }\quad\a=-\b.\\ \end{cases}$$ \end{lem} \begin{cor}\label{cor2sig} Assume $a$ has constant sign. Under the assumptions of (C2) and Theorem \ref{thmcases123}, $G_1$ has constant sign if $k<-1$ or $$|A(T)|<-\frac{\ln(k-\sqrt{k^2-1})}{2\sqrt{k^2-1}}.$$ Furthermore, $\sign(G_1)=\sign(k\,a)$. \end{cor} \begin{proof} The solutions of \eqref{gen-eq} for the case (C2), as seen before, are given by $$u(t) =\a e^{-B_e(t)}\[\cosh\(\sqrt{k^2-1}A(t)\)- \frac{1+k}{\sqrt{k^2-1}}\sinh\(\sqrt{k^2-1}A(t)\)\].$$ If $k>1$, then $1<\frac{1+k}{\sqrt{k^2-1}}$, so, using Lemma \ref{hyppha}, $$u(t)=-\a e^{-B_e(t)}\sqrt{\frac{2k}{k-1}}\sinh\(\frac{1}{2}\ln\left|k-\sqrt{k^2-1}\right|+\sqrt{k^2-1}A(t)\),$$ In order for the hypothesis of Proposition \ref{proredpro} to be satisfied, it is enough and sufficient to ask that $0\not\in u(I)$ for some $\a\ne0$. Equivalently, that $$\frac{1}{2}\ln(k-\sqrt{k^2-1})+\sqrt{k^2-1}A(t)\neq 0\nkp\fa t\in I,$$ That is, $$A(t)\neq -\frac{\ln(k-\sqrt{k^2-1})}{2\sqrt{k^2-1}}\nkp\fa t\in I.$$ Since $A$ is odd and injective, this is equivalent to $$|A(T)|<\sigma(k):=-\frac{\ln(k-\sqrt{k^2-1})}{2\sqrt{k^2-1}},\quad k>1.$$ Now, if $k<-1$, then $\left|\frac{1+k}{\sqrt{k^2-1}}\right|<1$, so using Lemma \ref{hyppha}, $$u(t)=\a e^{-B_e(t)}\sqrt{\frac{2k}{k-1}}\cosh\(\frac{1}{2}\ln\left|k-\sqrt{k^2-1}\right|+\sqrt{k^2-1}A(t)\)\ne0\quad\text{for all}\quad t\in I,\ \a\ne0,$$ so the hypothesis of Proposition \ref{proredpro} are satisfied.\par Using $|k|>1$, $a+b=(k^{-1}+1)b_e+b_o$ and the continuity of $G_1$ on $a$ and $b$ we can prove that the sign of the Green's function is given by Proposition \ref{proredpro}. \end{proof} \begin{rem} If we consider $\sigma$ defined piecewise as in Corollaries \ref{cor1sig} and \ref{cor2sig} and continuously continued through $1/2$, we get $$\sigma(k):=\begin{cases} \frac{\arccos(k)}{2\sqrt{1-k^2}} & \text{ if } k\in(-1,1) \\ \frac{1}{2} & \text{ if } k=1 \\ -\frac{\ln(k-\sqrt{k^2-1})}{2\sqrt{k^2-1}} & \text{ if } k>1\end{cases}$$ This function is not only continuous (it is defined thus), but also analytic. In order to see this it is enough to consider the extended definition of the logarithm and the square root to the complex numbers. Remember that $\sqrt{-1}:=i$ and that the principal branch of the logarithm is defined as $\ln_0(z)=\ln|z|+i\theta$ where $\theta\in[-\pi,\pi)$ and $z=|z|e^{i\theta}$ for all $z\in{\mathbb C}\backslash\{0\}$. Clearly, $\ln_0|_{(0,+\infty)}=\ln$.\par Now, for $|k|<1$, $\ln_0(k-\sqrt{1-k^2}i)=i\theta$ with $\theta\in[-\pi,\pi)$ such that $\cos\theta=k$, $\sin\theta=-\sqrt{1-k^2}$, that is, $\theta\in[-\pi,0]$. Hence, $i\ln_0(k-\sqrt{1-k^2}i)=-\theta\in[0,\pi]$. Since $\cos(-\theta)=k$, $\sin(-\theta)=\sqrt{1-k^2}$, it is clear that $$\arccos(k)=-\theta=i\ln_0(k-\sqrt{1-k^2}i).$$ We thus extend $\arccos$ to ${\mathbb C}$ by $$\arccos(z):=i\ln_0(z-\sqrt{1-z^2}i),$$ which is clearly an analytic function. So, if $k>1$, $$\sigma(k)=-\frac{\ln(k-\sqrt{k^2-1})}{2\sqrt{k^2-1}}=-\frac{\ln_0(k-i\sqrt{1-k^2})}{2i\sqrt{1-k^2}}=\frac{i\ln_0(k-i\sqrt{1-k^2})}{2\sqrt{1-k^2}}=\frac{\arccos(k)}{2\sqrt{1-k^2}}.$$ $\sigma$ is positive, strictly decreasing and $$\lim_{k\to -1^+}\sigma(k)=+\infty,\quad\lim_{k\to+\infty}\sigma(k)=0.$$ \end{rem} In a similar way to Corollaries \ref{cor1sig},\ref{corC3sign} and \ref{cor2sig}, we can prove results not assuming $a$ to be a constant sign function. The result is the following. \begin{cor} Under the assumptions of Theorem \ref{thmcases123} and conditions (C1), (C2) or (C3) (let $k$ be the constant involved in such conditions), $G_1$ has constant sign if $\max A(I)<\sigma(k)$. \end{cor} \section{The cases (C4) and (C5)} Consider the following problem derived from the nonhomogeneous problem \eqref{eq2cp}. \begin{align}\label{eoparts}\begin{pmatrix}x_o' \\ x_e'\end{pmatrix} & =\begin{pmatrix}a_o-b_o & -a_e-b_e \\ a_e-b_e & -a_o-b_o\end{pmatrix}\begin{pmatrix}x_o \\ x_e\end{pmatrix}+\begin{pmatrix}h_e \\ h_o\end{pmatrix}. \end{align} The following theorems tell us what happens when we impose the boundary conditions. \begin{thm} If condition (C4) holds, then problem \eqref{eq2cp} has solution if and only if $$\int_0^Te^{B_e(s)}h_e(s)\dif s=0,$$ and in that case the solutions of \eqref{eq2cp} are given by \begin{equation} \label{e-c4} u_c(t)=e^{-B_e(t)}\[c+\int_0^t\(e^{B_e(s)}h(s)+2a_e(s)\int_0^se^{B_e(r)}h_e(r)\dif r\)\dif s\]\enskip\text{for } c\in{\mathbb R}.\end{equation} \end{thm} \begin{proof} We know that any solution of problem \eqref{eq2cp} has to satisfy \eqref{eoparts}. In the case (C4), the matrix in \eqref{eoparts} is lower triangular \begin{align}\label{eoparts3}\begin{pmatrix}x_o' \\ x_e'\end{pmatrix} & =\begin{pmatrix}-b_o & 0 \\ 2a_e & -b_o\end{pmatrix}\begin{pmatrix}x_o \\ x_e\end{pmatrix}+\begin{pmatrix}h_e \\ h_o\end{pmatrix}. \end{align} so, the solutions of \eqref{eoparts3} are given by \begin{align*}x_o(t) & =e^{-B_e(t)}\[\tilde c+\int_0^te^{B_e(s)}h_e(s)\dif s\] ,\\x_e(t) & =e^{-B_e(t)}\[c+\int_0^t\(e^{B_e(s)}h_o(s)+2a_e(s)\[\tilde c+\int_0^se^{B_e(r)}h_e(r)\dif r\]\)\dif s\],\end{align*} where $c$, $\tilde c\in{\mathbb R}$. $x_e$ is even independently of the value of $c$. Nevertheless, $x_o$ is odd only when $\tilde c=0$. Hence, a solution of \eqref{eq2cp}, if it exists, it has the form \eqref{e-c4}.\par To show the second implication it is enough to check that $u_c$ is a solution of the problem \eqref{eq2cp}. \begin{align*} u'_c(t) = & -b_o(t)e^{-B_e(t)}\[c+\int_0^t\(e^{B_e(s)}h(s)+2a_e(s)\int_0^se^{B_e(r)}h_e(r)\dif r\)\dif s\] \\& +e^{-B_e(t)}\(e^{B_e(t)}h(t)+2a_e(t)\int_0^te^{B_e(r)}h_e(r)\dif r\)=h(t)-b_o(t)u(t)+2a_e(t)e^{-B_e(t)}\int_0^te^{B_e(r)}h_e(r)\dif r. \end{align*} Now, \begin{align*} & a_e(t)(u_c(-t)-u_c(t))+2a_e(t)e^{-B_e(t)}\int_0^te^{B_e(r)}h_e(r)\dif r\\= & a_e(t)e^{-B_e(t)}\[c-\int_0^t\(e^{B_e(s)}h(-s)-2a_e(s)\int_0^se^{B_e(r)}h_e(r)\dif r\)\dif s\]\\ & - a_e(t)e^{-B_e(t)}\[c+\int_0^t\(e^{B_e(s)}h(s)+2a_e(s)\int_0^se^{B_e(r)}h_e(r)\dif r\)\dif s\]+2a_e(t)e^{-B_e(t)}\int_0^te^{B_e(r)}h_e(r)\dif r\\= & -2a_e(t)e^{-B_e(t)}\int_0^te^{B_e(r)}h_e(r)\dif s+2a_e(t)e^{-B_e(t)}\int_0^te^{B_e(r)}h_e(r)\dif r= 0. \end{align*} Hence, $$u_c'(t)+a_e(t)u_c(-t)+(-a_e(t)+b_o(t))u_c(t)=h(t),\ a.\,e. t\in I.$$ The boundary condition $x(-T)-x(T)=0$ is equivalent to $x_o(T)=0$, this is, $$\int_0^Te^{B_e(s)}h_e(s)\dif s=0$$ and the result is concluded. \end{proof} \begin{thm} If condition (C5) holds, then problem \eqref{eq2cp} has solution if and only if \begin{equation} \label{conodd2} \int_0^T e^{B(s)-A(s)}h_e(s)\dif s=0,\end{equation} and in that case the solutions of \eqref{eq2cp} are given by \begin{equation} \label{e-c5} u_c(t)=e^{A(t)}\int_0^te^{-A(s)}h_e(s)\dif s+e^{-A(t)}\[c+\int_0^te^{A(s)}h_o(s)\dif s\]\enskip\text{for } c\in{\mathbb R}.\end{equation} \end{thm} \begin{proof} In the case (C5), $b_o=b$ and $a_o=a$. Also, the matrix in \eqref{eoparts} is diagonal \begin{align}\label{eoparts5}\begin{pmatrix}x_o' \\ x_e'\end{pmatrix} & =\begin{pmatrix}a_o-b_o & 0 \\ 0 & -a_o-b_o\end{pmatrix}\begin{pmatrix}x_o \\ x_e\end{pmatrix}+\begin{pmatrix}h_e \\ h_o\end{pmatrix}. \end{align} and the solutions of \eqref{eoparts5} are given by \begin{align*}x_o(t) & =e^{A(t)-B(t)}\[\tilde c+\int_0^te^{B(s)-A(s)}h_e(s)\dif s\],\\x_e(t) & =e^{-A(t)-B(t)}\[c+\int_0^te^{A(s)+B(s)}h_o(s)\dif s\],\end{align*} where $c$, $\tilde c\in{\mathbb R}$. Since $a$ and $b$ are odd, $A$ and $B$ are even. So, $x_e$ is even independently of the value of $c$. Nevertheless, $x_o$ is odd only when $\tilde c=0$. In such a case, since we need, as in the previous Theorem, that $x_o(T)=0$, we get condition \eqref{conodd2}, which allows us to deduce the first implication of the Theorem.\par Any solution $u_c$ of \eqref{eq2cp} has the expression \eqref{e-c5}.\par To show the second implication, it is enough to check that $u$ is a solution of the problem \eqref{eq2cp}. $$ u_c'(t)=(a(t)-b(t))e^{A(t)-B(t)}\int_0^te^{B(s)-A(s)}h_e(s)\dif s-(a(t)+b(t))e^{-A(t)-B(t)}\[c+\int_0^te^{A(s)+B(s)}h_o(s)\dif s\]+h(t).$$ Now, \begin{align*}\ & a(t)u_c(-t)+b(t)u_c(t)=a(t)\(-e^{A(t)-B(t)}\int_0^te^{B(s)-A(s)}h_e(s)\dif s+e^{-A(t)-B(t)}\[c+\int_0^te^{A(s)+B(s)}h_o(s)\dif s\]\)\\ & +b(t)\(e^{A(t)-B(t)}\int_0^te^{B(s)-A(s)}h_e(s)\dif s+e^{-A(t)-B(t)}\[c+\int_0^te^{A(s)+B(s)}h_o(s)\dif s\]\)\\ = &-(a(t)-b(t))e^{A(t)-B(t)}\int_0^te^{B(s)-A(s)}h_e(s)\dif s+(a(t)+b(t))e^{-A(t)-B(t)}\[c+\int_0^te^{A(s)+B(s)}h_o(s)\dif s\]. \end{align*} So clearly, $$u_c'(t)+a(t)u_c(-t)+b(t)u_c(t)=h(t)\quad\text{for a.e. } t\in I.$$ which ends the proof. \end{proof} \section{The mixed case} When we are not on the cases (C1)-(C5), since the fundamental matrix of $M$ is not given by its exponential matrix, it is more difficult to precise when problem \eqref{eq2cp} has a solution. Here we present some partial results. \par Consider the following ODE \begin{equation}\label{usualp}x'(t)+[a(t)+b(t)]x(t)=0,\quad x(-T)=x(T).\end{equation} The following lemma gives us the explicit Green's function for this problem. Let $\upsilon=a+b$. \begin{lem}\label{lem1} Let $h$, $a$ in problem \eqref{usualp} be in $L^1(I)$ and assume $\int_{-T}^{T}\upsilon(t)\dif t\ne0$. Then problem \eqref{usualp} has a unique solution given by $$u(t)=\int_{-T}^{T}G_3(t,s)h(t)\dif s,$$ where \begin{equation}\label{e-G3} G_3(t,s)=\begin{cases}\tau\,e^{\int_t^s\upsilon(r)\dif r}, & s\le t,\\(\tau-1)e^{\int_t^s\upsilon(r)\dif r}, & s> t,\end{cases}\quad \text{and}\quad \tau=\frac{1}{1-e^{-\int_{-T}^T\upsilon(r)\dif r}}.\end{equation} \end{lem} \begin{proof} $$\frac{\partial G_3}{\partial t}(t,s)=\begin{cases}-\tau\,\upsilon(t)\,e^{\int_t^s\upsilon(r)\dif r}, & s\le t,\\-(\tau-1)\upsilon(t)e^{\int_t^s\upsilon(r)\dif r}, & s> t,\end{cases}=-\upsilon(t)G_3(t,s).$$ Therefore, $$\frac{\partial G_3}{\partial t}(t,s)+\upsilon(t)G_3(t,s)=0,\ s\ne t.$$ Hence, \begin{align*} & u'(t)+\upsilon(t)u(t)=\frac{\dif}{\dif t}\int_{-T}^{t}G_3(t,s)h(s)\dif s+\frac{\dif}{\dif t}\int_{t}^{T}G_3(t,s)h(s)\dif s+\upsilon(t)\int_{-T}^{T}G_3(t,s)h(s)\dif s\\= & [G_3(t,t^-)-G_3(t,t^+)]h(t)+\int_{-T}^{T}\[\frac{\partial G_3}{\partial t}(t,s)+\upsilon(t)G_3(t,s)\]h(t)\dif s=h(t)\enskip\text{a.\,e. }t\in I. \end{align*} The boundary conditions are also satisfied. \begin{align*} & u(T)-u(-T)=\int_{-T}^{T}\[\tau\,e^{\int_T^s\upsilon(r)\dif r}-(\tau-1)e^{\int_{-T}^s\upsilon(r)\dif r}\]h(s)\dif s\\= & \int_{-T}^{T} \[\frac{e^{\int_T^s\upsilon(r)\dif r}}{1-e^{-\int_{-T}^T\upsilon(r)\dif r}}-\frac{e^{-\int_{-T}^T\upsilon(r)\dif r}\,e^{\int_{-T}^s\upsilon(r)\dif r}}{1-e^{-\int_{-T}^T\upsilon(r)\dif r}}\]h(s)\dif s\\= & \int_{-T}^{T} \[\frac{e^{\int_T^s\upsilon(r)\dif r}}{1-e^{-\int_{-T}^T\upsilon(r)\dif r}}-\frac{e^{-\int_{T}^s\upsilon(r)\dif r}}{1-e^{-\int_{-T}^T\upsilon(r)\dif r}}\]h(s)\dif s=0. \end{align*} \end{proof} \begin{lem} \begin{equation} \label{e-Fv} |G_3(t,s)|\le F(\upsilon):=\frac{e^{\|\upsilon\|_1}}{|e^{\|\upsilon^+\|_1}-e^{\|\upsilon^-\|_1}|}.\end{equation} \end{lem} \begin{proof} Observe that $$\tau=\frac{1}{1-e^{\|\upsilon^-\|_1-\|\upsilon^+\|_1}}=\frac{e^{\|\upsilon^+\|_1}}{e^{\|\upsilon^+\|_1}-e^{\|\upsilon^-\|_1}}.$$ Hence, $$\tau-1=\frac{e^{\|\upsilon^-\|_1}}{e^{\|\upsilon^+\|_1}-e^{\|\upsilon^-\|_1}}.$$ On the other hand, $$e^{\int_t^s\upsilon(r)\dif r}\le\begin{cases} e^{\|\upsilon^-\|_1}, & s\le t,\\epsilon^{\|\upsilon^+\|_1}, & s>t, \end{cases}$$ which ends the proof. \end{proof} The next result proves the existence and uniqueness of solution of $\eqref{eq2cp}$ when $\upsilon$ is `sufficiently small'. \begin{thm}\label{thmpn}Let $h$, $a$, $b$ in problem \eqref{eq2cp} be in $L^1(I)$ and assume $\int_{-T}^{T}\upsilon(t)\dif t\ne0$. Let $W:=\{(2T)^\frac{1}{p}(\|a\|_{p^*}+\|b\|_{p^*})\}_{p\in[1,+\infty]}$ where $p^{-1}+(p^*)^{-1}=1$. If $F(\upsilon)\|a\|_1(\inf W)<1$, $F(\upsilon)$ defined as in \eqref{e-Fv}, then problem \eqref{eq2cp} has a unique solution. \end{thm} \begin{proof} With some manipulation we get $$h(t) = x'(t)+a(t)\(\int_t^{-t}x'(s)\dif s+x(t)\)+b(t)x(t)=x'(t)+\upsilon(t)x(t)+a(t)\int_t^{-t}(h(s)-a(s)x(-s)-b(s)x(s))\dif s.$$ Hence, $$x'(t)+\upsilon(t)x(t)=a(t)\int_t^{-t}(a(s)x(-s)+b(s)x(s))\dif s+a(t)\int_{-t}^{t}h(s)\dif s+h(t).$$ Using $G_{3}$ defined as in \eqref{e-G3} and Lemma \ref{lem1}, it is clear that $$x(t)=\int_{-T}^TG_3(t,s)a(s)\int_s^{-s}(a(r)x(-r)+b(r)x(r))\dif r\dif s+\int_{-T}^TG_3(t,s)\[a(s)\int_{-s}^{s}h(r)\dif r+h(s)\]\dif s,$$ this is, $x$ is a fixed point of an operator of the form $Hx(t)+\beta(t)$, so, by Banach contraction Theorem, it is enough to prove that $\|H\|<1$ for some compatible norm of $H$. \par Using Fubini's Theorem, $$Hx(t)=-\int_{-T}^{T}\rho(t,r)(a(r)x(-r)+b(r)x(r))\dif r,$$ where $\rho(t,r)=\[\int_{|r|}^{T}-\int_{-T}^{-|r|}\]G_3(t,s)a(s)\dif s$. If $\int_{-T}^{T}\upsilon(t)\dif t=\|\upsilon^+\|_1-\|\upsilon^-\|_1>0$ then $G_3$ is positive and $$\rho(t,r)\le\int_{-T}^{T}G_3(t,s)|a(s)|\dif s\le F(\upsilon)\|a\|_1.$$ We have the same estimate for $-\rho(t,r)$.\par If $\int_{-T}^{T}\upsilon(t)\dif t<0$ we proceed with an analogous argument and arrive as well to the conclusion that $|\rho(t,s)|<F(\upsilon)\|a\|_1$.\par Hence, $|Hx(t)|\le F(\upsilon)\|a\|_1\int_{-T}^{T}|a(r)x(-r)+b(r)x(r)|\dif r=F(\upsilon)\|a\|_1\|a(r)x(-r)+b(r)x(r)\|_1$. Thus, it is clear that $$\|Hx\|_p \le (2T)^\frac{1}{p}F(\upsilon)\|a\|_1(\|a\|_{p^*}+\|b\|_{p^*})\|x\|_p,\ p\in[1,\infty],$$ which ends the proof. \end{proof} \begin{rem} In the hypothesis of Theorem \ref{thmpn}, realize that $F(\upsilon)\ge 1$. \end{rem} The following result will let us obtain some information on the sign of the solution of problem \eqref{eq2cp}. In order to prove it, we will use a theorem from \cite{Cab5} we cite below. \par Consider an interval $[w,d]\subset I$, the cone \begin{equation*}\label{eqcone-cs} K=\{u\in {\mathcal C}(I): \min_{t \in [w,d]}u(t)\geq c \|u\|\}, \end{equation*} and the following problem \begin{equation}\label{eqgenpro2} x'(t) =h(t,x(t),x(-t)),\, t\in I,\quad x(-T)=x(T), \end{equation} where $h$ is an $L^1$-Caratheodory function. Consider the following conditions. \begin{enumerate} \item[$(\mathrm{I}_{\protect\rho,\omega}^{1})$] \label{EqB2} There exist $\rho> 0$ and $\omega\in\(0,\frac{\pi}{4T}\]$ such that $f^{-\rho,\rho}_\omega <\omega$ where $$ f^{{-\rho},{\rho}}_\omega:=\sup \left\{\frac{h(t,u,v)+\omega v}{\rho }:\;(t,u,v)\in [ -T,T]\times [ -\rho,\rho ]\times [-\rho,\rho ]\right\}.$$ \item[$(\mathrm{I}_{\protect\rho,\omega}^{0})$] There exists $\rho >0$ such that $$ f_{(\rho ,{\rho /c})}^\omega\cdot\inf_{t\in [w,d]}\int_{w}^{d}\overline G(t,s)\,ds>1, $$ where $$ f_{(\rho ,{\rho /c})}^\omega =\inf \left\{\frac{h(t,u,v)+\omega v}{\rho }% :\;(t,u,v)\in [w,d]\times [\rho ,\rho /c]\times [-\rho /c,\rho /c]\right\}.$$ \end{enumerate} \begin{thm}\textrm{\cite[Theorem 5.15]{Cab5}}\label{thmgen} Let $\omega\in\(0,\frac{\pi}{2}T\]$. Let $[w,d]\subset I$ such that $w=T-d\in(\max\{0,T-\frac{\pi}{4\omega}\},\frac{T}{2})$. Let \begin{equation}\label{e-c}c=\frac{[1-\tan(\omega d)][1-\tan(\omega w)]}{[1+\tan(\omega d)][1+\tan(\omega w)]}.\end{equation} Problem \eqref{eqgenpro2} has at least one non-zero solution in $K$ if either of the following conditions hold. \begin{enumerate} \item[$(S_{1})$] There exist $\rho _{1},\rho _{2}\in (0,\infty )$ with $\rho _{1}/c<\rho _{2}$ such that $(\mathrm{I}_{\rho _{1},\omega}^{0})$ and $(\mathrm{I}_{\rho _{2},\omega}^{1})$ hold. \item[$(S_{2})$] There exist $\rho _{1},\rho _{2}\in (0,\infty )$ with $\rho _{1}<\rho _{2}$ such that $(\mathrm{I}_{\rho _{1},\omega}^{1})$ and $(\mathrm{I}% _{\rho _{2},\omega}^{0})$ hold. \end{enumerate} \end{thm} \begin{thm}\label{thmmix2} Let $h\in L^\infty(I)$, $a,b\in L^1(I)$ be such that $0<|b(t)|<a(t)<\omega<\frac{\pi}{2}T$ for a.\,e. $t\in I$ and $\inf h>0$. Then there exists a solution $u$ of \eqref{eq2cp} such that, $u>0$ in $\(\max\{0,T-\frac{\pi}{4\omega}\},\min\{T,\frac{\pi}{4\omega}\}\)$. \end{thm} \begin{proof} Problem \eqref{eq2cp} can be rewritten as \begin{equation*}\label{eq2cp3} x'(t)=h(t)-b(t)\,x(t)-a(t)\,x(-t),\enskip t\in I,\quad x(-T)=x(T). \end{equation*} With this formulation, we can apply Theorem \ref{thmgen}. Since $0<a(t)-|b(t)|<\omega$ a.\,e., take $\rho_2\in{\mathbb R}^+$ large enough such that $h(t)<(a(t)-|b(t)|)\rho_2$ a.\,e. Hence, $h(t)<(a(t)-\omega)\rho_2-|b(t)|\rho_2+\rho_2\omega$ for a.\,e. $t\in I$, in particular, $$h(t)<(a(t)-\omega)v-|b(t)|u+\rho_2\omega\le(a(t)-\omega)v+b(t)\,u+\rho_2\omega\text{ for a.\,e. }t\in I;\ u,v\in[-\rho_2,\rho_2].$$ Therefore, $$\sup \left\{\frac{h(t)-b(t)u-a(t)v+\omega v}{\rho_2 }:\;(t,v)\in [ -T,T]\times [-\rho_2,\rho_2]\right\}<\omega,$$ and thus, $(\mathrm{I}_{\rho _{2},\omega}^{1})$ is satisfied.\par Let $[w,d]\subset I$ be such that $[w,d]\subset\(T-\frac{\pi}{4\omega},\frac{\pi}{4\omega}\)$. Let $c$ be defined as in \eqref{e-c} and $\epsilon=\omega\int_w^d\overline G(t,s)\dif s$.\par Choose $\d\in(0,1)$ such that $h(t)>\[\(1+\frac{c}{\epsilon}\)\omega-(a(t)-|b(t)|)\]\rho_2\d$ a.\,e. and define $\rho_1:=\d c\rho_2$. Therefore, $h>\[(a(t)-\omega)v+b(t)\,u(t)\]\frac{\omega}{\epsilon}\rho_1$ for a.\,e. $t\in I$, $u\in[\rho_1,\frac{\rho_1}{c}]$ and $v\in[-\frac{\rho_1}{c},\frac{\rho_1}{c}]$. Thus, $$\inf \left\{\frac{h(t)-b(t)u-a(t)v+\omega v}{\rho_1 }:\;(t,v)\in [w,d]\times [-\rho_1/c,\rho_1/c]\right\}>\frac{\omega}{\epsilon},$$ and hence, $(\mathrm{I}_{\rho _{1},\omega}^{0})$ is satisfied. Finally, $(S_1)$ in Theorem \ref{thmgen} is satisfied and we get the desired result. \end{proof} \begin{rem} In the hypothesis of Theorem \ref{thmmix2}, if $\omega<\frac{\pi}{4}T$, we can take $[w,d]=[-T,T]$ and continue with the proof of Theorem \ref{thmmix2} as done above. This guarantees that $u$ is positive. \end{rem} \section{Appendix: Further considerations} \subsection{The general case} The equation $\eqref{proinv1}$, for the case $\phi(t)=-t$, can be reduced to the following system \begin{align*}\L\begin{pmatrix}x_o' \\ x_e'\end{pmatrix} & =\begin{pmatrix}a_o-b_o & -a_e-b_e \\ a_e-b_e & -a_o-b_o\end{pmatrix}\begin{pmatrix}x_o \\ x_e\end{pmatrix}+\begin{pmatrix}h_e \\ h_o\end{pmatrix}, \end{align*} where \begin{align*} \L=\begin{pmatrix}c_e+d_e & c_o-d_o \\ c_o+d_o & c_e-d_e\end{pmatrix}. \end{align*} Hence, if $\det(\L(t))=c(t)c(-t)-d(t)d(-t)\ne 0$ for a.\,e. $t\in I$, $\L(t)$ is invertible a.\,e. and \begin{align*}\begin{pmatrix}x_o' \\ x_e'\end{pmatrix} & =\L^{-1}\begin{pmatrix}a_o-b_o & -a_e-b_e \\ a_e-b_e & -a_o-b_o\end{pmatrix}\begin{pmatrix}x_o \\ x_e\end{pmatrix}+\L^{-1}\begin{pmatrix}h_e \\ h_o\end{pmatrix}. \end{align*} So the general case where $c\not\equiv0$ is reduced to the case studied on Section 3, taking $$\L^{-1}\begin{pmatrix}a_o-b_o & -a_e-b_e \\ a_e-b_e & -a_o-b_o\end{pmatrix}$$ as coefficient matrix. \subsection{Computing the matrix exponential} It is very well known that, in general, it is difficult to compute the exponential of a functional matrix and it is deeply related to the property of the matrix commuting with its integral. Here we summarize the findings of \cite{Kot} on this behalf. \begin{dfn} Let $S\subset{\mathbb R}$ be an interval. Define ${\mathcal M}\subset{\mathcal C}^1({\mathbb R},{\mathcal M}_{n\times n}({\mathbb R}))$ such that for every $M\in{\mathcal M}$, \begin{itemize} \item there exists $P\in{\mathcal C}^1({\mathbb R},{\mathcal M}_{n\times n}({\mathbb R}))$ such that $M(t)=P^{-1}(t)J(t)P(t)$ for every $t\in S$ where $P^{-1}(t)J(t)P(t)$ is a Jordan decomposition of $M(t)$; \item the superdiagonal elements of $J$ are independent of $t$, as well as the dimensions of the Jordan boxes associated to the different eigenvalues of $M$; \item two different Jordan boxes of $J$ correspond to different eigenvalues; \item if two eigenvalues of $M$ are ever equal, they are identical in the whole interval $S$. \end{itemize} \end{dfn} It is straightforward to check that the functional matrices appearing in cases (C1)--(C5) belong to ${\mathcal M}$. \begin{thm}[\cite{Kot}] Let $M\in{\mathcal M}$. Then, the following statements are equivalent. \begin{itemize} \item $M$ commutes with its derivative. \item $M$ commutes with its integral. \item $M$ commutes functionally, that is $M(t)M(s)=M(s)M(t)$ for all $t,s\in S$. \item $M=\sum_{k=0}^r\gamma_k(t)C^k$ For some $C\in{\mathcal M}_{n\times n}({\mathbb R})$ and $\gamma_k\in{\mathcal C}^1(S,{\mathbb R})$, $k=1,\dots,r$. \end{itemize} Furthermore, any of the last properties imply that $M(t)$ has a set of constant eigenvectors, i.e. a Jordan decomposition $P^{-1}J(t)P$ where $P$ is constant. \end{thm} When we first studied the case (C1) --with the other cases we need similar considerations-- we needed to compute the exponential of the matrix $$\overline M=\begin{pmatrix}B_e & -(1+k)A_o\\ (1-k)A_o & -B_e\end{pmatrix}.$$ $\overline M$ has two complex conjugate eigenvalues. What is more, it functionally commutes, so it has a basis of constant eigenvectors given by the constant matrix $$Y:=\frac{1}{k-1}\begin{pmatrix}i\sqrt{1-k^2} & -i\sqrt{1-k^2} \\ k-1 & k-1 \end{pmatrix}.$$ We have that $$Y^{-1}\overline M(t)Y=Z(t):=\begin{pmatrix} -B_e-i\,A_o\sqrt{1-k^2} & 0 \\ 0 & -B_e+i\,A_o\sqrt{1-k^2} \end{pmatrix}.$$ Hence, $$e^{\overline M(t)}=e^{YZ(t)Y^{-1}}=Ye^{Z(t)}Y^{-1}=e^{-B_e(t)}\begin{pmatrix}\cos\(\sqrt{1-k^2}A(t)\) & -\frac{1+k}{\sqrt{1-k^2}}\sin\(\sqrt{1-k^2}A(t)\)\\ \frac{\sqrt{1-k^2}}{1+k}\sin\(\sqrt{1-k^2}A(t)\) & \cos\(\sqrt{1-k^2}A(t)\)\end{pmatrix}.$$

https://arxiv.org/abs/2105.12550

Commuting probability in algebraic groups

We introduce the notion of commuting probability, $p(G)$, for an algebraic group $G$. This notion is inspired by the corresponding notions in finite groups and compact groups. The computation of $p(G)$ for reductive groups is readily done using the notion of $z$-classes. We introduce two generalisations of this relation, $iz$-equivalence and $dz$-equivalence. These notions lead us naturally to the notion of a regular element in $G$. Finally, with the help of this notion of regular elements, we compute $p(G)$ for a connected, linear algebraic group $G$. We also compute the set of limit points of the numbers $p(G)$ as $G$ varies over the classes of reductive groups, solvable groups and nilpotent groups.

\section{Introduction} \emph{What is the probability that two randomly chosen elements of a finite group commute?} \vskip2mm This was the question asked by Erd\"{o}s and Turan in \cite{ET} where they formulated the notion of commuting probability for a finite group. If $C$ denotes the set of ordered pairs of commuting elements of a finite $G$ then $p(G)$, the commuting probability of $G$, is defined to be the ratio $|C|/|G \times G|$. It is proved in the same paper that this number is equal to $k/|G|$ where $k$ is the number of conjugacy classes in $G$. Almost immediately, Gustafson (\cite{G}), generalised the notion of commuting probability to compact groups where he used the Haar measure in the definition. There has been a flurry of activity around this topic for many years since then. The purpose of the present paper is to introduce the notion of commuting probability in the case of algebraic groups. In algebraic geometry, the natural measure of a certain algebraic set is its dimension. If $G$ is a linear algebraic group then we define the set $C(G)$ to be the set of ordered pairs of commuting elements in $G$, $C(G) := \{(a, b) \in G \times G: ab = ba\}$. This is a Zariski closed subset of $G \times G$. We define the commuting probability of $G$, denoted by $p(G)$, as follows $$p(G) := \frac{\dim(C(G))}{\dim(G \times G)} = \frac{\dim(C(G))}{2\dim(G)} .$$ For a finite group $G$, which can be considered as an algebraic group with $\dim(G) = 0$, we define $p(G) = 1$. We note a basic lemma which computes the commuting probability of a direct product. \begin{lemma}\label{direct} If $G = H_1 \times H_2$ then $$p(G) = \dfrac{\dim(H_1)p(H_1) + \dim(H_2)p(H_2)}{\dim (H_1) + \dim(H_2)} .$$ \end{lemma} \begin{proof} Since $G = H_1 \times H_2$, it follows that $C(G)$ is isomorphic to $C(H_1) \times C(H_2)$ and therefore $\dim(C(G)) = \dim(C(H_1)) + \dim(C(H_2))$ which further equals $2\dim(H_1)p(H_1) + 2\dim(H_2)p(H_2)$. As $\dim(G) = \dim(H_1) + \dim(H_2)$, the lemma now follows. \end{proof} If $G$ is a connected, linear algebraic group then so is $G \times G$ and then the dimension of every proper closed subset of $G \times G$ is smaller than $\dim(G \times G)$. Hence for a connected $G$, $p(G) = 1$ if and only if $G$ is abelian. If $G$ is not connected, then $p(G)$ can be 1 without $G$ being abelian. Consider $G = G_1 \times G_2$ where $G_1$ is abelian and $G_2$ is a finite non-abelian group. Then by the above lemma, $p(G) = 1$ but $G$ is not abelian. To avoid such cases, we will work with connected groups in the remainder of this paper. In any case, $p(G) = p(G^0)$, so we are not losing out on the generality. We also choose to work over $\mathbb{C}$, though the results remain valid over an algebraically closed field, indeed over a perfect field. We begin the study of $p(G)$ with the case of reductive groups in Section 2. If $G$ is a reductive group of dimension $n$ and rank $r$ then we prove in Theorem \ref{reductive} that $p(G) = (n+r)/2n$. We also list the numbers $p(G)$ for all simple groups. These calculations are used later while computing the limit points of the set of the numbers $p(G)$. The main tool used in the proof of Theorem \ref{reductive} is the notion of $z$-classes which is not useful in the case of a non-reductive connected $G$. Hence we introduce the notions of $iz$-equivalence and $dz$-equivalence, generalising the $z$-equivalence in the third section. We study the corresponding equivalence classes and make some interesting observations, which may be of independent interest. These notions lead us to the notion of a regular element in $G$. The dimension of the centralizer of a regular element in $G$ is called the regular rank of $G$. We prove in Theorem \ref{general} that if $G$ is a connected, linear algebraic group with dimension $n$ and regular rank $r$ then $p(G) = (n+r)/2n$. In this section, Section 4, we also list the numbers $p(U)$ where $U$ is the unipotent radical of a fixed Borel subgroup in a simple algebraic group. By the Theorem \ref{general}, $p(G)$ is a rational number $> 1/2$. In Section 5, we compute all rational numbers that can occur as $p(G)$ for various classes of $G$, namely the reductive groups, solvable groups and nilpotent groups. As a consequence, we compute the limit points of these numbers for all these above classes. The set of limit points in all these cases is equal to the interval $[1/2, 1]$. We then compare the results on $p(G)$ for algebraic groups with the corresponding results in finite groups. We point out some similarities and many differences between these two notions. This occupies the Section 6. In the seventh section, we point out the compatibilities of $p(G)$ for a semisimple group $G$ defined over $\mathbb{F}_q$ with the corresponding finite groups of Lie type, in the asymptotic sense. We close the paper with some concluding remarks in the last section. We also indicate some questions that may of interest to the community. \section*{Acknowledgements} It is a pleasure to thank Dipendra Prasad, Saurav Bhaumik, Anuradha Garge and Anupam Kumar Singh for many useful discussions. Anupam, in particular, has been a catalyst for this paper in more than one ways. \section{Reductive groups} Let $G$ be a complex reductive group. In this section, we compute the commuting probability $p(G)$ of $G$ in terms of the dimension and rank of $G$. The rank of a reductive group is the dimension of a maximal torus in $G$. Since we compute dimensions, it will be sufficient to work with the group $G(\mathbb{C})$. In the computation, we will use the notion of a $z$-class. Let us first recall it. Two elements $g, h \in G(\mathbb{C})$ are said to be \emph{$z$-equivalent} if their centralisers, $Z_{G(\mathbb{C})}(g)$ and $Z_{G(\mathbb{C})}(h)$, are conjugate within $G(\mathbb{C})$. This is an equivalence relation and the corresponding equivalence classes are called \emph{$z$-classes}. This equivalence is weaker than the one given by the conjugacy relation. Each $z$-class is a union of certain conjugacy classes. It is known that if $G$ is a reductive group defined over $\mathbb{C}$ then the number of $z$-classes in $G(\mathbb{C})$ is finite. The result is true in a more general setting and we refer the reader to \cite{GS} for a detailed discussion. For the present paper, the finiteness result over $\mathbb{C}$ is sufficient. \begin{theorem}\label{reductive} Let $G$ be a complex reductive group with dimension $n$ and rank $r$. The commuting probability of $G$, $p(G)$, is equal to $$\frac{n + r}{2n} = \frac{1}{2} + \frac{r}{2n}.$$ \end{theorem} \begin{proof} We will be working throughout the proof with the set of $\mathbb{C}$-rational points without mentioning the field $\mathbb{C}$ explicitly. So, whenever we talk about a subset of an algebraic set $X$, we will always mean a subset of $X(\mathbb{C})$. Let $\mathcal{C}$ denote the set of conjugacy classes in $G$ and let $\mathfrak{c}$ be a conjugacy class in $G$. The set $C(G)$ is equal to the set $$\cup_{g \in G} \{g\} \times Z_G(g) = \bigcup_{\mathfrak{c} \in \mathcal{C}} \bigg(\cup_{g \in \mathfrak{c}} \big(\{g\} \times Z_G(g)\big)\bigg) = \bigcup_{\mathfrak{c} \in \mathcal{C}} C(G)_{\mathfrak{c}} .$$ If $g, h \in \mathfrak{c}$ then $Z_G(g)$ and $Z_G(h)$ are conjugate in $G$, in particular their dimensions are the same hence $\dim C(G)_{\mathfrak{c}}$ is equal to $\dim(G)$ for each $\mathfrak{c} \in \mathcal{C}$. Let $\mathcal{Z}$ denote the set of $z$-classes in $G$. This is a finite set and $$\mathcal{C} = \cup_{\mathfrak{z} \in \mathcal{Z}} \big(\mathfrak{z}/\mathfrak{c_{\mathfrak{z}}}\big)$$ where the set $\mathfrak{z}/\mathfrak{c_{\mathfrak{z}}}$ is the set of conjugacy classes in $\mathfrak{z}$. Since $\mathcal{Z}$ is a finite set, there is one $z$-class in $G$ which is dense in $G$. Indeed, the set of regular semisimple elements forms a $z$-class in $G$ which is dense in $G$. We denote this $z$-class by $\mathfrak{z}_0$. It then follows that $\bigcup_{\mathfrak{c} \subseteq \mathfrak{z}_0} C(G)_{\mathfrak{c}}$ is dense in $C(G)$ and hence $\dim(C(G))$ is equal to the dimension of $\bigcup_{\mathfrak{c} \subseteq \mathfrak{z}_0} C(G)_{\mathfrak{c}}$. Let us fix a conjugacy class $\mathfrak{c}$ in $\mathfrak{z}_0$. Then the dimension of $\bigcup_{\mathfrak{c} \subseteq \mathfrak{z}_0} C(G)_{\mathfrak{c}}$ is equal to $\dim(\mathfrak{z}_0/\mathfrak{c}) + \dim(G)$. Further, $\dim(\mathfrak{z}_0/\mathfrak{c})$ is $\dim(G) - \dim(\mathfrak{c})$ which is $\dim T$ for a maximal torus in $G$. Hence $\dim(C(G)) = \dim (G) + \dim (T) = n + r$ and $$p(G) = \frac{n + r}{2n} = \frac{1}{2} + \frac{r}{2n}= \frac{1}{2} + \frac{\mathrm{rank}(G)}{2\dim(G)}.$$ \end{proof} \begin{remark} The analysis in the above proof implies that the dimension of the variety of conjugacy classes in a reductive $G$ is equal to the dimension of a maximal torus in $G$. This is a well-known fact, see \cite[6.4]{St} for instance. However, we have used $z$-classes because we use a generalisation of this notion in the general case. The reason for that being that we have no information about the dimension of the variety of conjugacy classes in $G$ in the general case. We do not even know if the set of conjugacy classes forms a (quasi-projective) variety. \end{remark} \begin{remark}\label{simple} Using the above theorem, we note commuting probabilities of some groups, including all simple algebraic groups. \begin{enumerate} \item $p(GL_n) = \dfrac{1}{2} + \dfrac{n}{2n^2} = \dfrac{1}{2} + \dfrac{1}{2n} = \dfrac{n+1}{2n}$, for $n \geq 1$. \vskip1mm \item $p(SL_n) = \dfrac{1}{2} + \dfrac{n-1}{2(n^2-1)} = \dfrac{1}{2} + \dfrac{1}{2(n + 1)}$, for $n \geq 1$. \vskip1mm \item $p(SO_{2n+1}) = \dfrac{1}{2} + \dfrac{n}{2(2n^2 + n)} = \dfrac{1}{2} + \dfrac{1}{2(2n + 1)}$, for $n \geq 2$. \vskip1mm \item $p(Sp_{2n}) = \dfrac{1}{2} + \dfrac{n}{2(2n^2 + n)} = \dfrac{1}{2} + \dfrac{1}{2(2n + 1)}$, for $n \geq 3$. \vskip1mm \item $p(SO_{2n}) = \dfrac{1}{2} + \dfrac{n}{2(2n^2 - n)} = \dfrac{1}{2} + \dfrac{1}{2(2n - 1)}$, for $n \geq 4$. \vskip1mm \item $p(G_2) = \dfrac{4}{7}$, \hskip2mm $p(F_4) = p(E_6) = \dfrac{7}{13}$, \hskip2mm $p(E_7) = \dfrac{10}{19}$ and $p(E_8) = \dfrac{16}{31}$. \end{enumerate} \end{remark} \section{Regular elements and regular rank} We notice that the following three properties of $z$-classes have been crucial in the proof of Theorem \ref{reductive}. \begin{enumerate} \item A $z$-class is a union of conjugacy classes, \item a reductive $G$ contains a dense $z$-class, and, \item the centralisers of two elements in a $z$-class (are conjugate within $G$, hence are isomorphic as abstract groups, and hence they) have the same dimensions. \end{enumerate} \begin{example} A non-reductive $G$ need not contain a dense $z$-class. Let $U$ denote the group of $3 \times 3$ upper triangular matrices over $\mathbb{C}$. Then for $$g = \begin{pmatrix} 1 & a & b \\ & 1 & c \\ & & 1 \end{pmatrix}, \hskip5mm Z_U(g) = \left\{\begin{pmatrix} 1 & x & y \\ & 1 & z \\ & & 1 \end{pmatrix}: cx = az \right\} .$$ The group $U/Z(U)$ is abelian. If two centralisers in $U$ are conjugate then their images in $U/Z(U)$ must be the same. Thus, the $z$-classes in $U$ correspond to the ratios $\frac{a}{c} \in (\mathbb{C} \cup {\infty})$ along with the central $z$-class. The dimension of each $z$-class is less than $3$. In particular, there is no dense $z$-class in $U$. \end{example} Taking a cue from the above three properties we make the following two definitions generalising the notion of $z$-equivalence. \begin{definition} \noindent $(1)$ Two elements $x$ and $y$ in a group $G$ are called \emph{$iz$-equivalent} if the centralisers, $Z_G(x)$ and $Z_G(y)$, are isomorphic. \noindent $(2)$ Two elements $x$ and $y$ in an algebraic group $G$ are called \emph{$dz$-equivalent} if the centralisers, $Z_G(x)$ and $Z_G(y)$, have the same dimension. \end{definition} The $iz$-equivalence is also introduced by Dilpreet Kaur and Uday Bhaskar Sharma. We are borrowing the name, $iz$-equivalence, with their kind permission. The above two relations are indeed equivalence relations. The corresponding equivalence classes will be called $iz$-classes and $dz$-classes, respectively. We have the following hierarchy in terms of the strengths of these equivalences: \begin{itemize} \item a $dz$-class is a union of $iz$-classes, \item an $iz$-class is a union of $z$-classes and \item a $z$-class is a union of conjugacy classes. \end{itemize} These equivalence relations are, in general, different as the following examples show. \begin{example} The $z$-class of central elements in a group is not always a single conjugacy class. \end{example} \begin{example}\label{iz-but-not-z} We once again consider the group $U$ of $3 \times 3$ upper triangular matrices over $\mathbb{C}$. If we take $$x = \begin{pmatrix} 1 & 1 & 1 \\ & 1 & 0 \\ & & 1 \end{pmatrix}, y = \begin{pmatrix} 1 & 0 & 1 \\ & 1 & 1 \\ & & 1 \end{pmatrix} \hskip5mm \mathrm{then} \hskip5mm Z_U(x) = \begin{pmatrix} 1 & a & b \\ & 1 & 0 \\ & & 1 \end{pmatrix}, Z_U(y) = \begin{pmatrix} 1 & 0 & c \\ & 1 & d \\ & & 1 \end{pmatrix}$$ as $a, b, c, d \in \mathbb{G}_a$. These centralisers are isomorphic to $\mathbb{G}_a^2$. But they are not conjugate as their images in the quotient $U/Z(U)$, which is abelian, are different one dimensional subgroups of $U/Z(U)$. Thus, $x, y$ are $iz$-equivalent but not $z$-equivalent. \end{example} \begin{example}\label{dz-but-not-iz} Finally, we take $x$ to be a regular semisimple element in $G = SL_2(\mathbb{C})$ and $y$ to be a regular unipotent element in the same group. For instance, take $$x = \begin{pmatrix} \lambda & \\ & \lambda^{-1} \end{pmatrix}, \lambda \ne 1, \hskip5mm \mathrm{and} \hskip5mm y = \begin{pmatrix} 1 & 1 \\ & 1 \end{pmatrix}$$ then the centraliser of $x$ is isomorphic to $\mathbb{G}_m$ and the centraliser of $y$ is isomorphic to $\mathbb{G}_a$. Hence, $x$ and $y$ are $dz$-equivalent, $\dim(\mathbb{G}_m) = \dim(\mathbb{G}_a) = 1$, but not $iz$-equivalent as $\mathbb{G}_m$ and $\mathbb{G}_a$ are not isomorphic. \end{example} \begin{remark} Note that in the Example \ref{iz-but-not-z}, the two centralisers are conjugate in $SL_3$, a bigger group than $U$. There are two ways to see this. The elements $x$ and $y$ are themselves conjugate in $SL_3$ and hence their centralisers are conjugate in $SL_3$ which, in this case, agree with their centralisers in $U$. Alternatively, the two centralisers correspond to two different simple systems of roots in the root system $\Phi$ of $SL_3$, hence they are conjugate by a Weyl group element. If we were to consider only abstract groups, then the celebrated HNN-theory tells us that two isomorphic centralisers in a group become conjugate in a bigger group. \end{remark} Emboldened by the above remark and especially the Example \ref{iz-but-not-z}, we ask \begin{question} If $G$ is a linear algebraic group and $H_1$, $H_2$ are two isomorphic subgroups of $G$, then is there a linear algebraic group containing $G$ in which $H_1$ and $H_2$ become conjugate? Equivalently, if $H_1, H_2 \subseteq G$ are isomorphic then do we have an embedding of $G$ in some $GL_n$ such that $H_1, H_2$ become conjugate in $GL_n$? \end{question} Such a question for $iz$-equivalence vis-\'{a}-vis $z$-equivalence would be difficult to consider because, unlike Example \ref{iz-but-not-z}, the centralisers may change in a bigger group. We therefore only ask how the $iz$-equivalence in $G$ behaves with respect to the embeddings of $G$ in bigger groups. Do $iz$-equivalent elements in $G$ remain $iz$-equivalent in all such embeddings? Do they become $z$-equivalent in some embedding? What would be the situation for finite groups of Lie type? Such questions will not make sense for other equivalences. We may have a semisimple element $z$-equivalent to a unipotent element, in an abelian group for instance, but these two elements will never be conjugate in a bigger group. Indeed, they may not remain $z$-equivalent in a bigger group. Consider, for instance, the two distinct elements $1 = x, y$ in $B_2(\mathbb{F}_2)$, the Borel subgroup in $GL_2(\mathbb{F}_2)$. Since the group $B_2(\mathbb{F}_2)$ is abelian, $x$ and $y$ are $z$-equivalent but they don't remain $z$-equivalent in bigger groups, in $B_2(\mathbb{F}_4)$ for instance. Similarly, $x$ and $y$ in Example \ref{dz-but-not-iz} can never be $iz$-equivalent in a bigger group. \begin{remark} The number of $iz$-classes in group $U$ above is finite. However, we do not know if this holds in general. If we knew that the number of $iz$-classes in an algebraic group is finite, at least, for a class of groups (other than the reductive groups), then we would be able to generalise the proof of Theorem \ref{reductive} for this class of groups. However, we have no such information at present. We do not even know if there is a dense $iz$-class in a class of algebraic groups (other than the reductive groups). Hence we consider the $dz$-classes, whose number is finite for every algebraic group. It also follows that every algebraic group contains a dense $dz$-class. \end{remark} \begin{definition}\label{def-regular} Let $G$ be a linear algebraic group. An element $g \in G$ is called {\em regular} if the dimension of its centraliser, $Z_G(g)$, is minimum among such dimensions. Equivalently, an element $g \in G$ is regular if its conjugacy class has the largest possible dimension. \end{definition} We note that the set of regular elements in a $dz$-class, which is open (and hence dense) in $G$ and the centralisers of two elements in this class have the same dimensions. These were the three properties that we had listed at the beginning of this section. \begin{definition} Let $G$ be a linear algebraic group. The \emph{regular rank} of $G$ is the dimension of the centraliser of a regular element in $G$. \end{definition} We are now ready to prove the general version of Theorem \ref{reductive}. \section{Non-reductive groups} In this section, $G$ is a connected, linear algebraic group defined over $\mathbb{C}$. The set of regular elements in $G$ is denoted by $G_{reg}$. We note some basic results about the regular rank. The proofs are omitted. \begin{remark}\label{regular} \begin{enumerate} \item If $G$ is reductive then the regular rank of $G$ is the same as its rank, the dimension of a maximal torus in $G$. \item If $G_1 \subseteq G_2$ then the regular rank of $G_1$ is less than or equal to that of $G_2$. \item The regular rank of $G_1 \times G_2$ is the sum of the regular ranks of $G_i$. \item The regular rank of a semidirect product $G_1 \ltimes G_2$ need not be equal to the sum of the regular ranks of $G_i$. We will see an example of this, a Borel in a semisimple group, in Lemma \ref{Borel}. \end{enumerate} \end{remark} \begin{theorem}\label{general} Let $G$ be a complex, connected, linear algebraic group. If the dimension of $G$ is $n$ and the regular rank of $G$ is $r$ then the commuting probability of $G$ is $$\frac{n + r}{2n} = \frac{1}{2} + \frac{r}{2n}.$$ \end{theorem} \begin{proof} The proof develops on the similar lines as the proof of Theorem \ref{reductive}. We note that $C(G)$ is equal to the union $\cup_{\mathfrak{c} \in \mathcal{C}} C(G)_{\mathfrak{c}}$ where $\mathcal{C}$ is the set of conjugacy classes in $G$ and $ \dim(C(G)_{\mathfrak{c}}) = \dim(G)$ for every $\mathfrak{c} \in \mathcal{C}$. Since the set $G_{reg}$ is dense in $G$ and is a union of conjugacy classes in $G$, we have that $\dim(C(G))$ is equal to $\dim(G_{reg}/\mathfrak{c}) + \dim(G)$ where the set $G_{reg}/\mathfrak{c}$ is the set of conjugacy classes in $G_{reg}$. The dimension of $G_{reg}/\mathfrak{c}$ is equal to $\dim(G) - \dim(\mathfrak{c}) = \dim Z_G(g)$, where $g$ is a fixed element in $\mathfrak{c}$. Thus, $\dim(G_{reg}/\mathfrak{c})$ is the regular rank of $G$. Hence $\dim(C(G)) = n + r$ and $$p(G) = \frac{n + r}{2n} = \frac{1}{2} + \frac{r}{2n}.$$ \end{proof} \begin{remark} For a connected linear algebraic group $G$, if the set $\mathcal{C}$ of conjugacy classes forms a variety in a natural way then it follows from the above proof that $\dim(\mathcal{C})$ is the same as the regular rank of $G$. \end{remark} We now compute the regular ranks of certain subgroups of semisimple groups. \begin{lemma}\label{Borel} Let $G$ be a complex semisimple group. We fix a Borel subgroup $B$ in $G$ and let $U$ be the unipotent radical of $B$. Then the regular rank of $U$ is equal to the rank of the group $G$. The regular rank of a parabolic in $G$ is equal to the rank of $G$. In particular, the regular rank of $B$ is equal to the rank of $G$. \end{lemma} \begin{proof} Let $u$ be a regular unipotent element of $G$ which belongs to $U$. It is proved by Steinberg, \cite[$\S 4$]{St}, that such elements exist. The dimension of $Z_G(u)$ is equal to the rank of $G$ which is bigger than or equal to the regular rank of $U$. Further, since $u$ is in a unique Borel subgroup of $G$, it follows that $Z_G(u) = Z_U(u)$. Hence the regular rank of $U$ is equal to the rank of $G$. If $P$ is a parabolic in $G$ then, up to conjugacy in $G$, $U \subset P \subset G$. It follows, from Remark \ref{regular} (2), that the regular rank of $P$ is also equal to the rank of $G$. \end{proof} \begin{corollary} Let $G$ be as in the above lemma with dimension $n$ and rank $r$, and let $U$ be as in the above lemma. Then $$p(U) = \frac{(\frac{n-r}{2}) + r}{n - r} = \frac{1}{2} + \frac{r}{n-r} .$$ \end{corollary} \begin{proof} Follows from the above two results and that $2 \dim(U) + r = n$. \end{proof} \begin{remark}\label{unipotent} Using the above result, we note commuting probabilities of the unipotent radicals of Borel subgroups of all simple groups. For a simple $G$, the corresponding unipotent radical is denoted by $U(G)$. \begin{enumerate} \item $p(U(SL_n)) = \dfrac{1}{2} + \dfrac{n-1}{n^2 - n} = \dfrac{1}{2} + \dfrac{1}{n + 1}$, for $n \geq 1$. \vskip1mm \item $p(U(SO_{2n+1})) = \dfrac{1}{2} + \dfrac{n}{2n^2} = \dfrac{1}{2} + \dfrac{1}{2n}$, for $n \geq 2$. \vskip1mm \item $p(U(Sp_{2n})) = \dfrac{1}{2} + \dfrac{n}{2n^2} = \dfrac{1}{2} + \dfrac{1}{2n}$, for $n \geq 3$. \vskip1mm \item $p(U(SO_{2n})) = \dfrac{1}{2} + \dfrac{n}{2n^2 - 2n} = \dfrac{1}{2} + \dfrac{1}{2n - 2}$, for $n \geq 4$. \vskip1mm \item $p(U(G_2)) = \dfrac{2}{3}$, \hskip2mm $p(U(F_4)) = p(U(E_6)) = \dfrac{7}{12}$, \hskip2mm $p(U(E_7)) = \dfrac{5}{9}$ and \hskip2mm \noindent $p(U(E_8)) = \dfrac{27}{50}$. \end{enumerate} \end{remark} \section{Limit points} It follows from Theorem \ref{general} that $\frac{1}{2} < p(G) \leq 1$. We also have that $p(G) = 1$ if and only if $G$ is abelian. In this section, we investigate the possible values of $p(G)$ in $[\frac{1}{2}, 1]$ as $G$ varies over all linear algebraic groups. We will also compute the limit points of these numbers. \begin{lemma}\label{simplebounded} \begin{enumerate} \item An $\alpha \in [\frac{1}{2}, 1]$ is $p(G)$ for a simple $G$ if and only if $\alpha = \frac{1}{2} + \frac{1}{2m}$ for some $m > 1$. \item The number of simple groups $G$ with $p(G) > p/q > 1/2$ is finite. \item The set of these numbers, from $(1)$ above, has only one limit point which is $\frac{1}{2}$. \end{enumerate} \end{lemma} \begin{proof} The first part of the lemma is evident from Remark \ref{simple}. The inequality $p(G) > p/q$ gives $1/2 + r/2n > p/q$ which gives $r/n > (2p-q)/q > 0$ and $n/r < q/(2p-q)$. The number $n/r$ is an integer for a simple algebraic group and it follows, from Remark \ref{simple} for instance, that there are only finitely many simple $G$ whose $n/r$ is bounded above, for every bound. Thus, there are only finitely many simple $G$ with $p(G) > p/q > 1/2$. Finally, from (2) it follows that $1/2$ is the only possible limit point of the set of numbers $p(G)$ where $G$ varies over simple algebraic groups. We see that $1/2$ is indeed a limit point, $\lim_{n \to \infty}p(SL_n) = 1/2$. \end{proof} If we consider the reductive groups, instead of only the simple ones, then the above lemma does not hold. For example, if $G = GL_2 \times \mathbb{G}_m^a$ then $p(G) = \frac{3 + a}{4 + a}$ and the limit of these numbers is $1$! In fact, we have the following result: \begin{proposition} Every rational number from the set $(\frac{1}{2}, 1]$ is $p(G)$ for some reductive group $G$. \end{proposition} \begin{proof} Assume that we have a rational number $1/2 < p/q \leq 1$. Choose an $n$ with $1/2 < (n+1)/2n \leq p/q$. This is possible as $\{(n+1)/2n\}$ is a decreasing sequence whose limit is $1/2$. Let $G = GL_n^a \times \mathbb{G}_m^b$ where $a = q - p$ and $b = (2n^2p-n^2q - nq)/2$. Here, $a \geq 0$ as $p/q \leq 1$. Further, $b$ is an integer as $n^2 - n$ is always even and $b \geq 0$ if and only if $2np - (n+1) q \geq 0$ which holds as $(n+1)/2n \leq p/q$. The rank of $G$ is equal to $an + b$ and the dimension is equal to $an^2 + b$. Then $$p(G) = \frac{(an^2 + b) + (an + b)}{2(an^2 + b)} = \frac{(n^2-n)p}{(n^2-n)q} = \frac{p}{q} .$$ \end{proof} We have the same result for nilpotent groups. \begin{proposition} Every rational number from the set $(\frac{1}{2}, 1]$ is $p(G)$ for some nilpotent group $G$. \end{proposition} \begin{proof} The proof follows exactly in the similar way as the proof of the above proposition. The role of $\mathbb{G}_m$ in the above proof will be played by $\mathbb{G}_a$ here. \end{proof} \begin{corollary}\label{limit} \begin{enumerate} \item $\{p(G): G$ is reductive$\} = \{p(G): G$ is nilpotent$\} = \{p(G): G$ is solvable$\} = \{p(G): G$ is a linear algebraic group$\} = \mathbb{Q} \cap (1/2, 1]$. \item Every $\alpha \in [1/2, 1]$ is a limit point of each of the sets in $(1)$ above. \end{enumerate} \end{corollary} \section{Comparisons with the finite case} As mentioned in the introduction, the notion of $p(G)$ for finite groups has been studied extensively. A review of whether the analogous results hold for our notion of $p(G)$ is in order. \subsection{Simple groups} If $G$ is a non-abelian finite simple group then $p(G) \leq 1/12$, as was observed by J. Dixon (\cite[Introduction]{GR}). The probabilities $p(G)$ for simple algebraic groups are also well-behaved, Lemma \ref{simplebounded}. \subsection{Direct products} If $G_1$ and $G_2$ are finite groups then $p(G_1 \times G_2) = p(G_1) p(G_2)$. However, in the case of algebraic groups $p(G_1 \times G_2)$ is almost never equal to $p(G_1)p(G_2)$. One way to understand this is that $p(G)$ for an algebraic group is always bounded below by $1/2$. The multiplicativity of $p(G)$ for direct products of algebraic groups would say that for a non-abelian $G$ and a sufficiently large integer $r$, $p(G^r) = p(G)^r < 1/2$ giving a contradiction. \subsection{Solvable groups and non-solvable groups} If $G$ is a finite group with $p(G) > 3/40$ then Guralnick and Robinson prove that $G$ is either solvable or is isomorphic to $A_5 \times T$ for an abelian group $T$ (\cite[Theorem 11]{GR}). Thus, the numbers $p(G)$ for non-solvable groups are bounded above by $3/40$, except for $p(A_5) = 1/12$. In contrast, the numbers $p(G)$ for non-abelian reductive (hence non-solvable) algebraic groups $G$ take all rational values in $(1/2, 1)$, Corollary \ref{limit} (1). \subsection{Limit points} If $G$ is a finite non-abelian group then $p(G) \leq 5/8$ (\cite[Introduction]{G}). Further, there are gaps within the numbers $p(G)$. However, the numbers $p(G)$ for algebraic groups take all rational values in $(1/2, 1]$, Corollary \ref{limit} (1). As a result, every real number in $[1/2, 1]$ is a limit point of the set $\{p(G): G$ is algebraic$\}$ whereas the set of limit points in the finite case is a nowhere dense set of rational numbers (\cite[Corollary 1.3]{E}). \subsection{Isoclinism} Two groups $G$ and $H$ are called isoclic if there are isomorphisms $\phi:G/Z(G) \to H/Z(H)$ and $\psi:G' \to H'$ such that the diagram $$\begin{tikzcd} G/Z(G) \times G/Z(G) \arrow{r}{\phi \times \phi} \arrow[swap]{d}{\alpha_G} & H/Z(H) \times H/Z(H) \arrow{d}{\alpha_h} \\ G' \arrow{r}{\psi} & H' \end{tikzcd}$$ is commutative, where $\alpha_G(aZ(G), bZ(G)) = aba^{-1}b^{-1}$ and $\alpha_H$ is defined analogously. It is proved in \cite[Lemma 2.4]{L} that if $G$ and $H$ are finite isoclinic groups then $p(G) = p(H)$. We can define isoclinism in exactly the same way as above for algebraic groups. Then $GL_n$ and $SL_n$ are isoclinic groups with different commuting probabilities. \section{Compatibility with the finite groups of Lie type} It seems from the above section that there are more differences than similarities when we compare $p(G)$ for algebraic groups with the $p(G)$ for finite groups. This is not surprising as our definition of $p(G)$ for algebraic groups uses the dimension, which is additive for direct products, and then we take the ratios of the dimensions. This is the central reason for many differences noted in the previous section. In spite of this, we contend that our notion of $p(G)$ is a natural notion by comparing it with $p(G)$ of the finite groups of Lie type. Clearly, the characteristic of the field $\mathbb{C}$ played no role in the proofs of our results. Hence the results stand good for groups defined over an algebraically closed field. In fact, if $G$ is a connected, linear algebraic group defined over a perfect field $k$ then the dimension of $G$ and the regular rank of $G$ is defined over $k$. Hence the results proved in the paper hold good over the field $k$ as well. Let $k = \mathbb{F}_q$ be the finite field with $q$ elements and let $G$ be a semisimple group defined over $k$ of dimension $n$ and rank $r$. We recall a result of Steinberg. \begin{theorem}[Steinberg]\cite[14.11]{St1} The number of semisimple conjugacy classes in $G(k)$ is equal to $q^r$. \end{theorem} Further, the number of all conjugacy classes in $G(\mathbb{F}_q)$ is $O(q^r)$ and the cardinality of $G(\mathbb{F}_q)$ is $O(q^n)$. The commuting probability of the finite group $G(\mathbb{F}_q)$ is therefore $O(q^r)/O(q^n) = O(q^{n+r})/O(q^{2n})$ which is compatible with the commuting probability of the algebraic group $G$. \section{Concluding remarks} \subsection{Non-affine groups} The set $C(G)$ makes sense even when $G$ is not an affine algebraic group and hence we can define $p(G)$ in exactly the same way for a non-affine $G$ as in the affine case. Further, the notion of a regular element also makes sense for an algebraic group $G$, the set of regular elements is a dense subset of $G$ and hence our main result, Theorem \ref{general}, remains valid that $p(G) = (n+r)/2n$. However, at this point, we do not have any computation of $p(G)$ for a non-affine algebraic group, except when $G$ is an abelian variety with $p(G) = 1$. \subsection{Further possibilities} While the set $C(G)$ of commuting pairs in a reductive group $G$ has been studied previously, no one seems to have introduced the notion of the commuting probability of $G$. The notion in the case of finite groups has many applications and has been extensively investigated. We hope that this notion also gets investigated. There are many ways in which the probabilistic studies of algebraic groups can be done. We list some possibilities below. It is our hope that they will also be taken up. \begin{question} We prove in Section 5 that every rational number in $(1/2, 1]$ is $p(G)$ for a linear algebraic group $G$. However, in the proof we have used groups that have a large abelian direct factor. What are the value sets of $p(G)$ where $G$ has no abelian direct factor? What are the limit points of this set? \end{question} \begin{question} If $H$ is an algebraic subgroup of $G$ then we can define $p(G, H)$ as $$p(G, H) = \frac{\dim(C(G, H))}{\dim(G \times H)} $$ where $C(G, H) = \{(g, h) \in G \times H: gh = hg\}$. This $p(G, H)$ gives the probability that two randomly chosen elements $g \in G$ and $h\in H$ commute with each other. Following the sections 2--4, it does not seem difficult to compute $p(G, H)$ when $H$ is a parabolic in a reductive $G$. It would be interesting to study this notion in general, especially when $G$ and $H$ are unipotent. We have no information on the difficulty level of this question in the unipotent case. \end{question}

https://arxiv.org/abs/2006.09321

Interval parking functions

Interval parking functions (IPFs) are a generalization of ordinary parking functions in which each car is willing to park only in a fixed interval of spaces. Each interval parking function can be expressed as a pair $(a,b)$, where $a$ is a parking function and $b$ is a dual parking function. We say that a pair of permutations $(x,y)$ is \emph{reachable} if there is an IPF $(a,b)$ such that $x,y$ are the outcomes of $a,b$, respectively, as parking functions. Reachability is reflexive and antisymmetric, but not in general transitive. We prove that its transitive closure, the \emph{pseudoreachability order}, is precisely the bubble-sort order on the symmetric group $\Sym_n$, which can be expressed in terms of the normal form of a permutation in the sense of du~Cloux; in particular, it is isomorphic to the product of chains of lengths $2,\dots,n$. It is thus seen to be a special case of Armstrong's sorting order, which lies between the Bruhat and (left) weak orders.

\section{Introduction} We begin by briefly recalling the theory of parking functions, introduced in various contexts in~\cite{KW,Pyke,Riordan}; see \cite{Yan} for a comprehensive survey. Consider a parking lot with $n$ parking spots placed sequentially along a one-way street. A line of $n$ cars enters the lot, one by one. The $i^{th}$ car drives to its preferred spot $a(i)$ and parks there if possible; if the spot is already occupied then the car parks in the first available spot. The list of preferences $a=(a(1),\dots,a(n))$ is called a \defterm{parking function} if all cars successfully park; in this case the \defterm{outcome} is the permutation $\outcome(a)=w=(w(1),\dots,w(n))$, where the $i^{th}$ car parks in spot $w(i)$. It is well known that the number of parking functions for $n$ cars is $(n+1)^{n-1}$. Parking functions are an established area of research in combinatorics, with connections to labeled trees, non-crossing partitions, the Shi arrangement, symmetric functions, and other topics. In this paper, we study a generalization of parking functions in which the $i^{th}$ car is willing to park only in an interval $[a(i),b(i)]\subseteq\{1,\dots,n\}$. If all cars can successfully park then we say that the pair $(a,b)=( (a(1),\dots,a(n)), (b(1),\dots,b(n)) )$ is an \defterm{interval parking function}, or IPF. (If $b(i)=n$ for all $i$, then we recover the classical case described above.) It is easy to show that there are $n!(n+1)^{n-1}$ IPFs for $n$ cars, and that if $(a,b)$ is an IPF then the sequences $a$ and $b^*=(n+1-b(n),\dots,n+1-b(1))$ must both be parking functions, raising the question of the relationship between the permutations $\outcome(a)$ and $\outcome(b^*)$. We say that a pair of permutations $(x,y)\in\mathfrak{S}_n\times\mathfrak{S}_n$ is \defterm{reachable}, written $x\unrhd_R y$, if there exists an IPF $(a,b)$ such that $x=\outcome(a)$ and $y^*=\outcome(b^*)$. Reachability is \emph{not} a partial order on $\mathfrak{S}_n$ because it is not transitive; however, its transitive closure is a partial order, which we call \defterm{pseudoreachability}. The main result of this paper is that pseudoreachability order on~$\mathfrak{S}_n$ is precisely the \emph{bubble-sorting order} on $\mathfrak{S}_n$ (see \cite[Example 3.4.3]{BB}), which in turn is an instance of the more general \defterm{sorting order} defined by Armstrong~\cite{Armstrong} for Coxeter systems. In particular, pseudoreachability lies between Bruhat and (left) weak order in $\mathfrak{S}_n$, and it is a self-dual distributive lattice, poset-isomorphic to the product $C_2\times\cdots\times C_n$, where $C_i$ denotes the chain with $i$ elements. The proof proceeds as follows. The first significant result, Theorem~\ref{thm:bruhat}, states that $(x,y)$ is reachable only if $x\geq_By$, where $\geq_B$ denotes Bruhat order. By counting the fibers of the map $(a,b)\mapsto(x,y)$, we establish Theorem~\ref{thm:RC}, the Reachability Criterion, which is a key technical tool in what follows. Using this criterion, we show in \S\ref{sec:pseudoreach-order} that pseudoreachability is no weaker than left weak order, and use this result to show that it is graded by length, just like the Bruhat and weak orders. This grading is key for the proof in \S\ref{sorting} that pseudoreachability coincides with the bubble-sorting order. Initially, we had hoped to characterize reachability of a pair $(x,y)$ in terms of pattern-avoidance conditions on $x$ and $y$. This does not appear to be possible in general, but Section~\ref{sec:avoid} contains partial results in this direction: Theorems~\ref{thm:213-avoiding} and~\ref{thm:x} give sufficient conditions for a pair $(x,y)$ to be reachable, provided that $x\geq_By$. The authors thank Margaret Bayer for proposing the study of interval parking functions to EC and JLM at the KU Combinatorics Seminar in the spring of 2019. We are grateful to the Graduate Research Workshop in Combinatorics (GRWC) for providing the platform for this collaboration in 2019, and in particular we acknowledge helpful discussions with GRWC participants Sean English and Sam Spiro. We thank Bridget Tenner for her observant comments and Richard Stanley for his communications and suggestions for several directions of future investigation. \section{Preliminaries} \label{sec:notation} Square brackets always denote integer intervals: For $m,n\in\mathbb{Z}$ we put $[m,n]=\{m,\,\dots,\,n\}$ and $[n]=[1,n]$. Lists of positive integers (including permutations) will be regarded as functions: thus we will write $a=(a(1),\dots,a(n))$ rather than $a=(a_1,\dots,a_n)$. Thus notation such as $x[a,b]$ means $\{x(a),x(a+1),\dots,x(b)\}$. To simplify notation, we sometimes drop the parentheses and commas: e.g., $2431=(2,4,3,1)$. Let $a=(a(1),\dots,a(n))$ and $b=(b(1),\dots,b(n))\in\mathbb{Z}^n$. We write $a\leq_Cb$ if $a(i)\leq b(i)$ for all $i\in[n]$; this is the \defterm{componentwise partial order} on $\mathbb{Z}^n$. The \defterm{conjugate} (or reverse complement) of $x\in[n]^n$ is the vector $x^* = (n+1-x(n), \dots, n+1-x(1))$. Conjugation is an involution that reverses componentwise order. If $\geq$ is a partial ordering on a set $S$, then $\gtrdot$ denotes the corresponding covering relation: $x\gtrdot y$ if $x>y$ and there exists no $z$ such that $x>z>y$. It is elementary that if $\geq_1$ is a partial order at least as strong as $\geq_2$ (i.e., $x\geq_2y$ implies $x\geq_1y$), then $x>_2y$ and $x\gtrdot_1y$ together imply $x\gtrdot_2y$. The symmetric group of all permutations of $[n]$ is denoted by~$\mathfrak{S}_n$. We will as far as possible follow the notation and terminology for the symmetric group used in \cite{BB}. We set $e=(1,\dots,n)$ (the identity permutation) and $w_0=(n,n-1,\dots,1)$. The permutation transposing $i$ and $j$ and fixing all other values is denoted $t_{ij}$, and we set $s_i=t_{i,i+1}$; the elements $s_1,\dots,s_{n-1}$ are the \defterm{standard generators}. Our convention for multiplication is right to left, which is consistent with treating permutations as bijective functions $[n]\to[n]$. Thus $t_{ij}x$ is obtained by transposing the \textit{digits} $i,j$ wherever they appear in $x$, while $x t_{ij}$ is obtained by transposing the digits in the $i^{th}$ and $j^{th}$ \textit{positions}. We list some standard facts from the theory of $\mathfrak{S}_n$ as a Coxeter system of type~A, with generators $S=\{s_1,\dots,s_{n-1}\}$; see \cite{BB} for details. The \defterm{length} $\ell(x)$ of $x\in\mathfrak{S}_n$ is the smallest number $k$ such that $x$ can be written as a product $s_{i_1}\cdots s_{i_k}$ of standard generators; in this case $s_{i_1}\cdots s_{i_k}$ is called a \defterm{reduced word} for $x$. It is a standard fact that length equals number of inversions: \begin{equation} \label{length-inv} \ell(x)=\{(i,j):\ 1\leq i<j\leq n,\ x(i)>x(j)\}. \end{equation} The \defterm{Bruhat order} is the partial order $>_B$ on $\mathfrak{S}_n$ defined as the transitive closure of the relations $x>t_{ij}x$ whenever $\ell(x)>\ell(t_{ij}x)$. (Multiplying $x$ by $t_{ij}$ on the right rather than the left produces the same order, because $x t_{ij} x^{-1}$ is a transposition and $x t_{ij}=(x t_{ij} x^{-1})x$.) The \defterm{(left) weak order} $>_W$ is the transitive closure of the relations $x>s_ix$ whenever $s$ is a standard generator and $\ell(x)>\ell(sx)$. Both of these orders make $\mathfrak{S}_n$ into a graded poset with bottom element $e$ and top element $w_0$. \section{Parking functions and interval parking functions} \label{sec:intro} We begin by recalling the theory of parking functions, introduced in various contexts in~\cite{KW,Pyke,Riordan}; see \cite{Yan} for a comprehensive survey. Let $a=(a(1),\,\dots,\,a(n))\in[n]^n$. Consider a parking lot with $n$ parking spaces placed sequentially along a one-way street. Cars 1,\,\dots,\,$n$ enter the lot in order and try to park. {\bf Algorithm~A:} The $i^{th}$ car parks in the first available space in the range $[a(i),n]$. If no space in the range $[a(i),n]$ is available, the algorithm fails. If Algorithm~A succeeds in parking every car, then the preference vector $a$ is called a \defterm{parking function}. The set of all parking functions $a=(a(1),\,\dots,\,a(n))$ is denoted $\PF_{n}$. It is well known that $|\PF_n|=(n+1)^{n-1}$ and that \[\PF_n = \{a\in[n]^n:\ \tilde a(i)\leq i\ \ \forall i\}\] where $\tilde{a}$ is the unique non-decreasing rearrangement of $a$; in particular, every rearrangement of a parking function is a parking function. The \defterm{outcome} of a parking function $a\in\PF_n$ is the permutation $x=\outcome(a)=(x(1),\,\dots,\,x(n))$, where $x(i)$ is the spot in which car $i$ parks given the preference list $a$. We now modify Algorithm~A to obtain our central object of study. {\bf Algorithm~B:} Let $a,b\in[n]^n$ with $a\leq_Cb$. The $i^{th}$ car parks in the first available space in the range $[a(i),b(i)]$. If no space in the range $[a(i),b(i)]$ is available, the algorithm fails. \begin{definition} If Algorithm~B succeeds in parking every car, then $\mathbf{c}=(a,b)$ is called an \defterm{interval parking function}, or IPF. The set of all interval parking functions for $n$ cars is denoted $\IPF_n$. The \defterm{feasible interval} for the $i^{th}$ car is $[a(i),b(i)]$. \end{definition} For example, \[\IPF_2 = \{(11,12),\ (11,22),\ (12,12),\ (12,22),\ (21,21),\ (21,22)\}.\] Unlike ordinary parking functions, IPFs are \emph{not} invariant under the action of $\mathfrak{S}_2$ by permuting cars. For example, $(11,12)$ is an IPF but $(11,21)$ is not. \begin{prop} \label{IPF-characterization} Let $a,b\in[n]^n$. Then: \begin{enumerate} \item $a\in\PF_n$ if and only if $(a,(n,\dots,n))\in\IPF_n$. \item $(a,b)\in\IPF_n$ if and only if $a\in\PF_n$ and $\outcome(a) \le_C b$. \end{enumerate} \end{prop} \begin{proof} For (1), if $b(i)=n$ for all $i$ then Algorithm~B is identical to Algorithm~A. For (2), if the given conditions hold, then the execution of Algorithm~B mimics that of Algorithm~A. On the other hand, if $a$ is not a parking function, then some car will not find a spot, while if $\outcome(a)\not\leq_Cb$ then some car will not find a spot in its own feasible interval. \end{proof} As a consequence of the proof of (2), the outcome $\outcome(\mathbf{c})$ of $\mathbf{c}=(a,b)$ is just $\outcome(a)$. Moreover, for every $a\in\PF_n$, there are precisely $n!$ choices for $b$ such that $(a,b)\in\IPF_n$. (This fact was first observed by Sean English.) In particular, \begin{equation} \label{count-IPF} \left|\IPF_{n}\right| = n!(n+1)^{n-1}. \end{equation} \begin{prop} \label{lots-of-facts} Let $\mathbf{c} = (a,b)\in\IPF_n$. Then: \begin{enumerate} \item $b^*\in\PF_n$. \item $a \le_C \outcome(\mathbf{c}) \le_C b$ and $\outcome(b^*)^* \le_C b$. \end{enumerate} \end{prop} \begin{proof} \noindent \begin{enumerate} \item From $\outcome(\mathbf{c}) \leq_C b$, one has $b^* \leq_C \outcome(\mathbf{c})^*$, the latter is a permutation. Hence $b^*$ is a parking function. \item Evidently $a \le_C \outcome(\mathbf{c}) \le_C b$. By (1), $b^*$ is a parking function. Thus $b^* \le_C \outcome(b^*)$. Conjugation reverses the order $\leq_C$ and is an involution, so $\outcome(b^*)^* \le_C (b^*)^* = b$. \qedhere \end{enumerate} \end{proof} \section{The Bruhat property} \label{sec:Bruhat} In this section, we prove another property of interval parking functions related to Bruhat order on permutations. We use the following characterization of Bruhat order~\cite[Thm.~2.1.5, p.32]{BB}: $y\leq_Bx$ if and only if \begin{equation} \label{bruhat-criterion} y\langle i,j\rangle\leq x\langle i,j\rangle \qquad \forall i,j\in[n] \end{equation} where \begin{equation} \label{angle-brackets} u\langle i,j\rangle = \#\{k\in[i]:\ u(k)\geq j\}. \end{equation} (This quantity is notated $u[i,j]$ in \cite{BB}, but we reserve that notation for the image of an interval under a permutation.) For later use, we observe that by pigeonhole, it is always the case that \begin{equation} \label{bracket-ineq} x\langle i,j\rangle\geq i-j+1. \end{equation} Suppose that $\mathbf{c}=(a,b)$ is an IPF, and let $x=\outcome(a)$ and $y=\outcome(b^*)^*$. Then $x\langle i,j\rangle$ is the number of cars $1,\,\dots,\,i$ that park at or after spot $j$ under the parking function $a$. \begin{theorem} \label{thm:bruhat} Suppose that $\mathbf{c} = (a,b)$ is an IPF. Let $x=\outcome(a)$ and $y=\outcome(b^*)^*$. Then $x\geq_By$. \end{theorem} \begin{proof} First, we may assume without loss of generality that $x=a$, because replacing $a$ with $x$ doesn't change the execution of Algorithm~B (the $i^{th}$ car will have to drive to spot $x(i)$ anyway, and it is able to park there because $\mathbf{c}$ is an IPF). Fix $i,j\in[n]$, and let $p=x\langle i,j\rangle$ and $q=y\langle i,j\rangle$. By~\eqref{bruhat-criterion} we wish to show that $p\geq q$. By definition of $y\langle i,j\rangle$ we have \begin{equation} \label{bruhat:1} \Big|y[1,i]\cap[j,n]\Big|=q \end{equation} or equivalently \begin{equation} \label{bruhat:2} \Big|y^*[n-i+1,n]\cap[1,n+1-j]\Big|=q. \end{equation} Therefore, when Algorithm~A is run on the parking function $b^*$ with outcome $y^*$, the first $n-i$ cars must leave open at least $q$ spaces in the range $[1,n+1-j]$, so they cannot fill as many as $(n+1-j)-q+1=n-j-q+2$ of them. Therefore, $b^*[1,n-i]$ can contain no subset $\{v(1),\,\dots,\,v^*(n-j-q+2)\}$ such that \[(v(1),\,\dots,\,v^*(n-j-q+2))\leq_C (q,\,\dots,\,n+1-j).\] Equivalently, $\{b(i+1),\,\dots,\,b(n)\}$ can contain no subset $\{v(1),\,\dots,\,v(n-j-q+2)\}$ such that \[(v(1),\,\dots,\,v(n-j-q+2))\geq_C (j,\,\dots,\,n-q+1).\] It follows that when Algorithm~B is run on $\mathbf{c}$, no more than $n-j-q+1$ of the last $n-i$ cars will park in the spots $[j,n]$. On the other hand, since $x=\outcome(\mathbf{c})$, no more than $p=x\langle i,j\rangle$ of the first $i$ cars can park in the spots $[j,n]$. Therefore, the total number of cars that park in $[j,n]$ is at most \[(n+1-j-q)+p = |[j,n]|+(p-q).\] On the other hand, exactly $|[j,n]|$ cars park in $[j,n]$. It follows that $p\geq q$, as desired. \end{proof} Theorem~\ref{thm:bruhat} asserts that there is a well-defined \defterm{bioutcome} function \begin{equation} \label{define-bioutcome} \begin{array}{llll} \bioutcome:&\IPF_n&\to&\{(x,y)\in\mathfrak{S}_n\times\mathfrak{S}_n:\ x\geq_By\}\\ &(a,b)&\mapsto&(\outcome(a),\outcome(b^*)^*). \end{array} \end{equation} We say that a pair $(x,y)\in\mathfrak{S}_n\times\mathfrak{S}_n$ is \defterm{reachable} if it is in the image of $\bioutcome$; in this case we write $x\unrhd_R y$. (We use this notation rather than $x\geq_R y$ because reachability is not a partial order on $\mathfrak{S}_n$, as we will discuss shortly.) Then Theorem~\ref{thm:bruhat} asserts that all reachable pairs are related in Bruhat order. \begin{remark} \label{unreachable} If $a$ and $b^*$ are parking functions such that $\outcome(a)\geq_B\outcome(b^*)^*$, it does \emph{not} follow that $\mathbf{c}=(a,b)$ is an IPF. For example, if $a=w_0$ and $b$ is a permutation, then certainly $a=\outcome(a)\geq_B\outcome(b^*)^*=b$, but $(a,b)$ is an IPF only if $b=w_0$ as well. Moreover, if $x,y\in\mathfrak{S}_n$ with $x\geq_By$, there does not necessarily exist any IPF $\mathbf{c}=(a,b)$ such that $\bioutcome(\mathbf{c})=(x,y)$. For example, when $n=3$, take $(x,y)=(321,213)$, so that $y^*=132$. Then $a=321$ is the only parking function with $\outcome(a)=x$. By Prop.~\ref{lots-of-facts}(2) we must have $b\geq_Ca$, so $b\in\{321,331, 322, 332, 323,333\}$ and $b^*\in\{321,311,221,211,121,111\}$. But none of these parking functions have outcome $y^*=132$. \end{remark} The relation of reachability is reflexive (because $\bioutcome(x,x)=(x,x)$ for all $x\in\mathfrak{S}_n$) and antisymmetric (as a consequence of Theorem~\ref{thm:bruhat}). However, it is not transitive: for example, $321\mkern-1mu\not\mathrel{\mkern1mu\unrhd_R}\mkern1mu 213$, as just shown, but $(321,312)=\bioutcome(312,322)$ and $(312,213)=\bioutcome(312,313)$ are reachable. This observation motivates the following definition. \begin{definition} \label{def-pseu} We say that $(x,y)$ is \defterm{pseudoreachable}, written $x\geq_P y$, if there is a sequence $x=x_0\unrhd_R x_1\unrhd_R\cdots\unrhd_R x_k=y$. That is, pseudoreachability is the transitive closure of reachability. As such, it is a partial order on $\mathfrak{S}_n$, which by Theorem~\ref{thm:bruhat} is no stronger than Bruhat order. \end{definition} For reference, we summarize the various order-like relations that we will consider. \medskip \begin{center} {\renewcommand\arraystretch{1.2} \begin{tabular}{lll} \hline $a\geq_C b$ & Componentwise order & on $\mathbb{Z}^n$\\ $x\geq_B y$ & Bruhat order & \rdelim\}{4}{1em}[\ on $\mathfrak{S}_n$] \\ $x\geq_W y$ & Left weak order\\ $x\unrhd_R y$ & Reachability (not transitive)\\ $x\geq_P y$ & Pseudoreachability\\ \hline \end{tabular} } \end{center} \medskip \section{Reachability via counting fibers of the bioutcome map} \label{sec:reachable} Fix a pair of permutations $(x,y)\in\mathfrak{S}_n\times\mathfrak{S}_n$. How can we determine if $(x,y)$ is reachable? More generally, what is the number $\phi(x,y)=|\bioutcome^{-1}(x,y)|$ of IPFs $(a,b)$ with bioutcome~$(x,y)$? We can answer this enumerative question quickly, although the resulting formula is recursive and somewhat opaque. First, for each $i$, the number of possibilities $\mathsf{c}_i=\mathsf{c}_i(x,y)$ for $a(i)$ is the size of the largest block of spaces ending in $x(i)$ that are all occupied by one of the first $i$ cars. That is, \[\mathsf{c}_i=\mathsf{c}_i(x,y) = \max\left\{ j \in [1, x(i)]: x^{-1}(x(i)-k) \le i \text{ for all } 0 \le k \le j-1 \right\}.\] Second, given $a(1),\dots,a(i)$, the number of possibilities for $b(i)$ is $\mathsf{d}_i=\mathsf{d}_i(x,y) = \#\mathsf{D}_i(x,y)$, where \[\mathsf{D}_i(x,y)=\{k\in[0,J_i-1]:\ y(i) + k \ge x(i)\}\] and \[J_i =\max\{j\in[1,n+1-y(i)]:\ y^{-1}(y(i)+s) \ge i \text{ for all }0 \le s \le j-1\}.\] The definition of $J_i$ is analogous to that of $\mathsf{c}_i$: it is the size of the largest block of spaces ending in $n+1-y(i)$ that are all occupied by one of the first $n+1-i$ cars, so it is the number of possible values for $b^*_i$ under which $\outcome(b^*)=y^*$. The additional condition $y(i)+k\geq x(i)$ in the definition of $\mathsf{D}_i$ ensures that $(a,b)$ is an IPF because the upper bound on $x(i)$ given by $b(i)$ does not conflict with where the $i^{th}$ car parks under Algorithm~B. The sequences $\mathsf{c}=(\mathsf{c}_1,\dots,\mathsf{c}_n)$ and $\mathsf{d}=(\mathsf{d}_1,\dots,\mathsf{d}_n)$ then determine the size of the fibers of $\bioutcome$: \begin{equation} \phi(x,y)=\left|\bioutcome^{-1}(x,y)\right| = \prod_{i=1}^{n} \mathsf{c}_i\mathsf{d}_i. \end{equation} \begin{example} Let $x = 361245$ and $y = 341256$. Then $\mathsf{c}=(1,1,1,2,4,5)$ and $\mathsf{d}=(4,1,2,1,2,1)$, so there are $2^34^25^1=640$ IPF's with bioutcome $(x,y$). \end{example} It is clear from the definition that $1\leq\mathsf{c}_i\leq i$ for all $i$. On the other hand, one or more $\mathsf{d}_i$ may be zero. The pair $(x,y)$ is reachable if and only if $\mathsf{d}_i>0$ for all $i$; we refer to this as the \textbf{Count Criterion} for reachability. Evidently, the largest fiber occurs when $x$ and $y$ both equal the identity permutation in $\mathfrak{S}_n$. In this case $\mathsf{c}=(1,2,\dots,n)$ and $\mathsf{d}=(n,n-1,\dots,1)$, and the fiber size is $(n!)^2$. At the opposite end of the spectrum, if $x=y=(n,\dots,1)$, then $\phi(x,y) = 1$. Perhaps a better way to think about reachability is the following criterion. If we are solely interested in reachability and not the number of IPFs that achieve a given outcome, we can rephrase reachability more directly in terms of the permutations $x$ and $y$. \begin{theorem}[\textbf{Reachability Criterion}]\label{thm:RC} Let $x,y\in\mathfrak{S}_n$. Then \begin{equation}\label{RC} x\unrhd_R y \quad\iff\quad [y(i),x(i)] \subseteq y[i,n] \quad \forall\, i \in [n].\tag{\textbf{RC}} \end{equation} \end{theorem} \begin{proof} Let $i\in[n]$. We will show that $\mathsf{d}_i(x,y)>0$ if and only if $[y(i),x(i)] \subseteq y[i,n]$. Suppose that $[y(i),x(i)] \setminus y[i,n]\neq \emptyset$. That is, there is some $m\in[y(i),x(i)]$ such that $y^{-1}(m)<i$. Thus $J_i \leq m-y(i)$, so $y(i)+k<m\leq x(i)$ for all $k<J_i$, so $\mathsf{d}_i(x,y)=0$. Now assume that $[y(i),x(i)] \subseteq y[i,n]$. We wish to show that $\mathsf{D}_i\neq\emptyset$. If $y(i)\geq x(i)$, then $0\in\mathsf{D}_i$. On the other hand, if $y(i)<x(i)$, then $m=x(i)-y(i)>0$, and for all $0 \leq k \leq m$ we have $y^{-1}(y(i)+k) \ge i$. Therefore $J_i > m$ and $m\in\mathsf{D}_i$. \end{proof} It is worth emphasizing that the Reachability Criterion is sufficient, but not necessary, for showing that $x\geq_Py$. For example, the pair $(x,y)=(321,213)$ fails~\eqref{RC} for $i=2$, but nonetheless $x\geq_Py$. \begin{prop} \label{cf-df-facts} The sequence $\mathsf{d}(x,y)$ has the following properties. \begin{enumerate}[label=(\alph{enumi})] \item\label{dfone} $\mathsf{d}_1\geq1$. \item\label{dfi} For each $i$, if $y(i)\geq x(i)$, then $\mathsf{d}_i\geq1$. \item\label{dfn} If $x\geq_By$, then $\mathsf{d}_n=1$. \end{enumerate} \end{prop} \begin{proof} The first two assertions are direct consequences of~\eqref{RC}. For~\ref{dfone}, we have $[y(1),x(1)] \subseteq [n] = y[n]$, and for~\ref{dfi}, if $y(i)\geq x(i)$ then $[y(i),x(i)]\subseteq\{y(i)\}\subseteq y[i,n]$. For~\ref{dfn}, if $y\leq_B x$, then $y(n) \ge x(n)$ (a consequence of the inequalities~\eqref{bruhat-criterion} for $i=n-1$ and all $j$), so $\mathsf{d}_n>0$ by part~\ref{dfi}. Observe that \[J_n =\max\{j:\ y(n)+k\leq n \text{ and } y^{-1}(y(n)+k) \ge n \text{ for all }0 \le k \le j-1\} = 1\] because the conditions are true for $k=0$ but false for $k>0$. Therefore, $\mathsf{D}_n=\{k\in[0,0]:\ y(n)\geq x(n)\}=\{0\}$ and $\mathsf{d}_n=\#\mathsf{D}_n=1$. \end{proof} \section{Pseudoreachability order is graded} \label{sec:pseudoreach-order} In this section, we prove that the pseudoreachability order $\geq_P$ on $\mathfrak{S}_n$ is graded by length, just like the Bruhat and weak orders. Temporarily, we will use the notation $x{\,\mathrlap{\gtrdot}\rhd}_R\, y$ to mean that $x\unrhd_R y$ and $\ell(x) = \ell(y) + 1$. Note that if $x{\,\mathrlap{\gtrdot}\rhd}_R\, y$ then $x\gtrdot_Py$ (because $x\gtrdot_By$). Our goal is to prove the converse of the last statement, which will imply that pseudoreachability is graded by length. We have already shown that pseudoreachability order is no stronger than Bruhat order $\geq_B$. We next show that it is no weaker than left weak order $\geq_W$. \begin{prop} \label{lem:inv-1} If $x\gtrdot_Wy$, then $x{\,\mathrlap{\gtrdot}\rhd}_R\, y$. \end{prop} \begin{proof} Suppose that $x\gtrdot_Wy$, i.e., that $x=s_ay$, where $j=y^{-1}(a)<y^{-1}(a+1)=k$. Then Prop.~\ref{cf-df-facts}\ref{dfi} implies that $\mathsf{d}_i(x,y)>0$ for all $i\in[n]\setminus\{j\}$. Meanwhile $[y(j),x(j)]=\{a,a+1\}=\{y(j),y(k)\}\subseteq[y(j),y(n)]$, so~\eqref{RC} implies that $\mathsf{d}_j(x,y)>0$ as well. \end{proof} For each $x\in\mathfrak{S}_n$, let $\hat x$ be the permutation in $\mathfrak{S}_{n-1}$ defined by \begin{equation} \label{hats-on} \hat x(i) = \begin{cases} x(i) & \text{ if } x(i)<x(n),\\ x(i)-1 & \text{ if } x(i)>x(n).\end{cases} \end{equation} \begin{lemma} \label{lem:proj} Let $x,y \in \mathfrak{S}_n$ with $x(n) = y(n)$. Then $x\unrhd_R y$ if and only if $\hat x\unrhd_R\hat y$. \end{lemma} \begin{proof} By~\eqref{RC}, the proof reduces to showing that \begin{subequations} \begin{equation} \label{reach-xy} [y(i),x(i)] \subseteq y[i,n] \qquad \forall i\in[n] \end{equation} if and only if \begin{equation} \label{reach-proj} [\hat y(i),\hat x(i)] \subseteq \hat y[i,n] \qquad \forall i\in[n-1]. \end{equation} \end{subequations} ($\implies$) Assume that~\eqref{reach-xy} holds. Let $i\in[n-1]$ and $a \in [\hat y(i),\hat x(i)]$. There are two cases to consider. \textit{Case 1a}: $a < y(n)$. Then $\hat y(i)\leq a<y(n)$, so $\hat y(i)=y(i)$ (since~\eqref{hats-on} implies that if $\hat y(i)=y(i)-1$ then $\hat y(i)\geq y(n)$). Thus \[ [\hat y(i),a] = [y(i),a] \subseteq [y(i),x(i)] \subseteq y[i,n] \] because $a\leq\hat x(i)\leq x(i)$, and by~\eqref{reach-xy}. Therefore $a = y(k)=\hat y(k)$ for some $k\in[i,n-1]$. \textit{Case 1b}: $a \geq y(n)$. Then, since $\hat y(i) \geq y(i) - 1$ and $x(i) \geq \hat x(i) \geq y(n)$, $a \in [\hat y(i),\hat x(i)]$ implies that $a \in [y(i)-1,x(i)-1]$, i.e., $y(i) \leq a+1 \leq x(i)$. By~\eqref{reach-xy} there is some $k\in[i,n]$ such that $a+1 = y(k)$. In fact $k\neq n$ (since $a+1>y(n)$), so $\hat y(k) = y(k) - 1 = a$ and so $a \in \hat y[i,n-1]$. In both cases we have proved~\eqref{reach-proj}. \medskip ($\impliedby$) Assume that~\eqref{reach-proj} holds. It is immediate that~\eqref{reach-xy} holds when $i=n$, so fix $i\in[n-1]$ and $a \in [y(i),x(i)]$. We wish to show that $a=y(k)$ for some $k\in[i,n]$. This is clear if $a=y(n)$, so assume $a\neq y(n)$. \textit{Case 2a}: $a < y(n)$. Since $a\in[y(i),x(i)]$, either $a = x(i)$ or $a < x(i)$. If $a = x(i)$, then $a = x(i) = \hat x(i)$. If $a < x(i)$, then $a \leq \hat x(i)$ since $\hat x(i) \geq x(i) - 1$. In either case, \[ [y(i),a] = [\hat y(i),a] \subseteq [\hat y(i),\hat x(i)] \subseteq \hat y[i,n-1]. \] Thus $a=\hat y(k)=y(k)$ for some $k\in[i,n-1]$. \textit{Case 2b}: $a > y(n)$. Since $a\in[y(i),x(i)]$, either $a = y(i)$ or $a > y(i)$. If $a = y(i)$, then $a - 1 = y(i) - 1 = \hat y(i)$ since $y(i) > y(n)$. If $a > y(i)$, then we know that $a-1 \geq \hat y(i)$ since $y(i) \geq \hat y(i)$. It follows that $a-1\in[\hat y(i),\hat x(i)]$, so, by~\eqref{reach-proj}, there is some $k\in[i,n-1]$ such that $a-1 = \hat y(k)\geq y(n)$. Therefore, $a = y(k)$. In both cases we have proved~\eqref{reach-xy}. \end{proof} \begin{corollary} \label{cor:rcover} Let $x,y \in \mathfrak{S}_n$ with $x(n) = y(n)$. Then $x{\,\mathrlap{\gtrdot}\rhd}_R\, y$ if and only if $\hat x{\,\mathrlap{\gtrdot}\rhd}_R\,\hat y$. \end{corollary} \begin{proof} The definition of $\hat x$ implies that \begin{equation} \label{hat-inv} \ell(\hat x) = \ell(x)-(n-x(n)), \end{equation} which together with Lemma~\ref{lem:proj} produces the desired result. \end{proof} \begin{prop} \label{reachable-graded} Let $x,y\in\mathfrak{S}_n$ such that $x\unrhd_R y$, and let $m=\ell(x)-\ell(y)$. Then there exists a chain \begin{equation} \label{desired-chain} x_0=y{\,\mathrlap{\lessdot}\lhd}_R\, x_1{\,\mathrlap{\lessdot}\lhd}_R\,\cdots{\,\mathrlap{\lessdot}\lhd}_R\, x_m=x. \end{equation} \end{prop} \begin{proof} The proof proceeds by double induction on $n$ and $m$. The conclusion is trivial when $n\leq2$ or $m\leq1$. Accordingly, let $n>2$ and $m>1$, and assume inductively that the theorem holds for all $(n',m')<_C(n,m)$. First, suppose that $x(n) = y(n)$. Then $\hat{x}\unrhd_R\hat{y}$ by Lemma~\ref{lem:proj} where $\hat x,\hat y$ are defined by~\eqref{hats-on}. Moreover, $\ell(\hat{x}) - \ell(\hat{y}) = \ell(x) - \ell(y) = m$ by~\eqref{hat-inv}. Therefore, by the induction hypothesis, there is a chain $\hat{y} = \hat x_0 {\,\mathrlap{\lessdot}\lhd}_R\, \hat x_1 {\,\mathrlap{\lessdot}\lhd}_R\, \cdots {\,\mathrlap{\lessdot}\lhd}_R\, \hat x_m = \hat{x}$ in $\mathfrak{S}_{n-1}$, which by Corollary~\ref{cor:rcover} can be lifted to a chain of the form~\eqref{desired-chain}. Second, suppose that $x(n) \neq y(n)$. Since $x\geq_By$ by Theorem \ref{thm:bruhat}, in fact $x(n) < y(n)$ (as noted in the proof of Prop.~\ref{cf-df-facts}\ref{dfn}). Let $p=y(n)-1$; then $p\in[1,n-1]$, so we may set $q=y^{-1}(p)$ and $z=s_py=y t_{q,n}$. Then $z\gtrdot_Wy$ and so $z{\,\mathrlap{\gtrdot}\rhd}_R\, y$ by Prop.~\ref{lem:inv-1}. We will show that $x\unrhd_R z$ using~\eqref{RC}. \textit{Case 1}: $1\leq i\leq q$. Then $[z(i),x(i)]\subseteq[y[i],x(i)]$ and $y[i,n]=z[i,n]$, so $\mathsf{d}_n(x,y)\geq1$ implies $\mathsf{d}_n(x,z)\geq1$. \textit{Case 2}: $q<i<n$. Then $p=y(q)\not\in y[i,n]$, so by~\eqref{RC} $p\not\in[y(i),x(i)]$. Thus $p+1\not\in[y(i)+1,x(i)+1]$, and certainly $p+1=y(n)\neq y(i)$. Thus $[y(i),x(i)] \subseteq y[i,n]\setminus\{y(n)\}=y[i,n-1]$ and \[[z(i),x(i)] = [y(i),x(i)] \subseteq y[i,n-1]=z[i,n-1]\subseteq z[i,n]\] so again $\mathsf{d}_n(x,z)\geq1$. \textit{Case 3}: $i=n$. Then $x(n) \leq y(n)-1 =z(n)$, so $\mathsf{d}_{n}(x,z) \geq 1$ by Prop.~\ref{cf-df-facts}\ref{dfi}. Taken together, the three cases imply $x\unrhd_R z$. By induction there is a chain $x_1=z{\,\mathrlap{\lessdot}\lhd}_R\,\cdots{\,\mathrlap{\lessdot}\lhd}_R\, x_m=x$, and appending $x_0=y$ produces a chain of the form~\eqref{desired-chain}. \end{proof} \begin{theorem} \label{pseudoreachable-graded} Pseudoreachability order is graded by length. \end{theorem} \begin{proof} The definition of pseudoreachability as the transitive closure of reachability order implies that if $x_0<_P\cdots<_Px_m$ is a maximal chain, then in fact each $x_{i-1}\unlhd_R x_i$ for all $i$. Now, maximality together with Prop.~\ref{reachable-graded} implies in turn that in fact $x_{i-1}{\,\mathrlap{\lessdot}\lhd}_R\, x_i$. \end{proof} For comparison, the Hasse diagrams of Bruhat, pseudoreachability, and left weak orders on $\mathfrak{S}_3$ are shown in Figure~\ref{fig:s3}, together with the reachability relation (which is reflexive and antisymmetric, but not transitive). The three partial orders on $\mathfrak{S}_4$ are shown in Figure~\ref{fig:s4}. \begin{figure} \begin{center} \begin{tikzpicture} \newcommand{4.5}{4.5} \begin{scope}[shift={(0,0)}] \draw[black] (0,0)--(-1,1)--(-1,2)--(0,3)--(1,2)--(1,1)--cycle; \draw[black] (-1,1)--(1,2) (-1,2)--(1,1); \foreach \times/\y/\w in {0/0/123, -1/1/132, 1/1/213, -1/2/231, 1/2/312, 0/3/321} \node[fill=white] at (\times,\y) {\sf\w}; \node at (0,-.5) {Bruhat order $\geq_B$}; \end{scope} \begin{scope}[shift={(4.5,0)}] \draw[black] (0,0)--(-1,1)--(-1,2)--(0,3)--(1,2)--(1,1)--cycle; \draw[black] (-1,1)--(1,2); \foreach \times/\y/\w in {0/0/123, -1/1/132, 1/1/213, -1/2/231, 1/2/312, 0/3/321} \node[fill=white] at (\times,\y) {\sf\w}; \node at (0,-.5) {Pseudoreachability $\geq_P$}; \end{scope} \begin{scope}[shift={(2*4.5,0)}] \draw[black] (0,0)--(-1,1)--(-1,2)--(0,3)--(1,2)--(1,1)--cycle; \foreach \times/\y/\w in {0/0/123, -1/1/132, 1/1/213, -1/2/231, 1/2/312, 0/3/321} \node[fill=white] at (\times,\y) {\sf\w}; \node at (0,-.5) {Left weak order $\geq_W$}; \end{scope} \begin{scope}[shift={(3*4.5,0)}] \foreach \p in {(-1,1), (1,1), (-1,2), (1,2), (0,3)} \draw[black] (0,0)--\p; \foreach \p in {(-1,2), (1,2), (0,3)} \draw[black] (-1,1)--\p; \draw[black] (1,1)--(1,2)--(0,3)--(-1,2); \foreach \times/\y/\w in {0/0/123, -1/1/132, 1/1/213, -1/2/231, 1/2/312, 0/3/321} \node[fill=white] at (\times,\y) {\sf\w}; \node at (0,-.5) {Reachability $\unrhd_R$}; \end{scope} \end{tikzpicture} \caption{Bruhat, pseudoreachability, left weak order, and reachability on $\mathfrak{S}_3$\label{fig:s3}} \end{center} \end{figure} \begin{figure} \begin{center} \begin{tikzpicture}[scale=1.4] \coordinate (p4321) at (0,6); \coordinate (p3421) at (-1.5,5); \coordinate (p4231) at (0,5); \coordinate (p4312) at (1.5,5); \coordinate (p2431) at (-3,4); \coordinate (p3241) at (-1.5,4); \coordinate (p3412) at (0,4); \coordinate (p4132) at (1.5,4); \coordinate (p4213) at (3,4); \coordinate (p1432) at (-3.5,3); \coordinate (p2341) at (-2.25,3); \coordinate (p3142) at (-.75,3); \coordinate (p2413) at (.75,3); \coordinate (p3214) at (2.25,3); \coordinate (p4123) at (3.5,3); \coordinate (p1342) at (-3,2); \coordinate (p1423) at (-1.5,2); \coordinate (p2143) at (0,2); \coordinate (p2314) at (1.5,2); \coordinate (p3124) at (3,2); \coordinate (p1243) at (-1.5,1); \coordinate (p1324) at (0,1); \coordinate (p2134) at (1.5,1); \coordinate (p1234) at (0,0); \draw[very thick] (p1243)--(p1234) (p1324)--(p1234) (p1342)--(p1243) (p1423)--(p1324) (p1432)--(p1423) (p1432)--(p1342) (p2134)--(p1234) (p2143)--(p1243) (p2143)--(p2134) (p2314)--(p1324) (p2341)--(p1342) (p2413)--(p1423) (p2413)--(p2314) (p2431)--(p1432) (p2431)--(p2341) (p3124)--(p2134) (p3142)--(p2143) (p3214)--(p3124) (p3214)--(p2314) (p3241)--(p3142) (p3241)--(p2341) (p3412)--(p2413) (p3421)--(p3412) (p3421)--(p2431) (p4123)--(p3124) (p4132)--(p4123) (p4132)--(p3142) (p4213)--(p4123) (p4213)--(p3214) (p4231)--(p4132) (p4231)--(p3241) (p4312)--(p4213) (p4312)--(p3412) (p4321)--(p4312) (p4321)--(p4231) (p4321)--(p3421); \draw[thick,blue] (p1243)--(p1423) (p1324)--(p3124) (p1342)--(p3142) (p1423)--(p4123) (p1432)--(p3412) (p1432)--(p4132) (p2143)--(p4123) (p2413)--(p4213) (p2431)--(p4231) (p4132)--(p4312); \draw[red] (p1324)--(p1342) (p2134)--(p2314) (p2143)--(p2341) (p2143)--(p2413) (p2314)--(p2341) (p2413)--(p2431) (p3124)--(p3142) (p3142)--(p3412) (p3214)--(p3241) (p3214)--(p3412) (p3241)--(p3421) (p4213)--(p4231); \foreach \name/\loc in {1234/p1234, 1243/p1243, 1324/p1324, 1342/p1342, 1423/p1423, 1432/p1432, 2134/p2134, 2143/p2143, 2314/p2314, 2341/p2341, 2413/p2413, 2431/p2431, 3124/p3124, 3142/p3142, 3214/p3214, 3241/p3241, 3412/p3412, 3421/p3421, 4123/p4123, 4132/p4132, 4213/p4213, 4231/p4231, 4312/p4312, 4321/p4321} \node at (\loc) [rectangle,draw,fill=white] {\sf\scriptsize\name}; \draw[very thick] (5,3.5)--(6,3.5); \node at (7,3.5) {$\geq_B,\ \geq_P,\ \geq_W$}; \draw[thick, blue] (5,3)--(6,3); \node at (6.7,3) {$\geq_B,\ \geq_P$}; \draw[red] (5,2.5)--(6,2.5); \node at (6.44,2.5) {$\geq_B$}; \end{tikzpicture} \caption{Bruhat, pseudoreachability, and left weak order on $\mathfrak{S}_4$\label{fig:s4}} \end{center} \end{figure} \section{Pseudoreachability order and bubble-sorting order}\label{sorting} The theory of normal forms in a Coxeter system was introduced by du~Cloux~\cite{duCloux} and is described in~\cite[\S3.4]{BB}. We sketch here the facts we will need; see especially~\cite[Example 3.4.3]{BB}, which describes normal forms in the symmetric group in terms of bubble-sorting. Let $\sigma_k=s_1\cdots s_k$ and $\omega_n=\sigma_{n-1}\cdots\sigma_1$; then $\omega_n$ is a reduced word for $w_0\in\mathfrak{S}_n$. Every $x\in\mathfrak{S}_n$ has a unique \textbf{conormal form}: a reduced word $N(w)$ of the form $v_{n-1} v_{n-2} \cdots v_2 v_1$, where $v_k=s_j s_{j+1}\cdots s_k$ is a suffix of $\sigma_k$. The conormal form is the reverse of the lexicographically first reduced word for $x^{-1}$ (that is, of the normal form of $x^{-1}$, as described in~\cite{BB}). Thus $x$ is characterized by the sequence \[\lambda(x)=(\lambda_{n-1}(x),\dots,\lambda_1(x))=(|v_{n-1}|,\dots,|v_1|)\in[0,n-1]\times[0,n-2]\times\cdots\times[0,1].\] Armstrong~\cite{Armstrong} defined a general class of \emph{sorting orders} on a Coxeter system $(W,S)$: one fixes $w\in W$ and chooses a reduced word $\omega$ (the ``sorting word'') for $w\in W$, then partially orders all group elements expressible as a subword of~$\omega$ by inclusion between their lexicographically first such expressions. Armstrong proved that for every reduced word for the top element of a finite Coxeter group, the sorting order is a distributive lattice intermediate between the weak and Bruhat orders. In the case that $W=\mathfrak{S}_n$ and $\omega=\omega_n$, the sorting order is equivalent to comparing $\lambda(x)$ and $\lambda(y)$ componentwise, hence is isomorphic to $C_2\times\cdots\times C_n$, where $C_i$ denotes a chain with $i$ elements. \begin{prop} \label{v-reduction} Let $x,y\in\mathfrak{S}_n$ with $x(n)=y(n)=n$, and let $v=s_j s_{j+1} \cdots s_{n-1}$ be a suffix of $s_1 \; \cdots \; s_{n-1}$. Then $x\unrhd_R y$ if and only $vx\unrhd_R vy$. \end{prop} \begin{proof} If $v=e$, there is nothing to prove. Otherwise, by~\eqref{RC}, it suffices to show that for every $i\in[n]$, we have \begin{subequations} \begin{equation} \label{RC-for-xy} [y(i),x(i)]\subseteq y[i,n] \end{equation} if and only if \begin{equation} \label{RC-for-v} [(vy)(i),(vx)(i)]\subseteq vy[i,n]. \end{equation} \end{subequations} This is clear if $i=n$, so we assume henceforth that $i\neq n$. Moreover, \[v(k)=\begin{cases} k &\text{ if } k<j,\\ k+1 &\text{ if } j\leq k<n,\\ j &\text{ if } k=n\end{cases} \qquad\text{and}\qquad v^{-1}(k)=\begin{cases} k &\text{ if } k<j,\\ n &\text{ if } k=j,\\ k-1 &\text{ if } k>j.\end{cases} \] In particular, if $i\neq n$, then $x(i)>y(i)$ if and only if $v(x(i))>v(y(i))$. We assume henceforth that these two equivalent conditions hold, since if both fail then~\eqref{RC-for-xy} and~\eqref{RC-for-v} are both trivially true. The proofs of the two directions now proceed very similarly. \medskip $\eqref{RC-for-xy}\implies\eqref{RC-for-v}$:\quad There are three cases. \emph{Case 1a: $j>x(i)$.} Then $v$ fixes $[1,x(i)]$ pointwise, so $[(vy)(i),(vx)(i)]=v[y(i),x(i)]\subseteq vy[i,n]$ (applying $v$ to both sides of~\eqref{RC-for-xy}). \emph{Case 1b: $y(i) < j \leq x(i)$.} Then $(v x)(i) = x(i) + 1$ and $(v y)(i) = y(i)$, so \begin{align*} [(vy)(i),(vx)(i)] &= [y(i),j-1]\cup\{j\}\cup[j+1,x(i)+1]\\ &= v[y(i),j-1]\cup\{v(n)\}\cup v[j,x(i)]\\ &= v\left( [y(i),x(i)]\cup \{y(n)\}\right)\\ &\subseteq vy[i,n] \end{align*} establishing~\eqref{RC-for-v}. \emph{Case 1c: $j \leq y(i)$.} Similarly to Case~1a, we have $[(vy)(i),(vx)(i)] = [y(i)+1,x(i)+1] = v[y(i),x(i)]\subseteq vy[i,n]$, as desired. $\eqref{RC-for-v}\implies\eqref{RC-for-xy}$:\quad Applying $v^{-1}$ to both sides of~\eqref{RC-for-v} gives $v^{-1}[vy(i),vx(i)]\subseteq y[i,n]$, so in order to prove~\eqref{RC-for-xy} It is enough to show that \begin{equation} \label{enough} [y(i),x(i)]\subseteq v^{-1}[vy(i),vx(i)] \end{equation} Moreover, the earlier assumption $i\neq n$ implies that $vx(i)\neq j$ and $vy(i)\neq j$. \emph{Case 2a: $j > vx(i)$.} Then $v^{-1}$ fixes the set $[1,vx(i)]$ pointwise, so in particular $[y(i),x(i)] = [vy(i),vx(i)] = v^{-1}[vy(i),vx(i)]$, establishing~\eqref{enough}. \emph{Case 2b: $vy(i) < j < vx(i)$.} Then $y(i) = vy(i)$ and $x(i) = vx(i)-1$, so \begin{align*} [y(i),x(i)] &= [vy(i),j-1] \cup [j,vx(i)-1] \\ &= v^{-1}[vy(i),vy(n)-1] \cup v^{-1}[vy(n)+1,vx(i)] \\ &\subseteq v^{-1}[vy(i),vx(i)]. \end{align*} \emph{Case 2c: $j < vy(i)$.} Then $[y(i),x(i)] = [vy(i)-1,vx(i)-1] = v^{-1}[vy(i),vx(i)]$, again implying~\eqref{enough}. \end{proof} \begin{theorem} \label{thm:same} The pseudoreachability order coincides with the bubble-sorting order. \end{theorem} \begin{proof} It suffices to show that the two partial orders have the same covering relations, i.e., that \[x\gtrdot_P y \quad\iff\quad \lambda(x)\gtrdot_C\lambda(y).\] We induct on $n$; the base case $n=1$ is trivial. Let $x,y\in\mathfrak{S}_n$ with $n>1$, and let their conormal forms be \[ x = u\bar x = (s_i\cdots s_{n-1}) \bar x,\qquad y = v\bar y = (s_j\cdots s_{n-1}) \bar y \] where $i=x(n)=n-\lambda_{n-1}(x)$ and $j=y(n)=n-\lambda_{n-1}(y)$. \medskip ($\impliedby$)\quad Suppose that $\lambda(x)\gtrdot_C\lambda(y)$. Then either $i=j-1$ or $i=j$. If $i=j-1$, then $\lambda(\bar x)=\lambda(\bar y)$, so $\bar x=\bar y$ and $x=s_iy$, which by Prop~\ref{lem:inv-1} implies $x\gtrdot_P y$. If $i=j$, then $\lambda(\bar x)\gtrdot_C\lambda(\bar y)$. Then $\bar x\gtrdot_P\bar y$ by induction, so $v\bar x\gtrdot_P v\bar y=y$ by Prop.~\ref{v-reduction}. \medskip ($\implies$)\quad Suppose that $x\gtrdot_Py$. Then $x \gtrdot_B y$ by Theorem~\ref{thm:bruhat}, so $i\leq j$ (as noted in the proof of Prop.~\ref{cf-df-facts}). If $i<j$, then $v$ is a proper suffix of $u$. By the definition of Bruhat order it must be the case that $x=yt_{a,b}$ for some $a<b$; in fact $b=n$ (otherwise $x(n)=y(n)$). Then $x(n)=y(a)$ and $x(a)=y(n)$, and $x(k)=y(k)$ for $k\not\in\{a,n\}$. Moreover, $y(a)<x(a)$ (since $x \gtrdot_B y$ and not vice versa). On the other hand, if $y(a)\leq x(a)-2$, so that $y(a)<c<x(a)=y(n)$ for some $c$, then by~\eqref{RC} $c=y(k)$ for some $k\in[a+1,n-1]$, and in particular $x$ has at least three more inversions than $y$ --- not only $(a,n)$, but also $(a,k)$ and $(k,n)$, which contradicts the assumption $x\gtrdot_Py$. Therefore $y(a)=x(a)-1$, i.e., $x(n)=y(n)-1$. We conclude that $x=s_iy$, so $\lambda(x)\gtrdot_C\lambda(y)$ using the conormal forms above. If $i=j$, then $u=v$, so $\bar x\gtrdot_P\bar y$ by Prop.~\ref{v-reduction}. By induction $\lambda(\bar x)\gtrdot_C\lambda(\bar y)$, and prepending $n-i$ gives $\lambda(x)\gtrdot_C\lambda(y)$ as well. \end{proof} \section{Pattern avoidance and reachability} \label{sec:avoid} In this section, we establish two sufficient conditions for reachability using pattern avoidance. (It is dubious whether pattern avoidance conditions can completely characterize reachability.) Let $\pi\in\mathfrak{S}_n$ and $\sigma\in\mathfrak{S}_m$, where $m\leq n$. A \defterm{$\sigma$-pattern} is a subsequence $\pi(i_1),\dots,\pi(i_m)$ in the same relative order as $\sigma$, i.e., such that $1\leq i_1<\cdots<i_m\leq n$ and $\pi(i_j)<\pi(i_k)$ if and only if $\sigma(j)<\sigma(k)$. If $\pi$ contains no $\sigma$-pattern then we say that $\pi$ \defterm{avoids} $\sigma$. \begin{theorem} \label{thm:213-avoiding} If $x\geq_By$ and $y$ avoids $213$, then $x\unrhd_R y$. \end{theorem} \begin{proof} Suppose that $x\geq_By$ and $y$ avoids $213$, but $x\mkern-1mu\not\mathrel{\mkern1mu\unrhd_R}\mkern1mu y$. Let $i$ be any index such that $\mathsf{d}_i(x,y) = 0$. By Prop.~\ref{cf-df-facts} we know that $1<i<n$ and that $y(i) < x(i)$. In particular, $m\neq i$, where $m = y^{-1}(x(i))$; that is, $y(m)=x(i)$. First, suppose that $m > i$. We claim that there exists some $u<i$ such that $y(i) < y(u) < y(m)$. Otherwise, $J_i\geq y(m)-y(i)+1$, and then $k=y(m)-y(i)$ has the properties $k<J_i$ and $y(i)+k=y(m)=x(i)$, so $k\in\mathsf{D}_i(x,y)$, contradicting the assumption $\mathsf{d}_i(x,y) = 0$. Therefore $y(u), y(i), y(m)$ is a 213-pattern. Second, suppose that $m < i$. If $y(k) > y(m)$ for some $k>i$, then $y(m),y(i),y(k)$ is a 213-pattern. On the other hand, suppose that $y(k) < y(m) = x(i)$ for all $k > i$ (hence for all $k \ge i$). Then \[\{k\in[i,n]:\ y(k)<x(i)\}=[i,n]~\supsetneq~[i+1,n]\supseteq\{k\in[i,n]:\ x(k)<x(i)\}\subseteq[i+1,n]\] so \begin{align*} \#\{k\in[i,n]:\ y(k)<x(i)\} &> \#\{k\in[i,n]:\ x(k)<x(i)\}\\ \therefore\quad \#\{k\in[1,i-1]:\ y(k)<x(i)\} &< \#\{k\in[1,i-1]:\ x(k)<x(i)\}\\ \therefore\quad \#\{k\in[1,i-1]:\ y(k)\geq x(i)\} &> \#\{k\in[1,i-1]:\ x(k)\geq x(i)\}. \end{align*} That is, $y\langle i-1,x(i)\rangle>x\langle i-1,x(i)\rangle$, contradicting the assumption $x\geq_By$. \end{proof} Theorem~\ref{thm:213-avoiding} partially answers the question of when the converse of Theorem~\ref{thm:bruhat} holds, i.e., which Bruhat relations are also relations in pseudoreachability order. We next study if there is an analogous condition on $x$, rather than $y$, that suffices for reachability. One such condition that allows us to restrict $x$ instead of $y$ is to ensure that only very few entries $x(i)$ are large with respect to $i$. \begin{lemma} \label{lem:x1} Let $x\in\mathfrak{S}_n$. The following conditions are equivalent: \begin{enumerate} \item $x^{-1}(i) \leq i+1$ for all $i\in[n]$. \item $x\langle j,j\rangle = 1$ for all $j \in [n]$. \item $x$ avoids both 231 and 321. \item $x$ is of the form $s_{i_1}\cdots s_{i_k}$, where $n-1\geq i_1>\cdots>i_k\geq1$. \end{enumerate} \end{lemma} The number of these permutations is $2^{n-1}$, which is easiest to see from condition~(4). Conditions~(1) and~(3) were mentioned respectively by J.~Arndt (June 24, 2009) and M.~Riehl (August 5, 2014) respectively in the comments on sequence A000079 in \cite{OEIS}. Accordingly, we will call a permutation satisfying the condition of Lemma~\ref{lem:x1} an \defterm{AR permutation} (for Arndt--Riehl). \begin{proof} $(1)\iff(2)$: Formula~\eqref{angle-brackets} implies that \begin{align*} \forall j\in[n]:\ x\langle j,j\rangle = 1 &\iff \forall j\in[n]:\ [1,j-1]\subseteq x[1,j]\\ &\iff \forall j\in[n]:\ x^{-1}[1,j-1]\subseteq [1,j]\\ &\iff \forall i\in[n]:\ x^{-1}[1,i]\subseteq [1,i+1] \end{align*} since the last two statements differ only by the trivially true cases $i=0$ and $i=n$. $(3)\iff(1)$: Condition~(3) holds if and only if no digit $i\in[n]$ occurs later than position $i+1$, but this is precisely condition (1). $(4)\iff(1)/(3)$: Let $Y_n$ be the set of permutations in $\mathfrak{S}_n$ satisfying the equivalent conditions~(1) and~(3), and let $Z_n$ be the set satisfying condition (4). For $n\leq2$ we evidently have $Y_n=Z_n=\mathfrak{S}_n$. For $n\geq 3$, we proceed by induction. Observe that $Z_n=Z_{n-1}\cup s_{n-1}Z_{n-1}$, and that left-multiplication by $s_{n-1}$ (i.e., swapping the locations of $n-1$ and $n$) does not affect condition (1), which is always true for $i\in\{n-1,n\}$. Therefore $Z_n\subseteq Y_n$. On the other hand, if $w\in Y_n$ then $w_n\in\{n-1,n\}$, otherwise $w_n$, together with the digits $n-1$ and $n$, would form a 231- or 321-pattern. Therefore, $w'_n=n$, where either $w'=w$ or $w'=s_{n-1}w$. By induction $w'\in Z_{n-1}$, so $w\in Z_n$ as desired. \end{proof} \begin{corollary} \label{cor:arn-bru} If $x$ is AR and $y\leq_Bx$, then $y$ is AR as well. \end{corollary} \begin{proof} Lemma \ref{lem:x1} asserts that $x\langle i,i\rangle = 1$ for all $i \in [n]$. Since $y\leq_Bx$, $y\langle i,i\rangle = 1$ or $0$, but the latter could not happen by the pigeonhole principle. \end{proof} An \defterm{exceedance} of a permutation $x\in\mathfrak{S}_n$ is an index $k\in[n]$ such that $x(k)>k$. \begin{lemma} \label{lem:x2} Let $x \in \mathfrak{S}_{n}$ be an AR permutation. Suppose that $k$ is an exceedance of $x$, and let $i=x(k)$. Then $x(j)=j-1$ for all $j\in[k+1,i]$. \end{lemma} \begin{proof} The argument of Lemma~\ref{lem:x1} implies that $[1,k-1]\subseteq x[1,k]$; however, since $x(k)>k$ we have in fact $[1,k-1]=x[1,k-1]$. Now let $j\in[k+1,i]$. Lemma~\ref{lem:x1} also asserts that $x\langle j,j\rangle=\#A_j=1$, where $A_j=\{m\in[j]:\ x(m)\geq j\}$. Certainly $k\in A_j$, so $j\not\in A_j$, that is, $x(j)<j$. But since $x(j)\geq k$ for each such $j$, we can infer in turn that $x(k+1)=k$, $x(k+2)=k+1$, \dots, $x(i)=i-1$. \end{proof} \begin{theorem} \label{thm:x} If $x\geq_By$ and $x$ is AR, then $x\unrhd_R y$. \end{theorem} \begin{proof} Suppose that $x\geq_By$ and $x$ is AR, but $x\mkern-1mu\not\mathrel{\mkern1mu\unrhd_R}\mkern1mu y$. Let $i$ be some index such that $\mathsf{d}_{i}(x, y) = 0$. By~\eqref{RC}, there exists $j < i$ such that \begin{equation} \label{banana} y(i) < y(j) \le x(i). \end{equation} By Lemma \ref{lem:x1}, $x\langle i,i\rangle = 1$; that is, there exists some (unique) $k \leq i$ such that $x(k) \ge i$. First, suppose that $k=i$. Then $x\langle i-1,i\rangle = 0$, and $y\langle i-1,i\rangle = 0$ as well because $y \leq_B x$. Hence $y[1,i-1]=[i-1]$. But then~\eqref{banana} implies that $y(i)<y(j)\leq i-1$ as well, a contradiction. Second, suppose that $k < i$. Then $y(i)<x(i)<i$ by Lemma~\ref{lem:x2}, so $y(i)\leq i-2$. Set $p=y(i)$; then $y^{-1}(p)=i\geq k+2$. But then $y$ is not AR, which violates Corollary~\ref{cor:arn-bru}. \end{proof} \bibliographystyle{plain}

https://arxiv.org/abs/1202.2475

On the speed of convergence of Newton's method for complex polynomials

We investigate Newton's method for complex polynomials of arbitrary degree $d$, normalized so that all their roots are in the unit disk. For each degree $d$, we give an explicit set $\mathcal{S}_d$ of $3.33d\log^2 d(1 + o(1))$ points with the following universal property: for every normalized polynomial of degree $d$ there are $d$ starting points in $\mathcal{S}_d$ whose Newton iterations find all the roots with a low number of iterations: if the roots are uniformly and independently distributed, we show that with probability at least $1-2/d$ the number of iterations for these $d$ starting points to reach all roots with precision $\varepsilon$ is $O(d^2\log^4 d + d\log|\log \varepsilon|)$. This is an improvement of an earlier result in \cite{Schleicher}, where the number of iterations is shown to be $O(d^4\log^2 d + d^3\log^2d|\log \varepsilon|)$ in the worst case (allowing multiple roots) and $O(d^3\log^2 d(\log d + \log \delta) + d\log|\log \varepsilon|)$ for well-separated (so-called $\delta$-separated) roots.Our result is almost optimal for this kind of starting points in the sense that the number of iterations can never be smaller than $O(d^2)$ for fixed $\varepsilon$.

\section{Introduction} Newton's root finding method is an old and classical method for finding roots of a differentiable function; it goes back to Newton in the 18th century, perhaps earlier. It was one of the main reasons why A.\ Douady, J.\ Hubbard and others in the late 1970s studied iterations of complex analytic functions. The main question was to know where to start the Newton iterative method in order to converge to the roots of the polynomial. Newton's method is known as rapidly converging near the roots (usually with quadratic convergence), but had a reputation that its global dynamics was difficult to understand, so that in practice other methods for root finding were used. See \cite{JohannesNewtonSurvey} for an overview on recent results about Newton's method. Meanwhile, some small sets of good starting points are known: there are explicit deterministic sets with $O(d\log^2d)$ points that are guaranteed to find all roots of appropriately normalized polynomials of degree $d$ \cite{HSS}, and probabilistic sets with as few as $O(d(\log\log d)^2)$ points \cite{BLS}. We are interested in the question how many iterations are required until all roots are found with prescribed precision $\varepsilon$. In \cite{D}, it is shown that among a set of starting points as specified above, there are $d$ points that converge to the $d$ roots and require at most $O(d^4\log^2d+d^3\log^2d|\log\varepsilon|)$ to get $\varepsilon$-close to the $d$ roots in the worst case; for randomly placed roots (or for roots at mutual distance at least $\delta$ for some $\delta>0$), the required number of iterations is no more than $O(d^3\log^3d+d\log|\log\varepsilon|)$ (with the constant depending on $\delta$). This is about one power of $d$ away from the best possible bounds. In this paper, we show that Newton's method is about as fast as theoretically possible. We consider the space of polynomials of degree $d$, normalized so as to have all roots in the complex unit disk $\mathbb D$. Our main result is the following. \begin{theorem}[Quadratic Convergence in Expected Case] \label{Thm:Main} For every degree $d$, there is an explicit universal set $\mathcal{S}_d$ of points in ${\mathbb C}$, with $|\mathcal{S}_d|=3.33d\log^2d(1+o(1))$, with the following property: suppose that $\alpha_1, \ldots, \alpha_n$ are uniformly and independently distributed in the unit disk and put $p(z)= \prod_{j=1}^d (z-\alpha_j)$. Then there are $d$ starting points in $\mathcal{S}_d$ such that with probability $p_d \rightarrow 1$ as $d \rightarrow \infty$, the number of iterations needed to approximate all $d$ roots with precision $\varepsilon$ starting at these $d$ points is \[ O(d^2\log^4 d + d\log|\log \varepsilon|). \] \end{theorem} \begin{remark} As stated, the theorem deals with $d$ distinguishable (i.e., ordered) roots and their associated probability distribution. We prove that the same result holds if we identify our polynomials in terms of their sets of \emph{indistinguishable} roots, as two polynomials $p(z)= \prod_{j=1}^d (z-\alpha_j)$ and $q(z)= \prod_{j=1}^d (z-\beta_j)$ are the same if their unordered sets of roots $\{\alpha_1,\ldots, \alpha_d\}$ and $\{\beta_1,\ldots, \beta_d\}$ are equal. \end{remark} \begin{remark} This bound on the number of iterations is optimal in the sense that there is no bound on the number of iterations in the same generality that for fixed $\varepsilon$ has asymptotics in $o(d^2)$, so we are away from the best possible bound only by a factor of about $O(\log ^4d)$. \end{remark} \section{Good starting points for Newton's method} \label{sec: prelim results} Studying the geometry of the immediate basins outside the unit disk $\mathbb{D}$, in \cite{HSS} we proved the existence of a universal starting set with $1.11 d\log^2 d$ points depending only on $d$ such that for every polynomial of degree $d$ with all roots in the unit disk, and for every root, there is a point in the set which is in the immediate basin of this root. Enlarging the set by a factor of 3 approximately, in \cite{D} we obtained a set of starting points $\mathcal{S}_d$ which ensured that for each polynomial $p$ and each root $\alpha$ there is a point $z$ in $\mathcal{S}_d$ intersecting the immediate basin $U$ of $\alpha$ in the ``middle third`` of the ``thickest'' \textit{channel}, where a channel is an unbounded connected component of $U \setminus \overline{\mathbb{D}}$. Being in this middle third implies an upper bound on the displacement $d_U(z, N_p(z))$ in terms of the Poincar\'e metric of the immediate basin. It also turns out that the orbit of $z$ under iteration of Newton map does not leave $D_R(0)$, the disk of radius $R$ centered at the origin for some bounded value of $R$. We will refer to such points as having \textit{$R$-central orbits}. More precisely, let $\mathcal{S}_d$ be defined as follows. \begin{definition}[Efficient Grid of Starting Points] For each degree $d$, construct a circular grid $\mathcal{S}_d$ as follows. For $k = 1, \ldots, s = \lceil 0.4\log d\rceil$, set $$r_k = (1+\sqrt{2})\left(\frac{d-1}{d}\right)^\frac{2k-1}{4s},$$ and for each circle around 0 of radius $r_k$, choose $\lceil 8.33d\log d\rceil$ equidistant points (independently for all the circles). \end{definition} The set $\mathcal{S}_d$ thus constructed has $3.33(1+o(1))d\log^2 d$ points. The following theorem is proven in \cite[Theorem~8]{D}. \begin{theorem}\label{thm: set of starting points} For each degree $d$, the set $\mathcal{S}_d$ has the following universal property. If $p$ is any complex polynomial, normalized so that all its roots are in $\mathbb{D}$, then there are $d$ points $z^{(1)}, \ldots, z^{(d)}$ in $\mathcal{S}_d$ whose Newton iterations converge to the $d$ roots of $p$. If $\alpha$ is a root of $p$ and $U$ is the immediate basin of $\alpha$, then there is an index $i$ such that $z^{(i)}\in U$ with $d_U(z^{(i)}, N_p(z^{(i)}))< 2\log d$. In addition, $z^{(1)}, \ldots, z^{(d)}$ have $R$-central orbits for $$R \leq 5\left(\frac{d}{d-1}\right)^{\lceil 5\pi(\log d + 1)\rceil}.$$ \end{theorem} For $d = 100$, we have $R < 14;$ for $d = 1000, $ we have $R < 7.5;$ and asymptotically the upper bound on $R$ tends to $7$. The result provides an upper bound for $R$ that is uniform in $d$. This set of starting points will be the basis for the discussion which follows. \section{Uniformly distributed roots}\label{sec: result} In this manuscript we investigate the Newton map for complex polynomials with randomly distributed roots. In this section, we fix notation and give the strategy of the proof of our main result, Theorem~\ref{Thm:Main}. Let $\alpha$ be a simple root of the polynomial $p$ of degree $d$ and $U$ be the immediate basin of attraction of $\alpha$. By the discussion in the previous section, there exists $z_1\in \mathcal{S}_d$ with $R$-central orbit in $U$, i.e.\ under iteration of the Newton map $N_p$ the orbit converges to $\alpha$ and stays within $D_R(0)$. Let $z_{n+1} := N_p(z_{n})$ for $n\geq 1$. For any two consecutive points $z_n$ and $z_{n+1}$ along the orbit of $z_1$, in \cite[Section~4]{D} we constructed ``thick'' curves that, roughly speaking, ``use up'' area at least $|z_n - z_{n+1}|^2/(2\tau)$ with $\tau := d_U(z_1, z_2) < 2\log d$. In the region of quadratic convergence (near the root $\alpha$), only $\log_2|\log_2 \varepsilon - 5|$ iterations are sufficient to get $\varepsilon$-close to the root. Outside this region, two such curves with base points $z_n$ and $z_{n'}$ are disjoint if $n' - n \geq 2\tau + 6$ \cite[Lemma~11]{D}. The bound $O(d^3\log^3d + d\log|\log \varepsilon|)$ follows from lower bounds on the displacements $|z_n - z_{n+1}|$ along the orbit. The main improvement in this paper is on the lower bounds on the displacements when the roots are randomly distributed. As in \cite{D}, we partition $D_{R}(0)$ (the disk of radius $R$ centered at the origin) into domains \[ S_k := \left\{z\in D_{R}(0): \min_j|z-\alpha_j| \in \left(2^{-(k+1)}, 2^{-k}\right]\right\}, \ k \in \mathbb{Z} \;. \] It turns out that if the roots are randomly distributed in the unit disk, then with high probability the following holds: there exists a universal constant $C$ such that for every $n$ we have the following estimates \[ |z_n - z_{n+1}| \ge \left\{ \begin{array}{ll} \frac{C}{d\log d} & \mbox{if $z_n\in S_k$ with $k \leq \log_2 d$};\\ \frac{C}{2^k k } & \mbox{otherwise}.\end{array} \right. \] If $z_n\in S_k$ with $k\le \log_2d$, then we say that we are ``in the far case'', as $z_n$ is far from all the roots. Since each such iteration uses an area of at least $|z_n-z_{n+1}|^2/(2\tau)$, and one in $2\tau+6$ such areas are disjoint, the total number of orbit points in the far case is bounded by $O(d^2\log^4d)$. \cite[Lemma~16]{D} says that if the orbit gets very close to some root in comparison to the other roots, then it has entered the region of quadratic convergence of that root where only $\log_2|\log_2 \varepsilon - 5|$ are sufficient to approximate it within an $\varepsilon$-neighborhood. We call this ``the near case''. For randomly distributed roots, the mutual distance between roots is large enough so away from the region of quadratic convergence, we only need to consider $k \le 3+(2+\eta)\log_2 d$ for a certain $\eta>0$. We define the ``intermediate case'' as those $z_n\in S_k$ with $\log_2d< k\le 3+(2+\eta)\log_2 d$. Each domain $S_k$ has area $O(d 4^{-k})$, and each iteration with $z_n\in S_k$ uses area about $(C/2^kk)^2/2\tau\approx C^2/4^kk^2\tau$, so the number of orbit points in the intermediate case is at most $O(dk^2\tau)$ for each $k$, times the usual factor $2\tau+6$ to make the areas disjoint. But $\log_2d<k\leq 3+(2+\eta)\log_2 d$ and $\tau=O(\log d)$, so the total number of iterations in the intermediate case is $O(d\log^5d)$. In the subsequent sections we will make these arguments precise. \subsection{On the distribution of the roots}\label{sec: root distribution} In order to get a lower bound on the expected displacement, we will first investigate the distribution of the roots. We will be interested in two different kind of probability spaces. The first space $\mathcal{P}_d =\{(x_1, \ldots, x_d) : x_i\in \mathbb{D}\}$ consists of all polynomials with $d$ \textit{distinguishable} roots in the unit disk, normalized so as to have leading coefficients $1$, and the probability measure is induced by the Lebesgue measure on $\mathbb{D}^d$. The second space $\mathcal P_d/S_d$ consists of all polynomials with \textit{indistinguishable} roots in the unit disk, i.e.\ the quotient probability space of the standard action of the symmetric group $S_d$ on $\mathcal{P}_d$ defined by permuting the roots. The following lemma is the probabilistic ingredient of the main theorem. It certainly isn't new, but easier verified than looked up in the library. \begin{lemma}[Base-$d$ numbers]\label{lemma: numbers} Let $M_d$ be the set of all $d$-digit numbers in base $d$. (a) The probability that that a randomly chosen number $a\in M_d$ does not have a digit repeating more than $O(\log d)$ times is at least $1-1/d$. (b) Let $\sim$ be an equivalence relation on $M_d$ defined as follows: $a \sim b \Leftrightarrow \exists \sigma \in S_d$ with $a = \sigma b$, i.e.\ two elements are equivalent if they have the same sets of digits counted with multiplicities. Then the probability that a randomly chosen element $[a]\in M_d/\sim$ does not have a digit repeating more than $O(\log d)$ is at least $1-1/d$. \end{lemma} \begin{proof} (a) For fixed $i$, the number of $d$-digit numbers which contain at least $\alpha$ digits $i$ is at most $\binom{d}{\alpha}d^{d-\alpha}$. Thus the number of $d$-digit numbers which contain a symbol repeating at least $\alpha$ times is at most $$d \binom{d}{\alpha}d^{d-\alpha} < \frac{d}{\alpha !} d^{d}.$$ So the probability that a randomly selected number in $M_d$ contains at least $\alpha$ identical digits is at most $\frac{d}{\alpha !}$ since $|M_d| = d^d$. Therefore, with probability at least $1-\frac{d}{\alpha!}$, a randomly selected number in $M_d$ does not have a digit repeating more than $\alpha$ times. Note that if $\alpha!\ge d^2$ we have $1-\frac{d}{\alpha!} \geq 1 - \frac{1}{d}$. Therefore, by taking $\alpha$ such that $(\alpha - 1)! < d^2 \leq \alpha !$ (which implies that $\alpha \in O(\log d)$), we prove the first part of the claim. (b) Note that the elements of $ M_d/\sim$ can be bijectively mapped to the set $\hbox{Mult}_d = \{(x_0, \ldots, x_{d-1}) : x_i \in \mathbb{Z}_{\geq 0}, x_0 + \ldots + x_{d-1} = d\}$ as follows: for $[a]\in M_d/\sim$ let $x_i$ be the multiplicity of digit $i$ in every $a\in [a]$. It is well known and easy to see that \begin{equation}\label{eq: multiset coefficient} \left|\left\{\rule{0pt}{10pt}(x_0, \ldots, x_{r-1}) : x_i\in \mathbb{Z}_{\geq 0}, x_0 + x_1 + \ldots + x_{r-1} = n\right\}\right| = \binom{n + r - 1}{r - 1} \;. \end{equation} Thus we have $|\hbox{Mult}_d| = \binom{2d - 1}{d-1}$. On the other hand, the number of elements in $\hbox{Mult}_d$ with first component at least $\alpha$ is equal to the cardinality of \[ \left\{(x_2, \ldots, x_d) : x_i \in \mathbb{Z}_{\geq 0}, x_2 + \ldots + x_d \leq d-\alpha\right\} \] which has the same cardinality as \[ \{(y_1, x_2, \ldots, x_d) : y_1, x_i \in \mathbb{Z}_{\geq 0}, y_1 + x_2 + \ldots + x_d = d-\alpha\} \;. \] Again by (\ref{eq: multiset coefficient}) this quantity equals to $\binom{2d - \alpha - 1}{d-1}$. Therefore, the number of elements of $\hbox{Mult}_d$ with a component at least $\alpha$, i.e. the number of elements of $M_d/\sim$ with a digit repeating at least $\alpha$ times, is at most \[ d\binom{2d - \alpha - 1}{d-1}\;. \] Hence the probability that a number of $M_d/\sim$ has a digit repeating at least $\alpha$ times is at most $$\frac{d\binom{2d - \alpha - 1}{d-1}}{\binom{2d - 1}{d-1}} = \frac{d (2d - \alpha - 1)! (d-1)! d!}{(2d-1)! (d-1)! (d - \alpha)!} = $$ $$ = d\frac{(d-\alpha + 1)(d-\alpha + 2)\cdots d}{(2d - \alpha)(2d - \alpha + 1)\cdots (2d - 1)} \leq d \left(\frac{1}{2}\right)^{\alpha - 1}\frac{d}{2d - 1}.$$ Hence for $\alpha = \lceil 2\log_2 d + 1 \rceil\in O(\log d)$ the second part of the claim follows. \end{proof} \begin{remark} The statement remains true if we replace the probability $1-1/d$ by $1-1/d^c$ for any constant $c\geq 1$. \end{remark} If the roots are randomly distributed in the unit disk one should expect that the number of roots in a region is proportional to its area. The previous claim easily implies the following statement. \begin{lemma}\label{lemma:disk} Let a polynomial with roots $x_1, \ldots, x_d$ be randomly chosen in $\mathcal{P}_d$ or $\mathcal{P}_d/S_d$. Then there exists a constant $C_d \in O(\log d)$ such that with probability at least $1-1/d$ the following holds true: every disk in ${\mathbb C}$ with area $A = \pi r^2$ contains at most $k(A)$ points among $x_1, \ldots, x_d$ with $$k(A) = \left\{ \begin{array}{ll} C_ddA & \mbox{if $A \geq 1/d$}; \\ C_d & \mbox{otherwise}.\end{array} \right.$$ \end{lemma} \begin{proof} Without loss of generality we can assume that $d = (2k+1)^2$ for an integer $k$ (the general case follows by enlarging $C_{(2k+1)^2}$ by a bounded factor). Then the unit disk can be subdivided into $d$ pieces as follows (compare Fig.~\ref{fig: circle partitioned}): the first piece is a disk with center $0$ and radius $r_0 = 1/\sqrt{d};$ next, consider the annuli $A_s$ bounded between circles around $0$ of radii $(2s-1)r_0$ and $(2s+1)r_0$ for $s = 1, \ldots k$, and subdivide each annulus $A_s$ into exactly $8s$ pieces of equal area by drawing $8s$ radial segments. Thus we construct exactly $d$ pieces with equal area and diameters comparable with $r_0$. By Lemma \ref{lemma: numbers} it follows that each of the pieces contains at most $O(\log d)$ of the points $x_1, \ldots, x_d$ with probability at least $1-1/d$ (in both cases of distinguishable and indistinguishable roots): in the case of distinguishable roots, the $i$-th digit of a $d$-digit number specifies the number of the piece containing the $i$-th root; in the other case, the same symmetries apply on both sides of the equality. \begin{figure} \begin{center} \framebox{\includegraphics[width=0.5\textwidth]{DiskPartition1a.pdf}} \caption{Partition of the unit disk into smaller pieces of similar sizes.\label{fig: circle partitioned}} \end{center} \end{figure} Hence, the claim is true for that particular partition of the unit disk. This implies the general claim as follows. It is easy to see that each square of side length at most $r_0$ in the complex plane can intersect at most a constant number $C'$ of these pieces, where $C'$ does not depend on $d$. Consider a square $S$ for which the unit circle is inscribed, for example the one with sides parallel to the real and imaginary axes. Subdivide it into $d$ equal squares of side length $r_0 $ (using the fact that $d$ is a square and $r_0=1/\sqrt d$). Then each of these smaller squares will intersect at most $C'$ pieces from the partition of the unit disk (some squares will not intersect any). Therefore each of the small squares contains at most $C''\log d$ points for some constant $C''$ which does not depend on $d$. Since each square of side length $r_0$ (possibly rotated) intersects at most $9$ of these squares dividing $S$, we conclude that each square of side length $r_0$ contains, with probability at least $1-1/d$, at most $C\log d$ of the points $x_1,\ldots, x_d$ for $C = 9 C''$. If we group every 4 neighboring small squares (of side length $r_0$) and repeat the argument, we get that each square of side length $2r_0$ contains at most $4 C\log d$ points, and so on for squares of side length $4r_0, 8r_0,\ldots $. Thus an arbitrary square of side length $x\in[2^kr_0,2^{k+1}r_0]$ contains at most $2^{2k + 2} C\log d\le 4C(x^2/r_0^2)\log d \approx 4Cx^2d\log d$ points since it is contained some square of side length $2^{k+1}r_0$, and thus enlarging the constant by a factor of 4 the lemma will hold true for squares. Since each disk of radius $r$ is contained in a square of side length $2r$, the bound on the number of points in an arbitrary disk follows. \end{proof} Now we prove the following claim about the mutual distance for randomly distributed points in the unit disk. \begin{lemma}\label{lemma: mutual distance} Let the polynomial $p(z)$ be randomly chosen in $\mathcal{P}_d$ or $\mathcal{P}_d/S_d$. Then the mutual distance between any pair of its roots is at least ${1}/{d^{1+\eta}}$, for any fixed $\eta>0$, with probability at least $1-1/d^{2\eta}$. \end{lemma} \begin{proof} First, note that the claim for a randomly chosen polynomial in $\mathcal{P}_d/S_d$ follows from the claim for a randomly chosen polynomial in $\mathcal{P}_d$. Choosing randomly a polynomial in $\mathcal{P}_d$ is equivalent to choosing randomly and independently its roots. For a positive number $r$, the probability $p_{d,r}$ that $d$ uniformly and independently distributed points in the unit disk have mutual distance at least $r$ is at least \[ p_{d,r}\geq (1 - r^2)(1-2r^2)\ldots (1-(d-1)r^2) \] (the unit disk has area $\pi$, and after $k$ roots are selected, the $k+1$-st root must avoid an area of at most $k\pi r^2$; this has probability $(\pi-\pi kr^2)/\pi=1-kr^2)$. Since $\log(1+x) \ge x/(1+x)$ for $x>-1$, we get \[ \log p_{d,r}\ge \sum_{k=1}^{d-1} \log(1-k r^2) \ge \sum_{k = 1}^{d-1} \frac{-k r^2}{1-k r^2} \ge -r^2 \frac{\sum_{k = 1}^{d-1} k}{1-d r^2} \ge -r^2 \frac{d^2/2}{1-d r^2} \ge -d^2 r^2 \;, \] where the last inequality holds if $d r^2 < 1/2.$ Hence \[ p_{d, r}\ge \exp(-d^2 r^2)\ge 1 - d^2 r^2 \;. \] If $r = 1/d^{1+\eta}$ (which satisfies $d r^2 < 1/2$), then $p_{d,r}\ge 1-1/d^{2\eta}$ and thus the claim follows. \end{proof} \begin{remark} The precise value of $\eta$ in this lemma is not very important to us, as eventually it will only affect constants in the bounds we obtain. \end{remark} An immediate corollary of the lemma is an upper bound on the distance of $z_n$ to the closest root which guarantees quadratic convergence. \begin{corollary}\label{cor: bound on K} If the $d$ roots are randomly chosen and $z_n \in S_k$ with $2^{-k} < {1}/{8d^{2+\eta}}$ and $\eta>0$, then with probability at least $1-1/d^{2\eta}$ the orbit of $z_n$ converges to the closest root $\alpha$, and $\log_2|\log_2\varepsilon - 5|$ iterations of $z_n$ are sufficient to get $\varepsilon$-close to $\alpha$. \end{corollary} \begin{proof} Indeed, if $z_n\in S_k$ and $\alpha$ is the closest root to $z_n$, then $|z_n - \alpha| < {1}/{8d^{2+\eta}}$ and for every root $\alpha_j\neq\alpha$ we have \[ |z_n-\alpha_j| \geq |\alpha - \alpha_j| - |\alpha - z_n| > 1/ d^{1+\eta}-1/8d^{2+\eta} \ge (8d+1)/8d^{2+\eta} > (4d + 3)|z_n - \alpha| \] (under the conditions of Lemma~\ref{lemma: mutual distance}). Therefore by \cite[Lemma~16]{D}, we need no more than $\log_2|\log_2\varepsilon - 5|$ iterations to get $\varepsilon$-close to $\alpha$. \end{proof} We combine the previous two lemmas in the following claim. \begin{lemma}\label{lemma: probabilistic conditions} Let a polynomial with roots $x_1, \ldots, x_d$ be randomly chosen in $\mathcal{P}_d$ or $\mathcal{P}_d/S_d$. Then with probability $p_d\ge 1-O(d^{-2\eta})$ (for fixed $\eta\in(0,1/2)$), the following two statements simultaneously hold true: \begin{description} \item[Area Condition (AC)] There exists $C_d\in O(\log d)$ such that every disk in ${\mathbb C}$ with area $A = \pi r^2$ contains at most $k(A)$ roots among $x_1, \ldots, x_d$ with $$k(A) = \left\{ \begin{array}{ll} C_ddA & \mbox{if $A \geq 1/d$};\\ C_d & \mbox{otherwise}.\end{array} \right.$$ \item[Distance Condition (DC)] The mutual distance between any pair of roots is at least ${1}/{d^{1+\eta}}$. \end{description} \end{lemma} \begin{proof} We are interested in $P(AC = \texttt{true} \mbox{ and } DC = \texttt{true})$, which equals \[ 1 - P(AC = \texttt{false} \ \ \hbox{or}\ \ DC = \texttt{false}) \geq 1- P(AC = \texttt{false}) - P(DC = \texttt{false}). \] By Lemma \ref{lemma:disk} we have $P(AC = \texttt{false}) \leq 1/d$, and by Lemma \ref{lemma: mutual distance} $P(DC = \texttt{false})\le 1/d^{2\eta}$. Hence the claim follows. \end{proof} \subsection{Proof of the main theorem}\label{sec: proof} In this section we will use the two conditions \textbf{AC} and \textbf{DC} to prove Theorem \ref{Thm:Main}. While \textbf{DC} guarantees that proximity to a root implies fast convergence (Corollary \ref{cor: bound on K}), \textbf{AC} gives a lower bound on the displacements along an orbit far away from the roots. More precisely, the following statement holds true. \begin{lemma}\label{lemma: main} Suppose that the Area Condition in Lemma \ref{lemma: probabilistic conditions} holds true. If $z_n\in S_K \cap \mathbb{D}_2(0)$, then \[ |z_n - z_{n+1}|\geq \frac{1}{(1 + 2C_d)2^{K+1} + 16\pi C_dd } \;, \] where $C_d\in O(\log d)$. If $z_n \not \in \mathbb{D}_2(0)$, then $|z_n - z_{n+1}| > 1/d$. \end{lemma} \begin{proof} The fact that $z_n\in S_K$ means that the closest root, say $\alpha$, is at distance ${c}/{2^K}$ for some $c\in (0.5, 1]$, and all the other roots satisfy $ |z_n-\alpha_j|\geq {c}/{2^K}$. First suppose that $z_n\in S_K \cap \mathbb{D}_2(0)$. This implies that $K\geq -2$. Let $T_k := \{z\in \mathbb{C}: 2^{-k-1} < |z-z_n| \leq 2^{-k}\}$ for $k = -2, \ldots, K$. Then all the roots are contained in $\bigcup_{k = -2}^{K}T_k$. The Area Condition implies that there exists a constant $C_d\in O(\log d)$ such that the number of roots in $T_k$ is bounded by $\pi C_d d4^{-k}$ for $\pi 4^{-k}\geq 1/d$, and by $C_d$ otherwise. Thus we have \begin{align} \left|\sum\frac{1}{z_n-\alpha_j}\right| &\leq \left|\frac{1}{z_n-\alpha}\right| + \sum_{\alpha_j \not= \alpha}\left|\frac{1}{z_n-\alpha_j}\right| = \frac{2^K}{c} + \sum_{k = -2}^{K}\sum_{\alpha_j \not= \alpha \atop \alpha_j \in T_k}\left|\frac{1}{z_n-\alpha_j}\right| \nonumber \\ &\leq \frac{2^K}{c} + \sum_{k = -2}^{\lfloor 0.5\log_2 \pi d\rfloor}\sum_{\alpha_j \not= \alpha \atop \alpha_j \in T_k}\left|\frac{1}{z_n-\alpha_j}\right| + \sum_{k = 1+\lfloor 0.5\log_2\pi d\rfloor}^{K}\sum_{\alpha_j \not= \alpha \atop \alpha_j \in T_k}\left|\frac{1}{z_n-\alpha_j}\right| \nonumber \\ &\leq \frac{2^K}{c} + \sum_{k = -2}^{\lfloor 0.5\log_2 \pi d\rfloor} \frac{2\pi C_d d4^{-k}}{2^{-k}} + \sum_{k = 1+\lfloor 0.5\log_2\pi d\rfloor}^{K} C_d 2^{k+1} \nonumber \\ &\leq 2^{K+1} + 16\pi C_dd + C_d 2^{K+2}\;. \nonumber \end{align} Therefore \[ |z_n - z_{n+1}| = \frac{1}{\left|\sum\frac{1}{z_n-\alpha_j}\right|}\geq \frac{1}{(1 + 2C_d)2^{K+1} + 16\pi C_dd } \;. \] For the case $z_n \not \in \mathbb{D}_2(0)$ we have \[ |z_n - z_{n+1}|^{-1} = \left|\sum\frac{1}{z_n-\alpha_j}\right| < \sum_{\alpha_j}1 = d \;,\] and so $|z_n - z_{n+1}| > 1/d$. \end{proof} \begin{corollary}\label{cor: displacement} Suppose that the Area Condition in Lemma \ref{lemma: probabilistic conditions} holds true. Then there is a universal constant $C$ such that the following statements hold for any $z_n \in S_k$. \begin{enumerate} \item If $2^{-k} \geq 1/d$, then $|z_n - z_{n+1}| \geq \frac{C}{d\log d}$. \item If $1/8d^{2+\eta}\leq 2^{-k} < 1/d$, then $|z_n - z_{n+1}| \geq\frac{C}{k2^k }$. \end{enumerate} \end{corollary} \begin{proof} For $z_n\in \mathbb{D}_2(0)$, Lemma \ref{lemma: main} gives \[ |z_n - z_{n+1}| \geq \frac{1}{(1 + 2C_d)2^{k+1} + 16\pi C_dd } \;. \] If $2^{-k} \geq 1/d$, i.e.,\ $2^{k+1} \leq 2d$, the denominator is at most $O(C_d d)$, so the displacement is at least $C'/(d\log d)$ for some universal constant $C'$ (since $C_d\in O(\log d)$). On the other hand, if $1/8d^{2+\eta}\leq 2^{-k} < 1/d$ and thus $d < 2^{k}$, the denominator is at most $O(C_d2^k)$. In this case, $C_d\in O(\log d) = O(k)$, so the displacement is at least $C''/(k2^k)$. Therefore the claim follows if we take $C = \min \{C', C''\}$. Finally, $z_n\not \in \mathbb{D}_2(0)$ implies $k < -1$. Again Lemma \ref{lemma: main} gives $|z_n - z_{n+1}| > 1/d$, and thus we finish the proof by possibly decreasing the constant $C$. \end{proof} The final step towards proving our main result is in the following theorem. \begin{theorem}\label{thm:one root} Let the polynomial $p(z)$ be randomly chosen in $\mathcal{P}_d$ or $\mathcal{P}_d/S_d$ and let $(z_n)$ be an $R$-central orbit converging to a root $\alpha$ with $d_U(z_0, z_1)\leq \tau$ for $\tau < 2\log d$. Then with probability $p_d\ge 1-O(d^{-2\eta})$ (for fixed $\eta\in(0,1/2)$), the required number of iterations for $z_0$ to get $\varepsilon$-close to $\alpha$ is \[ O\left(d^2\log^4 d R^2 + \log|\log\varepsilon - 5\right|) \;. \] \end{theorem} Before proceeding to the proof of this statement, we will outline the main idea. As in \cite{D}, we construct ``thick'' curves connecting orbit points $z_n$ and $z_{n+1}$ that use up certain area contained in a bounded domain. Far from the root, two curves corresponding to $z_n$ and $z_{m}$ are disjoint provided that $|n-m| > 2\tau + 6.$ A lower bound on the area of the ``thick'' curves gives an upper bound on the number of iterations. Also, near the root the orbit enters the domain of quadratic convergence where only a few iterations are sufficient to approximate the root. More precisely, let $\varphi\colon U\to \mathbb{D}$ be the Riemann map with $\varphi(\alpha) = 0$ considered in \cite[Section~5]{D}. If $|\varphi(z_n)| < 1/2$ (``region of fast convergence''), then according to \cite[Lemma~11]{D} we need only $\log_2|\log_2 \varepsilon - 5|$ iterations to get $\varepsilon$-close to the root $\alpha$. For orbit points with $\varphi$-images having absolute values greater than $e ^{1/2} - 1$, we can prove the following. \begin{lemma}\label{lemma: area bound} For every $n$ with $|\varphi(z_n)| > e^{1/2} - 1$, there are open connected subsets $V_n\subset D_{2R+2}(0)$ with $z_n,z_{n+1}\in\partial V_n$ and $|V_n|\ge |z_n-z_{n+1}|^2/2\tau,$ having the following property: whenever $n$ and $m$ are such that $\min \{|\varphi(z_n)|, |\varphi(z_m)|\} > e^{1/2} - 1$ and $|n-m|\ge \lceil 2\tau+6\rceil$, we have $V_n\cap V_m=\emptyset$. \end{lemma} \begin{proof} Let $\gamma\colon[0,s]\to U$ be the hyperbolic geodesic within $U$ connecting $z_n$ to $z_{n+1}$. For each $z=\gamma(t)$, let $\eta(t)$ be the Euclidean distance from $\gamma(t)$ to $\partial U$, and let $X_t$ be the straight line segment (without endpoints) perpendicular to $\gamma(t)$ of Euclidean length $\eta(t)$, centered at $\gamma(t)$. Let $V_n:=\bigcup_{t\in(0,s)}X_t$. Then all $V_n$ are open and connected and $z_n,z_{n+1}\in\partial V_n$, and the area of $V_n$ is at least $|z_n-z_{n+1}|^2/2\tau$: this follows as in \cite[Lemma~9]{D} (in this reference, the areas restricted to certain domains $S_k$ are calculated; omitting this restriction, we obtain the result we need, and the computations only get simpler). Moreover, the orbit $(z_n)$ is $R$-central and the unit disk contains other roots than $\alpha$, and hence the length of $\gamma(t)$ for $t\in[0,s]$ is bounded by $R+1.$ This implies that all pieces $V_n$ are contained in $D_{2R+2}(0)$ by construction. The fact that $V_n\cap V_m$ are disjoint when $|n-m|>2\tau+6$ is proved in \cite[Lemma~12]{D} (again for restricted domains, but this is immaterial for the proof). \end{proof} \begin{proof}[Proof of Theorem \ref{thm:one root}] We only need to consider iteration points whose images under $\varphi$ have absolute values at least $e^{1/2} - 1$. Also, by Lemma~ \ref{lemma: probabilistic conditions}, the conditions \textbf{AC} and \textbf{DC} hold true with probability $p_d\ge 1-O(d^{-2\eta})$. Choose $M$ so that $ 2^M - 1>2R+2$. We distinguish the following three cases. \begin{description} \item[The Far Case] we have $z_n\in S_k$ with $2^{-k}\ge 1/d$. By Corollary \ref{cor: displacement} (1) we have $|z_n - z_{n+1}| \geq\frac{C}{d\log d }$. Lemma \ref{lemma: area bound} says that any Newton iteration $z_n\mapsto z_{n+1}$ with $z_n\in S_k$ needs area at least \[ \frac{|z_n-z_{n+1}|^2}{2\tau} \ge \frac{C^2}{2\tau d^2\log^2 d} \;. \] Moreover, the pieces of area for the iterations $z_n\mapsto z_{n+1}$ and $z_{n'}\mapsto z_{n'+1}$ are disjoint provided that $n-n' \geq 2\tau + 6$, and all these pieces of area are contained in the disk $D_{2R+2}(0)$ with $R$ universally bounded. The total number of such iterations $D_{2R+2}(0)$ can accommodate is thus at most \[ C'd^2(\log d)^2 \tau \lceil 2\tau + 6\rceil R^2 \] for a universal constant $C'$. \item[The Intermediate Case] we have $z_n\in S_k$ with $1/8d^{2+\eta}\leq 2^{-k} < 1/d$. Then $\log_2 d < k \leq 3 + (2+\eta) \log_2 d$. By Corollary \ref{cor: displacement} (2) we have $|z_n - z_{n+1}| \geq{C}/{k2^k }$. Thus by \cite[ Proposition 13]{D}, the set $S_k$ contains at most {\allowdisplaybreaks \begin{align} &\pi d\left(2^{-k+1} + \frac{C}{k2^k}\right)^2 \left(2\tau + 2^{k-1}\frac{C}{k2^k}\right)\lceil 2\tau + 6\rceil \frac{k^22^{2k}}{C^2} \nonumber \\ &= \pi d 2^{-2k}k^{-2}(2k + C)^2 \left(2\tau + \frac{C}{2k}\right)\lceil 2\tau + 6\rceil \frac{k^22^{2k}}{C^2} \nonumber \\ &= \pi d C^{-2}(2k + C)^2 \left(2\tau + \frac{C}{2k}\right)\lceil 2\tau + 6\rceil \nonumber \\ &\leq \pi d C^{-2}(6 + (4+2\eta)\log_2 d + C)^2 \left(2\tau + \frac{C}{2\log_2 d}\right)\lceil 2\tau + 6\rceil \nonumber \\ &\leq C''d \log^2 d (2\tau + 1)\lceil 2\tau + 6\rceil \nonumber \end{align} }% orbit points for some universal constant $C''$. There are $3+(1+\eta)\log d$ possible values of $k$ in the Intermediate Case, so $\bigcup_{k} S_k$ (for all $k$ in the Intermediate Case) can accommodate at most \[ (1+\eta)C''d \log^3 d (2\tau + 1)\lceil 2\tau + 6\rceil \] orbit points for some universal constant $C''$. \item[The Near Case] we have $z_n \in S_k$ with $2^{-k} < {1}/{8d^{2+\eta}}$. By Corollary \ref{cor: bound on K}, $\alpha$ is the closest root to $z_n$ and we need $\log_2|\log_2\varepsilon - 5|$ iterations to get $\varepsilon$-close to it. \hide{Thus it only remains to consider the orbit points $z_n\in S_k$ for $k \leq 3 + (2+\eta)\log_2 d$; these are not necessarily in the region of ``fast`` convergence (their image under the Riemann map $\varphi$ as constructed in \cite[Section 3.3]{D} has absolute value at least $e^{1/2} - 1$). Since the orbit $(z_n)$ is contained in ${D}_{2^M - 1}(0)$ by hypothesis, we have $z_n\in S_k$ with $k \geq -M$ for all $n\geq 0$. } \end{description} Since $\tau\in O(\log d)$ and the Far Case dominates the Intermediate Case, the claim follows. \end{proof} We now conclude the main statement. \begin{proof}[Proof of Theorem \ref{Thm:Main}] By Theorem \ref{thm: set of starting points}, for each root there is a starting point satisfying the conditions of the theorem. In particular, these orbits are $R$-central for a universally bounded value of $R$. Note that the $d$ roots have to compete for the available area in $D_{2R+2}(0)$. Since the estimates in the proof of Theorem \ref{thm:one root} are based on the area (except for the Near Case where the orbit gets to the region of quadratic convergence), we get the same estimate for the combined number of iterations (except that the estimate $\log |\log \varepsilon|$ applies for each root separately, thus it is multiplied by $d$). \end{proof} \begin{remark} This result is close to optimal in the sense that the power of $d$ cannot be reduced for our universal set of starting points that is bounded away from the unit disk. The reason is that outside the unit disk $N_p$ is conjugate to the linear map $w \mapsto \frac{d-1}{d} w$ by \cite[Lemma~4]{HSS}, so at least $O(d)$ iterations are required for each ``good`` starting point to get close to the unit disk where the roots are located, and at least $O(d^2)$ for all the $d$ starting points combined. \end{remark}

https://arxiv.org/abs/2211.09094

Guessing cards with complete feedback

We consider the following game that has been used as a way of testing claims of extrasensory perception (ESP). One is given a deck of $mn$ cards comprised of $n$ distinct types each of which appears exactly $m$ times: this deck is shuffled and then cards are discarded from the deck one at a time from top to bottom. At each step, a player (whose psychic powers are being tested) tries to guess the type of the card currently on top, which is then revealed to the player before being discarded. We study the expected number $S_{n,m}$ of correct predictions a player can make: one could always guess the exact same type of card which shows that one can achieve $S_{n,m}>m$. We prove that the optimal (non-psychic) strategy is just slightly better than that and find the first order correction when $n, m$ grows at suitable rates. This is very different from the case where $m$ is fixed and $n$ is large (He & Ottolini) and similar to the case of fixed $n$ and $m$ is large (Graham & Diaconis). The case $m=n$ answers a question of Diaconis.

\section{Introduction} \subsection{Zener cards} Sometimes people present claims of having powers of extrasensory perceptions: a natural framework (proposed by the psychologist K. Zener and the botanist J. Rhine) in which to test such a hypothesis is that of card guessing (so-called Zener cards). Consider a well-mixed deck of $mn$ cards comprised of $n$ distinct types of cards each of which appears exactly $m$ times. The deck is well shuffled and then placed in front of the player (who has full knowledge of the composition of the deck). The player then has to guess the type of the card on top; after the guess is made, the card is shown to the player and then discarded from the deck. The game continues until the deck runs over. Assuming the player does \textit{not} have psychic abilities, how many correct guesses can one expect? \begin{center} \begin{figure}[h!] \centering \includegraphics[width=0.5\textwidth]{zener.png} \caption{Zener cards: the figure shows the $n=5$ different types each of which appears $m=5$ times for a total of 25 cards.} \label{fig:my_label} \end{figure} \end{center} \vspace{-15pt} A simple strategy would be to always guess the same type (say, the circle card). One is guaranteed to make exactly $m$ correct guesses. The optimal strategy is to memorize all cards that have been discarded up to now and guess the type of card that has, so far, appeared the fewest amounts of times (the optimality of this strategy was proven by Diaconis \& Graham \cite{DG81}). The natural question is now, assuming no powers of extrasensory perception, how many correct guesses can be expected under this optimal strategy? This game has also been analyzed in connection with clinical trials \cite{Blackwell1957, EFRON1971} and is generally well studied: for other results and variations, we refer to \cite{C98,Diaconis1978,DG81,diaconis2020card,diaconis2020guessing,KP01,KT21,KPP09,L21,P09,P91,S21}. \subsection{Results.} For any given $m,n$, we consider the quantity $S_{n,m}$ describing the expected number of correctly guessed cards under the optimal strategy. Two types of regimes are well understood. The first regime deals with the case where $n$, the number of distinct types, is fixed and the multiplicity $m$ with which each type appears becomes larges. \begin{thm}[Diaconis \& Graham \cite{DG81}] For the number of different types $n$ fixed and the multiplicity $m$ going to infinity, \begin{equation}\label{largemsmalln} S_{n,m}=m+\frac{\pi}{2}M_n\sqrt{m}+o_n(\sqrt m), \end{equation} where $M_n$ denotes the expected value of the maximum of $n$ normal random variables. \end{thm} One way of interpreting the result is perhaps as follows: the frequency with which each card appears should behave roughly like a normal distribution. Exploiting the fluctuations of $n$ Gaussians and taking the one that deviates the most from its expectation suggests an asymptotic along the lines given by Diaconis \& Graham. Note that this is merely a heuristic: these card counts are not actually independent. In the opposite regime, fixing the multiplicity $m$ and assuming the number of different types $n$ becomes large, we obtain a very different result. \begin{thm}[He \& Ottolini \cite{he2021card}] For the multiplicity $m$ fixed and the number of different types $n$ going to infinity, \begin{equation}\label{largemsmalln} S_{n,m}=H_mH_n+\sum_{j=1}^{m-1}\frac{1}{j}\ln{m\choose j}+O_m(n^{-1/m}), \end{equation} where $H_n = 1 + \dots + 1/n$ denotes the $n$-th harmonic number. \end{thm} The leading order term is $H_m H_n \sim \ln{n} \ln{m}$ which, for $m$ fixed, is logarithmic growth in $n$. The second term in the expansion only depends on $m$ and is thus a constant. Empirically, the result is accurate even for small values of $m,n$ (see \cite{he2021card}).\\ A natural remaining question is what happens when both $m,n \rightarrow \infty$ with the original Zener setup $m=n$ being perhaps particularly interesting. Our main result covers a wide range of these parameters and is applicable as long as the number of different cards $n$ is slightly smaller than exponential in the number of different types $m$. Such a restriction is necessary: when $n$ becomes disproportionately large compared to $m$, the result of He \& Ottolini \cite{he2021card} shows the behavior to be different. \begin{thm}[Main Result] \label{mainthm} Let $c, \varepsilon > 0$. If $m,n \rightarrow \infty$ while $(\ln{n})^{3+\varepsilon} \leq c \cdot m$, then \begin{equation}\label{mnwhatever} S_{n,m}=m+\frac{\pi}{\sqrt{2}}\sqrt{m\ln n}+o_{c,\varepsilon}(\sqrt{m\ln n}). \end{equation} \end{thm} The result covers a wide range of parameters. It also suggests that there is a phase transition in the regime where there are a great many different types of cards each of which only appearing a relatively small number of times. In that regime, we expect a switch from \ref{mnwhatever} to \ref{largemsmalln}. The nature of this transition is currently not understood and appears to be an interesting problem: the proof of our main results suggests that this phasse transition may perhaps occur around $\ln{n} \sim m$. Of course, many other problems (variance or the existence of a central limit theorem) remain. There is a heuristic that motivates our main result. We use $X_i(t)\in \{0,1,\ldots , m\}$ to denote the numbers of cards of type $1\leq i\leq n$ that are left in the deck when there are $1\leq t\leq nm$ cards left in total. Linearity of expectation and the description of the optimal strategy imply that \begin{equation}\label{linearita} S_{n,m}=\sum_{t=1}^{nm}\frac{\mathbb E[\max_i X_i(t)]}{t}. \end{equation} We rescale $p = t/nm$ (thus $0 \leq p \leq 1$). The $X_i(t)$ should approximately obey a normal distribution and we could moreover assume that they are independent. This is certainly false because $X_1(t) + \dots +X_n(t) = t$ but it would simplify the problem. Pretending that the $X_i(t)$'s are independent normal random variables with the correct mean $mp$ and variance $mp(1-p)$, one would obtain \begin{align*} \mathbb E[\max_i X_i(t)]\approx m p+\sqrt{2mp(1-p)\ln n}. \end{align*} Plugging in, we obtain (after substituting $p = t/nm$) \begin{align*} S_{n,m}=\sum_{t=1}^{nm}\frac{\mathbb E[\max_i X_i(t)]}{t} &\approx m + \sum_{t=1}^{mn} \frac{\sqrt{2mp(1-p)\ln n}}{t} \\ &= m + \sqrt{2 m \ln{n}}\sum_{t=1}^{mn} \frac{\sqrt{p(1-p)}}{t} \\ &\approx m + \sqrt{2 m \ln{n}}\int_0^1\sqrt{\frac{1-p}{p}}dp \end{align*} and the integral evaluates to $\pi/2$. \section{Proof} \subsection{Outline} Our proof will be motivated by the heuristic sketched above. To justify the heuristic, we will exploit a well-known conditional representation of the $X_i(t)$'s in terms of conditionally independent binomial random variables $Y_i(t)$ given their sum. The intuition is that the maximum should only be mildly affected by the conditioning on the sum, which is indeed the case. One has to be careful in the case of of small and large $t$ (having selected almost none or almost all of the cards) where approximations degenerate. We will split the argument into two parts. In Section \ref{sec2}, we show how to reduce the problem to the case where the $X_i$s are independent binomials by means of a conditional representation. In Section \ref{sec3}, we prove the result in the case of independent binomials by exploiting sharp bounds on binomials tails. These two ingredients then establish the result. Section \ref{prooof} contains the proof of the main result. \subsection{Reduction to independence}\label{sec2} We start by explaining how to reduce the problem to that of independent random variables by means of a useful conditional representation. The use of conditional limit theory to deal with order statistics of discrete processes dates back to \cite{Levin1981}, where the author exploits a conditional representation of the multinomial distribution in terms of independent Poisson conditioned on their sum.\\ Recall that, for each $1 \leq i \leq n$ and each $1\leq t\leq mn$, the random variable $X_i(t)$ counts the number of cards of type $i$ that are in the remaining $t$ cards. Fixing a value of $t$ their joint distribution is given by a multivariate hypergeometric distribution \begin{equation} \mathbb P(X_1(t)=j_1,\ldots, X_n(t)=j_n)=\frac{\prod_{i=1}^n {m \choose j_i}}{{nm \choose t}}, \quad j_1+\ldots+j_n=t, \quad 0\leq j_i\leq m \end{equation} Note that, if it were not for the constraint of having a total of $t$ cards remaining $$ \sum_{i=1}^{n} X_i(t) = \sum_{i=1}^{n} j_i=t,$$ the $X_i$ would be independent. To overcome this issue, we consider $n$ independent and identically distributed random variables $Y_1(t), \ldots, Y_n(t)$ each of which follow an independent binomial distribution $$Y_i\sim \mbox{Bin}(m,p) \quad \mbox{with}~ p=\frac{t}{mn}.$$ We also introduce their sum $$\tilde Y(t)=\sum_{i=1}^{n} Y_i(t)$$ and note that $\tilde Y(t)\sim \mbox{Bin}(mn,p)$. We will use the fact that for having $t$ fixed, the distribution of cards can be realized via independent and identical random variables following a binomial distribution and conditioned on having the correct sum (in particular, the binomial random variables are also not yet independent). This well-known characterization of the hypergeometric distribution (see, e.g., \cite{Skibinsky1970}) has a simple proof that we report here for the sake of completeness.\\ \begin{lemma}\label{conditionalrep} For any $n,m$ and $1\leq t\leq mn$ fixed, we have \begin{align*} \mathcal L(X_1(t),\ldots, X_n(t))=\mathcal L(Y_1(t)\ldots, Y_n(t)|\tilde Y(t)=t), \end{align*} where $\mathcal{L}$ denotes the law of the random variable. \end{lemma} \begin{proof} Consider any $n$-tuple $0\leq j_i\leq m$ with $\sum_{i=1}^n j_i=t$, and let $p=t/mn$. By definition of conditional expectation \begin{align*} \mathbb{P}(Y_1(t)=j_1,\ldots, Y_n(t)=j_n|\tilde Y(t)=t)&=\frac{\mathbb P(Y_1(t)=j_1,\ldots, Y_n(t)=j_n)}{\mathbb P(\tilde Y(t)=t)}, \end{align*} where the condition $\tilde Y(t) = t$ can be omitted because $\sum_{i=1}^n j_i=t$ by design. We can now use the independence of the $Y_i$ to compute \begin{align*} \mathbb P(Y_1(t)=j_1,\ldots, Y_n(t)=j_n) &= \prod_{i=1}^{n} {m \choose j_i}p^{j_i}(1-p)^{m-j_i} \\ &= p^{\sum_{i=1}^{n} j_i} (1-p)^{mn - \sum_{i=1}^{n} j_i }\prod_{i=1}^{n} {m \choose j_i} \\ &= p^t (1-p)^{mn - t} \prod_{i=1}^{n} {m \choose j_i}. \end{align*} Simultaneously, since $\tilde Y(t)\sim \mbox{Bin}(mn,p)$, we have $$ \mathbb P(\tilde Y(t)=t) = \binom{mn}{t} p^t (1-p)^{mn-t}$$ from which we deduce the desired statement. \begin{align*} \mathbb{P}(Y_1(t)=j_1,\ldots, Y_n(t)=j_n|\tilde Y(t)=t)=\frac{\prod_{i=1}^n {m \choose j_i}}{{nm \choose t}}. \end{align*} \end{proof} We define $\tilde S_{n,m}$ to be the analogue of \eqref{linearita} where we replace the hypergeometric random variables with \textit{independent} binomial random variables, i.e. \begin{equation} \label{eq:better} \tilde S_{n,m}=\sum_{t=1}^{mn}\frac{\mathbb E[\max_i Y_i(t)]}{t}. \end{equation} Note that, according to Lemma 1, if we condition the binomial random variables on having the correct sum $t$, we recover the hypergeometric distribution exactly: the purpose of the next Lemma is to show that omitting this conditioning leads to a small error. This will the conclude the first part of the proof, the remainder of which is then dedicated to the study \ref{eq:better}. \begin{lemma}\label{randomizing} We have, for some universal $C>0$, \begin{align*} |S_{n,m}-\tilde S_{n,m}| \leq C(\sqrt m+\ln n). \end{align*} \end{lemma} \begin{proof} Owing to Lemma \eqref{conditionalrep}, we can replace the independent binomial random variables $Y_i$ by hypergeometric random variables $X_i$ provided that we condition on their sum $t$: this allows us to write \begin{align*} S_{n,m}-\tilde S_{n,m}&=\sum_{t=1}^m \frac{\mathbb E[\max_i X_i(t)]-\mathbb E[\max_i Y_i(t)]}{t}\\&=\sum_{t=1}^{nm}\frac{1}{t}\sum_{s=1}^{nm}\left(\mathbb E[\max_i X_i(t)]-\mathbb E[\max_i X_i(s)]\right)\mathbb P(\tilde Y(t)=s). \end{align*} One would of course now expect that $P(\tilde Y(t)=s)$ is tightly concentrated around its expectation: it thus suffices to understand how quickly the maximum can change for $|t-s|$ relatively small (without loss of generality, we assume from now on $s<t$). We therefore have to understand the likelihood $\mathbb P\left(\max_i X_i(t+1)>\max_i X_i(t)\right)$. The maximum can only increase if a card is picked that is already maximal before. This suggests on conditioning on the number of types that have appeared a maximal number of times and we note that \begin{equation}\label{excursion} \mathbb P\left(\max_i X_i(t+1)>\max_i X_i(t)\Big||\ell: X_{\ell}(t)=\max_j X_j(t)=k|=j\right)=\frac{(m-k)j}{nm-t}\leq \frac{j}{n}, \end{equation} where we used that $$ t = \sum_{i=1}^{n} X_i(t) \leq n \max_{1 \leq i \leq n} X_i(t)$$ implying $$ k = \max_{1\leq i \leq n} X_i(t) \geq\frac{t}{n}.$$ The only case in which \eqref{excursion} is saturated is the configuration in which $j$ cards appear with multiplicity $k$, and all other cards appear with multiplicity $k-1$. In this case the maximum increases with probability precisely $j/n$. One way of seeing this is as follows: if the other cards had only appeared rarely up to that point, we would be more likely to pick one of these cards since there are still more of them in the pile. The most likely transition to a new maximum happens if the chance of picking a card that is already chosen a maximal number of times is greatest. \\ We will now present an argument which, implicitly, works under the assumption that we are constantly in the worst case setting described above (in a suitable sense). We introduce a Markov chain whose role is to keep track of the number of different types of cards whose current occurence is given by maximal multiplicity $\max_{i} X_i(t)$. Note that, in particular, if the maximum increases then there exists exactly one card which arises with maximal multiplicity and the counter drops back to 1. The Markov chain will be operating on the state space $\{1,\ldots, n\}$: it is possible to move from each point $j$ to either $j+1$ or back to $1$. The corresponding transition probabilities $q_{i,j}$ are given by \begin{align*} q_{j, j+1}=\frac{n-j}{n} \quad \mbox{and} \quad q_{j,1}=\frac{j}{n} \quad \mbox{for} \qquad 1\leq j\leq n. \end{align*} The further the Markov chain is from 1, the more likely it is to return to 1 (corresponding to uncovering a new maximum). We also observe that the Markov chain is more likely to return to 1 than we are to uncover a new maximum (because the card deck will not always be in the worst case scenario assumed above): more formally, for all $1\leq k\leq m$, $1\leq t\leq mn$ and $1\leq j\leq j'\leq n$ \begin{align*} \mathbb P\left(\max_i X_i(t+1)>\max_i X_i(t)\Big||\ell: X_{\ell}(t)=\max_j X_j(t)=k|=j\right)\leq q_{j,1}\leq q_{j',1}, \quad \end{align*} The number of times at which $\max X_i(t)$ changes are bounded above by the number $N_{t-s}$ of excursions away from $1$ of the Markov chain. In particular, \begin{align*} \mathbb E[\max_i X_i(t)]-\mathbb E[\max_i X_i(s)]\leq \mathbb E[N_{t-s}]. \end{align*} It thus remains to understand how often the Markov chain is going to hit the state 1. Let now $T$ be the time to return back to the origin for the Markov chain $Z$. Renewal theory suggest that the latter should be approximately $(t-s)\mathbb E[T]$ for $t-s$ large. This is made precise in \cite{renewal}, whose result gives the estimate \begin{align*} \mathbb E[N_{t-s}]\leq \frac{t-s}{\mathbb E[T]}+O\left(\frac{\mathbb E[T^2]}{(\mathbb E[T])^2}\right). \end{align*} Note that $T$ is nothing but the expected time to observe a birthday coincidence in the classical birthday problem. In particular, $$\mathbb P(T\geq s)=\left(1-\frac{s}{n}\right)\ldots \left(1-\frac{1}{n}\right)$$ and using the estimate \begin{align*} \exp\left(-\frac{s^2}{2n}+O\left(\frac{s^3}{n^2}\right)\right)\leq \left(1-\frac{s}{n}\right)\ldots \left(1-\frac{1}{n}\right)\leq \exp\left(-\frac{s^2}{2n}\right), \end{align*} we obtain the well-known results $\mathbb E[T]=\Omega(\sqrt n)$ and $\mathbb E[T^2]=O(n)$ and thus \begin{align*} \mathbb E[\max_i X_i(t)]-\mathbb E[\max_i X_i(s)]=O\left(\frac{t-s}{\sqrt n}+1\right). \end{align*} This implies \begin{align*} |S_{n,m}-\tilde S_{n,m}|&=O\left(\frac{1}{\sqrt n}\sum_{t=1}^{nm}\frac{1}{t}\sum_{s=1}^{mn}\mathbb |t-s|P(\tilde Y(t)=s)+\sum_{t=1}^{nm}\frac{1}{t}\right). \end{align*} The second sum can be bounded by $\ln{(mn)}$. As for the first sum, we first note that by Cauchy-Schwarz for any random variable $$ \mathbb{E} \left| X - \mathbb{E}X \right| \leq \sqrt{\mathbb{V} X}.$$ We also observe that $\tilde Y(t)\sim \mbox{Bin}(nm,p)$ with $p=t/nm$ has standard deviation $\sqrt{mnp(1-p)}$ and thus $$ \sum_{s=1}^{mn}\mathbb |t-s|P(\tilde Y(t)=s) \leq \sqrt{m n p (1-p)}.$$ This simplifies the first sum and leads to the desired bound since \begin{align*} \frac{1}{\sqrt n}\sum_{t=1}^{nm}\frac{1}{t}\sum_{s=1}^{mn}\mathbb |t-s|P(\tilde Y(t)=s) &\leq \frac{1}{\sqrt n}\sum_{t=1}^{nm}\frac{1}{t} \sqrt{mn p (1-p)} \\ &=\sqrt m\sum_{t=1}^{nm}\frac{1}{t} \sqrt{\frac{t}{mn} \left(1 - \frac{t}{mn}\right)} \\ &\leq c\sqrt{m} \int_0^1 \frac{1}{x}\sqrt{x(1-x)} dx \leq C \sqrt{m}. \end{align*} \end{proof} \subsection{The independent case}\label{sec3} It now suffices to analyze $$\tilde S_{n,m}=\sum_{t=1}^{nm}\frac{\mathbb E[\max_i Y_i(t)]}{t}$$ where the $Y_i$ are \textit{independent} binomial random variables and $$Y_i\sim \mbox{Bin}(m,p) \quad \mbox{with}~ p=\frac{t}{mn}.$$ We start by rescaling these random variables by shifting them to have mean 0: we let $\overline Y_i(t)=Y_i(t)-t/n$ which reduces our problem to the study of \begin{equation}\label{tails} \tilde S_{n,m}=m+\sum_{t=1}^{nm}\frac{\mathbb E[\max_i \overline Y_i(t)]}{t}. \end{equation} These binomial random variables are well approximated by a normal distributions in regions where their variance is not too small: this naturally suggests splitting the problem into different regions. We write, for some $ 0 < s=s_{n,m} \ll 1$ to be determined later (which will ultimately tend to 0 at a suitable rate), \begin{align*} \tilde S_{n,m} =m &+ \sum_{t=snm}^{(1-s)nm}\frac{\mathbb E[\max_i \overline Y_i(t)]}{t} \\ &+\sum_{t=1}^{s nm}\frac{\mathbb E[\max_i \overline Y_i(t)]}{t} + \sum_{t= (1-s) nm}^{nm}\frac{\mathbb E[\max_i \overline Y_i(t)]}{t}. \end{align*} We treat the first sum by comparing with a normal distribution. The two remaining sums are tail events that will be treated separately. We start with the tail events.\\ \textbf{The tail sums.} In order to deal with the tails, we just need an upper bound of the right order, but we can afford to lose a factor of $\sqrt{1-p}$. This follows from a Chernoff type argument: for all $\theta>0$, by linearity of expectation, \begin{align*} \exp{\{\theta \mathbb ~\mathbb{E}[\max_i \overline Y_i(t)]\}} &\leq \sum_{i=1}^{n} \exp{\{\theta~ \mathbb E[ \overline Y_i(t)]\}} \\ &= n\exp{\{\theta~ \mathbb{E}[\overline Y_1(t)]\}}\leq n(1-p+pe^{\theta})^me^{-\theta m}, \end{align*} so that taking logarithm of both sides and using $\ln(1+x)\leq x$ we obtain \begin{align*} \mathbb E[\max_i \overline Y_i(t)]\leq \frac{\ln n}{\theta}+\frac{mp}{\theta}(e^{\theta}-1)-m. \end{align*} We start with the case where $p$ is close to 0, which is the hardest to deal with since it is the time where the feedback is the most relevant. This allows for many extra correct guesses owing to the detailed knowledge of the composition of the deck -- captured by the logarithmic singularity close to $t=0$ in \eqref{tails}. However, we deal with that considering the choice \begin{align*} \theta=\sqrt{\frac{2\ln n}{m}}\left(\ln\frac{1}{p}\right)^{1+\varepsilon} \end{align*} Since $p = t/(mn) \geq 1/(mn)$ we deduce $$\ln \frac{1}{p}\leq \ln n+\ln m$$ and thus \begin{align*} \theta=O\left(\sqrt{\frac{(\ln n+\ln m)^{3+2\varepsilon}}{m}}\right)=o(1) \end{align*} using the main assumption on $m$ and $n$ from the Main Theorem. Since $\theta$ is tending to 0, we can replace the exponential function $e^{\theta}$ by a second order Taylor expansion and \begin{align*} \mathbb E[\max_i \overline Y_i(t)]&\leq \frac{\ln n}{\theta}+\frac{mp\theta}{2}+o(\theta)\\&=O\left(\sqrt{m\ln n}\left[\left(\ln\frac{1}{p}\right)^{-1-\varepsilon}+p\left(\ln\frac{1}{p}\right)^{1+\varepsilon}\right]\right)\\ &= O\left( \sqrt{m\ln n}\left(\ln\frac{1}{p}\right)^{-1-\varepsilon} \right). \end{align*} Therefore, we conclude \begin{align*} \sum_{t=1}^{s nm}\frac{\mathbb E[\max_i \overline Y_i(t)]}{t} &\lesssim \sqrt{m \ln{n}} \sum_{t=1}^{s m n} \frac{1}{t} \left(\ln \frac{mn}{t} \right)^{-1 - \varepsilon}. \end{align*} Comparing to the integral, we have $$\sum_{t=1}^{s m n} \frac{1}{t} \left(\ln \frac{mn}{t} \right)^{-1 - \varepsilon} \lesssim \int_0^{s} \frac{1}{x} \left(\ln \frac{1}{x} \right)^{-1 - \varepsilon} dx.$$ The integrand has an antiderivative in closed form $$ \int \frac{1}{x} \left(\ln \frac{1}{x} \right)^{-1 - \varepsilon} dx = \frac{1}{\varepsilon} \left( \ln\frac{1}{x} \right)^{-\varepsilon}$$ allowing us to deduce that $$ \int_0^{s} \frac{1}{x} \left(\ln \frac{1}{x} \right)^{-1 - \varepsilon} dx =\frac{1}{\varepsilon} \left( \ln \frac{1}{s} \right)^{-\varepsilon}$$ which, for fixed $\varepsilon$ tends to 0 provided that $s$ tends to 0. In the second regime, $p$ close to 1, we choose \begin{align*} \theta=\sqrt{\frac{2\ln n}{m}} \end{align*} which is guaranteed to converge to 0 as $m,n \rightarrow \infty$. A Taylor expansion and the observation $p\leq 1$ shows \begin{align*} \mathbb E[\max_i \overline Y_i(t)]\leq \frac{\ln n}{\theta}+\frac{mp\theta}{2}+o(\theta)=O(\sqrt{m\ln n}). \end{align*} From here, we conclude since \begin{align*} \sum_{t=(1-s)mn}^{ nm -1}\frac{\mathbb E[\max_i \overline Y_i(t)]}{t} &\lesssim \sqrt{m \ln{n}} \sum_{t=(1-s)mn}^{m n -1}\frac{1}{t} \lesssim s \sqrt{m\ln n} \end{align*} which again tends to 0 as $s \rightarrow 0$. \subsection{The main term} As for the main term, we need sharp asymptotics for the maximum of independent binomials. This amounts to controlling their tail probabilities. The following is a corollary of a more general result by Feller \cite{feller}. \begin{lemma}[Feller \cite{feller}]\label{fellermagic} Let $Y_{i}(t) \sim \emph{Bin}(m,p)$ be i.i.d. and $p = t/(nm)$. Let \begin{align*} Z_i(t)=\frac{Y_i(t)-mp}{\sqrt{mp(1-p)}} \end{align*} and assume that $p, 1-p\geq s_{n,m}$, where $s=s_{n,m}$ is chosen so that \begin{align*} \frac{\ln^3n}{ms(1-s)}=o(1), \quad \frac{m^{6/7}}{ms(1-s)}=o(1). \end{align*} Then, one has \begin{align*} \mathbb E[\max Z_i(t)]= \sqrt{2\ln n} + \mathcal{O}(1) \end{align*} with a uniform error bound on all such $p$s. \end{lemma} The assumption in the Main Theorem guarantees that we can find a sequence $s_{n,m}=m^{-\varepsilon'}$ for some $\varepsilon'=\varepsilon'(\varepsilon)$ sufficiently small. Therefore, we will use the notation $O_{\varepsilon}(1)$ to indicate that the error will be a function of $\varepsilon$. \begin{proof} We follow the notation of Theorem $1$ by Feller \cite{feller} applied to binomial random variables, which gives an uniform estimate \begin{align*} \mathbb P(Z_i(t)\geq x)=(1-\Phi(x))\left(1+O\left(\frac{x^3}{\sqrt{mp(1-p)}}\right)\right) \end{align*} where $\Phi$ denotes the cumulative distribution of a standard normal random variable. In particular, for $x=\sqrt{2\ln n}$ the error is small by our assumption and we can write the error term as $o_{\varepsilon}(1)$. Using, for $x=\sqrt{2\ln n}$, the elementary estimate \begin{align*} n(1-\Phi(x))=n\frac{e^{-x^2/2}}{\sqrt{2\pi}x}\left(1+O\left(1/x^2\right)\right)\geq \frac{c}{\sqrt{\ln n}}, \end{align*} for some absolute constant $c$ and all $n\geq 2$, we derive the bound \begin{align*} \mathbb P(\max Z_i(t)\leq x)&=\left(\mathbb P(Z_i(t)\leq x)\right)^n\\&=\left(1-\frac{n\mathbb P(Z_i(t)\geq x)}{n}\right)^n\\&=1-O\left(\frac{1}{\sqrt{\ln n}}(1+o_{\varepsilon}(1))\right) \end{align*} From this, we get a lower bound on expectation \begin{align*} \mathbb E[\max Z_i(t)]&\geq x \cdot \mathbb P(\max \mathbb Z_i\leq x)=\sqrt{2\ln n}+O_{\varepsilon}(1). \end{align*} As for the upper bound, we slightly extend the range and consider values of $x$ up to $x\leq \sqrt{2(\ln n+m^{1/7})}$. This range is still admissible since $$\frac{x^3}{\sqrt{mp(1-p)}}\rightarrow 0$$ owing to our assumptions. Using the standard bound $$ 1 - \Phi(x) \leq \frac{e^{-\frac{x^2}{2}}}{\sqrt{2\pi} x}$$ for cumulative distribution function of the Gaussian, we infer that if $x=\sqrt{2(\ln n+c)}$ for some $0\leq c\leq m^{1/7}$, then \begin{align*} n(1-\Phi(x))\leq \frac{ne^{-\frac{x^2}{2}}}{\sqrt{2\pi}x}=O\left(\frac{e^{-c}}{\sqrt{2\ln n}}\right) \end{align*} from which we obtain \begin{align*} n \cdot\mathbb P(Z_i(t)\geq x)= n(1-\Phi(x))(1+o_{\varepsilon}(1))=O_{\varepsilon}\left(\frac{e^{-c}}{\sqrt{\ln n}}\right). \end{align*} In order to bound the expectation, let $$M=\sqrt{m\max\left(\frac{1-p}{p},\frac{p}{1-p}\right)}$$ be the maximum value of $|Z_i|$ Notice that, for instance, $M\leq m$ owing to our assumption on $p$ and $1-p$. Then we obtain \begin{align*} \mathbb E[\max Z_i(t)]&\leq \int_0^{M}\mathbb P(\max Z_i(t)\geq x)dx\\&\leq \sqrt{2\ln n}+\int_{{\sqrt{2\ln n}}}^{\sqrt{2\ln n+m^{1/7}}}n\mathbb P(Z_i(t)\geq x)dx+Me^{-m^{1/7}} \end{align*} The last term is certainly $o(1)$, while a change of variable shows that the second term is \begin{align*} \int_{{\sqrt{2\ln n}}}^{\sqrt{2\ln n+m^{1/7}}}n\mathbb P(Z_i(t)\geq x)dx=O_{\varepsilon}\left( \int_0^{\infty}\frac{e^{-c}}{\sqrt{2\ln n+c}}dc\right)=O_{\varepsilon}\left(\frac{1}{\sqrt{\ln n}}\right). \end{align*} Collecting all the pieces, we have \begin{align*} \mathbb E[\max_i Z_i(t)]=\sqrt{2\ln n}+O_{\varepsilon}(1). \end{align*} \end{proof} \subsection{Proof of the Main Result} \label{prooof} \begin{proof} Using Lemma \ref{fellermagic}, we have \begin{align*} \mathbb E\left[\max_i Y_i(t)-\frac{t}{n}\right] = \sqrt{m p(1-p)} \sqrt{2 \ln n } + \mathcal{O}(\sqrt{m p(1-p)}) \end{align*} with error term uniform for all $p, 1-p\geq s\rightarrow 0$. Combining Lemma \ref{randomizing} with the control on the tail, we have \begin{align*} S_{n,m}-m &= \tilde S_{n,m}-m + \mathcal{O}(\sqrt m+\ln n)\\ &= \sum_{t=1}^{nm}\frac{\mathbb E[\max_i Y_i(t)]}{t} - m + \mathcal{O}(\sqrt m+\ln n)\\ &= \sum_{t= 1}^{nm}\frac{\mathbb E[\max_i \overline Y_i(t)]}{t}+ \mathcal{O}(\sqrt m+\ln n). \end{align*} At this point, we split the sum into the three regions $$ \sum_{t=1}^{mn} = \sum_{t=1}^{smn} + \sum_{t=smn}^{(1-s)mn} + \sum_{t=(1-s)mn}^{mn}.$$ As was shown in Section \S 2.3, as long as $s \rightarrow 0$, the first and the third sum are $o(\sqrt{m \ln{n}})$. The sum in the middle, provided $s$ does not tend to 0 too quickly, will then contribute $$ \sum_{t= smn}^{(1-s)nm}\frac{\mathbb E[\max_i \overline Y_i(t)]}{t} = \left(1+\mathcal{O}\left( \frac{1}{\sqrt{\ln{n}}}\right)\right)\sum_{t= smn}^{(1-s)nm}\frac{ \sqrt{m p(1-p)} \sqrt{2 \ln n } }{t}.$$ The sum can now be simplified to $$ \sum_{t= smn}^{(1-s)nm}\frac{ \sqrt{m p(1-p)} \sqrt{2 \ln n } }{t} = \sqrt{2 m \ln{n}} \sum_{t= smn}^{(1-s)nm}\frac{ \sqrt{ p(1-p)} }{t} $$ which, recalling $p = t/mn$ leads, as $s \rightarrow 0$, to the Riemann sum $$\int_0^1\sqrt{\frac{1-p}{p}}dp = \frac{\pi}{2}.$$ \end{proof} \section*{Acknowledgment} We warmly thank Persi Diaconis for suggesting the problem and for some useful references. \bibliographystyle{abbrv}

https://arxiv.org/abs/1802.08855

Minimax Distribution Estimation in Wasserstein Distance

The Wasserstein metric is an important measure of distance between probability distributions, with applications in machine learning, statistics, probability theory, and data analysis. This paper provides upper and lower bounds on statistical minimax rates for the problem of estimating a probability distribution under Wasserstein loss, using only metric properties, such as covering and packing numbers, of the sample space, and weak moment assumptions on the probability distributions.

\section{Preliminary Lemmas and Proof Sketch of Theorem~\ref{thm:expectation_bound_appendix}} \label{sec:lemmas} In this section, we outline the proof of Theorem~\ref{thm:expectation_bound}, our upper bound for the case of totally bounded metric spaces. The proof of the more general Theorem~\ref{thm:unbounded_upper_bound} for unbounded metric spaces, which is given in the next section, builds on this. We begin by providing a few basic lemmas; these lemmas are not fundamentally novel, but they will be used in the subsequent proofs of our main upper and lower bounds, and also help provide intuition for the behavior of the Wasserstein metric and its connections to other metrics between probability distributions. The proofs of these lemmas are given later, in Appendix~\ref{app:proofs}. Our first lemma relates Wasserstein distance to the notion of resolution of a partition. \begin{lemma} Suppose $\S \in \SS$ is a countable Borel partition of $\Omega$. Let $P$ and $Q$ be Borel probability measures such that, for every $S \in \S$, $P(S) = Q(S)$. Then, for any $r \geq 1$, $W_r(P, Q) \leq \operatorname{Res}(\S)$. \label{lemma:measures_agree_on_partition} \end{lemma} Our next lemma gives simple lower and upper bounds on the Wasserstein distance between distributions supported on a countable subset $\X \subseteq \Omega$, in terms of $\Diam(\X)$ and $\Sep(\X)$. Since our main results will utilize coverings and packings to approximate $\Omega$ by finite sets, this lemma will provide a first step towards approximating (in Wasserstein distance) distributions on $\Omega$ by distributions on these finite sets. Indeed, the lower bound in Inequality~\eqref{ineq:countable_support_bound} will suffice to prove our lower bounds, although a tighter upper bound, based on the upper bound in~\eqref{ineq:countable_support_bound}, will be necessary to obtain tight upper bounds. \begin{lemma} Suppose $(\Omega, \rho)$ is a metric space, and suppose $P$ and $Q$ are Borel probability distributions on $\Omega$ with countable support; i.e., there exists a countable set $\X \subseteq \Omega$ with $P(\X) = Q(\X) = 1$. Then, for any $r \geq 1$, \begin{equation} (\Sep(\X))^r \sum_{x \in \X} \left| P(\{x\}) - Q(\{x\}) \right| \leq W_r^r(P,Q) \leq (\Diam(\X))^r \sum_{x \in \X} \left| P(\{x\}) - Q(\{x\}) \right|. \label{ineq:countable_support_bound} \end{equation} \label{lemma:countable_support_bound} \end{lemma} \begin{remark} Recall that the term $\sum_{x \in \X} \left| P(\{x\}) - Q(\{x\}) \right|$ in Inequality~\eqref{ineq:countable_support_bound} is the $\L_1$ distance \[\|p - q\|_1 := \sum_{x \in \X} \left| p(x) - q(x) \right|\] between the densities $p$ and $q$ of $P$ and $Q$ with respect to the counting measure on $\X$, and that this same quantity is twice the total variation distance \[TV(P,Q) := \sup_{A \subseteq \Omega} \left| P(A) - Q(A) \right|.\] Hence, Lemma~\ref{lemma:countable_support_bound} can be equivalently written as \[\Sep(\Omega) \left( \|p - q\|_1 \right)^{1/r} \leq W_r(P,Q) \leq \Diam(\Omega) \left( \|p - q\|_1 \right)^{1/r}\] and as \[\Sep(\Omega) \left( 2 TV(P,Q) \right)^{1/r} \leq W_r(P,Q) \leq \Diam(\Omega) \left( 2 TV(P,Q) \right)^{1/r},\] bounding the $r$-Wasserstein distance in terms of the $\L_1$ and total variation distance. As noted in Example~\ref{ex:discrete_bound}, equality holds in \eqref{ineq:countable_support_bound} precisely when $\rho$ is the unit discrete metric given by $\rho(x,y) = 1_{\{x \neq y\}}$ for all $x,y \in \Omega$. On metric spaces that are discrete (i.e., when $\Sep(\Omega) > 0$), the Wasserstein metric is (topologically) at least as strong as the total variation metric (and the $\L_1$ metric, when it is well-defined), in that convergence in Wasserstein metric implies convergence in total variation (and $\L_1$, respectively). On the other hand, on bounded metric spaces, the converse is true. In either of these cases, \emph{rates} of convergence may differ between metrics, although, in metric spaces that are both discrete \textit{and} bounded (e.g., any finite space), we have $W_r \asymp TV^{1/r}$. \label{remark:Wasserstein_L1_TV} \end{remark} To obtain tight bounds as discussed below, we will require not only a partition of the sample space $\Omega$, but a nested sequence of partitions, defined as follows. \begin{definition}[Refinement of a Partition, Nested Partitions] Suppose $\S, \T \in \SS$ are partitions of $\Omega$. $\T$ is said to be a \emph{refinement of $\S$} if, for every $T \in \T$, there exists $S \in \S$ with $T \subseteq S$. A sequence $\{\S_k\}_{k \in \N}$ of partitions is called \emph{nested} if, for each $k \in \N$, $\S_k$ is a refinement of $\S_{k + 1}$, \end{definition} While Lemma~\ref{lemma:countable_support_bound} gave a simple upper bound on the Wasserstein distance, the factor of $\Diam(\Omega)$ turns out to be too large to obtain tight rates for a number of cases of interest (such as the $D$-dimensional unit cube $\Omega = [0,1]^D$, discussed in Example~\ref{ex:unit_cube_lower_bound}). The following lemma gives a tighter upper bound, based on a hierarchy of nested partitions of $\Omega$; this allows us to obtain tighter bounds (than $\Diam(\Omega)$) on the distance that mass must be transported between $P$ and $Q$. Note that, when $K = 1$, Lemma~\ref{lemma:nested_partitions_Wasserstein_bound} reduces to a trivial combination of Lemmas~\ref{lemma:measures_agree_on_partition} and \ref{lemma:countable_support_bound}; indeed, these lemmas are the starting point for proving Lemma~\ref{lemma:nested_partitions_Wasserstein_bound} by induction on $K$. Note that the idea of such a ``multi-resolution'' upper bound has been utilized extensively before, and numerous versions have been proven before (see, e.g., Fact 6 of \citet{do2011sublinearTimeEMD}, Lemma 6 of \citet{fournier2015rate}, or Proposition 1 of \citet{weed2017sharp}). Most of these versions have been specific to Euclidean space; to the best of our knowledge, only Proposition 1 of \citet{weed2017sharp} applies to general metric spaces. However, that result also requires that $(\Omega,\rho)$ is totally bounded (more precisely, that $m_x^\infty(P) < \infty$, for some $x \in \Omega$). \begin{lemma} Let $K$ be a positive integer. Suppose $\{\S_k\}_{k \in \N}$ is a nested sequence of countable Borel $\delta$-partitions of $(\Omega,\rho)$. Then, for any $r \geq 1$ and Borel probability measures $P$ and $Q$ on $\Omega$, \begin{equation} W_r^r(P, Q) \leq (\operatorname{Res}(\S_0))^r + \sum_{k = 1}^\infty \left( \operatorname{Res}(\S_k) \right)^r \left( \sum_{S \in \S_{k + 1}} \left| P(S) - Q(S) \right| \right). \label{ineq:multiresolution_bound} \end{equation} \label{lemma:nested_partitions_Wasserstein_bound} \end{lemma} Lemma~\ref{lemma:nested_partitions_Wasserstein_bound} requires a sequence of partitions of $\Omega$ that is not only multi-resolution but also nested. While the $\epsilon$-covering number implies the existence of small partitions with small resolution, these partitions need not be nested as $\epsilon$ becomes small. For this reason, we now give a technical lemma that, given any sequence of partitions, constructs a \textit{nested} sequence of partitions of the same cardinality, with only a small increase in resolution. \begin{lemma} Suppose $\S$ and $\T$ are partitions of $(\Omega,\rho)$, and suppose $\S$ is countable. Then, there exists a partition $\S'$ of $(\Omega,\rho)$ such that: \begin{enumerate}[label=\alph*)] \item $|\S'| \leq |\S|$. \item $\operatorname{Res}(\S') \leq \operatorname{Res}(\S) + 2\operatorname{Res}(\T)$. \item $\T$ is a refinement of $\S'$. \end{enumerate} \label{lemma:fine_refinement} \end{lemma} Lemmas~\ref{lemma:nested_partitions_Wasserstein_bound} and \ref{lemma:fine_refinement} are the main tools needed to bound the expected Wasserstein distance $\E[W_r^r(P, \hat P)]$ of the empirical distribution from the true distribution into a sum of its expected errors on each element of a nested partition of $\Omega$. Then, we will need to control the total expected error across these partition elements, which we will show behaves similarly to the $\L_1$ error of the standard maximum likelihood (mean) estimator a multinomial distribution from its true mean. Thus, the following result of \citet{han2015minimax} will be useful. \begin{lemma}[Theorem 1 of \citep{han2015minimax}] Suppose $(X_1,...,X_K) \sim \operatorname{Multinomial}(n,p_1,...,p_K)$. Let \[Z := \|X - n p\|_1 = \sum_{k = 1}^K \left| X_k - n p_k \right|.\] Then, $\E \left[ Z/n \right] \leq \sqrt{(K - 1)/n}$. \label{lemma:multinomial_expectation} \end{lemma} Finally, we are ready to prove Theorem~\ref{thm:expectation_bound_appendix}. \begin{customthm}{\ref{thm:expectation_bound}} Let $(\Omega,\rho)$ be a metric space on which $P$ is a Borel probability measure. Let $\hat P$ denote the empirical distribution of $n$ IID samples $X_1,...,X_n \stackrel{IID}{\sim} P$, give by \[\hat P(S) := \frac{1}{n} \sum_{i = 1}^n 1_{\{X_i \in S\}}, \quad \forall S \in \Sigma.\] Then, for any sequence $\{\epsilon_k\}_{k \in [K]} \in (0,\infty)^K$ with $\epsilon_0 = \Diam(\Omega)$, \[\E \left[ W_r^r(P, \hat P) \right] \leq \epsilon_K^r + \frac{1}{\sqrt{n}} \sum_{k = 1}^K \left( \sum_{j = k - 1}^K 2^{j - k} \epsilon_j \right)^r \sqrt{N(\epsilon_k) - 1}.\] \label{thm:expectation_bound_appendix} \end{customthm} \begin{proof} By recursively applying Lemma~\ref{lemma:fine_refinement}, there exists a sequence $\{\S_k\}_{k \in [K]}$ of partitions of $(\Omega,\rho)$ satisfying the following conditions: \begin{enumerate} \item for each $k \in [K]$, $|\S_k| = N(\epsilon_k)$. \item for each $k \in [K]$, $\displaystyle \operatorname{Res}(\S_k) \leq \sum_{j = k}^K 2^{j - k} \epsilon_j$. \item $\{S_k\}_{k \in [K]}$ is nested. \end{enumerate} Note that, for any $k \in [K]$, the vector $n\hat P(S)$ (indexed by $S \in \S_k$) follows an $n$-multinomial distribution over $|\S_k|$ categories, with means given by $P(S)$; i.e., \[(n\hat P(S_1),...,n\hat P(S_k)) \sim \operatorname{Multinomial}(n,P(S_1),...,P(S_k)).\] Thus, by Lemma~\ref{lemma:multinomial_expectation}, for each $k \in [K]$, \[\E \left[ \sum_{S \in \S_k} \left| P(S) - \hat P(S) \right| \right] \leq \sqrt{\frac{|\S_k| - 1}{n}} = \sqrt{\frac{N(\epsilon_k) - 1}{n}}.\] Thus, by Lemma~\ref{lemma:nested_partitions_Wasserstein_bound}, \begin{align*} \E \left[ W_r^r(P, Q) \right] & \leq \E \left[ \epsilon_K^r + \sum_{k = 1}^K \left( \sum_{j = k}^K 2^{j - k} \epsilon_j \right)^r \left( \sum_{S \in \S_k} \left| P(S) - Q(S) \right| \right) \right] \\ & \leq \epsilon_K^r + \sum_{k = 1}^K \left( \sum_{j = k}^K 2^{j - k} \epsilon_j \right)^r \E \left[ \sum_{S \in \S_k} \left| P(S) - Q(S) \right| \right] \\ & \leq \epsilon_K^r + \frac{1}{\sqrt{n}} \sum_{k = 1}^K \left( \sum_{j = k}^K 2^{j - k} \epsilon_j \right)^r \sqrt{N(\epsilon_k) - 1} \end{align*} \end{proof} \section{Proof Sketch of Theorem~\ref{thm:unbounded_upper_bound_appendix}} In this section, we prove our more general upper bound, Theorem~\ref{thm:unbounded_upper_bound_appendix}, which applies to potentially unbounded metric spaces $(\Omega,\rho)$, assuming that $P$ is sufficiently concentrated (i.e., has at least $\ell > 0$ finite moments). The basic idea is to partition the potentially unbounded metric space $(\Omega,\rho)$ into countably many totally bounded subsets $B_1,B_2,...$, and to decompose the Wasserstein error into its error on each $B_i$, weighted by the probability $P(B_i)$. Specifically, fixing an arbitrary base point $x_0$, $B_1,B_2,...$ will be spherical shells, such that $x_0 \in B_1$, and both the distance between $B_i$ and $x_0$, as well as the size (covering number) of $B_i$, increase with $i$. For large $i$, the assumption that $P$ has $\ell$ bounded moments implies (by Markov's inequality) that $P(B_i)$ is small, whereas, for small $i$, we adapt our previous result Theorem~\ref{thm:expectation_bound_appendix} in terms of the covering number. To carry out this approach, we will need two new lemmas. The first decomposes Wasserstein distance into the sum of its distances on each $B_i$, and can be considered an adaptation of Lemma 2.2 of \citet{lei2018convergence} (for Banach spaces) to general metric spaces. \begin{lemma} Fix a reference point $x_0 \in \Omega$ and a non-decreasing real-valued sequence $\{w_k\}_{k \in \N}$ with $w_0 = 0$ and $\lim_{k \to \infty} w_k = \infty$. For each $k \in \N$, define \[B_k := \left\{x \in \Omega : w_k \leq \rho(x_0, x) < w_{k + 1} \right\}.\] Then, there exists a constant $C_r$ depending only on $r$ such that, for any Borel probability measures $P$ and $Q$ on $\Omega$, \[W_r^r(P,Q) \leq C_r \sum_{k = 0}^\infty w_k^r \min \left\{ P(B_k), Q(B_k) \right\} W_r^r(P_{B_k},Q_{B_k}) + \left| P(B_k) - Q(B_k) \right|.\] where, for any sets $A, B \subseteq \Omega$, \[P_A(B) = \frac{P(A \cap B)}{P(B)}\] (under the convention that $\frac{0}{0} = 0$) denotes the conditional probability of $B$ given $A$, under $P$. \label{lemma:sigma_finite_partition} \end{lemma} The second lemma is more nuanced variant of Lemma~\ref{lemma:multinomial_expectation} (albeit, leading to slightly looser constants). When $i$ is large the covering number of $B_i$ can become quite large, but the total probability $P(B_i)$ is quite small. Whereas Lemma~\ref{lemma:multinomial_expectation} depends only on the size of the partition, the following result will allow us to control the total error using both of these factors. \begin{lemma}[Theorem 1 of \citet{berend2013binomialMAD}] Suppose $X \sim \operatorname{Binomial}(n,p)$. Then, we have the bound \begin{equation} \E \left[ \left| X - n p \right| \right] \leq n \min \left\{ 2P(A), \sqrt{P(A)/n} \right\}. \label{ineq:binomial_MAD} \end{equation} on the mean absolute deviation of $X$. \label{lemma:binomial_MAD} \end{lemma} Finally, we are ready to prove our main upper bound result for unbounded metric spaces. \begin{customthm}{\ref{thm:unbounded_upper_bound}}[General Upper Bound for Unbounded Metric Spaces] Let $x_0 \in \Omega$ and suppose $m_{\ell,x_0}(P) \in [1, \infty)$. Let $J$ be a positive integer. Fix two non-decreasing real-valued sequences $\{w_k\}_{k \in \N}$ and $\{\epsilon_j\}_{j \in \N}$, of which $\{w_k\}_{k \in \N}$ is non-decreasing with $w_0 = 0$ and $\lim_{k \to \infty} w_k = \infty$ and $\{\epsilon_j\}_{j \in [J]}$ is non-increasing. For each $k \in \N$, define \[B_k(x_0) := \left\{ y \in \Omega : w_k \leq \rho(x_0, x) < w_{k + 1} \right\}.\] Then, \begin{align*} \E \left[ W_r^r(P, \hat P) \right] & \leq m_{\ell,x_0}^\ell \sum_{k \in \N} w_k^{-\ell} \left( \epsilon_J \right)^r + 2^r w_k^{r - \ell/2} \min \left\{ 2w_k^{-\ell/2}, \sqrt{\frac{1}{n}} \right\} \\ & \hspace{2cm} + \sum_{j = 1}^J \left( \sum_{t = j}^J 2^{J - t} \epsilon_t \right)^r \min \left\{ 2w_k^{-\ell}, \sqrt{\frac{w_k^{-\ell}}{n} N(B_k,\rho,\epsilon_j)} \right\}. \end{align*} \label{thm:unbounded_upper_bound_appendix} \end{customthm} \begin{proof} As in the proof of Theorem~\ref{thm:expectation_bound_appendix}, by recursively applying Lemma~\ref{lemma:fine_refinement}, for each $k \in \N$, we can construct a nested sequence $\{\S_{k,j}\}_{j \in [J_k]}$ of partitions of $B_k$ such that, for each $j \in [J_k]$, \begin{equation} |\S_{k,j}| = N(B_k,\rho,\epsilon_{k,j}) \quad \text{ and } \quad \operatorname{Res}(\S_{k,j}) \leq \sum_{t = 0}^j 2^t \epsilon_{k,j}. \label{eq:recursive_fine_refinement} \end{equation} Since each $P_{B_k}$ and $\hat P_{B_k}$ are supported only on $B_k$, plugging the bound Lemma~\ref{lemma:nested_partitions_Wasserstein_bound} into the bound in Lemma~\ref{lemma:sigma_finite_partition} gives \begin{align*} & W_r^r(P, \hat P) \\ & \leq \sum_{k \in \N} \min \left\{ P(B_k), \hat P(B_k) \right\} \left( \left( \operatorname{Res}(\S_{k,0}) \right)^r + \sum_{j = 1}^{J_k} \left( \operatorname{Res}(\S_{k,j}) \right)^r \sum_{S \in \S_{k,j + 1}} \left| P_{B_k}(S) - \hat P_{B_k}(S) \right| \right) \\ & \hspace{1cm} + 2^r w_k^r \left| P(B_k) - \hat P(B_k) \right| \\ & \leq \sum_{k \in \N} 2^r w_k^r \left| P(B_k) - \hat P(B_k) \right| + P(B_k) \left( \operatorname{Res}(\S_{k,0}) \right)^r + \sum_{j = 1}^J \left( \operatorname{Res}(\S_{k,j}) \right)^r \sum_{S \in \S_{k,j + 1}} \left| P(S) - \hat P(S) \right|. \end{align*} Since each $\hat P(S) \sim \operatorname{Binomial}(n, P(S))$, for each $k \in \N$ and $j \in [J_k]$, Lemma~\ref{lemma:binomial_MAD} followed by Cauchy-Schwarz gives \begin{align*} \E \left[ \sum_{S \in \S_{k,j}} \left| P(S) - \hat P(S) \right| \right] & \leq \sum_{S \in \S_{k,j + 1}} \min \left\{ 2P(S), \sqrt{P(S)/n} \right\} \\ & \leq \min \left\{ 2P(B_k), \sqrt{\frac{P(B_k)}{n} |\S_{k,j}|} \right\}. \end{align*} Therefore, taking expectations (over $X_1,...,X_n$), applying Inequality~\ref{eq:recursive_fine_refinement}, and applying Lemma~\ref{lemma:binomial_MAD} once more gives \begin{align*} \E \left[ W_r^r(P, \hat P) \right] & \leq \sum_{k \in \N} P(B_k) \left( \epsilon_{k,0} \right)^r + 2^r w_k^r \min \left\{ 2P(B_k), \sqrt{P(B_k)/n} \right\} \\ & \hspace{1cm} + \sum_{j = 1}^{J_k} \left( \sum_{t = 0}^j 2^t \epsilon_{k,j} \right)^r \min \left\{ 2P(B_k), \sqrt{\frac{P(B_k)}{n} N(B_k,\rho,\epsilon_{k,j + 1})} \right\}. \end{align*} Now note that, by Markov's inequality, \begin{equation} P(B_k) \leq \pr_{X \sim P} \left[ \rho(x_0, X) \geq w_k \right] = \pr_{X \sim P} \left[ \rho^\ell (x_0, X) \geq w_k^\ell \right] \leq \frac{m_{\ell,x_0}^\ell(P)}{w_k^\ell}. \end{equation} Therefore, assuming that each $m_{\ell,x_0}^\ell \geq 1$, so that $m_{\ell,x_0}^\ell \geq m_{\ell,x_0}^{\ell/2}$, \begin{align*} \E \left[ W_r^r(P, \hat P) \right] & \leq m_{\ell,x_0}^\ell \sum_{k \in \N} w_k^{-\ell} \left( \epsilon_{k,0} \right)^r + 2^r w_k^r \min \left\{ 2w_k^{-\ell}, \sqrt{w_k^{-\ell}/n} \right\} \\ & \hspace{1cm} + \sum_{j = 1}^{J_k} \left( \sum_{t = 0}^j 2^t \epsilon_{k,j} \right)^r \min \left\{ 2w_k^{-\ell}, \sqrt{\frac{w_k^{-\ell}}{n} N(B_k,\rho,\epsilon_{k,j + 1})} \right\}, \end{align*} proving the theorem. \end{proof} \section{Proofs of Lemmas} \label{app:proofs} \begin{customlemma}{\ref{lemma:measures_agree_on_partition}} Suppose $\S \in \SS$ is a countable Borel partition of $\Omega$. Let $P$ and $Q$ be Borel probability measures such that, for every $S \in \S$, $P(S) = Q(S)$. Then, for any $r \geq 1$, $W_r(P, Q) \leq \operatorname{Res}(\S)$. \end{customlemma} \begin{proof} This fact is intuitively obvious; clearly, there exists a transportation map $\mu$ from $P$ to $Q$ that moves mass only within each $S \in \S$ and therefore without moving any mass further than $\delta$. For completeness, we give a formal construction. Let $\mu : \Sigma^2 \to [0,1]$ denote the coupling that is conditionally independent given any set $S \in \S$ with $P(S) = Q(S) > 0$ (that is, for any $A, B \in \Sigma$, $\mu(A \times B \cap S \times S) P(S) = P(A \cap S) Q(B \cap S)$).\footnote{The existence of such a measure can be verified by the Hahn-Kolmogorov theorem, similarly to that of the usual product measure (see, e.g., Section IV.4 of \citet{doob2012measure}).} It is easy to verify that $\mu \in \mathcal{C}(P,Q)$. Since $\S$ is a countable partition and $\mu$ is only supported on $\bigcup_{S \in \S} S \times S$, \begin{align*} W_r(P, Q) & \leq \left( \int_{\Omega \times \Omega} \rho^r(x,y) \, d\mu(x,y) \right)^{1/r} \\ & = \left( \sum_{S \in \S} \int_{S \times S} \rho^r(x,y) \, d\mu(x,y) \right)^{1/r} \\ & \leq \left( \sum_{S \in \S} \int_{S \times S} \delta^r \, d\mu(x,y) \right)^{1/r} \\ & = \delta \left( \sum_{S \in \S} \mu(S \times S) \right)^{1/r} = \delta \left( \sum_{S \in \S} \frac{P(S) Q(S)}{P(S)} \right)^{1/r} = \delta \left( \sum_{S \in \S} Q(S) \right)^{1/r} = \delta. \end{align*} \end{proof} \begin{customlemma}{\ref{lemma:countable_support_bound}} Suppose $(\Omega, \rho)$ is a metric space, and suppose $P$ and $Q$ are Borel probability distributions on $\Omega$ with countable support; i.e., there exists a countable set $\X \subseteq \Omega$ with $P(\X) = Q(\X) = 1$. Then, for any $r \geq 1$, \[(\Sep(\X))^r \sum_{x \in \X} \left| P(\{x\}) - Q(\{x\}) \right| \leq W_r^r(P,Q) \leq (\Diam(\X))^r \sum_{x \in \X} \left| P(\{x\}) - Q(\{x\}) \right|.\] \end{customlemma} \begin{proof} The term $\sum_{x \in \X} \left| P(\{x\}) - Q(\{x\}) \right| = TV(P, Q)$ is precisely the (unweighted) amount of mass that must be transported to transform between $P$ and $Q$. Hence, the result is intuitively fairly obvious; all mass moved has a cost of at least $\Sep(\Omega)$ and at most $\Diam(\Omega)$. However, for completeness, we give a more formal proof below. To prove the lower bound, suppose $\mu \in \Pi(P, Q)$ is any coupling between $P$ and $Q$. For $x \in \X$, \[\mu(\{x\} \times \{x\}) + \mu(\{x\} \times (\Omega \sminus \{x\})) = \mu(\{x\} \times \Omega) = P(\{x\})\] and, similarly, \[\mu(\{x\} \times \{x\}) + \mu((\Omega \sminus \{x\}) \times \{x\}) = \mu(\Omega \times \{x\}) = Q(\{x\}).\] Since $P(\{x\}), Q(\{x\}) \in [0,1]$, it follows that \[\mu(\{x\} \times (\Omega \sminus \{x\})) + \mu(\mu((\Omega \sminus \{x\}) \times \{x\})) \geq \left| P(\{x\} - Q(\{x\}) \right|.\] Therefore, since $\rho(x,y) = 0$ whenever $x = y$ and $\rho(x, y) \geq \Sep(\Omega)$ whenever $x \neq y$, \begin{align*} \int_{\Omega \times \Omega} \rho^r(x, y) \, d\mu(x,y) & = \int_{\X \times \X} \rho^r(x, y) \, d\mu(x,y) \\ & = \sum_{x \in \X} \int_{\{x\} \times (\Omega \sminus \{x\})} \rho^r(x, y) \, d\mu(x,y) + \int_{(\Omega \sminus \{x\}) \times \{x\}} \rho^r(x, y) \, d\mu(x,y) \\ & \geq (\Sep(\Omega))^r \sum_{x \in \X} \mu(\{x\} \times (\Omega \sminus \{x\})) + \mu((\Omega \sminus \{x\}) \times \{x\}) \\ & \geq (\Sep(\Omega))^r \sum_{x \in \X} \left| P(\{x\}) - Q(\{x\}) \right|. \end{align*} Taking the infimum over $\mu$ on both sides gives \[(\Sep(\Omega))^r \sum_{x \in \X} \left| P(\{x\}) - Q(\{x\}) \right| \leq W_r^r(P, Q).\] To prove the upper bound, since $\rho$ is upper bounded by $\Diam(\Omega)$, it suffices to construct a coupling $\mu$ that only moves mass into or out of each given point, but not both; that is, for each $x \in \X$, \[\min\{\mu(\{x\} \times (\Omega \sminus \{x\})), \mu((\Omega \sminus \{x\}) \times \{x\})\} = 0.\] One way of doing this is as follows. Fix an ordering $x_1,x_2,...$ of the elements of $\X$. For each $i \in \N$, define \[X_i := \sum_{\ell = 1}^i (P(x_\ell) - Q(x_\ell))_+ \quad \text{ and } \quad Y_i := \sum_{\ell = 1}^i (Q(x_\ell) - P(x_\ell))_+,\] and further define \[j_i := \min \{ j \in \N : X_i \leq Y_j \} \quad \text{ and } \quad k_i := \min \{ k \in \N : X_j \geq Y_i \}.\] Then, for each $i \in \N$, move $X_i$ mass from $\{x_1,...,x_i\}$ to $\{y_1,...,y_{j_i}\}$ and move $Y_i$ mass from $\{y_1,...,y_i\}$ to $\{x_1,...,x_{k_i}\}$. As $i \to \infty$, by construction of $X_i$ and $Y_i$, the total mass moved in this way is \[\mu((\X \times \X) \sminus \{(x,x) : x \in \X\}) = \lim_{i \to \infty} X_i + Y_i = \sum_{x \in \X} \left| P(x) - Q(x) \right|.\] \end{proof} \begin{customlemma}{\ref{lemma:nested_partitions_Wasserstein_bound}} Let $K$ be a positive integer. Suppose $\{\S_k\}_{k \in [K]}$ is a sequence of nested countable Borel partitions of $(\Omega,\rho)$, with $\S_0 = \Omega$. Then, for any $r \geq 1$ and any Borel probability distributions $P$ and $Q$ on $\Omega$, \[W_r^r(P, Q) \leq (\operatorname{Res}(\S_K))^r + \sum_{k = 1}^K \left( \operatorname{Res}(\S_{k - 1}) \right)^r \left( \sum_{S \in \S_k} \left| P(S) - Q(S) \right| \right).\] \end{customlemma} \begin{proof} Our proof follows the same ideas as and slightly generalizes of the proof of Proposition 1 in \citet{weed2017sharp}. Intuitively, to prove Lemma~\ref{lemma:nested_partitions_Wasserstein_bound} it suffices to find a transportation map such that For each $k \in [K]$, recursively define \[P_k := P - \sum_{j = 0}^{k - 1} \mu_k \quad \text{ and } \quad Q_k := Q - \sum_{j = 0}^{k - 1} \nu_k,\] where, for each $k \in [K]$, $\mu_k$ and $\nu_k$ are Borel measures on $\Omega$ defined for any $E \in \Sigma$ by \[\mu_k(E) := \sum_{S \in \S_k : P_k(S) > 0} \left( P_k(S) - Q_k(S) \right)_+ \frac{P_k(E \cap S)}{P_k(S)}\] and \[\nu_k(E) := \sum_{S \in \S_k : Q_k(S) > 0} \left( Q_k(S) - P_k(S) \right)_+ \frac{Q_k(E \cap S)}{Q_k(S)}.\] By construction of $\mu_k$ and $\nu_k$, each $\mu_k$ and $\nu_k$ is a non-negative measure and $\sum_{k = 1}^K \mu_k \leq P$ and $\sum_{k = 1}^K \nu_k \leq Q$. Furthermore, for each $k \in [K - 1]$, for each $S \in \S_k$, $\mu_{k + 1}(S) = \nu_{k + 1}(S)$, and \[\mu_k(\Omega) = \nu_k(\Omega) \leq \sum_{S \in \S_k} \left| P(S) - Q(S) \right|.\] Consequently, although $\mu$ and $\nu$ are not probability measures, we can slightly generalize the definition of Wasserstein distance by writing \[W_r^r \left( \mu_k, \nu_k \right) := \mu(\Omega) \inf_{\tau \in \Pi \left( \frac{\mu_k}{\mu_k(\Omega)}, \frac{\nu_k}{\nu_k(\Omega)}\right)} \E_{(X,Y) \sim \mu} \left[ \rho^r \left( X, Y \right) \right]\] (or $W_r^r(\mu_k, \nu_k) = 0$ if $\mu_k = \nu_k = 0$). In particular, this is convenient because we one can easily show that, by construction of the sequences $\{P_k\}_{k \in [K]}$ and $\{Q_k\}_{k \in [K]}$, \begin{equation} W_r^r(P, Q) \leq W_r^r \left( P_K, Q_K \right) + \sum_{k = 1}^K W_r^r \left(\mu_k, \nu_k \right). \label{ineq:decomposition} \end{equation} For each $k \in [K]$, Lemma~\ref{lemma:countable_support_bound} implies that \begin{align*} W_r^r(\mu_k,\nu_k) & \leq \sum_{S \in \S_{k - 1}} \left( \Diam(S) \right)^r \sum_{T \in \S_k : T \subseteq S} \left| P(T) - Q(T) \right| \\ & \leq \left( \operatorname{Res}(\S_{k - 1}) \right)^r \sum_{S \in \S_{k - 1}} \sum_{T \in \S_k : T \subseteq S} \left| P(T) - Q(T) \right| \\ & = \left( \operatorname{Res}(\S_{k - 1}) \right)^r \sum_{T \in \S_k} \left| P(T) - Q(T) \right|. \end{align*} Furthermore, for each $S \in \S_K$, $P_K = Q_K$, Lemma~\ref{lemma:measures_agree_on_partition} gives that \[W_r^r \left( P_K, Q_K \right) \leq \left( \operatorname{Res}(\S_K) \right)^r\] Plugging these last two inequalities into Inequality~\eqref{ineq:decomposition} gives the desired result: \[W_r^r(P, Q) \leq \left( \operatorname{Res}(\S_K) \right)^r + \sum_{k = 1}^K \left( \operatorname{Res}(\S_{k - 1}) \right)^r \sum_{S \in \S_k} \left| P(S) - Q(S) \right|.\] \end{proof} \begin{customlemma}{\ref{lemma:fine_refinement}} Suppose $\S$ and $\T$ are partitions of $(\Omega,\rho)$, and suppose $\S$ is countable. Then, there exists a partition $\S'$ of $(\Omega,\rho)$ such that: \begin{enumerate}[label=\alph*)] \item $|\S'| \leq |\S|$. \item $\operatorname{Res}(\S') \leq \operatorname{Res}(\S) + 2\operatorname{Res}(\T)$. \item $\T$ is a refinement of $\S'$. \end{enumerate} \end{customlemma} \begin{proof} Enumerate the elements of $\S$ as $S_1,S_2,...$. Define $S_0' := \emptyset$, and then, for each $i \in \{1,2,...\}$, recursively define \[S_i' := \left. \left( \bigcup_{T \in \T : T \cap S_i \neq \emptyset} T \right) \middle \sminus \left( \bigcup_{j = 1}^{i - 1} S_j' \right) \right.,\] and set $\S' = \{S_1',S_2',...\}$. Clearly, $|\S'| \leq |\S|$ (equality need not hold, as we may have some $S_i' = \emptyset$). By the triangle inequality, each \[\Diam(S_i') \leq \Diam \left( \bigcup_{T \in \T : T \cap S_i \neq \emptyset} T \right) \leq \delta_\S + 2 \delta_T.\] Finally, since $\T$ is a partition and we can write \[S_i' = \left. \left( \bigcup_{T \in \T : T \cap S_i \neq \emptyset} T \right) \middle \sminus \left( \bigcup_{j = 1}^{i - 1} \bigcup_{T \in \T : T \cap S_j' \neq \emptyset} T \right) \right.,\] $\T$ is a refinement of $\S'$. \end{proof} \section{Proof of Lower Bound} In this section, we provide a proof of our main lower bound, Theorem~\ref{thm:Wasserstein_distribution_estimation_lower_bound} in the main text. The proof consists of two main steps. First, we show that the minimax error of estimation in Wasserstein distance can be lower bounded by a product of two terms, one depending on the packing radius $R$ and the other depending on the minimax risk of estimating a particular discrete (i.e., multinomial) distribution under $\L_1$ loss. The second step is then to apply a minimax lower bound on the risk of estimating a multinomial distribution under $\L_1$ loss. These two steps respectively rely on two lemmas, Lemma~\ref{lemma:wasserstein_projections} and Lemma~\ref{lemma:multinomial_minimax_lower_bound} given below. The first lemma implies that, when a distribution $P$ is supported on a finite subset $\D$ of the sample space, then there exists an estimator $\hat P_\D$ of $\hat P$ that is supported on $\D$ is minimax optimal, up to a small constant factor. While this fact is relatively obvious for measure-theoretic metrics such as $\L_p$ distances, it is somewhat less obvious for Wasserstein distances, which also depend on metric properties of the space. This observation is key to lower bounding the minimax rate in terms of the minimax rate for estimating a discrete distribution. \begin{lemma}[Wasserstein Projections] Let $(\X,\rho)$ be a metric space and let $\D \subseteq \X$ be finite. Let $\P$ denote the family of all Borel probability distributions on $\X$, and let \[\P_\D := \{P \in \P : P(\D) = 1\}\] denote the set of distributions supported only on $\D$. Suppose $P \in \P_\D$ and $Q \in \P$. Then, \[\mathop{\arg\!\min}_{\tilde Q \in \P_\D} W_r(Q, \tilde Q) \neq \emptyset \quad \text{ and, for any } \quad Q' \in \mathop{\arg\!\min}_{\tilde Q \in \P'} W_r(Q, \tilde Q),\] we have $W_r(P, Q') \leq 2W_r(P, Q)$. \label{lemma:wasserstein_projections} \end{lemma} \begin{proof} Let $\{\S_x\}_{x \in \D}$ denote the Voronoi diagram of $\X$ with respect to $\D$; that is, for each $x \in \D$, let \[\S_x := \{y \in \X : x \in \mathop{\arg\!\min}_{z \in \D} \rho(x,y) \}.\] Since $\{S_x\}_{x \in \D}$ is a finite cover of $\X$, we can disjointify it (see Remark~\ref{remark:disjointification}) while retaining the property that, for every $x \in \D$ and every $y \in \S_x$, $\rho(x,y) = \min_{z \in \D} \rho(z,y)$; hence, we assume without loss of generality that $\{\S_x\}_{x \in \D}$ is a partition of $\X$. Then, there is a unique distribution $Q' \in \P_D$ such that, for each $x \in \D$, $Q'(\{x\}) = Q(\S_x)$. It is easy to see by definition of the Voronoi diagram that $Q' \in \mathop{\arg\!\min}_{\tilde Q \in \P_\D} W_r(Q, \tilde Q)$; the unique transportation map $\mu_* \in \Pi(Q,Q')$ such that each $\mu(\S_x,\{x\}) = Q(\S_x)$ clearly minimizes \[\E_{(X,Y) \sim \mu} \left[ \rho^r(X,Y) \right]\] over all $\mu \in \bigcup_{\tilde Q \in \P_\D} \Pi(Q, \tilde Q)$. Moreover, since $P \in \P_D$, by the triangle inequality and the definition of $Q'$, $W_r(P, Q') \leq W_r(P, Q) + W_r(Q, Q') \leq 2 W_r(P, Q)$. \end{proof} The second lemma is a simple minimax lower bound for the problem of estimating the mean vector of a multinomial distribution, under $\L_1$ loss. \begin{lemma}[Minimax Lower Bound for Mean of Multinomial Distribution] Suppose $k \leq 32 n$. Let $p \in \Delta^k$, and suppose $X_1,...,X_n \stackrel{IID}{\sim} \operatorname{Categorical}(p_1,...,p_k)$ are distributed IID according to a categorical distribution on $[k]$, with mean vector $p$. Then, we have the following minimax lower bound for estimating $p$ under $\L^1$-loss: \[\inf_{\hat p} \sup_{p \in \Delta^k} \E \left[ \|p - \hat p\|_1 \right] \geq \frac{3\log 2}{4096} \sqrt{\frac{k - 1}{n}},\] where the infimum is taken over all estimators (i.e., all (potentially randomized) functions $\hat p : [k]^n \to \Delta^k$ of the data). \label{lemma:multinomial_minimax_lower_bound} \end{lemma} Note that, while the above result is phrased for categorical distributions to simplify notation in the proof, the result is equivalent to a statement for multinomial distributions, since $\sum_{i = 1}^n X_i \sim \text{Multinomial}(n,p_1,...,p_k)$ and $X_1,...,X_n$ are assumed to be IID and therefore exchangeable. \begin{proof} We follow a standard procedure for proving minimax lower bounds based on Fano's inequality, as outlined in Section 2.6 of \citet{tsybakov2009introduction}. Let $p_0 = \left( 1/k, ...., 1/k \right) \in \Delta^K$ denote the uniform vector in $\Delta^k$. Let $\I := \left[ \lfloor \frac{k}{2} \rfloor \right]$. For each $j \in \I$, define $\phi_j : [k] \to \R^k$ by \[\phi_j := 1_{\{2j - 1\}} - 1_{\{2j\}},\] and, for each $\tau \in \{-1,1\}^\I$, define \[p_\tau := p_0 + \frac{c}{k} \sum_{j \in \I} \tau_j \phi_j,\] where \[c = \frac{1}{16} \sqrt{\frac{k - 1}{n}\log 2} \leq \frac{1}{2}.\] Note that, since $|c| \leq 1$ and, for each $j \in \I$, $\sum_{\ell \in [k]} \phi_j(\ell) = 0$, each $p_\tau \in \Delta^k$. Observe that, for any $\tau,\tau' \in \{-1,1\}^\I$, we have \[\|p_\tau - p_{\tau'}\|_1 = \frac{4 c \omega(\tau,\tau')}{k}, \quad \text{ where } \quad \omega(\tau,\tau') = \sum_{i \in \I} 1_{\{\tau_i \neq \tau_i'\}}\] denotes the Hamming distance between $\tau$ and $\tau'$. By the Varshamov-Gilbert bound (see, e.g., Lemma 2.9 of \citet{tsybakov2009introduction}), there exists a subset $T \subseteq \{-1,1\}^\I$ such that $\log |T| \geq \frac{\lfloor k/2 \rfloor \log 2}{8}$ and, for every $\tau, \tau' \in T$, \[\omega(\tau,\tau') \geq \frac{|\I|}{8} = \frac{\lfloor k/2 \rfloor}{8}, \quad \text{ and hence } \quad \|p_\tau - p_{\tau'}\|_1 \geq c \frac{\lfloor k/2 \rfloor}{2k}.\] Also, for any $\tau \in T$, \begin{align*} D_{KL}(p_\tau^n,p_0^n) & = n D_{KL}(p_\tau,p_0) \\ & = n \sum_{j \in [k]} p_{\tau,j} \log \left( \frac{p_{\tau,j}}{p_{0,j}}\right) \\ & = n \sum_{j \in \I} p_{\tau,2j - 1} \log \left( \frac{p_{\tau,2j - 1}}{1/k} \right) + p_{\tau,2j} \log \left( \frac{p_{\tau,2j}}{1/k} \right) \\ & = \frac{n |\I|}{k} \left( (1 - c) \log \left( 1 - c \right) + (1 + c) \log \left( 1 + c \right) \right) \end{align*} One can check (e.g., by Taylor expansion) that, for any $c \in (0,1/2)$, \[(1 - c) \log \left( 1 - c \right) + (1 + c) \log \left( 1 + c \right) < 2c^2.\] Thus, since $|\I| \leq k/2$, \[D_{KL}(p_\tau^n,p_0^n) \leq \frac{2 n |\I| c^2}{k} \leq n c^2.\] It follows that from the choice of $c$ (and noting that, by the assumptions that $k \leq 32n$, $c \in (0,1/2)$) that \[\frac{1}{|T|} \sum_{\tau \in T} D_{KL}(p_\tau^n, p_0^n) \leq nc^2 \leq \frac{\lfloor k/2 \rfloor \log 2}{128} \leq \frac{1}{16} \log |T|.\] Therefore, by Fano's method for lower bounds (see, e.g., Theorem 2.5 of \citet{tsybakov2009introduction}, with $\alpha = 1/16$ and \[s := \frac{c}{16} \leq c \frac{\lfloor k/2 \rfloor}{4k} \leq \frac{1}{2} \|p_\tau - p_{\tau'}\|_1,\] we have \begin{align*} \inf_{\hat p} \sup_{p \in \Delta^k} \E \left[ \|p - \hat p\|_1 \right] & \geq \inf_{\hat p} \sup_{p \in \Delta^k} c \frac{\lfloor k/2 \rfloor}{4k} \pr \left[ \|p - \hat p\|_1 \geq c \frac{\lfloor k/2 \rfloor}{4k} \right] \\ & \geq c \frac{\lfloor k/2 \rfloor}{4k} \frac{3}{16} \\ & \geq \frac{3\log 2}{4096} \sqrt{\frac{k - 1}{n}}. \end{align*} \end{proof} \begin{customthm}{\ref{thm:Wasserstein_distribution_estimation_lower_bound}} Let $(\Omega,\rho)$ be a metric space, and let $\P$ denote the set of Borel probability measures on $(\Omega,\rho)$. \[\inf_{\hat P : \X^n \to \P} \sup_{P \in \P} \E_{X_1,...,X_n \stackrel{IID}{\sim} P} \left[ W_r^r(P, \hat P(X_1,...,X_n)) \right] \geq c_r \sup_{k \in [32n]} R^r(k) \sqrt{\frac{k - 1}{n}},\] where \[c_r = \frac{3\log 2}{4096\cdot 2^r}.\] is independent of $n$ and the infimum is taken over all estimators (i.e., all (potentially randomized) functions $\hat P : \X^n \to \P$ of the data). \label{thm:Wasserstein_distribution_estimation_lower_bound_appendix} \end{customthm} \begin{proof} Let $k \leq 32n$, and let $\D$ be an $R(k)$-packing $\D$ of $(\Omega,\rho)$ with $|\D| = k$. Let $\P_\D$ denote the class of (discrete) distributions over $\D$. By Lemma~\ref{lemma:countable_support_bound}, Lemma~\ref{lemma:wasserstein_projections}, Lemma~\ref{lemma:multinomial_minimax_lower_bound}, and the definition of the packing radius (in that order) \begin{align*} \inf_{\hat P : \X^n \to \P} \sup_{P \in \P} \E \left[ W_r^r(\hat P, P) \right] & \geq \left( \Sep(\D) \right)^r \inf_{\hat P : \X^n \to \P} \sup_{P \in \P} \E \left[ \|\hat P - P\|_1 \right] \\ & \geq \left( \Sep(\D) \right)^r \inf_{\hat P : \X^n \to \P} \sup_{P \in \P_\D} \E \left[ \|\hat P - P\|_1 \right] \\ & \geq \left( \frac{\Sep(\D)}{2} \right)^r \inf_{\hat P : \X^n \to \P_\D} \sup_{P \in \P_\D} \E \left[ \|\hat P - P\|_1 \right] \\ & \geq \frac{3\log 2}{4096 \cdot 2^r} \left( \Sep(\D) \right)^r \sqrt{\frac{|\D| - 1}{n}} \\ & \geq \frac{3\log 2}{4096\cdot 2^r} R^r(k) \sqrt{\frac{k - 1}{n}}. \end{align*} The theorem follows by taking the supremum over $k \leq 32n$ on both sides. \end{proof} \section{Introduction} The Wasserstein metric is an important measure of distance between probability distributions, based on the cost of transforming either distribution into the other through mass transport, under a base metric on the sample space. Originating in the optimal transport literature,\footnote{The Wasserstein metric has been variously attributed to Monge, Kantorovich, Rubinstein, Gini, Mallows, and others; see Chapter 3 of \citep{villani2008optimalTransport} for detailed history.} the Wasserstein metric has, owing to its intuitive and general nature, been utilized in such diverse areas as probability theory and statistics, economics, image processing, text mining, robust optimization, and physics~\citep{villani2008optimalTransport,fournier2015rate,esfahani2015robustOptimization,gao2016distributionallyRobust}. In the analysis of image data, the Wasserstein metric has been used for various tasks such as texture classification and face recognition~\citep{sandler2011NMFImageAnalysis}, reflectance interpolation, color transfer, and geometry processing~\citep{solomon2015imageOptimalTrans}, image retrieval~\citep{rubner2000imageRetrieval}, and image segmentation~\citep{ni2009imageSegmentation}, and, in the analysis of text data, for tasks such as document classification~\citep{kusner2015documentDistances} and machine translation~\citep{zhang2016machineTranslation}. In contrast to a number of other popular notions of dissimilarity between probability distributions, such as $\L_p$ distances or Kullback-Leibler and other $f$-divergences~\citep{morimoto1963divergences,csiszar1964divergences,ali1966divergences}, which require distributions to be absolutely continuous with respect to each other or to a base measure, Wasserstein distance is well-defined between \emph{any} pair of probability distributions over a sample space equipped with a metric.\footnote{Hence, we use ``distribution estimation'' in this paper, rather than the more common ``density estimation''.} As a particularly important consequence, Wasserstein distances between discrete (e.g., empirical) distributions and continuous distributions are well-defined, finite, and informative (e.g., can decay to $0$ as the distributions become more similar). Partly for this reason, many central limit theorems and related approximation results~\citep{ruschendorf1985wasserstein,johnson2005central,chatterjee2008normalApproximation,rio2009upper,rio2011asymptotic,chen10SteinsMethod,reitzner2013central} are expressed using Wasserstein distances. Within machine learning and statistics, this same property motivates a class of so-called \emph{minimum Wasserstein distance estimates}~\citep{del1999CLT,del2003correction,bassetti2006minimum,bernton2017inferenceUsingWasserstein} of distributions, ranging from exponential distributions~\citep{baillo2016exponentialWasserstein} to more exotic models such as restricted Boltzmann machines (RBMs)~\citep{montavon2016wassersteinRBMs} and generative adversarial networks (GANs)~\citep{arjovsky2017wassersteinGAN}. This class of estimators also includes $k$-means and $k$-medians, where the hypothesis class is taken to be discrete distributions supported on at most $k$ points~\citep{pollard1982quantization}; more flexible algorithms such as hierarchical $k$-means~\citep{ho2017multilevel} and $k$-flats~\citep{tseng2000kFlats} can also be expressed in this way, using a more elaborate hypothesis classes. PCA can also be expressed and generalized to manifolds using Wasserstein distance minimization~\citep{boissard2015template}. These estimators are conceptually equivalent to empirical risk minimization, leveraging the fact that Wasserstein distances between the empirical distribution and distributions in the relevant hypothesis class are well-behaved. Moreover, these estimates often perform well in practice because they are free of both tuning parameters and strong distributional assumptions. For many of the above applications, it is important to understand how quickly the empirical distribution converges to the true distribution in Wasserstein distance, and whether there exist distribution estimators that converge more quickly. For example, \citet{canas2012learning} use bounds on Wasserstein convergence to prove learning bounds for $k$-means, while \citet{arora2017generalization} used the slow rate of convergence in Wasserstein distance in certain cases to argue that GANs based on Wasserstein distances fail to generalize with fewer than exponentially many samples in the dimension. To this end, the {\bf main contribution} of this paper is to identify, in a wide variety of settings, the minimax convergence rate for the problem of estimating a distribution using Wasserstein distance as a loss function. Our setting is very general, relying only on metric properties of the support of the distribution and the number of finite moments the distribution has; some diverse examples to which our results apply are given in Section~\ref{sec:examples}. Specifically, we assume only that the distribution is has some number of finite moments in a given metric. We then prove bounds on the minimax convergence rates of distribution estimation, utilizing covering numbers of the sample space for upper bounds and packing numbers for lower bounds. It may at first be surprising that positive results can be obtained under such mild assumptions; this highlights that the Wasserstein metric is quite a weak metric (see our Lemma 11 and the subsequent remark for discussion of this). Moreover, our results imply that, without further assumptions on the population distribution, the empirical distribution is typically minimax rate-optimal. Note that, while there has been previous work on upper bounds (discussed in Section~\ref{sec:related_work}), this paper is the first to study minimax lower bounds for this problem. \textbf{Organization: } The remainder of this paper is organized as follows. Section~\ref{sec:notation} provides notation required to formally state both the problem of interest and our results, while Section~\ref{sec:related_work} reviews previous work studying convergence of distributions in Wasserstein distance. Sections~\ref{sec:upper_bounds} and \ref{sec:lower_bounds} respectively contain our main upper and lower bound results. Since the proofs of the upper bounds, are fairly long, Appendices A and B provide high-level sketches of the proofs, followed by detailed proofs in Appendix C. The lower bound is proven in Appendix D Finally, in Section~\ref{sec:examples}, we apply our upper and lower bounds to identify minimax convergence rates in a number of concrete examples. Section~\ref{sec:conclusion} concludes with a summary of our contributions and suggested avenues for future work. \section{Notation and Problem Setting} \label{sec:notation} For any positive integer $n \in \N$, $[n] = \{1,2,...,n\}$ denotes the set of the first $n$ positive integers. For sequences $\{a_n\}_{n \in \N}$ and $\{b_n\}_{n \in \N}$ of non-negative reals, $a_n \lesssim b_n$ and, equivalently $b_n \gtrsim a_n$, indicate the existence of a constant $C > 0$ such that $\limsup_{n \to \infty} \frac{a_n}{b_n} \leq C$. $a_n \asymp b_n$ indicates $a_n \lesssim b_n \lesssim a_n$. \subsection{Problem Setting} For the remainder of this paper, fix a metric space $(\Omega,\rho)$, over which $\Sigma$ denotes the Borel $\sigma$-algebra, and let $\P$ denote the family of all Borel probability distributions on $\Omega$. The main object of study in this paper is the Wasserstein distance on $\P$, defined as follows: \begin{definition}[$r$-Wasserstein Distance] Given two Borel probability distributions $P$ and $Q$ over $\Omega$ and $r \in [1,\infty)$, the $r$-\emph{Wasserstein distance} $W_r(P,Q) \in [0,\infty]$ between $P$ and $Q$ is defined by \[W_r(P,Q) := \inf_{\mu \in \Pi(P,Q)} \left( \E_{(X,Y) \sim \mu} \left[ \rho^r \left( X, Y \right) \right] \right)^{1/r},\] where $\Pi(P,Q)$ denotes all couplings between $X \sim P$ and $Y \sim Q$; that is, \[\Pi(P,Q) := \left\{ \mu : \Sigma^2 \to [0,1] \middle| \text{ for all } A \in \Sigma, \mu(A \times \Omega) = P(A) \text{ and } \mu(\Omega \times A) = Q(A) \right\},\] is the set of joint probability measures over $\Omega \times \Omega$ with marginals $P$ and $Q$. \end{definition} Intuitively, $W_r(P,Q)$ quantifies the $r$-weighted total cost of transforming mass distributed according to $P$ to be distributed according to $Q$, where the cost of moving a unit mass from $x \in \Omega$ to $y \in \Omega$ is $\rho(x,y)$. $W_r(P,Q)$ is sometimes defined in terms of equivalent (e.g., dual) formulations; these formulations will not be needed in this paper. $W_r$ it is symmetric in its arguments and satisfies the triangle inequality, and, for all $P \in \P$, $W_r(P, P) = 0$. Thus, $W_r$ is always a pseudometric. Moreover, it is a proper metric (i.e., $W_r(P,Q) = 0 \Rightarrow P = Q$) if and only if $\rho$ is as well. This paper studies the following problem: \textbf{Formal Problem Statement:} Suppose $(\Omega,\rho)$ is a known metric space. Suppose $P$ is an unknown Borel probability distribution on $\Omega$, from which we observe $n$ IID samples $X_1,...,X_n \stackrel{IID}{\sim} P$. We are interested in studying the minimax rates at which $P$ can be estimated from $X_1,...,X_n$, in terms of the ($r^{th}$ power of the) $r$-Wasserstein loss. Specifically, we are interested in deriving finite-sample upper and lower bounds, in terms of only properties of the space $(\Omega,\rho)$, on the quantity \begin{equation} \inf_{\hat P} \sup_{P \in \P} \E_{X_1,...,X_n \stackrel{IID}{\sim} P} \left[ W_r^r \left( P, \hat P(X_1,...,X_n) \right) \right], \label{exp:minimax_error} \end{equation} where the infimum is taken over all estimators $\hat P$ (i.e., (potentially randomized) functions $\hat P : \Omega^n \to \P$ of the data). In the sequel, we suppress the dependence of $\hat P = \hat P(X_1,...,X_n)$ in the notation. \subsection{Definitions for Stating our Results} Here, we give notation and definitions needed to state our main results in Sections~\ref{sec:upper_bounds} and \ref{sec:lower_bounds}. Let $2^\Omega$ denote the power set of $\Omega$. Let $\SS \subseteq 2^{2^{\Omega}}$ denote the family of all Borel partitions of $\Omega$: \[\SS := \left\{ \S \subseteq \Sigma \quad : \quad \Omega \subseteq \bigcup_{S \in \S} S \quad \text{ and } \quad \forall S,T \in \S, S \cap T = \emptyset \right\}.\] We now define some metric notions that will later be useful for bounding Wasserstein distances: \begin{definition}[Diameter and Separation of a Set, Resolution of a Partition] For any set $S \subseteq \Omega$, the \emph{diameter $\Diam(S)$ of $S$} is defined by $\Diam(S) := \sup_{x,y \in S} \rho(x,y)$, and the \emph{separation $\Sep(S)$ of $S$} is defined by $\Sep(S) := \inf_{x \neq y \in S} \rho(x,y)$. If $\S \in \SS$ is a partition of $\Omega$, then the \emph{resolution $\operatorname{Res}(\S)$ of $\S$} defined by $\operatorname{Res}(\S) := \sup_{S \in \S} \Diam(S)$ is the largest diameter of any set in $\S$. \end{definition} We now define the covering and packing number of a metric space, which are classic and widely used measures of the size or complexity of a metric space \citep{dudley1967coveringNumbers,haussler1995sphere,zhou2002covering,zhang2002covering}. Our main convergence results will be stated in terms of these quantities, as well as the packing radius, which acts, approximately, as the inverse of the packing number. \begin{definition}[Covering Number, Packing Number, and Packing Radius of a Metric Space] The \emph{covering number $N : (0,\infty) \to \N$ of $(\Omega,\rho)$} is defined for all $\epsilon > 0$ by \[N(\epsilon) := \min \left\{ |\S| : \S \in \SS \quad \text{ and } \quad \operatorname{Res}(\S) \leq \epsilon \right\}.\] The \emph{packing number $M : (0,\infty) \to \N$ of $(\Omega,\rho)$} is defined for all $\epsilon > 0$ by \[M(\epsilon) := \max \left\{ |S| : S \subseteq \Omega \quad \text{ and } \quad \Sep(S) \geq \epsilon \right\}.\] Finally, the \emph{packing radius $R : \N \to [0,\infty]$} is defined for all $n \in \N$ by \[R(n) := \sup\{ \Sep(S) : S \subseteq \Omega \quad \text{ and } \quad |S| \geq n\}.\] Sometimes, we use the covering or packing number of a metric space, say $(\Theta, \tau)$, other than $(\Omega,\rho)$; in such cases, we write $N(\Theta;\tau;\epsilon)$ or $M(\Theta;\tau;\epsilon)$ rather than $N(\epsilon)$ or $M(\epsilon)$, respectively. For specific $\epsilon > 0$, we will also refer to $N(\Theta;\tau;\epsilon)$ as the \emph{$\epsilon$-covering number of $(\Theta,\tau)$}. \end{definition} \begin{remark} The covering and packing numbers of a metric space are closely related. In particular, for any $\epsilon > 0$, we always have \begin{equation} M(\epsilon) \leq N(\epsilon) \leq M(\epsilon/2). \label{ineq:covering_packing_relationship} \end{equation} The packing number and packing radius also have a close approximate inverse relationship. In particular, for any $\epsilon > 0$ and $n \in \N$, we always have \begin{equation} R(M(\epsilon)) \geq \epsilon \quad \text{ and } \quad M(R(n)) \geq n. \label{ineq:packing_number_radius_relationship} \end{equation} However, it is possible that $R(M(\epsilon)) > \epsilon$ or $M(R(n)) > n$. \end{remark} Finally, when we consider unbounded metric spaces, we will require some sort of concentration conditions on the probability distributions of interest, to obtain useful results. Specifically, we an appropriately generalized version of the moment of the distribution: \begin{remark} We defined the covering number slightly differently from usual (using partitions rather than covers). However, the given definition is equivalent to the usual definition, since (a) any partition is itself a cover (i.e., a set $\C \subseteq 2^\Omega$ such that $\Omega \subseteq \bigcup_{C \in \C} C$), and (b), for any countable cover $\C := \{C_1,C_2,...\} \subseteq 2^\Omega$, there exists a partition $\S \in \SS$ with $|\S| \leq |\C|$ and each $S_i \subseteq C_i$, defined recursively by $S_i := C_i \sminus \bigcup_{j = 1}^{i - 1} S_i$. $\S$ is often called the \emph{disjointification} of $\C$. \label{remark:disjointification} \end{remark} \begin{definition}[Metric Moments of a Probability Distribution] For any $\ell \in [0,\infty]$, probability measure $P \in \P$, and $x \in \Omega$, the \emph{$\ell^{th}$ metric moment $m_{\ell,x}(P)$ of $P$ around $x$} is defined by \[m_{\ell,x}(P) := \left( \E_{Y \sim P} \left[ \left( \rho(x,Y) \right)^\ell \right] \right)^{1/\ell} \in [0,\infty],\] using the appropriate limit if $\ell = \infty$. The chosen reference point $x$ only affects constant factors since, \[\text{ for all } x,x' \in \Omega, \quad \left| m_{\ell,x}^\ell(P) - m_{\ell,x'}^\ell(P) \right| \leq \left( \rho(x,x') \right)^\ell.\] Note that, if $\Omega$ has linear structure with respect to which $\rho$ is translation-invariant (e.g., if $(\Omega,\rho)$ is a Fr\'echet space), we can state our results more simply in terms of $m_\ell(P) := \inf_{x \in \Omega} m_{\ell,x}(P)$. As an example, if $\Omega = \R$ and $\rho(x,y) = |x - y|$, then $m_2(P)$ is precisely the standard deviation of $P$. \end{definition} \section{Related Work} \label{sec:related_work} A long line of work~\citep{dudley1969speed,ajtai1984optimalMatchings,canas2012learning,dereich2013constructive,boissard2014mean,fournier2015rate,weed2017sharp,lei2018convergence} has studied the rate of convergence of the empirical distribution to the population distribution in Wasserstein distance. In terms of upper bounds, the most general and tight upper bounds are the recent works of \citep{weed2017sharp} and \citep{lei2018convergence}. As we describe below, while these two papers overlap significantly, neither supersedes the other, and our upper bound combines the key strengths of those in \citep{weed2017sharp} and \citep{lei2018convergence}. The results of \citep{weed2017sharp} are expressed in terms of a particular notion of dimension, which they call the \textit{Wasserstein dimension} $s$, since they derive convergence rates of order $n^{-r/s}$ (matching the $n^{-r/D}$ rate achieved on the unit cube $[0,1]^D$). The definition of $s$ is complex (e.g., it depends on the sample size $n$), but \citep{weed2017sharp} show that, in many cases, $s$ converges to certain common definitions of the intrinsic dimension of the support of the distribution. This paper overcomes three main limitations of \citep{weed2017sharp}: \begin{enumerate}[nolistsep,leftmargin=2em] \item The upper bounds of \citep{weed2017sharp} apply only to totally bounded metric spaces. In contrast, our upper bounds permit unbounded metric spaces under the assumption that the distribution $P$ has some finite moment $m_\ell(P) < \infty$. The results of \citep{weed2017sharp} correspond to the special case $\ell = \infty$. \item Their main upper bound (their Proposition 10) only holds when $s > 2r$, with constant factors diverging to infinity as $s \downarrow 2r$. Hence, their rates are loose when $r$ is large or when the data have low intrinsic dimension. In contrast, our upper bound is tight even when $s \leq 2r$. \item As we discuss in our Example~\ref{ex:Lipschitz_Ball}, the upper bound of \citep{weed2017sharp} becomes loose as the Wasserstein dimension $s$ approaches $\infty$, limiting its utility in infinite-dimensional function spaces. In contrast, we show that our upper and lower bounds match for several standard function spaces. \end{enumerate} Intuitively, we find that the finite-sample bounds of \citep{weed2017sharp} are tight when the intrinsic dimension of the data lies in an interval $[a, b]$ with $2r < a < b < \infty$, but they can be loose outside this range. In contrast, we find our results give tight rates for a larger class of problems. On the other hand, \citep{lei2018convergence} focuses on the case where $\Omega$ is a (potentially unbounded and infinite-dimensional) Banach space, under moment assumptions on the distributions. Thus, while the results of \citep{lei2018convergence} cover interesting cases such as infinite-dimensional Gaussian processes, they do not demonstrate that convergence rates improve when the intrinsic dimension of the support of $P$ is smaller than that of $\Omega$ (unless this support lies within a \textit{linear} subspace of $\Omega$). As a simple example, if the distribution is in fact supported on a finite set of $k$ linearly independent points, the bound of \citep{lei2018convergence} implies only a convergence rate, whereas we give a bound of order $O(\sqrt{k/n})$. Although we do not delve into this here, our results (unlike those of \citep{lei2018convergence}) should also benefit from the multi-scale behavior discussed in Section 5 of \citep{weed2017sharp}; namely, much faster convergence rates are often observed for small $n$ than for large $n$. This may help explain why algorithms such as functional $k$-means~\citep{garcia2015functionalKMeans} work in practice, even though the results of \citep{lei2018convergence} imply only a slow convergence rate of $O\left( (\log n)^{-p} \right)$, for some constant $p > 0$, in this case. Under similarly general conditions, \citep{sriperumbudur2010integralProbabilityMetrics,sriperumbudur2012empirical} have studied the related problem of estimating the Wasserstein distance between two unknown distributions given samples from those two distributions. Since one can estimate Wasserstein distances by plugging in empirical distributions, our upper bounds imply upper bounds for Wasserstein distance estimation. These bounds are tighter, in several cases, than those of \citep{sriperumbudur2010integralProbabilityMetrics,sriperumbudur2012empirical}; for example, when $\X = [0,1]^D$ is the Euclidean unit cube, we give a rate of $n^{-1/D}$, whereas they give a rate of $n^{-\frac{1}{D + 1}}$. Minimax rates for this problem are currently unknown, and it is presently unclear to us under what conditions recent results on estimation of $\L_1$ distances between discrete distributions~\citep{jiao2017minimaxL1} might imply an improved rate as fast as $\left( n \log n \right)^{-1/D}$ for estimation of Wasserstein distance. To the best of our knowledge, minimax lower bounds for distribution estimation under Wasserstein loss remain unstudied, except in the very specific case when $\Omega = [0,1]^D$ is the Euclidean unit cube and $r = 1$~\citep{liang2017well}. As noted above, most previous works have focused on studying convergence rate of the empirical distribution to the true distribution in Wasserstein distance. For this rate, several lower bounds have been established, matching known upper bounds in many cases. However, many distribution estimators besides the empirical distribution can be considered. For example, it is tempting (especially given the infinite dimensionality of the distribution to be estimated) to try to reduce variance by techniques such as smoothing or importance sampling~\citep{bucklew2013introduction}. Our lower bound results, given in Section~\ref{sec:lower_bounds}, imply that the empirical distribution is already minimax optimal, up to constant factors, in many cases. \section{Upper Bounds} \label{sec:upper_bounds} In this section, we present our main upper bounds on the convergence rate of the empirical distribution to the true distribution in Wasserstein distance. We begin by presenting a simpler result for the case of totally bounded metric spaces, followed by a more complex but general result for arbitrary metric spaces under finite-moment assumptions on the distribution. \begin{theorem} Let $(\Omega,\rho)$ be a metric space on which $P$ is a Borel probability measure. Let $\hat P$ denote the empirical distribution of $n$ IID samples $X_1,...,X_n \stackrel{IID}{\sim} P$, give by \begin{equation} \hat P(S) := \frac{1}{n} \sum_{i = 1}^n 1_{\{X_i \in S\}}, \quad \forall S \in \Sigma. \label{eq:empirical_distribution} \end{equation} Then, for any non-increasing sequence $\{\epsilon_k\}_{k \in [K]} \in (0,\infty)^K$ with $\epsilon_0 = \Diam(\Omega)$, \[\E \left[ W_r^r(P, \hat P) \right] \leq \epsilon_K^r + \frac{1}{\sqrt{n}} \sum_{k = 1}^K \left( \sum_{j = k}^K 2^{K - j} \epsilon_j \right)^r \sqrt{N(\epsilon_k) - 1}.\] \label{thm:expectation_bound} \end{theorem} In the proof of the above theorem, the sequence $\{\epsilon_k\}_{k \in [K]} \in (0,\infty)^K$ gives the resolutions of a sequence of increasingly fine partitions of $\Omega$. The basic idea of the proof is to recursively bound the error over each partition at resolution $\epsilon_j$ in terms of $\epsilon_j$ and the error over the partition of resolution $\epsilon_{j + 1}$. The parameter $K$ restricts us to a particular finite resolution, with optimal value typically increasing with $n$. Note that this ``multi-resolution'' proof approach has been utilized in several special cases, apparently originating in the analysis of Our Theorem~\ref{thm:expectation_bound} is most comparable to the upper bound (Proposition 10) of \citet{weed2017sharp}. Theorem~\ref{thm:expectation_bound} requires $(\Omega,\rho)$ to be totally bounded in order for $N(\epsilon)$ to be finite. Next, we present a more complex bound, which, under the additional assumption that $P$ has some number of finite moments, is often finite even when $(\Omega,\rho)$ is \textit{not} totally bounded. The key idea of the proof is to partition $\Omega = \bigcup_{k = 0}^\infty B_k$ into bounded subsets, over each of which we can apply a bound similar to Theorem~\ref{thm:expectation_bound}. Thus, instead of the covering number $N(\epsilon)$ of $(\Omega,\rho)$, this result uses covering numbers $N(B_k,\rho,\epsilon)$ of a partition $\Omega = \bigcup_{k = 0}^\infty B_k$ into totally bounded subsets. \begin{theorem}[General Upper Bound for Unbounded Metric Spaces] Let $x_0 \in \Omega$ and suppose $m_{\ell,x_0}(P) \in [1, \infty)$. Let $J \in \N$. Fix two non-decreasing real-valued sequences $\{w_k\}_{k \in \N}$ and $\{\epsilon_j\}_{j \in \N}$, of which $\{w_k\}_{k \in \N}$ is non-decreasing with $w_0 = 0$ and $\lim_{k \to \infty} w_k = \infty$ and $\{\epsilon_j\}_{j \in [J]}$ is non-increasing. For each $k \in \N$, define $B_k(x_0) := \left\{ y \in \Omega : w_k \leq \rho(x_0, x) < w_{k + 1} \right\}$. Then, \begin{align*} \E \left[ W_r^r(P, \hat P) \right] & \leq m_{\ell,x_0}^\ell(P) \sum_{k \in \N} w_k^{-\ell} \left( \epsilon_J \right)^r + 2^r w_k^{r - \ell/2} \min \left\{ 2w_k^{-\ell/2}, \sqrt{\frac{1}{n}} \right\} \\ & \hspace{2cm} + \sum_{j = 1}^J \left( \sum_{t = j}^J 2^{J - t} \epsilon_t \right)^r \min \left\{ 2w_k^{-\ell}, \sqrt{\frac{w_k^{-\ell}}{n} N(B_k,\rho,\epsilon_j)} \right\}. \end{align*} \label{thm:unbounded_upper_bound} \end{theorem} In the above, $w_k$ corresponds to radii of the partition of $\Omega = \bigcup_{k = 0}^\infty B_k$ into a sequence of ``spherical shells'', whereas $\epsilon_j$, as in the previous result, corresponds to resolutions of partitions of the $B_k$'s. As with $K$ in the previous result, $J$ is used to ensure that we restrict ourselves to a particular finite resolution. The $\min$ terms appear because, for large $k$, the error is controlled by the fact that $P(B_k)$ is small (due to the moment assumption), rather than using a covering of $B_k$. \section{Lower Bounds} \label{sec:lower_bounds} In this section, we provide a minimax lower bound (over the family $\P$ of all Borel distributions on $\Omega$) for density estimation in Wasserstein distance (that is, the quantity \begin{equation} \inf_{\hat P : \X^n \to \P} \sup_{P \in \P} \E_{X_1,...,X_n \stackrel{IID}{\sim} P} \left[ W_r^r \left( P, \hat P \right) \right], \label{exp:distribution_estimation_minimax} \end{equation} where the infimum is over all estimators $\hat P$ of $P$ (i.e., all (potentially randomized) functions $\hat P : \Omega^n \to \P$)). Our bound depends primarily on the packing radius $R$ of $(\Omega,\rho)$, and, presently, we handle only the case without finite-moment assumptions on $P$. However, we show in the next section that this often implies tight lower bounds when enough (roughly, $\ell \geq \max\{D,2r\}$) moments exist. \begin{theorem} Let $(\Omega,\rho)$ be a metric space, on which $\P$ is the set of Borel probability measures. Then, \[\inf_{\hat P : \X^n \to \P} \sup_{P \in \P} \E_{X_1,...,X_n \stackrel{IID}{\sim} P} \left[ W_r^r(P, \hat P(X_1,...,X_n)) \right] \geq c_r \sup_{k \in [32n]} R^r(k) \sqrt{\frac{k - 1}{n}},\] where $c_r = \frac{3\log 2}{2^{r + 12}}$ depends only on $r$. \label{thm:Wasserstein_distribution_estimation_lower_bound} \end{theorem} \section{Example Applications} \label{sec:examples} Our theorems in the previous sections are quite abstract and have many tuning parameters. Thus, we conclude by exploring applications of our results to cases of interest. In each of the following examples, $P$ is an unknown Borel probability measure over the specified $\Omega$, from which we observe $n$ IID samples. For upper bounds, $\hat P$ denotes the empirical distribution~\eqref{eq:empirical_distribution} of these samples. \begin{example}[Finite Space] Consider the case where $\Omega$ is a finite set, over which $\rho$ is the discrete metric given, for some $\delta > 0$, by $\rho(x, y) = \delta 1_{\{x = y\}}$, for all $x,y \in \Omega$. Then, for any $\epsilon \in (0,\delta)$, the covering number is $N(\epsilon) = |\Omega|$. Thus, setting $K = 1$ and sending $\epsilon_1 \to 0$ in Theorem~\ref{thm:expectation_bound} gives \[\E \left[ W_r^r(P, \hat P) \right] \leq \delta^r \sqrt{\frac{|\Omega| - 1}{n}}.\] On the other hand, $R(|\Omega|) = \delta$, and so, setting $k = |\Omega|$ in Theorem~\ref{thm:Wasserstein_distribution_estimation_lower_bound} yields \[\inf_{\hat P} \sup_{P \in \P} \E_{X_1,...,X_n \stackrel{IID}{\sim} P} \left[ W_r^r(P, \hat P) \right] \gtrsim \delta^r \sqrt{\frac{|\Omega| - 1}{n}}.\] \label{ex:discrete_bound} \end{example} \begin{example}[Unit Cube, Euclidean Metric] Consider the case where $\Omega = \R^D$ is the unit cube and $\rho$ is the Euclidean metric. Assuming $\ell > r$, using the fact that $N \left( B_k, \rho, \epsilon \right) \leq \left( \frac{3 w_k}{\epsilon} \right)^D$~\citep{pollard1990empirical} and plugging $\epsilon_j = 2^{-2j}$ and $w_k = 2^k$ into Theorem~\ref{thm:unbounded_upper_bound} gives (after a straightforward but very tedious calculation) a constant $C_{D,r,\ell}$ depending only on $D$, $r$, and $\ell$ such that \begin{equation} \E \left[ W_r^r(P, \hat P) \right] \leq C_{D,\ell,r} m_\ell^\ell(P) \left( n^{\frac{\ell - r}{\ell}} + 2^{-2Jr} + \sum_{j = 1}^J 2^{(D - 2r)j} \right). \label{ineq:general_Euclidean_bound} \end{equation} Of these three terms, the first depends only on the number $\ell$ of finite moments $P$ is assumed to have and the order $r$ of the Wasserstein distance, whereas the second and third terms depend on choosing the parameter $J$. The optimal choice of $J$ scales with the sample size $n$ at a rate depending on the quantity $D - 2r$. Specifically, if $D = 2r$, then setting $J \asymp \frac{1}{4r} \log_2 n$ gives a rate of $\E \left[ W_r^r(P, \hat P) \right] \lesssim n^{\frac{\ell - r}{\ell}} + n^{-1/2} \log n$. If $D \neq 2r$, then~\eqref{ineq:general_Euclidean_bound} reduces to \[\E \left[ W_r^r(P, \hat P) \right] \leq C_{D,\ell,r} m_\ell^\ell(P) \left( n^{\frac{\ell - r}{\ell}} + 2^{-2Jr} + \frac{2^{(D - 2r)J} - 1}{2^{D - 2r} - 1} \right).\] Then, if $D > 2r$, sending $J \to \infty$ gives $\E \left[ W_r^r(P, \hat P) \right] \lesssim n^{\frac{\ell - r}{\ell}} + n^{-1/2}$. Finally, if $D < 2r$, then setting $J \asymp \frac{1}{2D} \log n$ gives $\E \left[ W_r^r(P, \hat P) \right] \lesssim n^{\frac{\ell - r}{\ell}} + n^{-\frac{r}{D}}$. To summarize \[\E \left[ W_r^r(P, \hat P) \right] \lesssim n^{\frac{\ell - r}{\ell}} + \left\{ \begin{array}{ll} n^{-1/2} & \text { if } 2r > D \\ n^{-1/2} \log n & \text { if } 2r = D \\ n^{-r/D} & \text { if } 2r < D \end{array} \right.\] (reproducing Theorem 1 of \citep{fournier2015rate}). On the other hand, it is easy to check that the packing radius $R$ satisfies $R(n) \geq n^{-1/D}$ and $R(2) \geq \sqrt{D}$. Thus, Theorem~\ref{thm:Wasserstein_distribution_estimation_lower_bound} with $k = n$ and $k = 2$ yields \[\inf_{\hat P} \sup_{P \in \P} \E \left[ W_r^r(\hat P, P) \right] \gtrsim \max \left\{ (n + 1)^{-r/D}, D^{r/2} n^{-1/2} \right\}.\] Together, these bounds give the following minimax rates for density estimation in Wasserstein loss: \[\inf_{\hat P} \sup_{P \in \P} \E \left[ W_r^r(\hat P, P) \right] \asymp \left\{ \begin{array}{ll} n^{-1/2} & \text { if } \ell > 2r > D \\ n^{-r/D} & \text { if } 2r < D, \ell > \frac{Dr}{D - r} \end{array} \right.\] When $2r = D$ and $\ell > 2r$, our upper and lower bounds are separated by a factor of $\log n$. The main result of \citep{ajtai1984optimalMatchings} implies that, for the case $D = 2$ and $r = 1$, the empirical distribution converges as $n^{-1/2} \log n$, suggesting that the $\log n$ factor in our upper bound may be tight. Further generalization of Theorem~\ref{thm:Wasserstein_distribution_estimation_lower_bound} is needed to give lower bounds when both $D, \ell \leq 2r$ or when $D > 2r$ and $\ell \leq \frac{Dr}{D - r}$. \label{ex:unit_cube_lower_bound} \end{example} The next example demonstrates how the rate of convergence in Wasserstein metric depends on properties of the metric space $(\Omega,\rho)$ at both large and small scales. Specifically, if we discretize $\Omega$, then the phase transition at $2r = D$ disappears. \begin{example} Suppose $\Omega = \mathbb{Z}^D$ is a $D$-dimensional grid of integers and $\rho$ is $\ell_\infty$-metric (given by $\rho(x,y) = \max_{j \in [D]} |x_j - y_j|$). Since $\Z^D \subseteq \R^D$ and the $\ell_\infty$ and Euclidean metrics are topologically equivalent, the upper bounds from Example~\ref{ex:unit_cube_lower_bound} clearly apply, up to a factor of $\sqrt{D}$. However, we also have the fact that, whenever $\epsilon < 1$, $N(B_k,\rho,\epsilon) = w_k^D$. Therefore, setting $J = 0$, $\epsilon_0 = 0$, and $w_k = 2^k$ in Theorem~\ref{thm:unbounded_upper_bound} gives, for a constant $C_{D,\ell,r}$ depending only on $D$, $\ell$, and $r$, \begin{align*} \E \left[ W_r^r(P, \hat P) \right] & \leq C_{D,\ell,r} m_\ell^\ell(P) \left( n^{\frac{\ell - r}{\ell}} + \sum_{k \in \N} \sqrt{\frac{2^{(D - \ell)k}}{n}} \right). \end{align*} When $\ell > D$, this reduces to $\E \left[ W_r^r(P, \hat P) \right] \lesssim n^{\frac{\ell - r}{\ell}} + n^{-1/2}$, giving a tighter rate than in Example~\ref{ex:unit_cube_lower_bound} when $2r \leq D < \ell$. To the best of our knowledge, no prior results in the literature imply this fact. \end{example} Finally, we consider distributions over an infinite dimensional space of smooth functions. \begin{example}[H\"older Ball, $\L_\infty$ Metric] Suppose that, for some $\alpha \in (0,1]$, \[\Omega := \left\{ f [0,1]^D \to [-1,1] \quad \middle| \quad \forall x,y \in [0,1]^D, \quad |f(x) - f(y)| \leq \|x - y\|_2^\alpha \right\}\] is the class of unit $\alpha$-H\"older functions on the unit cube and $\rho$ is the $\L_\infty$-metric given by \[\rho(f,g) = \sup_{x \in [0,1]^D} |f(x) - g(x)|, \quad \text{ for all } f,g \in \Omega.\] The covering and packing numbers of $(\Omega,\rho)$ are well-known to be of order $\exp \left( \epsilon^{-D/\alpha} \right)$ \citep{devore1993approximation}; specifically, there exist positive constants $0 < c_1 < c_2$ such that, for all $\epsilon \in (0,1)$, \[c_1 \exp \left( \epsilon^{-D/\alpha} \right) \leq N(\epsilon) \leq M(\epsilon) \leq c_2 \exp \left( \epsilon^{-D/\alpha} \right).\] Since $\Diam(\Omega) = 2$, applying Theorem~\ref{thm:expectation_bound} with $K = 1$ and \[\epsilon_1 = \left( 2\log n - (\alpha r/D) \log \log n \right)^{-\frac{\alpha r}{D}} \quad \text{ gives } \quad \E \left[ W_r^r(P, \hat P) \right] \lesssim \left( \log n \right)^{\frac{- \alpha r}{D}}.\] Conversely, Inequality~\eqref{ineq:packing_number_radius_relationship} implies $R(n) \geq \left( \log(n/c_1) \right)^\frac{-\alpha}{D}$, and so setting $k = n$ in Theorem~\ref{thm:Wasserstein_distribution_estimation_lower_bound} gives \[\inf_{\hat P} \sup_{P \in \P} \E_{X_1,...,X_n \stackrel{IID}{\sim} P} \left[ W_r^r(P, \hat P) \right] \gtrsim \left( \frac{1}{\log(n/c_1)} \right)^\frac{\alpha r}{D},\] showing that distribution estimation over $(\P,W_r^r)$ has the extremely slow minimax rate $\left( \log n \right)^\frac{-\alpha r}{D}$. Although we considered only $\alpha \in (0,1]$ (due to the notational complexity of defining higher-order H\"older spaces), analogous rates hold for all $\alpha > 0$. Also, since our rates depend only on covering and packing numbers of $\Omega$, identical rates can be derived for related Sobolev and Besov classes. Note that the Wasserstein dimension used in the prior work \citep{weed2017sharp} is of order $\frac{D}{\alpha} \log n$, and so their upper bound (their Proposition 10) gives a rate of $n^{-\frac{\alpha r}{D \log n}} = \exp \left( -\frac{\alpha r}{D} \right)$, which fails to converge as $n \to \infty$. \label{ex:Lipschitz_Ball} \end{example} One might wonder why we are interested in studying Wasserstein convergence of distributions over spaces of smooth functions, as in Example~\ref{ex:Lipschitz_Ball}. Our motivation comes from the historical use of smooth function spaces have been widely used for modeling images and other complex naturalistic signals \citep{mallat1999wavelet,peyre2011numerical,sadhanala2016totalVariation}. Empirical breakthroughs have recently been made in generative modeling, particularly of images, based on the principle of minimizing Wasserstein distance between the empirical distribution and a large class of models encoded by a deep neural network~\citep{montavon2016wassersteinRBMs,arjovsky2017wassersteinGAN,gulrajani2017improved}. However, little is known about theoretical properties of these methods; while there has been some work studying the optimization landscape of such models~\citep{nagarajan2017gradient,liang2018interaction}, we know of far less work exploring their \textit{statistical} properties. Given the extremely slow minimax convergence rate we derived above, it must be the case that the class of distributions encoded by such models is far smaller or sparser than $\P$. An important avenue for further work is thus to explicitly identify stronger assumptions that can be made on distributions over interesting classes of signals, such as images, to bridge the gap between empirical performance and our theoretical understanding. \section{Conclusion} \label{sec:conclusion} In this paper, we derived upper and lower bounds for distribution estimation under Wasserstein loss. Our upper bounds generalize prior results and are tighter in certain cases, while our lower bounds are, to the best of our knowledge, the first minimax lower bounds for this problem. We also provided several concrete examples in which our bounds imply novel convergence rates. \subsection{Future Work} We studied minimax rates over the very large entire class $\P$ of all distributions with some number of finite moments. It would be useful to understand how minimax rates improve when additional assumptions, such as smoothness, are made (see, e.g., \citep{liang2017well} for somewhat improved upper bounds under smoothness assumptions when $(\Omega,\rho)$ is the Euclidean unit cube). Given the slow convergence rates we found over $\P$ in many cases, studying minimax rates under stronger assumptions may help to explain the relatively favorable empirical performance of popular distribution estimators based on empirical risk minimization in Wasserstein loss. Moreover, while rates over all of $\P$ are of interest only for very weak metrics such as the Wasserstein distance (as stronger metrics may be infinite or undefined), studying minimax rates under additional assumptions will allow for a better understanding of the Wasserstein metric in relation to other commonly used metrics. \newpage \subsubsection*{Acknowledgments} This work was partly supported by a NSF Graduate Research Fellowship DGE-1252522 to S.S. {\small \bibliographystyle{plainnat}

https://arxiv.org/abs/1310.2972

Hypergraph Colouring and Degeneracy

A hypergraph is "$d$-degenerate" if every subhypergraph has a vertex of degree at most $d$. A greedy algorithm colours every such hypergraph with at most $d+1$ colours. We show that this bound is tight, by constructing an $r$-uniform $d$-degenerate hypergraph with chromatic number $d+1$ for all $r\geq2$ and $d\geq1$. Moreover, the hypergraph is triangle-free, where a "triangle" in an $r$-uniform hypergraph consists of three edges whose union is a set of $r+1$ vertices.

\section{Introduction} \citet{EL75} proved the following fundamental result about colouring hypergraphs\footnote{A \emph{hypergraph} $G$ consists of a set $V(G)$ of \emph{vertices} and a set $E(G)$ of subsets of $V(G)$ called \emph{edges}. A hypergraph is \emph{$r$-uniform} if every edge has size $r$. A \emph{graph} is a $2$-uniform hypergraph. A hypergraph $H$ is a \emph{subhypergraph} of a hypergraph $G$ if $V(H)\subseteq V(G)$ and $E(H)\subseteq E(G)$. A \emph{colouring} of a hypergraph $G$ assigns one colour to each vertex in $V(G)$ such that no edge in $E(G)$ is monochromatic. The \emph{chromatic number} of $G$, denoted by $\chi(G)$, is the minimum number of colours in a colouring of $G$. A colouring of $G$ can be thought of as a partition of $V(G)$ into \emph{independent sets}, each containing no edge. The \emph{degree} of a vertex $v$ is the number of edges that contain $v$. See the textbook of \citet{BergeBook} for other notions of degree in a hypergraph.} \begin{thm}[\citep{EL75}] \label{EL75} For fixed $r$, every $r$-uniform hypergraph with maximum degree $\Delta$ has chromatic number at most $O(\Delta^{1/(r-1)})$. \end{thm} Theorem~\ref{EL75} implies that every $r$-uniform hypergraph with maximum degree $\Delta$ has an independent set of size at least $\Omega(n/\Delta^{1/(r-1)})$. \citet{Spencer} proved the following stronger bound. \begin{thm}[\citep{Spencer}] \label{Spencer} For fixed $r$, every $r$-uniform hypergraph with $n$ vertices and average degree $d$ has an independent set of size at least $\Omega(n/d^{1/(r-1)})$. \end{thm} A hypergraph is \emph{$d$-degenerate} if every subhypergraph has a vertex of degree at most $d$. A minimum-degree-greedy algorithm colours every $d$-degenerate hypergraph with at most $d+1$ colours. This bound is tight for graphs ($r=2$) since the complete graph on $d+1$ vertices is $d$-degenerate, and of course, has chromatic number $d+1$. However, this observation does not generalise for $r\geq3$. In particular, for the complete $r$-uniform hypergraph on $n$ vertices, every vertex has degree $\binom{n-1}{r-1}$, yet the chromatic number is $\ceil{\frac{n}{r-1}}$. Thus for $r\geq 3$, the degeneracy is much greater than the chromatic number. Given Theorems~\ref{EL75} and \ref{Spencer}, it seems plausible that for $r\geq 3$, every $r$-uniform $d$-degenerate hypergraph is $o(d)$-colourable. It even seems possible that every $r$-uniform $d$-degenerate hypergraph is $O(d^{1/(r-1)})$-colourable. This natural strengthening of Theorems~\ref{EL75} and \ref{Spencer} would (roughly) say that $G$ can be partitioned into independent sets, whose average size is that guaranteed by Theorem~\ref{Spencer}. This note rules out these possibilities, by showing that the naive upper bound $\chi\leq d+1$ is tight for all $r$. This is the main conclusion of this paper. Moreover, we prove it for triangle-free hypergraphs, where a \emph{triangle} in an $r$-uniform hypergraph consists of three edges whose union is a set of $r+1$ vertices. Observe that this definition with $r=2$ is equivalent to the standard notion of a triangle in a graph (although there are other notions of a triangle in a hypergraph \citep{CM13}). \begin{thm} \label{Main} For all $r\geq2$ and $d\geq 1$ there is a triangle-free $d$-degenerate $r$-uniform hypergraph with chromatic number $d+1$. \end{thm} Theorem~\ref{Main} and its proof is a generalisation of a result of \citet{AKS99} who proved it for graphs ($r=2$). Of course, the complete graph $K_{d+1}$ is $d$-degenerate with chromatic number $d+1$. The triangle-free property was the main conclusion of their result. See \citep{KR-RSA10,Alon85} for other related results. \section{Proof} Theorem~\ref{Main} is a corollary of the following: \begin{lem} \label{MainMain} Fix $r\geq 2$. For all $d\geq 1$ there is a triangle-free $d$-degenerate $r$-uniform hypergraph $G_d$ with chromatic number $d+1$, such that in every $(d+1)$-colouring of $G_d$ each colour is assigned to at least $r-1$ vertices. \end{lem} \begin{proof} We proceed by induction on $d$. First consider the base case $d=1$. Let $n:=r(r-1)$. Let $V(G_1):=\{v_1,\dots,v_n\}$ and $E(G_1):=\{e_i:1\leq i\leq n-r+1\}$, where $e_i:=\{v_i,v_{i+1},\dots,v_{i+r-1}\}$. If $S\subseteq V(G_1)$ and $i$ is minimum such that $v_i\in S$, then $v_i$ has degree at most 1 in the subhypergraph induced by $S$. Thus $G_1$ is $1$-degenerate. If $e_i,e_j,e_k$ are three edges in $G_1$ with $i<j<k$, then $e_i\cup e_j\cup e_k$ includes the $r+2$ distinct vertices $v_i,v_{i+1},\dots,v_{i+r-1},v_{j+r-1},v_{k+r-1}.$ Hence $G_1$ is triangle-free. Consider a 2-colouring of $G_1$. Clearly, $G_1$ contains $r-1$ pairwise disjoint edges, each of which contains vertices of both colours. Hence each colour is assigned to at least $r-1$ vertices. This completes the base case. Now assume that $G_{d-1}$ is a triangle-free $(d-1)$-degenerate $r$-uniform hypergraph with chromatic number $d$, such that in every $d$-colouring of $G_{d-1}$ each colour is assigned to at least $r-1$ vertices. Initialise $G_d$ to consist of $d+r-2$ disjoint copies $H_1,\dots,H_{d+r-2}$ of $G_{d-1}$. Let $S$ be a set of $(r-1)d$ vertices in $H_1\cup\dots\cup H_{d+r-2}$ such that $|S\cap V(H_i)|\in\{0,r-1\}$ for $1\leq i\leq d+r-2$. That is, $S$ contains exactly $r-1$ vertices from exactly $d$ of the $H_i$, and contains no vertices from the other $r-2$. Now, for each such set $S$, add $r-1$ \emph{new} vertices $v_1,\dots,v_{r-1}$ to $G_d$ and add the \emph{new} edge $(S\cap V(H_i))\cup\{v_j\}$ to $G_d$ whenever $|S\cap V(H_i)|=r-1$. Thus each new vertex has degree $d$. Since $H_1\cup \dots\cup H_{d+r-2}$ is $d$-degenerate, $G_d$ is also $d$-degenerate. Suppose on the contrary that $G_d$ contains a triangle $T$. Since $G_{d-1}$ is triangle-free, at least one edge in $T$ is a new edge, which is contained in $V(H_i)\cup\{v\}$ for some $i\in[1,d+r-2]$ and some new vertex $v$. Each vertex in a triangle is in at least two of the edges of the triangle. However, by construction, $v$ is contained in only one edge contained in $V(H_i)\cup\{v\}$. Thus $G_d$ is triangle-free. Since $H_1\cup\dots\cup H_{d+r-2}$ is $d$-colourable, and no edge contains only new vertices, assigning all the new vertices a $(d+1)$-th colour produces a $(d+1)$-colouring of $G_d$. Thus $\chi(G_d)\leq d+1$. Suppose on the contrary that $G_d$ has a $(d+1)$-colouring with at most $r-2$ vertices of some colour, say `blue'. Say the other colours are $1,\dots,d$. At most $r-2$ copies of the $H_i$ contain blue vertices. Hence, without loss of generality, $H_1,\dots,H_d$ contain no blue vertices. That is, $H_1,\dots,H_d$ are $d$-coloured with colours $1,\dots,d$. By induction, $H_i$ contains a set $S_i$ of $r-1$ vertices coloured $i$ for $1\leq i\leq d$. By construction, there are $r-1$ vertices $v_1,\dots,v_{r-1}$ in $G_d$, such that $S_i\cup\{v_j\}$ is an edge of $G_d$ for $1\leq i\leq d$ and $1\leq j\leq r-1$. Since each such edge is not monochromatic, each vertex $v_j$ is coloured blue. In particular, there are at least $r-1$ blue vertices, which is a contradiction. Therefore, in every $(d+1)$-colouring of $G_d$, each colour class has at least $r-1$ vertices, as claimed. (In particular, $G_d$ has no $d$-colouring.) \end{proof} \section{An Open Problem} We conclude with an open problem. The \emph{girth} of a graph (that contains some cycle) is the length of its shortest cycle. \citet{Erdos59} proved that there exists a graph with chromatic number at least $k$ and girth at least $g$, for all $k\geq 3$ and $g\geq 4$. (\citet{EH66} proved an analogous result for hypergraphs). Theorem~\ref{Main} strengthens this result for triangle-free graphs (that is, with girth $g=4$). This leads to the following question: Does there exist a $d$-degenerate graph with chromatic number $d+1$ and girth $g$, for all $d\geq 2$ and $g\geq4$? Odd cycles prove the $d=2$ case. An affirmative answer would strengthen the above result of \citet{Erdos59}. A negative answer would also be interesting---this would provide a non-trivial upper bound on the chromatic number of $d$-degenerate graphs with girth $g$. \subsection*{Note} After this paper was written the author discovered the beautiful paper by \citet{KN99} which proves a strengthening of Theorem~\ref{Main} and includes the positive solution of the above open problem. \subsection*{Acknowledgement} Thanks to an anonymous referee for pointing out an error in an earlier version of this paper. \def\soft#1{\leavevmode\setbox0=\hbox{h}\dimen7=\ht0\advance \dimen7 by-1ex\relax\if t#1\relax\rlap{\raise.6\dimen7 \hbox{\kern.3ex\char'47}}#1\relax\else\if T#1\relax \rlap{\raise.5\dimen7\hbox{\kern1.3ex\char'47}}#1\relax \else\if d#1\relax\rlap{\raise.5\dimen7\hbox{\kern.9ex \char'47}}#1\relax\else\if D#1\relax\rlap{\raise.5\dimen7 \hbox{\kern1.4ex\char'47}}#1\relax\else\if l#1\relax \rlap{\raise.5\dimen7\hbox{\kern.4ex\char'47}}#1\relax \else\if L#1\relax\rlap{\raise.5\dimen7\hbox{\kern.7ex \char'47}}#1\relax\else\message{accent \string\soft \space #1 not defined!}#1\relax\fi\fi\fi\fi\fi\fi}

https://arxiv.org/abs/math/0609845

On a Balanced Property of Compositions

Let $S$ be a finite set of positive integers with largest element $m$. Let us randomly select a composition $a$ of the integer $n$ with parts in $S$, and let $m(a)$ be the multiplicity of $m$ as a part of $a$. Let $0\leq r<q$ be integers, with $q\geq 2$, and let $p_{n,r}$ be the probability that $m(a)$ is congruent to $r$ modulo $q$. We show that if $S$ satisfies a certain simple condition, then $\lim_{n\to \infty} p_{n,r} =1/q$. In fact, we show that an obvious necessary condition on $S$ turns out to be sufficient.

\section{Introduction} A {\em composition} of the positive integer $n$ is a sequence $(a_1,a_2,\cdots ,a_k)$ of positive integers so that $\sum_{i=1}^k a_i=n$. The $a_i$ are called the {\em parts} of the composition. It is well-known \cite{bonaw} that the number of compositions of $n$ into $k$ parts is ${n-1\choose k-1}$. From this fact, it is possible to prove the following. Let $0\leq r<q$ be integers, with $q\geq 2$, and let $P_{n,r}$ be the probability of the event that the number of parts of a randomly selected composition of $n$ is congruent to $r$ modulo $q$. Then $\lim_{n\rightarrow \infty} P_{n,r} =1/q$. In other words, \[\lim_{n\rightarrow \infty} \frac{\sum_{i=0}^{\lfloor (n-1)/q \rfloor} {n-1\choose iq+r}}{2^{n-1}}= \frac{1}{q}.\] For $q=2$, this follows from the well-known fact that the number of even-sized subsets of a non-empty finite set is equal to the number of odd-sized subsets of that set. For $q>2$, the statement can be proved, for example, by the method we will use in this paper. The special cases of $q=3$ and $q=4$ appear as Exercises 4.41 and 4.42 in \cite{bonaw}. In other words, all residue classes are equally likely to occur. We will refer to this phenomenon by saying that the number of part sizes of a randomly selected composition of $n$ is {\em balanced}. Now let us impose a restriction on the {\em part sizes} of the compositions of $n$ that form our sample space by requiring that all part sizes come from a finite set $S$. Is it still true that the number of part sizes of a randomly selected composition of $n$ is balanced? That will certainly depend on the restriction we impose on the part sizes. For instance, if the set $S$ of allowed parts consists of odd numbers only, then the number of part sizes will not be balanced. Indeed, if $q=2$, then $P_{n,0}=1$ if $n$ is even, and $P_{n,0}=0$ if $n$ is odd. Let $m$ be the largest element of the set $S$ of allowed parts. It turns out that it is easier to (directly) work with the multiplicity $m(a)$ of $m$ as a part of the randomly selected composition $a$ than with its number of parts. For the special case when $S$ has only two elements, the results obtained for $m(a)$ can then be translated back to results on the number of parts of $a$. In this paper, we prove that if $S$ satisfies a certain obviously necessary condition, then the parameter $m(a)$ is balanced, as described in the abstract. That is, the remainder of $m(a)$ modulo $q$ is equally likely to take all possible values. \section{The Strategy} Let $S=\{s_1,s_2,\cdots ,s_k=m\}$ be a finite set of positive integers with at least two elements. Let us assume without loss of generality that no integer larger than 1 divides all $k$ elements of $S$. Clearly, if $s_1,s_2, \cdots ,s_{k-1}$ are all divisible by a certain prime $h>1$, then the multiplicity $m(a)$ of $m=s_k$ as a part of a composition $a$ of $n$ is restricted. Indeed, $n-m(a)m$ must be divisible by $h$. In particular, if $n$ is divisible by $h$, then $m(a)m$ is divisible by $h$, and so $m(a)$ must be divisible by $h$. Therefore, the parameter $m(a)$ is not balanced. Indeed, if $n$ is divisible by $h$, and we choose $q=h$, then $p_{n,0}=1$, and $p_{n,r}=0$ for $r\neq 0$, while if $n$ is not divisible by $h$ and $q=h$, then $p_{n,0}=0$. So for $m(a)$ to be a balanced parameter, it is {\em necessary} for $S$ to satisfy the condition that its smallest $k-1$ elements do not have a proper common divisor. In the rest of this paper, we prove that this condition is at the same time {\em sufficient} for $m(a)$ to be balanced. {\em Unless otherwise stated}, let $S$ be a finite set of positive integers with at least two elements, and let $S=\{s_1,s_2,\cdots ,s_k=m\}$, where the $s_i$ are listed in increasing order. So $m$ is the largest element of $S$. {\em Unless otherwise stated}, let us also assume that no integer larger than 1 is a divisor of all of $s_1,s_2, \cdots ,s_{k-1}$. Note that this means that if $|S|=2$, then $s_1=1$. For a fixed positive integer $n$, let $A_{S,n}(x)$ be the ordinary generating function of all compositions of the integer $n$ into parts in $S$ according to their number of parts equal to $m$. In other words, \[A _{S,n}(x)=\sum_a x^{m(a)}= \sum_{d}a_{S,n,d}x^d,\] where $a$ ranges over all compositions of $n$ into parts in $S$, and $m(a)$ is the multiplicity of $m$ as a part in $a$. On the far right, $a_{S,n,d}$ is the number of compositions $a$ of $n$ into parts in $S$ so that $m(a)=d$. \begin{example} Let $S=\{1,3\}$. Then the first few polynomials $A_n(x)= A_{S,n}(x)$ are as follows. \begin{itemize} \item $A_0(x)=A_1(x)=A_2(x)=1$, \item $A_3(x)=x$, $A_4(x)=2x+1$, $A_5(x)=3x+1$. \item $A_6(x)=x^2+4x+1$, $A_7(x)=3x^2+5x+1$, $A_8(x)=6x^2+6x+1$. \end{itemize} \end{example} Let $q\geq 2$ be a positive integer, and let $0\leq r\leq q-1$. Let $A_{S,n,r}$ be the number of compositions $a$ of $n$ with parts in $S$ so that $m(a)$ is congruent to $r$ modulo $q$. So \[A_{S,n,r}=a_{S,n,r}+a_{S,n,q+r}+\cdots +a_{S,n,\lfloor n/q \rfloor q+r} .\] In order to simplify the presentations of our results, we will first discuss the special case when $n$ is divisible by $q$ and $r=0$. Let $w$ be a primitive $q$th root of unity. Then \begin{eqnarray} \label{unitroots} \sum_{t=0}^{q-1} A_{S,n}(w^t) & = & \sum_{t=0}^{q-1} \sum_{d= 0}^{n/m} a_{S,n,d} w^{td} \\ & = & \sum_{d=0}^{n/m} \sum_{t=0}^{q-1} a_{S,n,d} w^{td} \\ & = & \sum_{d=0}^{n/m} a_{S,n,d} \sum_{t=0}^{q-1} w^{td}. \end{eqnarray} Using the summation formula of a geometric progression, we get that \[ \sum_{t=0}^{q-1} (w^{d})^t= \left\{ \begin{array}{l@{\ }l} 0 \hbox{ if $w^d\neq 1$, that is, if $q\nmid d$}, \\ \\ q \hbox{ if $w^d =1$, that is, if $q|d$. } \end{array}\right. \] Therefore, (\ref{unitroots}) reduces to \[ \sum_{t=0}^{q-1} A_{S,n}(w^t) = q\cdot \sum_{j=1}^{n/q} a_{S,n,jq} = qA_{S,n,0},\] \begin{equation} \label{forma0} \frac{1}{q} \sum_{t=0}^{q-1} A_{S,n}(w^t)=A_{S,n,0}. \end{equation} So in order to find the approximate value of $A_{S,n,0}$, it suffices to find the approximate values of $A_n(w^t)$, for $0\leq t\leq r-1$, and for a primitive root of unity $w$. The number $A_{S,n}$ of {\em all} compositions of $n$ into parts in $S$ is equal to $A_n(1)$, so we will need that value as well, in order to compute the ratio $A_{S,n,0}/A_{S,n}$. Finally, note that if $n$ is not divisible by $q$, but $r=0$, then the same argument applies, and $\frac{1}{q} \sum_{t=0}^{q-1} A_{S,n}(w^t) =A_{S,n,0}$ still holds. If $r\neq 0$, then instead of computing $\sum_{t=0}^{q-1} A_{S,n}(w^t)$, we compute \[T_n(w)=\sum_{t=0}^{q-1} A_{S,n}(w^t)w^{-rt} =\sum_{d= 0}^{n/m} a_{S,n,d} \sum_{t=0}^{q-1} w^{t(d-r)}.\] This shows that the coefficient of $w^k$ in $T_n(w)$ is 0, unless $w^{d-r}=1$, that is, unless $d-r$ is divisible by $q$. If $d-r$ is divisible by $q$, then this coefficient is $q$. This shows again that \begin{equation} \label{generalc} \frac{1}{q} \sum_{t=0}^{q-1} A_{S,n}(w^t)w^{-tr}=A_{S,n,r}. \end{equation} Therefore, computing $A_{S,n}(w^t)$ will be useful in the general case as well. \section{Linear Recurrence Relations} In order to compute the values of $A_{S,n}(x)$ for various values of $x$, we can keep $x$ fixed, and let $n$ grow. The connection among the polynomials $A_{S,n}(x)$ is explained by the following Proposition. \begin{proposition} Let $S=\{s_1,s_2,\cdots ,s_k=m\}$, with $k\geq 2$. Then for all positive integers $n\geq 3$, the polynomials $A_{S,n}(x)$ satisfy the recurrence relation \begin{equation} \label{recur} A_{S,n}(x)=A_{S,n-s_1}(x)+A_{S,n-s_2}(x)+\cdots +A_{S,n-s_{k-1}}(x)+ xA_{S,n-m}(x). \end{equation} \end{proposition} \begin{proof} Let $a$ be a composition of $n$ with parts in $S$. If the first part of $a$ is $s_i$, for some $i\in [1,k-1]$, then the rest of $a$ forms a composition of $n-s_i$ with parts in $S$ in which the multiplicity of $m$ as a part is still $m(a)$. These compositions of $n$ are counted by $A_{S,n-s_i}(x)$. If the first part of $a$ is $m$, then the rest of $a$ forms a composition of $n-m$ with parts in $S$ in which the multiplicity of $m$ as a part is $m(a)-1$. These compositions of $n$ are counted by $xA_{S,n-m}(x)$. \end{proof} \begin{example} If $S=\{1,3\}$, then (\ref{recur}) reduces to \begin{equation} \label{srecur} A_{S,n}(x)= A_{S,n-1}(x)+x A_{S,n-3}(x). \end{equation} \end{example} For a {\em fixed} real number $x$, the recurrence relation (\ref{recur}) becomes a recurrence relation on real numbers. The solutions of such recurrence relations are described by the following well-known theorem. (See, for instance, \cite{rosen}, Section 7.2. ) \begin{theorem} \label{recthe} Let \begin{equation} \label{grec} a_n=c_1a_{n-1}+c_2a_{n-2}+\cdots +c_ka_{n-k} \end{equation} be a recurrence relation, where the $c_i$ are complex constants. Let $\alpha_1,\alpha_2,\cdots, \alpha_t$ be the distinct roots of the characteristic equation \begin{equation} \label{chareq} z^k-c_1z^{k-1}-c_2z^{k-2}-\cdots -c_k=0,\end{equation} and let $M_i$ be the multiplicity of $\alpha_i$. Then the sequence $a_0,a_1,\cdots $ of complex numbers satisfies (\ref{grec}) if and only if there exist constants $b_1,b_2, \cdots ,b_k$ so that for all $n\geq 0$, we have \begin{equation} \label{fgrec} a_n=b_1\alpha_1^n+b_2 n\alpha_1^n+ \cdots +b_{M_1}n^{M_1-1}\alpha_1^n+ b_{M_1+1}\alpha_2^n, \cdots \end{equation} \begin{equation} \cdots +b_{M_1+M_2}n^{M_2-1}\alpha_2^n, \cdots, \cdots +b_kn^{M_k-1}\alpha_k^n.\end{equation} In other words, the solutions of (\ref{grec}) form a $k$-dimensional vector space. \end{theorem} We will need the following consequence of Theorem \ref{recthe}. \begin{corollary} \label{nonzero} Let us assume that the sequence $\{a_n\}$ is a solution of (\ref{grec}) and that there is no linear recurrence relation of a degree less than $k$ that is satisfied by $ \{a_n\}$. Let us further assume that the characteristic equation (\ref{chareq}) of (\ref{grec}) has a unique root $\alpha_1$ of largest modulus. Then there is a {\em nonzero} constant $C$ so that \[a_n=C\alpha_1^n + o(\alpha_1^n).\] \end{corollary} \begin{proof} As $\{a_n\}$ does not satisfy a recurrence relation of a degree less than $k$, we must have $c_1\neq 0$. As $|\alpha_1|>|\alpha_i|$ for $i\neq 1$, the statement follows. \end{proof} Let us now apply Theorem \ref{recthe} to find the solution of (\ref{recur}) for a fixed $x$. The characteristic equation of (\ref{recur}) is \begin{equation} \label{polyeq} f(z)=z^{m}-\sum_{i=1}^{k-1} z^{m-s_{i}} - x =0.\end{equation} As explained in Section 2, we will need to compute $A_{S,n}(1)$ and also, $A_{S,n}(w^t)$ for the case when $w\neq 1$ is a $q$th primitive root of unity. To that end, we need to find the roots of the corresponding characteristic equations. That is, we will compare the root of largest modulus of the characteristic equation \begin{equation} \label{forone} f_1(z)= z^{m}-\sum_{i=1}^{k-1} z^{m-s_{i}} - 1=0 \end{equation} and the root(s) of the largest modulus of the characteristic equation \begin{equation} \label{forw} f_w(z)= z^{m}-\sum_{i=1}^{k-1} z^{m-s_{i}} - w=0 \end{equation} The following lemma, helping to compute root of largest modulus of $f_1(z)$, is a special case of Exercise III.16 in \cite{polya}. \begin{lemma} \label{first} The polynomial $f_1(z)=z^{m}-\sum_{i=1}^{k-1} z^{m-s_{i}} - 1$ has a unique positive real root $\alpha$. \end{lemma} \begin{proof} Let $\alpha$ be the smallest positive real root of $f_1(z)$. We know such a root exists since $f(0)=-1$ and $\lim_{z\rightarrow \infty}f(z)=\infty$. We claim that then $f$ is strictly monotone increasing on $[\alpha, \infty )$, implying that $f$ cannot have another positive real root. Indeed, if $r>1$, then \begin{eqnarray*} f(r\alpha)+1 & = & (r\alpha)^{m} - \sum_{i=1}^{k-1} (r\alpha)^{m-s_{i}}\\ & > & r^{m} \left(\alpha^{m}-\sum_{i=1}^{k-1} \alpha^{m-s_{i}} \right ) \\ & = & r^m \\ & > & 1, \end{eqnarray*} and so $f(r\alpha)>0$. \end{proof} Now we address the problem of finding the roots of the characteristic equation (\ref{polyeq}) in the case when $w\neq 1$ is a root of unity. It turns out that it suffices to assume that $|w|=1$. (The following Lemma is similar to Exercise III.17 in \cite{polya}.) \begin{lemma} Let $\alpha$ be defined as in Lemma \ref{first}. Let $w$ be any complex number satisfying $w\neq 1$ and $|w|=1$. Then all roots of the polynomial $f_w(z)$ are of smaller modulus than $\alpha$. \end{lemma} \begin{proof} Let $y$ be a root of $f$. Then \begin{eqnarray*} |y|^m & = & \left |w + \sum_{i=1}^{k-1} y^{m-s_i} \right |\\ & \leq & 1+ \sum_{i=1}^{k-1} \left | y^{m-s_i} \right |. \end{eqnarray*} Therefore, $f(|y|)\leq 0$. This implies that $|y|\leq \alpha$ since we have seen in the proof of Lemma \ref{first} that $f(t)>0$ if $t>\alpha$. Furthermore, in the last displayed inequality, the inequality is strict unless for all $i$ so that $1\leq i\leq k-1$, the complex numbers $y^{m-s_i}$ have the same argument as $w$, and that argument is the same as the argument of $y^m$. That happens only if the complex numbers $y^{s_1},y^{s_2-s_1},\cdots y_{s_{k-1}-s_{k-2}}$ all have argument 0, that is, when these numbers are positive real numbers. However, that happens precisely when $s_1, s_2-s_1, \cdots, s_{k-1}-s_{k-2}$ are all multiples of the multiplicative order $o_y$ of $y/|y|$ as a complex number. That implies that $s_1,s_2,\cdots ,s_{k-1}$ are all divisible by $o_y$, contradicting our hypothesis on $S$. \end{proof} The previous two lemmas show that the largest root of the characteristic equation for the sequence $\{A_{S,n}(1)\}_{n\geq 0}$ is larger than the largest root(s) of the characteristic equation for the sequence $\{A_{S,n}(w)\}_{n\geq 0}$ for any complex number $w\neq 1$ with absolute value 1. Given formula (\ref{fgrec}), in order to see that the first sequence indeed grows faster than the second, all we need to show is that the {\em coefficient} $b_1$ of $\alpha^n$ in (\ref{grec}) is not 0. (Here $\alpha$, the largest root of $f_1(z)$, plays the role of $\alpha_1$ in (\ref{fgrec})). This is the content of the next lemma. \begin{lemma} \label{noshorter} Let $S=\{s_1,s_2,\cdots ,s_{k-1}\}$ be any finite set of positive integers (so for this Lemma, we do not require that $s_1,s_2,\cdots s_{k-1}$ do not have a proper common divisor). Then the sequence $\{A_{S,n}\}_{n\geq 0}=\{A_{S,n}(1)\}_{n\geq 0}$ does not satisfies a linear recurrence relation with constant coefficients and less than $|S|+1$ terms. In other words, if $|S|=k$, then there do not exist constants $c_2,c_3,\cdots ,c_k$ and positive integers $j_1,j_2,\cdots , j_{k-1}$ so that for all $n\geq 0$, \[A_{S,n}=\sum_{i=1}^{k-1} c_i A_{S,n-j_i}.\] \end{lemma} \begin{proof} Let us assume that $S$ is a minimal counterexample. It is then straightforward to verify that $|S|>2$. Let $S'=S-m$, that is, the set obtained from $S$ by removing the largest element of $S$. Then \begin{equation} \label{reduction} A_{S',n}=\sum_{i=1}^{k-1} c_i' A_{S',n-s_i},\end{equation} and there is no shorter recurrence satisfied by $\{A_{S',n}\}$. Now crucially, $A_{S',n}=A_{S,n}$ for all $n$ satisfying $0\leq n< m$. So these sequences agree in $m-1\geq k-1$ values. So if $\{A_{S,n}\}$ satisfied a linear recurrence relation of degree $k-1$, that would have to be the recurrence relation (\ref{reduction}). Indeed, by Theorem \ref{recthe}, the solutions of (\ref{reduction}) form a $k$-dimensional vector space, so knowing $k$ elements of a solution determines the whole solution. However, $\{A_{S,n}\}$ does not satisfy (\ref{reduction}) since $A_{S,m}=A_{S',m}+1\neq A_{S',m}$, where the difference is caused by the one-part composition $m$. \end{proof} Now we are in position to express the growth rate of $A_{S,n}=A_{S,n}(1)$. \begin{proposition} Let $\alpha$ be defined as in Lemma \ref{first}. Then \[A_{S,n}(1)=C \alpha^n + o(\alpha^n),\] for some nonzero constant $C$. \end{proposition} \begin{proof} Immediate from Corollary \ref{nonzero} and Lemma \ref{noshorter}. \end{proof} We can now compare the growth rates of $A_{S,n}(w)$ and $A_{S,n}(1)$. \begin{lemma} \label{grrate} Let $w\neq 1$ be any complex number so that $|w|=1$. Then \[\lim_{n\rightarrow \infty} \frac{A_{S,n}(w)}{A_{S,n}(1)} =0. \] \end{lemma} \begin{proof} Lemma (\ref{first}) shows that the unique positive root of the characteristic equation (\ref{forone}) is larger than the absolute value of all roots of the characteristic equation (\ref{forw}). Therefore, $A_{S,n}(w)=O(n^k\beta^k)$, with $\beta < \alpha$. \end{proof} Finally, we can use the results of this section to prove the balanced properties of the numbers $A_{S,n,r}$. \begin{theorem} Let $q\geq 2$ be an integer, and let $r$ be an integer satisfying $0\leq r \leq q-1$. Then \[\lim_{n\rightarrow \infty} p_{n,r}=\lim_{n\rightarrow \infty} \frac{A_{S,n,r}}{A_{S,n}}=\frac{1}{q}.\] \end{theorem} \begin{proof} Let us first address the case of $r=0$. Dividing (\ref{forma0}) by $A_{S,n}$, we get that \[A_{S,n,0}=\frac{1}{q} \sum_{t=0}^{q-1}\frac{A_{S,n}(w^t)}{A_{S,n}}.\] However, Lemma \ref{grrate} shows that all but one of the $q$ summands on the right-hand side converge to 0, and the remaining one (the first summand) is equal to 1. For general $r$, the only change is that instead of dividing both sides of (\ref{forma0}) by $A_{S,n}$, we divide both sides of (\ref{generalc}) by $A_{S,n}$. As $|w|=1$, the result follows in the same way. \end{proof} \section{Further Directions} Let $S=\{1,3\}$. Numerical evidence suggest that for all $n$, the polynomials $A_{S,n}(x)$ have real roots only. Furthermore, numerical evidence also suggests that the sequences of polynomials $\{A_{S,3n+r}(x)\}_{n\geq 0}$ form a Sturm sequence for each of $r=0,1,2$. (See \cite{wilfb} for the definition and importance of Sturm sequences.) This raises the following intriguing questions. \begin{question} For which sets $S$ is it true that the polynomials $A_{S,n}(x)$ have real roots only? \end{question} \begin{question} For which sets $S$ is it true that the polynomials $A_{S,n}(x)$ can be partitioned into a few Sturm sequences? \end{question} Herb Wilf \cite{wilf} proved that the set $S=\{1,2\}$ does have both of these properties. If $A_{S,n}(x)$ has real roots only, then its coefficients form a log-concave (and therefore, unimodal) sequence. (See Chapter 8 of \cite{bonaint} for an introduction into into unimodal and log-concave sequences.) This raises the following questions. \begin{question} Let us assume that $A_{S,n}(x)$ has real roots only. Is there a combinatorial proof for the log-concavity of its coefficients? \end{question} \begin{question} Let us assume that $A_{S,n}(x)$ has real roots only. Where is the peak (or peaks) of the unimodal sequence of its coefficients? \end{question} Another interesting question is the following. \begin{question} For what $S$ and $n$ does the equality $A_{S,n}(-1)=0$ hold? When it does, the number of compositions of $n$ with parts in $S$ and with $m(a)$ even equals the number of compositions of $n$ with parts in $S$ and with $m(a)$ odd. Is there a combinatorial proof of that fact? \end{question} Finally, our methods rested on the finiteness of $S$, but we can still ask what can be said for {\em infinite} sets of allowed parts. \begin{center} {\bf Acknowledgment} I am grateful to Herb Wilf for valuable discussions and advice. \end{center}

https://arxiv.org/abs/1805.05025

Cutoff for product replacement on finite groups

We analyze a Markov chain, known as the product replacement chain, on the set of generating $n$-tuples of a fixed finite group $G$. We show that as $n \rightarrow \infty$, the total-variation mixing time of the chain has a cutoff at time $\frac{3}{2} n \log n$ with window of order $n$. This generalizes a result of Ben-Hamou and Peres (who established the result for $G = \mathbb{Z}/2$) and confirms a conjecture of Diaconis and Saloff-Coste that for an arbitrary but fixed finite group, the mixing time of the product replacement chain is $O(n \log n)$.

\section{Introduction} Let $G$ be a finite group, and let $[n] := \{1, 2, \dots, n\}$. We consider the set $G^n$ of all functions $\sigma: [n] \to G$ (or ``configurations''). We may define a Markov chain $(\sigma_t)_{t\ge0}$ on $G^{n}$ as follows: if we have a current state $\sigma$, then uniformly at random, choose an ordered pair $(i, j)$ of distinct integers in $[n]$, and change the value of $\sigma(i)$ to $\sigma(i) \sigma(j)^{\pm 1}$, where the signs are chosen with equal probability. We will restrict the chain $(\sigma_t)_{t\ge 0}$ to the space of {\it generating $n$-tuples}, i.e.\ the set of $\sigma$ whose values generate $G$ as a group: \[ {\mathcal S} := \left\{ \sigma \in G^n \ : \ \langle \sigma(1), \dots, \sigma(n) \rangle=G \right\}. \] It is not hard to see that for fixed $G$ and large enough $n$, the chain on ${\mathcal S}$ is irreducible (see \cite[Lemma 3.2]{DSC96}). We will always assume $n$ is large enough so that this irreducibility holds. Note that the chain is also symmetric, and it is aperiodic because it has holding on some states. Thus, the chain has a uniform stationary distribution $\pi$ with $\pi(\sigma)=1/|{\mathcal S}|$. This Markov chain was first considered in the context of computational group theory---it models the \emph{product replacement algorithm} for generating random elements of a finite group introduced in \cite{Celler}. By running the chain for a long enough time $t$ and choosing a uniformly random index $k \in [n]$, the element $\sigma_t(k)$ is a (nearly) uniformly random element of $G$. The product replacement algorithm has been found to perform well in practice \cite{Celler, Holt-Rees}, but the question arises: how large does $t$ need to be in order to ensure near uniformity? One way of answering the question is to estimate the mixing time of the Markov chain. It was shown by Diaconis and Saloff-Coste that for any fixed finite group $G$, there exists a constant $C_G$ such that the $\ell^2$-mixing time is at most $C_G n^2 \log n$ \cite{DSC96, DSC98} (see also Chung and Graham \cite{Chung-Graham-Group} for a simpler proof of this fact with a different value for $C_G$). In another line of work, Lubotzky and Pak \cite{Lubotzky-Pak} analyzed the mixing of the product replacement chain in terms of Kazhdan constants (see also subsequent quantitative estimates for Kazhdan constants by Kassabov \cite{Kassabov}). We also mention a result of Pak \cite{Pak} which shows mixing in $\text{polylog}(|G|)$ steps when $n = \Theta(\log |G| \log \log |G|)$. The reader may consult the survey \cite{Pak-Survey} for further background on the product replacement algorithm. Diaconis and Saloff-Coste conjectured that the mixing time bound can be improved to $C_G n \log n$ \cite[Remark 2, Section 7, p.\ 290]{DSC98}, based on the observation that at least $n \log n$ steps are needed by the classical coupon-collector's problem. This was confirmed in the case $G = {\mathbb Z}/2$ by Chung and Graham \cite{Chung-Graham-Cube} and recently refined by Ben-Hamou and Peres, who show that when $G={\mathbb Z}/2$, the chain in fact exhibits a cutoff at time $\frac{3}{2}n \log n$ in total-variation with window of order $n$ \cite{BHP}. In this paper, we extend the result of Ben-Hamou and Peres to all finite groups. Note that this also verifies the conjecture of Diaconis and Saloff-Coste for a fixed finite group. To state the result, let us denote the total variation distance between $\Pr_\sigma(\sigma_t \in \cdot \ )$ and $\pi$ by \[ d_\sigma(t) := \max_{A \subseteq {\mathcal S}}|\Pr_\sigma(\sigma_t \in A)- \pi (A)|. \] \begin{theorem}\label{Thm:main} Let $G$ be a finite group. Then, the Markov chain $(\sigma_t)_{t \ge 0}$ on the set of generating $n$-tuples of $G$ has a total-variation cutoff at time $\frac{3}{2}n\log n$ with window of order $n$. More precisely, we have \begin{equation}\label{Eq:UB} \lim_{\b\to\infty} \limsup_{n\to\infty} \max_{\sigma \in {\mathcal S}} d_\sigma\left(\frac{3}{2}n\log n + \b n\right) = 0 \end{equation} and \begin{equation}\label{Eq:LB} \lim_{\b\to\infty} \liminf_{n\to\infty} \max_{\sigma \in {\mathcal S}} d_\sigma\left(\frac{3}{2}n\log n - \b n\right) = 1. \end{equation} \end{theorem} \subsection{A connection to cryptography} We mention another motivation for studying the product replacement chain in the case $G=({\mathbb Z}/q)^m$ for a prime $q \ge 2$ and integers $m \ge 1$. It comes from a public-key authentication protocol proposed by Sotiraki \cite{Sotiraki}, which we now briefly describe. In the protocol, a verifier wants to check the identity of a prover based on the time needed to answer a challenge. First, the prover runs the Markov chain with $G = ({\mathbb Z}/q)^m$ and $n = m$, which can be interpreted as performing a random walk on $SL_n({\mathbb Z}/q)$, where $\sigma(k)$ is viewed as the $k$-th row of a $n \times n$ matrix. (In each step, a random row is either added to or subtracted from another random row.) After $t$ steps, the prover records the resulting matrix $A \in SL_n({\mathbb Z}/q)$ and makes it public. To authenticate, the verifier gives the prover a vector $x \in ({\mathbb Z}/q)^n$ and challenges her to compute $y := Ax$. The prover can perform this calculation in $O(t)$ operations by retracing the trajectory of the random walk. Without knowing the trajectory, if $t$ is large enough, an adversary will not be able to distinguish $A$ from a random matrix and will be forced to perform the usual matrix-vector multiplication (using $n^2$ operations) to complete the challenge. Thus, the question is whether $t \ll n^2$ is large enough for the matrix $A$ to become sufficiently random, so that the prover can answer the challenge much faster than an adversary. Note that when $n > m$, the product replacement chain on $G = ({\mathbb Z}/q)^m$ amounts to the projection of the random walk on $SL_n({\mathbb Z}/q)$ onto the first $m$ columns. Thus, Theorem \ref{Thm:main} shows that when $m$ is fixed and $n \rightarrow \infty$, the mixing time for the first $m$ columns is around $\frac{3}{2} n \log n$. One then hopes that the mixing of several columns is enough to make it computationally intractable to distinguish $A$ from a random matrix; this would justify the authentication protocol, as $n \log n \ll n^2$. We remark that when $t$ is much larger than the mixing time of the random walk on $SL_n({\mathbb Z}/q)$ generated by row and additions and subtractions, it is information theoretically impossible for an adversary to distinguish $A$ from a random matrix. However, the diameter of the corresponding Cayley graph on $SL_n({\mathbb Z}/q)$ is known to be of order $\Theta\left( \frac{n^2}{\log_q n} \right)$ \cite{AHM, Christofides}, so a lower bound of the same order necessarily holds for the mixing time. Diaconis and Saloff-Coste \cite[Section 4, p.\ 420]{DSC96} give an upper bound of $O(n^4)$, which was subsequently improved to $O(n^3)$ by Kassabov \cite{Kassabov}. Closing the gap between $n^3$ and $\frac{n^2}{\log n}$ remains an open problem. \subsection{Outline of proof} The proof of Theorem \ref{Thm:main} analyzes the mixing behavior in several stages: \begin{itemize} \item an initial ``burn-in'' period lasting around $n \log n$ steps, after which the group elements appearing in the configuration are not mostly confined to any proper subgroup of $G$; \item an averaging period lasting around $\frac{1}{2} n \log n$ steps, after which the counts of group elements become close to their average value under the stationary distribution; and \item a coupling period lasting $O(n)$ steps, after which our chain becomes exactly coupled to the stationary distribution with high probability. \end{itemize} The argument is in the spirit of \cite{BHP}, but a more elaborate analysis is required in the second and third stages. To analyze the first stage, for a fixed proper subgroup $H$, the number of group elements in $H$ appearing in the configuration is a birth-and-death process whose transition probabilities are easy to estimate. The analysis of the resulting chain is the same as in \cite{BHP}, and we can then union bound over all proper subgroups $H$. In the second stage, for a given starting configuration $\sigma_0 \in {\mathcal S}$, we consider quantities $n_{a,b}(\sigma)$ counting the number of sites $k$ where $\sigma_0(k) = a$ and $\sigma(k) = b$. A key observation (which also appears in \cite{BHP}) is that by symmetry, projecting the Markov chain onto the values $(n_{a,b}(\sigma_t))_{a, b \in G}$ does not affect the mixing behavior. Thus, it is enough to understand the mixing behavior of the counts $n_{a,b}$. One expects these counts to evolve towards their expected value ${\mathbb E}_{\sigma \sim \pi} n_{a,b}(\sigma)$ as the chain mixes. To carry out the analysis rigorously, we write down a stochastic difference equation for the $n_{a,b}$ and analyze it via the Fourier transform. Intuitively, as $n \rightarrow \infty$, the process approaches a ``hydrodynamic limit'' so that it becomes approximately deterministic. It turns out that after about $\frac{1}{2} n \log n$ steps, the $n_{a,b}$ are likely to be within $O(\sqrt{n})$ of their expected value. Our analysis requires a sufficiently ``generic'' initial configuration, which is why the first stage is necessary. Finally, in the last stage, we show that if the $(n_{a,b}(\sigma))_{a,b\in G}$ and $(n_{a,b}(\sigma'))_{a,b\in G}$ for two configurations are within $O(\sqrt{n})$ in $\ell^1$ distance, they can be coupled to be exactly the same with high probability after $O(n)$ steps of the Markov chain. A standard argument involving coupling to the stationary distribution then implies a bound on the mixing time. The main idea to prove the coupling bound is that even if the $\ell^1$ distance evolves like an unbiased random walk, there is a good chance that it will hit $0$ due to random fluctuations. A similar argument is used to prove cutoff for lazy random walk on the hypercube \cite[Chapter 18]{LevinPeresWilmer}. However, some careful accounting is necessary in our setting to ensure that in fact the $\ell^1$ distance does not increase in expectation and to ensure sufficient fluctuations. \subsection{Organization of the paper} The rest of the paper is organized as follows. In Section \ref{Sec:UB}, we state (without proof) the key lemmas describing the behavior in each of the three stages and use these to prove the upper bound \eqref{Eq:UB} in Theorem \ref{Thm:main}. Sections \ref{Sec:DE-Proofs} and \ref{Sec:Coupling-Proof} contain the proofs of these lemmas. Finally, in Section \ref{Sec:LB}, we prove the lower bound \eqref{Eq:LB} in Theorem \ref{Thm:main}; this is mostly a matter of verifying that the estimates used in the upper bound were tight. \subsection{Notation} Throughout this paper, we use $c, C, C', \dots$, to denote absolute constants whose exact values may change from line to line, and also use them with subscripts, for instance, $C_G$ to specify its dependency only on $G$. We also use subscripts with big-$O$ notation, e.g.\ we write $O_G(\,\cdot\,)$ when the implied constant depends only on $G$. \section{Proof of Theorem \ref{Thm:main} (\ref{Eq:UB})}\label{Sec:UB} Let us fix a finite group $G$ and denote its cardinality by ${\mathcal Q} := |G|$. For a configuration $\sigma \in {\mathcal S}$, let $n_a(\sigma)$ denote the number of sites having group element $a$, i.e., \[ n_a(\sigma) := |\{i \in [n] \ : \ \sigma(i)=a\}|. \] \subsection{The burn-in period} For a proper subgroup $H \subseteq G$, let \[ n_{non}^H(\sigma) := \sum_{a \in G \setminus H} n_a(\sigma) \] denote the number of sites not in $H$, and define for $c \in (0, 1)$ the set \[ \Snon{c} := \{\sigma \in {\mathcal S} \ : \ n_{non}^H(\sigma) \ge cn \text{ for all proper subgroups $H \subseteq G$} \}. \] Thus, $\Snon{c}$ is the set of states $\sigma$ where the group elements appearing in $\sigma$ are not mostly confined to any particular proper subgroup of $G$. The next lemma shows that we reach $\Snon{1/3}$ in about $n \log n$ steps, and once we reach $\Snon{1/3}$, we remain in $\Snon{1/6}$ for $n^2$ steps with high probability. Note that $n^2$ is much larger than the overall mixing time, so we may essentially assume that we are in $\Snon{1/6}$ for all of the later stages. \begin{lemma}\label{Lem:Burn} Let $\t_{1/3} := \min\{t \ge 0 : \sigma_t \in \Snon{1/3}\}$ be the first time to hit $\Snon{1/3}$. Then for all large enough $n$ and for all large enough $\b > 0$, \[ \max_{\sigma \in {\mathcal S}} \Pr_\sigma(\t_{1/3} > n \log n + \b n) \le \frac{120 {\mathcal Q}}{\b^2}. \] Moreover, there exists a constant $C_G$ depending only on $G$ such that \[ \max_{\sigma \in \Snon{1/3}}\Pr_\sigma \left(\sigma_t \notin \Snon{1/6} \ \text{\rm for some $t \le n^2$}\right) \le C_G n^2e^{-n/10}. \] \end{lemma} \begin{proof} Fix a proper subgroup $H \subset G$, and consider what happens to $n_{non}^H(\sigma_t)$ at time $t$. Suppose our next step is to replace $\sigma(i)$ with $\sigma(i)\sigma(j)$. If $\sigma(j) \in H$, then $n_{non}^H(\sigma_{t+1}) = n_{non}^H(\sigma_t)$. If $\sigma(j) \not\in H$ and $\sigma(i) \in H$, then $n_{non}^H(\sigma_{t+1}) = n_{non}^H(\sigma_t) - 1$. Finally, if $\sigma(j), \sigma(i) \not\in H$, then $\sigma(i)\sigma(j)$ may or may not be in $H$, so $n_{non}^H(\sigma_{t+1}) \ge n_{non}^H(\sigma_t) - 1$. Let $(N_t)_{t \ge 0}$ be the birth-and-death chain with the following transition probabilities for $1 \le k \le n$: \begin{align*} \Pr(N_{t+1} = k+1 \mid N_t = k) &= \frac{k(n-k)}{n(n-1)} \\ \Pr(N_{t+1} = k-1 \mid N_t = k) &= \frac{k(k-1)}{n(n-1)} \\ \Pr(N_{t+1} = k \mid N_t = k) &= \frac{n-k}{n}. \end{align*} We start this chain at $N_0 = n^H_{non}(\sigma_0)$; note that because the elements appearing in $\sigma_0$ generate $G$, we are guaranteed to have $n^H_{non}(\sigma_0) > 0$. The above birth-and-death chain corresponds to the behavior of $(n^H_{non}(\sigma_t))$ if whenever $\sigma(j), \sigma(i) \not\in H$, it always happened that $\sigma(i)\sigma(j) \in H$. Thus, $(n^H_{non}(\sigma_t))$ stochastically dominates $(N_t)$. The chain $(N_t)$ is precisely what is analyzed in \cite{BHP} for the case $G = {\mathbb Z}/2$. Let \[T_k := \min\{t \ge 0 : N_t=k\}.\] Then, we have ${\mathbb E}_{k-1}T_k \le \frac{n^2}{k(n-2k)}$ \cite[(2) in the proof of Lemma 1]{BHP} and thus ${\mathbb E}_1 (T_{n/3}) =\sum_{k=2}^{n/3}{\mathbb E}_{k-1}T_k \le n \log n + n$. On the other hand, setting $v_k={\rm Var}_{k-1}(T_k)$, we have $v_2 \le n^2$, \[v_{k+1}\le \frac{k}{n-k}v_k + \frac{54 n^2}{k^2},\] and ${\rm Var}_1 (T_{n/3}) = \sum_{k=2}^{n/3}v_k \le 110 n^2$ \cite[The proof of Lemma 1]{BHP}. Hence by Chebyshev's inequality for all large enough $\b > 0$, \[\Pr_1(T_{n/3} > n \log n + \b n) \le \frac{120}{\b^2}.\] Moreover, we have $\Pr_{n/3} \left( T_{n/6} \le n^2 \right) \le n^2e^{-n/10}$. Indeed, this follows from the fact that for $m<k$, we have \[\Pr_k(T_m \le n^2) \le n^2\frac{\pi_{\rm BD}(m)}{\pi_{\rm BD}(k)},\] where $\pi_{\rm BD}(k)={n \choose k}/(2^n-1)$ \cite[(5) and the following in the proof of Proposition 2]{BHP}. We now take a union bound over all the proper subgroups $H$. \end{proof} \subsection{The averaging period} In the next stage, the counts $n_a(\sigma_t)$ go toward their average value. We actually analyze this stage in two substages, looking at a ``proportion vector'' and ``proportion matrix'', as described below. \subsubsection{Proportion vector chain} For a configuration $\sigma \in {\mathcal S}$, we consider the ${\mathcal Q}$-dimensional vector $(n_a(\sigma)/n)_{a \in G}$, which we call the {\it proportion vector} of $\sigma$. One may check that for a typical $\sigma \in {\mathcal S}$, each $n_a(\sigma)/n$ is about $1/{\mathcal Q}$. For each $\d > 0$, we define the $\d$-{\it typical set} \[ {\mathcal S}_\ast(\d) := \left\{\sigma \in {\mathcal S} \ : \ \left\|\left(\frac{n_a(\sigma)}{n}\right)_{a \in G} - \left(\frac{1}{{\mathcal Q}}\right)_{a \in G}\right\| \le \d \right\}, \] where $\| \cdot \|$ denotes the $\ell^2$-norm in ${\mathbb R}^G$. The following lemma implies that starting from $\sigma \in \Snon{1/3}$, we reach $\Sast{\d}$ in $O_\d(n)$ steps with high probability. The proof is given in Section \ref{Subsec:vector}. \begin{lemma}\label{Lem:1stDE} Consider any $\sigma \in \Snon{1/3}$ and any constant $\d>0$. There exists a constant $C_{G, \d}$ depending only on $G$ and $\d$ such that for any $T \ge C_{G, \d} n$, we have \[ \Pr_\sigma\left(\sigma_T \notin \Sast{\d} \right) \le \frac{1}{n} \] for all large enough $n$. \end{lemma} \subsubsection{Proportion matrix chain} We actually need a more precise averaging than what is provided by Lemma \ref{Lem:1stDE}. Fix a configuration $\sigma_0 \in {\mathcal S}$. For any $\sigma \in {\mathcal S}$ and for any $a, b \in G$, define \[ n_{a,b}^{\sigma_0}(\sigma) := |\{i \in [n] \ : \ \sigma_0(i)=a, \sigma(i)=b \}|. \] If we run the Markov chain $(\sigma_t)_{t\ge 0}$ with initial state $\sigma_0$, then $n_{a,b}^{\sigma_0}(\sigma_t)$ is the number of sites that originally contained the element $a$ (at time 0) but now contain $b$ (at time $t$). Note that \[ \sum_{b \in G} n_{a,b}^{\sigma_0}(\sigma) = n_a(\sigma_0) \quad \text{and} \quad \sum_{a \in G} n_{a,b}^{\sigma_0}(\sigma) = n_b(\sigma).\] We can then associate with $(\sigma_t)_{t \ge 0}$ another Markov chain $\left(n_{a,b}^{\sigma_0}(\sigma_t)/n_a(\sigma_0)\right)_{a, b \in G}$ for $t \ge 0$, which we call the {\it proportion matrix chain} ({\it with respect to $\sigma_0$}). The state space for the proportion matrix chain is $\{0, 1, \dots, n\}^{G \times G}$, and the transition probabilities depend on $\sigma_0$. The proportion matrix acts like a ``sufficient statistic'' for analyzing our Markov chain started at $\sigma_\ast$, because of the permutation invariance of our dynamics. In fact, as the following lemma shows, the distance to stationarity of the proportion matrix chain is equal to the distance to stationarity of the original chain. \begin{lemma}\label{Lem:nij} Let $\sigma_\ast \in {\mathcal S}$ be a configuration. For the Markov chain $(\sigma_t)_{t \ge 0}$ with initial state $\sigma_\ast$, we consider $\left(n_{a, b}^{\sigma_\ast}(\sigma_t)\right)_{a, b \in G}$. Let $\overline{\pi}^{\sigma_\ast}$ be the stationary measure for the Markov chain $\{(n_{a, b}^{\sigma_\ast}(\sigma_t))_{a, b \in G}\}_{t \ge 0}$ on $\left\{0, 1, \dots, n\right\}^{G \times G}$. Then, for every $t \ge 0$, we have \[ \normTV{ \Pr_{\sigma_\ast}(\sigma_t \in \cdot \ ) - \pi } = \normTV{ \Pr_{\sigma_\ast}\left( (n_{a, b}^{\sigma_\ast}(\sigma_t))_{a, b \in G} \in \cdot \ \right) - \overline{\pi}^{\sigma_\ast} }. \] \end{lemma} \begin{proof} For any matrix $N = (N_{a,b})_{a,b \in G} \in \{0, 1, \ldots , n\}^{G \times G}$, write \[ {\mathcal X}_{(N)} := \left\{\sigma \in {\mathcal S} \ : \ n_{a, b}^{\sigma_\ast}(\sigma)=N_{a, b} \ \text{for all $a, b \in G$}\right\} \] for the set of configurations with $N$ as their proportion matrix. Since the distribution of $\sigma_t$ is invariant under permutations on sites $i \in [n]$ preserving the set $\{ i : \sigma_\ast(i) = a\}$ for every $a \in G$, the conditional probability measures $\Pr_{\sigma_\ast}\left(\sigma_t \in \cdot \mid \sigma_t \in {\mathcal X}_{(N)} \right)$ and $\pi( \ \cdot \mid {\mathcal X}_{(N)})$ are both uniform on ${\mathcal X}_{(N)}$. This implies that for each $\sigma \in {\mathcal X}_{(N)}$, \[ |\Pr_{\sigma_\ast}(\sigma_t =\sigma) - \pi(\sigma)| = \frac{1}{\left|{\mathcal X}_{(N)}\right|}\left|\Pr_{\sigma_\ast}\left((n^{\sigma_\ast}_{a, b}(\sigma_t))_{a, b \in G}=N\right)- \overline{\pi}^{\sigma_\ast}(N) \right|, \] and summing over all $\sigma \in {\mathcal X}_{(N)}$ and all $N$, we obtain the claim. \end{proof} For $\sigma_0 \in {\mathcal S}$ and $r > 0$, define the set of configurations \[ \Sast{\sigma_0, r} := \left\{ \sigma \in {\mathcal S} \ : \ \left\| \left(\frac{n^{\sigma_0}_{a, b}(\sigma)}{n_a(\sigma_0)}\right)_{b \in G} - \left(\frac{1}{{\mathcal Q}}\right)_{b \in G}\right\| \le r \text{ for all $a \in G$} \right\}. \] Roughly speaking, the following lemma shows that starting from a typical configuration $\sigma_\ast \in \Sast{\frac{1}{4{\mathcal Q}}}$, we need about $\frac{1}{2}n \log n$ steps to reach $\Sast{\sigma_\ast, \frac{R}{\sqrt{n}}}$, where $R$ is a constant. We show this fact in a slightly more general form where the initial state need not be $\sigma_\ast$; the proof is given in Section \ref{Subsec:matrix}. \begin{lemma}\label{Lem:2ndDE} Consider any $\sigma_\ast, \sigma'_\ast \in \Sast{\frac{1}{4{\mathcal Q}}}$, and let $T := \ceil{\frac{1}{2} n \log n}$. There exists a constant $C_G > 0$ depending only on $G$ such that for any given $R > 0$, we have \[ \Pr_{\sigma'_\ast}\left(\sigma_T \notin \Sast{\sigma_\ast, \frac{R}{\sqrt{n}}}\right) \le C_G e^{-R} + \frac{1}{n} \] for all large enough $n$. \end{lemma} \subsection{The coupling period} After reaching $\Sast{\sigma_\ast, \frac{R}{\sqrt{n}}}$, we show that only $O(n)$ additional steps are needed to mix in total variation distance. The main ingredient in the proof is a coupling of proportion matrix chains so that they coalesce in $O(n)$ steps when they both start from configurations $\sigma, \tilde\sigma \in \Sast{\sigma_\ast, \frac{R}{\sqrt{n}}}$. We construct such a coupling and prove the following lemma in Section \ref{Sec:Coupling-Proof}. \begin{lemma}\label{Lem:RW} Consider any $\sigma_\ast \in \Sast{\frac{1}{5{\mathcal Q}^3}}$, and let $R > 0$. Suppose $\sigma, \tilde \sigma \in \Sast{\sigma_\ast, \frac{R}{\sqrt{n}}}$. Then, there exists a coupling $(\sigma_t, \tilde \sigma_t)$ of the Markov chains with initial states $(\sigma, \tilde \sigma)$ such that for a given $\b > 0$ and all large enough $n$, \[ \Pr_{\sigma, \tilde \sigma}(\t > \b n) \le \frac{32{\mathcal Q}^2 R}{\sqrt{\b}}, \] where $\t:=\min\{t \ge 0 : n^{\sigma_\ast}_{a, b}(\sigma_t) = n^{\sigma_\ast}_{a, b}(\tilde \sigma_t) \ \text{\rm for all $a, b \in G$}\}$. \end{lemma} To translate this coupling time into a bound on total variation distance, we need also the simple observation that the stationary measure $\pi$ concentrates on $\Sast{\sigma_\ast, \frac{R}{\sqrt{n}}}$ except for probability $O(1/R^2)$, as given in the next lemma. \begin{lemma}\label{Lem:stationary} For the stationary distribution $\pi$ of the chain $(\sigma_t)_{t \ge 0}$, for every $R>0$ and for all $n > m$, \[ \pi\left(\sigma \notin \Sast{\frac{R}{\sqrt{n}}}\right) \le \frac{C_G}{R^2}. \] Moreover for every $\d<1/(2{\mathcal Q})$, for every $R>0$ and for all $n>m$, \[ \max_{\sigma_\ast \in \Sast{\d}}\pi\left(\sigma \notin \Sast{\sigma_\ast, \frac{R}{\sqrt{n}}}\right) \le \frac{2 C_G {\mathcal Q}}{R^2}, \] where $C_G$ and $m$ are constants depending only on $G$. \end{lemma} \begin{proof} Observe that since the stationary distribution $\pi$ is uniform on ${\mathcal S}$, it is given by the uniform distribution ${\rm Unif}$ on $G^n$ conditioned on ${\mathcal S}$. Note that we can always generate $G$ using each of its $|G|$ elements, so we have an easy lower bound of $|{\mathcal S}| \ge |G|^{n - |G|}$. Consequently, we have \begin{align*} \pi\left(\sigma \notin \Sast{\frac{R}{\sqrt{n}}}\right) &\le |G|^{|G|} {\rm Unif}\left(\sigma \notin \Sast{\frac{R}{\sqrt{n}}}\right) \\ &\le |G|^{|G|} \sum_{a \in G}{\rm Unif}\left(\left| \frac{n_a(\sigma)}{n}-\frac{1}{{\mathcal Q}}\right| \ge \frac{R}{\sqrt{n}}\right) \le \frac{|G|^{|G|}}{R^2}\left(1-\frac{1}{{\mathcal Q}}\right). \end{align*} Concerning the second assertion, we note that $n_a(\sigma_\ast) \ge (1/{\mathcal Q} -\d)n$ for each $a \in G$; the rest follows similarly, so we omit the details. \end{proof} \begin{remark}\label{Rem:1} In Lemma \ref{Lem:stationary} above, we have given a very loose bound on $C_G$ for sake of simplicity. Actually, it is not hard to see that holding $G$ fixed, we have $\lim_{n\rightarrow \infty} |{\mathcal S}|/|G|^n = 1$. See also \cite[Section 6.B.]{DSC98} for more explicit bounds for various families of groups. \end{remark} Together, Lemmas \ref{Lem:2ndDE}, \ref{Lem:RW}, and \ref{Lem:stationary} imply the following bound for total variation distance. \begin{lemma} \label{Lem:coupling-tv} Let $\b > 0$ be given, and let $T := \ceil{\frac{1}{2} n \log n} + \ceil{\b n}$. Then, for any $\sigma_\ast \in \Sast{\frac{1}{5{\mathcal Q}^3}}$, we have \[ \normTV{ \Pr_{\sigma_\ast}(\sigma_T \in \cdot \ ) - \pi } \le \frac{C_G}{\b^{1/4}}, \] where $C_G$ is a constant depending only on $G$. \end{lemma} \begin{proof} Let $\tilde\sigma$ be drawn from the stationary distribution $\pi$. Define \[ \tau = \min \left\{ t \ge 0 : n^{\sigma_\ast}_{a,b}(\sigma_t) = n^{\sigma_\ast}_{a,b}(\tilde\sigma_t) \text{ for all $a, b \in G$}\right\}, \] where $(\tilde\sigma_t)$ is a Markov chain started at $\tilde\sigma$. Let $\overline{\pi}^{\sigma_\ast}$ denote the stationary distribution for the proportion matrix with respect to $\sigma_\ast$. Since $\tilde\sigma$ was drawn from $\pi$, the proportion matrix of $\tilde\sigma_t$ remains distributed as $\overline{\pi}^{\sigma_\ast}$ for all $t$. We first run $\sigma$ and $\tilde\sigma$ independently up until time $T_1 := \ceil{\frac{1}{2} n \log n}$. For a parameter $R$ to be specified later, consider the events \[ {\mathcal G} := \left\{ \sigma_{T_1} \in \Sast{\sigma_\ast, \frac{R}{\sqrt{n}}} \right\}, \qquad \tilde{{\mathcal G}} := \left\{ \tilde\sigma_{T_1} \in \Sast{\sigma_\ast, \frac{R}{\sqrt{n}}} \right\}. \] Lemma \ref{Lem:2ndDE} implies that $\Pr({\mathcal G}^{\sf c}) \le C_G e^{-R} + \frac{1}{n}$, and Lemma \ref{Lem:stationary} implies that $\Pr(\tilde{\mathcal G}^{\sf c}) \le \frac{2 C_G {\mathcal Q}}{R^2}$. Let $T_2 := \ceil{\b n}$. Starting from time $T_1$, as long as both ${\mathcal G}$ and $\tilde{\mathcal G}$ hold, we may use Lemma \ref{Lem:RW} to form a coupling $(\sigma_t, \tilde\sigma_t)$ so that \[ \Pr_{\sigma_\ast, \sigma_\ast} \Big( n^{\sigma_\ast}_{a, b}(\sigma_{T_1+T_2}) \ne n^{\sigma_\ast}_{a, b}(\tilde\sigma_{T_1+T_2}) \text{ for some $a, b \in G$} \,\Big|\, {\mathcal G} \cap \tilde{\mathcal G} \Big) \le \frac{C{\mathcal Q}^2 R}{\sqrt{\b}}. \] Setting $R = \b^{1/4}$, we conclude that \begin{align*} &\Pr_{\sigma_\ast, \sigma_\ast} \Big( n^{\sigma_\ast}_{a, b}(\sigma_{T_1+T_2}) \ne n^{\sigma_\ast}_{a, b}(\tilde\sigma_{T_1+T_2}) \text{ for some $a, b \in G$} \Big) \le \frac{C{\mathcal Q}^2 R}{\sqrt{\b}} + \Pr({\mathcal G}^{\sf c}) + \Pr(\tilde{\mathcal G}^{\sf c}) \\ &\qquad\le \frac{C{\mathcal Q}^2 R}{\sqrt{\b}} + \left(C_G e^{-R} + \frac{1}{n}\right) + \frac{2C_G {\mathcal Q}}{R^2} = O_G\left(\frac{1}{\b^{1/4}}\right). \end{align*} We have $T = T_1 + T_2$, and recall that the proportion matrix for $\tilde\sigma$ is stationary for all time. This yields \[ \normTV{ \Pr_{\sigma_\ast}\left( (n^{\sigma_\ast}_{a, b}(\sigma_T))_{a, b \in G} \in \cdot \ \right) - \overline{\pi}^{\sigma_\ast} } = O_G\left(\frac{1}{\b^{1/4}}\right). \] The result then follows by Lemma \ref{Lem:nij}. \end{proof} \subsection{Proof of the main theorem} We now combine the lemmas from the burn-in, averaging, and coupling periods to complete the proof of the upper bound in Theorem \ref{Thm:main}. \begin{proof}[Proof of Theorem \ref{Thm:main} (\ref{Eq:UB})] Define $T_1 := \ceil{n \log n + \b n}$, $T_2 := \ceil{\b n}$, and $T_3 := \ceil{\frac{1}{2} n \log n} + \ceil{\b n}$. Let $\tau_{1/3}$ be the first time to hit $\Snon{1/3}$ as in Lemma \ref{Lem:Burn}. Then, Lemma \ref{Lem:Burn} implies that for any $\sigma_1 \in {\mathcal S}$ and any $t \ge 0$, we have \begin{align} d_{\sigma_1}(T_1 + t) &\le \Pr_{\sigma_1}\left( \tau_{1/3} > T_1 \right) + \max_{\sigma \in \Snon{1/3}} d_\sigma(t) \nonumber \\ &\le \frac{120 {\mathcal Q}}{\b^2} + \max_{\sigma \in \Snon{1/3}} d_\sigma(t). \label{Eq:stage1} \end{align} Next, by Lemma \ref{Lem:1stDE}, for any $\sigma_2 \in \Snon{1/3}$ and when $\b$ and $n$ are sufficiently large, we have that $\Pr_{\sigma_2} \left( \sigma_{T_2} \not\in \Sast{\frac{1}{5{\mathcal Q}^3}} \right) \le \frac{1}{n}$. Consequently, for $\sigma_2 \in \Snon{1/3}$, we have \begin{equation} \label{Eq:stage2} d_{\sigma_2}(T_2 + t) \le \frac{1}{n} + \max_{\sigma_\ast \in \Sast{\frac{1}{5{\mathcal Q}^3}}} d_{\sigma_\ast}(t). \end{equation} Finally, Lemma \ref{Lem:coupling-tv} states that \begin{equation} \label{Eq:stage3} \max_{\sigma_\ast \in \Sast{\frac{1}{5{\mathcal Q}^3}}} d_{\sigma_\ast}(T_3) \le \frac{C_G}{\b^{1/4}}. \end{equation} Thus, combining \eqref{Eq:stage1}, \eqref{Eq:stage2}, and \eqref{Eq:stage3}, we obtain for any $\sigma \in {\mathcal S}$ that \begin{align*} d_\sigma\left(\frac{3}{2} n \log n + 4 \b n \right) &\le d_\sigma\left(T_1 + T_2 + T_3 \right) \nonumber \\ &\le \frac{120 {\mathcal Q}}{\b^2} + \frac{1}{n} + \frac{C_G}{\b^{1/4}} \end{align*} sending $n \rightarrow \infty$ and then $\b \rightarrow \infty$ yields \eqref{Eq:UB}. \end{proof} \section{Proofs for the averaging period} \label{Sec:DE-Proofs} In this section, we prove Lemmas \ref{Lem:1stDE} and \ref{Lem:2ndDE}. The proofs are based on analyzing stochastic difference equations satisfied by the Fourier transform of the proportion vector or matrix. \subsection{The Fourier transform for $G$} We first establish some notation and preliminaries for the Fourier transform. Let $G^\ast$ be a complete set of non-trivial irreducible representations of $G$. In other words, for each $\r \in G^\ast$, we have a finite dimensional complex vector space $V_\r$ such that $\r: G \to GL(V_\r)$ is a non-trivial irreducible representation, and any non-trivial irreducible representation of $G$ is isomorphic to some unique $\r \in G^\ast$. Moreover, we may equip each $V_\r$ with an inner product for which $\r \in G^\ast$ is unitary. For a configuration $\sigma \in {\mathcal S}$ and for each $\r \in G^\ast$, we consider the matrix acting on $V_\r$ given by \[ x_\r(\sigma) := \sum_{a \in G} \frac{n_a(\sigma)}{n}\r(a), \] so that $x_\r(\sigma)$ is the Fourier transform of the proportion vector at the representation $\r$. We write $x(\sigma):=(x_\r(\sigma))_{\r \in G^\ast}$. Let $\wt{V} := \bigoplus_{\r \in G^\ast}{\rm End}_{\mathbb C}(V_\r)$, and write $d_\r := \dim_{\mathbb C} V_\r$. For an element $x = (x_\r)_{\r \in G^\ast} \in \wt{V}$, we define a norm $\| \cdot \|_{\wt{V}}$ given by \begin{equation*}\label{Eq:norm} \|x\|_{\wt{V}}^2 := \frac{1}{{\mathcal Q}}\sum_{\r \in G^\ast}d_\r \|x_\r\|_{{\rm HS}}^2, \end{equation*} where $\langle A, B \rangle_{{\rm HS}} =\Tr(A^\ast B)$ denotes the Hilbert-Schmidt inner product in ${\rm End}_{\mathbb C}(V_\r)$ and $\|\cdot\|_{{\rm HS}}$ denotes the corresponding norm. (Note that $\langle \cdot, \cdot \rangle_{{\rm HS}}$ and $\|\cdot\|_{{\rm HS}}$ depend on $\r$, but for sake of brevity, we omit the $\r$ when there is no danger of confusion.) The Peter-Weyl theorem \cite[Chapter 2]{DiaconisRep} says that \begin{equation*} \label{Eq:Peter-Weyl} L^2(G) \cong {\mathbb C} \oplus \wt{V}, \end{equation*} where the isomorphism is given by the Fourier transform. The Plancherel formula then reads \begin{equation}\label{Eq:Plancherel} \|x(\sigma)\|_{\wt{V}}^2 = \left\|\left(\frac{n_a(\sigma)}{n}\right)_{a \in G} - \left(\frac{1}{{\mathcal Q}}\right)_{a \in G}\right\|^2. \end{equation} Thus, in order to show that $\sigma \in \Sast{\d}$, it suffices to show that $\|x(\sigma)\|_{\wt{V}}$ is small. A similar argument may be applied to the proportion matrix instead of the proportion vector. Finally, for an element $A \in {\rm End}_{\mathbb C}(V_\r)$, we will at times also consider the \emph{operator norm} $\|A\|_{op} := \sup_{v \in V_\r, v \neq 0} \|Av\| / \|v\|$. We will also sometimes use the following (equivalent) variational characterization of the operator norm: \begin{align*} \sup_{\substack{X \in {\rm End}_{\mathbb C}(V_\r) \\ \|X\|_{{\rm HS}} = 1}} \|XA\|^2_{{\rm HS}} &= \sup_{\substack{X \in {\rm End}_{\mathbb C}(V_\r) \\ \|X\|_{{\rm HS}} = 1}} \Tr (XAA^*X^*) = \sup_{\substack{X \in {\rm End}_{\mathbb C}(V_\r) \\ \|X\|_{{\rm HS}} = 1}} \Tr (X^*XAA^*) \\ &= \sup_{\substack{Y \in {\rm End}_{\mathbb C}(V_\r) \\ Y = Y^*, \;\; \Tr Y = 1}} \langle Y , AA^* \rangle_{{\rm HS}} = \|AA^*\|_{op} = \|A\|_{op}^2. \end{align*} \subsubsection{The special case of $G = {\mathbb Z}/q$} On a first reading of this section, the reader may wish to consider everything for the special case of $G = {\mathbb Z}/q$ for some integer $q \ge 2$. In that case, each representation is one-dimensional, and the representations can be indexed by $\ell = 0, 1, 2, \ldots , q - 1$. The Fourier transform is then particularly simple: the coefficients are scalar values \[ x_\ell(\sigma) = \sum_{a = 0}^{q - 1} \frac{n_a(\sigma)}{n} \o^{a \ell}, \] where $\o := e^{\frac{2\pi i}{q}}$ is a primitive $q$-th root of unity. This special case already illustrates most of the main ideas while simplifying the estimates in some places (e.g.\ matrix inequalities we use will often be immediately obvious for scalars). \subsection{A stochastic difference equation for the $n_a$} For $a \in G$, we next analyze the behavior of $n_a(\sigma_t)$ over time. For convenience, we write $n_a(t) = n_a(\sigma_t)$. Let ${\mathcal F}_t$ denote the $\sigma$-field generated by the Markov chain $(\sigma_t)_{t \ge 0}$ up to time $t$. Then, our dynamics satisfy the equation \begin{equation}\label{Eq:DE0} {\mathbb E}[n_a(t+1)-n_a(t) \mid {\mathcal F}_t] = \sum_{b \in G} \frac{n_{ab^{-1}}(t) n_b(t) }{2n(n-1)}+\sum_{b \in G} \frac{n_{ab}(t) n_b(t)}{2n(n-1)} - \frac{n_a(t)}{n}. \end{equation} Note that $|n_a(t + 1) - n_a(t)| \le 1$ almost surely. Thus, for each $a \in G$, we can write the above as a stochastic difference equation \begin{equation}\label{Eq:DE} n_a(t+1) - n_a(t) = \sum_{b \in G} \frac{n_{ab^{-1}}(t) n_b(t) }{2n(n-1)}+\sum_{b \in G} \frac{n_{ab}(t) n_b(t)}{2n(n-1)} - \frac{n_a(t)}{n} + M_a(t+1), \end{equation} where ${\mathbb E}[M_a(t+1) \mid {\mathcal F}_t] = 0$ and $|M_a(t)| \le 2$ almost surely. It is easiest to analyze this equation through the Fourier transform. Writing $x_\r(t) = x_\r(\sigma_t)$, we calculate from \eqref{Eq:DE} that \[ x_\r(t+1) - x_\r(t) = \frac{1}{n-1}x_\r(t) \left(\frac{x_\r(t) + x_\r(t)^\ast}{2} -\frac{n-1}{n}\right) + \wh{M}_\r(t+1), \] where $\wh{M}_\r(t) := \frac{1}{n}\sum_{a \in G}M_a(t) \r(a)$. For convenience, write \[X_\r(t) = \frac{1}{n - 1}\left(\frac{x_\r(t) + x_\r(t)^\ast}{2} - \frac{n-1}{n}\right),\] so that our equation becomes \begin{equation} \label{Eq:DEfourier} x_\r(t+1) - x_\r(t) = x_\r(t) X_\r(t) + \wh{M}_\r(t+1). \end{equation} Note that we have \[ \|x_\r(t)\|_{{\rm HS}} \le \sqrt{d_\r}, \qquad {\mathbb E}[\wh{M}_\r(t+1) \mid {\mathcal F}_t] = 0, \qquad\text{and}\qquad \|\wh{M}_\r(t)\|_{{\rm HS}} \le \frac{2 {\mathcal Q}\sqrt{d_\r}}{n}, \] and thus, \[ \|x(t)\|_{\wt{V}} \le 1 \qquad\text{and}\qquad \|\wh{M}(t)\|_{\wt{V}}\le \frac{2{\mathcal Q}}{n},\] where $\wh{M}=(\wh{M}_\r)_{\r \in G^\ast}$. \subsection{A general estimate for stochastic difference equations} Before proving Lemma \ref{Lem:1stDE}, we also need a technical lemma for controlling the behavior of stochastic difference equations, which will be used to analyze \eqref{Eq:DEfourier} as well as other similar equations. \begin{lemma} \label{Lem:GenDE} Let $(z(t))_{t \ge 0}$ be a sequence of $[0,1]$-valued random variables adapted to a filtration $({\mathcal F}_t)_{t \ge 0}$. Let $\varepsilon \in (0, 1)$ be a small constant, and let $\varphi : {\mathbb R}^+ \to (0,1]$ be a non-decreasing function. Suppose that there are ${\mathcal F}_t$-measurable random variables $M(t)$ for which \begin{equation} \label{Eq:zDE} z(t+1) - z(t) \le -\varepsilon\varphi(t+1) z(t) + M(t+1) \end{equation} and which, for some constant $D$, satisfy the bounds \[ {\mathbb E}[ M(t+1) \mid {\mathcal F}_t ] \le D\varepsilon\sqrt{\varepsilon}, \qquad |M(t)| \le D\varepsilon. \] Then, for each $t$ and each $\lambda > 0$, we have \[ \Pr\left( z(t) \ge \lambda \sqrt{\varepsilon} + e^{- \varepsilon \int_0^t \varphi(s)\,ds} \cdot z(0) \right) \le C_{D,\varphi} e^{-c_{D,\varphi} \lambda^2} \] for constants $c_{D,\varphi}, C_{D,\varphi}$ depending only on $D$ and $\varphi$. \end{lemma} \begin{proof} Let us define for integers $t \ge 1$, \[ \Phi(t) := \varepsilon^{-1} \sum_{k = 1}^t \log \frac{1}{1-\varepsilon \varphi(k)}, \qquad \text{and} \qquad \Phi(0) := 0.\] Taking conditional expectations in the inequality relating $z(t+1)$ to $z(t)$, we have \[ {\mathbb E}[ z(t+1) \mid {\mathcal F}_t ] \le (1 - \varepsilon \varphi(t+1)) z(t) + D \varepsilon\sqrt{\varepsilon}. \] Rearranging and using the fact that $\varphi(t)$ is non-decreasing, we have \begin{align*} {\mathbb E}[ z(t+1) \mid {\mathcal F}_t ] - \frac{D\sqrt{\varepsilon}}{\varphi(0)} &\le (1 - \varepsilon \varphi(t+1)) z(t) - \frac{D\sqrt{\varepsilon}(1 - \varepsilon \varphi(t+1))}{\varphi(0)} \\ &\le (1 - \varepsilon \varphi(t+1)) \left( z(t) - \frac{D\sqrt{\varepsilon}}{\varphi(0)} \right). \end{align*} Consequently, \[ Z_t := e^{\varepsilon \Phi(t)} \left( z(t) - \frac{D\sqrt{\varepsilon}}{\varphi(0)} \right) \] is a supermartingale, and its increments are bounded by \begin{equation} \label{Eq:Z-increment-bound} |Z_{t+1}-Z_t| \le e^{\varepsilon \Phi(t+1)}\left(|M(t+1)|+D \varepsilon\right) \le 2D\varepsilon e^{\varepsilon\Phi(t+1)}. \end{equation} Recall that $\varphi$ is non-decreasing, so that for all $t \ge s \ge 0$, we have \[ \Phi(t) = \Phi(s) + \varepsilon^{-1} \sum_{k = s + 1}^t \log \frac{1}{1 - \varepsilon \varphi(k)} \ge \Phi(s) + (t - s) \varphi(0). \] Using this with \eqref{Eq:Z-increment-bound}, we see that the sum of the squares of the first $t$ increments is at most \begin{align*} \sum_{s = 1}^{t} 4D^2 \varepsilon^2 e^{2\varepsilon\Phi(s)} &\le 4D^2\varepsilon^2 \sum_{s = 1}^t e^{2\varepsilon \Phi(t) - 2\varepsilon\varphi(0)(t - s)} \le 4D^2\varepsilon^2 e^{2\varepsilon\Phi(t)} \cdot \frac{1}{1 - e^{-2\varepsilon\varphi(0)}} \\ &\le 4D^2\varepsilon^2 e^{2\varepsilon\Phi(t)} \cdot \frac{1}{1 - (1 - \frac{1}{2}\varepsilon\varphi(0))} = \frac{8D^2 \varepsilon}{\varphi(0)} \cdot e^{2\varepsilon\Phi(t)}. \end{align*} By the Azuma-Hoeffding inequality, this yields \[ \Pr\left(Z_t \ge \lambda \sqrt{\varepsilon} e^{\varepsilon\Phi(t)} + Z_0 \right) \le \exp\left( - \frac{\varphi(0) \lambda^2 \varepsilon \cdot e^{2\varepsilon\Phi(t)}}{16D^2 \varepsilon \cdot e^{2\varepsilon\Phi(t)}} \right) = \exp\left( -\frac{\varphi(0) \lambda^2}{16D^2} \right), \] which in turn implies \[ \Pr\left( z(t) \ge \frac{D\sqrt{\varepsilon}}{\varphi(0)} + e^{-\varepsilon\Phi(t)} z(0) + \lambda \sqrt{\varepsilon} \right) \le \exp\left( -\frac{\varphi(0)\lambda^2}{16D^2} \right). \] Finally, observe that $\Phi(t) \ge \sum_{k = 1}^t \varphi(k) \ge \int_0^t \varphi(s)\, ds$. The result then follows upon shifting and rescaling of $\lambda$. \end{proof} \subsection{Proportion vector chain: Proof of Lemma \ref{Lem:1stDE}}\label{Subsec:vector} We first prove a bound for the Fourier coefficients $x_\r(t)$. \begin{lemma} \label{Lem:1stDE-fourier} Consider any $\sigma \in \Snon{1/3}$ and any $\r \in G^\ast$. We have a constant $c_G$ depending only on $G$ for which \[ \Pr_\sigma\left( \bigcup_{t = 1}^{n^2} \left\{ \|x_\r(t)\|_{{\rm HS}} \ge \frac{1}{n^{1/8}} + e^{-c_G t/n}\cdot \|x_\r(0)\|_{{\rm HS}} \right\} \right) \le \frac{1}{n^3}. \] for all large enough $n$. \end{lemma} This immediately implies Lemma \ref{Lem:1stDE}. \begin{proof}[Proof of Lemma \ref{Lem:1stDE}] With $c_G$ defined as in Lemma \ref{Lem:1stDE-fourier}, take $C_{G, \d}$ large enough so that for any $T \ge C_{G, \d} n$, \[ \frac{1}{n^{1/8}} + e^{-c_G T/n}\sqrt{d_\r} \le \d. \] Then, Lemma \ref{Lem:1stDE-fourier} and Plancherel's formula yield \begin{align*} \Pr_\sigma\left(\sigma_T \notin \Sast{\d} \right) &\le \Pr_\sigma\left( \|x_\r(T)\|_{{\rm HS}} \ge \d \text{ for some $\r \in G^\ast$} \right) \\ &\le \frac{{\mathcal Q}}{n^3} \le \frac{1}{n}, \end{align*} for large enough $n$, as desired. \end{proof} We are now left with proving Lemma \ref{Lem:1stDE-fourier}, which relies on the following bound on the operator norm. \begin{lemma} \label{Lem:gap} There exists a positive constant $\gamma_G$ depending on $G$ such that for any $\r \in G^\ast$ and any $\sigma \in \Snon{1/6}$, \[ \|I_{d_\r}+X_\r(\sigma)\|_{op} \le 1-\frac{\gamma_G}{n}. \] \end{lemma} \begin{proof} Let $\Delta_G$ denote the set of all probability distributions on $G$, and for $c \in (0, 1)$, let $\Delta_G(c) \subset \Delta_G$ denote the set of all probability distributions $\mu$ such that $\mu(H) \le 1 - c$ for all proper subgroups $H \subset G$. Consider a representation $\r \in G^\ast$, and consider the function $h : \Delta_G(1/6) \to {\rm End}_{\mathbb C}(V_\r)$ given by \[ h(\mu) = \sum_{a \in G} \mu(a) \frac{\r(a)+\r(a)^\ast}{2}. \] Then, $h(\mu)$ is hermitian, and since $\r$ is unitary, we clearly have \[\l(\mu):=\max_{v \in V_\r, \|v\|=1}\langle h(\mu)v, v\rangle \le 1. \] We claim that $\l(\mu) < 1$ for each $\mu \in \Delta_G(c)$. Indeed, suppose the contrary. Then, there exists a non-zero vector $v \in V_\r$ such that $\Re \langle \r(a)v, v \rangle=1$ for all $a \in G$ with $\mu(a)>0$. This implies that the support of $\mu$ is included in the subgroup \[H=\{a \in G \ : \ \r(a)v=v\}.\] Since $\r$ is a (non-trivial) irreducible representation, $H$ is a proper subgroup of $G$, and thus $\mu(H) \le 1- c$, contradicting the assumption that $\mu \in \Delta_G(c)$. Note that $\mu \mapsto \l(\mu)$ is continuous. We may define \[\gamma_\r:=\max_{\mu \in \Delta_G(1/6)}\l(\mu) < 1 \qquad\text{and}\qquad \tilde \gamma_G:=\max_{\r \in G^\ast}\gamma_\r<1.\] Then, we have for any $\sigma \in \Snon{1/6}$, \[ \frac{x_\r(\sigma) + x_\r(\sigma)^\ast}{2} = \sum_{a \in G} \frac{n_a(\sigma)}{n}\frac{\r(a)+\r(a)^\ast}{2} \preceq \tilde \gamma_G I_{d_\r}. \] Taking $0<\gamma_G < 1-\tilde \gamma_G$, and plugging this into the definition of $X_\r$ gives $X_\r(\sigma) \preceq -\frac{\gamma_G}{n}I_{d_\r}$. Note that $X_\r(\sigma) \succeq -\frac{2}{n-1}I_{d_\r}$. Combining these together gives the result. \end{proof} \begin{remark} A much more direct approach is possible in the case $G = {\mathbb Z}/q$. The condition $\sigma \in \Snon{1/6}$ implies that $n_0(\sigma) \le \frac{5}{6}$. Then, we have \[ \Re x_\ell(\sigma) := \Re \sum_{a = 0}^{q - 1} \frac{n_a(\sigma)}{n} \o^{a \ell} \le \frac{5}{6} + \frac{1}{6} \max_{1 \le a \le q - 1} \Re \o^{a \ell} = \frac{5}{6} + \frac{1}{6} \cos \frac{2\pi}{q} < 1 - \gamma_G \] for some positive $\gamma_G$. Some rearranging of equations then yields the desired result. \end{remark} \begin{proof}[Proof of Lemma \ref{Lem:1stDE-fourier}] Fix $\r \in G^\ast$. Let ${\mathcal G}_t$ denote the event where for all $0 \le s \le t$, we have $\|I_{d_\r}+X_\r(s)\|_{op} \le 1 - \frac{\gamma_G}{n}$, where $\gamma_G$ is taken as in Lemma \ref{Lem:gap}. Since our chain starts at $\sigma \in \Snon{1/3}$, Lemmas \ref{Lem:Burn} and \ref{Lem:gap} together imply that \[ \Pr_\sigma({\mathcal G}_{n^2}^{\sf c}) \le C_G n^2 e^{-n/10}. \] Next, we turn to \eqref{Eq:DEfourier}. Rearranging \eqref{Eq:DEfourier} and squaring, we have \begin{align} \|x_\r(t+1)\|_{{\rm HS}}^2 &= \|x_\r(t)(I_{d_\r} + X_\r(t))\|_{{\rm HS}}^2 + \|\wh{M}_\r(t+1)\|_{{\rm HS}}^2 \nonumber \\ &\hphantom{==} + 2\Re \langle x_\r(t)(I_{d_\r} + X_\r(t)), \wh{M}_\r(t+1) \rangle_{{\rm HS}} \label{Eq:1stDEsquared} \end{align} Let $z_t := {\bf 1}_{{\mathcal G}_t} \|x_\r(t)\|_{{\rm HS}}^2$ and \[ M'(t+1) := \|\wh{M}_\r(t+1)\|_{{\rm HS}}^2 + 2\Re \langle x_\r(t)(I_{d_\r} + X_\r(t)), \wh{M}_\r(t+1) \rangle_{{\rm HS}}. \] Substituting into \eqref{Eq:1stDEsquared}, we obtain \[ z_{t+1} \le \|I_{d_\r} + X_\r(t)\|_{op}^2 \cdot z_t + {\bf 1}_{{\mathcal G}_t} M'(t+1) \le \left(1 - \frac{\gamma_G}{n}\right)^2 z_t + {\bf 1}_{{\mathcal G}_t} M'(t+1). \] Note that we have the bounds \[ {\mathbb E}[ M'(t+1) \mid {\mathcal F}_t ]= {\mathbb E}[ \|\wh{M}_\r(t+1)\|_{{\rm HS}}^2 \mid {\mathcal F}_t ] \le \frac{4{\mathcal Q}^2d_\r}{n^2} \] \[ |M'(t+1)| \le \|\wh{M}_\r(t+1)\|_{{\rm HS}}^2 + 2\sqrt{d_\r}\left(1+\frac{1}{n(n-1)}\right) \|\wh{M}_\r(t+1)\|_{{\rm HS}} \le \frac{6{\mathcal Q}^2 d_\r}{n}. \] We now apply Lemma \ref{Lem:GenDE} with $\varepsilon = \frac{1}{n}$, $\varphi(t) = \gamma_G$, $D = 6{\mathcal Q}^2 d_\r$, and $\lambda = n^{1/4}$. This yields \[ \Pr\left( z_t \ge n^{-1/4} + e^{-\gamma_G t/n}\cdot z_0 \right) \le C'_G e^{-c'_G \sqrt{n}}. \] Consequently, \[ \Pr_\sigma\left( \|x_\r(t)\|_{{\rm HS}} \ge n^{-1/8} + e^{-\gamma_G t/2n} \cdot \|x_\r(0)\|_{{\rm HS}} \right) \le C'_G e^{-c'_G \sqrt{n}} + C_G n^2 e^{-n/10}. \] The lemma with $c_G = \gamma_G/2$ then follows from union bounding over all $1 \le t \le n^2$ and taking $n$ sufficiently large. \end{proof} \subsection{Proportion matrix chain: Proof of Lemma \ref{Lem:2ndDE}}\label{Subsec:matrix} We carry out a similar albeit more refined strategy to analyze the proportion matrix. Throughout this section, we assume our Markov chain $(\sigma_t)_{t \ge 0}$ starts at an initial state $\sigma_\ast \in \Sast{\frac{1}{4{\mathcal Q}}}$. We again write $n_a(t)=n_a(\sigma_t)$ and $n_{a, b}(t)=n^{\sigma_\ast}_{a, b}(\sigma_t)$, and similar to before, the $n_{a, b}(t)$ satisfy the difference equation \begin{equation}\label{Eq:de2} n_{a, b}(t+1)-n_{a, b}(t)=\sum_{c \in G}\frac{n_{a, bc^{-1}}(t)n_c(t)}{2n(n-1)}+\sum_{c \in G}\frac{n_{a, bc}(t)n_c(t)}{2n(n-1)} - \frac{n_{a, b}(t)}{n} + M_{a, b}(t+1), \end{equation} where ${\mathbb E}[M_{a, b}(t+1) \mid {\mathcal F}_t]=0$ and $|M_{a, b}(t)| \le 2$ for all $t \ge 0$. We can again analyze this equation via the Fourier transform. In this case, for each $a \in G$, we take the Fourier transform of $\left(n_{a, b}(t)/n_a(\sigma_\ast)\right)_{b \in G}$. For $\r \in G^\ast$, let \[ y_{a,\r}(t) = y_{a,\r}^{\sigma_\ast}(t) := \sum_{b \in G}\frac{n_{a, b}(t)}{n_a(\sigma_\ast)}\r(b) \] denote the Fourier coefficient at $\r$. Let $\wh{M}_{a, \r}(t) := \frac{1}{n_a(\sigma_\ast)}\sum_{b \in G}M_{a, b}(t)\r(b)$. Then, \eqref{Eq:de2} becomes \begin{equation}\label{Eq:DEy} y_{a, \r}(t+1) - y_{a, \r}(t) = y_{a, \r}(t) X_\r(t) + \wh{M}_{a, \r}(t+1). \end{equation} Note that ${\mathbb E}_\sigma[\wh{M}_{a, \r}(t+1) \mid {\mathcal F}_t]=0$. Also, since we assumed $\sigma_\ast \in \Sast{\frac{1}{4{\mathcal Q}}}$, it follows that $\frac{n_a(\sigma_\ast)}{n} \ge \frac{1}{2{\mathcal Q}}$. Thus, we also know $\|\wh{M}_{a, \r}(t+1)\|_{{\rm HS}} \le \frac{4{\mathcal Q}^2\sqrt{d_\r}}{n}$. Again, our main step is a bound on the Fourier coefficients $y_{a, \r}(t)$, which will also be useful later in proving Lemma \ref{Lem:RW}. \begin{lemma} \label{Lem:2ndDE-fourier} Consider any $\sigma_\ast, \sigma'_\ast \in \Sast{\frac{1}{4{\mathcal Q}}}$. There exist constants $c_G, C_G > 0$ depending only on $G$ such that for all large enough $n$, we have \[ \Pr_{\sigma'_\ast}\left( \|y^{\sigma_\ast}_{a, \r}(t)\|_{{\rm HS}} \ge R \left(\frac{1}{\sqrt{n}} + e^{-t/n} \|y^{\sigma_\ast}_{a, \r}(0)\|_{{\rm HS}} \right) \right) \le e^{-\O_G(R^2) + O_G(1)} + \frac{2}{n^3} \] for all $t$ and $R > 0$. \end{lemma} The above lemma directly implies Lemma \ref{Lem:2ndDE}. \begin{proof}[Proof of Lemma \ref{Lem:2ndDE}] We apply Lemma \ref{Lem:2ndDE-fourier} to each $a \in G$ and $\r \in G^\ast$. Recall that $T = \ceil{\frac{1}{2} n \log n}$, so that \[ \frac{1}{\sqrt{n}} + e^{-T/n} \|y^{\sigma_\ast}_{a, \r}(0)\|_{{\rm HS}} \le \frac{2\sqrt{d_\r}}{\sqrt{n}}. \] Then, Lemma \ref{Lem:2ndDE-fourier} implies \[ \Pr_{\sigma'_\ast}\left( \|y^{\sigma_\ast}_{a, \r}(T)\|_{{\rm HS}} \ge \frac{R}{\sqrt{n}} \right) \le e^{-\O_G(R^2) + O_G(1)} + \frac{2}{n^3}. \] Union bounding over all $a \in G$ and $\r \in G^\ast$ and using the Plancherel formula, this yields \begin{align*} &\Pr_{\sigma'_\ast}\left( \sigma_\ast \not\in \Sast{\sigma_\ast, \frac{R}{\sqrt{n}}} \right) \le \Pr_{\sigma'_\ast}\left( \max_{a, \r} \|y^{\sigma_\ast}_{a, \r}(T)\|_{{\rm HS}} \ge \frac{R}{\sqrt{n}} \right) \\ &\qquad\qquad \le e^{-\O_G(R^2) + O_G(1)} + \frac{2 {\mathcal Q}^2}{n^3} \le C_G e^{-R} + \frac{1}{n} \end{align*} for sufficiently large $C_G$ and $n$. \end{proof} We now prove Lemma \ref{Lem:2ndDE-fourier}. Before proceeding with the main proof, we need the following routine estimate as a preliminary lemma. \begin{lemma} \label{Lem:theta} Let $\theta_n : {\mathbb R}^d \to {\mathbb R}^+$ be the function given by $\theta_n(x) = \|x\| + \frac{1}{\sqrt{n}}e^{-\sqrt{n}\|x\|} - \frac{1}{\sqrt{n}}$. Then, we have the inequalities \[ \|\nabla \theta_n(x)\| \le 1, \qquad \theta_n(x + h) \le \theta_n(x) + \langle h, \nabla \theta_n(x) \rangle + \frac{\sqrt{n}}{2} \|h\|^2. \] \end{lemma} \begin{proof} We can write $\theta_n(x) = f(\|x\|)$, where $f(r) = r + \frac{1}{\sqrt{n}} e^{-\sqrt{n} r} - \frac{1}{\sqrt{n}}$. By spherical symmetry, we have \[ \|\nabla \theta_n(x)\| = f'(\|x\|) = 1 - e^{-\sqrt{n}\|x\|} \le 1, \] which is the first inequality. Again by spherical symmetry, the eigenvalues of the Hessian $\nabla^2 \theta_n(x)$ can be directly computed to be $f''(\|x\|)$ and $f'(\|x\|) / \|x\|$. But these are bounded by \[ f''(r) \le \sqrt{n} e^{-\sqrt{n}r} \le \sqrt{n}, \qquad f'(r)/r \le \frac{1 - e^{-\sqrt{n}r}}{r} \le \sqrt{n}. \] Thus, $\nabla^2 \theta_n(x) \preceq \sqrt{n} I$, and the second inequality follows from Taylor expansion. \end{proof} \begin{proof}[Proof of Lemma \ref{Lem:2ndDE-fourier}] Let $\gamma_G$ and $c_G$ be the constants from Lemmas \ref{Lem:gap} and \ref{Lem:1stDE-fourier}, respectively. Define the events \[ {\mathcal G}_t := \bigcap_{s = 0}^t \left\{ X_\r(\sigma_s) \preceq -\frac{\gamma_G}{n} \right\}, \qquad {\mathcal G}'_t := \bigcap_{s = 0}^t \left\{ X_\r(\sigma_s) \preceq -\frac{1 - \sqrt{d_\r} e^{-c_G s/n} - 2n^{-1/8}}{n} \right\}. \] Note that $\sigma'_\ast \in \Sast{\frac{1}{4{\mathcal Q}}} \subseteq \Snon{1/3}$. Hence, by Lemmas \ref{Lem:Burn} and \ref{Lem:gap}, we have $\Pr({\mathcal G}^{\sf c}_{n^2}) \le C_G n^2 e^{-n/10}$. We also have \begin{align*} X_\r(s) &= \frac{1}{n - 1} \left( \frac{x_\r(s)+x_\r(s)^\ast}{2} - \frac{n - 1}{n}I_{d_\r} \right) \preceq -\frac{1}{n}\left( 1 - \frac{n \|x_\r(s)\|_{{\rm HS}}}{n - 1} \right) I_{d_\r} \\ &\preceq -\frac{1}{n}\left( 1 - \|x_\r(s)\|_{{\rm HS}} - \frac{\sqrt{d_\r}}{n-1} \right) I_{d_\r}, \end{align*} where we have used the fact that $\left\|\frac{x_\r(s)+x_\r(s)^\ast}{2}\right\|_{op} \le \|x_\r(s)\|_{op} \le \|x_\r(s)\|_{{\rm HS}}$. Lemma \ref{Lem:1stDE-fourier} then implies that $\Pr({\mathcal G}'^{\sf c}_{n^2}) \le \frac{1}{n^3}$. Thus, setting \[ \varphi(t) := \max(\gamma_G, 1 - \sqrt{d_\r} e^{-c_G t/n} - 2 n^{-1/8}), \] \[ {\mathcal H}_t := {\mathcal G}_t \cap {\mathcal G}'_t = \bigcap_{s = 0}^t \left\{ X_\r(\sigma_s) \preceq -\frac{\varphi(t)}{n} \right\}, \] we conclude that \[ \Pr({\mathcal H}^{\sf c}_{n^2}) \le \Pr({\mathcal G}^{\sf c}_{n^2}) + \Pr({\mathcal G}'^{\sf c}_{n^2}) \le \frac{2}{n^3} \] for all large enough $n$. Next, we turn to \eqref{Eq:DEy} and apply $\theta_n$ to both sides, where we identify ${\mathbb C}^{d_\r^2}$ with ${\mathbb R}^{2d_\r^2}$. Using Lemma \ref{Lem:theta} and taking the conditional expectation, we obtain \begin{align*} {\mathbb E}\left[ \theta_n\left( y_{a, \r}(t+1) \right) \,\middle|\, {\mathcal F}_t \right] &\le \theta_n\left( y_{a, \r}(t) (I_{d_\r} + X_\r(t))\right) + \frac{8 {\mathcal Q}^4 d_\r}{n \sqrt{n}} \\ &\le \theta_n(\|I_{d_\r} + X_\r(t)\|_{op} \cdot y_{a, \r}(t)) + \frac{8 {\mathcal Q}^4 d_\r}{n \sqrt{n}} \\ &\le \|I_{d_\r} + X_\r(t)\|_{op} \cdot \theta_n(y_{a, \r}(t)) + \frac{8 {\mathcal Q}^4 d_\r}{n \sqrt{n}}, \end{align*} where the second inequality follows from the variational formula for operator norm (i.e. that $\|BA\|_{{\rm HS}} \le \|A\|_{op} \|B\|_{{\rm HS}}$), and the third inequality follows from the fact that $\theta_n$ is convex with $\theta_n(0) = 0$. Thus, we may write \[ \theta_n(y_{a, \r}(t+1)) \le \|I_{d_\r} + X_\r(t)\|_{op} \cdot \theta_n(y_{a, \r}(t)) + M'(t+1) \] where \[ {\mathbb E}[ M'(t+1) \mid {\mathcal F}_t ] \le \frac{8 {\mathcal Q}^4 d_\r}{n\sqrt{n}}, \qquad |M'(t+1)| \le \frac{8{\mathcal Q}^2 \sqrt{d_\r}}{n}. \] Now, let $z_t := {\bf 1}_{{\mathcal H}_t} \theta_n(y_{a, \r}(t))$, and note that since $X_\r(\sigma) \succeq -\frac{2}{n-1} I_{d_\r}$, we have $\|I_{d_\r} + X_\r(t)\|_{op} \le 1-\frac{\varphi(t)}{n}$ whenever ${\mathcal H}_t$ holds. Thus, \[ z_{t+1} \le \|I_{d_\r} + X_\r(t)\|_{op} \cdot z_t + {\bf 1}_{{\mathcal H}_t}M'(t+1) \le \left(1 - \frac{1}{n} \varphi(t) \right) z_t + {\bf 1}_{{\mathcal H}_t}M'(t+1). \] We may then apply Lemma \ref{Lem:GenDE} with $\varepsilon = \frac{1}{n}$ and $D = 8{\mathcal Q}^4 d_\r$. Note that \begin{align*} \int_0^t \varphi(s) \,ds &\ge \left(1 - 2 n^{-\frac{1}{8}}\right)t - \sqrt{d_\r} \int_0^\infty e^{-\frac{c_G s}{n}} \,ds \ge t - O_G(n) \end{align*} for all large enough $n$. Thus, Lemma \ref{Lem:GenDE} implies that \begin{equation} \label{Eq:2ndDE-fourier-z_t} \Pr\left(z_t \ge \frac{\lambda}{\sqrt{n}} + C_G e^{-t/n} \cdot z_0 \right) \le C'_G e^{-c'_G \lambda^2}. \end{equation} Consequently, \begin{align*} &\Pr\left( \|y_{a, \r}(t)\|_{{\rm HS}} \ge R\left( \frac{1}{\sqrt{n}} + e^{-\frac{t}{n}}\|y_{a, \r}(0)\|_{{\rm HS}} \right) \right) \\ &\qquad\qquad \le \Pr\left( \theta_n(y_{a, \r}(t)) \ge \frac{R - 1}{\sqrt{n}} + Re^{-\frac{t}{n}}\|y_{a, \r}(0)\|_{{\rm HS}} \right) \\ &\qquad\qquad \le \Pr\left( \theta_n(y_{a, \r}(t)) \ge \frac{R - 1}{\sqrt{n}} + Re^{-\frac{t}{n}}\theta_n(y_{a, \r}(0)) \right) \\ &\qquad\qquad \le \Pr\left( z_t \ge \frac{R - 1}{\sqrt{n}} + Re^{-\frac{t}{n}}z_0 \right) + \Pr({\mathcal H}^{\sf c}_{n^2}) \\ &\qquad\qquad \le e^{-\O_G(R^2) + O_G(1)} + \frac{2}{n^3}, \end{align*} as desired. \end{proof} \section{Construction of the coupling: Proof of Lemma \ref{Lem:RW}}\label{Sec:Coupling-Proof} For each $\d>0$, we define a subset of $\{0, 1, \dots, n\}^{G \times G}$ by \[ \Mc_\d:=\left\{ (n_{a, b})_{a, b \in G} \ : \ n_{a, b} \ge \frac{(1-\d) n}{{\mathcal Q}^2} \ \text{for every $a, b \in G$ and}\ \sum_{a, b \in G}n_{a, b}=n\right\}. \] \begin{lemma}\label{Lem:coupling} Consider a configuration $\sigma_\ast \in {\mathcal S}$ and a constant $0<\d \le \frac{1}{2{\mathcal Q}^2}$, and assume that $(1 - \d)n/{\mathcal Q}^2$ is an integer. Let $(\sigma_t)_{t \ge 0}$ and $(\tilde\sigma_t)_{t \ge 0}$ be two product replacement chains started at $\sigma$ and $\tilde\sigma$, respectively. Then, there exists a coupling $(\sigma_t, \tilde \sigma_t)$ of the Markov chains satisfying the following: Let \[ D_t:=\frac{1}{2}\sum_{a, b \in G}|n^{\sigma_\ast}_{a, b}(\sigma_t) - n^{\sigma_\ast}_{a, b}(\tilde \sigma_t)|. \] Then, on the event $\{(n^{\sigma_\ast}_{a, b}(\sigma_t))_{a, b \in G}, (n^{\sigma_\ast}_{a, b}(\tilde \sigma_t))_{a, b \in G} \in \Mc_\d\}$ and $\{D_t > 0\}$, one has \begin{align} {\mathbb E}_{\sigma, \tilde \sigma}[D_{t+1}-D_t \mid \sigma_t, \tilde \sigma_t] &\le 0, \label{Eq:D-drift-0} \\ \Pr_{\sigma, \tilde \sigma}\left(D_{t+1} - D_t \neq 0 \mid \sigma_t, \tilde \sigma_t \right) &\ge \frac{(1-\d)^2}{4{\mathcal Q}^3}. \label{Eq:D-fluctuate} \end{align} \end{lemma} \begin{proof} Let us abbreviate $n_{a, b}(t) = n^{\sigma_\ast}_{a, b}(\sigma_t)$ and $\tilde n_{a, b}(t) = n^{\sigma_\ast}_{a, b}(\tilde \sigma_t)$. Let $m_{a, b}(t):=\min(n_{a, b}(t), \tilde n_{a, b}(t))$. For each $a \in G$, we define the quantity \[ d_a(t) := \frac{1}{2}\sum_{b \in G} |n_{a, b}(t) - \tilde{n}_{a, b}(t)| = \sum_{b \in G} n_{a, b}(t) - \sum_{b \in G} m_{a, b}(t), \] so that $D_t = \sum_{a \in G} d_a(t)$. For accounting purposes, it is helpful to introduce two sequences \[ (x_1, x_2, \ldots , x_n) \text{ and } (\tilde{x}_1, \tilde{x}_2, \ldots , \tilde{x}_n) \] of elements of $G \times G$. These sequences are chosen so that the number of $x_k$ equal to $(a, b)$ is exactly $n_{a, b}$, and similarly the number of $\tilde{x}_k$ equal to $(a, b)$ is $\tilde{n}_{a, b}$. Moreover, we arrange their indices in a coordinated fashion, as described below. We define three families of disjoint sets: $P_{a, b}$, $Q_a$, and $R_a \subset [n]$. \begin{itemize} \item For each $a, b \in G$, let $P_{a, b}$ be a set of size $(1 - \delta)n/{\mathcal Q}^2$ such that for any $k \in P_{a, b}$, we have $x_k =\tilde{x}_k = (a, b)$. (This is possible provided that $(n_{a, b}(t)), (\tilde n_{a, b}(t)) \in \Mc_\d$ holds.) \item For each $a \in G$, let $Q_a$ be a set of size $\sum_{b \in G}(m_{a, b} - |P_{a, b}|)$ such that for any $k \in Q_a$, $x_k =\tilde{x}_k= (a, b)$ for some $b$. (Note that $Q_a$ may be empty.) \item For each $a \in G$, let $R_a$ be a set of size $d_a$ such that for any $k \in R_a$, $x_k$ and $\tilde{x}_k$ both have $a$ as their first coordinate. (This $R_a$ is well-defined since $\sum_b n_{a, b} = \sum_b \tilde n_{a, b}$ for each $a$; it may also be empty.) \end{itemize} Define \[ P := \bigsqcup_{a, b \in G} P_{a, b}, \qquad Q := \bigsqcup_{a \in G} Q_a, \qquad R := \bigsqcup_{a \in G} R_a. \] Suppose that $D_t > 0$, so that for some $a_*, b_*, b_*' \in G$ we have $n_{a_*, b_*} > \tilde{n}_{a_*, b_*}$ and $n_{a_*, b'_*} < \tilde{n}_{a_*, b'_*}$. Let us consider all possible ways to sample a pair of indices and a sign $(k, l, s) \in \{1, 2, \ldots , n\}^2 \times \{\pm 1\}$ with $k \neq l$. Suppose $x_k = (a_k, b_k)$ and $x_l = (a_l, b_l)$. We think of $(k, l, +1)$ as corresponding to a move on $(n_{a, b}(t))$ where $n_{a_k, b_k}$ is decremented and $n_{a_k, (b_k \cdot b_l)}$ is incremented. Similarly, $(k, l, -1)$ corresponds to a move where $n_{a_k, b_k}$ is decremented and $n_{a_k, (b_k \cdot b_l^{-1})}$ is incremented. We may also think of $(k, l, \pm 1)$ as corresponding to moves on $(\tilde n_{a, b}(t))$ in an analogous way. \begin{figure} \includegraphics{pqr_figure.pdf} \caption{Illustration of cases (i) through (iv).} \label{fig:coupling-cases} \end{figure} We now analyze four cases, as illustrated in Figure \ref{fig:coupling-cases}. \paragraph{(i) {\bf Case $(k, l) \in (P \sqcup Q) \times (P \sqcup Q)$.}} For all but an exceptional situation described below, we apply the move corresponding to $(k, l, s)$ to both states $(n_{a, b}(t))$ and $(\tilde n_{a, b}(t))$. In these cases, $D_{t+1}=D_t$. We now describe the exceptional situation. Define \[ S = P_{a_*, b_*} \times \left(\bigsqcup_{c \in G} P_{c, (b_*^{-1} \cdot b'_*)}\right) \qquad\text{and}\qquad S' = P_{a_*, b'_*} \times \left(\bigsqcup_{c \in G} P_{c, {\rm id}}\right). \] Then, the exceptional situation occurs when $s = +1$ and $(k, l) \in S \sqcup S'$. Take any bijection $\tau$ from $S$ to $S'$. If $(k, l) \in S$, then we apply $(k, l, +1)$ to $(n_{a, b}(t))$ while applying $(\tau(k, l), +1)$ to $(\tilde n_{a, b}(t))$. This increments $n_{a_*, b'_*}$, decrements $n_{a_*, b_*}$, and has no effect on the $(\tilde{n}_{a, b}(t))$. The overall effect is that $D_{t+1} = D_t - 1$. If instead $(k, l) \in S'$, then we apply $(k, l, +1)$ to $(n_{a, b}(t))$ and $(\tau^{-1}(k, l), +1)$ to $(\tilde n_{a, b}(t))$. A similar analysis shows that in this case $D_{t+1} = D_t + 1$. The exceptional event occurs with probability $\frac{(1 - \d)^2}{2{\mathcal Q}^3}$, and when it occurs, $D_t$ increases or decreases by $1$ with equal probability. Thus, the exceptional situation plays the role of introducing some unbiased fluctuation in $D_t$ and gives us \eqref{Eq:D-fluctuate}. \\ \paragraph{(ii) {\bf Case $(k, l) \in (Q \sqcup R) \times (Q \sqcup R)$ but $(k, l) \not\in Q \times Q$.}} This occurs with probability \[ \frac{1}{n(n-1)}\left((|Q|+|R|)(|Q|+|R|-1)-|Q|(|Q|-1)\right) \] which is at most \[ \frac{2}{n(n - 1)}(|Q| + |R|) |R| = \frac{2 \d}{n - 1}D_t. \] Apply the move corresponding to $(k, l, s)$ to both states. This increases $D_t$ by at most $1$. We will see later that the effect of this case is small compared to the other cases. \\ \paragraph{(iii) {\bf Case $(k, l) \in P \times R$.}} This occurs with probability \[ \frac{1}{n(n - 1)}|P| |R| =\frac{1 - \d}{n - 1}D_t. \] Apply the move corresponding to $(k, l, s)$ to both states. Again, this increases $D_t$ by at most $1$, but there is also a chance not to increase. Suppose that $x_l = (a_1, b_1)$ and $\tilde{x}_l = (a_1, \tilde{b}_1)$, and suppose that $k \in P_{a_2, b_2}$. Then the move has the effect of decreasing $n_{a_2, b_2}$ and $\tilde{n}_{a_2, b_2}$ while increasing $n_{a_2, (b_2 \cdot b_1^s)}$ and $\tilde{n}_{a_2, (b_2\cdot \tilde{b}_1^s)}$. Note that conditioned on this case happening, $(a_2, b_2)$ is distributed uniformly over $G \times G$. When $(a_2, (b_2\cdot \tilde{b}_1^s)) = (a_*, b_*)$ or $(a_2, (b_2\cdot b_1^s)) = (a_*, b'_*)$, the move does not increase $D_t$. Therefore there is at least a $2/{\mathcal Q}^2$ chance that $D_t$ is actually not increased. Hence, the probability that $D_t$ is increased by $1$ is at most \[ \left(1 - \frac{2}{{\mathcal Q}^2}\right)\frac{1 - \d}{n-1}D_t. \] \paragraph{(iv) {\bf Case $(k, l) \in R \times P$.}} This occurs with probability \[ \frac{1}{n(n - 1)} |R| |P| = \frac{1 - \d}{n - 1} D_t. \] Suppose that $x_k = (a, b)$ and $\tilde{x}_k = (a, \tilde{b})$. Let $\tau$ be a permutation of $P$ such that for $l \in P_{a, c}$, one has $\tau(l) \in P_{a, \tilde{b}^{-1}\cdot b \cdot c^s}$. Then apply $(k, l, s)$ to $(n_{a, b}(t))$ and apply $(k, \tau(l), s)$ to $(\tilde n_{a, b}(t))$. This always decreases $D_t$ by $1$. \\ Let us now summarize what we know when $(n_{a, b}(t)), (\tilde n_{a, b}(t)) \in \Mc_\d$ and $D_t>0$. From Cases (i), (ii), and (iii), we have \[ \Pr_{\sigma, \tilde \sigma}(D_{t+1} = D_t + 1 \mid \sigma_t, \tilde \sigma_t) \le \left(1-\frac{2(1-\d)}{{\mathcal Q}^2}+\d\right)\frac{D_t}{n-1} + \frac{(1 - \d)^2}{4{\mathcal Q}^3}. \] From Cases (i) and (iv), we have \[ \Pr_{\sigma, \tilde \sigma}(D_{t+1} = D_t - 1 \mid \sigma_t, \tilde \sigma_t) \ge (1-\d)\frac{D_t}{n-1} + \frac{(1 - \d)^2}{4{\mathcal Q}^3}. \] Therefore, if $0<\d \le \frac{1}{2{\mathcal Q}^2}$, then \[{\mathbb E}_{\sigma, \tilde \sigma}[D_{t+1}-D_t \mid \sigma_t, \tilde \sigma_t] \le 0, \] verifying \eqref{Eq:D-drift-0}. To fully define the coupling, when $D_t = 0$, we can couple $\sigma_t$ and $\sigma_t$ to be identical, and if either $(n_{a, b}(t)) \notin \Mc_\d$ or $(\tilde n_{a, b}(t)) \notin \Mc_\d$, we may run the two chains independently. \end{proof} \begin{proof}[Proof of Lemma \ref{Lem:RW}] Since $\sigma \in \Sast{\sigma_\ast, \frac{R}{\sqrt{n}}}$, we must have for each $a \in G$ and $\r \in G^\ast$ that $\|y^{\sigma_\ast}_{a, \r}(\sigma)\|_{{\rm HS}} \le \frac{R}{\sqrt{n}}$. Note that for large enough $n$, we have $\Sast{\sigma_\ast, \frac{R}{\sqrt{n}}} \subseteq \Sast{\frac{1}{5{\mathcal Q}^3}}$. Thus, we may apply Lemma \ref{Lem:2ndDE-fourier} to obtain \begin{equation} \label{Eq:RW-1} \Pr\left( \bigcup_{t = 0}^{n^2} \left\{ \|y^{\sigma_\ast}_{a, \r}(\sigma_t)\|_{{\rm HS}} \ge \frac{1}{5{\mathcal Q}^3} \right\} \right) \le n^2 \left( e^{-\O_{G}(n) + O_{G}(1)} + \frac{2}{n^3} \right) \le \frac{3}{n} \end{equation} for large enough $n$. Define the event \[ {\mathcal G}_t := \left\{ \sigma_s \in \Sast{\sigma_\ast, \frac{1}{5{\mathcal Q}^3}} \text{ for all $1 \le s \le t$} \right\}. \] The Plancherel formula applied to \eqref{Eq:RW-1} implies that $\Pr({\mathcal G}^{\sf c}_{n^2}) \le \frac{3{\mathcal Q}^2}{n}$. We may analogously define an event $\tilde{\mathcal G}_t$ for $\tilde\sigma$ and let ${\mathcal A}_t := {\mathcal G}_t \cap \tilde{\mathcal G}_t$. Thus, $\Pr({\mathcal A}_{n^2}^{\sf c}) \le \frac{6{\mathcal Q}^2}{n}$. Pick $\d' \in \left(\frac{2}{5{\mathcal Q}^2}, \frac{3}{7{\mathcal Q}^2}\right)$ so that $(1 - \d')n/{\mathcal Q}^2$ is an integer. Note that when ${\mathcal A}_t$ holds, we have \[ \sigma_t \in \Sast{\sigma_\ast, \frac{1}{5{\mathcal Q}^3}} \quad \text{and} \quad \sigma_\ast \in \Sast{\frac{1}{5{\mathcal Q}^3}}\implies (n_{a, b}(t)) \in \Mc_{\frac{2}{5{\mathcal Q}^2}} \subseteq \Mc_{\d'}, \] and similarly $\tilde\sigma_t \in \Mc_{\d'}$. Thus, we may invoke Lemma \ref{Lem:coupling} to give a coupling between $\sigma$ and $\tilde\sigma$ where on the event ${\mathcal A}_t$, the quantity $D_t$ is more likely to decrease than increase. Letting ${\bf D}_t:={\bf 1}_{{\mathcal A}_t}D_t$, we see that $({\bf D}_t)$ is a supermartingale with respect to $({\mathcal F}_t)$. Define \[ \t := \min\{ t \ge 0 : D_t=0 \}, \qquad {\tilde \t} := \min\{ t \ge 0 : {\bf D}_t=0 \}. \] Then, Lemma \ref{Lem:coupling} ensures that on the event $\{\tilde\t > t\}$, we have ${\rm Var}({\bf D}_{t+1}\mid {\mathcal F}_t) \ge \a^2$, where $\a^2:=\left(1- \frac{1}{{\mathcal Q}^2}\right)\frac{(1-\d')^2}{4{\mathcal Q}^3}$. By \cite[Proposition 17.20]{LevinPeresWilmer}, for every $u > 12/\a^2$, \begin{equation}\label{Eq:RW} \Pr(\tilde \t > u) \le \frac{4 {\bf D}_0}{\a\sqrt{u}}. \end{equation} Recall that $T = \ceil{\b n}$ and $D_0 \le \sqrt{{\mathcal Q}} R\sqrt{n}$. As long as $\b$ is large enough, we may apply \eqref{Eq:RW} with $u = T$ to get \[ \Pr_{\sigma, \tilde \sigma}(\t > T) \le \frac{16 {\mathcal Q}^2 R}{(1-\d')\sqrt{\b}} + \Pr({\mathcal A}_T^{\sf c}) \le \frac{32 {\mathcal Q}^2 R}{\sqrt{\b}} \] for all large enough $n$, as desired. \end{proof} \section{Proof of Theorem \ref{Thm:main} (\ref{Eq:LB})}\label{Sec:LB} \newcommand{n^{\{\id\}}_{non}}{n^{\{{\rm id}\}}_{non}} The lower bound is proved essentially by showing that the estimates of Lemmas \ref{Lem:Burn} and \ref{Lem:2ndDE} cannot be improved. Let $a_1, a_2, \ldots , a_k$ be a set of generators for $G$. Let $\sigma_\star \in {\mathcal S}$ be the configuration given by \[ \sigma_\star(i) = \begin{cases} a_i & \text{if $i \le k$,} \\ 0 & \text{otherwise}. \end{cases} \] We will analyze the Markov chain started at $\sigma_\star$ and show that it does not mix too fast. Recall from Section \ref{Sec:UB} the notation \[ n^{\{\id\}}_{non}(\sigma) = |\{ i \in [n] : \sigma(i) \ne {\rm id} \}| \] for the number of sites in $\sigma$ that do not contain the identity. We first show that if we run the chain for slightly less than $n \log n$ steps, most of the sites will still contain the identity. \begin{lemma} \label{Lem:Burn-lower} Let $T := \floor{n \log n - Rn}$. Then, \[ \Pr_{\sigma_\star}\left( n^{\{\id\}}_{non}(\sigma_T) \ge \frac{n}{3} \right) \le \frac{4{\mathcal Q}^2}{R^2}. \] \end{lemma} \begin{proof} Recall that in one step of our Markov chain, we pick two indices $i, j \in [n]$ and replace $\sigma(i)$ with $\sigma(i) \cdot \sigma(j)$ or $\sigma(i) \cdot \sigma(j)^{-1}$. The only way for $n^{\{\id\}}_{non}(\sigma_t)$ to increase after this step is if $\sigma(j) \ne {\rm id}$. Thus, \begin{equation} \label{Eq:nnon} \Pr( n^{\{\id\}}_{non}(\sigma_{t+1}) = n^{\{\id\}}_{non}(\sigma_t) + 1 \mid n^{\{\id\}}_{non}(\sigma_t)) \le \frac{n^{\{\id\}}_{non}(\sigma_t)}{n}. \end{equation} Let $\t := \min\{ t \ge 0 : n^{\{\id\}}_{non}(\sigma_t) \ge \frac{n}{3} \}$ be the first time that $n^{\{\id\}}_{non}(\sigma_t)$ is at least $\frac{n}{3}$. We have that $n^{\{\id\}}_{non}(\sigma_\star) = k$, so it follows from \eqref{Eq:nnon} that $\t$ stochastically dominates the sum \[ G := \sum_{s = k}^{\floor{n/3}} G_s, \] where the $G_s$ are independent geometric variables with success probability $\frac{s}{n}$. Note that we have the bounds \[ {\mathbb E} G = \sum_{s = k}^{\floor{n/3}} \frac{n}{s} \ge n \left( \log \floor{\frac{n}{3}} - \log k \right), \qquad {\rm Var}(G) = \sum_{s = k}^{\floor{n/3}} \frac{n(n-s)}{s^2} \le n^2. \] Hence, \begin{align*} \Pr(\t < T) &\;\le\; \Pr(G < T) \;\le\; \Pr(G < {\mathbb E} G + n\log(3k) - Rn ) \\ &\;\le\; \frac{n^2}{n^2(R - \log(3k))^2} \le \frac{4}{R^2} \end{align*} for $R \ge 2 {\mathcal Q} \ge 2 \log (3k)$. On the other hand, the bound claimed in the lemma is trivial for $R \le 2 {\mathcal Q}$, so we have completed the proof. \end{proof} Next, we show that it really takes about $\frac{1}{2}n \log n$ steps for the Fourier coefficients $x_\r$ to decay to $O\left( \frac{1}{\sqrt{n}} \right)$, as suggested by Lemma \ref{Lem:2ndDE}. Note that it suffices here to analyze the $x_\r$ instead of the $y_{a, \r}$, which simplifies our analysis. Actually, it suffices to consider (the real part of) the trace of $x_\r$. Here the orthogonality of characters reads $\frac{1}{{\mathcal Q}}\sum_{a \in G} \Tr \r(a)=0$, and it takes about $\frac{1}{2}n \log n$ steps for $\Re \Tr x_\r(t)$ to decay to $O\left( \frac{1}{\sqrt{n}} \right)$. \begin{lemma} \label{Lem:DE-lower} Consider any $\r \in G^\ast$ and any $R > 5$. Let $T := \floor{\frac{1}{2}n \log n - Rn}$, and suppose that $\sigma \in {\mathcal S}$ satisfies $n^{\{\id\}}_{non}(\sigma) \le \frac{n}{3}$. Then, \[ \Pr_{\sigma}\left( \|x_\r(\sigma_T)\|_{{\rm HS}} \le \frac{R}{\sqrt{n}} \right) \le \frac{4{\mathcal Q}^2}{R^2}. \] \end{lemma} \begin{proof} Let $z(t) := (1/d_\r) \Tr (x_\r(t)+x_\r(t)^\ast)/2$. Then, noting that \eqref{Eq:DEfourier} also holds for $x_\r(t)^\ast$ since $x_{\r^\ast}(t)=x_\r(t)^\ast$, we have \[ z(t+1)-z(t) = \frac{1}{n-1}\frac{1}{d_\r} \Tr \left(\frac{x_\r(t)+x_\r(t)^\ast}{2}\right)^2-\frac{1}{n}z(t) + M(t+1), \] where \[ {\mathbb E}[M(t+1) \mid {\mathcal F}_t]=0 \qquad \text{and} \qquad |M(t)| \le \frac{2{\mathcal Q}}{n}. \] Here we have \[\frac{1}{d_\r} \Tr \left(\frac{x_\r(t)+x_\r(t)^\ast}{2}\right)^2 \ge z(t)^2.\] We compare $z(t)$ to another process $(w(t))_{t \ge 0}$ defined by $w(0) := \frac{1}{3}$ and \begin{equation} \label{Eq:w-equation} w(t+1) := \left(1-\frac{1}{n}\right)w(t) + M(t+1). \end{equation} We will show by induction that $z(t) \ge w(t)$ for all $t$. For the base case, note that since $n^{\{\id\}}_{non}(\sigma) \le \frac{n}{3}$, we have \[ z(0) = \frac{1}{d_\r}\Re \Tr \sum_{a \in G} \frac{n_a(t)}{n} \cdot \r(a) \ge \frac{2}{3} - \frac{1}{3} = \frac{1}{3}. \] Suppose now that $z(t) \ge w(t)$. Then, \begin{align*} z(t+1) &\ge z(t) + \frac{1}{n-1}z(t)^2-\frac{1}{n}z(t) + M(t+1) \\ &\ge \left(1 - \frac{1}{n}\right) w(t) + M(t+1) = w(t+1), \end{align*} completing the induction. It now suffices to lower bound $w(T)$. To this end, we first note that applying \eqref{Eq:w-equation} repeatedly and taking expectations, we obtain \[ {\mathbb E} w(T) = \left(1 - \frac{1}{n}\right)^T \cdot \frac{1}{3} \ge \frac{e^R}{6\sqrt{n}} \ge \frac{2R}{\sqrt{n}}. \] In order to calculate the variance of $w(T)$, we can also square \eqref{Eq:w-equation} and take the expectation, which gives us \begin{align*} {\rm Var}(w(T)) &= {\mathbb E} w(T)^2 - ({\mathbb E} w(T))^2 \\ &\le \left(1 - \frac{1}{n}\right)^{2T} \cdot \frac{1}{9} + n \cdot \left( \frac{2{\mathcal Q}}{n} \right)^2 - \left( \left(1 - \frac{1}{n}\right)^T \cdot \frac{1}{3} \right)^2 \\ &= \frac{4{\mathcal Q}^2}{n}. \end{align*} Then, by Chebyshev's inequality, we have \begin{align*} \Pr_{\sigma}\left( \|x_\r(\sigma_T)\|_{{\rm HS}} \le \frac{R}{\sqrt{n}} \right) &\le \Pr\left( z(T) \le \frac{R}{\sqrt{n}} \right) \le \Pr\left( w(T) \le \frac{R}{\sqrt{n}} \right) \\ &\le \frac{4{\mathcal Q}^2/n}{(R/\sqrt{n})^2} = \frac{4{\mathcal Q}^2}{R^2}, \end{align*} as desired. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm:main} (\ref{Eq:LB})] Let $T = T_1 + T_2$, where $T_1 := \floor{n \log n - \b n}$ and $T_2 := \floor{\frac{1}{2} n \log n - \b n}$. Fix any $\r \in G^\ast$. By Lemma \ref{Lem:Burn-lower} followed by Lemma \ref{Lem:DE-lower}, we have for large enough $\b$ that \[ \Pr_{\sigma_\star} \left(\sigma_T \in \Sast{\frac{\b}{\sqrt{n}}} \right) \le \Pr_{\sigma_\star}\left( \|x_\r(T)\|_{{\rm HS}} \le \sqrt{\frac{{\mathcal Q}}{d_\r}}\frac{\b}{\sqrt{n}} \right) \le \frac{8{\mathcal Q}^2}{\b^2}. \] On the other hand, Lemma \ref{Lem:stationary} tells us that \[ \pi\left({\mathcal S}_\ast\left(\frac{\b}{\sqrt{n}}\right)\right) \ge 1- \frac{c_G}{\b^2}. \] Consequently, \[ d_{\sigma_\star}(T) \ge 1 - \frac{c_G}{\b^2} - \frac{8{\mathcal Q}^2}{\b^2}, \] which tends to $1$ as $\b \rightarrow \infty$, establishing \eqref{Eq:LB}. \end{proof} \section*{Acknowledgements} This work was initiated while R.T. and A.Z. were visiting Microsoft Research in Redmond. They thank Microsoft Research for the hospitality. R.T. was also visiting the University of Washington in Seattle and thanks Professor Christopher Hoffman for making his visit possible. R.T. is supported by JSPS Grant-in-Aid for Young Scientists (B) 17K14178. A.Z. is supported by a Stanford Graduate Fellowship. \bibliographystyle{alpha}

https://arxiv.org/abs/2005.08899

Non-asymptotic Results for Singular Values of Gaussian Matrix Products

This article concerns the non-asymptotic analysis of the singular values (and Lyapunov exponents) of Gaussian matrix products in the regime where $N,$ the number of term in the product, is large and $n,$ the size of the matrices, may be large or small and may depend on $N$. We obtain concentration estimates for sums of Lyapunov exponents, a quantitative rate of convergence of the empirical measure of the normalized squared singular values to the uniform distribution on $[0,1]$, and results on the joint normality of Lyapunov exponents when $N$ is sufficiently large as a function of $n.$ Our technique consists of non-asymptotic versions of the ergodic theory approach at $N=\infty$ due originally to Furstenberg and Kesten in the 1960's, which were then further developed by Newman and Isopi-Newman as well as by a number of other authors in the 1980's. Our key technical idea is that small ball probabilities for volumes of random projections give a way to quantify convergence in the multiplicative ergodic theorem for random matrices.

\section{Introduction} This article is about the spectral theory of random matrix products \begin{equation}\label{E:X-def} X_{N,n} := A_N\cdots A_1, \end{equation} where $A_i$ are independent $n\times n$ matrices with independent real Gaussian entries $\lr{A_i}_{\alpha\beta}\sim \mathcal N(0,1/n)$ of mean zero and variance $1/n.$ We are primarily interested in the situation when $N$ is large and finite, while $n$ may depend on $N$ and may be either small or large. Our results concern the singular values of $X_{N,n}:$ \begin{equation}\label{E:sing-def} s_1(X_{N,n})\geq \cdots \geq s_n(X_{N,n}) \end{equation} and can be summarized informally as follows: \begin{enumerate} \item We prove that as $N,n$ tend to infinity \textit{at any relative rate} the global distribution of the normalized squared singular values $\set{s_i(X_{N,n})^{2/N},\, i = 1,\ldots, n}$ converges to the uniform distribution on $[0,1]$ (see \S \ref{S:global-intro} and Theorem \ref{T:global}). Unlike previous results, we obtain quantitative concentration estimates valid for all $N,n$ larger than a fixed constant. See also \S \ref{S:why-triangle} for a heuristic explanation of why the uniform distribution appears in this context. \item We prove that as long as $N$ is sufficiently large as a function of $n,$ the Lyapunov exponents \begin{equation}\label{E:lyapunov-def} \lambda_i = \lambda_i(X_{N,n}) :=\frac{1}{N}\log s_i(X_{N,n}) \end{equation} of $X_{N,n}$ are approximately independent and Gaussian (see Theorem \ref{T:normal} in \S \ref{S:normal}). Unlike previous results, our estimates simultaneously treat all the Lyapunov exponents and provide quantitative concentration estimates when $N$ is large but finite even when $n$ grows with $N.$ \item The statements listed above derived from our main technical result, Theorem \ref{T:sums}, which gives quantitative deviation estimates on sums of Lyapunov exponents of $X_{N,n}$: \[\mathbb P\lr{\abs{\frac{1}{n}\sum_{i=m}^k(\lambda_i -\mu_{n,i})}\geq s}\leq c_1 e^{-c_2nNs\min\set{1, ng_{n,k}(s)}},\quad s\geq \frac{k}{nN}\log\lr{\frac{en}{k}},\] where $\mu_{n,i}$ is defined in \eqref{E:mu-def} and $g_{n,k}(s)$ is a function defined in \eqref{E:g-def}. It is known that (e.g. equations (1) and (7) in \cite{newman1986distribution}) $\mu_{n,i}$ is the almost sure limit of $\lambda_i$ when $N\rightarrow\infty$. \end{enumerate} In this article, we exclusively treat the case of $A_i$ having iid real Gaussian entries. This simplifies a number of arguments, but we conjecture that similar results hold if we assume only that the distribution of the entries of $A_i$ have finite fourth moments and bounded density. We leave this for future work. \subsection{Main Technical Result} Let us set some notation. Denote as in \eqref{E:lyapunov-def} by $\lambda_i=\lambda_i(X_{N,n})$ the Lyapunov exponents of $X_{N,n}$. Further, define for any $s>0$ \begin{equation} \label{E:g-def} g_{n,k}(s) = \left\{ \begin{array}{cc} \min\left\{ 1, \frac{ n s}{ k}\right\} \ ,& \hspace{5mm} k\leq \frac{n}{2} \\ \min\left\{\delta_{n,k}, \frac{ s}{ \log{1/\delta_{n,k}}} \right\} \ ,& \hspace{5mm} \frac{n}{2}<k\leq n \\ \end{array} \right., \end{equation} where for $ k \geq n/2$ we've set \[ \delta_{n,k}:= \frac{ n-k+1}{n}\in \left[ \frac{1}{n} , \frac{n-1}{n}\right].\] Finally, write \begin{equation}\label{E:mu-def} \mu_{n,k} := \E{\frac{1}{2}\log\lr{\frac{1}{n}\chi_{n-k+1}^2}}=\frac{1}{2}\lr{\log\lr{\frac{2}{n}}+\psi\lr{\frac{n-k+1}{2}}}, \end{equation} where $\psi(z) = \frac{d}{dz}\log\Gamma(z)$ is the digamma function and $\chi_m^2$ is a chi-squared random variable with $m$ degrees of freedom. Our main technical result is the following \begin{theorem}[Deviation Estimates for Sums of Lyapunov Exponents]\label{T:sums} \label{T:sums} \noindent There exist universal constants $c_1,c_2,c_3>0$ with the following property. Fix $ 1\leq m \leq k \leq n $ as well as $N\geq 1$. Then, \begin{equation} \label{E:sums} \mathbb P\lr{\abs{\frac{1}{n}\sum_{i=m}^k (\lambda_i - \mu_{n,i})}\geq s} \leq c_2 \exp\lr{-c_3nNs\min\set{1,ng_{n,k}(s)}} {\color{purple} } \end{equation} provided $s\geq c_1\frac{k}{nN}\log(en/k).$ \end{theorem} Theorem \ref{T:sums} holds for every $n,N\geq 1$ and reveals a great deal about the singular values and Lyapunov exponents of $X_{N,n}$. For instance, in the bulk (i.e. when $k$ is comparable to $n$), the restriction on $s$ in \eqref{E:sums} reduces simply to $s> C/ N$, giving information about $X_{N,n}$ as soon as $N$ is large, regardless of $n$. This turns out to be enough to prove Theorem \ref{T:global}, given in \S \ref{S:global-intro} below, which states that the squared singular values of $X_{N,n}$ approximate the uniform distribution on $[0,1]$ when $N,n$ tend to infinity at any relative rate. Theorem \ref{T:sums} also gives precise information about the top Lyapunov exponents of $X_{N,n}$. Indeed, taking $k$ to be fixed in \eqref{E:sums} gives non-trivial information on $\lambda_1,\ldots,\lambda_k$ as soon as $N\gg \log(n)$. Further, note that standard estimates for the digamma function $\psi$ yield \begin{equation}\label{E:mu-est} \mu_{n,k}= \log\lr{1-\frac{k-1}{n}} - \frac{1}{n-k+1} + O\lr{\frac{1}{(n-k+1)^2}}. \end{equation} This shows that the difference between the means $\mu_{n,1}$ and $\mu_{n,2}$ of $\lambda_1$ and $\lambda_2$ is on the order of $1/n$. As soon as $N\gg n\log(n)$, we may apply \eqref{E:sums} with $s\ll \lambda_1-\lambda_2$ to conclude that \[ \frac{s_1(X_{N,n})}{s_2(X_{N,n})}=e^{N(\lambda_1-\lambda_2)}\geq e^{cN/n}\quad \text{with high probability}.\] Hence, we find that $X_{N,n}$ begins to have a large spectral gap in the ``near ergodic'' regime $N\gg n\log(n)$. In fact, in Theorem \ref{T:normal}, we prove that in this regime $\lambda_1,\ldots,\lambda_k$ are also approximately independent Gaussians. We refer the reader to \S \ref{S:normal} for the details. A notable aspect of Theorem \ref{T:sums} is that it applies to any finite $n,N\geq 1$, allowing us to ``interpolate'' between the ergodic $N\gg n$ and free $n\gg N$ regimes. To explain this point, note that matrix products of the form \eqref{E:X-def} have been studied primarily in two setting. The first, which we refer to as the free probability regime occurs when $N$ is fixed and $n\rightarrow \infty.$ This is a kind of maximum entropy regime in which the global distribution of singular values can be characterized in terms of maximizing the non-commutative entropy (cf eg \cite{arous1997large,banica2011free}). The second, which we call the ergodic regime, occurs when $n$ is fixed and $N\rightarrow \infty$. This is a kind of minimal entropy regime in which the Lyapunov exponents (and singular values of $X_{N,n}$) tend to almost sure limits. In both the ergodic and the free regimes, it is often difficult to obtain finite size corrections. Theorem \ref{T:sums} supplies such information. Moreover, since the ergodic and free regimes are usually treated by rather different means, it is unclear which techniques can give information that can interpolate between them. Our approach extends the ergodic techniques pioneered by Furstenburg-Kesten \cite{furstenberg1960products}, further developed in connection to random Schr\"odinger operators by Carmona \cite{carmona1982exponential} and Le Page \cite{le1982theoremes} (cf also \cite{bougerol1985concentration}), and applied in a very similar context as ours by Newman \cite{newman1986distribution} and Isopi-Newman \cite{isopi1992triangle}. It is therefore not surprising that in all of our results, we need $N$ to be in some sense large. Although we do not take this approach in the present article, it is also natural to study spectra of random matrix products by adapting techniques originally developed to treat the case when $N=1$. Indeed, in this setting, there has been considerable effort to obtain non-asymptotic analogs of classical random matrix theory results when $n=\infty$ \cite{rudelson2014lecture,vershynin12,rudelson2017delocalization}, culminating in the resolution of a number of long-standing open problems \cite{rudelson2008littlewood,rudelson2009smallest,adamczak2010quantitative,tikhomirov2020singularity}. More recently, several groups of authors \cite{henriksen2020concentration,huang2020matrix,kathuria2020concentration} have started to extend techniques for obtaining concentration for random matrices tailored (see \cite{tropp2015introduction}) to the small $N$ regime for understanding the kinds of matrix products considered in this article. From this point of view, our article takes a complementary approach, finding extensions of techniques originally coming from the ergodic theory used to analyze the case when $N=\infty$. \subsection{Convergence of Squared Singular Values to the Uniform Distribution}\label{S:global-intro} Prior work \cite{isopi1992triangle,kargin2008lyapunov, tucci2010limits,gorin2018gaussian,liu2018lyapunov,ahn2019fluctuations} shows that in a variety of settings where $n,N\rightarrow \infty$, the global distribution of singular values of $X_{N,n}^{1/N}$ converges to the so-called triangle law after proper normalization. Informally, this means \begin{equation}\label{E:triangle-informal} \lim_{N,n\rightarrow \infty} \frac{1}{n}\#\set{j\leq n~|~ s_i^{1/N}(X_{N,n})\leq t} ~=~ \int_{-\infty}^t 2s{\bf 1}_{\set{s\in [0,1]}} ds~=:~TL(t). \end{equation} The graph of the density $2s{\bf 1}_{\set{s\in [0,1]}}$ of $\TL$ has the shape of a triangle, giving the distribution its name. With the exception of the articles \cite{liu2018lyapunov,gorin2018gaussian}, which obtain much more precise information for products of \textit{complex} Gaussian matrices and the article \cite{ahn2019fluctuations} concerning $\beta-$Jacobi products as well as the real Gaussian case, the majority of prior results about \eqref{E:triangle-informal} (e.g. \cite{isopi1992triangle,kargin2008lyapunov,tucci2010limits}) do not allow $n,N$ to tend to infinity simultaneously. Moreover, all prior results we are aware of do not give quantitative rates of convergence. Theorem \ref{T:global} provides both for the real Gaussian case we consider here. To state it, we note that if a random variable $T$ is distributed according to the triangle law, then $T^2$ is uniformly distributed on $[0,1]$, i.e. has the following cumulative distribution function: \[\mathcal U(t):=\int_{-\infty}^t {\bf 1}_{[0,1]}(t)dt.\] \begin{theorem}[Global Convergence to Triangle Law]\label{T:global} There exist universal constants $c_1,c_2,c_3,c_4>0$ with the following property. For all $\varepsilon\in (0,c_1)$, if $N>c_2/\varepsilon^{2}$ and $n> c_3\log(1/\varepsilon)/\varepsilon$, then the probability \begin{align*} &\mathbb P\lr{\sup_{t\in \mathbb R}\abs{{\frac{1}{n}\#\set{1\leq i\leq n~|~ s_i^{2/N}(X_{N,n})\leq t}} - \mathcal U(t)}\geq \varepsilon } \end{align*} that the cumulative distribution for the squared singular values of $X_{N,n}$ deviates from the uniform distribution by more than $\varepsilon$ is bounded above by \[ 4\exp\left[-c_4n N \varepsilon^{2}\min\set{1,ng_{n,k}(\varepsilon^{2})} \right].\] \end{theorem} In the next section we use the circular law \eqref{E:circular-law} for the (complex) \textit{eigenvalues} of $X_{N,n}^{1/N}$ to give an intuitive but heuristic explanation for why the uniform distribution (or equivalently the triangle law) should appear as the limiting distribution of singular values on $X_{N,n}$. Before doing so, we briefly discuss the dependence of Theorem \ref{T:global} on $N,n$, starting with the former. For fixed $N$, consider an iid random sequence $\set{X_{N,n}}_{n=1}^\infty$ with the product measure. Taking $\varepsilon =2(c_2/N)^{1/2}=:CN^{-1/2}$, Theorem \ref{T:global} shows that \[ \mathbb P \lr{\sup_{t\in \mathbb R}\abs{{\frac{1}{n}\#\set{1\leq i\leq n~|~ s_i^{2/N}(X_{N,n})\leq t}} - \mathcal U(t)}\geq \frac{C}{\sqrt{N}}} \leq 4e^{-cn},\quad c>0. \] Thus, by Borel-Cantelli, we find that \begin{equation}\label{E:N-rate} \sup_{t\in \mathbb R}\abs{\lim_{n\rightarrow \infty}{\frac{1}{n}\#\set{1\leq i\leq n~|~ s_i^{2/N}(X_{N,n})\leq t}} - \mathcal U(t)}\leq \frac{C}{\sqrt{N}}\qquad \text{with probability }1, \end{equation} where by $\lim_{n\rightarrow\infty} a_n$ we mean any limit point of the sequence $a_n$. This $1/\sqrt{N}$ can be seen as a Berry-Esseen-type estimate. To make this precise, consider \[\rho_{N,\infty}:=\lim_{n\rightarrow \infty}\frac{1}{n}\sum_{i=1}^n \delta_{s_{i}(X_{N,n})^{1/N}},\] the large matrix limit for the empirical distribution of normalized singular values for $X_{N,n}.$ It is known \cite[Thm 6.1]{banica2011free} that \[\rho_{N,\infty} = \mathrm{qc}^{\boxtimes N},\qquad \mathrm{qc}(x):=\frac{1}{2\pi}\sqrt{x(2-x)}{\bf 1}_{\set{[0,2]}}(x),\] where $\mathrm{qc}$ is the quarter circle law and $\boxtimes$ is the multiplicative free convolution. Kargin \cite{kargin2008lyapunov} and Tucci \cite{tucci2010limits} show that, consistent with Theorem \ref{T:global}, \[\lim_{N\rightarrow \infty} \rho_{N,\infty} = \mathrm{TL}.\] As far as we know, the optimal rate of convergence for such repeated multiplicative free convolution is unknown. However, from this point of view, \eqref{E:N-rate} shows that the rate of convergence is at least as fast as in the usual central limit theorem. To understand the dependence of Theorem \ref{T:global} on $n,$ we send $N$ to infinity in Theorem \ref{T:global} to obtain as before that there is $C>0$ so that \begin{equation}\label{E:n-rate} \sup_{t\in \mathbb R}\abs{\lim_{N\rightarrow \infty}{\frac{1}{n}\#\set{1\leq i\leq n~|~ s_i^{2/N}(X_{N,n})\leq t}} - \mathcal U(t)}\leq \frac{C\log(n)}{n}\qquad \text{with probability }1, \end{equation} Apart from the $\log(n)$, this estimate is sharp. Indeed, the empirical distribution \[\rho_{\infty,n}:=\lim_{N\rightarrow \infty}\frac{1}{n}\sum_{i=1}^n \delta_{s_{i}(X_{N,n})^{1/N}}\] of singular values in the large number of matrices limit exists almost surely and is deterministic by the Multiplicative Ergodic Theorem. Among other things, Theorem \ref{T:normal} below computes, in agreement with the early work of Newman \cite{newman1986distribution}, this limit in our Gaussian case. The subsequent work of Isopi-Newman \cite{isopi1992triangle} showed that, under minimal assumptions, \[\lim_{n\rightarrow \infty}\rho_{\infty,n}=\mathrm{TL}.\] This of course agrees with Theorem \ref{T:global}, which via \eqref{E:n-rate} provides a natural rate of convergence. This rate is optimal, perhaps up to the $\log(n),$ because the spacing of the atoms in $\rho_{\infty,n}$ is approximately $1/n$. Hence, the distance between $\rho_{\infty,n}$ and triangle law $\mathrm{TL}$, which is a continuous distribution, is bounded below by a constant times $1/n$. \subsection{Why the Uniform Distribution in Theorem \ref{T:global}?}\label{S:why-triangle} A number of articles \cite{newman1986distribution,isopi1992triangle,kargin2008lyapunov,tucci2010limits,liu2018lyapunov, gorin2018gaussian} show as in \eqref{E:triangle-informal} that in the limit where $n,N$ tend to infinity, the singular values $s_i(X_{N,n})^{1/N}$ (or for similar matrix products) converge to the triangle law (and hence their squares converge to the uniform distribution on $[0,1]$). These articles use a variety of techniques ranging from free probability to ergodic theory and special functions. Why does the uniform distribution appear? The purpose of this section to give an intuitive explanation for this phenomenon. After writing an initial draft of this article, we learned from G. Akemann that an explanation similar to the one below can be found on pages 3,4 in \cite{akemann2014universalb}. We also refer the reader to the work of Kieberg-K\"osters \cite{kieburg2016exact} about an exact relation between eigenvalues and singular values for products of complex Ginibre matrices. Since $X_{N,n}$ is not normal with probability $1$, its spectral properties are captured not only its singular values but also by its eigenvalues \begin{equation}\label{E:spec-def} \abs{\zeta_1(X_{N,n})}\geq \cdots \geq \abs{\zeta_n(X_{N,n})},\qquad \zeta_i(X_{N,n})\in \mathbb C. \end{equation} Our argument for why the triangle law appears in Theorem \ref{T:global} relates the singular values and eigenvalues of $X_{N,n}$ and consists of two observations. First, consider the (complex) eigenvalues of $X_{N,n}^{1/N}$ as defined in \eqref{E:spec-def}. It is shown in \cite{gotze2010asymptotic, o2011products} that for each fixed $N$ the empirical distribution of the eigenvalues of $X_{N,n}^{1/N}$ converges weakly almost surely to the uniform measure on the unit disk in $\mathbb C.$ This result is often called the circular law. Informally, it reads \begin{equation}\label{E:circular-law} \lim_{n\rightarrow \infty}\frac{1}{n} \sum_{i=1}^{n} \delta_{\zeta_i^{1/N}}(z)~~=~~ \frac{1}{\pi}{\bf 1}_{\set{\abs{z}\leq 1}},\qquad z\in \mathbb C \end{equation} Precise results on the rate of convergence can be found in \cite{gotze2018rate,jalowy2019rate} and local limit theorems are obtained in \cite{nemish2017local}. Since in polar coordinates $(r,\theta)$ the radial part of the uniform measure on the unit disk is $2rdr$, a corollary of the circular law is that \begin{equation}\label{E:evals-TL} \text{For }N\text{ fixed, as }n\rightarrow \infty,\text{ squared eigenvalue moduli }\abs{\zeta_i}^{2/N}\text{ of }X_{N,n}^{1/N}\text{ converge to }\mathcal U. \end{equation} Thus, the uniform distribution $\mathcal U$ appears naturally as the distribution of the squared moduli of eigenvalues of $X_{N,n}^{1/N}$ for every $N!$ On the other hand, it has been proved that for any fixed finite $n$ \cite{reddy2016lyapunov,reddy2019equality} that when $N$ is large \[\forall i=1,\ldots, n \qquad \abs{\zeta_i}^{1/N}\approx s_i^{1/N}.\] Thus, we extract another piece of intuition: \begin{equation}\label{E:singval-eval} \text{For }n\text{ fixed, as }N\rightarrow \infty,\text{ eigenvalue moduli and singular values of }X_{N,n}^{1/N}\text{ coincide}. \end{equation} Putting together \eqref{E:evals-TL} and \eqref{E:singval-eval}, we conclude heuristically that if both $n,N$ tend to infinity then the distribution of the singular values $s_{i}^{1/N}$ should converge to the triangle law. This is precisely the content of Theorem \ref{T:global}. While the heuristic for \eqref{E:singval-eval} was previously established only when $n$ is fixed, we believe it can also be proved in the regime where $n$ is allowed to grow with $N$ but leave this for future work. \subsection{Distribution of Lyapunov Exponents in the Near Ergodic Regime}\label{S:normal} In addition to studying the global distribution of singular values of $X_{N,n}$, we also obtain in Theorem \ref{T:normal} precise estimates for the joint distribution of the Lyapunov exponents \begin{equation}\label{E:lambda-def} \lambda_i=\lambda_i(X_{N,n})=\frac{1}{N}\log s_i(X_{N,n}) \end{equation} of $X_{N,n}$ in the regime when $N\gg n\log^2(n).$ To state it, we need some notation. Recall first that for each $1\leq k \leq n$ we had set \begin{equation}\label{E:mu-def-second} \mu_{n,k} = \E{\frac{1}{2}\log\lr{\frac{1}{n}\chi_{n-k+1}^2}}=\frac{1}{2}\lr{\log\lr{\frac{2}{n}}+\psi\lr{\frac{n-k+1}{2}}}, \end{equation} where $\psi(z) = \frac{d}{dz}\log\Gamma(z)$ is the digamma function and $\chi_m^2$ is a chi-squared random variable with $m$ degrees of freedom. We also recall the estimate \eqref{E:mu-est}: \begin{equation*} \mu_{n,k}= \log\lr{1-\frac{k-1}{n}} - \frac{1}{n-k+1} + O\lr{\frac{1}{(n-k+1)^2}}. \end{equation*} The quantity $\mu_{n,k}$ already appears in \cite{newman1986distribution,isopi1992triangle} as the mean of $\lambda_k$ when $N\rightarrow \infty.$ We futher define \begin{equation}\label{E:sigma-est} \sigma_{n,k}^2:=\Var\left[\frac{1}{2}\log\lr{\frac{1}{n}\chi_{n-k+1}^2}\right]= \psi'\lr{\frac{n-k+1}{2}} = \frac{1}{2(n-k+1)}+O\lr{\frac{1}{(n-k+1)^2}}, \end{equation} and set \begin{equation}\label{E:mu-sigma-def} \mu_{n,\leq k}:= \lr{\mu_{n,1},\ldots, \mu_{n,k}},\qquad \sigma_{n,\leq k}^2:=\lr{\sigma_{n,1}^2,\ldots, \sigma_{n,k}^2}. \end{equation} Finally, we will consider for two $\mathbb R^k$-valued random variables $X,Y$ the following high-dimensional generalization of the usual Kolmogorov-Smirnov distance: \begin{equation}\label{E:dist-def} d( X, Y ) := \sup_{ C \in {\cal{C}}_{k} } \left| \mathbb P ( X\in C ) -\mathbb P ( Y\in C) \right|, \end{equation} where $ {\cal{C}}_{k}$ is the collection of all convex subsets of $\mathbb R^{k}$. \begin{theorem}[Asymptotic Normality of Lyapunov Exponents]\label{T:normal} There exist constants $C_1,C_2>0$ with the following property. Suppose $X_{N,n}$ is as in \eqref{E:X-def}, fix $1\leq k \leq n$, and write \[\Lambda_k = \lr{\lambda_1,\ldots, \lambda_k}\] for the vector of the top $k$ Lyapunov exponents of $X_{N,n}.$ Then, $\lambda_1,\ldots, \lambda_k$ are approximately independent and Gaussian when $N$ is sufficiently large as a function of $k,n$: \begin{equation}\label{E:dist-est}d\lr{\Lambda_k, \,\mathcal N\lr{\mu_{n,\leq k},\, \frac{1}{N}\mathrm{Diag}\lr{\sigma_{n,\leq k}^2}}} \leq C_2\lr{\frac{k^{7/2}n\log^2(n)\log^2(N/n)}{N}}^{1/2}. \end{equation} Here $\mathcal N(\mu, \Sigma)$ denotes a Gaussian with mean $\mu$ and co-variance $\Sigma$ and for any $v=\lr{v_1,\ldots,v_k}\in \mathbb R^k$ we have written $\mathrm{Diag}(v)$ for the diagonal matrix with $\mathrm{Diag}(v)_{ii}=v_i.$ \end{theorem} \begin{remark} The arguments in \cite{akemann2012universal,akemann2012universal,akemann2014universalb, akemann2019integrable} strongly suggest (see \S \ref{S:intution-IPS}) that for $k$ fixed and independent of $n$, a necessarily and sufficient condition for $\lambda_1,\ldots, \lambda_k$ to be close to independent and Gaussian is $N\gg n.$ Thus, the $\log^2(n)\log^2(N/n)$ in \eqref{E:dist-est} is likely sub-optimal. It is not clear whether the power $k^{7/2}$ can be improved. \end{remark} For $k\geq 1$ fixed independent of $n,N$, Theorem \ref{T:normal} shows that the top $k$ Lyapunov exponents of $X_{N,n}$ are close to independent Gaussian as soon as $N\gg n\log^2(n)\log^2(N/n).$ This is a significant refinement of the result in \cite{carmona1982exponential} (see also Theorem 5.4 in \cite{bougerol1985concentration}), which states that when $n$ is fixed $\lambda_1$ is asymptotically normal. It also refines the recent result of Reddy \cite[Theorem 11]{reddy2019equality}, which holds only for fixed finite $n$ and does not give estimates at finite $N$. The advantage of Theorem \ref{T:normal} is that it treats simultaneously any number of Lyapunov exponents and gives a rate of convergence. For example, taking $k=n,$ we find that if $N\gg n^{9/2}\log^2(n)\log^2(N/n)$, then \textit{all} Lyapunov exponents of $X_{N,n}$ are approximately independent Gaussians. However, results in articles such as \cite{carmona1982exponential} are for matrix products $A_N\cdots A_1$ in which the entries of $A_i$ have mean zero, variance $1/n$ and satisfy some mild regularity assumptions, whereas our results hold only for the Gaussian case. We conjecture that Theorem \ref{T:normal} holds in this more general setting as well but leave this to future work. \section{Prior Work and Intuitions} The purpose of this section is to give an exposition of prior work and provide several intuitions for thinking about the matrix products $X_{N,n}$, especially about the differences between the near-ergodic $N\gg n$ and the near-free $n\gg N$ regimes. We do this by first giving in \S \ref{S:intuition-dynamics} a basic intuition from dynamical systems, which suggests that one can think of $N$ as a time variable and $n$ as a system size. This intuition dovetails with the multiplicative ergodic theorem. We proceed in \S \ref{S:intution-IPS} to explain an exact correspondence derived in \cite{akemann2012universal,akemann2014universal,akemann2014universalb,akemann2019integrable} at a physical level of rigor in which $n/N$ plays the role of a time parameter for the evolution of the $n$ singular values of $X_{N,n}$. This helps to explain why even simple linear statistics behave differently depending on the relative size of $n,N$. \subsection{$X_{N,n}$ at Fixed $n$ as a Dynamical System}\label{S:intuition-dynamics} One way to intuitively think of $X_{N,n}=A_N\cdots A_1$ is as defining the time $0$ to time $N$ map for a dynamical system in which the time one dynamics are very chaotic and are modelled as multiplication by an iid random matrix. In this analogy, $N$ takes on the role of a time parameter, whereas $n$ denotes the system size. Since large systems take longer to come to equilibrium, we should expect that $N$ and $n$ are ``in tension.'' If we fix $n$ and let $N$ tend to infinity, then the size of the long time image $\norm{X_{N,n}u}$ of an unit length input $u\in \mathbb R^n$ satisfies a pointwise ergodic theorem: \begin{equation}\label{E:LLN} \lim_{N\rightarrow \infty}\frac{1}{N}\log \norm{X_{N,n}u} = E, \end{equation} where $E$ is a constant (independent of $u$) depending on the measure $\mu$ according to which the entries of the matrices $A_i$ making up the matrix product $X_{N,n}$ are distributed. This can be proved in a variety of ways (e.g. Corollary 3.2 in \cite{cohen1984stability}). In fact, much more is true. It was shown by Kesten-Furstenberg in \cite{furstenberg1960products}, that this statement tolerates taking a supremum over $u$: \[\lim_{N\rightarrow \infty} \lambda_1(X_{N,n}) = E\qquad \text{almost surely}.\] Later, in his seminal work \cite{oseledets1968multiplicative} Oseledets proved the multiplicative ergodic theorem. In the context of iid products of $N$ matrices of size $n\times n$, it says that under some mild conditions on $\mu$ if $n$ is fixed, then the full list of Lyapunov exponents $\lambda_1(X_{N,n}),\ldots, \lambda_n(X_{N,n})$ converges almost surely to a deterministic limit. We refer the reader to \cite{filip2019notes} for a review of the vast literature on this subject and to \cite{bougerol1985concentration} for an exposition specifically about matrix products. Determining the values of the limiting Lyapunov exponents in the multiplicative ergodic theorem is in general quite difficult and has applications to Anderson localization for random Schr\"odinger operators \cite{bougerol1985concentration,damanik2011short}. Moreover, the work of LePage \cite{le1982theoremes} as well as subsequent analysis \cite{carmona1982exponential,bougerol1985concentration} showed that the top Lyapunov exponent of matrix products such as $X_{N,n}$ (not necessarily Gaussian) is asymptotically normal in the sense that there exist $a_{n},b_{N,n}\in \mathbb R$ so that \[b_{N,n}\lr{\lambda_1(X_{N,n})- a_{n}}\quad \stackrel{d}{\longrightarrow}\quad \mathcal N(0,1),\] where the $d$ indicates that the convergence is in the sense of distribution. As far as we are aware, all known mathematical proofs of asymptotic normality results hold only for finite fixed $n$, for the top Lyapunov exponent $\lambda_1$ and do not include quantitative rates of convergence. For the real Gaussian case we study, our Theorem \ref{T:normal} overcomes these deficiencies. However, at the physical level of rigor, we refer the reader to the excellent articles \cite{akemann2012universal,akemann2014universal,akemann2014universalb,akemann2015recent, akemann2019integrable} that derive in the case of complex Gaussian matrix products asymptotic normality and much more for the top Lyapunov exponents (cf \S \ref{S:intution-IPS}). While the preceding discussion concerned matrix products with any entry distribution $\mu$ with mean $0$ and variance $1/n$, the Gaussian $\mu= \mathcal N(0,1/n)$ considered in this article leads to some significant simplifications. For instance, Newman \cite{newman1986distribution} computed the exact expression, which can be written in terms of the digamma function, for the limiting Lyapunov exponents. Similarly, \eqref{E:LLN} is a simple fact in this case since $\frac{1}{N}\log \norm{X_{N,n}u}$ turns out to be sum of iid random variables (see Lemma \ref{L:pointwise-iid}). These simplifications stem from the fact that the distribution of each matrix $A_i$ is left and right-invariant under multiplication by an orthogonal matrix. \subsection{$n/N$ as a Time Parameter in an Interacting Particle System}\label{S:intution-IPS} In the regime where $n/N$ is bounded away from $0$ and $\infty$ as $n,N\rightarrow \infty$, even the behavior of an innocuous seeming log-linear statistic depends very much on the ratio of $n$ and $N$. Informally, \begin{equation}\label{E:linear-stats-intro} \log \norm{X_{N,n}u}~\approx~ \mathcal N\lr{-\frac{N}{4n},\frac{N}{4n}} + O\lr{\frac{N}{n^2}}, \qquad u\in \mathbb R^n,\,\, \norm{u}=1, \end{equation} where $\mathcal N(\mu, \Sigma)$ denotes a Gaussian with mean $\mu$ and covariance $\Sigma.$ As mentioned above, in the Gaussian case we consider in this article, this approximate normality is easy to see since $\log \norm{X_{N,n}u}$ is a sum of iid variables (see Lemma \ref{L:pointwise-iid}). Precise versions of \eqref{E:linear-stats-intro} also hold true when the matrices $A_i$ in the definition \eqref{E:X-def} of $X_{N,n}$ have symmetric but non-Gaussian entries (see Theorem 1 in \cite{hanin2019products}). It is interesting to compare the almost sure convergence to a constant in \eqref{E:LLN} (note that $1/N$ normalization) with the asymptotic normality in \eqref{E:linear-stats-intro}. The relation \eqref{E:linear-stats-intro} already suggests that $t=n/N$ is an important parameter for interpolating between the ergodic regime, defined by $t=0$ and the asymptotically free regime, in which $t=\infty$. A number of remarkable articles \cite{akemann2012universal,akemann2014universal,akemann2014universalb} and especially \cite{akemann2019integrable} establish a correspondence between $t$ and the time parameter in the stochastic evolution of an interacting particle system. {This correspondence between singular values for products of complex Ginibre matrices and DBM appears to be initially due to Maurice Duits.} The particles in question are the limiting Lyapunov exponents $\lambda_i$ of $X_{N,n}$. When $t=0$, they are approximately uniformly spaced (see Theorem \ref{T:normal}) and are interpreted as an initial condition for Dyson Brownian motion (DBM) \begin{equation}\label{E:DBM} d\lambda_i = dB_i + \sum_{j\neq i} \frac{dt}{\lambda_i-\lambda_j},\qquad i=1,\ldots,n \end{equation} the dynamics induced on the spectrum of a matrix by allowing each entry to evolve for time $t$ under and independent Brownian motion \cite{dyson1962statistical}. The surprising observation is that, at least in the bulk of the spectrum (i.e. $\lambda_k$ with $k$ proportional to $n$) the joint distribution of the Lyapunov exponents of $X_{N,n}$ satisfies \eqref{E:DBM} at time $t$ with an equally spaced initial condition in the limit when $n/N=t$ and $n,N\rightarrow \infty.$ The idea of the derivations in \cite{akemann2012universal,akemann2014universal,akemann2014universalb, akemann2019integrable} is to use that when $X_{N,n}$ is a product of complex Ginibre matrices, the joint distribution of all of its singular values, at any finite $n,N$, is given by a determinental point process. One may then study the scaling limit of the corresponding determinental kernel at any fixed $t=n/N$. This kernel coincides with the solution to DBM from equally spaced initial conditions, which is also determinental \cite{johansson2004determinantal}. A rigorous analysis of the determinental kernel for the joint distribution of singular values for products of complex Gaussian matrices was undertaken in a variety of articles \cite{forrester2013lyapunov, forrester2014eigenvalue, forrester2016singular, liu2016bulk,liu2018lyapunov}. In particular, \cite{liu2016bulk} shows that when $N$ is arbitrary but fixed and $n\rightarrow \infty$, the determinental kernel for singular values in products of $N$ iid complex Gaussian matrices of size $n\times n$ converges to the familiar sine and Airy kernels that arise in the local spectral statistics of large GUE matrices in the bulk and edge, respectively. This agrees with the prediction from \cite{akemann2019integrable}. Indeed, in this regime, the time parameter $t=n/N$ is infinite and the limiting distribution of DBM is that of the eigenvalues for a large GUE matrix. Moreover, \cite{liu2018lyapunov} rigorously obtained an expression for the limiting determinental kernel when $t=n/N$ is arbitrary in the context of products of complex Ginibre matrices.{We refer the reader also to the subsequent article of Liu-Wang \cite{liu2019phase} that performs a similar analysis for the \textit{eigenvalues} in the same setting. } Also in the regime where $n/N$ is fixed while $n,N\rightarrow \infty$, we refer the reader to Gorin-Sun \cite{gorin2018gaussian}. This article shows that the fluctuations of the singular values of $X_{N,n}$ around the triangle law always converge to a Gaussian field. We also refer the reader to \cite{ahn2019fluctuations}, which obtains a CLT for linear statistics of top singular values when $n/N$ is fixed and finite. \section{Idea of Proof: Reduction to Small Ball Estimates} Before turning to the formal proofs of Theorems \ref{T:global} and \ref{T:normal}, we give a brief overview of our approach, which begins with the following representation (cf e.g. \cite{newman1986distribution,isopi1992triangle}) for sums of Lyapunov exponents from Lemma \ref{lem-Grass-3} (see \S \ref{S:wedge-background}): \begin{equation}\label{E:wedge-rep-informal} \lambda_1+\cdots + \lambda_k = \sup_{\substack{\Theta \in \mathrm{Fr}_{n,k}}} \frac{1}{N}\log \norm{X_{N,n}(\Theta)}. \end{equation} In the previous line, we've denoted by $\mathrm{Fr}_{n,k}$ the collection of all orthonormal $k-$frames in $\mathbb R^n$ (i.e. collections of $k$ orthonormal vectors $v_1,\ldots, v_k$). We have also set \[X_{N,n}(\Theta)=X_{N,n}\theta_1\wedge\cdots\wedge X_{N,n}\theta_k,\qquad \Theta = (\theta_1,\ldots,\theta_k)\in \mathrm{Fr}_{n,k},\] where we recall that for $a,b\in \mathbb R^n$ $a_1\wedge \cdots \wedge a_k$ is the anti-symmetrization of $a_1\otimes \cdots \otimes a_k$. We refer the reader to \S \ref{S:wedge-background} for more background and to the start of \S 7 in \cite{spivak1970comprehensive} and to \S 2.6.1 in \cite{tao2010epsilon} for more on wedge products. As pointed out in \cite{isopi1992triangle}, information about the sums $\lambda_1+\cdots+\lambda_k$ can easily be translated into the information about their cumulative distribution function, ultimately resulting in Theorem \ref{T:global}. Similarly, the vector of the top $k$ Lyapunov exponents considered in Theorem \ref{T:normal} can be obtained by an affine transformation of the vector of partial sums $\lambda_1,\lambda_1+\lambda_2,\ldots,\lambda_1+\cdots + \lambda_k$. Thus, the focus of our proofs is to obtain precise concentration estimates for the expression on the right hand side of \eqref{E:wedge-rep-informal}. An important idea for analyzing \eqref{E:wedge-rep-informal}, which goes back to the work of Furstenburg-Kesten \cite{furstenberg1960products} is that when $N$ is large, one can almost drop the supremum: \begin{equation}\label{E:pointwise-informal} \lim_{N\rightarrow \infty}\abs{\frac{1}{N}\log\norm{X_{N,n}(\Theta)}-\sup_{\substack{\Theta' \in \mathrm{Fr}_{n,k}}} \frac{1}{N}\log\norm{X_{N,n}(\Theta')} } = 0 \end{equation} for any fixed $\Theta\in \mathrm{Fr}_{n,k}$. As explained below this is plausible since the ratio $s_k(X_{N,n})/s_{k+1}(X_{N,n})$ of the $k^{th}$ and $(k+1)^{st}$ singular values grows exponentially with $N$, causing the wedge product $X_{N,n}(\Theta)$ to align almost entirely with the wedge product of the top $k$ singular vectors of $X_{N,n}$ for almost every $\Theta$. The ``pointwise'' quantity $\frac{1}{N}\log \norm{X_{N,n}(\Theta)}$ is a sum of iid random variables (Lemma \ref{L:pointwise-iid}) and can be analyzed using a result of \L ata\l a \cite{latala1997estimation} (see Theorem \ref{Latala}). It then remains to obtain quantitative versions of \eqref{E:wedge-rep-informal} valid for large but finite $N,n$. One possible approach is via energy-entropy estimates using $\varepsilon$-nets on $\mathrm{Fr}_{n,k}$. However, while this gives some results, this approach is suboptimal for large $N$. The reason that $\varepsilon$-nets fail is that, due to the $N^{-1}$ normalization, \[N~~ \text{large} \quad \Rightarrow\quad \Var_\Theta\left[\frac{1}{N}\log \norm{X_{N,n}(\Theta)}\right]\ll \Var_{X_{N,n}}\left[\frac{1}{N}\log \norm{X_{N,n}(\Theta)}\right]\] by which we mean that the variance of $\frac{1}{N}\log \norm{X_{N,n}(\Theta)}$ over $\Theta\in \mathrm{Fr}_{n,k}$ for a typical realization of $X_{N,n}$ is much smaller than its variance over the randomness in $X_{N,n}$ for any fixed $\Theta$, causing the optimal net to have constant cardinality. The main technical novelty of our proofs is that we quantify \eqref{E:pointwise-informal} not through net arguments but rather via small ball probabilities for volumes of random projections, which are already known (cf Proposition \ref{Gaussian-matrix-det-est}). The key result is the following: \begin{proposition}\label{P:small-ball-bounds} For any $\varepsilon\in (0,1)$ and any $\Theta \in \mathrm{Fr}_{n,k}$ we have \[\mathbb P\lr{\abs{\frac{1}{N}\log \norm{X_{N,n}\Theta} -\sup_{\Theta' \in \mathrm{Fr}_{n,k}}\frac{1}{N}\log \norm{X_{N,n}\Theta'}}\geq \frac{1}{N}\log\lr{\frac{1}{\varepsilon}}}\leq \mathbb P\lr{\norm{P_F(\Theta)}\leq \varepsilon},\] where $F$ is a Haar distributed $k-$dimensional subspace of $\mathbb R^n$ and \[P_F(\Theta)=P_F\theta_1\wedge \cdots\wedge P_F\theta_k\] with $P_F$ denoting the orthogonal onto $F.$ \end{proposition} For the proof of Proposition \ref{P:small-ball-bounds} see Lemma \ref{L:small-ball-bounds}. The appearance of small ball probabilities is natural, although perhaps somewhat unexpected. Let us briefly describe why in the simplest case of $k=1$. Denote by $v^{(i)}$ the right eigenvector of $X_{N,n}$ corresponding to the singular value $s_{i}(X_{N,n})$. For any $\theta \in S^{n-1}$, we may write \[\norm{X_{N,n}\theta}^2 = \sum_{i=1}^n \abs{\inprod{X_{N,n}\theta}{v^{(i)}}}^2.\] When $N\gg n$, the matrix $X_{N,n}$ is highly degenerate in the sense that there exists a universal constant $C>0$ so that \[\frac{s_{1}(X_{N,n})}{s_{2}(X_{N,n})}\geq e^{CN/n}.\] This is easy to see intuitively since in this regime $\lambda_1-\lambda_2 \approx \frac{1}{n}$ (cf Theorem \ref{T:normal}). Hence, \begin{equation}\label{E:degen} \norm{X_{N,n}\theta}^2 \approx \abs{\inprod{X_{N,n}\theta}{v^{(1)}}}^2 \end{equation} unless $\abs{\inprod{\theta}{v^{(1)}}}$ is unusually small. In fact, for all $\theta$ \[0\geq \frac{1}{N}\log\norm{X_{N,n}\theta}-\lambda_1 = \frac{1}{2N}\log \lr{\frac{\norm{X_{N,n}\theta}^2}{s_1^2(X_{N,n})}}\geq \frac{1}{N}\log\abs{\inprod{\theta}{v^{(1)}}}.\] This lower bound is essentially sharp by \eqref{E:degen} unless $\theta$ has small overlap with $v^{(1)}$, an event whose probability is controlled precisely by a small ball estimate. \section{Acknowledgements} We are grateful to Vadim Gorin, Maurice Duits, and Gernot Akemann for pointing us to a number of interesting references. We would also like to thank a referee for a very careful and helpful reading of an earlier draft that pointed out a number of inaccuracies and ultimately lead to a substantially improved exposition. \section{Organization of the Rest of the Article} The rest of this article is structured as follows. First, in \S \ref{S:wedge-background} we collect some well-known results on the relation between the exterior algebra of $\mathbb R^n$ and the singular values of any linear map $A:\mathbb R^n\rightarrow \mathbb R^n.$ We also record several elementary observations (Lemmas \ref{L:polar} and \ref{L:haar-flags}) about polar decompositions and Haar measures on orthonormal frames and their flags. We will use this formalism throughout our proofs. Next, in \S \ref{S:iid-background} recalls two kinds of results. The first, Theorem \ref{Latala}, is a result of \L ata\l a \cite{latala1997estimation} that gives precise information on moments (and hence tail behavior) for sums of independent centered random variables. The second is a set of results related to the multivariate central limit theorem (Theorem \ref{MultiCLT-theo}) and the Gaussian content of boundaries of convex sets (Theorem \ref{T:Nazarov-Ball}). The latter allows us to prove Proposition \ref{Dist-lem}, a stability result for the Kolmogorov-Smirnov-type distance function $d$ used in the statement of Theorem \ref{T:normal}. Section \ref{S:roadmap} follows, containing a brief road map to the proofs of Theorems \ref{T:sums}, \ref{T:global} and \ref{T:normal}. Then, \S \ref{S:small-ball-use} is devoted to explaining how to use small ball estimates on volumes of random projections to formalize the ergodicity \eqref{E:pointwise-informal}. Further, the results in \S \ref{S:concentration} are used in all our proofs. The main result there is Proposition \ref{P:pointwise}, which together with Proposition \ref{P:pointwise-small-ball} and Lemma \ref{L:pointwise-iid} explains the appearance of chi-squared random variables in the statement of Theorem \ref{T:normal}. The proof of Proposition \ref{P:moments} is the most technical part of our arguments. Next, we complete the proof of Theorem \ref{T:sums} in \S \ref{S:sums-pf}. We then use Theorem \ref{T:sums} to complete in \S \ref{S:global-pf} and \S \ref{S:normal-outline} the proofs of Theorems \ref{T:global} and \ref{T:normal}, respectively. \section{Singular Values via Wedge Products}\label{S:wedge-background} In this section, we recall some background on wedge products and refer the reader to the start of \S 7 in \cite{spivak1970comprehensive} and to \S 2.6.1 in \cite{tao2010epsilon} for more details. The usual $\ell_2$-structure on $\mathbb R^n$ gives rise in a functorial way to an $\ell_2$ structure on the exterior powers $\Lambda^k \mathbb R^n$. If $ x_{1}, \cdots , x_{k} $ are in $ \mathbb R^{n}$ (e.g. are a frame for an element of $G_{n,k}$) we denote the resulting norm by \[\norm{x_1\wedge\cdots\wedge x_k}.\] If we denote by $ X^{\ast}$ the $ n\times k$ matrix $( x_{1}, \cdots , x_{k})$, the Gram identity reads \begin{equation} \label{wedge-def} \| x_{1} \wedge \cdots \wedge x_{k} \| = \sqrt{{\rm det} ( X X^{\ast})} = \vol_k\lr{\mathcal P \lr{x_1,\ldots, x_k}}, \end{equation} where $\mathcal P \lr{x_1,\ldots, x_k}$ is the parallelopiped spanned by $x_1,\ldots, x_k$. The following Lemma gives a well-known characterization of products of singular values in terms of norms of wedge products, which we will use repeatedly in the proofs of our results. \begin{lemma} \label{lem-Grass-3} \noindent Let $ A$ be an $ n\times n$ real matrix with singular values $ s_{1}(A)\geq s_2(A)\geq \cdots \geq s_n(A)$. If $ \theta_{1}, \cdots , \theta_{k}$ are unit vectors in $ \mathbb R^{n}$, then \begin{equation} \label{wedge-4} \| A\theta_{1}\wedge \cdots \wedge A\theta_{k} \| \leq \sup_{\theta_1',\ldots, \theta_k' \in S^{n-1}}\| A\theta_{1}'\wedge \cdots \wedge A\theta_{k}' \| =\prod_{i=1}^{k} s_{i} (A) , \end{equation} with equality if and only if $\theta_{i}$ are orthonormal and $ {\rm span} \{ \theta_{i} , i\leq k \} $ is the subspace spanned by the eigenvectors of $AA^{\ast}$ that correspond to the largest singular values of $A.$ \end{lemma} \begin{proof} The inequality on the left is clear. To derive the equality, note that, for any $\theta_1,',\ldots,\theta_k'\in S^{n-1}$, \[ \theta_1'\wedge\cdots\wedge \theta_k' = \theta_1''\wedge \cdots \wedge \theta_k'', \] where \[ \theta_j'' = \Pi_{\leq j-1}^{\perp} \theta_j',\qquad \theta_1'' = \theta_1 \] and $\Pi_{\leq j-1}^{\perp}$ is the projection onto the orthogonal complement of the span of $\theta_1',\ldots, \theta_{j-1}'.$ This follows immediately from the fact that $a_1\wedge \cdots \wedge a_k$ is zero if $\set{a_j}$ is linearly dependent. Thus, the supremum in \eqref{wedge-4} can be taken over $\theta_1',\ldots, \theta_k'$ that are orthogonal. Over such collections, the supremum is obtained by letting $\theta_1',\ldots, \theta_k'$ be any permutation of the $k$ right singular vectors of $A$. \end{proof} \noindent Next, we record in Lemma \ref{lem-Grass-2} some basic properties of this norm of wedge products that we will use. \begin{lemma} \label{lem-Grass-2} Let $x, x_{1}, \cdots , x_{k}$ be vectors in $ \mathbb R^{n}$. Then we have the following basic properties: \begin{enumerate} \item{ \bf Homogeneity}: If $\lambda_i>0$ \begin{equation} \label{wedge-3} \| \lambda x_{1} \wedge \cdots \wedge \lambda_{k} x_{k} \| = \left( \prod_{i=1}^{n} \lambda_{i} \right) \| x_{1} \wedge \cdots \wedge x_{k} \| \end{equation} \item{\bf Projection formula}: Let $P_{V_{i}^{\perp}}$ be the orthogonal projection onto the orthogonal complement of $ V_{i} := {\rm span} \{ x_{1}, \cdots , x_{k} \} , V_{0} = \{0\} , 1\leq k \leq n-1$. We have \begin{equation} \label{wedge-1.5} \| x_{1} \wedge \cdots \wedge x_{k} \| = \prod_{i=1}^{k} \| P_{V_{i-1}^{\perp}} x_{i} \|_{2}. \end{equation} \item{\bf Pythagorean Theorem}: Let $e_1,\ldots, e_n$ be any orthonormal basis of $\mathbb R^n$, and define for each multi-index $I=(i_1,\ldots, i_k)$ \[e_I := e_{i_1}\wedge \cdots \wedge e_{i_k}.\] Then, \begin{equation}\label{E:pyth} \norm{x_1\wedge\cdots\wedge x_k}^2 = \sum_{\substack{I=\lr{i_1,\ldots, i_k}\\ 1\leq i_1<\cdots<i_k\leq n}} \inprod{x_1\wedge \cdots \wedge x_k}{e_I}^2. \end{equation} \item {\bf Generalized Gram Identity}: Let $\Theta =\lr{\theta_1,\ldots,\theta_k}$ be an orthonormal system of $k$ vectors in $\mathbb R^n$ and write $P_\Theta$ for the orthogonal projection onto the span of the $\theta_i$. Consider arbitrary linearly independent vectors $v_1,\ldots, v_k$ in $\mathbb R^n$, and denote by $V$ the $n\times k$ matrix whose columns are $v_i$. Then \begin{equation}\label{E:generalized-gram} \inprod{v_1\wedge\cdots \wedge v_k}{\theta_1\wedge\cdots\wedge \theta_k}^2= \det(P_{\Theta}VV^*P_\Theta) = \norm{P_{\Theta}v_1\wedge\cdots\wedge P_{\Theta}v_k}^2. \end{equation} \end{enumerate} \end{lemma} \begin{proof} \noindent Homogeneity is immediate from the multi-linearity of the determinant \eqref{wedge-def}. The projection formula \eqref{wedge-1.5} follows from \eqref{wedge-def} and the fact that $\sqrt{\det(XX^*)}$ is the volume of the parallelopiped spanned by $x_1,\ldots, x_k.$. Next, the Pythagorean theorem follows from the fact that in the definition of the $\ell_2$ structure on $\Lambda^k\mathbb R^n,$ \[\set{e_I,\, I = \lr{i_1,\ldots,i_k},\, 1\leq i_1<\cdots< i_k\leq n}\] is an orthonormal basis. Finally, to show \eqref{E:generalized-gram}, assume first that \[\theta_j=e_j,\qquad j=1,\ldots,k\] are the first $k$ standard unit vectors. Then the right equality follows immediately from the Gram identity \eqref{wedge-def}. To see the left equality, write \[v_{j}=\sum_{i=1}^n v_{j,i}e_i.\] We have \begin{align*} v_1\wedge\cdots \wedge v_k &= \sum_{i_1,\ldots,i_k} \prod_{j=1}^k v_{j,i_j} e_{i_1}\wedge \cdots \wedge e_{i_k}= \sum_{\substack{I=\lr{i_1,\ldots,i_k}\\ 1\leq i_1<\cdots<i_k\leq n}} \lr{\sum_{\sigma \in S_k} (-1)^{\mathrm{sgn}(\sigma)} \prod_{j=1}^k v_{j,\sigma(j)}} e_I. \end{align*} Hence, writing $V_k$ for the matrix obtained from $V$ by keeping only the first $k$ rows, we find from the Pythagorean theorem \eqref{E:pyth}, \[\inprod{v_1\wedge\cdots \wedge v_k}{\theta_1\wedge\cdots\wedge \theta_k}^2 = \det(V_k)^2=\det(V_kV_k^*).\] The case of general $\theta_i$ follows by considering any orthogonal matrix $U$ satisfying \[\theta_i = U e_i,\qquad i = 1,\ldots,k.\] Then \[\inprod{v_1\wedge\cdots \wedge v_k}{\theta_1\wedge\cdots\wedge \theta_k}^2 = \det((U^T V)_k(U^TV)_k^*)=\det(P_\Theta V (P_\Theta V)^*).\] \end{proof} \subsection{Haar Measure on Frames} It well-know that if $\xi$ is a standard Gaussian on $\mathbb R^n$, then $\widehat{\xi}=\xi/\norm{\xi}$ is independent of $\norm{\xi}$ and that $\widehat{\xi}$ is uniform on the unit sphere. We will need natural generalizations of these facts to orthonormal frames, Lemma \ref{L:polar} and \ref{L:haar-flags}. \begin{lemma}[Polar Decomposition for Haar Measure on Flags]\label{L:polar} Fix integers $n\geq k\geq 1.$ Let $\xi_1,\ldots, \xi_k\in \mathbb R^n$ be independent standard Gaussian random vectors. The following collections of random variables are independent: \begin{equation}\label{E:norms-angles} \set{\norm{\xi_1},\norm{\xi_1\wedge \xi_2},\ldots, \norm{\xi_1\wedge \cdots \wedge \xi_k}},\qquad \set{\frac{\xi_1}{\norm{\xi_1}},\ldots, \frac{\xi_1\wedge \cdots \wedge \xi_k}{\norm{\xi_1\wedge \cdots \wedge \xi_k}}}. \end{equation} Moreover, denote by $P_{\leq i}$ the orthogonal projection onto the complement of the span of $\xi_1,\ldots, \xi_i.$ Then, the random variables terms $\norm{P_{\leq i-1} \xi_i}$ are joiltly independent. \end{lemma} \begin{proof} We begin by recalling a fact from elementary probability. Namely, let $X,Y,Z$ be any random variables defined on the same probability space. Then, \begin{equation}\label{E:fact} X\perp Y\qquad \text{and}\qquad X\perp Z | Y\qquad \Rightarrow\qquad X \perp (Y,Z). \end{equation} In words, if $X$ is independent of $Y$ and $Z$ is independent of $X$ given $Y$, then, $(Y,Z)$ is independent of $X.$ We proceed by induction on $k.$ The case $k=1$ follows from the fact that the radial and angular parts of a standard Gaussian are independent. Suppose now $k\geq 2$ and we have proved the statement for $k-1.$ For any $\xi \in \mathbb R^k\backslash \set{0}$, let us write $P_{\xi^\perp}$ for the orthogonal projection onto the orthogonal complement to the line spanned by $\xi.$ Define for $\ell=2,\ldots,k$ \[\xi_\ell':=P_{\xi_1^\perp}\xi_\ell.\] Note that \[\norm{\xi_1\wedge\cdots\wedge \xi_{\ell}}=\norm{\xi_1}\norm{\xi_2'\wedge\cdots \wedge\xi_\ell'}\] and that \[\xi_1\wedge\cdots\wedge \xi_{\ell}=\xi_1\wedge \xi_2' \wedge \cdots \wedge \xi_\ell'.\] With this notation, it is enough to show that the collections \[\set{\norm{\xi_1},\norm{\xi_2'},\ldots,\norm{\xi_2'\wedge\cdots\wedge \xi_k'}},\qquad \set{\frac{\xi_1}{\norm{\xi_1}}, \frac{\xi_2'}{\norm{\xi_2'}},\ldots, \frac{\xi_2'\wedge\cdots\wedge \xi_k'}{\norm{\xi_2'\wedge\cdots\wedge \xi_k'}}}\] are independent since if $X,Y$ are independent, then so are $f(X),g(Y)$ for any measurable functions $f,g$. To see this, first observe that, aside from $\norm{\xi_1}$, all other random variables in all both collections are measurable functions of $\xi_1/\norm{\xi_1},\xi_2,\ldots, \xi_k$. Hence, $\norm{\xi_1}$ is independent of all other variables in both collections. It therefore remains to check only that \[\mathcal A = \set{\norm{\xi_2'},\ldots,\norm{\xi_2'\wedge\cdots\wedge \xi_k'}}, \qquad \mathcal B = \set{\frac{\xi_1}{\norm{\xi_1}},\frac{\xi_2'}{\norm{\xi_2'}},\ldots, \frac{\xi_2'\wedge\cdots\wedge \xi_k'}{\norm{\xi_2'\wedge\cdots\wedge \xi_k'}}}\] are independent. Observe that, given, $\xi_1/\norm{\xi_1}$, the random variables $\xi_\ell',\,\ell=2,\ldots,k$ are projections of iid standard Gaussians onto a fixed dimension $k$ subspace and hence are themselves iid standard Gaussians. By the inductive hypothesis, conditional on $\xi_1/\norm{\xi_1}$, the collection $\mathcal A$ is independent of the collection $\mathcal B\backslash \set{\xi_1/\norm{\xi_1}}$. Moreover, the random variables in $\mathcal A$ are independent of $\xi_1$. Invoking \eqref{E:fact} therefore completes the proof of the statement that the two collections in \eqref{E:norms-angles} are independent. To finish the proof of this Lemma, let us check that the terms \begin{equation}\label{E:norms} \norm{\xi_1},\norm{P_{\leq 1}\xi_2},\ldots, \norm{P_{\leq k-1} \xi_k} \end{equation} are independent by induction on $k.$ When $k=1$ the claim is trivial. For the inductive step, use the projection formula \eqref{wedge-1.5} to we write \[\norm{P_{\leq k-1} \xi_k}=\norm{\frac{\xi_1\wedge\cdots\wedge \xi_{k-1}}{\norm{\xi_1\wedge\cdots\wedge \xi_{k-1} }}\wedge \xi_k}.\] By the first part of this Lemma, we know that \[\frac{\xi_1\wedge\cdots\wedge \xi_{k-1}}{\norm{\xi_1\wedge\cdots\wedge \xi_{k-1} }},\qquad \set{\norm{\xi_1\wedge \cdots \wedge \xi_j},\qquad j\leq k-1}\] are independent. Hence, since $\xi_k$ is independent of $\xi_1,\ldots, \xi_{k-1}$, we conclude that the terms in \eqref{E:norms} are indepenedent as well. \end{proof} \noindent We will also use this following result: \begin{lemma}[Haar Measure on Flags]\label{L:haar-flags} Fix integers $n\geq k\geq 1.$ Let $\xi_1,\ldots, \xi_k\in \mathbb R^n$ be independent standard Gaussian random vectors. For each $i=1,\ldots, k$ define \[\xi_i' = \begin{cases}\xi_1,&\quad i=1\\ P_{V_{i-1}} \xi_i,&\quad i>1\end{cases},\] where $V_{i-1}=\mathrm{Span}\set{\xi_j,\, 1\leq j<i }^\perp$ and $P_{V_{i-1}}$ is the orthogonal projection onto $V_{i-1}$. Then $\set{\xi_i'/\norm{\xi_i'},\, i=1,\ldots, k}$ is distributed according to Haar measure on the space of such flags of orthonormal frames. In particular, if $v_1,\ldots, v_k$ is Haar-distributed in the space of $k-$frames in $\mathbb R^n$, then \begin{equation}\label{E:gaussian-frame} v_1\wedge \cdots \wedge v_k \stackrel{d}{=} \frac{\xi_1\wedge \cdots \wedge \xi_k}{\norm{\xi_1\wedge \cdots \wedge \xi_k }} \end{equation} \end{lemma} \begin{proof} The random variable $\lr{\xi_i',\, i=1,\ldots,k}$ clearly takes values in the set of $k$-frames in $\mathbb R^n.$ Moreover, it is invariant under the action of the orthogonal group on such frames and hence must be distributed according to the Haar measure. Indeed, since the angular part of a standard Gaussian is uniform on the sphere, $\xi_1'/\norm{\xi_i'}$ is uniform on $S^{n-1}$. Similarly, $\xi_2'$ is a standard Gaussian in the orthogonal complement ${\xi_1'}^\perp$ to $\xi_i'.$ Hence, $\xi_2'/\norm{\xi_2'}$ is uniform on the unit sphere in ${\xi_1'}^\perp$ and so on. Finally, to derive \eqref{E:gaussian-frame}, note that \[ \xi_1\wedge \cdots \wedge \xi_k = \xi_{1}'\wedge \cdots \xi_k' \] since the wedge product of any linearly dependent set of vectors vanishes. Combining this with the projection formula \eqref{wedge-1.5}, we conclude that \[ \frac{\xi_1\wedge \cdots \wedge \xi_k }{\norm{\xi_1\wedge \cdots \wedge \xi_k }} = \frac{\xi_1'}{\norm{\xi_1'}}\wedge\cdots \wedge \frac{\xi_k'}{\norm{\xi_k'}} \] and \eqref{E:gaussian-frame} follows from the first part of this Lemma. \end{proof} \section{Background on Sums of Independent Random Variables}\label{S:iid-background} \subsection{A Result of \L ata\l a: Precise Behavior for Moments of Sums} \noindent In the proof of our pointwise esimate Proposition \ref{P:pointwise}, we will use the following result of R. \L ata\l a \cite[Thm. 2, Cor. 2, Rmk. 2]{latala1997estimation}: \begin{theorem} \label{Latala} \noindent Let $ X_{1}, \cdots , X_{N}$ be mean zero, independent r.v. and $ p\geq 1$. Then \begin{equation} \label{Latala-1} \left( \mathbb E\left[ \bigg| \sum_{j=1}^{N} X_{j} \bigg|^{p}\right] \right)^{\frac{1}{p}} \simeq \inf\left\{ t>0 : \sum_{j=1}^{N} \log \left[ \mathbb E | 1+ \frac{X_{j}}{ t}|^{p} \right] \leq p \right\}, \end{equation} where $a\simeq b$ means there exist universal constants $c_1,c_2$ so that $c_1a \leq b \leq ac_2.$ Moreover, if $ X_{i}$ are also identically distributed then \begin{equation} \label{Latala-2} \left( \mathbb E \left[\bigg| \sum_{j=1}^{N} X_{j} \bigg|^{p}\right] \right)^{\frac{1}{p}} \simeq \sup_{ \max\{ 2, \frac{ p}{ N} \} \leq s \leq p } \frac{p}{s} \left( \frac{ N}{p}\right)^{\frac{1}{s}} \| X_{i} \|_{s} . \end{equation} \end{theorem} \subsection{Small Ball Estimates for Sums of iid Random Variables with Bounded Density} We will have occasion to use the following standard result (e.g. Lemma 2.2 in \cite{rudelson2008littlewood}). \begin{theorem}\label{T:iid-small-ball} Let $\zeta_k,\, k=1,\ldots, n$ be iid non-negative random variables and suppose there exists $K$ such that $\mathbb P\lr{\zeta_k < \varepsilon}\leq K \varepsilon$ for all $\varepsilon\geq 0$, then for all $\varepsilon \geq 0$ \[ \mathbb P\lr{\sum_{k=1}^n \zeta_k^2 < n\varepsilon^2}\leq (CK\varepsilon)^n, \] where $C$ is an absolute constant. \end{theorem} \subsection{Quantitative Multivariate CLT}\label{S:MCLT} One of our goals in this article is to measure the approximate normality for Lyapunov exponents of $X_{N,n}$ when $N\gg n$ (see Theorem \ref{T:normal}). As explained in the introduction, we will measure normality using a distance function that is a natural high-dimensional generalization of the usual Kolmogorov-Smirnov distance: \begin{equation}\label{E:dist-def} d( X, Y ) := \sup_{ C \in {\cal{C}}_{k} } \left| \mathbb P ( X\in C ) -\mathbb P ( Y\in C) \right|, \end{equation} where $ {\cal{C}}_{k}$ is the collection of all convex subsets of $\mathbb R^{k}$. The distance function $d$ has three desirable properties. First, it is affine invariant in sense that if $T$ is any invertible affine transformation and $A$ is any convex set, then $T^{-1}A$ is also convex and hence for any random variables $X,Y$ on the same probability space \begin{align} \label{E:affine-inv}d(TX,TY)&=d(X,Y). \end{align} The second desirable property of $d$ is that it is stable to small $\ell_2$ perturbations, as explained in Proposition \ref{Dist-lem} below. To state this result, we write \[\mathcal N(\mu, \Sigma),\qquad \mu \in \mathbb R^k,\quad \Sigma \in \mathrm{Sym}_k^+\] for a Gaussian with mean $\mu$ and invertible covariance $\Sigma$. \begin{proposition}\label{Dist-lem} There exists $c>0$ with the following property. Suppose that $X,Y$ are $\mathbb R^k$-valued random variables defined on the same probability space. For all $\mu\in \mathbb R^k$, invertible symmetric matrices $\Sigma\in \mathrm{Sym}_k^+$, and $\delta>0$ we have \begin{equation} \label{Dist-lemma-1} d(X+Y, \mathcal N\lr{\mu, \Sigma})\leq 3 d(X,\mathcal N\lr{\mu, \Sigma})+c\delta \sqrt{\norm{\Sigma^{-1}}_{HS}}+2\mathbb P\lr{\norm{Y}_2>\delta}. \end{equation} \end{proposition} We prove Proposition \ref{Dist-lem} in \S \ref{S:dist-lem-pf} below. Before doing so, however, we state the third desirable property of the distance $d$, namely, that it measures convergence in the multivariate CLT. We follow the notation in Bentkus \cite{bentkus2005lyapunov} and define \[S:= S_{N}= X_{1} + \cdots + X_{N},\] where $ X_{1}, \ldots, X_{N}$ are independent random vectors in $ \mathbb R^{k}$ with common mean $ \mathbb E X_{j} = 0$. We set \[ C := {\rm cov} (S) \] to be the covariance matrix of $S$, which we assume is invertible. With the definition \begin{equation} \label{Multi-CLT-0} \beta_{j} := \mathbb E \| C^{-\frac{1}{2}} X_{j} \|_{2}^{3} , \ \beta:= \sum_{j=1}^{N} \beta_{j}, \end{equation} we have the following \cite{bentkus2005lyapunov}: \begin{theorem}[Multivariate CLT with Rate] \label{MultiCLT-theo} \noindent There exists an absolute constant $c>0$ such that \begin{equation} \label{Multi-CLT-1} d( S, C^{\frac{1}{2}}Z ) \leq c k^{\frac{1}{4}} \beta , \end{equation} where $Z\sim \mathcal N(0,\mathrm{Id}_k)$ denotes a standard Gaussian on $\mathbb R^k.$ \end{theorem} \subsection{Proof of Proposition \ref{Dist-lem}}\label{S:dist-lem-pf} We rely on the following result (Theorem \ref{T:Nazarov-Ball}) of Nazarov \cite{nazarov2003maximal}, which was an extension of a Theorem of Ball \cite{ball1993reverse}. \\ \noindent Let $ Z_{k}$ be the standard Gaussian in $ \mathbb R^{k}$. Let $ C$ be a positive definite symmetric matrix and let $ \gamma_{C}$ be the density of the $C^{-\frac{1}{2}}Z_{k}$, i.e. $$ d\gamma_{C} ( y ) := \frac{\sqrt{|\det C|}}{ (2\pi)^{\frac{k}{2}}} e^{ - \frac{ \langle C y, y\rangle }{2}} d y . $$ Let \begin{equation} \label{def-K-e} K_{\epsilon} := \{ x\in \mathbb R^{k} : \exists y \in K , \| x-y\|_{2} <\epsilon\} \ {\rm and } \ K_{-\epsilon} : \{ x\in K := x+ \epsilon B_{2}^{k} \subseteq K\} \end{equation} We define \begin{equation} \label{Naz-1} \Gamma(C ) := \sup_{K\in {\cal{C}}} \left\{ \lim_{\epsilon\rightarrow 0_+ }\frac{ \gamma_{C} \left( K_{\epsilon} \setminus K\right) }{ \epsilon} \right\}. \end{equation} Our proof of Proposition \ref{Dist-lem} crucially relies on the following result of Nazarov. \begin{theorem}[\cite{nazarov2003maximal}]\label{T:Nazarov-Ball} \noindent There exists absolute constants $ c_{1}, c_{2}>0$ such that \begin{equation} \label{Naz-1} c_{1} \sqrt{\| C\|_{HS} } \leq \Gamma(C) \leq c_{2} \sqrt{\| C\|_{HS} }. \end{equation} \end{theorem} \noindent We will need an elementary corollary of \eqref{Naz-1}. \begin{corollary}\label{C:Naz} For any $\varepsilon\in (0,1)$ and $K\in \mathcal D_k$, we have \begin{equation} \label{Naz-2} \left| \mathbb P ( C^{-\frac{1}{2}} Z_{k} \in K_{\epsilon} ) - \mathbb P ( C^{-\frac{1}{2}} Z_{k} \in K_{ -\epsilon} ) \right| \leq c_{2} \epsilon \sqrt{\| C\|_{HS} }.\\ \end{equation} \end{corollary} \begin{proof} Our argument is along the lines of the proof in \cite{bentkus2005lyapunov} of equations (1.3), (1.4) or \cite{bentkus2003dependence} equation (1.2), which obtain similar statements in the special case where $C=\mathrm{Id}$. We will use a standard estimate (see e.g. Lemma 10.5 in \cite{kallenberg2002foundations}) that if $K$ is any convex set in $\mathbb R^k$ and $(\partial K)_\epsilon$ is the $\epsilon$-neighborhood of the boundary of $K$, then \begin{equation}\label{E:Kal} \mathrm{vol}((\partial K)_\epsilon) \leq 2\left[\lr{1+\frac{\epsilon}{r(K)}}^k-1\right]\mathrm{vol}(B_{r(K)}), \end{equation} where $r(K)$ is the radius of the smallest ball containing $K$, $\mathrm{vol}$ is the Euclidean volume, and $B_{r}$ is a ball of radius $r$. The crucial point is that the right hand side is bounded for all $\epsilon\in (0,1)$ by $c\epsilon$ with $c$ depending only on $r(K)$ and the ambient dimension $k$. To derive from this estimate \eqref{Naz-2}, observe that for every $ \epsilon>0$, if $ K\in {\cal{C}}_{k}$ then also $ K_{\epsilon } $ and $K_{-\epsilon}$ are in $ {\cal{C}}_{k}$. Note also that the difference of probabilities on the left-hand side of \eqref{Naz-2} can be bounded above by $ \mathbb P (W\in K_{\epsilon} \setminus K )+ \mathbb P (W\in K\setminus K_{-\epsilon} ) $ where $ W:= C^{-\frac{1}{2}} Z_{k} $. It is therefore enough to check that each of these probabilities is in turn bounded above by the right hand side of \eqref{Naz-2}. To see this for $\mathbb P (W\in K_{\epsilon} \setminus K )$, denote for $K$ convex and $t>0$, \[ \omega_{K}(t) := \mathbb P \left( W \in K_{t} \setminus K\right). \] Since $ K_{t+ \epsilon} \setminus K = ( K_{t+ \epsilon} \setminus K_{t}) \cup (K_{t} \setminus K) $ we have that \begin{equation}\label{E:semi-group} \omega_{K} (t+ \epsilon) - \omega_{K} (t) = \omega_{K_{t} } (\epsilon). \end{equation} This relation, together with \eqref{E:Kal}, implies that $\omega_K(t)$ is absolutely continuous. Indeed, we may write \begin{align*} \omega_{K_t}(\epsilon) = \mathbb P\lr{Z_k\in C^{1/2}K_{t+\epsilon}\backslash C^{1/2}K_t} \leq \frac{\mathrm{vol}\lr{C^{1/2}K_{t+\epsilon}\backslash C^{1/2} K_t}}{(2\pi)^{k/2}} . \end{align*} Denoting by $\lambda_{\text{max}}$ the maximal eigenvalue of $C^{1/2}$, we may write \[ C^{1/2}K_{t+\epsilon} = C^{1/2}\lr{K_{t}}_\epsilon\subseteq (C^{1/2}K_{t})_{\lambda_{\text{max}}\epsilon}. \] Thus, \begin{align*} \omega_{K_t}(\epsilon) \leq \frac{\mathrm{vol}\lr{(C^{1/2}K_t)_{\lambda_{\text{max}}\epsilon}\backslash C^{1/2}K_t}}{(2\pi)^{k/2}} . \end{align*} Denoting by $R$ the radius of the smallest ball containing $(C^{1/2}K_1)_{\lambda_{\text{max}}}$, we obtain from \eqref{E:Kal} that there exists a constant $c>0$ depending on $C$, $k$ and $R$ so that for any $0\leq t,\epsilon \leq 1$ \[ \omega_{K_t}(\epsilon) \leq c\epsilon. \] Thus, using \eqref{E:semi-group}, we indeed see that $\omega_K(t)$ is absolutely continuous on $[0,1]$. Hence, its derivative $\omega_K'(t)$ exists almost everywhere and \[ \omega_K(t)=\int_0^t \omega_K'(s)ds. \] Moreover, combining \eqref{E:semi-group} with \eqref{Naz-1}, we find for any $t\in (0,1)$ that \[ 0\leq \omega_K'(t)= \lim_{\epsilon\rightarrow 0_+}\frac{\omega_{K} (t+ \epsilon) - \omega_{K} (t) }{\epsilon} = \lim_{\epsilon\rightarrow 0_+}\frac{\omega_{K_t}(\epsilon)}{\epsilon}\leq c\sqrt{||C|| _{HS}} \] Hence, since $\omega_K(0)=0$, we find that $ \omega_{K} (t) \leq ct\sqrt{\| C\|_{HS}} $, for all $t\leq 1$. Using a similar argument for $\bar{\omega}_{K}(t) := \mathbb P (W\in K\setminus K_{- t} )$, we conclude that \eqref{Naz-2} indeed holds. \end{proof} We are now ready to prove Proposition \ref{Dist-lem}. Note that if $S,T$ are any events on the same probability space, then \[\abs{\mathbb P\lr{S}-\mathbb P\lr{T}}\leq \mathbb P\lr{S\Delta T},\] where $S\Delta T$ is the symmetric difference. For any convex set $A\subseteq \mathbb R^k$, we have \[\abs{\mathbb P\lr{X+Y\in A}- \mathbb P\lr{X\in A}}\] is bounded above by \[\abs{\mathbb P\lr{X+Y\in A,\, \norm{Y}_2\leq \delta}- \mathbb P\lr{X\in A,\, \norm{Y}_2\leq \delta}}+2\mathbb P\lr{\norm{Y}_2>\delta}.\] Note that \[\set{X+Y\in A,\, \norm{Y}_2\leq \delta}\Delta\set{X\in A,\, \norm{Y}_2\leq \delta}~\subseteq~\{ X\in A_{\delta}\setminus A_{-\delta}\}.\] Thus, we find \begin{align*} \abs{\mathbb P\lr{X+Y\in A}- \mathbb P\lr{X\in A}}&\leq \mathbb P\lr{X\in A_\delta} - \mathbb P\lr{X\in A_{-\delta}} + 2 \mathbb P\lr{\norm{Y}_2>\delta}\\ &\leq | \mathbb P\lr{X\in A_\delta} - \mathbb P\lr{ \mathcal N\lr{\mu, \Sigma}\in A_\delta} |\\ &+ | \mathbb P\lr{X\in A_{-\delta}} - \mathbb P\lr{ \mathcal N\lr{\mu, \Sigma}\in A_{-\delta}} |\\ &+ \mathbb P\lr{ \mathcal N\lr{\mu, \Sigma}\in A_\delta} - \mathbb P\lr{ \mathcal N\lr{\mu, \Sigma}\in A_{-\delta}} + 2 \mathbb P\lr{\norm{Y}_2>\delta}\\ &\leq 2d(X, \mathcal N\lr{\mu, \Sigma}) + c_{0}\delta \sqrt{\norm{\Sigma^{-1}}_{HS}} + 2\mathbb P\lr{\norm{Y}_2>\delta} \end{align*} where we have used \eqref{Naz-2}. Putting this all together, we find \begin{align*} d(X+Y, \mathcal N\lr{\mu, \Sigma}) &\leq d(X, \mathcal N\lr{\mu, \Sigma}) + d(X+Y, X)\\ &\leq 3d(X, \mathcal N\lr{\mu, \Sigma}) + c_{0}\delta \sqrt{\norm{\Sigma^{-1}}_{HS}} + 2\mathbb P\lr{\norm{Y}_2>\delta}. \end{align*} \hfill $\square$ \section{Roadmap for Proofs of Theorems \ref{T:global} and \ref{T:normal}}\label{S:roadmap} In this section, we explain the organization of the proofs of Theorems \ref{T:global} and \ref{T:normal}. Our starting point is in \S \ref{S:small-ball-use}. There, in Proposition \ref{P:pointwise-small-ball} we explain how to provide surprisingly useful bounds on the size of the difference \begin{equation}\label{E:local-global-diff} \frac{1}{N}\log \norm{X_{N,n}(\Theta)}-\lr{\lambda_1+\cdots + \lambda_k} \end{equation} using small ball estimates on determinants of volumes of random projections. This makes precise \eqref{E:pointwise-informal}. We remind the reader that $\lambda_1,\ldots,\lambda_k$ are the top $k$ Lyapunov exponents for $X_{N,n}$ and that \[X_{N,n}(\Theta) = X_{N,n}\theta_1\wedge\cdots X_{N,n}\theta_k,\] where $\theta_j$ are is a fixed orthonormal $k-$frame in $\mathbb R^n.$ We think of $X_{N,n}(\Theta)$ as a pointwise analog of $\lambda_1+\cdots+ \lambda_k$ since by Lemma \ref{lem-Grass-3} the supremum over $\Theta$ of $\frac{1}{N}\log\norm{X_{N,n}(\Theta)}$ equals $\lambda_1+\cdots+ \lambda_k.$ Using Proposition \ref{P:pointwise-small-ball}, we analyze in \S \ref{S:concentration} the concentration properties of $\frac{1}{N}\log\norm{X_{N,n}(\Theta)}.$ By Lemma \ref{L:pointwise-iid} it is a sum of independent random variables, allowing us to apply Theorem \ref{Latala} several times. The main result is Proposition \ref{P:pointwise}, whose proof is the most technical part of this article. Combining these concentration estimates for $\frac{1}{N}\log\norm{X_{N,n}(\Theta)}$ with the bounds on \eqref{E:local-global-diff} derived in Proposition \ref{P:pointwise-small-ball}, we derive Theorem \ref{T:global} in \S \ref{S:global-pf}, giving quantitative estimates about convergence of the global distribution of singular values of $X_{N,n}$ to the Triangle Law. Finally, in \S \ref{S:normal-outline}, we combine Theorem \ref{T:sums} with Proposition \ref{P:pointwise} and the multivariate CLT (see \S \ref{S:MCLT}) to prove the approximate normality of Lyapunov exponents stated in Theorem \ref{T:normal}. \section{Lyapunov Sums Via Small Ball Estimates}\label{S:small-ball-use} The purpose of this section is to explain how to use small ball estimates on volumes of random projections to obtain concentration estimates on the difference \[\frac{1}{N}\log \norm{X_{N,n}(\Theta)}-\sup_{\Theta'\in \mathrm{Fr}_{n,k}}\frac{1}{N}\log \norm{X_{N,n}(\Theta')}=\frac{1}{N}\log \norm{X_{N,n}(\Theta)}-\lr{\lambda_1+\cdots + \lambda_k}\] between the sum of the first $k$ Lyapunov exponents of $X_{N,n}$ and the ``pointwise'' analog of this quantity evaluated at any fixed orthonormal system $\Theta=\lr{\theta_1,\dots,\theta_k}$ of $k$ vectors in $\mathbb R^n.$ Our main result is the following \begin{proposition}\label{P:pointwise-small-ball} There exists $C>0$ with the following property. For any $\varepsilon\in (0,1)$ and any $\Theta \in \mathrm{Fr}_{n,k}$ we have \[\mathbb P\lr{\abs{\frac{1}{N}\log \norm{X_{N,n}(\Theta)} -\sup_{\Theta' \in \mathrm{Fr}_{n,k}}\frac{1}{N}\log \norm{X_{N,n}(\Theta')}}\geq \frac{k}{2N}\log\lr{\frac{n}{k\varepsilon^2}}}\leq \lr{C\varepsilon}^{k/2}.\] \end{proposition} \begin{proof} The key observation is the following: \begin{lemma}\label{L:small-ball-bounds} For any $\varepsilon\in (0,1)$ and any $\Theta \in \mathrm{Fr}_{n,k}$ we have \[\mathbb P\lr{\abs{\frac{1}{N}\log \norm{X_{N,n}(\Theta)} -\sup_{\Theta' \in \mathrm{Fr}_{n,k}}\frac{1}{N}\log \norm{X_{N,n}(\Theta')}}\geq \frac{1}{N}\log\lr{\frac{1}{\varepsilon}}}\leq \mathbb P\lr{\norm{P_F(\Theta)}\leq \varepsilon},\] where $F$ is a Haar distributed $k-$dimensional subspace of $\mathbb R^n$ and \[P_F(\Theta)=P_F\theta_1\wedge \cdots\wedge P_F\theta_k\] with $P_F$ denoting the orthogonal onto $F.$ \end{lemma} \begin{proof}[Proof of Lemma \ref{L:small-ball-bounds}] Denote by $v^{(1)},\ldots, v^{(n)}$ the right singular vectors of $X_{N,n}$ corresponding to its singular values $s_1\geq \cdots \geq s_n$. By abuse of notation, we will write $X_{N,n}:\Lambda^k\mathbb R^n\rightarrow \Lambda^k\mathbb R^n$ for the linear transformation given by \[X_{N,n}(x_1\wedge\cdots\wedge x_k)=X_{N,n}x_1\wedge\cdots \wedge X_{N,n}x_k.\] The right singular vectors of $X_{N,n}$ acting on $\Lambda^k\mathbb R^n$ are \[v^{(I)}:=v^{(i_1)}\wedge\cdots \wedge v^{(i_k)},\qquad I = \lr{i_1,\ldots, i_k},\, i_1<\cdots <i_k\] and the corresponding singular values are \[s_I:=\prod_{i\in I}s_i.\] Hence, the Pythagorean theorem for wedge products and the generalized Gram identity (see Lemma \ref{lem-Grass-2}) yield \begin{align*} \norm{X_{N,n}(\Theta)}^2 &= \sum_{\substack{I=\lr{i_1,\ldots,i_k}\\ 1\leq i_1<\cdots < i_k \leq n}} s_I^2 \inprod{v^{(I)}}{\Theta}^2\geq \lr{\prod_{i=1}^k s_i^2} \inprod{v^{(1,\ldots,k)}}{\Theta}^2 = \lr{\prod_{i=1}^k s_i^2}\norm{P_k(\Theta)}^2, \end{align*} where in the last equality we've denote by $P_k$ the orthogonal projection into the span of the top $k$ right singular vectors of $X_{N,n}$. We therefore obtain, using Lemma \ref{lem-Grass-3}: \begin{align*} 0&\geq \frac{1}{N}\log \norm{X_{N,n}(\Theta)} -\sup_{\Theta' \in \mathrm{Fr}_{n,k}}\frac{1}{N}\log \norm{X_{N,n}\Theta'}\\ &= \frac{1}{2N}\log\lr{\frac{\norm{X_{N,n}(\Theta)}^2}{\prod_{i=1}^k s_i^2}}\\ &=\frac{1}{N}\log\norm{P_k(\Theta)}. \end{align*} Since $X_{N,n}$ is invariant under right multiplication by a Haar orthogonal matrix, we see that $P_k$ is equal in distribution to the orthogonal projection onto a Haar distributed $k$-dimensional subspace of $\mathbb R^n.$ \end{proof} \noindent In order to apply Lemma \ref{L:small-ball-bounds} we need small ball estimates on $\norm{P_F(\Theta)}$. Gaussian analogs of such estimates are essentially available in the literature, but are phrased in the language of determinants of random matrices. To reduce to this case, note that if $F$ is Haar distributed among $k-$dimensional subspaces of $\mathbb R^n$, an orthonormal basis $v_1,\ldots,v_k$ for $F$ is Haar distributed on the space of such $k-$frames in $\mathbb R^n$. Thus, by \eqref{E:gaussian-frame} from Lemma \ref{L:haar-flags}, we find \begin{align} \notag \norm{P_F(\Theta)} &= \| P_{F} \theta_{1} \wedge \cdots \wedge P_{F} \theta_{k} \|\\ \notag &= \abs{\inprod{v_1\wedge\cdots\wedge v_k}{\theta_1\wedge\cdots\wedge \theta_k}}\\ \notag &\stackrel{d}{=} \abs{\inprod{\frac{g_1\wedge \cdots\wedge g_k}{\norm{g_1\wedge \cdots \wedge g_k}}}{\theta_1\wedge\cdots\wedge \theta_k}}\\ \notag&\stackrel{d}{=} \frac{ \| G \theta_{1} \wedge \cdots \wedge G \theta_{k} \| }{ {\rm det} ( G G^{\ast} )^{\frac{1}{2}} }\\ \label{E:wedge-det}&=\lr{\frac{ \det(G_kG_k^*) }{ {\rm det} ( G G^{\ast} ) }}^{\frac{1}{2}} \end{align} where $\stackrel{d}{=}$ denotes equality in distribution, $G$ is a $k\times n$ matrix with iid standard Gaussian columns $g_i$, $G_k$ is the obtained from $G$ by keeping only the first $k$ columns, and we have used the Gram identity \eqref{wedge-def} and \eqref{E:generalized-gram} in the last two lines. The relation \eqref{E:wedge-det}, combined in Lemma \ref{L:small-ball-bounds}, therefore gives that \[ \mathbb P\lr{\abs{\frac{1}{N}\log \norm{X_{N,n}(\Theta)} -\sup_{\Theta' \in \mathrm{Fr}_{n,k}}\frac{1}{N}\log \norm{X_{N,n}(\Theta')}}\geq \frac{1}{N}\log\lr{\frac{1}{\varepsilon}}} \] is bounded above by \begin{equation}\label{E:small-ball-det} \mathbb P\lr{\lr{\frac{ \det(G_kG_k^*) }{ {\rm det} ( G G^{\ast} ) }}^{\frac{1}{2}}\leq \varepsilon}, \end{equation} To complete the proof, we recall the following result. \begin{proposition}[Lemma 4.2 in \cite{paouris2013small}] \label{Gaussian-matrix-det-est} \noindent There exist universal constants $c,C>0$ with the following property. Let $ G$ be a $k\times n $ matrix with iid standard $\mathcal N(0,1)$ Gaussian entries. Then \begin{equation} \label{G-M-D-E-1} \left( \mathbb E \left[{\det} ( G G^{\ast} )^{\frac{p}{2k}}\right]\right)^{\frac{1}{p}} \leq C \sqrt{n} , \ 0< p \leq k n . \end{equation} and \begin{equation} \label{G-M-D-E-2} \left( \mathbb E \left[{\det} ( G G^{\ast} )^{-\frac{p}{2k}}\right]\right)^{-\frac{1}{p}} \geq c \sqrt{n} , \ 0< p \leq k (n-k+1- e^{-k(n-k+1)}). \end{equation} \end{proposition} \noindent This allows use to estimate the probability in \eqref{E:small-ball-det} via the following corollary, which in combination with $\eqref{E:small-ball-det}$ and Lemma \ref{L:small-ball-bounds} completes the proof of Proposition \ref{P:pointwise-small-ball}. \begin{corollary}\label{C:small-ball} \noindent There exists a universal constant $c>0$ with the following property. Let $G$ be a $k\times n$ matrix with iid Standard Gaussian entries, and denote by $G_k$ the matrix obtained from $G$ by keeping only the first $k$ columns. \begin{equation} \label{invar-cor-2} \mathbb P\left( \lr{\frac{\det(G_kG_k^*)}{\det(GG^*)}}^{\frac{1}{2k}} \leq \varepsilon \sqrt{ \frac{k}{n}}\right) \leq \left( c \varepsilon \right)^{\frac{k}{2}} , \varepsilon >0 \end{equation} \end{corollary} \begin{proof} Using \eqref{G-M-D-E-1} and \eqref{G-M-D-E-2} and Markov's inequality, we have that for $t\geq 1$, $$ \mathbb P \left( {\det} ( G G^{\ast} )^{\frac{1}{2k}} \geq t C \sqrt{n} \right) \leq \frac{1}{ t^{nk}} \ {\rm and } \ \mathbb P \left( {\det} ( G G^{\ast} )^{\frac{1}{2k}} \leq \varepsilon c \sqrt{n} \right) \leq (c \varepsilon)^{k(n-k+1-e^{-k(n-k+1)}) } .$$ Note that $ G_{k}$ has the same distribution as a $k\times k$ matrix with iid standard $\mathcal N(0,1)$ Gaussian entries. So, we have that \begin{align*} \mathbb P\left( \lr{\frac{\det(G_kG_k^*)}{\det(GG^*)}}^{\frac{1}{2k}} \leq \varepsilon \sqrt{\frac{k}{n}} \right) &\leq \mathbb P \left( \frac{ {\rm det} ( G_{k} G_{k}^{\ast} )^{\frac{1}{2k}} }{ {\rm det} ( G G^{\ast} )^{\frac{1}{2k}} } \leq \varepsilon \sqrt{\frac{k}{n}} \ {\rm and } \ {\rm det} ( G G^{\ast} )^{\frac{1}{2k}} \leq t C\sqrt{n} \right)\\ &+ \mathbb P \left( {\rm det} ( G G^{\ast} )^{\frac{1}{2k}} \geq t C\sqrt{n} \right) \\ &\leq \mathbb P \left( {\rm det} ( G_{k} G_{k}^{\ast} )^{\frac{1}{2k}} \leq \varepsilon t C\sqrt{k} \right) + \mathbb P \left( {\rm det} ( G G^{\ast} )^{\frac{1}{2k}} \geq t C\sqrt{n} \right)\\ &\leq (c^{\prime}t\varepsilon)^{k(1-e^{-k}) } + \frac{1}{ t^{kn}} \leq \left( c \varepsilon \right)^{k/2}, \end{align*} where in the last line we've taken $ t= \varepsilon^{-1/2n}$. \end{proof} \end{proof} \section{Concentration for $\frac{1}{N}\log\norm{X_{N,n}(\Theta)}$}\label{S:concentration} As mentioned above, a key step towards proving Theorems \ref{T:global} and \ref{T:normal} is to obtain precise concentration estimates for \begin{equation} \label{E:XnN-theta-def}\frac{1}{N}\log\norm{X_{N,n}(\Theta)}=\frac{1}{N}\log\norm{X_{N,n}\theta_1\wedge\cdots\wedge X_{N,n}\theta_k}, \end{equation} where $\Theta = \lr{\theta_1,\ldots,\theta_k}$ is a fixed orthonormal system in $\mathbb R^n.$ Define \begin{equation}\label{E:M-xi-def} M_j:=n-j+1,\qquad\xi_{n,k}=\frac{1}{n}\sum_{j=1}^k \frac{1}{M_j} \end{equation} and as in \eqref{E:mu-def} set \[\mu_{n,j}:=\frac{1}{2}\E{\log\lr{\frac{1}{n}\chi_{n-j+1}^2}}.\] Our main result about the concentration for $\log\norm{X_{N,n}(\Theta)}$ is the following. \begin{proposition}\label{P:pointwise} There exists a universal constant $c>0$ with the following property. Fix any orthonormal system $\Theta$ of $k$ vectors in $\mathbb R^n$. With $X_{N,n}(\Theta)$ as in \eqref{E:XnN-theta-def}, we have \begin{equation} \label{proposition9-1} \mathbb P\left( \abs{ \frac{ 1 }{ nN }\log \norm{X_{N,n}(\Theta)} - \frac{1}{n}\sum_{j=1}^k \mu_{n,j}}\geq s \right) \leq 2 \exp\left\{ - c n N \min\{ M_{k} s , \frac{s^{2}}{ \xi_{n,k}}\} \right\}, \qquad s>0 . \end{equation} \end{proposition} \begin{remark} The double behavior in the exponent of the estimates \eqref{proposition9-1} is of Bernstein-type. We do not use any off-the-shelf Bernstein estimates for deriving it however, relying instead on Theorem \ref{Latala} of \L ata\l a \cite{latala1997estimation}. One advantage of our approach is that \L ata\l a's estimates are all reversible (i.e. have matching upper and lower bounds). Hence, with a bit more work it is possible to show that the estimate \eqref{proposition9-1} is sharp. We will not need this fact, however, and will provide only a proof of the upper bound. \end{remark} \begin{remark} Although we focus in this article on the Gaussian case, we believe it is possible to prove that Proposition \ref{P:pointwise} holds under minimal assumptions on the distribution of the entries of $A_i.$ Somewhat weaker results in this directions are proved in \cite[Thms. 7,8]{le1982theoremes} and \cite[Thm. 5.1]{bougerol1985concentration}. \end{remark} \begin{remark} By Lemma \ref{L:pointwise-iid} below, we have \[ \E{\frac{ 1 }{ N }\log \norm{X_{N,n}(\Theta)}} = \sum_{j=1}^k \mu_{n,j}.\] \end{remark} The proof of Proposition \ref{P:pointwise} proceeds from the observation that for the Gaussian case we consider here, $\log X_{N,n}(\Theta)$ is a sum of independent random variables. \begin{lemma}\label{L:pointwise-iid} Fix $n,N\geq 1$ and $1\leq k \leq n$ as well as a collection $\Theta$ of $k$ orthonormal vectors in $\mathbb R^n.$ We have \[\log\norm{X_{N,n}(\Theta)}\stackrel{d}{=}\sum_{i=1}^N \sum_{j=1}^k Y_{i,j},\] where $\stackrel{d}{=}$ denotes equality in distribution, $Y_{i,j}$ are independent, and for each $i,j$ the random variable $Y_{i,j}$ is distributed like the logarithm $\frac{1}{2}\log(\frac{1}{n}\chi_{n-j+1}^2)$ of a normalized chi-squared random variable with $n-j+1$ degrees of freedom. \end{lemma} \begin{proof} We have \begin{align} \notag \log\norm{X_{N,n}(\Theta)} &= \log \norm{A_N\cdots A_1\lr{\Theta}}\\ \label{E:one-term} &= \log \norm{A_N\cdots A_2 \frac{A_1\lr{\Theta}}{\norm{A_1(\Theta)}}}+\log\norm{A_1(\Theta)}. \end{align} Note that by Lemma \ref{L:polar}, we have that \[\norm{A_1(\Theta)} = \norm{A_1\theta_1\wedge \cdots \wedge A_1 \theta_k}\] is independent of \[\frac{A_1(\Theta)}{\norm{A_1(\Theta)}}=\frac{A_1\theta_1\wedge \cdots \wedge A_1 \theta_k}{\norm{A_1\theta_1\wedge \cdots \wedge A_1 \theta_k}}.\] Hence the two terms in \eqref{E:one-term} are independent. Moreover, $A_2\lr{\frac{A_1(\Theta)}{\norm{A_1(\Theta)}}}$ is independent of $A_3,\ldots, A_N$ and, in distribution, we have \begin{equation}\label{E:dist-wedge} A_2\lr{\frac{A_1(\Theta)}{\norm{A_1(\Theta)}}}\stackrel{d}{=}A_2(\Theta). \end{equation} Indeed, we may write \[\frac{A_1(\Theta)}{\norm{A_1(\Theta)}}=\frac{A_1\theta_1\wedge \cdots \wedge A_k\theta_k}{\norm{A_1\theta_1\wedge \cdots \wedge A_k\theta_k}}= \frac{A_1\theta_1}{\norm{A_1\theta_1}}\wedge \frac{\Pi_{\leq 1}A_1\theta_1}{\norm{\Pi_{\leq 1}A_1\theta_1}}\wedge \cdots \wedge \frac{\Pi_{\leq k-1} A_1\theta_k}{\norm{\Pi_{\leq k-1} A_1\theta_k}},\] where we've written $\Pi_{\leq i}$ for the projection onto the complement of the span of $A_1\theta_1,\ldots,A_1\theta_i.$ Next, we may choose an orthogonal matrix $M$ so that \[\Pi_{\leq i-1}A_1\theta_i = M e_i, \] where $e_i$ is the $i^{th}$ standard basis vector. For this choice of $M,$ we find \[\frac{A_1(\Theta)}{\norm{A_1(\Theta)}}=Me_1\wedge \cdots \wedge Me_k= M(e_1\wedge \cdots \wedge e_k).\] Since $A_2 \stackrel{d}{=}A_2M$, we conclude that \eqref{E:dist-wedge} holds. In particular, we find that, in distribution, \[\log \norm{X_{N,n}(\Theta)}\stackrel{d}{=}\sum_{i=1}^N \log \norm{A_i(\Theta)}\] is a sum of iid terms. Finally, for any fixed $i=1,\ldots, N$ \[\norm{A_i(\Theta)}\stackrel{d}{=}\norm{\xi_1\wedge \cdots \wedge \xi_k},\] where $\xi_i$ are iid $n$-dimensional standard Gaussians. Hence, by the projection formula \eqref{wedge-1.5}, we find that \[\norm{A_i(\Theta)}\stackrel{d}{=}\prod_{i=1}^k \norm{P_{\leq i-1} \xi_i},\] where we've denoted by $P_{\leq j}$ the projection onto the orthogonal complement of the span of $\xi_1,\ldots,\xi_j$. The terms in the product are independent by Lemma \ref{L:polar}, and the distribution of the $i^{th}$ term is precisely the same as that of $\sqrt{\frac{1}{n}\chi_{n-i+1}^2}$, completing the proof. \end{proof} Lemma \ref{L:pointwise-iid} allows us to obtain precise estimates on the rate of growth of moments of $\log X_{N,n}(\Theta)$ using the result of \L ata\l a \cite{latala1997estimation} (Theorem \ref{Latala} above). These moment estimates, in turn, yield Proposition \ref{P:pointwise} via Markov's inequality applied to the optimal power of $\log \norm{X_{N,n}(\Theta)}$. We carry out these details in \S \ref{S:pointwise-pf}. \subsection{Details for Proof of Proposition \ref{P:pointwise}}\label{S:pointwise-pf} The purpose of this section is to prove Proposition \ref{P:pointwise}. Throughout this section $C,C^{\prime},c,c^{\prime}$ etc will be universal constants that may change from line to line. Recalling from \eqref{E:M-xi-def} the notation \begin{equation}\label{E:M-xi-mu-def} M_j=n-j+1,\quad \xi_{n,k}=\frac{1}{n}\sum_{j=1}^k\frac{1}{M_j}, \end{equation} we seek to show that for $s>0$ \begin{equation}\label{E:pointwise-goal-1} \mathbb P\lr{\abs{\frac{1}{nN}\log\norm{X_{N,n}(\Theta)}- \frac{1}{n}\sum_{j=1}^k\mu_{n,j}}\geq s}\leq 2 \exp\lr{-cnN\min\set{M_ks,\frac{s^2}{\xi_{n,k}}}}. \end{equation} where we remind the reader that as in \eqref{E:mu-def}, we've set \[\mu_{n,j}:=\frac{1}{2}\E{\log\lr{\frac{1}{n}\chi_{n-j+1}^2}}.\] According to Lemma \ref{L:pointwise-iid}, we have in distribution that \[\frac{2}{N}\log\norm{X_{N,n}(\Theta)} =\frac{1}{N}\sum_{i=1}^N T_i,\] where $T_i$ are independent and \begin{equation}\label{E:T-def} T_i =\sum_{j=1}^k t_{i,j},\qquad t_{i,j}\sim \log\lr{\frac{1}{n}\chi_{n-j+1}^2}\quad iid. \end{equation} Hence, we find in particular that \begin{equation}\label{E:T-mean} \E{\frac{1}{N}\log\norm{X_{N,n}(\Theta)}}= \sum_{j=1}^k \mu_{n,j} \end{equation} and see that Proposition \ref{P:pointwise} is equivalent to showing that for any $s>0$ \begin{equation}\label{E:pointwise-goal-2} \mathbb P\lr{\abs{\frac{1}{nN}\sum_{i=1}^N \overline{T}_i}\geq s}\leq 2 \exp\lr{-cnN\min\set{M_ks,\frac{s^2}{\xi_{n,k}}}}, \end{equation} where for any random variable $Y$ we will use the shorthand \[\overline{Y}:=Y- \E{Y}.\] We will obtain \eqref{E:pointwise-goal-2} by Markov's inequality applied to certain moments of the sum of the $T_i$'s. Specifically, we will prove the following estimate. \begin{proposition}\label{P:moments} There exists a universal constant $C$ so that for any $n,N,k$ and $p\geq 1$ \[\lr{\E{\bigg|\sum_{i=1}^N \overline{T}_i\bigg|^p}}^{1/p} \leq C\lr{\sqrt{pN \sum_{j=1}^k \frac{1}{M_j}} + \frac{p}{M_k}}.\] \end{proposition} The proof of Proposition \ref{P:moments}, which is straightforward but tedious, is given in \S \ref{S:moments-pf} below. We assume it for now and complete the proof of \eqref{E:pointwise-goal-2}. Write \begin{equation}\label{E:p_0-def} p_0 := M_k^2 \sum_{j=1}^k \frac{1}{M_j} \end{equation} and note that \[p\leq Np_0\quad \Longleftrightarrow \quad \frac{p}{M_k}\leq \sqrt{pN \sum_{j=1}^k \frac{1}{M_j}}.\] Thus, applying Markov's inequality to Proposition \ref{P:moments} shows that there exists $C>0$ so that for $1\leq p \leq Np_0$ \[\mathbb P\lr{\bigg|\frac{1}{nN}\sum_{i=1}^N \overline{T}_i\bigg| \geq \frac{C}{n\sqrt{N}}\sqrt{p\sum_{j=1}^k\frac{1}{M_j}}}\leq e^{-p}.\] Equivalently, recalling that \[\xi_{n,k}=\frac{1}{n}\sum_{j=1}^k \frac{1}{M_j},\] we see that there exists $c>0$ so that \[\mathbb P\lr{\bigg|\frac{1}{nN}\sum_{i=1}^N \overline{T}_i\bigg| \geq s}\leq 2e^{-cnN\frac{s^2}{\xi_{n,k}}},\qquad 0\leq s \leq C \xi_{n,k}M_k.\] This establishes \eqref{E:pointwise-goal-2} in this range of $s.$ To treat $s\geq CM_k\xi_{n,k}$, we again apply Markov's inequality to Proposition \ref{P:moments} to see that there exists $C>0$ so that \[p\geq N p_0\quad \Longrightarrow\quad \mathbb P\lr{\bigg|\sum_{i=1}^N \overline{T}_i \bigg| > C\frac{p}{M_k}}\leq e^{-p}.\] Hence, there exists $c>0$ so that \[\mathbb P\lr{\bigg|\frac{1}{nN}\sum_{i=1}^N \overline{T}_i\bigg|\geq s}\leq e^{-cnNM_ks},\quad s \geq CM_k \xi_{n,k},\] completing the proof of \eqref{E:pointwise-goal-2}. \subsection{Proof of Proposition \ref{P:moments}}\label{S:moments-pf} We seek to estimate the moments of \[\sum_{i=1}^N \overline{T_i} = \sum_{i=1}^N \sum_{j=1}^k \overline{t}_{i,j},\qquad t_{i,j}\sim \log(\frac{1}{n}\chi_{n-j+1}^2).\] By Theorem \ref{Latala}, we have \begin{equation}\label{E:Latala-iid} \lr{\E{\bigg|\sum_{i=1}^N \overline{T}_i\bigg|^p}}^{1/p}\simeq \sup_{\max\set{2,\tfrac{p}{N}}\leq s \leq p}\frac{p}{s}\lr{\frac{N}{p}}^{\frac{1}{s}}\E{\abs{\overline{T}_i}^s}^{\frac{1}{s}}. \end{equation} Our strategy is therefore to estimate $\E{\abs{\overline{T}_i}^s}^{1/s}$ and optimize over $s$. In particular, we seek to show that there exists $C>0$ so that for every $p\geq 1$, we have \begin{equation}\label{E:moments-goal} \lr{\E{\abs{\overline{T}_i}^p}}^{1/p}\leq C\lr{\sqrt{p\sum_{j=1}^k M_j^{-1}} +\frac{p}{M_k}}. \end{equation} Our first step towards showing \eqref{E:moments-goal} is to obtain the following estimates on the moments of $\overline{t}_{i,j}$. \begin{lemma}\label{L:tij-est} There exist $C_1>0$ such that \[ \lr{\E{\abs{\overline{t}_{i,j}}^p}}^{1/p}\leq C_1 \max\set{\sqrt{\frac{p}{M_j}}, \frac{p}{M_j}}.\] \end{lemma} \begin{proof} We first make a reduction. Namely, let us check that the estimates in Lemma \ref{L:tij-est} for $\overline{t}_{i,j}= t_{i,j}-\E{\log(n^{-1}\chi_{M_j}^2)}$ follow from the same estimates for \[ \widehat{t}_{i,j}:=t_{i,j}-\log(n^{-1}M_j). \] To see this, recall that by \eqref{E:mu-def-second} and \eqref{E:mu-est}, we have \[ \E{\log\lr{\frac{1}{n}\chi_{M_j}^2}} = \log\lr{\frac{2}{n}} + \psi\lr{\frac{M_j}{2}} = \log\lr{\frac{M_j}{n}}+ \varepsilon_j,\qquad \varepsilon_j = O(M_j^{-1}) \] where $\psi$ is the digamma function, and we have used its asymptotic expansion $\psi(z)\sim \log(z)+O(z^{-1})$ for large arguments. Thus, we have for each $i$ that \begin{align*} \E{\abs{\overline{t}_{i,j}}^p} &= \E{\abs{\widehat{t}_{i,j}+\varepsilon_j}^p}\leq \sum_{k=0}^p\binom{p}{k} \E{\abs{\widehat{t}_{i,j}}^k} \abs{\varepsilon_j}^{p-k}. \end{align*} So assuming that $\widehat{t}_{i,j}$ satisfy the conclusion of Lemma \ref{L:tij-est}, we find \begin{align*} \E{\abs{\overline{t}_{i,j}}^p}&\leq \sum_{k=0}^p\binom{p}{k}\zeta_{k,j}^k\abs{\varepsilon_j}^{p-k}, \qquad \zeta_{k,j}:=C\max\set{\sqrt{\frac{k}{M_j}},\frac{k}{M_j}}. \end{align*} Since for $0\leq k \leq p$ we have $\zeta_{k,j}\leq \zeta_{p,j}$, we see that \[ \E{\abs{\overline{t}_{i,j}}^p}\leq \sum_{k=0}^p\binom{p}{k}\zeta_{p,j}^k\abs{\varepsilon_j}^{p-k} \leq \lr{\zeta_{p,j}+\abs{\varepsilon_j}}^p. \] Finally, since $\varepsilon_j=O(M_j^{-1})=O(\zeta_{p,j})$, we find that there exists $C>0$ so that \[ \lr{\E{\abs{\overline{t}_{i,j}}^p}}^{1/p}\leq C \zeta_{p,j}\leq C\max\set{\sqrt{\frac{p}{M_j}},\, \frac{p}{M_j}}, \] as desired. It therefore remains to show that $\widehat{t}_{i,j}=t_{i,j} - \log\lr{n^{-1}M_j}$ satisfies the conclusion of Lemma \ref{L:tij-est}. To do this, we begin by checking that there exists $c_1>0$ so that, with $M_j=n-j+1$, for all $s\geq 0$ \begin{equation}\label{E:tij-est} \mathbb P\lr{\abs{t_{i,j} - \log\lr{\frac{M_j}{n}}}\geq s}\leq 4e^{-c_1M_j\min\set{s,s^2}}. \end{equation} We have \begin{align*} \mathbb P\lr{\abs{t_{i,j} - \log\lr{\frac{M_j}{n}}}\geq s } &= \mathbb P\lr{\abs{\log\lr{\frac{1}{n}\chi_{M_j}^2} - \log\lr{\frac{M_j}{n}}}\geq s }\\ &=\mathbb P\lr{\abs{\log\lr{\frac{1}{M_j}\chi_{M_j}^2}}\geq s }\\ &=\mathbb P\lr{\chi_{M_j}^2\geq M_j e^s }+\mathbb P\lr{\chi_{M_j}^2\leq M_j e^{-s} }. \end{align*} Let us first bound $\mathbb P\lr{\chi_{M_j}^2\geq M_j e^s }$. Note that the mean of $\chi_{M_j}^2$ is $M_j$ and that $\chi_{M_j}^2 - M_j$ is a sub-exponential random variable with parameters $(4M_j,4)$. Thus, Bernstein's tail estimates for sub-exponential random variables yield the existence of $c>0$ such that for all $t\geq 0$ \begin{equation}\label{E:bernstein} \mathbb P\lr{\abs{M_j^{-1}\chi_{M_j}^2 - 1}\geq t}\leq 2e^{-cM_j\min\set{t,t^2}}. \end{equation} In particular, \[ \mathbb P\lr{\chi_{M_j}^2 \geq M_je^s} =\mathbb P\lr{M_j^{-1}\chi_{M_j}^2-1 \geq e^s-1} \leq 2 e^{-cM_j\min\set{e^s-1,(e^s-1)^2}} \leq 2 e^{-cM_j\min\set{s,s^2}}, \] where in the last inequality we used that $e^s-1 \geq s.$ This gives the first half of \eqref{E:tij-est}. Let us now obtain a similar estimate on $\mathbb P\lr{\chi_{M_j}^2\leq M_j e^{-s} }$, which is a small ball estimate for a sum of iid random variables with bounded density. We need to consider two cases. Theorem \ref{T:iid-small-ball} on small ball estimates for sums of iid random variables shows that there exists a universal constant $C>0$ so that \[ \mathbb P\lr{\chi_{M_j}^2\leq M_j e^{-s} } \leq (Ce^{-s/2})^{M_j}. \] Hence, \[ s > s_*:=2\log(C)\quad \Rightarrow \quad \exists c>0 \text{ s.t. } \mathbb P\lr{\chi_{M_j}^2\leq M_j e^{-s} }\leq e^{-csM_j}, \] giving the desired estimate in this range. Finally, suppose $s\leq s_*$. Then, \[ e^{-s} \leq 1- e^{-s_*}s \] since both sides equal $1$ at $s=0$ and the derivative $-e^{-s}$ of the left hand side is more negative than the derivative $-e^{-s_*}$ of the right hand side for all $s\in (0,s_* )$. Thus, we find for $s\in (0,s_*)$ that \[ \mathbb P\lr{\chi_{M_j}^2\leq M_j e^{-s} } \leq \mathbb P\lr{\chi_{M_j}^2\leq M_j (1-e^{-s_*}s)} \leq \mathbb P\lr{\abs{M_j^{-1}\chi_{M_j}^2-1}> e^{-s_*}s} \] Using the Bernstein inequality \eqref{E:bernstein}, we obtain that there exists $c>0$ depending only on $s_*$ such that for $s\in (0,s_*)$, \[ \mathbb P\lr{\chi_{M_j}^2\leq M_j e^{-s} } \leq 2e^{-cM_j\min\set{s,s^2}}, \] giving the desired estimate in this range as well. This establishes \eqref{E:tij-est}. To complete the proof that $\widehat{t}_{i,j}$ satisfy Lemma \ref{L:tij-est}, we use \eqref{E:tij-est} to write \begin{align*} \E{\abs{\widehat{t}_{i,j}}^p}&=\int_0^\infty \mathbb P\lr{\abs{\widehat{t}_{i,j}}>x}px^{p-1}dx\\ &\leq p \left[\int_0^1 e^{-cM_jx^2} x^{p-1}dx + \int_1^\infty e^{-cM_jx}x^{p-1}dx\right]. \end{align*} The first term can be estimated by comparing to the moments of a Gaussian as follows: \begin{align*} p \int_0^1 e^{-cM_jx^2} x^{p-1}dx &= p\lr{ 2cM_j}^{-p/2} \int_0^{(2cM_j)^{1/2}} e^{-x^2}x^{p-1}dx\\ &\leq p\lr{ 2cM_j}^{-p/2} \int_0^{\infty} e^{-x^2}x^{p-1}dx\\ &\leq p\lr{ 2cM_j}^{-p/2} 2^{\frac{p}{2}}\Gamma\lr{\frac{p}{2}}\\ &\leq p\lr{ \frac{p}{2cM_j}}^{p/2}, \end{align*} where we used that for $z>0$ we have $\Gamma(z)\leq z^z.$ The second term can similarly be estimated by comparing the moments of an exponential: \begin{align*} p \int_1^\infty e^{-cM_jx} x^{p-1}dx &= p (cM_j)^{-p}\int_{cM_j}^\infty e^{-x}x^{p-1}dx\\ &=p(cM_j)^{-p} (p-1)!\\ &\leq p\lr{\frac{p}{cM_j}}^p. \end{align*} Putting these two estimates together and taking $1/p$ powers, we find that there exists $C>0$ so that \[ \lr{\E{\abs{\widehat{t}_{i,j}}^p}}^{1/p}\leq C \max\set{\sqrt{\frac{p}{M_j}}, \frac{p}{M_j}}\] for all $p\geq 1.$ This completes the proof. \end{proof} \noindent With Lemma \ref{L:tij-est} in hand, we are now in a position to show \eqref{E:moments-goal}. Since \[\overline{T}_i=\sum_{j=1}^k \overline{t}_{i,j},\] we have by Theorem \ref{Latala} that \begin{equation}\label{E:Latala-noniid} \left( \mathbb E |\overline{T}_i|^{p} \right)^{\frac{1}{p}} \simeq \inf\left\{ t>0 : \sum_{j=1}^{k} \log \left[ \mathbb E | 1+ \frac{\overline{t}_{i,j}}{ t}|^{p} \right] \leq p \right\}, \end{equation} where $\simeq$ means bounded above and below by absolute constants. We will use the notation from \eqref{E:p_0-def}: \[p_0=M_k^2\sum_{j=1}^k M_j^{-1}.\] Since \[\sqrt{p\sum_{j=1}^k M_j^{-1}} \leq \frac{p}{M_k}\qquad \Longleftrightarrow\qquad p\geq p_0,\] we will show \eqref{E:moments-goal} by breaking into two cases. Namely, we will show that there exists $C>0$ so that \begin{equation}\label{E:moments-goal-1} p\leq p_0\quad \Longrightarrow\quad \left( \mathbb E \left| \overline{T}_i \right|^{p} \right)^{\frac{1}{p}} \leq C \sqrt{p \sum_{j=1}^{k} M_{j}^{-1} } \end{equation} as well as \begin{equation} \label{E:moments-goal-2} p\geq p_0\quad\Longrightarrow\quad\left( \mathbb E \left| \overline{T}_i \right|^{p} \right)^{\frac{1}{p}} \leq C\frac{p}{M_{k}}. \end{equation} We may assume without loss of generality that $p$ is an even integer. Indeed, once we've show \eqref{E:moments-goal-1} and \eqref{E:moments-goal-2} for even integers $p$ (and a uniform constant $C$), we may use that \[\lr{\E{\abs{\overline{T}_i}^p}}^{1/p}\leq \lr{\E{\abs{\overline{T}_i}^{p'}}}^{1/p'},\] where $p'$ is the smallest even integer greater than or equal to $p.$ This yields \eqref{E:moments-goal-1} and \eqref{E:moments-goal-2} for any $p$ with $C$ replaced by $2C.$ We now turn to showing that \eqref{E:moments-goal-1} holds with $p$ an even integer. To do this, we will need to evaluate the expectation $\E{|1+t^{-1}t_{i,j}|^p}$ appearing in \eqref{E:Latala-noniid}. A key point is to use that $\overline{t}_{i,j}$ are centered. Since $p$ is even, we may bring this to bear most directly by noting that the absolute value in $\E{|1+t^{-1}t_{i,j}|^p}$ is unnecessary and using that $\E{\overline{t}_{i,j}}=0$. Lemma \ref{L:tij-est} and the well-known estimate \begin{equation}\label{E:binom-est} \lr{\frac{n}{k}}^k\leq \binom{n}{k}\leq \lr{\frac{n}{k}}^k e^k,\qquad k\geq 1 \end{equation} yield that for all $i,j$ we have: \begin{align} \notag \E{\lr{ 1 + \frac{ \overline{t}_{i,j}}{ t} }^{p}} &= 1 + \sum_{\ell=2}^{p} { p\choose \ell} \frac{ \E{ \overline{t}_{i,j}^{\ell}} }{ t^{\ell}}\\ \notag &\leq 1 + \sum_{\ell=2}^{\min\set{p,M_j}} \binom{p}{\ell} \lr{\frac{C\ell}{t^2M_j}}^{\ell/2}+ \sum_{\ell=M_j+1}^p\binom{p}{\ell} \lr{\frac{C\ell}{t M_j}}^\ell\\ \notag &\leq 1 + \sum_{\ell=2}^{\min\set{p,M_j}} \lr{\frac{p}{\ell}}^\ell \lr{\frac{C\ell}{t^2M_j}}^{\ell/2}+ \sum_{\ell=M_j+1}^p \lr{\frac{Cp}{t M_j}}^\ell\\ \label{E:t-t-est}&\leq 1 + \sum_{\ell=2}^{\min\set{p,M_j}} \lr{\frac{p}{\ell}}^{\ell/2} \lr{\frac{Cp}{t^2M_j}}^{\ell/2}+ \sum_{\ell=M_j+1}^p \lr{\frac{Cp}{t M_j}}^\ell. \end{align} We now bound the first two terms in the previous line by breaking into the terms where $\ell$ is even and odd. When $\ell$ is even the terms in \eqref{E:t-t-est} are bounded above by \begin{align} \label{E:even-est-1} 1+ \sum_{\substack{\ell=2\\ \ell\text{ even}}}^{\min\set{p,M_j}} \lr{\frac{p}{\ell}}^{\ell/2} \lr{\frac{C'p}{t^2M_j}}^{\ell/2}&\leq 1+\sum_{\substack{\ell=2\\ \ell\text{ even}}}^{\min\set{p,M_j}} \binom{p/2}{\ell/2} \lr{\frac{C'p}{t^2 M_j}}^{\ell/2}\leq \lr{1+\frac{C'p}{t^2M_j}}^{p/2}. \end{align} To bound the odd terms in \eqref{E:t-t-est}, let us first note that for any $0\leq m \leq \ell\leq p$, we have \[ \lr{\frac{p}{\ell}}^\ell \leq p^m \binom{p}{\ell-m}. \] Indeed, when $\ell=m$, this equality simply reads \[ \lr{\frac{p}{\ell}}^{\ell} \leq p^{\ell+1}, \] while if $m<\ell,$ we note that the inequality is equivalent to \[ \frac{p^{\ell-m}}{\ell^\ell} \leq \binom{p}{\ell-m}, \] which follows by estimating the expression on the left hand side using \eqref{E:binom-est} as follows: \[ \lr{\frac{p}{\ell}}^{\ell-m} \frac{1}{\ell^{m}}=\lr{\frac{p}{\ell-m}}^{\ell-m}\lr{\frac{\ell-m}{\ell}}^{\ell-m} \frac{1}{\ell^{m}}\leq \binom{p}{\ell-m}. \] Thus, the odd terms in \eqref{E:t-t-est} are bounded above by \begin{align} \notag \sum_{\substack{\ell=3\\ \ell\text{ odd}}}^{\min\set{p,M_j}} \lr{\frac{p}{\ell}}^{\ell/2} \lr{\frac{Cp}{t^2 M_j}}^{\ell/2} &\leq \min_{m=1,3}\set{\lr{\frac{Cp^2}{t^2M_j}}^{\frac{m}{2}}\sum_{\substack{\ell=3\\ \ell\text{ odd}}}^{p-1} {p\choose \ell-m}^{\frac{1}{2}} \lr{\frac{Cp}{t^2M_j}}^{\frac{\ell-m}{2}}}. \end{align} To proceed, note that for any $0\leq b\leq a$ \[ \binom{2a}{2b} \leq 2^b \binom{a}{b}^2. \] This inequality follows by observing that for any $j=0,\ldots, b-1$, we have \[ \frac{(2a-2j)(2a-2j-1)}{(2b-2j-1)} = \frac{(a-j)(a-j-1/2)}{(b-j)(b-j-1/2)}\leq 2 \lr{\frac{a-j}{b-j}}^2 \] and repeatedly applying this estimate to the terms in $\binom{2a}{2b}$. Thus, we obtain \begin{align} \notag \sum_{\substack{\ell=3\\ \ell\text{ odd}}}^{\min\set{p,M_j}} \lr{\frac{p}{\ell}}^{\ell/2} \lr{\frac{Cp}{t^2 M_j}}^{\ell/2} &\leq \min_{m=1,3}\set{\lr{\frac{Cp^2}{t^2M_j}}^{\frac{m}{2}}}\sum_{\ell=0}^{p/2} \binom{p/2}{\ell}\lr{\frac{Cp}{t^2M_j}}^{\ell}\\ \notag &=\min_{m=1,3}\set{\lr{\frac{Cp^2}{t^2M_j}}^{\frac{m}{2}}} \lr{1+\frac{Cp}{t^2 M_j}}^{p/2}\\ &\leq \lr{\frac{Cp^2}{t^2M_j}} \lr{1+\frac{Cp}{t^2 M_j}}^{p/2}\label{E:odd-est-1}, \end{align} where in the last inequality we've used that $\min\set{x^{1/2}, x^{3/2}}\leq x$ for all $x\geq 0.$ Putting together \eqref{E:even-est-1} and \eqref{E:odd-est-1} we see that there exists $C>0$ so that \begin{equation}\label{E:gen-moment-est} \E{ \lr{1+\frac{\overline{t}_{i,j}}{t}}^p}\leq \lr{1+\frac{Cp^2}{t^2 M_j}}\lr{1+\frac{Cp}{t^2M_j}}^{p/2}+ \sum_{\ell=M_j+1}^p \lr{\frac{Cp}{t M_j}}^\ell. \end{equation} Let us now verify \eqref{E:moments-goal-1}. Recall that for $j\leq k$ we have $M_{j} \geq M_{k}$ and that $p\leq p_{0} = M_{k}^{2} \sum_{j=1}^{k} \frac{1}{M_{j}}$. We set \begin{equation}\label{E:t-setting} t = \sqrt{C'p \sum_{i=1}^k M_i^{-1}} \end{equation} where \[ C^{\prime}= \max\set{\lr{16C}^2,\, 2C^{1/2}} \] is an absolute constant depending only on the constant $C$ appearing in \eqref{E:gen-moment-est}. For this choice of $C'$, we have \[\frac{Cp}{tM_j} = C\sqrt{\frac{p^2}{C'pM_j^{2} \sum_{i=1}^kM_i^{-1}}} \leq C\sqrt{\frac{p}{C'M_k^{2} \sum_{i=1}^kM_i^{-1}}} = C\sqrt{\frac{p}{C'p_0}}\leq \frac{1}{16}, \forall j \leq k \] and $$ \sum_{j=1}^{k} \frac{Cp}{ t^{2} M_{j}} \leq \frac{C}{(C^{\prime})^{2}} \leq \frac{1}{4}.$$ In particular for $j\leq k $ such that $ M_{j} \leq p $, $$ \sum_{\ell= M_{j}+ 1}^{p} \left( \frac{Cp}{ t M_{j}}\right)^{\ell} \leq \sum_{\ell= M_{j}+ 1}^{p} \frac{1}{ 16^{\ell}} \leq \frac{1}{ 4^{M_{j}}} $$ Hence, since $ \log{(a+b)} \leq (\log{a}) + b $ for $ a\geq 1$ and $ b>0$, \begin{align} \notag\sum_{j=1}^k\log \E{\lr{1+\lr{\frac{\overline{t}_{i,j}}{t}}^p}} &\leq \frac{p}{2}\sum_{j=1}^k \log\lr{1+\frac{Cp}{t^2M_j}} + \sum_{j=1}^k\log\lr{1+\frac{Cp^2}{t^2M_j}}+\sum_{j=1}^k \lr{\frac{1}{4}}^{M_j}\\ \notag&\leq \frac{p}{2}\sum_{j=1}^k \frac{Cp}{t^2M_j} + \sum_{j=1}^k \frac{Cp^2}{t^2M_j}+ \sum_{s= n-k+1}^{n} \frac{1}{4^{s}} \\ \notag&= \frac{3p}{8}+\frac{1}{2}\\ \notag&\leq p. \end{align} Hence, \eqref{E:moments-goal-1} follows from \eqref{E:Latala-noniid}. We now turn to the case when $p\geq p_0$ and seek to show \eqref{E:moments-goal-2}. Rather than \eqref{E:t-setting}, to show \eqref{E:moments-goal-2}, we set \begin{equation}\label{E:t-setting-2} t = \frac{C'p}{M_k} \end{equation} with \[ C' = \max\set{4C,\, 2C^{1/2}}. \] Then, \[\frac{Cp}{tM_j} = \frac{CM_k}{C'M_j}<\frac{C}{C'}\leq \frac{1}{4}\] and \[\sum_{j=1}^k \frac{Cp}{t^2M_j} = \sum_{j=1}^k \frac{CM_k^2}{(C')^2pM_j}= \frac{C}{(C')^2} \frac{p_0}{p}\leq \frac{C}{(C')^2}\leq\frac{1}{4} .\] Hence from \eqref{E:gen-moment-est}, we find \begin{align*} \sum_{j=1}^k\log \E{\lr{1+\lr{\frac{\overline{t}_{i,j}}{t}}^p}} &\leq \frac{p}{2}\sum_{j=1}^k \log\lr{1+\frac{Cp}{t^2M_j}} + \sum_{j=1}^k\log\lr{1+\frac{Cp^2}{t^2M_j}}+\sum_{j=1}^k \lr{\frac{1}{4}}^{M_j}\\ &\leq \frac{p}{2}\sum_{j=1}^k \frac{Cp}{t^2M_j} + \sum_{j=1}^k\frac{Cp^2}{t^2M_j}+\frac{1}{2}\\ &\leq \frac{3Cp}{2(C')^2} +\frac{1}{2}\\ &\leq p. \end{align*} Thus, we see that relation \eqref{E:moments-goal-2} also follows from \eqref{E:Latala-noniid}. We are now in a position to finish the proof of Proposition \ref{P:pointwise} by combining \eqref{E:Latala-iid} with \eqref{E:moments-goal-1} and \eqref{E:moments-goal-2}. We find from \eqref{E:Latala-iid} that \begin{align} \notag \lr{\E{\abs{\sum_{i=1}^N \overline{T}_i}^p}}^{1/p} &\leq \sup_{\max\set{2,\frac{p}{N}}\leq s \leq p} \frac{p}{s}\lr{\frac{N}{p}}^{1/s}\E{\abs{\overline{T}_i}^s}^{1/s}. \end{align} If $p\leq p_0$, then \eqref{E:moments-goal-1} implies \begin{align} \label{E:p-small} \lr{\E{\abs{\sum_{i=1}^N \overline{T}_i}^p}}^{1/p} &\leq C \sup_{\max\set{2,\frac{p}{N}}\leq s \leq p} \frac{p}{s}\lr{\frac{N}{p}}^{1/s}\sqrt{s\sum_{j=1}^kM_j^{-1}}. \end{align} Define \[f(s)=\log\lr{\frac{p}{s}\lr{\frac{N}{p}}^{1/s}\sqrt{s}}=\log(p) - \frac{1}{2}\log(s) + \frac{1}{s}\log\lr{\frac{N}{p}}.\] Note that \[f'(s) = -\frac{1}{2s}-\frac{1}{s^2}\log\lr{\frac{N}{p}}.\] When $p\leq N$ the function $f$ is manifestly monotone decreasing in $s>0$. Therefore, taking $s=2$, we find \[ p\leq N \quad \Rightarrow \quad \lr{\E{\abs{\sum_{i=1}^N \overline{T}_i}^p}}^{1/p} \leq C \sqrt{pN\sum_{j=1}^k M_j^{-1}} \] On the other hand, when $p\geq N$, we have \[ \frac{p}{s}\lr{\frac{N}{p}}^{1/s}\sqrt{s\sum_{j=1}^k M_j^{-1}} \leq ps^{-1/2}\sqrt{\sum_{j=1}^k M_j^{-1}}, \] which is strictly decreasing in $s>0$. Hence, taking $s=p/N$, we again find \[ p\geq N \quad \Rightarrow \quad \lr{\E{\abs{\sum_{i=1}^N \overline{T}_i}^p}}^{1/p} \leq C \sqrt{pN\sum_{j=1}^k M_j^{-1}}. \] Hence, we find that if $p\leq p_0$, then \begin{align*} \lr{\E{\abs{\sum_{i=1}^N \overline{T}_i}^p}}^{1/p} &\leq C \sqrt{pN\sum_{j=1}^kM_j^{-1}}, \end{align*} as desired. Finally, if $p\geq p_0$, then \eqref{E:moments-goal-2} implies \begin{align} \label{E:p-small} \lr{\E{\abs{\sum_{i=1}^N \overline{T}_i}^p}}^{1/p} &\leq C \sup_{\max\set{2,\frac{p}{N}}\leq s \leq p} \frac{p}{s}\lr{\frac{N}{p}}^{1/s}\frac{s}{M_k}=\frac{Cp}{M_k} \sup_{\max\set{2,\frac{p}{N}}\leq s \leq p} \lr{\frac{N}{p}}^{1/s}. \end{align} Note that if $p/N\geq 1$, then for every $s\geq 2$ we have $(N/p)^{1/s}\leq 1$ and hence \[\frac{Cp}{M_k}\sup_{\max\set{2,\frac{p}{N}}\leq s \leq p} \lr{\frac{N}{p}}^{1/s}\leq \frac{Cp}{M_k}.\] Further, if $p/N\leq 1$, then $(N/p)^{1/s}$ is monotonically decreasing with $s$ and hence \[\frac{Cp}{M_k}\sup_{\max\set{2,\frac{p}{N}}\leq s \leq p} \lr{\frac{N}{p}}^{1/s}\leq \frac{Cp}{M_k}\lr{\frac{N}{p}}^{1/2}=C\sqrt{pNM_k^{-2}}\leq C\sqrt{pN\sum_{j=1}^k M_j^{-1}}.\] Thus, in all cases we find \[\lr{\E{\bigg|\sum_{i=1}^N \overline{T}_i\bigg|^p}}^{1/p} \leq C\lr{\sqrt{pN \sum_{j=1}^k \frac{1}{M_j}} + \frac{p}{M_k}},\] which is precisely the statement of Proposition \ref{P:moments}.\hfill $\square$ \section{Lyapunov Sums: Proof of Theorem \ref{T:sums}}\label{S:sums-pf} To prove Theorem \ref{T:sums}, we start by combining previous estimates for $\frac{1}{N}\log\norm{X_{N,n}(\Theta)}$ from Proposition \ref{P:pointwise} with the deviation estimates in Proposition \ref{P:pointwise-small-ball}. Let recall the definition of the function $g$ (cf \eqref{E:g-def}) \begin{equation} g_{n,k}(s) = \left\{ \begin{array}{cc} \min\left\{ 1, \frac{ n s}{ k}\right\} \ ,& \hspace{5mm} k\leq \frac{n}{2} \\ \min\left\{ \delta_{n,k}, \frac{ s}{ \log{1/\delta_{n,k}}} \right\} \ ,& \hspace{5mm} k>\frac{n}{2} \\ \end{array} \right., \end{equation} where we recall that for $ k \geq \frac{n}{2}$ \[ \delta_{n,k}:= \frac{ n-k+1}{n}.\] Let $ s>0$ and $ 1\leq k<m \leq n. $ Writing $$ p_{s,k, m} := \mathbb P \left( \abs{ \frac{1}{n} \sum_{i=m}^{k}\lr{ \lambda_{i} -\mu_{n,i}} } \geq s \right), $$ we seek to show that there exist universal constants $c_1,c_2,c_3>0$ such that for any $ 1\leq m \leq k \leq n $ and every $ s\geq c_1 \frac{k}{nN} \log{\frac{en}{k}} $ we have \begin{equation} \label{conc-L-1-new} p_{s,k,m} \leq c_2 \exp{ \left\{ - c_3\min\set{nNs, n^2 N s g_{n,k}(s) } \right\}}. \end{equation} To see this, note that the triangle inequality yields \[p_{s,k,m}\leq p_{s/2,k,1} + p_{s/2,m-1,1}.\] Hence, it suffices to prove \eqref{conc-L-1-new} with $m=1.$ To do this, fix $\Theta \in \mathrm{Fr}_{n,k}$, an orthonormal $k-$frame in $\mathbb R^n$. We may write for any $s>0$ \begin{align} \label{E:prob-1} p_{s,k, 1} & \leq \mathbb P\lr{\abs{\frac{1}{n}\sum_{i=1}^k \lambda_i - \frac{1}{nN}\log\norm{X_{N,n}(\Theta)}}\geq \frac{s}{2}}\\ \label{E:prob-2} &+ \mathbb P\lr{\abs{\frac{1}{nN}\log\norm{X_{N,n}(\Theta)} - \frac{1}{n}\sum_{i=1}^k \mu_{n,i}}\geq \frac{s}{2}}. \end{align} We will show separately that the probabilities in \eqref{E:prob-1} and \eqref{E:prob-2} are both bounded above by the right hand side of \eqref{conc-L-1-new}. To check this for \eqref{E:prob-2}, note that Proposition \ref{P:pointwise} guarantees \begin{equation} \label{proposition9-1-new-00} \mathbb P\left( \left| \frac{ \log{ \| X_{N,n} ( \Theta)\| } }{ nN } - \frac{1}{n} \sum_{j=1}^{k} \mu_{n,j} \right|\geq s \right) \leq 2 \exp\left\{ - c n N \min\{ M_{k} s , \frac{s^{2}}{ \xi_{n,k}}\} \right\} , s>0 , \end{equation} where we remind the reader that $$ M_{j} := n-j+1 , \ \xi_{n,k} := \frac{1}{n} \sum_{j=1}^{k} \frac{ 1}{M_{j}} , \ \mu_{n,k} := \frac{1}{2}\mathbb E \left[ \log\left( \frac{1}{n} \chi_{n-j+1}^{2}\right)\right] .$$ Some routine algebra reveals \begin{equation} \label{xi-mu-estimates-1} k\leq \frac{n}{2}\qquad \Rightarrow \qquad n\xi_{n,k} \simeq \frac{ k}{n} ,\qquad M_{k} \simeq n \end{equation} and \begin{equation} \label{xi-mu-estimates-1} k\geq \frac{n}{2}\qquad \Rightarrow \qquad n\xi_{n,k} \simeq \log{\left( \frac{1}{\delta_{n,k}}\right)} ,\qquad M_{k}=\delta_{n,k}n, \end{equation} where $a \simeq b$ means that there exists $c_1,c_2>0$ so that $c_1a\leq b\leq c_2 a.$ Hence, \[ k\leq \frac{n}{2}\quad \Rightarrow \quad \min\set{M_ks,\frac{s^2}{\xi_{n,k}}} \simeq ns\min\set{1, \frac{ns}{k}} \] and similarly \[ k\geq \frac{n}{2}\quad \Rightarrow \quad \min\set{M_ks,\frac{s^2}{\xi_{n,k}}} \simeq ns\set{\delta_{n,k}, \frac{s}{\log(1/\delta_{n,k})}}. \] Putting these two estimates together, we find that \eqref{proposition9-1-new-00} yields for any $s>0$ \begin{align} \label{proposition9-1-new-1} \mathbb P\left( \left| \frac{ \log \norm{X_{N,n}(\Theta)} }{ nN } - \frac{1}{n} \sum_{j=1}^{k} \mu_{n,j} \right| \geq s \right) \leq 2 \exp\left\{ - c n^{2} N sg_{n,k}(s) \right\}, \end{align} as desired. Turning to the probability in \eqref{E:prob-1}, recall that in Proposition \ref{P:pointwise-small-ball}, we have shown that for every $ \varepsilon\in (0,1) $, \[\mathbb P\lr{\abs{\frac{1}{nN}\log \norm{X_{N,n}(\Theta)} - \frac{1}{n}\sum_{i=1}^{k} \lambda_{i} }\geq \frac{k}{2Nn}\log\lr{\frac{n}{k\varepsilon^2}}}\leq \lr{C\varepsilon}^\frac{k}{2}.\] If we set $ s:= \frac{k}{nN} \log{ \frac{en}{k\varepsilon^{2}}} $, then \[ (C\varepsilon)^{k/2} = \exp\left[-\frac{1}{4}snN + \frac{k}{4}\log\lr{\frac{en}{k}}+\frac{k}{2}\log(C)\right]. \] Hence, assuming that \[ s\geq C'\frac{k}{nN}\log\lr{\frac{en}{k}} \] for $C'$ sufficiently large, we arrive to the following expression: \begin{equation} \label{conc-by-sb} \mathbb P \left(\abs{ \frac{1}{n} \sum_{i=1}^{k} \lambda_{i} - \frac{ \log \norm{X_{N,n}(\Theta)}}{nN} }\geq s \right) \leq e^{ - \frac{snN }{4}}, \quad s\geq C'\frac{ k}{nN} \log\lr{ \frac{en}{k}}. \end{equation} Thus, putting together \eqref{proposition9-1-new-1} and \eqref{conc-by-sb}, we find that \[ p_{s,k,1}\leq c_2 \exp\set{-c_3 \min\set{nNs, n^2 Ns g_{n,k}(s)}}, \] completing the proof. \hfill $\square$ \section{Convergence to the Triangle Law: Proof of Theorem \ref{T:global}}\label{S:global-pf} In this section, we derive Theorem \ref{T:global} from Theorem \ref{T:sums}. We will need the following elementary result. \begin{lemma}\label{L:digamma} Fix positive integers $n,q,m$ satisfying $4\leq m\leq q\leq n$. Then \[\frac{m}{2n}\log\lr{q/n} - \frac{1}{n}\sum_{j=n-q+1}^{n-q+m} \mu_{n,j}\geq \frac{(m-1)^2}{4nq}.\] Further, assuming that $n-q-m\geq 0$, we also have \[\frac{m}{2n}\log(q/n)-\frac{1}{n}\sum_{j=n-q-m+1}^{n-q}\mu_{n,j}\leq -\frac{m(m-3)}{3nq}.\] \end{lemma} \begin{proof} Let us first prove the lower bound. Recall that \begin{equation}\label{E:mu-form} \mu_{n,j}=\frac{1}{2}\lr{\log\lr{\frac{2}{n}}+\psi\lr{\frac{n-j+1}{2}}}. \end{equation} Moreover, a well-known estimate \cite[eq.6.3.18, p.259]{abramowitz+stegun} for the digamma function is: \[\psi(x)< \log(x).\] Using this we obtain \[\mu_{n,j}< \frac{1}{2}\log\lr{1-\frac{j-1}{n}} ,\] which allows us to write \begin{align*} \frac{m}{2n}\log\lr{q/n}- \frac{1}{n}\sum_{j=n-q+1}^{n-q+m} \mu_{n,j} &\geq \frac{1}{2n}\sum_{j=n-q+1}^{n-q+m}\log\lr{\frac{q}{n-j+1}}=\frac{1}{2n}\sum_{j=1}^{m-1}\log\lr{\frac{1}{1-j/q}}. \end{align*} Since $q,n$ are fixed, let us temporarily introduce \[\xi:=\frac{q}{n}.\] With this notation, because $\log(1/(1-t))$ is monotonically increasing for $t\in [0,1)$, we have \[\frac{1}{2n}\sum_{j=1}^{m-1}\log\lr{\frac{1}{1-j/q}}=\frac{1}{2}\sum_{j=1}^{m-1}\frac{1}{n}\log\lr{\frac{1}{1-(j/n)/\xi}}\geq \frac{1}{2}\int_0^{\frac{m-1}{n}} \log\lr{\frac{1}{1-t/\xi}}dt.\] Some routine calculus therefore reveals that \begin{align*} \frac{m}{2n}\log\lr{q/n}- \frac{1}{n}\sum_{j=n-q+1}^{n-q+m} \mu_{n,j}&\geq \frac{\xi}{2}\left[\lr{1-\varepsilon/\xi}\log\lr{1-\varepsilon/\xi} + \varepsilon/\xi\right], \end{align*} where we've set $\varepsilon := (m-1)/n.$ Finally, note that for $x>0$ \[(1-x)\log(1-x) + x = \sum_{k\geq 2} \frac{x^k}{k(k-1)} \geq x^2/2.\] Hence, we obtain \begin{align*} \frac{m}{2n}\log\lr{q/n}- \frac{1}{n}\sum_{j=n-q+1}^{n-q+m} \mu_{n,j}&\geq \frac{\varepsilon^2}{4\xi}=\frac{(m-1)^2}{4nq}, \end{align*} as claimed. Let us now derive the upper bound. Using \eqref{E:mu-form}, we get \begin{align*} \frac{m}{2n}\log\lr{q/n}- \frac{1}{n}\sum_{j=n-q-m+1}^{n-q} \mu_{n,j} &= \frac{1}{2n}\sum_{j=n-q-m+1}^{n-q}\left\{\log\lr{\frac{q}{2}} - \psi\lr{\frac{n-j+1}{2}}\right\}\\ &= \frac{1}{2n}\sum_{j=1}^{m}\left\{\log\lr{\frac{q}{2}} - \psi\lr{\frac{q+j}{2}}\right\}. \end{align*} Using the inequality (see again \cite[eq.6.3.18, p.259]{abramowitz+stegun}) \[\psi(x)>\log\lr{x} - 1/x,\] we arrive at the estimate \begin{align*} \frac{m}{2n}\log\lr{q/n}- \frac{1}{n}\sum_{j=n-q-m+1}^{n-q} \mu_{n,j} &\leq \frac{1}{2n}\sum_{j=1}^{m}\left\{\log\lr{\frac{q}{q+j}}+\frac{2}{q+j}\right\}. \end{align*} As before, we will estimate this sum above by an integral. Still writing $\xi=q/n$, we have as before \[\frac{1}{2n}\sum_{j=1}^m \log\lr{\frac{1}{1+j/q}} \leq \frac{\xi}{2}\int_0^{\varepsilon/\xi} \log\lr{\frac{1}{1+t}}dt = \frac{\xi}{2}\left[-(1+\varepsilon/\xi)\log\lr{1+\varepsilon/\xi}+ \varepsilon/\xi\right]\] where we've now set $\varepsilon = m/n$ (which is slightly different than above). For $x\in (0,1)$, we have \[-(1+x)\log(1+x)+x = \sum_{k\geq 2}(-1)^{k+1}\frac{x^k}{k(k-1)}\leq -\frac{x^2}{2}+\frac{x^3}{6}\leq -\frac{x^2}{3},\] we therefore find \[\frac{1}{2n}\sum_{j=1}^m \log\lr{\frac{1}{1+j/q}}\leq -\frac{\varepsilon^2}{3\xi} .\] Next, \[\frac{1}{q}\sum_{j=1}^m \frac{1}{n}\frac{1}{1+\xi^{-1}(j/n)}\leq \frac{1}{q}\int_0^{\varepsilon} \frac{dt}{1+t/\xi} =\frac{\xi}{q}\log\lr{1+ \varepsilon/\xi}\leq \frac{\varepsilon}{q}.\] So all together we find the upper bound \[ \frac{m}{2n}\log\lr{q/n}- \frac{1}{n}\sum_{j=n-q-m+1}^{n-q} \mu_{n,j}\leq -\frac{m^2}{3nq}+\frac{m}{nq}=-\frac{m(m-3)}{3nq}.\] This completes the proof. \end{proof} \noindent We now conclude the proof of Theorem \ref{T:global}. To do this, fix $\varepsilon>0$ and assume that \[n>\frac{c_1\sqrt{\log(1/\varepsilon)}}{\varepsilon},\qquad N>\frac{c_2}{\varepsilon^2}\] for some constants $c_1,c_2>1$ that we will fix later. To prove Theorem \ref{T:global} note that the bound above on $n$ guarantees that $\varepsilon>c_1/n.$ Hence, we need only consider such $\varepsilon.$ Moreover, we may always assume that \[\varepsilon = \frac{m}{n}\] for some integer $5\leq m\leq n$ since $\mathcal U(t)$ is $1$-lipschitz, and will use $\varepsilon$ and $m/n$ interchangeably. Next, recall the following notation for the cumulative distributions \begin{align*} \mathcal H_{N,n}(s):=\frac{1}{n}\#\set{j\leq n~\big |~ s_j(X_{N,n})^{2/N} \leq s},\qquad \mathcal U(s):=\begin{cases}0,&\quad s<0\\ s,&\quad 0\leq s\leq 1\\ 1,&\quad s>1\end{cases} \end{align*} of the squared singular values of $X_{N,n}$ and the uniform distribution on $[0,1]$. Let us define the event \[S_{n,m}:=\set{\forall q\in \set{m+1,\ldots, n}~ \abs{\mathcal H_{N,n}\lr{q/n}-\mathcal U\lr{q/n}}\leq \varepsilon}.\] On this event, since $\mathcal H_{N,n}$ and $\mathcal U$ are both monotone we have for $t\leq (m+1)/n$ that \begin{align*} \mathcal H_{N,n}(t) &\leq \mathcal H_{N,n}\lr{(m+1)/n}\leq \varepsilon + \mathcal U\lr{(m+1)/n}=\varepsilon + (m+1)/n \leq 3\varepsilon. \end{align*} Similarly if $t>1$ \begin{align*} 1-\mathcal H_{N,n}(t)\leq 1-\mathcal H_{N,n}\lr{1} \leq 1-\mathcal U\lr{1}+\varepsilon = \varepsilon \end{align*} Using the same idea we may write for any $t\in [(m+1)/n,1]$ \begin{align*} \mathcal H_{N,n}(t)- \mathcal U(t) &\leq \mathcal H_{N,n}((j+1)/n) - \mathcal U((j+1)/n) + \mathcal U(t) - \mathcal U((j+1)/n)\\ &\leq \varepsilon + 1/n\\ &\leq 2\varepsilon, \end{align*} where $m+1\leq j\leq n$ is the unique integer for which \[\frac{j}{n}\leq t < \frac{j+1}{n}.\] Hence, \[\mathbb P\lr{\sup_{t\in \mathbb R}\abs{\mathcal H_{N,n}(t)- \mathcal U(t)}> 3\varepsilon}\leq \mathbb P\lr{S_{n,m}^c}\] and we must therefore bound $\mathbb P\lr{S_{n,m}^c}$ from above. We will do this by rewriting all events involving the singular values $s_i$ in terms of the Lyapunov exponents $\lambda_j.$ Moving forward, let us agree that any event that involves $\lambda_{-s}$ or $\lambda_{n+s}$ for $ s>0$ is by definition empty. Since \[ s_j(X_{N,n})^{2/N}\leq \frac{q}{n}\quad \Leftrightarrow\quad \lambda_j\leq \frac{1}{2}\log\lr{\frac{q}{n}}, \] we find \begin{align*} \left| {\cal{H}}_{N,n} \left(q/n\right) - {\mathcal U}\left( q/n \right) \right| &= \left| \frac{1}{n} \#\left\{ j\leq n : \lambda_{j} \leq \frac{1}{2} \log\left(q/n\right) \right\} - q/n \right|\\ &=\left| \frac{1}{n} \#\left\{ j\leq n : \lambda_{j} > \frac{1}{2} \log\left( q/n\right) \right\} - (n-q)/n \right| . \end{align*} For any positive integer $m+1\leq q\leq n$, define \begin{equation} \label{tria-lem-1} p:=p(n,m,q,N)=\mathbb P \left( \left| \frac{1}{n} \#\left\{ j\leq n : \lambda_{j} > \frac{1}{2} \log\left(q/n\right) \right\} - (n-q)/n \right| \geq \frac{m}{ n} \right) . \end{equation} We have \[\mathbb P\lr{S_{n,m}^c} \leq \sum_{q=m+1}^n p(n,m,q,N),\] and the proof of Theorem \ref{T:global} therefore reduces to estimating the probabilities in this sum. To do this, we fix $n,m,q,N$ and observe that since $\lambda_j$ are decreasing, the event whose probability we've denoted by $p(n,m,q,N)$ is equal to $$ \left\{ \lambda_{n-q+m} > \frac{1}{2} \log\left(q/n\right) \right\} \ \cup \ \left\{ \lambda_{n-q-m+1} \leq \frac{1}{2} \log\left( q/n\right) \right\},$$ where we remind the reader that the second event is empty if $q\geq n-m+1.$ Again using the monotonicity of $\lambda_j$, this implies $$ \left\{ \frac{1}{n}\sum_{j=n-q+1}^{n-q+m} \lambda_{j}> \frac{m}{2n} \log\left( q/n\right) \right\} \ \cup \ \left\{ \frac{1}{n} \sum_{j=n-q-m+1}^{n-q} \lambda_{j} \leq \frac{m}{2n} \log\left(q/n\right) \right\}.$$ So, the probability $p(n,m,q,N)$ we seek to bound is itself bounded above by $$p_{1} + p_{2} := \mathbb P \left( \frac{1}{n}\sum_{j=n-q+1}^{n-q+m} \lambda_{j}> \frac{m}{2n} \log\left( q/n\right) \right) + \mathbb P \left( \frac{1}{n} \sum_{j=n-q-m+1}^{n-q} \lambda_{j} \leq \frac{m}{2n} \log\left(q/n\right) \right) . $$ \noindent To estimate $p_{1}$ note that by Lemma \ref{L:digamma}, \[\frac{m}{2n}\log\lr{q/n} - \frac{1}{n}\sum_{j=n-q+1}^{n-q+m} \mu_{n,j}\geq \frac{(m-1)^2}{4nq}.\] Hence, we obtain \[p_1\leq \mathbb P\lr{\abs{\frac{1}{n}\sum_{j=n-q+1}^{n-q+m}\left\{\lambda_j- \mu_{n,j} \right\}}\geq \frac{(m-1)^2}{4nq}}.\] We will bound the right hand side by using Theorem \ref{T:sums}. To do this, we must ensure that for $c_2$ sufficiently large, our assumption $N>c_2/\varepsilon^2$ implies that for the constant $c_1$ in Theorem \ref{T:sums}, we have \begin{equation}\label{E:apply-thm} \frac{(m-1)^2}{4nq}\geq c_1 \frac{n-q+m}{nN}\log\lr{\frac{en}{n-q+m}}. \end{equation} To check this, note that since $x\log(e/x)\leq 1$ for $x\in [0,1]$ this estimate holds as soon as \begin{equation}\label{E:inq1} N\geq c_1 \frac{4nq}{(m-1)^2}. \end{equation} Recall that by construction, we have \[q\leq n,\qquad (m-1)^2 \geq\frac{1}{2}m^2= \frac{1}{2}\varepsilon^2 n^2.\] Hence, \eqref{E:inq1} is satisfied once \[N\geq 8c_1\varepsilon^{-2},\] as claimed. Thus, we may apply Theorem \ref{T:sums} to conclude that \[p_1(n,m,q,N)\leq c_3\exp\lr{c_4\min\set{nNs, n^2Nsg_{n,k}(s)}},\qquad s=s(n,m,q)=\frac{(m-1)^2}{4nq}.\] Since \[\inf_{q=m+1,\ldots,n}s(n,m,q) \geq \frac{\varepsilon^2}{8},\] we find that \[\sum_{q=m+1}^n p_1(m,n,q,N) \leq c_3\exp\lr{-c_4\min\set{nN\varepsilon^2, n^2N\varepsilon^2 g_{n,k}(\varepsilon^2)} + \log(n)}\] for some universal constants $c_3,c_4>0$. Further, note that by assumption, \[ nN\varepsilon^2 > c_2 n > \log(n) \] as soon as $c_2>1.$ Hence, at the cost of replacing $c_4$ by a slightly larger constant $c_4'$, we find that \[\sum_{q=m+1}^n p_1(m,n,q,N) \leq c_3\exp\lr{-c_4'\min\set{nN\varepsilon^2, n^2N\varepsilon^2 g_{n,k}(\varepsilon^2)}}.\] Essentially the identical argument (but this time the upper bound from Lemma \ref{L:digamma}) implies that this same upper bound holds for $p_2$ as well, completing the proof of Theorem \ref{T:global}.\hfill $\square$ \section{Asymptotic Normality: Proof for Theorem \ref{T:normal}}\label{S:normal-outline} Theorem \ref{T:normal} concerns \[\Lambda_k = \lr{\lambda_1,\ldots, \lambda_k}=\lr{\lambda_1(X_{N,n}),\ldots, \lambda_k(X_{N,n})},\] the vector of the first $k$ Lyapunov exponents of $X_{N,n}$. Our aim is to prove that there exist universal constants $C,C'>0$ so that once $N\geq Cn\log(n)$ we have \begin{equation}\label{E:normal-goal} d\lr{\Lambda_k, \mathcal N\lr{\mu_{n,\leq k}, \Sigma_{n,k,N}}}\leq C'\lr{\frac{k^{7/2}n\log^2(n)\log(N/n)}{N}}^{1/2},\qquad \Sigma_{n,k,N}:=\frac{1}{N}\mathrm{Diag}(\sigma_{n,\leq k}^{2}) \end{equation} where $\mu_{n,\leq k},\sigma_{n,\leq k}^{2}$ are the vectors of means and variances of $\lr{\frac{1}{2}\log\lr{\frac{1}{n}\chi_{n-m+1}^2},\, m=1,\ldots,k}$ (see \eqref{E:mu-sigma-def}) and $d$ is the distance function defined in \eqref{E:dist-def}. To prove \eqref{E:normal-goal}, we introduce \[S_k = \lr{\lambda_1,\lambda_1+\lambda_2,\ldots, \lambda_1+\cdots+\lambda_k}^*\] and note that \begin{equation}\label{E:T-def} S_k= T\Lambda_k, \end{equation} where $T$ is a lower triangular matrix all of whose lower-triangular entries are equal to $1$. The explain the strategy for proving Theorem \ref{T:normal}, let us fix $\Theta=\theta_1\wedge\cdots \wedge \theta_k$, where $\set{\theta_j}$ form an orthonormal $k$-frame in $\mathbb R^n.$ For $1\leq m \leq k,$ we will abbreviate \[\Theta_{\leq m}= \theta_1\wedge\cdots \wedge \theta_m.\] The idea of the proof is to compare $S_k,\Lambda_k$ to their ``pointwise'' analogs \[\widehat{S}_k: = \frac{1}{N}\lr{\log \norm{X_{N,n}(\Theta_{\leq 1})}, \ldots, \log \norm{X_{N,n}(\Theta_{\leq k})}}^*\] and \begin{equation}\label{E:T-def-0} \widehat{\Lambda}_k := T^{-1}\widehat{S}_k, \end{equation} where $\Theta = \lr{\theta_1,\ldots,\theta_k}$ is any fixed collection of $k$ orthonormal vectors in $\mathbb R^n.$ Specifically, by Proposition \ref{Dist-lem} and the affine invariance \eqref{E:affine-inv} of $d$, we find that there exists $c_0>0$ so that for all $\delta>0$ \begin{align} \notag d\lr{\Lambda_k, \mathcal N\lr{\mu_{n,k}, \Sigma_{n,k}}} & = d\lr{S_k, \mathcal N\lr{T\mu_{n,k}, T\Sigma_{n,k,N} T^*}} \\ \notag &\leq 3d\lr{\widehat{S}_k,\mathcal N\lr{T\mu_{n,k}, T\Sigma_{n,k,N} T^*}} +c_0\delta \norm{\lr{T\Sigma_{n,k} T^*}^{-1}}_{HS}^{1/2}\\ \notag &+ 2 \mathbb P\lr{\norm{S_k - \widehat{S}_k}>\delta}\\ \notag &= 3d\lr{\widehat{\Lambda}_k,\mathcal N\lr{\mu_{n,k}, \Sigma_{n,k,N} }} +c_0\delta \norm{\lr{T\Sigma_{n,k,N} T^*}^{-1}}_{HS}^{1/2}\\ \label{E:KS-terms}&+ 2 \mathbb P\lr{\norm{S_k - \widehat{S}_k}>\delta}. \end{align} The remainder of the proof consists of bounding each of these three terms and then optimizing over $\delta$. To start, let us check that the first term in \eqref{E:KS-terms} is small: \begin{lemma}\label{L:pointwise-guassian} In distribution, \begin{equation}\label{E:pointwise-gaussian} \widehat{\Lambda}_k = \frac{1}{N}\sum_{i=1}^N \lr{Y_{i,1},\ldots, Y_{i,k}}, \end{equation} where $\set{Y_{i,j},~1\leq i \leq N,\,\, 1\leq j\leq k}$ are independent with \[Y_{i,j}\sim \frac{1}{2} \log \lr{\frac{1}{n}\chi_{n-j+1}^2}.\] Consequently, by the multivariate central limit theorem, there exists $C>0$ so that \begin{equation}\label{E:first-term} d\lr{\widehat{\Lambda}_k,\mathcal N\lr{\mu_{n,\leq k}, \Sigma_{n,k,N} }}\leq \frac{C k^{7/4}}{N^{1/2}} \end{equation} where $ \Sigma_{n,k,N} = \frac{1}{N} {\mathrm{ Diag}(\sigma_{n,\leq k }^{2} )}$, $ \sigma_{n,j}^{2} :=\Var\left[\frac{1}{2}\log\lr{\frac{1}{n}\chi_{n-j+1}^2}\right]$. \end{lemma} \begin{proof} Fix integers $N,n\geq 1$ and $1\leq k \leq n$ and recall that $X_{N,n}=A_N\cdots A_1$ with $A_i$ iid $n\times n$ Gaussian matrices. Note that for each $1\leq m\leq k$, we have \begin{equation}\label{E:vector-log} \log\norm{X_{N,n}(\Theta_{\leq m})} = \sum_{i=1}^N \log \norm{A_i(\Theta_{\leq m}^{(i)})}, \end{equation} where \[ \Theta_{\leq m}^{(1)}= \Theta_{\leq m},\qquad \Theta_{\leq m}^{(i+1)} = \frac{A_i\lr{\Theta_{\leq m}^{(i)}}}{\norm{A_i\lr{\Theta_{\leq m}^{(i)}}}}. \] Repeatedly applying Lemma \ref{L:haar-flags}, we therefore conclude that in distribution \[\widehat{S}_k=\frac{1}{N}\sum_{i=1}^N \lr{\log\norm{A_i(\Theta_{\leq 1})},\ldots, \log \norm{A_i(\Theta_{\leq k})}}^*\] is equal to a sum of iid random vectors. Thus, using the definition \eqref{E:T-def-0} of $\widehat{\Lambda}_k$, we find that in distribution \[\widehat{\Lambda}_k=\frac{1}{N}\sum_{i=1}^N\widehat{\Lambda}_{k,i},\qquad \widehat{\Lambda}_{k,i}:= T^{-1}\lr{\log\norm{A_i(\Theta_{\leq 1})},\ldots, \log \norm{A_i(\Theta_{\leq k})}}^*,\] where we recall that $T$ is a lower triangular matrix with all lower triangular entries equal to $1.$ Namely, \[T = \lr{\begin{array}{ccccc} 1 & 0 & 0&\cdots & 0 \\ 1 & 1 & 0& \cdots & 0 \\ \vdots & \cdots & \ddots &\ddots &\vdots\\ 1 & \cdots & 1 & 1& 0 \\ 1 & \cdots & 1 & 1& 1 \\\end{array}},\qquad T^{-1}= \lr{\begin{array}{ccccc} 1 & 0 & 0&\cdots & 0 \\ -1 & 1 & 0& \cdots & 0 \\ \vdots & \ddots & \ddots &\ddots &\vdots\\ 0 & \cdots & -1 & 1& 0 \\ 0 & \cdots & 0 & -1& 1 \\ \end{array}}.\] Note that $\set{\widehat{\Lambda}_{k,i},\, m=1,\ldots, k}$ are independent collection for different $i$. Next, the $m^{th}$ component of $\widehat{\Lambda}_{k,i}$ is \begin{align} \label{E:components} \lr{\widehat{\Lambda}_{k,i}}_m=\log\norm{A_i(\Theta_{\leq m})} -\log\norm{A_i(\Theta_{\leq m-1})} = \log \norm{\frac{A_i(\Theta_{\leq m-1})}{\norm{A_i(\Theta_{\leq m-1})}}\wedge A\theta_m} \end{align} Since $\set{\theta_i}$ are orthonormal, the collection $\set{A\theta_i}$ are iid Gaussians. In particular, we see that $A\theta_m$ is independent of $\set{A(\Theta_{\leq j}), 1\leq j\leq m-1}.$ Also, by Lemma \ref{L:polar}, the following collections of random variables are independent: \[\set{\norm{A(\Theta_{\leq 1})},\ldots, \norm{A(\Theta_{\leq m-1})}},\qquad \set{\frac{A(\Theta_{\leq 1})}{\norm{A(\Theta_{\leq 1})}}, \ldots,\frac{A(\Theta_{\leq m-1})}{\norm{A(\Theta_{\leq m-1})}}}.\] The left hand side of relation \eqref{E:components} shows that the $1,\ldots,m-1^{st}$ components of $\Lambda_{k,i}$ depend only $\set{\norm{A(\Theta_{\leq j})},\, j=1,\ldots, m-1}$, whereas the right hand side of \eqref{E:components} shows that the $m^{th}$ component of $\Lambda_{k,i}$ depends only on $A(\Theta_{\leq m-1})/\norm{A(\Theta_{\leq m-1})}$ and on $A\theta_m$. Therefore, the $m^{th}$ component of $\widehat{\Lambda}_{k,i}$ is independent of all the previous components. Proceeding in this way for $m=k,k-1,\ldots,1$, we find that the components of $\widehat{\Lambda}_{k,i}$ are independent. Finally, let us denote by $\Pi_{\leq m-1}$ the orthogonal projection onto the orthogonal complement of the span of $\set{\theta_1,\ldots,\theta_{m-1}}.$ We have by Lemma \ref{lem-Grass-2} that in distribution \[\lr{\widehat{\Lambda}_{k,i}}_m = \log \norm{ \Pi_{\leq m-1}(A\theta_m)}.\] Note that $A\theta_m$ is independent of $\Pi_{\leq m-1}$ since the latter depends only on $A\theta_1,\ldots,A\theta_{m-1}$. Hence, we have the following equality in distribution: \[\lr{\widehat{\Lambda}_{k,i}}_m = \frac{1}{2}\log\lr{ \frac{1}{n}\chi_{n-m+1}^2}.\] This completes the proof of \eqref{E:pointwise-gaussian}. To conclude \eqref{E:first-term}, we apply the multivariate CLT (Theorem \ref{MultiCLT-theo}) to \[\widehat{\Lambda}_k-\E{\widehat{\Lambda}_k}=\sum_{i=1}^N \frac{1}{N}\lr{\widehat{\Lambda}_{k,i} -\mu_{n,\leq k}}.\] Since the covariance matrix of $ (Y_{i,1}, \cdots , Y_{i,k} ) $ is ${\mathrm{ Diag}(\sigma_{n,\leq k }^{2} )}$ by independence we have that $C:= {\rm Cov} ( \widehat{\Lambda}_{K}) := \frac{1}{N} {\rm Diag} ( \sigma_{n,1}^{2}, \cdots , \sigma_{n,k}^{2})$. Recall that $ \beta_{i} := \mathbb E \| C^{-\frac{1}{2}} (\overline{Y}_{i,1}, \cdots , \overline{Y}_{i,k})\|_{2}^{3} $. It is not difficult to check that $ \log\chi_{m}^{2} $ is a $\log$-concave random variable (i.e. its density is a log-concave function). Moreover, since $ \sigma_{n,j}^{-1} \overline{Y}_{i,j}$ have mean zero and variance $1$, $ D:=( \sigma_{i,1}^{-1} \overline{Y}_{i,1}, \cdots, \sigma_{i,k}^{-1} \overline{Y}_{i,k})$ is a log-concave random vector in $\mathbb R^{k}$ with covariance matrix equals to the identity. Therefore $\mathbb E \|D\|_{2}^{2} = k $. It is known that the Euclidean norm of such vectors satisfies a reverse H\"older inequality with a universal constant, and in particular (see e.g. \cite{paouris2006concentration} or [\cite{artstein2015asymptotic} Theorem 10.4.6] for a stronger result) that $$ \left( \mathbb E \| D\|_{2}^{3} \right)^{\frac{1}{3}} \leq C \left( \mathbb E \| D\|_{2}^{2} \right)^{\frac{1}{2}}= C\sqrt{k} , $$ where $C>0$ is an absolute constant. So, $$ \beta_{i} =\frac{1}{N^{\frac{3}{2}}} \mathbb E \| D \|_{2}^{3} \leq \frac{C^{3} k^{\frac{3}{2}} }{N^{\frac{3}{2}}} \ 1\leq i \leq N.$$ Therefore, $$ \beta:= \sum_{j=1}^{N} \beta_{j} \leq \frac{C^{3} k^{\frac{3}{2}} }{N^{\frac{1}{2}}} $$ and we conclude that there exists an absolute constant $c>0$ so that \[d(\widehat{\Lambda}_k, \mathcal N(\mu_{n,\leq k}, \Sigma_{n,k}))\leq c k^{7/4}N^{-1/2}. \] \end{proof} \noindent Having bounded the first term in \eqref{E:KS-terms}, we write \[\lr{T\Sigma_{n,k,N} T^*}^{-1} = \lr{T^*}^{-1}\Sigma_{n,k,N}^{-1}T^{-1}\] and bound the second term using that the matrix $\Sigma$ is diagonal and that $T^{-1}$ a bi-diagonal: \begin{lemma}\label{L:HS-est} There exists $C>0$ so that \begin{equation}\label{E:second-term} \norm{\lr{T\Sigma_{n,k,N} T^*}^{-1}}_{HS}^{1/2}\leq Ck^{1/4}(nN)^{1/2}. \end{equation} \end{lemma} \begin{proof} We have \[T = \lr{\begin{array}{ccccc} 1 & 0 & 0&\cdots & 0 \\ 1 & 1 & 0& \cdots & 0 \\ \vdots & \cdots & \ddots &\ddots &\vdots\\ 1 & \cdots & 1 & 1& 0 \\ 1 & \cdots & 1 & 1& 1 \\\end{array}},\qquad T^{-1}= \lr{\begin{array}{ccccc} 1 & 0 & 0&\cdots & 0 \\ -1 & 1 & 0& \cdots & 0 \\ \vdots & \ddots & \ddots &\ddots &\vdots\\ 0 & \cdots & -1 & 1& 0 \\ 0 & \cdots & 0 & -1& 1 \\ \end{array}}.\] Thus, recalling that \[\Sigma_{n,k,N}= \frac{1}{N}\mathrm{Diag}\lr{\sigma_{n,\leq k}}=\frac{1}{N}\lr{\sigma_{n,1}^{2},\ldots, \sigma_{n,k}^{2}},\] we find \[(T^*)^{-1}\Sigma_{n,k,N}^{-1}T^{-1} = N\lr{\begin{array}{ccccc} \sigma_{n,1}^{-2}+\sigma_{n,2}^{-2} & - \sigma_{n,2}^{-2}& 0&\cdots & 0 \\ -\sigma_{n,2}^{-2} & \sigma_{n,2}^{-2}+\sigma_{n,3}^{-2} & -\sigma_{n,3}^{-2}& \cdots & 0 \\ \vdots & \cdots & \ddots &\ddots &\vdots\\ 0 & \cdots & -\sigma_{n,k-1}^{-2} &\sigma_{n,k-1}^{-2}+\sigma_{n,k}^{-2} & -\sigma_{n,k}^{-2} \\ 0 & \cdots & 0 &-\sigma_{n,k}^{-2} & \sigma_{n,k}^{-2} \\\end{array}}\] Hence, using \eqref{E:sigma-est}, we find that for some $C>0$ \[\norm{(T^*)^{-1}\Sigma_{n,k,N}^{-1}T^{-1}}_{HS} \leq 2N \lr{\sum_{j=1}^k \sigma_{n,k,j}^{-4}}^{1/2}\leq CN\lr{\sum_{j=1}^k (n-k+1)^2}^{1/2}\leq CNnk^{1/2},\] and Lemma \ref{L:HS-est} follows. \end{proof} \noindent Thus far, combining the previous two Lemma with \eqref{E:KS-terms}, we've shown that \begin{align} \label{E:first-two}d\lr{\Lambda_k, \mathcal N\lr{\mu_{n,k}, \Sigma_{n,k}}} &\leq \frac{Ck^{7/4}}{N^{1/2}} +c_0\delta k^{1/4}(nN)^{1/2}+ 2 \mathbb P\lr{\norm{S_k - \widehat{S}_k}>\delta}. \end{align} So it remains to estimate \[ \mathbb P\lr{\norm{S_k - \widehat{S}_k}_2\geq\delta} \] and optimize over $\delta$. To do this, write $S_{k,j}, \widehat{S}_{k,j}$ for the $j^{th}$ components of $S_k,\widehat{S}_k$. By \eqref{conc-by-sb}, there exists $C>0$ so that for $ 1\leq j \leq k \leq n$, \begin{equation*} \mathbb P \left( | S_{k,j} -\widehat{S}_{k,j}|\geq s \right) \leq 2e^{ - sN/4 } ,\qquad s\geq C\frac{ j}{N} \log\lr{ \frac{en}{j}}. \end{equation*} For any collection positive real numbers $\delta_j>C\frac{j}{N}\log\lr{\frac{en}{j}}$ we therefore have, \[ \mathbb P\lr{\norm{S_k - \widehat{S}_k}_2\geq\lr{\sum_{j=1}^k \delta_j^2}^{1/2}}\leq \sum_{j=1}^k \mathbb P \left( | S_{k,j} -\widehat{S}_{k,j}|\geq \delta_j \right) \leq 2\sum_{j=1}^ke^{-\delta_j N/4 }. \] Setting \[ \delta_j:=\frac{Cj}{N}\log\lr{\frac{en}{j}} \log\lr{\frac{N}{n}}, \] for a sufficiently large constant $C$ we find \[ \mathbb P\lr{\abs{S_{k,j} - \widehat{S}_{k,j}}\geq\delta_j}\leq 2e^{-Cj\log(en /j )\log(N/n)}\leq 2(n/N)^{j/2}. \] Hence, as soon as $N> n$, we have \[ \mathbb P\lr{\norm{S_k - \widehat{S}_k}_2\geq\delta}\leq C\lr{\frac{n}{N}}^{1/2} \] where \[ \delta :=\lr{\sum_{j=1}^{k}\delta_j^2}^{1/2} \leq \frac{C k^{3/2}\log(n)\log(N/n)}{N}. \] In conjunction with \eqref{E:first-two} yields \begin{align*} d\lr{\Lambda_k, \mathcal N\lr{\mu_{n,k}, \Sigma_{n,k}}} &\leq \frac{Ck^{7/4}}{N^{1/2}} +\lr{\frac{Ck^{7/2}n \log^2(n)\log^2(N/n)}{N}}^{1/2} + C\lr{\frac{n}{N}}^{1/2}\\ &\leq \lr{\frac{4Ck^{7/2}n \log^2(n)\log^2(N/n)}{N}}^{1/2}, \end{align*} as claimed. \bibliographystyle{alpha}

https://arxiv.org/abs/1808.04163

Approximation in FEM, DG and IGA: A Theoretical Comparison

In this paper we compare approximation properties of degree $p$ spline spaces with different numbers of continuous derivatives. We prove that, for a given space dimension, $\smooth {p-1}$ splines provide better a priori error bounds for the approximation of functions in $H^{p+1}(0,1)$. Our result holds for all practically interesting cases when comparing $\smooth {p-1}$ splines with $\smooth {-1}$ (discontinuous) splines. When comparing $\smooth {p-1}$ splines with $\smooth 0$ splines our proof covers almost all cases for $p\ge 3$, but we can not conclude anything for $p=2$. The results are generalized to the approximation of functions in $H^{q+1}(0,1)$ for $q<p$, to broken Sobolev spaces and to tensor product spaces.

\section{Introduction} The aim of this work is to compare the approximation properties of different piecewise polynomial spaces commonly employed in Galerkin methods for PDEs. Following the well known Lemmas of C\'ea and Strang such approximation results imply a priori error estimates for these numerical methods. In particular we consider the tensor product spaces used in Discontinuous Galerkin (DG), Finite Element Methods (FEM) and IsoGeometric Analysis (IGA) that differ only in their global smoothness. As such our comparison provides an answer to the following question: ``does smoothness impede or favour approximation?''. It was noticed by Hughes, Cottrell and Bazilevs \cite{Hughes:2005} that smoother spaces have better approximation properties in their numerical tests. Spline approximation in the IGA setting was first studied by Bazilevs, Beir\~ao da Veiga, Cottrell, Hughes and Sangalli \cite{MR2250029}. Later, Evans, Bazilevs, Babuska and Hughes \cite{Evans:2009} numerically computed approximation constants and observed that the maximally smooth splines provide better a priori bounds on the approximation error. Beir\~ao da Veiga, Buffa, Sangalli and Rivas \cite{MR2800710} studied how the approximation depends on the mesh-size, the degree and the global smoothness. Takacs and Takacs \cite{Takacs:2016} proved an upper bound for the approximation error with an explicit constant. Recently, Floater and Sande \cite{Floater:2017,Floater:2018} provided optimal constants on which we base our results. The comparison in this paper is related to the $n$-width theory \cite{Kolmogorov:36,Pinkus:85}, i.e., looking at approximation properties of a space of fixed dimension. Our results can be seen as a partial answer to an $n$-width problem constrained to piecewise polynomial spaces on uniform partitions. We will first look at the univariate setting before extending the results to general tensor product spaces. Let $\poly p$ be the space of polynomials of degree at most $p$, and $L^2=L^2(0,1)$ and $H^{q+1}=H^{q+1}(0,1)$ be the standard Sobolev spaces. For a given $n\in \mathbb{N}$ let $I_j$ be the interval $[\frac jn,\frac{j+1}n)$ and define the spline space $\spline p k n$, of degree $p$, smoothness $k$ and on $n$ segments by \begin{equation}\label{eq:def-spline} \spline p k n =\big\{f \in\smooth k([0,1]): f|_{I_j} \in\poly p,\, j=0,\dots,n-1\big\}. \end{equation} Here $k=-1$ means that jumps are allowed at the internal breakpoints. Furthermore, let $\proj pkn$ be the $L^2$ projection onto $\spline pkn$ and $\cf pknq$ be the smallest real number such that \begin{equation}\label{eq:estimate-form} \norm{f- \proj pkn f}\le \cf pknq \norm{\partial^{q+1} f} \end{equation} holds for all $f\in H^{q+1}$. Here $\norm{\cdot}$ denotes the $L^2$ norm. Finally for $q=p$ we let $\const pkn:=\cf pknp$. The studied $n$-width problem can then be formulated as follows. Given the space dimension $N$ and Sobolev regularity $q$, find the degree $p$, smoothness $k$, and number of segments $n$ such that the constant $\cf pknq$ is minimized. Note that for each $N$ only finitely many $(p,k,n)$ fulfill \begin{equation}\label{eq:spline-dim} N=\dim\spline p k n =(n-1)(p-k)+p+1. \end{equation} It is then possible to try an exhaustive approach. The difficulty of such a strategy is that the constants $\cf pknq$ are solutions of eigenvalue problems that are badly conditioned. Any conclusion based on this strategy would then have to take into consideration the reliability of the method used to compute the constant. Our approach is to first provide lower and upper bounds for $\const pkn$ and base the conclusions on provable properties of these bounds. In particular we compare $\const p{p-1}m$ with $\const pkn$ for $k<p-1$ under the constraint $$\dim\spline {p} {p-1} {m} = \dim\spline {p} {k} {n}, $$ i.e., for \begin{equation}\label{eq:m-formula} m=(n-1)(p-k)+1. \end{equation} Note that for a fixed number of segments $n$ we have $\spline p{k_1}n \supseteq \spline p{k_2}n$ whenever $k_1\le k_2$ so that necessarily $\const p{k_1}n \le \const p{k_2}n$ under the same condition. However, for a fixed dimension $N$ the smoother space is defined on a finer partition. We show in Section \ref{sec:comparison-1d} that $\const p{p-1}m$ is smaller than $\const pkn$ in almost all cases of practical interest for $k=0$ (see Theorem~\ref{thm:fem}) and $k=-1$ (see Theorem~\ref{thm:dg}). In Section~\ref{sec:low} we extend our results to the case of $p>q$ in \eqref{eq:estimate-form}. Here we compare the approximation of maximally smooth spline spaces of a ``too high degree'', $\spline {p} {p-1} {m}$, with spline spaces of lower smoothness, $\spline {q} {k} {n}$, under the constraint $\dim\spline {p} {p-1} {m} = \dim\spline {q} {k} {n}$. The main result of this section is contained in Theorem~\ref{thm:lower}. A comparison in the case of Broken Sobolev spaces is then performed in Section~\ref{sec:broken} and extensions to tensor product cases are considered in Section~\ref{sec:tens}. The fact that smoother spaces provide better approximation could be surprising to people not familiar with the $n$-width theory, indeed it could seem reasonable that smoother spaces are more ``rigid'' and thus that they can not approximate functions that are less regular. This is not true: for instance it was shown by Kolmogorov \cite{Kolmogorov:36} that $$ \SPAN \{1,\cos(\pi x),\ldots,\cos((N-1)\pi x)\} $$ is optimal for $H^1$ in the $n$-width sense, meaning that no other $N$-dimensional subspace of $L^2$ can provide a better a priori error estimate for $H^1$ functions. Based on results of Melkman and Micchelli \cite{Melkman:78} it was proved in \cite{Floater:2017} that for all $q$ and $N$ there exists an optimal $\smooth {\ell (q+1)-2}$ spline space of degree $\ell (q+1)-1$ for any $\ell=1,2,\dots$. In fact, for $q=0$ and with even degrees $p$, the knots of the optimal spline spaces are uniform and so they are subspaces of $\spline p{p-1}n$. \section{Upper and lower bounds for $\const pkn$}\label{sec:univ} We prove the following bounds on the best constants $\const {p}{k}n$, $k=-1,0, \ldots, p-1$. \begin{theorem}\label{thm:bounds} For all $p\ge 0$, $k\in\{-1,0,\ldots,p-1\}$ and $n\ge 1$ we have \begin{align} \label{est:-1} \frac{(p+1)!}{(2p+2)!\sqrt{2p+3}} n^{-p-1}\le &\const {p}{k}n \le (n\pi)^{-p-1} \end{align} \end{theorem} The above inequalities are shown in the following lemmas. \begin{lemma}\label{lem:legendre-poly} For discontinuous spline approximation we have $$ \frac{(p+1)!}{(2p+2)!\sqrt{2p+3}}n^{-p-1} \le \const p{-1}n. $$ \end{lemma} \begin{proof} It is enough to show that there exists an $f\in H^{p+1}$ such that $$ \norm{f-\proj p{-1}n f}\ge \frac{(p+1)!}{(2p+2)!\sqrt{2p+3}} n^{-p-1} \norm{\partial^{p+1}f}. $$ This is the case for $f(x)=x^{p+1}$. The projection $\proj p{-1}n$ acts independently on each $I_j$, $j=0,\dots,n-1$ and on $I_j$ we have $$ x^{p+1}= \sum_{i=0}^{p+1} c_{i,j}\,\ell_i (n x-j ) ,$$ where $\ell_i$ is the $i$-th Legendre polynomial on $[0,1]$. Since $\ell_{p+1}(nx-j)$ is orthogonal to the polynomials of degree $p$ on $I_j$ we have $$x^{p+1} -\proj p{-1}n x^{p+1} = \sum_{j=0}^{n-1} c_{p+1,j} \ell_{p+1}( n x-j ) . $$ Since $$ \norm{\ell_{p+1}}=(\sqrt{2p+3})^{-1}\qquad \text{and}\qquad \norm{\partial^{p+1}\ell_{p+1}}=\frac{(2p+2)!}{(p+1)!}. $$ by taking the derivative of $\ell_{p+1}( n x-j )$ we obtain $$ \norm{x^{p+1}-\proj p{-1}n x^{p+1}}_{I_j}= \frac{(p+1)!}{(2p+2)!\sqrt{2p+3}}n^{-p-1} \norm{\partial^{p+1} x^{p+1}}_{I_j}. $$ Summing over $j$ the squares of the left and right hand sides yields the result. \end{proof} \begin{lemma}\label{lem:univk} For maximally smooth spline approximation we have $$\const {p}{p-1}n \le (n\pi)^{-p-1}.$$ \end{lemma} \begin{proof} This is a corollary of the results in \cite{Floater:2018}. Let $$E^{p}=\{f\in H^{p+1}:\quad \partial^{s} f(0)=\partial^{s} f(1)=0,\quad 0\leq s< p,\quad s+p\text{ is odd} \}.$$ Observe that for $p$ odd, $E^p$ coincides with the non-standard Sobolev space $H^{p+1}_0$ defined in \cite{Floater:2018}, and for $p$ even it coincides with the space $H^{p+1}_1$ in that paper. Then \cite[Theorems 1 and 2]{Floater:2018} states that for all $v\in E^p$ \begin{equation}\label{eq:Floater-result} \norm{v-\Pi_E v}\le \Big(\frac{1}{n\pi}\Big)^{p+1}\norm{\partial^{p+1}v}, \end{equation} where $\Pi_E :E^p\rightarrow E^p\cap \spline p{p-1}n$ is the $L^2$ projection. Note that the $n$ in \cite{Floater:2018} is the space dimension, what we call $N$, and not the number of segments. Given $f\in H^{p+1}$ let $g\in \poly p$ be a polynomial such that $f-g\in E^p$. In other words, for $0\le s < p$ with $s+p$ odd, we have $$ \left\{\begin{aligned} \partial^s g(0)&=\partial^s f(0)\\ \partial^s g(1)&=\partial^s f(1). \end{aligned}\right . $$ This $g$ exists according to Lemmas \ref{lem:poly-interp} and \ref{lem:poly-interp2} in the appendix. Then, since $g\in\spline p{p-1}n$ and $f-g\in E^p$ we have $$ \begin{aligned} \norm{f-\proj p{p-1}n f}&= \norm{(f-g) - \proj p{p-1}n (f-g)}\\ &\le \norm{(f-g) - \Pi_E (f-g)}\\ &\le \Big(\frac{1}{n\pi}\Big)^{p+1}\norm{\partial^{p+1}(f-g)}\\ &= \Big(\frac{1}{n\pi}\Big)^{p+1}\norm{\partial^{p+1}f}. \end{aligned} $$ \end{proof} Theorem \ref{thm:bounds} now follows from the observation that $\spline p {k+1}n\subset\spline p {k}n$ for all $k=-1,0,\ldots,p-2$ and so $\const pkn\le \const p{k+1}n$. \section{Univariate comparisons} \label{sec:comparison-1d} Here we compare the space of maximally smooth splines, $\spline p{p-1}m$, commonly used in IGA, with the space $\spline pkn$ of smoothness $k<p-1$ where $m$ depends on $n$ as in \eqref{eq:m-formula}, i.e., such that $\dim\spline p{p-1}m=\dim\spline pkn$. This means that the smoother space is defined on a finer grid. See Figure \ref{fig:deg1-3} for an example of this. Note that the case $k=p-1$ and the case $n=1$ are uninteresting since we would then be comparing $\spline pkn$ with itself. The estimates in Section \ref{sec:univ} are sharp enough to prove that smooth splines will eventually provide a better approximation in the number of degrees of freedom. This is stated in the following theorem. More precise statements for the IGA-FEM comparison ($k=0$) and the IGA-DG comparison ($k=-1$) are contained in subsections 3.1 and 3.2. \begin{figure} \includegraphics{deg1.pdf} \includegraphics{deg3.pdf \caption{Example of pairs of functions in $\spline p{-1}n$ (blue) and $\spline p{p-1}m$ (red) for $p=1$ and $p=3$, $n=3$. Note how the maximally smooth spline space is defined on a finer grid.}\label{fig:deg1-3} \end{figure} \begin{theorem}\label{thm:comparison-general} For all $k\geq -1$ and $n\ge 2$ there exists $\bar p$ such that for all $p\ge \bar p$ $$ \const p{p-1}m<\const pkn, $$ where $m=(n-1)(p-k)+1$. \end{theorem} This theorem follows from studying the bounds in Theorem \ref{thm:bounds}, which is done in Lemma \ref{thm:ratio} and Proposition \ref{prop:B}. \begin{lemma}\label{thm:ratio} For all $p\ge 0$, $k\in\{-1,\dots,p-1\}$ and $n,m\ge 1$ we have \begin{equation}\label{eq:ratio} \frac{\const p{p-1}m}{\const pkn} \le \Big(\frac{4}{e\pi}\Big)^{p+1}\Big(\frac{n}{m}\Big)^{p+1} (p+1)^{p+1}\sqrt{4p+6}. \end{equation} \end{lemma} \begin{proof} From Theorem \ref{thm:bounds} we have for $k=-1,\dots,p-1$, that \begin{equation*} \frac{\const p{p-1} m}{\const pkn} \le \Big(\frac{n}{m\pi}\Big)^{p+1} \frac{(2p+2)!}{(p+1)!}\sqrt{2p+3}. \end{equation*} Now, using the error bounds of the Stirling's approximation~\cite{Robbins:55} \begin{equation}\label{ineq:Stirling} \sqrt{2\pi}r^{r+\frac{1}{2}}e^{-r}e^{\frac{1}{12r+1}}\le r! \le \sqrt{2\pi}r^{r+\frac{1}{2}}e^{-r}e^{\frac{1}{12r}}, \end{equation} we find that \begin{align*} \frac{(2p+2)!}{(p+1)!} &\le \frac{\sqrt{2\pi}(2p+2)^{2(p+1)+\frac{1}{2}}e^{-2(p+1)}e^{\frac{1}{12(2p+2)}}}{\sqrt{2\pi}(p+1)^{p+1+\frac{1}{2}}e^{-p-1}e^{\frac{1}{12(p+1)+1}}},\\ &=2^{2(p+1)+\frac{1}{2}} \frac{(p+1)^{2(p+1)+\frac{1}{2}}}{(p+1)^{p+1+\frac{1}{2}}} \frac{e^{-2(p+1)}}{e^{-p-1}} \frac{e^{\frac{1}{12(2p+2)}}}{e^{\frac{1}{12(p+1)+1}}},\\ &=4^{p+1}\sqrt{2}(p+1)^{p+1} e^{-p-1} e^{\frac{1}{24(p+1)}-\frac{1}{12(p+1)+1}},\\ &=\Big(\frac{4}{e}\Big)^{p+1}\sqrt{2}(p+1)^{p+1}e^{\frac{1-12(p+1)}{24(p+1)(1+12(p+1))}},\\ &\le \Big(\frac{4}{e}\Big)^{p+1}\sqrt{2}(p+1)^{p+1}, \end{align*} and the result follows. \end{proof} For $m$ as in \eqref{eq:m-formula}, we let $\rt pnk$ be the bound in \eqref{eq:ratio}, now given as \begin{align}\label{eq:ratio-m} \rt pkn=(\B pkn)^{p+1} \sqrt{4p+6}\qquad\text{with}\\ \label{eq:base-m} \B pkn = \frac{4}{e\pi} \frac{n (p+1)}{(p-k)(n-1)+1} . \end{align} The next step of our analysis is the study of $\B pkn$. \begin{proposition}\label{prop:B} For $-1\leq k <p-1$ and $n\ge 2$ the following statements hold \begin{enumerate} \item \label{p-lim} for fixed $n$ and $k $ $$\lim_{p\rightarrow\infty} B_{p,k,n}= \frac {4 } {e\pi}\frac n{n-1}<1;$$ \item \label{n-lim} for fixed $p$ and $k$, $$\lim_{n\rightarrow\infty} B_{p,k,n}= \frac {4 } {e\pi }\frac{p+1}{p-k};$$ \item\label{n-decreasing} $B_{p,k,n}$ is strictly decreasing in $n$; \item\label{p-decreasing} $B_{p,k,n}$ is decreasing in $p$ for $k\ge 0$. \end{enumerate} \end{proposition} \begin{proof} The limits in points \ref{p-lim} and \ref{n-lim} are straightforward. For \ref{n-decreasing} it is sufficient to show that $\B p{n+1}k<\B {p}nk$, i.e., \begin{align*} &\frac{n+1}{(p-k)n+ 1}<\frac{n}{(p-k)(n-1) + 1}, \end{align*} which is equivalent to $k<p-1$, since the denominators are positive. For \ref{p-decreasing} we prove that $\B {p+1}nk \leq \B pnk$, i.e., \begin{align*} \frac{p+2}{(p-k+1)(n-1)+1} \leq \frac{p+1}{(p-k)(n-1)+1}. \end{align*} This is equivalent to $(k+1)(n-1)\geq 1$, which holds for $n\geq 2$ and $k\geq 0$. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:comparison-general}] From point \ref{p-lim} of Proposition \ref{prop:B} we deduce that for $p>\hat p$ we have $\B pkn \le t<1$ and $$ \rt pkn = (\B pkn)^{p+1}\sqrt{4p+6}\le t^{p+1}\sqrt{4p+6}. $$ Thus there exists $\bar p\ge \hat p$ such that for all $p>\bar p$, $\rt pkn <1$. \end{proof} \begin{remark}\label{rem:exp} Using Proposition \ref{prop:B} we can obtain an estimate of how much better the approximation with smooth splines is in Theorem \ref{thm:comparison-general}. For a fixed $k$, and any $\bar p,\bar n$ satisfying $\B {\bar p}k{\bar n}\leq \frac{4}{e\pi}\gamma<1$ we have \begin{equation}\label{ineq:exp} \rt pkn \leq \Big ( \frac 4{e\pi}\gamma \Big)^{p+1} \sqrt{4p+6},\qquad \forall n\geq\bar n, \, \forall p\geq\bar p, \end{equation} i.e. $\rt pkn$ gets exponentially smaller as $p$ increases. The level set $\B pkn=\frac{4}{e\pi}\gamma$ is the hyperbola \begin{align*} 0 = \big(n -\frac {\gamma}{\gamma-1} \big)\big(p-\frac{\gamma k+1}{\gamma-1} \big) + \frac{\gamma(\gamma-k-2)}{(\gamma-1)^2} \end{align*} and has asymptotes \begin{align*} p= \frac{\gamma k+1}{\gamma-1}, \qquad n=\frac \gamma{\gamma-1}. \end{align*} This tells us that for each $\bar n\geq \frac \gamma{\gamma-1}$ there is a corresponding $\bar p$ such that we obtain the exponential improvement in \eqref{ineq:exp}. \end{remark} \begin{corollary}\label{thm:decreasing-n} For all $p\ge1$ and $k=-1,\dots,p-2$, the ratio $\rt pkn$ in \eqref{eq:ratio-m} is strictly decreasing in $n$. \end{corollary} \begin{proof} By definition $$ \rt pkn=(\B pkn)^{p+1} \sqrt{4p+6} $$ and $\B pkn^{p+1}$ is strictly decreasing in $n$ by point \ref{n-decreasing} of Proposition \ref{prop:B}. \end{proof} \begin{corollary}\label{thm:decreasing-p} For all $k\ge 0$, $\rt pkn$ is strictly decreasing in $p$ for all $p\ge \bar p$ where $\bar p$ is such that $\rt {\bar p}kn\le1$. \end{corollary} \begin{proof} From point \ref{p-decreasing} of Proposition \ref{prop:B}, $\B pkn$ is decreasing in $p$. Moreover, $(4p+6)^{1/(2p+2)}$ is strictly decreasing in $p$. Thus $(\rt pkn)^{1/(p+1)}$ is also strictly decreasing in $p$ and the result follows. \end{proof} \begin{remark}\label{rem:quarter} For fixed $k\geq 0$ and given $\bar p$ and $\bar n$ such that $\rt{\bar p}k{\bar n}<1$ then from the above corollaries we find that $$\const p{ p-1} m < \const p{k}{ n},\qquad \forall p\ge \bar p,\quad \forall n \ge \bar n.$$ This means that there cannot be isolated values for which this inequality holds. A similar result is true for $k=-1$, but it requires a more technical argument that is postponed until later. \end{remark} \subsection{IGA-FEM comparison} \begin{theorem}[IGA-FEM comparison]\label{thm:fem} For $p\geq 3$ and sufficiently large $n$, more precisely $$ \begin{aligned} &n\geq 7 &&\text{for} &&p=3, \\&n\geq 4 &&\text{for} &&p\in\{4,5\}, \\&n\geq 3 &&\text{for} &&p\in\{6,...,37\}, \\&n\geq 2 &&\text{for} &&p\ge 38, \end{aligned} $$ we have $$ \const p{p-1}m<\const p0n. $$ \end{theorem} Note that for $n=1$ or $p=0$ the spaces are the same and $\const p{p-1}m=\const p0n$. Note further that no conclusion can be drawn for $p=2$. Indeed we have $$ \rt 20n=\Big(\frac {4}{e\pi}\frac{3n}{2n-1}\Big)^3 \sqrt{14} >\Big(\frac {6}{e\pi}\Big)^3 \sqrt{14}> 1, \quad \forall n\ge 2. $$ All cases are summarized in Fig.~\ref{fig:iga-fem}. \begin{proof} Using Corollary \ref{thm:decreasing-n} and Corollary \ref{thm:decreasing-p} as explained in Remark \ref{rem:quarter} it is enough to show that $\rt {38}0{2}$, $\rt 603$, $\rt 404$ and $\rt 307$ are less than $1$. We have \begin{align*} \rt {38}02 &= \Big(\frac8{e\pi }\Big)^{39}\sqrt{158} = 0.9851\ldots\\ \rt 603&= \Big(\frac4{e\pi }\frac {21}{13} \Big)^{7}\sqrt{30} = 0.7776\ldots\\ \rt 404&= \Big(\frac4{e\pi }\frac {20}{13} \Big)^{5}\sqrt{22} = 0.9114\ldots\\ \rt 307&= \Big(\frac4{e\pi }\frac {28}{19} \Big)^{4}\sqrt{18} = 0.9632\ldots \end{align*} \end{proof} \begin{figure} \begin{center} \begin{tikzpicture} \fill[purple!15] (0,0)--(0,1)--(1,1)--(1,7)--(2,7)--(2,1) --(9.5,1)--(9.5,0); \fill[white] (0,1) rectangle (1,6.8); \fill[blue!50] (3,6)--(4,6)--(4,3)--(6,3)--(6,2)--(8.5,2)--(8.5,1)--(9.5,1)--(9.5,7)--(3,7); \fill[red!50] (2,1) -- (8.5,1)--(8.5,2)--(6,2)--(6,3)--(4,3)--(4,6)--(3,6)--(3,7)--(2,7); \draw[gray] (0,0) grid (7,7); \begin{scope}[xshift=7.5cm] \draw[gray] (-.9,0) grid (2,7); \end{scope} \draw[->] (-.1,0)--(9.6,0) node[below] {$p$}; \draw[->] (0,-.1)--(0,7.1) node[left] {$n$}; \node at (4,.5) {\large SAME SPACE}; \node[rotate=-90] at (.5,3) {\large FEM=$\mathbb{R}$}; \foreach \x[evaluate=\x as \xp using \x+.1] in {7.1,7.3,7.5} { \fill[white] (\x,-.1) rectangle (\xp,7.1); } \foreach \p[evaluate=\p as \l using \p+.5] in {0,1,...,6} {\draw (\l,.1)--(\l,-.1) node[below] {$\p$};} \foreach \n[evaluate=\n as \l using \n-.5] in {1,2,...,7} {\draw (.1,\l)--(-.1,\l) node[left] {$\n$};} \foreach \p/\l in {37/8, 38/9} {\draw (\l,.1)--(\l,-.1) node[below] {$\p$};} \node at (5.5,4.5) {\large $\displaystyle\frac {\const p{p-1}m}{\const p0n}<1$}; \node at (3.5,6.5) {\Large $\color{green}\checkmark$}; \node at (4.5,3.5) {\Large $\color{green}\checkmark$}; \node at (6.5,2.5) {\Large $\color{green}\checkmark$}; \node at (9,1.5) {\Large $\color{green}\checkmark$}; \end{tikzpicture} \end{center} \caption{The blue area indicates for which $p$ and $n$ we can conclude that IGA approximation is better than FEM approximation. The red area indicates where no conclusion can be obtained from the estimate. The two spaces coincide in the pink area.}\label{fig:iga-fem} \end{figure} \subsection{IGA-DG comparison}\label{sec:dg} Similarly to the previous subsection we note that for $n=1$ or $p=0$ the spaces are the same and $\const p{p-1}m=\const p{-1}n$. \begin{lemma}\label{thm:decreasing-p-1} For $n=2$, $p\ge 22$ and $n=3$, $p\ge 2$ the function $\rt p{-1}n$ is strictly decreasing in $p$. \end{lemma} \begin{proof} First note that $\rt p{-1}n$ is decreasing whenever $\rt p{-1}n^2$ is decreasing. We now let $s=p+1$ and compute the derivative of $\rt {s-1}{-1}n^2$ with respect to $s$ and show that it is negative. $$ \begin{aligned} \partial_s&( \rt {s-1}{-1}n)^2= \underbrace{\frac 4{ns-s+1}\Big(\frac 4{e\pi} \frac{ns}{ns-s+1}\Big )^{2s} }_{> 0}\\ &\Big( 1-2 s^2 (n-1) + \underbrace{(1+2s)(ns-s+1)}_{>0} \underbrace{\ln \big(\frac 4{\pi} \frac {ns}{ns-s+1} \big)}_{\leq L} \Big), \end{aligned} $$ where $L=\ln \big(\frac 4{\pi} \frac {n}{n-1} \big)<1$ is an upper bound on the logarithm. It follows that $\partial_s \rt {s-1}{-1}n <0$ if $$ 2(n-1)(L-1)s^2 + (n+1)L s + (L+1)<0, $$ i.e., for $$ s> \frac { -(n+1)L -\sqrt{(n+1)^2L^2-8(L^2-1)(n-1)} }{4(n-1)(L-1)}. $$ For $n=2$ we have $L=\ln\frac 8\pi < 0.935$ and $\rt p{-1}2$ is strictly decreasing for $$ p=s-1\ge \frac {3L+\sqrt{L^2+8}}{4(1-L)}-1\approx 21.14\ldots. $$ For $n=3$ we have $L=\ln\frac 6\pi < 0.648$ and $\rt p{-1}3$ is strictly decreasing for $$ p=s-1\ge\frac12\frac{1+L}{1-L}-1\approx 1.33\ldots. $$ \end{proof} \begin{theorem}[IGA-DG comparison]\label{thm:dg} For$$ \begin{aligned} &n\ge 3 &&\text{and}&&p\in\{1,\dots,17\}, \\&n\ge 2 &&\text{and}&&p\ge 18, \end{aligned}$$ we have $$ \const p{p-1}m<\const p{-1}n. $$ \end{theorem} \begin{proof} Using Lemma \ref{thm:decreasing-n} and Lemma \ref{thm:decreasing-p-1} it is enough to show that $\rt 1{-1}3$, $\rt 2{-1}3$ and $\rt {22}{-1}2$ are less than $1$ to cover all cases but $\rt {18}{-1}2$, $\rt {19}{-1}2$, $\rt {20}{-1}2$, $\rt {21}{-1}2$. The latter are also checked. We have \begin{align*} \rt 1{-1}3&=\Big(\frac4{e\pi }\frac 65 \Big)^{2}\sqrt{10} = 0.9990\ldots\\ \rt 2{-1}3&=\Big(\frac {4}{e\pi }\frac 97 \Big)^{3}\sqrt{14} = 0.8172\ldots\\ \rt {18}{-1}2&=\Big(\frac4{e\pi }\frac {19}{10} \Big)^{19}\sqrt{ 78} = 0.9639 \ldots\\ \rt {19}{-1}2&=\Big(\frac4{e\pi }\frac {40}{21} \Big)^{20}\sqrt{ 82} = 0.9247 \ldots\\ \rt {20}{-1}2&=\Big(\frac4{e\pi }\frac {21}{11} \Big)^{21}\sqrt{ 86} = 0.8862 \ldots\\ \rt {21}{-1}2&=\Big(\frac4{e\pi }\frac {44}{23} \Big)^{22}\sqrt{ 90} = 0.8484 \ldots\\ \rt {22}{-1}2&=\Big(\frac4{e\pi }\frac {23}{12} \Big)^{23}\sqrt{ 94} = 0.8115 \ldots \end{align*} \end{proof} Note that nothing can be concluded for $n=2$ and $p\in\{1,\dots,17\}$ since the estimate $\rt p{-1}2>1$ in these cases. All cases are summarized in Fig.~\ref{fig:iga-dg}. \begin{figure} \begin{center} \begin{tikzpicture} \fill[purple!15] (0,0)--(0,7)--(1,7)--(1,1)--(12,1)--(12,0); \fill[blue!50] (1,2)--(6,2)--(6,1)--(12,1)--(12,7)--(1,7); \fill[red!50] (1,1) rectangle(6,2); \draw[gray] (0,0) grid (12,7); \draw[->] (-.1,0)--(12.1,0) node[below] {$p$}; \draw[->] (0,-.1)--(0,7.1) node[left] {$n$}; \node at (2,.5) {\large SAME SPACE}; \node at (8.5,4) {\large $\displaystyle\frac {\const p{p-1}m}{\const p{-1}n}<1$}; \foreach \p[evaluate=\p as \l using \p+.5] in {0,1,...,3} {\draw (\l,.1)--(\l,-.1) node[below] {$\p$};} \foreach \n[evaluate=\n as \l using \n-.5] in {1,2,...,6} {\draw (.1,\l)--(-.1,\l) node[left] {$\n$};} \foreach \x[evaluate=\x as \xp using \x+.1] in {4.3,4.5,4.7} { \fill[white] (\x,-.1) rectangle (\xp,7.1); } \foreach \p/\l in {17/5.5,18/6.5,19/7.5,20/8.5,21/9.5,22/10.5,23/11.5} {\draw (\l,.1)--(\l,-.1) node[below] {$\p$};} \node at (1.5,2.5) {\Large $\color{green}\checkmark$}; \node at (2.5,2.5) {\Large $\color{green}\checkmark$}; \node at (6.5,1.5) {\Large $\color{green}\checkmark$}; \node at (7.5,1.5) {\Large $\color{green}\checkmark$}; \node at (8.5,1.5) {\Large $\color{green}\checkmark$}; \node at (9.5,1.5) {\Large $\color{green}\checkmark$}; \node at (10.5,1.5) {\Large $\color{green}\checkmark$}; \end{tikzpicture} \end{center} \caption{The blue area indicates for which $p$ and $n$ we can conclude that IGA approximation is better than DG approximation. The red area indicates where no conclusion can be obtained from the estimate. The two spaces coincide in the pink area.}\label{fig:iga-dg} \end{figure} \section{Lower order Sobolev spaces} \label{sec:low} In this section we consider an approximand $f$ in $H^{q+1}$ and compare the approximation by smooth splines of degree $p>q$, $\spline p{p-1}m$, with that by $\smooth k$ splines of degree $q$, $\spline qkn$. Both spaces provide the same approximation order, but the smoother space has a degree higher than the regularity of the approximand. In IGA the degree of the spline space is sometimes determined by the parametrisation of the domain, and not by the Sobolev regularity of the solution. Our aim is to show that, for practical purposes, smooth spline spaces of degree higher than the Sobolev regularity have better approximation properties than lower smoothness spaces of the optimal degree. Recalling that $\cf pknq$, $0\le q\le p$ is the best constant in $$ \norm {f-\proj pkn f}\le \cf pknq \norm{\partial^{q+1}f}, $$ we compare $\cf p{p-1} m q$ with $\const q kn $ under the constraint $\dim \spline p{p-1}m = \dim \spline qkn,$ which corresponds to \begin{equation}\label{eq:m-formula-low} m=(q-k)(n-1)+1+q-p. \end{equation} \begin{theorem}\label{thm:lower-norm} For all $0\le q \le p$, $k\in\{-1,0,\ldots,p-1\}$ and $n\ge 1$ we have $$ \cf pknq \le \Big(\frac1{n\pi}\Big)^{q+1}. $$ \end{theorem} The proof is done only for $k=p-1$ and using induction starting from $q=0$. The base case is proved in the following lemma. \begin{lemma}\label{lem:ineqH1} For all $p\geq 0$ and $n\geq 1$ we have \begin{equation}\label{ineq:H1} \cf p{p-1}n0 \le \frac1{n\pi}. \end{equation} \end{lemma} \begin{proof} If $p$ is even, then this follows directly from \cite[Theorem~2]{Floater:2018} (originally shown in \cite[Theorem~2]{Floater:2017}) where it is proved for a subspace of $\spline p{p-1}{n}$. If $p$ is odd, \cite[Theorem~2]{Floater:2018} states approximation results for splines on a different partition. We obtain the desired result by extending the domain to $$ \widetilde I=(-\frac{1}{2n},1+\frac{1}{2n}) $$ and considering the spaces \begin{align*} \widetilde E&=\{ f\in \smooth {p-1}(\widetilde I) : \partial^s f(-\frac 1{2n} )= \partial^s f(1+\frac 1{2n} )=0, \ 1\le s\le p,\,s\text{ odd} \}\\ \widetilde{\mathcal S}&=\{f\in \widetilde E : f|_{[-\frac 1{2n},0)},\, f|_{[1,1+\frac 1{2n})},\,f|_{I_j}\in\poly p, \ j=0,\ldots,n-1 \} \end{align*} where we recall $I_j=[\frac jn,\frac{j+1}n)$. Note that $\spline pkn$ is the restriction of $\widetilde{\mathcal S}$ to $[0,1]$ and that $\dim \widetilde {\mathcal S}=n+1$. Furthermore let $\widetilde \Pi: L^2(\widetilde I) \rightarrow \widetilde {\mathcal S} $ be the orthogonal projection and $\mathcal{E}: H^1(I)\rightarrow H^1(\widetilde I)$ be the extension operator $$ \mathcal{E} f(x) =\begin{cases} f(0)& x\le 0,\\ f(x)& 0<x\le 1,\\ f(1)& x>1. \end{cases} $$ Then using \cite[Theorem~2]{Floater:2018} on $\widetilde I$ we get \begin{align*} \norm{f-\proj p{p-1}n f}&\le \norm{f- (\widetilde \Pi \circ \mathcal{E} f)|_{I}} \le \norm{\mathcal{E} f- \widetilde \Pi \circ \mathcal{E} f}_{\widetilde I}\\ &\le \frac{n+1}{n}\frac 1{(n+1)\pi}\norm{\partial \mathcal{E} f}_{\widetilde I} = \frac{1}{n\pi}\norm{\partial f}, \end{align*} where the factor $(n+1)/n$ is the length of $\widetilde I$, $n+1$ is $\dim \widetilde{\mathcal S}$ and $\norm{\cdot}_{\widetilde I}$ denotes the $L^2$ norm on the interval $\widetilde I$. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:lower-norm}] The proof is by induction. The cases $p\ge q=0$ are proved in Lemma \ref{lem:ineqH1}. The case $(p,q)$ is proved assuming the result is true for $(p-1,q-1)$, namely that for $f\in H^q$ we have $$\norm {f - \proj {p-1}{p-2}n f }\le \Big(\frac 1 {n\pi}\Big)^{q}\norm{\partial^q f}.$$ Let $Q:H^{1}\rightarrow \spline p{p-1}n$ be the projection defined by \begin{equation}\label{eq:H1proj} Q f (x) =c(f) + \int_0^x \proj {p-1}{p-2}n \partial f(z)\,dz \end{equation} where $c(f)\in\mathbb{R}$ is uniquely determined by requiring that $Q$ is a projection. Since $\proj p{p-1}n$ is a projection and using Lemma \ref{lem:ineqH1}, we have for $f\in H^{q+1}$ \begin{align*} \norm{f-\proj p{p-1}n f }&= \norm{ (f-Q f) -\proj p{p-1}n ( f-Q f) } \le \Big(\frac 1 {n\pi}\Big) \norm{\partial (f-Q f)}. \end{align*} Using \eqref{eq:H1proj} and the induction hypothesis on $\partial f\in H^q$ we obtain \begin{align*} \Big(\frac 1 {n\pi}\Big) \norm{\partial (f-Q f)} = \Big(\frac 1 {n\pi}\Big) \norm{\partial f - \proj {p-1}{p-2}n \partial f} \le \Big(\frac 1 {n\pi}\Big)^{q+1} \norm{\partial^{q+1}f}. \end{align*} \end{proof} \begin{theorem}\label{thm:lower} Let $q$ and $k<q-1$ be given. If $\rt qk{\bar n}<1$ for some $\bar n$, then for all $p\ge q$, \begin{equation}\label{ineq:lower} n\ge \frac{p-k-1}{q-k-1}\bar n \end{equation} and with $m$ as in \eqref{eq:m-formula-low}, it holds $$ \cf p{p-1}mq<\const qkn. $$ \end{theorem} Observe that for fixed $n, k$ and $q$, the degree $p$ can only be increased until it reaches $p=(q-k)(n-1) +q $. At that point $m=1$ and $\spline p{p-1}m=\poly p$. \begin{proof} Similar to Section \ref{sec:comparison-1d} we have $$ \frac {\cf p{p-1}mq}{\const qkn}\le \rf pknq $$ where \begin{align*} \rf pknq&= (\Bf pknq)^{q+1}\sqrt{4q+6}\quad\text{with}\\ \label{eq:bf-pknq}\Bf pknq&=\frac{4}{e\pi} \frac{n (q+1)}{(q-k)(n-1)+1+q-p}. \end{align*} Moreover, $$\Bf pknq \le \B qk{\bar n}\quad \Rightarrow\quad\rf pknq\le \rt qk{\bar n}$$ and $\Bf pknq \le \B qk{\bar n}$ is equivalent to \eqref{ineq:lower}. \end{proof} Comparing $\Bf pknq$ in the above proof with $\B pkn$ for the case $p=q$ in Section~\ref{sec:comparison-1d}, there is only an additional $q-p$ in the denominator. \begin{example} IGA-DG comparison in $H^2$. It follows from Theorem \ref{thm:dg} that for $k=-1$ and $q=1$ we can choose $\bar n=3$ in \eqref{ineq:lower}. Thus the IGA space of degree $p\geq 1$ gives better approximation in $H^2$ than the DG space of degree $1$ for all $n\geq 3p$. \end{example} \begin{example} IGA-FEM comparison in $H^4$. It follows from Theorem \ref{thm:fem} that for $k=0$ and $q=3$ we can choose $\bar n=7$ in \eqref{ineq:lower}. Thus the IGA space of degree $p\geq 3$ gives better approximation in $H^4$ than the FEM space of degree $3$ for all $n\geq 7(p-1)/2$. \end{example} \section{Broken Sobolev spaces} \label{sec:broken} In numerical methods for PDEs, and especially in IGA \cite{Buffa:14}, it is common to consider broken Sobolev spaces, i.e., spaces of functions that are piecewise in $H^{p+1}$. The problem of approximating functions in broken Sobolev spaces arises in DG, PDEs with discontinuous coefficients and in isoparametric methods where the parametrization is only piecewise regular. The aim of this section is to show that smooth spline spaces have better approximation properties, provided the discontinuities are representable in the spline space and that the partitions are sufficiently fine. Let $\Xi=\{\xi_1<\dots<\xi_T\}\subset (0,1)$ be a set of breakpoints and $S=(s_1,\dots,s_{T})$, $s_i\in\{-1,\dots,p-1\}$, be the corresponding smoothness parameters. For notational reasons let $\xi_0=0$ and $\xi_{T+1}=1$. We consider the broken Sobolev space $\mathfrak H^{p+1}(\Xi,S)$ defined by \begin{equation} \mathfrak H^{p+1}(\Xi,S) = \left\{ \begin{aligned}f:\ &\partial^{p+1} f \in L^2(\xi_i,\xi_{i+1}),\, i=0,\dots,T \\& \partial^{s_i+1} f \in L^2(\xi_{i-1},\xi_{i+1}),\,i=1,\dots,T \end{aligned} \right\}. \end{equation} See Figure \ref{fig:broken} for an example. We will consider error estimates of the type $$ \norm{f-\Pi f} \le C \norm{\partial^{p+1}f}_\Xi $$ where $\norm{\cdot}_\Xi$ is the piecewise $L^2$ norm: $$ \norm{g}_\Xi = \Big (\sum_{i=0}^T \norm{ g}^2_{L^2(\xi_i,\xi_{i+1})} \Big)^{\frac 12}. $$ To achieve the expected approximation order it is necessary that the spline spaces can represent the discontinuities of the derivatives of the functions in $\mathfrak H^{p+1}(\Xi,S)$. Because of this we enrich the spline space $\spline pkn$ by adding $$ \mathfrak J^p_k(\Xi,S)= \{f: f|_{(\xi_i,\xi_{i+1})}\in\poly p,\, f\in \smooth {\min \{k, s_i\}}(\xi_{i-1},\xi_{i+1}),\ i=1,\dots,T \} $$ and obtaining $$ \brkS pkn(\Xi,S)= \spline pkn + \mathfrak J^p_k(\Xi,S). $$ Thus $\brkS pkn(\Xi,S)$ is a spline space having varying degree of smoothness at the breakpoints. An example is shown in Figure~\ref{fig:broken}. \begin{figure} \begin{center} \begin{tikzpicture} \draw (0,0) node[above] {$0$} -- (7,0) node[above] {$1$}; \draw (0,-.1)--(0,.1)(7,-.1)--(7,.1); \node[red] (CR) at (6,-1.5) {$\smooth{s_i}$}; \foreach \x[count=\i] in {2.3, 4.3, 5.5} { \fill[red] (\x,0) circle (1.5pt) node (Q\i) {}; \draw[red,->] (CR)--(Q\i); \node[above,red] at (Q\i) {$\xi_\i$}; } \end{tikzpicture} \begin{tikzpicture} \draw (0,0) node[above] {$0$} -- (7,0) node[above] {$1$}; \draw (0,-.1)--(0,.1)(7,-.1)--(7,.1); \node (CK) at (3,-1.5) {$\smooth k$}; \node[red] (CR) at (6,-1.5) {$\smooth {\min\{s_i,k\}}$}; \foreach \x[count=\i] in {1,...,6} { \fill (\x,0) circle (1.5pt) node (K\i){}; \draw[gray,->] (CK)--(K\i); \node[above] at (K\x) {$\frac \i7$}; } \foreach \x[count=\i] in {2.3, 4.3, 5.5} { \fill[red] (\x,0) circle (1.5pt) node (Q\i) {}; \draw[red,->] (CR)--(Q\i); \node[above,red] at (Q\i) {$\xi_\i$}; } \end{tikzpicture} \end{center} \caption{Above, the breakpoints and corresponding regularities defining $\mathfrak H^{p+1}(\Xi,S)$. Below, those defining $\brkS pk7(\Xi,S)$.}\label{fig:broken} \end{figure} Let $\brkC pkn(\Xi,S)$ be the smallest constant such that for all $f\in\mathfrak H^{p+1}(\Xi,S)$ we have $$ \norm{f- \mathfrak P_{p,k,n} f} \le \brkC pkn(\Xi,S) \norm{\partial^{p+1}f}_\Xi, $$ where $\mathfrak P_{p,k,n}$ is the orthogonal projection onto $\brkS pkn(\Xi,S)$. As in the previous sections we compare $\brkC p{p-1}m(\Xi,S)$ with $\brkC pkn(\Xi,S)$. In this case it is not always possible to choose $m$ such that $\dim \brkS p{p-1}m(\Xi,S)=\dim\brkS pkn(\Xi,S)$ because an increment of $1$ in $m$ does not necessarily correspond to an increment of $1$ in $\dim \brkS p{p-1}m(\Xi,S)$, e.g., when some of the breakpoints of $\spline p{p-1}m$ align with the points in $\Xi$. The dimension of $\brkS pkn(\Xi,S)$ is $$ \dim \brkS pkn(\Xi,S)= (p-k)(n-1) +p+1 + \sum_{i=1}^T \sigma_i $$ where $$\sigma_i=\begin{cases} p-\min\{k, s_i\}& \xi_i\not\in\{j/n,j=1,\dots,n-1\}\\ \max\{k-s_i, 0\}&\xi_i\in \{j/n,j=1,\dots,n-1\}. \end{cases} $$ In particular for $k=p-1$ we have $k\le s_i$ and $k-s_i=(p-s_i)-1$ giving $$ \dim \brkS p{p-1}m(\Xi,S)= m +p + \sum_{i=1}^T (p-s_i) - \#(M\cap \Xi) $$ where $M=\{i/m:\,i=1,\dots,m-1 \}$. Our choice of $m$ is \begin{equation}\label{eq:brk-m} m= (n-1)(p-k)+1+ \sum_{i=1}^T(\sigma_i+s_i-p) \end{equation} for which we have $$\dim \brkS p{p-1}m(\Xi,S)=\dim \brkS pkn(\Xi,S)-\#(M\cap \Xi)\leq \dim \brkS pkn(\Xi,S).$$ \begin{lemma} For all $\Xi$ and $S$ we have \begin{equation}\label{eq:broken-bounds} \frac{(p+1)!}{(2p+2)!\sqrt{2p+3}} (n+T)^{-p-1}\le \brkC pkn(\Xi,S) \le \const pkn \end{equation} \end{lemma} \begin{proof} The lower bound is obtained by looking at $k=-1$. In this case $\brkS pkn(\Xi,S)$ is a space of discontinuous piecewise polynomials on the non-uniform partition containing the intersections $(\xi_i,\xi_{i+1})\cap I_j$. This partition has at most $n+T$ elements, moreover for a given number of elements the approximation error of $x^{p+1}$ is minimized by the uniform partition. We can thus use $\const p{-1}{n+T}$ as a lower bound. Next we look at the upper bound. Given any $f \in \mathfrak H^{p+1}(\Xi,S)$ we can choose a $g\in \mathfrak J^p_k(\Xi,S)$ such that $f-g\in H^{p+1}$ and \begin{align*} \Vert f -\mathfrak P_{p,k,n} f\Vert^2 &=\Vert (f-g) -\mathfrak P_{p,k,n} (f-g) \Vert^2\\ &\le\norm{(f-g) -\proj pkn (f-g)}^2\\ &\le (\const pkn \norm{\partial^{p+1}(f-g)})^2\\ &=(\const pkn)^2\sum_{i=0}^T \norm{\partial^{p+1}(f-g)}^2_{L^2(\xi_i,\xi_{i+1})}\\ &=(\const pkn \norm{\partial^{p+1}f}_\Xi)^2. \end{align*} The result then follows by taking the square-root of both sides. \end{proof} Similarly to Section \ref{sec:low} we deduce the following result: \begin{theorem}\label{thm:broken} Let $\Xi$ and $S$ be given. If $\rt pk{\bar n}<1$ for some $\bar n$, then for all \begin{equation}\label{ineq:2} n\ge \Big(1 + \frac{T(p-k)-\sum_{i=1}^T (\sigma_i+s_i-p)}{p-k-1} \Big)\bar n - T \end{equation} and $m$ as in \eqref{eq:brk-m} we have $$ \brkC p{p-1}m<\brkC pkn. $$ \end{theorem} \begin{proof} Reasoning as in Section \ref{sec:comparison-1d} we have $$ \frac {\brkC p{p-1}m}{\brkC pkn}\le \brkR pkn. $$ where \begin{align*} \brkR pkn&= (\brkB pkn)^{p+1}\sqrt{4q+6}\quad\text{with}\\ \brkB pkn&=\frac{4}{e\pi} \frac{(n+T) (p+1)}{(p-k)(n-1)+1+\sum_{i=1}^T(\sigma_i+s_i-p)}. \end{align*} Moreover, $$\brkB pkn \le \B pk{\bar n}\quad \Rightarrow\quad \brkR pkn\le \rt pk{\bar n}$$ and $\brkB pkn \le \B pk{\bar n}$ is equivalent to \eqref{ineq:2}. \end{proof} \section{The multivariate case}\label{sec:tens} In this section we consider the unit hypercube $\Omega=(0,1)^d$ and a tensor product space \begin{equation}\label{eq:tensor} {\bf V}= V_1 \otimes\dots\otimes V_d. \end{equation} Let $\Pi_{\bf V}$ be the $L^2(\Omega)$ projection onto $\bf V$, $\Pi_{i}$ the $L^2(0,1)$ projection onto $V_i$ and $C_{i}$ be the smallest constant in the univariate estimate $$ \Vert f-\Pi_{i} f\Vert \le C_i \Vert \partial^{q_i+1}f\Vert,\qquad\forall f\in H^{q_i+1}(0,1). $$ \begin{theorem}\label{thm:multivariate} For all $f\in L^2(\Omega)$, such that $\partial_i^{q_i+1} f \in L^2(\Omega)$ we have \begin{equation}\label{ineq:kd} \norm{f-\Pi_{\bf V} f} \le \sum_{i=1}^d C_i \norm{\partial_i^{q_i+1}f} \end{equation} and the result is sharp. \end{theorem} \begin{proof} First of all, we remind that $L^2(\Omega)=L^2(0,1)\otimes\dots\otimes L^2(0,1)$ and that $\Pi_{\bf V}$ factorizes as $$ \Pi_{\bf V} = \Pi_1\otimes \dots \otimes \Pi_d. $$ These factors commute as in $\Pi_i\otimes \Pi_j=(\Pi_i\otimes \mathrm{I})\circ(\mathrm{I}\otimes \Pi_j)=(\mathrm{I}\otimes \Pi_j)\circ (\Pi_i\otimes \mathrm{I})$ where $\mathrm I$ is in the identity operator. The theorem is proved by induction on $d$. For $d=1$ it is the definition of $C_i$. Now suppose the result is true for dimension $d-1$, i.e., that for $\Pi_*=\Pi_1\otimes\dots\otimes\Pi_{d-1}$ we have $$ \norm{f-\Pi_* f}\le \sum_{i=1}^{d-1}C_i\norm{\partial_i^{q_i+1}f}. $$ Using the triangle inequality and that $\norm{\mathrm{I}\otimes\Pi_d}=1$ we find \begin{align*} \norm{f- \Pi_1\otimes\dots\otimes \Pi_d f}&\le \norm{f-\mathrm{I}\otimes\Pi_d f} + \norm{\mathrm{I}\otimes\Pi_df- \Pi_*\otimes\Pi_d f}\\ &\le C_d \norm{\partial_{d}^{q_d+1}f} + \norm{\mathrm{I}\otimes\Pi_d} \norm{f - \Pi_*\otimes \mathrm{I} f}\\ &\le \sum_{i=1}^d C_i \norm{\partial_i^{q_i+1}f}. \end{align*} To see that the result is sharp we consider $f(x_1,\dots,x_d)=g(x_i)$ and notice that the statement is false for any constant smaller than $C_i$. \end{proof} From Theorem \ref{thm:multivariate} we deduce that all conclusions obtained in the univariate comparisons extend to the tensor product setting by considering each direction separately. \section{Conclusions} In this paper we have provided a mathematical justification for the numerically observed phenomena that $\smooth {p-1}$ splines of degree $p$, the so-called $k$-\emph{method} in IGA, provide better approximation in degrees of freedom than splines of smoothness $\smooth {-1}$ (DG) and $\smooth {0}$ (FEM). Specifically, we have shown in Section~\ref{sec:comparison-1d} that for sufficiently refined uniform partitions, $\smooth{p-1}$ splines yield better a priori error estimates than $\smooth{-1}$ splines for $p\ge 2$, and $\smooth0$ splines for $p\ge 3$, when approximating functions in the Sobolev space $H^{p+1}$. For $p=2$ it is an open problem whether $\smooth {1}$ spline spaces provide better a priori error estimates than $\smooth0$ spline spaces of the same dimension. Sharper estimates on the approximation constants could complete the result for this case. Since we have used the lower bound for discontinuous spline approximation also as the lower bound for continuous spline approximation, it seems reasonable to look for an improved lower bound on the approximation constants for $\smooth0$ splines. It is worth mentioning that the combination of the techniques in Sections~\ref{sec:low} and \ref{sec:broken} allow also the comparison for lower order broken Sobolev spaces. This comparison has not been included. \section*{Appendix} \begin{lemma}\label{lem:poly-interp} Let $p\geq 1$ be any odd number. Then the interpolation problem: find $g \in \poly p$ such that for all $s=0,2,\ldots,p-1,$ \begin{equation}\label{poly-interp} \left\{\begin{aligned} \partial^s g(0)&=a_s\\ \partial^s g(1)&=b_s \end{aligned}\right .\ \end{equation} admits a solution for all $\{a_s\}$, $\{b_s\}$. \end{lemma} \begin{proof} We proceed by induction on $p$. If $p=1$ then the linear interpolant $g(x)=a_0 + x (b_0-a_0)$ satisfies $g(0)=a_0$ and $g(1)=b_0$ and is a solution. Now, let $p\geq 3$ be any odd number and assume the result is true for $p-2$. Let $f\in \poly {p-2}$ be the solution of \begin{equation*} \partial^s f(0)=a_{s+2},\quad \partial^s f(1)=b_{s+2}\quad s=0,2,\ldots,p-3, \end{equation*} which we know exist by the induction hypothesis. We then define the function $g$ by integrating $f$ twice, i.e., $$g(x)=cx+d+\int_0^x\int_0^yf(z)dzdx.$$ This function then satisfies the cases $s\geq2$ of \eqref{poly-interp} for all $c,d\in\mathbb{R}$. We finish the proof by picking the constants $c$ and $d$ such that the case $s=0$, meaning $g(0)=a_0$ and $g(1)=b_0$, is also satisfied. \end{proof} \begin{lemma}\label{lem:poly-interp2} Let $p\geq 0$ be any even number. Then the interpolation problem: find $g \in \poly p$ such that for all $s=1,3,\ldots,p-1,$ \begin{equation}\label{poly-interp2} \left\{\begin{aligned} \partial^s g(0)&=a_k\\ \partial^s g(1)&=b_k \end{aligned}\right .\ \end{equation} admits a solution for all $\{a_s\}$, $\{b_s\}$. \end{lemma} \begin{proof} For $p=0$ there is nothing to prove, and so we consider an even number $p\geq 2$. We then let $f\in\poly{p-1}$ be the solution of \begin{equation*} \partial^s f(0)=a_{s-1},\quad \partial^s f(1)=b_{s-1}\quad s=0,2,\ldots,p-2, \end{equation*} which we know exist by Lemma \ref{lem:poly-interp}. The function $g(x)=c+\int_0^xf(y)dy$ is then a solution of \eqref{poly-interp2} for any $c\in\mathbb{R}$. \end{proof} \section*{Acknowledgements} The research leading to these results has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement 339643.

https://arxiv.org/abs/1202.2023

Surprising symmetries in 132-avoiding permutations

We prove that the total number $S_{n,132}(q)$ of copies of the pattern $q$ in all 132-avoiding permutations of length $n$ is the same for $q=231$, $q=312$, or $q=213$. We provide a combinatorial proof for this unexpected threefold symmetry. We then significantly generalize this result to show an exponential number of different pairs of patterns $q$ and $q'$ of length $k$ for which $S_{n,132}(q)=S_{n,132}(q')$ and the equality is non-trivial.

\section{Introduction} \subsection{Background and Definitions} Let $q=q_1 q_2\ldots q_k$ be a permutation in the symmetric group $S_k$. We say that the permutation $p=p_1 p_2 \ldots p_n\in S_n$ {\it contains a $q$-pattern\/} if and only if there is a subsequence $p_{i_1}p_{i_2}\ldots p_{i_k}$ of $p$ whose elements are in the same relative order as those in $q$, that is, $$ p_{i_t}<p_{i_u} \mbox{ if and only if } q_t<q_u $$ whenever $1\leq t,u\leq k$. If $p$ does not contain $q$, then we say that $p$ {\em avoids} $q$. For instance, 214653 contains 231 (consider the third, fourth, and sixth entries), but avoids 4321. See Chapter 14 of \cite{bona} for an introduction to pattern avoiding permutations, and Chapters 4 and 5 of \cite{combperm} for a somewhat more detailed treatment. It is straightforward to compute, using the linear property of expectation, that the average number of $q$-patterns in a randomly selected permutation of length $n$ is $\frac{1}{k!}{n\choose k}$, where $k$ is the length of $q$. Joshua Cooper \cite{cooper} has raised the following interesting family of questions. Let $r$ be a given permutation pattern. What can be said about the average number of occurrences of $q$ in a randomly selected $r$-avoiding permutation of a given length? Equivalently, can we determine the {\em total number} $S_{n,r}(q)$ of all $q$-patterns in all $r$-avoiding permutations of length $n$? \subsection{Earlier Results} In \cite{occurrences}, present author found formulae for the generating functions of the sequence $S_{132,n}(q)$ for the cases of monotone $q$, that is, for $q=12\cdots k$ and $q=k(k-1)\cdots 1$, for any $k$. He also proved that if $n$ is large enough, then for any fixed $k$, among all patterns $q$ of length $k$, it is the monotone decreasing pattern that maximizes $S_{132,n}(q)$ and it is the monotone increasing pattern that minimizes $S_{132,n}(q)$. \subsection{The Outline of our Paper} In this paper, we first present a computational proof of the surprising fact that for all $n$, the equalities \begin{equation} \label{triple} S_{132,n}(231)=S_{132,n}(312)=S_{132,n}(213) \end{equation} hold. The first equality is trivial, since taking the inverse of a 132-avoiding permutation keeps that permutation 132-avoiding, and turns 231-patterns into 312-patterns. However, the second equality is non-trivial. (The reverse or complement of a 132-avoiding permutation is not necessarily 132-avoiding.) In particular, if $a(p)$ denotes the number of 213-copies in $p$, and $b(p)$ denotes the number of 231-copies in $p$, then the statistics $a(p)$ and $b(p)$ are {\em not} equidistributed over the set of all 132-avoiding permutations of length $n$, but their average values are equal over that set. In other words, we will prove that a randomly selected non-monotonic pattern of length three in a 132-avoiding permutation is equally likely to be a 231-pattern, a 312-pattern, or a 213-pattern. It is well-known (see Chapter 14 of \cite{bona}) that 132-avoiding permutations of length $n$ are counted by the Catalan numbers $c_n={2n\choose n}/(n+1)$, and as such, they are one of more than 150 distinct kinds of objects counted by those numbers. However, we do not know of any other example when a natural statistic on objects counted by Catalan numbers shows a similar threefold symmetry. In the next part of the paper we provide a bijective proof of (\ref{triple}). Finally, we will significantly generalize this result by showing more than $c_{h-2}$ pairs of patterns $q$ and $q'$ of length $h$ that behave as 213 and 231, that is, for which $S_{n,132}(q)= S_{n,132}(q')$, and the equality is non-trivial. \section{Arguments Using Generating Functions} Let $d_n$ be the total number of inversions (in other words, copies of the pattern 21) in all 132-avoiding $n$-permutations. It is proved in \cite{occurrences} that \begin{equation} \label{inversions} D(x)= \sum_{n\geq 1}d_nx^n=\frac{x}{1-4x} \cdot \left (\frac{1}{\sqrt{1-4x}} - \frac{1-\sqrt{1-4x}}{2x} \right). \end{equation} \subsection{Counting Copies of 213} Let $a_n$ be the total number of all 213-patterns in all 132-avoiding permutations of length $n$. Clearly, then $a_0=a_1=a_2=0$. There are three ways that a 132-avoiding permutation $p$ of length $n$ can contain a 213-pattern $q$. Either $q$ is entirely on the left of the entry $n$, or $q$ is entirely on the right of $n$, or $q$ ends in $n$. For $n\geq 3$, this leads to the recurrence relation \[a_n= \sum_{i=1}^n a_{i-1}c_{n-i} + \sum_{i=1}^{n} c_{n-1}a_{n-i} + \sum_{i=3}^n d_{i-1} c_{n-i}.\] Let $A(x)$ (resp. $C(x)$) be the ordinary generating function for the sequence of the numbers $a_n$ (resp. $c_n$). Then the last displayed formula yields the functional equation \[A(x)= 2xA(x)C(x)+xD(x)C(x) ,\] which is equivalent to \begin{equation} \label{explicitA} A(x)=\frac{xD(x)C(x)}{1-2xC(x)}=\frac{x}{2(1-4x)^2} +\frac{x-1}{2(1-4x)^{3/2}} + \frac{1}{2(1-4x)}.\end{equation} From here, we get that if $n\geq 3$, then \[a_n=\frac{n}{2}4^{n-1}+\frac{1}{2}4^n-(2n+1){2n-1\choose n-1}+ (2n-1){2n-3\choose n-2}, \] which simplifies to \begin{equation} \label{exactfora} a_n=(n+4)\cdot 2^{2n-3}-(2n+1){2n-1\choose n-1}+ (2n-1){2n-3\choose n-2}. \end{equation} \subsection{Counting Copies of 231} Let $h_n$ be the total number of all non-inversions (in other words, copies of the pattern 12) in all 132-avoiding permutations of length $n$. It is proved in \cite{occurrences} that \begin{equation} \label{non-inversions} H(x)=\sum_{n\geq 0}h_nx^n=\frac{1}{2(1-4x)} + \frac{1}{2x} -\frac{1-x}{2x\sqrt{1-4x}}. \end{equation} Let $b_n$ be the total number of all 231-copies in all 132-avoiding permutations of length $n$, and let $B(x)=\sum_{n\geq 0}b_nx^n$. Let \begin{equation} \label{explicitZ} Z(x)=\sum_{n\geq 0}nc_nx^n= \sum_{n\geq 0}{2n\choose n}\frac{n}{n+1}x^n= \frac{1}{\sqrt{1-4x} }- \frac{1-\sqrt{1-4x}}{2x}.\end{equation} Note that $Z(x)$ is the generating function for the number of entries (which are copies of the pattern 1) in all 132-avoiding $n$-permutations. If $p$ is a 132-avoiding $n$-permutation, and $q$ is a 231-pattern contained in $p$, then either $q$ is entirely on the left of the entry $n$, or $q$ is entirely on the right of the entry $n$, or the entry $n$ is the largest entry of $q$, or the first and second entries of $q$ form a 12-pattern on the left of $n$, while the third entry of $q$ is on the right of $n$. For $n\geq 3$, this leads to the recurrence relation \[b_n=\sum_{i=1}^n b_{i-1}c_{n-i} + \sum_{i=1}^{n} c_{n-1}b_{n-i} + \sum_{i=2}^{n-1}(i-1)(n-i)c_{i-1}c_{n-i} + \sum_{i=3}^{n-1} h_{i-1}c_{n-i}(n-i).\] In terms of generating functions, this yields \[B(x)=2xB(x)C(x) + xZ^2(x)+xH(x)Z(x),\] \[B(x)=\frac{ xZ^2(x)+xH(x)Z(x)}{1-2xC(x)}= \frac{ xZ^2(x)+xH(x)Z(x)}{\sqrt{1-4x}}.\] Given the explicit formulae (\ref{non-inversions}) and (\ref{explicitZ}) for $H(x)$ and $Z(x)$, the last displayed equation yields the formula \begin{equation} \label{explicitB} B(x)=\frac{xD(x)C(x)}{1-2xC(x)}=\frac{x}{2(1-4x)^2} +\frac{x-1}{2(1-4x)^{3/2}} + \frac{1}{2(1-4x)}. \end{equation} The proof of the main result of this section is now immediate. \begin{theorem} For all positive integers $n$, the equalities \[ S_{132,n}(231)=S_{132,n}(312)=S_{132,n}(213)\] hold. \end{theorem} \begin{proof} As we mentioned in the Introduction, the first equality is trivially true since there is a natural bijection between the 231-copies of the 132-avoiding permutation $p$ and the 312-copies of the 132-avoiding permutation $p^{-1}$. Indeed, let $p=p_1p_2\cdots p_n$ be a 132-avoiding permutation, and let $1\leq i<j<k\leq n$. Then $p_ip_jp_k$ is a 231-copy in $p$ if and only if $ijk$ is a 312-copy in $p^{-1}$. The equality $S_{n,132}(231)=S_{n,132}(213)$ holds since we have seen in formulae (\ref{explicitA}) and (\ref{explicitB}) that the two sides of this equality have identical generating functions. \end{proof} \section{A Bijective Proof} In this section we provide a bijective proof for the surprising identity $S_{n,132}(213)=S_{n,132}(231)$. \subsection{Binary Plane Trees} In our proof, we will identify a 132-avoiding permutation $p$ with its {\em binary plane tree} $T(p)$ using a very well-known bijection. We will briefly describe this bijection now. For more details, the reader may consult Chapter 14 of \cite{bona}. The tree $T(p)$ will be a binary plane tree, that is, a rooted unlabeled tree in which each vertex has at most two children, and each child is a left child or a right child of its parent, even if it is the only child of its parent. The root of $T(p)$ corresponds to the entry $n$ of $p$, the left subtree of the root corresponds to the string of entries of $p$ on the left of $n$, and the right subtree of the root corresponds to the string of entries of $p$ on the right of $n$. Both subtrees are constructed recursively, by the same rule. Note that since $p$ is 132-avoiding, the position of the entry $n$ of $p$ determines the set of entries that are on the left (resp. on the right) of $n$. In fact, if $n$ is in the $i$th position, the set of entries on the left of $n$ must be $\{n-i+1,n-i+2,\cdots ,n-1\}$, and the set of entries on the right of $n$ must be $\{1,2,\cdots ,n-i\}$. We point out that in the process of constructing $T(p)$, each vertex of $T(p)$ is associated to an entry of $p$. Indeed, each vertex is added to $T(p)$ as the root of a subtree $S$, and so each vertex is associated to the entry that is the largest among the entries that belong to $S$. However, it is important to point out that $T(p)$ is an {\em unlabeled tree} since the way in which the entries of $p$ correspond to the vertices of $T(p)$ is completely determined by the unlabeled tree $T(p)$ as long as $p$ is 132-avoiding. See Figure \ref{binplane} for an illustration. \begin{figure}[ht] \begin{center} \epsfig{file=binplane.eps} \label{binplane} \caption{The tree $T(p)$ for $p=67823415$, and the entries of $p$ associated to the vertices of $T(p)$.} \end{center} \end{figure} Note that in order to get $p$ from $T(p)$, it suffices to read the vertices of $T(p)$ {\em in-order}, that is, by first reading the left subtree of the root, then the right subtree of the root, and then the right subtree of the root. The respective subtrees are read recursively, by this same rule. Therefore, it is meaningful to talk about the first, second, etc, last vertex of $T(p)$, since that means the first, second, etc, last vertex of $T(p)$ in the {\em in-order} reading. A {\em left descendant} (resp. {\em right descendant} of a vertex $x$ in a binary plane tree is a vertex in the left (resp. right) subtree of $x$. The left (resp. right) subtree of $x$ does {\em not} contain $x$ itself. It is straightforward to see that $p_ip_j$ is a 12-pattern in $p$ if and only if $p_i$ is a left-descendant of $p_j$ in $T(p)$. On the other hand, $p_jp_i$ is a 21-pattern in $p$ if and only if either $p_i$ is a right descendant of $p_j$ in $T(p)$ or there is a vertex $x$ in $T(p)$ so that $p_j$ is a left descendant of $x$ and $p_i$ is a right descendant of $x$. In the previous section we gave an exhaustive list of the ways in which 213-patterns and 231-patterns can occur in a 132-avoiding permutation. The reader is invited to translate that list into the language of binary plane trees. \subsection{Our Bijection} Let $p$ be a 132-avoiding $n$-permutation, and let $Q$ be an occurrence of the pattern 213 in $p$. Let $Q_2,Q_1,Q_3$ be the three vertices of $T(p)$ that correspond to $Q$, going left to right. Let us color these three entries black. There are then two possibilities. \begin{enumerate} \item Either $Q_1$ is a right descendant of $Q_2$ and $Q_2$ is a left descendant of $Q_3$, or \item there exists a lowest left descendant $Q_x$ of $Q_3$ so that $Q_2$ is a left descendant of $Q_x$ and $Q_1$ is a right descendant of $Q_x$. \end{enumerate} Let $A_n$ be the set of all binary plane trees on $n$ vertices in which three vertices forming a 213-pattern are colored black. Let $B_n$ be the set of all binary plane trees on $n$ vertices in which three vertices forming a 231-pattern are colored black. Now we are going to define a map $f:A_n\rightarrow B_n$. We will then prove that $f$ is a bijection. The map $f$ will be defined differently in the two cases described above. \begin{itemize} \item {\em Case 1.} If $T\in A_n$ is in the first case, then let $f(T)$ be the pair obtained by interchanging the right subtree of $Q_2$ and the right subtree of $Q_3$. Keep all three black vertices $Q_i$ black, even as $Q_1$ gets moved. See Figure \ref{firstmove} for an illustration. \begin{figure}[ht] \begin{center} \epsfig{file=firsttreemove.eps} \label{firstmove} \caption{Interchanging the right subtrees of $Q_2$ and $Q_3$.} \end{center} \end{figure} Note that in $f(T)$, in the set of black vertices, there is one that is an ancestor of the other two, namely $Q_3$. \item {\em Case 2.} If $T\in A_n$ is in the second case, then let $f(T)$ be the tree obtained by interchanging the right subtrees of the vertices $Q_x$ and $Q_3$, and coloring $Q_2$, $Q_x$ and $Q_1$ black. See Figure \ref{secondmove} for an illustration. \begin{figure}[ht] \begin{center} \epsfig{file=secondtreemove.eps} \label{secondmove} \caption{Interchanging the right subtrees of $Q_x$ and $Q_3$.} \end{center} \end{figure} Note that in $f(T)$, there is no black vertex that is an ancestor of the other two black vertices. Also note that in $f(T)$, the lowest common ancestor of $Q_x$ and $Q_1$ is $Q_3$. \end{itemize} It is a direct consequence of our definitions that if $T\in A_n$, then $f(T)=B_n$. Now we are in a position to prove the main result of this section. \begin{theorem} \label{bijective} The map $f:A_n\rightarrow B_n$ defined above is a bijection. \end{theorem} \begin{proof} Let $U\in B_n$. We will show that there is exactly one $T\in A_n$ so that $f(T)=U$ holds. This will show that $f$ has an inverse, proving that $f$ is a bijection. By definition, three nodes of $U$ are colored black, and the entries of the permutation corresponding to $U$ form a 231-pattern. Let $K_2$, $K_3$, and $K_1$ denote these three vertices, from left to right. There are two possibilities for the location of the $K_i$ relative to each other. We will show that in both cases, $U$ has a unique preimage under $f$, essentially because swapping two subtrees is an involution. \begin{enumerate} \item If $K_3$ is an ancestor of both other black vertices, then $f(T)=U$ implies that $T$ belongs to Case 1. In this case, the unique $T\in A_n$ satisfying $f(T)=U$ is obtained by swapping the right subtrees of $K_3$ and $K_2$, and keeping all three black vertices black, even if $K_1$ got moved. \item If $K_3$ is not an ancestor of both other black vertices and then $f(T)=U$ implies that $T$ belongs to Case 2. In this case, let $K_x$ be the smallest common ancestor of $U_3$ and $U_1$. Then the unique $T\in A_n$ satisfying $f(T)=U$ is obtained by swapping the right subtrees of $K_3$ and $K_x$, and coloring $K_x$ black instead of $K_3$, while keeping $K_1$ and $K_2$ black. \end{enumerate} This completes the proof. \end{proof} \section{A Generalization} In this section, we will significantly generalize the result of the previous section. The key observation is that in the proof of Theorem \ref{bijective}, the {\em left} subtrees of $Q_1$ and $Q_2$ never changed. In order to state our result, we announce the following definitions. \begin{definition} Let $q$ be a pattern of length $k$ and let $t$ be a pattern of length $m$. Then $q\oplus t$ is the pattern of length $k+m$ defined by \[ (q\oplus t)_i= \left\{ \begin{array}{l@{\ }l} q_i \hbox{ if $i\leq k$},\\ \\ t_{i-k} +k \hbox{ if $i>k$. } \end{array}\right. \] \end{definition} In other words, $q\oplus t$ is the concatenation of $q$ and $t$ so that all entries of $t$ are increased by the size of $q$. \begin{example} If $q=3142$ and $t=132$, then $q\oplus t=3142576$. \end{example} \begin{definition} Let $q$ be a pattern of length $k$ and let $t$ be a pattern of length $m$. Then $q\ominus t$ is the pattern of length $k+m$ defined by \[ (q\ominus t)_i= \left\{ \begin{array}{l@{\ }l} q_i + m \hbox{ if $i\leq k$},\\ \\ t_{i-k} \hbox{ if $i>k$. } \end{array}\right. \] \end{definition} In other words, $q\ominus t$ is the concatenation of $q$ and $t$ so that all entries of $q$ are increased by the size of $t$. \begin{example} If $q=3142$ and $t=132$, then $q\ominus t=6475132$. \end{example} Now we are ready to state and prove the most general result of this paper. \begin{theorem} \label{general} Let $q$ and $t$ be any non-empty patterns that end in their largest entry. Let $i_u$ denote the increasing pattern $12\cdots u$. Then for all positive integers $n$, we have \[S_{n,132}((q\ominus t)\oplus i_u) = S_{n,132}((q\oplus i_u)\ominus t),\] where 1 denotes the pattern consisting of one entry. \end{theorem} In particular, the result of the previous section is the special case of Theorem \ref{general} in which $q=t=i_u=1$ (the one-entry pattern 1). \begin{example} If $q=3124$, $t=213$, and $u=2$, then Theorem \ref{general} says that \[S_{n,132}(645721389)=S_{n,132}(645789213).\] \end{example} \begin{proof} (of Theorem \ref{general}) Note that we can assume that $q$ and $t$ are both 132-avoiding, since otherwise the statement of Theorem \ref{general} is trivially true as both sides are equal to 0. Let $k$ denote the length of $q$, let $m$ denote the length of $t$. Similarly to the proof of Theorem \ref{bijective}, let $A_n$ be the set of all binary plane trees on $n$ vertices in which $h$ vertices forming a $((q\ominus t)\oplus i_u)$-pattern are colored black, and let $B_n$ be the set of all binary plane trees on $n$ vertices in which $h$ vertices forming a $((q\oplus i_u)\ominus t)$-pattern are colored black. Let $T\in A_n$. Let $Q_b$ be the $k$th black vertex of $T$ in the in-order reading, let $Q_a$ be the $(k+m)$th black vertex of $T$, and let $Q_c$ be the rightmost black vertex of $T$. We are now going to construct a bijection $F:A_n\rightarrow B_n$. The construction is analogous to the one that we saw before Theorem \ref{bijective} \begin{enumerate} \item If $Q_a$ is a right descendant of $Q_b$, then let $F(T)$ be the tree obtained from $T$ by swapping the right subtree of $Q_b$ and the right subtree of $Q_c$. Note that in $F(T)$, the black vertices form a $(q\ominus t)\oplus i_u$-pattern, and that $Q_c$ is an ancestor of all other black vertices in $F(T)$. See Figure \ref{firstg} for an illustration. \begin{figure}[ht] \begin{center} \epsfig{file=firstg.eps} \label{firstg} \caption{Interchanging the right subtrees of $Q_b$ and $Q_c$, and turning a copy of 341256 into a copy of 345612.} \end{center} \end{figure} \item Otherwise, there exists a lowest vertex $Q_x\in T$ so that $Q_b$ is a left descendant of $Q_x$ and $Q_a$ is a right descendant of $Q_x$. Note that in this case, it follows that $Q_x$ is not black. Now let $F(T)$ be the tree obtained from $T$ by swapping the right subtree of $Q_x$ and the right subtree of $Q_c$, and by coloring $Q_x$ black, instead of $Q_c$. Note that again, in $F(T)$, the black vertices form a $(q\ominus t)\oplus i_u$-pattern. Also note that there is no black vertex in $F(T)$ that would be an ancestor of all other black vertices. See Figure \ref{secondg} for an illustration. \begin{figure}[ht] \begin{center} \epsfig{file=secondg.eps} \label{secondg} \caption{Interchanging the right subtrees of $Q_x$ and $Q_c$, coloring $Q_x$ black instead of $Q_c$, and turning a copy of 341256 into a copy of 345612.} \end{center} \end{figure} \end{enumerate} It is straightforward to show that $F:A_n \rightarrow B_n$ is a bijection. Indeed, let $U\in B_n$. If there is a black vertex in $U$ that is an ancestor of all other black vertices, then $U$ could only be obtained by the first rule, otherwise $U$ could only be obtained by the second rule. The unique preimage $F^{-1}(U)$ is then obtained by swapping the appropriate right subtrees. In the first case, swap the right subtrees of $U_b$ and $U_c$, where $U_b$ is the $k$th and $U_c$ is the $(k+u)$th black vertex of $U$ in the in-order reading. In the second case, let $U_x$ be the $(k+1)$st black vertex of $U$, let $U_a$ be the last black vertex of $U$, and let $U_c$ be the lowest common ancestor of $U_x$ and $U_a$. Then the unique preimage $F^{-1}(U)$ is obtained by swapping the right subtrees of $U_x$ and $U_c$, and coloring $U_c$ black instead of $U_x$. \end{proof} Note that by transitivity, Theorem \ref{general} implies the following. \begin{corollary} Let $q$, $t$, and $i_u$ be as in Theorem \ref{general}, and let $1\leq v<u$. Then we have \[S_{n,132}((q\oplus i_v)\ominus t)\oplus i_{u-v})=S_{n,132}((q\ominus t)\oplus i_u).\] \end{corollary} \begin{proof} Theorem \ref{general} shows that both sides are equal to $S_{n,132}((q\oplus i_u)\ominus t)$. \end{proof} \section{Further Directions} Formula (\ref{exactfora}) implies that $S_{n,132}(213)\sim C_1 4^n n$, while the generating functions computed in \cite{occurrences} imply that $S_{n,132}(321)\sim C_2 4^n n^{3/2}$ and $S_{n,132}(123)\sim C_3 4^n n^{1/2}$, where the $C_i$ are positive constants. So occurrences of non-monotone patterns of length three are infinitely rare compared to occurrences of 321, and infinitely frequent compared to occurrences of 123; the frequency of non-monotone patterns is halfway between the two extremes. While precise formulae like the ones given in earlier sections of this paper may not be obtainable for longer patterns, comparative results as the ones described in the previous paragraph may be possible to prove even for such patterns. If we set $u=1$ and $h=k+m+1$, then Theorem \ref{general} provides $\sum_{i=2}^{h-1}c_{i-2}c_{h-i-1}=c_{h-2}$ non-trivial examples of two patterns $s$ and $s'$ for which $S_{n,132}(s)=S_{n,132}(s')$ for all $n$. Other choices of $u$ provide additional such pairs. However, it seems that there are other pairs of patterns whose total number of copies in all 132-avoiding permutations agree. We hope to discuss such pairs in an upcoming paper. Are there any other such pairs? Are there any such pairs when 132 is replaced by another pattern $r$? Are there any patterns $r$ and $r'$ for which $S_{n,r}(u)=S_{n,r'}(u')$ for all $n$ and the equality is non-trivial? Finally, how do the permutation statistics studied in this paper translate to the other 150 families of objects counted by the Catalan numbers listed in \cite{stanley}?

https://arxiv.org/abs/1911.12464

Words With Few Palindromes, Revisited

In 2013, Fici and Zamboni proved a number of theorems about finite and infinite words having only a small number of factors that are palindromes. In this paper we rederive some of their results, and obtain some new ones, by a different method based on finite automata.

\section{Introduction} In this paper we are concerned with certain avoidance properties of finite and infinite words. Recall that a word $x$ is said to be a {\it factor} of a word $w$ if there exist words $y,z$ such that $w = yxz$. For example, the word {\tt act} is a factor of the English word {\tt factor}. We sometimes say $w$ {\it contains\/} $x$. Another term for {\it factor} is {\it subword}, although this latter term sometimes refers to a different concept entirely. We say a (finite or infinite) word $x$ {\it avoids} a set $S$ if no element of $S$ is a factor of $x$. The reverse of a word $x$ is written $x^R$. Thus, for example, ${\tt (drawer)}^R = {\tt reward}$. A word is a {\it palindrome\/} if $x = x^R$, such as the English word {\tt radar}. A palindrome is called {\it even\/} if its length is even, and {\it odd\/} if its length is odd. For example, the English word {\tt noon} is even, while {\tt madam} is odd. Fici and Zamboni \cite{Fici&Zamboni:2013} studied avoidance of palindromes. In particular, they were interested in constructing infinite words with the minimum possible number of distinct palindromic factors, and infinite words that minimize the length of the largest palindromic factor. In both cases these minima crucially depend on the size of the underlying alphabet. In this paper we revisit their results using a different approach. The crucial observation is in Section~\ref{two}: the set of finite words over a finite alphabet containing at most $n$ distinct palindromic factors (resp., whose largest palindromic factor is of length at most $n$) is regular. The companion paper to this one is \cite{Fleischer&Shallit:2019}, where some of the same ideas are used. \section{Palindromes and regularity} \label{two} Let $x$ be a finite or infinite word. The set of all of its factors is written $\Fac(x)$, and the set of its factors that are palindromes is written $\PalFac(x)$. Let $P_\ell (\Sigma)$ (resp., $P_{\leq \ell} (\Sigma)$) be the set of all palindromes of length $\ell$ (resp., length $\leq \ell$) over $\Sigma$. Of course, since both of these sets are finite, they are regular. \begin{theorem} Let $S$ be a finite set of palindromes over an alphabet $\Sigma$. Then the language $$ C_\Sigma(S) := \{ x \in \Sigma^* \ : \ \PalFac(x) \subseteq S \} $$ is regular. \label{one} \end{theorem} \begin{proof} Let $\ell$ be the length of the longest palindrome in $S$. We claim that $ \overline{C_\Sigma (S)} = L$, where $$L = \bigcup_{t \in P_{\leq \ell+2} - S} \Sigma^* \, t \ \Sigma^* . $$ \noindent $\overline{C_\Sigma(S)} \subseteq L$: If $x \in \overline{C_\Sigma(S)}$, then $x$ must have some palindromic factor $y$ such that $y \not\in S$. If $|y| \leq \ell+2$, then $y \in P_{\leq \ell+2} - S$. If $|y| > \ell+2$, we can write $y = uvu^R$ for some palindrome $v$ such that $|v| \in \{ \ell+1, \ell+2 \}$. Hence $x$ has the palindromic factor $v$ and $v \in P_{\leq \ell+2} - S$. In both cases $x \in L$. \bigskip \noindent $L \subseteq \overline{C_\Sigma(S)}$: Let $x \in L$. Then $x \in \Sigma^* \, t \ \Sigma^* $ for some $t \in P_{\leq \ell+2} - S$. Hence $x$ has a palindromic factor outside the set $S$ and so $x \not\in C_\Sigma(S)$. \bigskip Thus we have written $\overline{C_\Sigma(S)}$ as the finite union of regular languages, and so $C_\Sigma(S)$ is also regular. \end{proof} \begin{remark} The set $P_{\leq \ell+2} (\Sigma) - S$ can be fairly large. However, because $$ \Sigma^* \, x \, \Sigma^* \subseteq \Sigma^* \, y \, \Sigma^*$$ if $y$ is a factor of $x$, we can replace $P_{\leq \ell+2} (\Sigma) - S$ in Theorem~\ref{one} with the subset of its minimal elements under the factor ordering. (An element $x \in T$ is {\it minimal\/} if $x, y \in T$ with $y$ a factor of $x$ implies that $x = y$.) This typically will have many fewer elements. \end{remark} \begin{corollary} \leavevmode \begin{itemize} \item[(a)] Let $D_{\ell} (\Sigma)$ be the set of finite words over $\Sigma$ containing at most $\ell$ distinct palindromes as factors (including the empty word). Then $D_{\ell} (\Sigma)$ is regular. \item[(b)] Let $E_{\ell} (\Sigma)$ be the set of finite words over $\Sigma$ containing no palindrome of length $> \ell$ as a factor. Then $E_{\ell} (\Sigma)$ is regular. \item[(c)] Let $R_{\ell,m} (\Sigma)$ be the set of finite words over $\Sigma$ containing no even palindrome of length $>\ell$ nor any odd palindrome of length $>m$ as factors. Then $R_{\ell,m}(\Sigma)$ is regular. \item[(d)] Let $T_{\ell,m} (\Sigma)$ be the set of finite words over $\Sigma$ containing at most $\ell$ even palindromes and $m$ odd palindromes. Then $T_{\ell,m}(\Sigma)$ is regular. \end{itemize} \label{palreg} \end{corollary} \begin{proof} \leavevmode \begin{itemize} \item[(a)] Note that no word in $D_{\ell} (\Sigma)$ cannot contain a palindrome of length $r \geq 2\ell$ as a factor, because then it would also contain palindromes of length $0, 2,\ldots r-2$ as factors ($r$ even) or length $1,3, \ldots r-2$ as factors ($r$ odd). In both cases this gives at least $\ell+1$ distinct palindromes. Hence $$ D_\ell (\Sigma) = \bigcup_{ {|S| \leq \ell} \atop {S \subseteq P_{\leq 2\ell-1} (\Sigma)} } C_{\Sigma} (S) , $$ the union of a finite number of regular languages. \item[(b)] We have $E_\ell (\Sigma) = C_{\Sigma} (P_{\leq \ell} (\Sigma)) $. \item[(c)] We have $R_{\ell,m} (\Sigma) = C_{\Sigma} \biggl( \bigl(P_{\leq \ell} (\Sigma) \, \cap \, (\Sigma^2)^* \bigr) \ \cup \ \bigl(P_{\leq m} (\Sigma) \, \cap \, \Sigma(\Sigma^2)^* \bigr) \biggr) $. \item[(d)] We have $$T_{\ell, m} (\Sigma) = \bigcup_{{|S_1| \leq \ell} \atop {S_1 \subseteq \bigcup_{0 \leq i < \ell} P_{2i} (\Sigma)}} C_\Sigma (S_1) \ \cup \ \bigcup_{{|S_2| \leq m} \atop {S_1 \subseteq \bigcup_{0 \leq i < m} P_{2i+1} (\Sigma)}} C_\Sigma (S_2).$$ \end{itemize} \end{proof} Theorem~\ref{one} and Corollary~\ref{palreg} implicitly provide an algorithm for actually finding the DFA's accepting the languages $D_\ell (\Sigma)$, $E_\ell (\Sigma)$, $R_{\ell,m} (\Sigma)$, and $T_{\ell,m} (\Sigma)$: namely, construct automata for each term of the unions and intersections, and combine them using standard techniques (e.g., \cite[Sect.~3.2]{Hopcroft&Ullman:1979}), possibly using minimization at each step. This can be carried out, for example, using a software package such as {\tt Grail} \cite{Raymond&Wood:1994,Campeanu:2019}. However, our experience shows that the intermediate automata so generated can be quite large. Instead, we use a different approach to construct the automata directly, which we now illustrate for the case of $D_\ell(\Sigma)$, as follows. The states are of the form $\Sigma^{\leq 2\ell-1} \times 2^U$, where $U$ is the set of the nonempty palindromes of length at most $2\ell-1$. Given a state of the form $(x, S)$, upon reading the letter $a$, we go to the new state $(y, T)$, where $y = xa$ (if $|xa| \leq 2\ell-1$) or the suffix of length $2\ell-1$ of $xa$ (if $|xa| = 2\ell$), and $T = S \ \cup \ \PalFac(xa)$. If $|T| > \ell$, it is labeled as a rejecting state. The resulting automaton, as described, still can be rather large. However, many states will not be reachable from the start state. Instead, we construct all reachable states using a queue, in a breadth-first manner starting from the initial state $(\varepsilon, \emptyset)$. As soon as we reach a state $(x,S)$ with $|S| > \ell$, the state is labeled as a dead state and we do not append it to the queue. We implemented this idea in Dyalog APL. Our program creates an automaton in {\tt Grail} format which can then be minimized using {\tt Grail}. Our approach allows us to recover many of the results of Fici and Zamboni, and even more. For example, the DFA's we compute give us a complete description of {\it all\/} words, both finite and infinite, containing at most $\ell$ distinct palindromic factors. It provides an easy and efficient way to determine whether or not there exist infinite words containing a given avoidance property, and if so, whether some of these words are aperiodic. As corollaries, we can computably determine a linear recurrence giving the number $a(n)$ of such words of length $n$, and the asymptotic growth rate of the sequence $(a(n))_{n \geq 0}$. Finally, our approach replaces a long case-based argument that can be difficult to follow, and is prone to error, with a machine computation that can be verified mechanically. \section{Linear recurrences and automata} \label{three} We summarize some well-known techniques for enumerating the number of length-$n$ words accepted by deterministic finite automata that we use in this paper. For more details, see, for example, \cite[Sect.~3.8]{Shallit:2009} and \cite{Everest&vdp&Shparlinski&Ward:2003}. We introduce some notation and terminology: if $q(X) = q_t X^t + q_{t-1} X^{t-1} + \cdots + q_1X + q_0$ is a polynomial and ${\bf a} = (a(n))_{n \geq 0}$ is a sequence, then $q \circ {\bf a}$ denotes the sequence $(q_t a(t+i) + q_{t-1} a(t+i-1) + \cdots + q_1 a(i+1) + q_0 a(i))_{i \geq 0}$ obtained by taking the dot product of the coefficients of $q$ with sliding ``windows'' of the sequence $\bf a$. If $q \circ {\bf a}$ is the sequence $(0,0,0,\ldots)$, we call $q$ an {\it annihilator\/} of $(a(n))_{n \geq 0}$. It is now easy to verify that if $q, r$ are polynomials, then $(qr) \circ {\bf a} = q \circ (r \circ {\bf a})$. We also define $\Lead(q) = q_t$ to be the leading coefficient of $q$. Suppose $Q = \{ q_0, q_1, \ldots, q_{r-1} \}$ and $A = (Q, \Sigma, \delta, q_0, F)$ is an $r$-state DFA. From this we can compute an $n \times n$ matrix $M$ such that $M[i,j] = \{ a \in \Sigma \ : \ \delta(q_i, a) = q_j \}$. Let $v = [1 \ 0 \ 0 \ \cdots 0]$ be the row vector with a $1$ in the first position and $0$'s elsewhere, and let $w$ be the column vector with $1$'s in positions corresponding to the final states $F$ and $0$'s corresponding to $Q-F$. Then $a(n)$, the number of length-$n$ words accepted by $A$, is $v M^n w$. We can find a linear recurrence for the sequence $(a(n))_{n \geq 0}$ as follows: first, we compute the minimal polynomial $p(X) = X^t + p_{t-1} X^{t-1} + \cdots + p_1X + p_0$ of $M$ using standard techniques. Then $p(M) = 0$, so $M^t + p_{t-1} M^{t-1} + \cdots + p_1M + p_0I = 0$. By multiplying by $M^i$, we get $M^{t+i} + p_{t-1} M^{t+i-1} + \cdots + p_1M^{i+1} + p_0 M^i = 0$. By premultiplication by $v$ and postmultiplication by $w$, we get $v M^{t+i} w + p_{t-1} v M^{t+i-1} w + \cdots + p_1 v M^{i+1} w + p_0 vM^i w = 0$. Hence $a(t+i) + p_{t-1} a(t+i-1) + \cdots + p_1 a(i+1) + p_0 a(i) = 0$, and hence $(a(n))_{n \geq 0}$ satisfies a linear recurrence with constant coefficients given by the $p_i$. Using our terminology, the polynomial $p$ annihilates $(a(n))_{n \geq 0}$. However, $p$ may not be the lowest-degree annihilator of $(a(n))_{n \geq 0}$. A lower degree annihilator will necessarily be a divisor of the polynomial $p$. The lowest degree annihilator can be determined using an algorithm based on the following theorem, which seems to be new. \begin{theorem} Suppose the polynomial $p(X)$, with leading coefficient nonzero, annihilates the sequence $(a(n))_{n \geq 0}$ and suppose $q(x) \, | \, p(x)$. If the polynomial ${p \over q}$ also annihilates the sequence $(a(n))_{n \geq 0}$ for the first $\deg q$ consecutive windows of $(a(n))_{n \geq 0}$, then it annihilates all of $(a(n))_{n \geq 0}$. \end{theorem} \begin{proof} Suppose $p \over q$ annihilates ${\bf a} = (a(n))_{n \geq 0}$ for the first $s := \deg q$ consecutive windows of $\bf a$, but not all of $\bf a$. Write ${p \over q} = d_t X^t + \cdots + d_1 X + d_0$. Define $(f(n))_{n \geq 0} = {p \over q} \circ (a(n))_{n \geq 0}$. Thus $ f(n) = \sum_{0 \leq i \leq t} d_i a(n+i)$. Then by hypothesis we have $f(n) = 0$ for $n = 0, 1, \ldots , s-1$. Now $p$ annihilates $(a(n))_{n \geq 0}$, so $q$ annihilates $(f(n))_{n \geq 0}$. Let $r$ be the least index such that $f(r) \not= 0$. So $(f(0), f(1), \ldots, f(r)) = (0,0, \ldots ,0, e)$ for some $e \not= 0$. But $q$ annihilates $(f(n))_{n \geq 0}$, so if $r \geq s$ then $q \circ ( \overbrace{0,0, \ldots , 0}^{s-1}, e) = 0$. But $q \circ ( \overbrace{0,0,\ldots ,0}^{s-1} ,e) = e \Lead(q) \not= 0$, a contradiction. \end{proof} This gives us the following algorithm for finding the lowest-degree annihilator of a recurrence. \bigskip\hrule \begin{tabbing} {\sc Algorithm LDA}$(p, {\bf a})$ \\ \ \\ Write $p := q_1 q_2 \cdots q_m$, the product of (not necessarily distinct) irreducible factors. \\ For \= $i := 1$ to $m$ do \\ \> $r := p/q_i$ \\ \> If $r$ annihilates the first $\deg r$ windows of $\bf a$, set $p := p/q_i$. \\ return($p$); \end{tabbing} \smallskip\hrule \bigskip Terms of the form $X^n$ in an annihilator can be removed if one assumes that the recurrence begins at $a(n)$ instead of $a(0)$. For this reason, in this paper, we do not report such terms in our annihilators. In our computations, we used {\tt Maple} to compute minimal polynomials (via the {\tt LinearAlgebra} package) and factor them. \section{Automata and infinite words} We recall some material from the companion paper \cite{Fleischer&Shallit:2019}. The DFA's generated in this paper are for regular languages $L$ that are defined by avoidance of a finite set $S$ of finite words. Such languages are called {\it factorial}; that is, every factor of a word of $L$ is also a word of $L$. The minimal DFA $M = (Q, \Sigma, \delta, q_0, F)$ for a factorial language $L \not= \Sigma^*$ has exactly one nonaccepting state, which is the dead state. (A state is {\it dead\/} if it is nonaccepting and transitions to itself on all letters of the alphabet $\Sigma$.) {\it In this paper, we do not display this dead state in our figures, nor count it in our discussion of the cardinality of a DFA's states.} The (one-sided) infinite words with the given avoidance property are then given by the infinite paths through $M$, starting at the start state $q_0$. A state $q$ is called {\it recurrent} if there is a nonempty word $w$ such that $\delta(q,w) = q$. A state $q$ is called {\it birecurrent} if there are two noncommuting words $x_0, x_1$ such that $\delta(q,x_0) = \delta(q,x_1) = q$. As shown in \cite{Fleischer&Shallit:2019}, an infinite word having the desired avoidance property exists iff $M$ has a recurrent state, and aperiodic infinite words exist iff $M$ has a birecurrent state. In this latter case, there are actually uncountably many such words. As shown in \cite{Fleischer&Shallit:2019}, these correspond to the image, under the morphism $h: 0 \rightarrow x_0$ and $1 \rightarrow x_1$ of an aperiodic binary word. Furthermore, we can find infinite words avoiding $S$ that are (a) uniformly recurrent and aperiodic (b) linearly recurrent and aperiodic and (c) $k$-automatic for any $k \geq 2$ and uniformly recurrent and aperiodic and (d) the fixed point of a primitive uniform morphism, which is uniformly recurrent. To see this, note that the image under a nonerasing morphism of a uniformly recurrent infinite word is uniformly recurrent. So it suffices to apply $h$ to any uniformly recurrent binary word, such as the Thue-Morse word $\bf t$ \cite{Allouche&Shallit:1999}. Similarly, the image under a nonerasing morphism of a linearly recurrent infinite word is linearly recurrent. To see that we can find a $k$-automatic word with the desired properties, note that we can start with any $k$-automatic word that is uniformly recurrent and aperiodic (for example, the fixed point of $0 \rightarrow 0^{k-1} 1$ and $ 1 \rightarrow 1 0^{k-1}$) and apply the morphism $h$ to it. Finally, assume that $x_0$ and $x_1$ are chosen such that for some $a\in \{ 0, 1\}$ we have $x_a$ starts with $a$. Let $b = \{ 0, 1 \} - \{a\}$. Write $g(a) = x_a x_b$ and $g(b) = x_b x_a$. Then $g^\omega(a)$, the infinite fixed point of $g$ starting with $a$, is uniformly recurrent. In what follows, we use the alphabet $\Sigma_k = \{ 0, 1, \ldots, k-1 \}$. A-numbers in the paper refer to sequences from the {\it On-Line Encyclopedia of Integer Sequences} \cite{Sloane:2019}. \section{Minimizing the number of palindromes} We define $d_{k,\ell} (n)$ to be the number of length-$n$ words in $D_\ell(\Sigma_k)$. \subsection{Alphabet size 2} \begin{theorem}{(Fici-Zamboni)} There are infinite binary words containing at most 9 palindromes. All are periodic, and of the form $x^\omega$ for $x$ a conjugate of either $001011$ or $001101$. There are no infinite binary words containing at most 8 palindromes. \label{thm2-9} \end{theorem} \begin{proof} We construct the DFA for $D_9 (\Sigma_2)$ as in Section~\ref{two}. It has 611 states before minimization and 98 after minimization, and we omit it here. No state is birecurrent, but there are 12 recurrent states. Examining the associated paths easily gives the result. To see the result for 8 palindromes, we can construct the DFA for $D_8(\Sigma_2)$. It has 259 states before minimization and 23 after minimization. No state is recurrent. The longest word accepted is of length $8$. Alternatively, one can prove this result using a simple breadth-first search of the space of words. \end{proof} \begin{theorem}{(Restatement of Fici-Zamboni)} There are exactly 40 infinite binary words containing exactly 10 palindromes. All are ultimately periodic, and are of the following forms: \begin{itemize} \item $x^\omega$ for $x$ a conjugate of $0001011$, $0001101$, $0010111$, or $0011101$; \item $y (001011)^\omega$ for $y \in \{ 0, 01, 111, 0011, 11011, 101011 \}$; \item $y (001101)^\omega$ for $y \in \{ 0, 11, 001, 0101, 11101, 101101 \}$. \end{itemize} \end{theorem} \begin{proof} We create the automaton for $D_{10} (\Sigma_2)$ as in Section~\ref{two}. It has 1655 states before minimization and 280 after. None of these states are birecurrent. By examining the possible infinite paths, we see these include those of Theorem~\ref{thm2-9} and the ones listed above. \end{proof} \begin{theorem} There are uncountably many aperiodic, uniformly recurrent infinite binary words containing exactly 11 palindromes. \end{theorem} \begin{proof} Using the method in Section~\ref{two}, we can construct the DFA for $D_{11} (\Sigma_2)$. It has 5253 states before minimization, and 810 states afterwards, for $D_{11} (\Sigma_2)$. We do not give the latter automaton here, as it is too large to display in a reasonable way, but it can be downloaded from the second author's website at State 738 is birecurrent, with two paths labeled $x_0 = 0001011001011$ and $x_1 = 001011001011$. \end{proof} \begin{corollary} The number of binary words containing at most 11 distinct palindromic factors (including the empty word) is $(d_{2,11} (n))_{n \geq 0}$, where \begin{displaymath} \begin{split} &(d_{2,11}(0), \ldots, d_{2,11}(41)) = (1,2,4,8,16,32,64,128,256,512,1024,292,270,268, 276,276,288, \\ & 320,340, 364, 388,404,428,476,512,560,610,644,692,768,840,924, 1020,1100,1190,1316, \\ & 1452,1612, 1786,1952,2134,2348) \end{split} \end{displaymath} and \begin{align*} d_{2,11}(n) &= -d_{2,11}({n-1}) - d_{2,11}({n-2}) - d_{2,11}({n-3}) -d_{2,11}({n-4}) - d_{2,11}({n-5}) + 2d_{2,11}({n-6}) \\ &\quad + 4d_{2,11}({n-7}) + 5d_{2,11}({n-8}) + 5 d_{2,11}({n-9}) + 5 d_{2,11}({n-10}) + 5 d_{2,11}({n-11}) \\ & \quad + 2 d_{2,11}({n-12}) -3 d_{2,11}({n-13}) + -6d_{2,11}({n-14}) -8 d_{2,11}({n-15}) -8 d_{2,11}({n-16}) \\ & \quad -8 d_{2,11}({n-17}) -7 d_{2,11}({n-18}) - 3 d_{2,11}({n-19}) + 3 d_{2,11}({n-21}) + 4d_{2,11}({n-22}) \\ & \quad + 4 d_{2,11}({n-23}) + 4 d_{2,11}({n-24}) + 3 d_{2,11}({n-25}) +2 d_{2,11}({n-26}) + d_{2,11}({n-27}) \end{align*} for $n \geq 42$. Asymptotically, $d_{2,11}(n) \sim c \cdot \alpha^n$, where $\alpha \doteq 1.1127756842787054706297$ is the largest positive real zero of $X^7 - X - 1$ and $c \doteq 20.665$. \end{corollary} \begin{proof} Using {\tt Maple}, we computed the minimal polynomial for the matrix of the 811-state DFA described above. It is \begin{multline*} X^{15} (X-1)(X-2)(X+1)(X^2 + 1)(X^2 + X + 1)(X^2 - X + 1)(X^7 - X - 1) \\ (X^4 + 1)(X^6 + X^5 + X^4 + X^3 + X^2 + X + 1)(X^8 - X^2 - 1) . \end{multline*} Next, using the procedure described in Section~\ref{three}, we can find the minimal annihilator of the recurrence. It is \begin{equation} (X-1)(X+1)(X^2+X+1)(X^2-X+1)(X^7-X-1)(X^6 + X^5 + X^4 + X^3 + X^2 + X + 1)(X^8 - X^2 - 1). \label{polys} \end{equation} When expanded, this gives the coefficients of the annhiliator of the sequence $(d_{2,11}(n))_{n \geq 0}$, which are given above. To get the asymptotic behavior of the recurrence, we must find the largest real zero of the polynomials given in \eqref{polys}. It is the largest real zero of $X^7-X-1$, which is approximately $1.1127756842787054706297$. \end{proof} \begin{remark} This is sequence \seqnum{A330127} in the OEIS. \end{remark} In their paper, Fici and Zamboni constructed a uniformly recurrent aperiodic binary word containing 13 palindromic factors, and whose set of factors is closed under reversal. We achieve the same result using a different construction and a different proof. \begin{theorem} Define $G_0 = 001101000110$ and $G_{n+1} = G_n 01 G_n^R$ for $n \geq 0$. Then $G_{\infty} = \lim_{n \rightarrow \infty} G_n$ is uniformly recurrent, aperiodic, and has 13 palindromic factors. \end{theorem} \begin{proof} We start by constructing the DFA for the language $D_{13}(\Sigma_2)$ using the method described in Section~\ref{two}. This DFA $M$ has 93125 states before minimization and 6522 states after minimization. The unique dead state is numbered 3012. Next, we look at the transformations $\tau_n$ of states induced by the words $G_n$. We claim that \begin{itemize} \item $\tau_{G_n} = \tau_{G_{n+1}}$ for $n \geq 2$; \item $\tau_{G_n} = \tau_{G_n^R}$ for $n \geq 1$. \end{itemize} which can be easily verified by induction using the transition function for $M$. The resulting transformations of states for $n \geq 2$ are as follows: \begin{align*} 0 & \xrightarrow{\ G_n\ } 4882 \xrightarrow{\ 01 \ } 5058 \xrightarrow{\ G_n^R\ } 4882 \\ 0 & \xrightarrow{\ G_n^R\ } 4882 \xrightarrow{\ 10 \ } 5059 \xrightarrow{\ G_n\ } 4882 \end{align*} Since these paths do not end in the unique nonaccepting state, the corresponding words contain at most $13$ palindromes. It is easy to see that the word $G_\infty$ is uniformly recurrent and closed under reversal. This is left to the reader. The fact that $G_\infty$ is not ultimately periodic follows from \cite[Thm.~4]{Shallit:1982b}. \end{proof} \subsection{Alphabet size 3} \begin{theorem}{(Fici-Zamboni)} If a ternary infinite word contains $4$ palindromes (including the empty word), it is necessarily of the form $(abc)^\omega$ for distinct letters $a,b, c$. No ternary infinite word can contain $3$ or fewer palindromes. \end{theorem} \begin{proof} We construct the DFA for $D_4(\Sigma_3)$ using the algorithm suggested in Section~\ref{two}. It has 52 states and, when minimized, has 18 states. It is depicted below in Figure~\ref{pal3num4}. Only the states numbered $12, 13, 14, 15, 16, 17$ are recurrent, and none of them are birecurrent. The desired result now easily follows from examining the possible paths through these states. \begin{figure}[H] \begin{center} \includegraphics[width=5.5in]{pal3num4min.pdf} \end{center} \caption{Automaton for ternary words containing at most $4$ palindromes} \label{pal3num4} \end{figure} To see that no ternary infinite word can contain $3$ or fewer palindromes, we can perform the same construction as above, but for $3$ palindromes. The resulting automaton has 13 states (3 when minimized) and no recurrent states. We omit it here. Alternatively, one can prove this result with a simple breadth-first search of the space of words. \end{proof} \begin{theorem} There are uncountably many aperiodic ternary words containing at most $5$ palindromic factors. \end{theorem} \begin{proof} We can construct the automaton for $D_5(\Sigma_3)$ as described in Section~\ref{two}. It has 319 states before minimization and 69 states after. We do not depict it here, as it is too large to visualize clearly. The state 39 is birecurrent, with paths labeled $x_0 = 0012$ and $x_1 = 012$. \end{proof} \begin{corollary} The number of ternary words containing at most $5$ palindromic factors is $d_{3,5}(n) $, where $(d_{3,5}(0), \ldots, d_{3,5}(8)) = (1,3,9,27,81,42,54,66,78)$ and $d_{3,5}(n) = d_{3,5}(n-3) + d_{3,5}(n-4)$ for $n \geq 9$. Asymptotically we have $d_{3,5}(n) \sim c \alpha^n$ where $\alpha \doteq 1.2207440846$ and $c \doteq 16.07007$. \end{corollary} \begin{proof} The minimal polynomial of the corresponding matrix is $$ X^5 (X-1) (X-3) (X^2 + X + 1) (X^4 - X - 1) .$$ Using the method in Section~\ref{three}, we can find the minimal annihilator of the sequence, which is $X^4 - X - 1$. The result now follows. \end{proof} \begin{remark} This is sequence \seqnum{A329023} in the OEIS. We have $d_{3,5}(n) = 6 \cdot$\seqnum{A164317}$(n)$ for $n \geq 5$. \end{remark} \section{Lengths of palindromes} Instead of minimizing the total number of palindromes, Fici and Zamboni also considered minimizing the length of the longest palindrome. We can also do that with our method. We define $e_{k,\ell} (n)$ to be the number of length-$n$ words in $E_\ell (\Sigma_k)$. \subsection{Alphabet size 2} \begin{theorem}{(Restatement of Fici-Zamboni)} There are exactly 20 infinite binary words having no palindromes of length $>4$, and all are ultimately periodic. They are as follows: \begin{itemize} \item $x^\omega$ for $x$ a conjugate of $001011$; \item $x^\omega$ for $x$ a conjugate of $001101$; \item $(0+00+111+1111)(001011)^\omega$; \item $(0+00+11101+111101)(001101)^\omega$. \end{itemize} \label{pal2len4-thm} \end{theorem} \begin{proof} The automaton for $E_4(\Sigma_2)$ is depicted in Figure~\ref{pal2len4}, and the only infinite paths are those given. (There are no birecurrent states.) \end{proof} \begin{figure}[H] \begin{center} \includegraphics[width=6.5in]{pal2len4m.pdf} \end{center} \caption{Automaton for binary words containing no palindromes of length $>4$} \label{pal2len4} \end{figure} \begin{theorem} There are uncountably many uniformly recurrent binary words containing no palindromes of length $>5$. They are the labels of the paths through the automaton in Figure~\ref{pal2len5}. \end{theorem} \begin{figure}[H] \begin{center} \includegraphics[width=6.5in]{pal2len5m.pdf} \end{center} \caption{Automaton for binary words containing no palindromes of length $>5$} \label{pal2len5} \end{figure} \begin{proof} As before. There are 719 states in the unminimized automaton for $E_5 (\Sigma_2)$ and 62 states in the minimized one. State 44 is birecurrent, with paths $x_0 = 01010110$ and $x_1 = 0010101110$. \end{proof} \begin{theorem} The sequence $(e_{2,5} (n))_{n \geq 0}$ counting the number of binary words of length $n$ containing no palindromes of length $>5$ satisfies the recurrence $$e_{2,5}(n) = 3 e_{2,5}(n-6) + 2 e_{2,5}(n-7) + 2 e_{2,5}(n-8) + 2 e_{2,5}(n-9) + e_{2,5}(n-10)$$ for $n \geq 20$. Asymptotically $e_{2,5}(n) \sim c \alpha^n$ where $\alpha \doteq 1.36927381628918060784\cdots$ is the positive real zero of the equation $X^{10}-3X^4-2X^3-2X^2-2X-1$, and $c = 9.8315779\cdots$. \end{theorem} \begin{proof} The minimal polynomial of the corresponding matrix is \begin{equation} X^{10} (X-2) (X^{10} + X^4 - 2X^3 -2X^2 -2X - 1)(X^{10} -3X^4 -2X^3 -2X^2 -2X - 1). \label{pal35} \end{equation} The technique described in Theorem~\ref{three} can be used to find the minimal annihilator for the recurrence. It is the last term in the factorization \eqref{pal35}. \end{proof} \begin{remark} The sequence $e_{2,5} (n)$ is sequence \seqnum{A329824} in the OEIS. \end{remark} \subsection{Alphabet size 3} \begin{theorem}{(Fici-Zamboni)} The only infinite ternary words having no palindromes of length $>1$ are those of the form $(abc)^\omega$ for distinct letters $a,b,c$. \end{theorem} \begin{proof} The automaton for $E_1(\Sigma_3)$ has 16 states before minimization and 10 states after. We omit it here. There are no birecurrent states, and the only infinite paths are those given. \end{proof} \begin{theorem} There are uncountably many ternary words containing no palindromes of length $>2$. \label{pal3len2-thm} \end{theorem} \begin{proof} We can construct the automaton for $E_2(\Sigma_3)$ as in Section~\ref{two}. It has 67 states unminimized and 19 states when minimized. State $6$ is birecurrent, with paths labeled $x_0 = 211002$ and $x_1 = 11002$. \begin{figure}[H] \begin{center} \includegraphics[width=6in]{pal3len2m.pdf} \end{center} \caption{Automaton for ternary words containing no palindromes of length $>2$} \label{pal3len2} \end{figure} \end{proof} The Fibonacci numbers $F_n$ are defined by $F_0 = 0$ and $F_1 = 1$ and $F_n = F_{n-1} + F_{n-2}$. \begin{corollary} The number $e_{3,2}(n)$ of length-$n$ ternary words containing no palindromes of length $>2$ is $6F_{n+1}$ for $n \geq 3$. \end{corollary} \begin{proof} The minimal polynomial of the matrix is $X^3 (X-3)(X^2 - X - 1)(X^4 + X^3 + 2X^2 + 2X + 1)$. The minimal annihilator is $X^2 - X - 1$. The result now follows easily. \end{proof} \begin{remark} The sequence $e_{3,2}(n)$ is sequence \seqnum{A330010} in the OEIS. \end{remark} \subsection{Alphabet size 4} Fici and Zamboni proved that, over the alphabet $\Sigma_4$, there is an infinite aperiodic uniformly recurrent word whose only palindromes are $\varepsilon, 0, 1, 2, 3$. We show how to handle this using our method. \begin{theorem} There is an infinite aperiodic uniformly recurrent word over $\Sigma_4$ whose only palindromes are $\varepsilon, 0, 1, 2, 3$. \end{theorem} \begin{proof} To find the words avoiding all palindromes as factors except these $5$, we can use Theorem~\ref{one}. After computing the minimal elements, it suffices to avoid the factors $$ \{00,11,22,33,010,020,030,101,121,131,202,212,232,303,313,323\} .$$ The minimal DFA is depicted in Figure~\ref{fig1}. \begin{figure}[H] \begin{center} \includegraphics[width=6.5in]{auts4.pdf} \end{center} \caption{Automaton for $4$-letter alphabet. The dead state, numbered 5, is omitted.} \label{fig1} \end{figure} The state numbered $6$ is birecurrent, with two paths labeled $2301$ and $301$. Let ${\bf x}$ be an aperiodic uniformly recurrent word over $\{ 0, 1\}$ and define the morphism $h(0) = 2301$ and $h(1) = 301$. For example, we can take $\bf x$ to be the Thue-Morse word. Then $h({\bf x})$ has the desired properties. \end{proof} \begin{corollary} The number $e_{4,1}(n)$ of finite words over $\Sigma_4$ having all their palindromic factors contained in $\{ \varepsilon, 0, 1, 2, 3 \}$ is $3 \cdot 2^n$ for $n \geq 2$. \end{corollary} \begin{proof} The minimal polynomial of the matrix corresponding to the automaton is $X^2 (X-1) (X-2)(X-4)(X+1)(X^2 + X + 2)$. Using the procedure in Section~\ref{three} we can determine the minimal annihilator,which is $X-2$. It follows that $e_{4,1} (n) = 3 \cdot 2^n$ for $n \geq 2$. \end{proof} Berstel, Boasson, Carton, and Fagnot \cite{Berstel&Boasson&Carton&Fagnot:2009} constructed an infinite word over $\Sigma_4$ that is uniformly recurrent, has exactly $5$ palindromic factors, and further is closed under reversal, as follows: define $B_0 = 01$ and $B_{n+1} = B_n 23 B_n^R$. This is an example of {\it perturbed symmetry}; see \cite{Dekking&MendesFrance&vanderPoorten:1982} for more details. We can verify their construction using our method. Consider the DFA in Figure~\ref{fig1}; then each word $w$ induces a transformation $\tau_w$ of the states given by $q \rightarrow \delta(q,w)$. We claim that \begin{itemize} \item[(a)] $\tau_{B_n} = \tau_{B_n^R} = (9, 5, 5, 9, 9, 5, 5, 5, 5, 5, 9, 9, 5, 5, 9, 5, 5, 9)$ for $n \geq 1$; \item[(b)] $\tau_{23} = (17, 17, 17, 5, 5, 5, 17, 5, 5, 17, 5, 5, 17, 17, 5, 5, 5, 5)$. \item[(c)] $\tau_{32} = (14, 14, 14, 5, 5, 5, 14, 5, 5, 14, 5, 5, 5, 5, 5, 14, 14, 5)$. \end{itemize} The claims about $\tau_{B_1}$, $\tau_{B_1^R}$, $\tau_{23}$, and $\tau{32}$ are easily verified. We now prove the claim about $B_n$ by induction. The reader can now check that $\tau_{B_{n+1}} = \tau_{B_n 23 B_n^R} = \tau_{B_n}$ and $\tau_{B_{n+1}^R} = \tau_{B_n 32 B_n^R} = \tau_{B_n}$. Since $0$ is mapped to accepting state $9$ by $B_n$, it follows that each $B_n$ has the desired properties. \section{Odd and even palindromes} In order to illustrate that the technique in this paper has wider applicability, we now turn to a topic not covered in the paper of Fici and Zamboni. Because an odd palindrome factor of length $\ell$ implies the existence of odd palindrome factors of all shorter lengths, and the same for even palindrome factors, it makes sense to consider minimizing the lengths of odd and even palindrome factors separately. This is what we do in this section. We define $r_{k,\ell, m} (n)$ to be the number of length-$n$ words in $R_{\ell,m} (\Sigma_k)$. \subsection{Alphabet size 2} \begin{theorem} There are uncountably many uniformly recurrent binary words having longest even palindrome factor of length $\leq 2$ and longest odd palindrome of length $\leq 5$. \end{theorem} \begin{figure}[H] \begin{center} \includegraphics[width=6.5in]{a2palm2-5.pdf} \end{center} \caption{Automaton for binary words with longest even palindrome factor of length $\leq 2$ and longest odd palindrome of length $\leq 5$.} \label{palm2-25} \end{figure} \begin{proof} We construct the automaton for $R_{2,5}(\Sigma_2)$ as discussed above. Before minimization it has 155 states. After minimization it has 44 states. State 18 is birecurrent, with cycles labeled $x_0 = 10100011$ and $x_1 = 1010100011$. \end{proof} \begin{theorem} Let $(r_{2,2,5}(n))_{n \geq 0}$ denote the number of finite binary words containing longest even palindrome factor of length $\leq 2$ and longest odd palindrome of length $\leq 5$. Then $r_{2,2,5}(n) = r_{2,2,5}({n-8}) + r_{2,2,5}({n-10})$ for $n \geq 16$. Furthermore, $r_{2,2,5}(n) \sim C_1 \alpha^n + C_2 (-\alpha)^n$, $C_1 \doteq 15.991809$, $C_2 \doteq 0.023895$, and $\alpha \doteq 1.0804184273981 $ is the largest real zero of $X^{10} - X^2 - 1$. \end{theorem} \begin{proof} The minimal polynomial of the corresponding matrix is $$X^6(X-2)(X^{10} - X^2 - 1).$$ The minimal annihilator of the recurrence can be determined by using the ideas in Section~\ref{three}; it is $X^{10} - X^2 - 1$. \end{proof} \begin{remark} The sequence $r_{2,2,5} (n)$ is sequence \seqnum{A330130} in the OEIS. \end{remark} The case of longest even palindrome factor of length $\leq 4$ and longest odd palindrome of length $\leq 3$ is already covered in Theorem~\ref{pal2len4-thm}. \begin{theorem} There are uncountably many uniformly recurrent binary words over having longest even palindrome factor of length $\leq 6$ and longest odd palindrome of length $\leq 3$. \end{theorem} \begin{figure}[H] \begin{center} \includegraphics[width=6.5in]{a2palm6-3.pdf} \end{center} \caption{Automaton for binary words with longest even palindrome factor of length $\leq 6$ and longest odd palindrome of length $\leq 3$.} \label{palm2-63} \end{figure} \begin{proof} We construct the automaton for $R_{6,3}(\Sigma_2)$ as discussed above. Before minimization it has 477 states. After minimization it has 60 states. State 17 is birecurrent, with cycles labeled $x_0 = 110010$ and $x_1 = 1111000010$. \end{proof} \begin{theorem} Let $(r_{2,6,3}(n))_{n \geq 0}$ denote the number of finite binary words containing longest even palindrome factor of length $\leq 6$ and longest odd palindrome of length $\leq 3$. Then $r_{2,6,3}(n) = r_{2,6,3}({n-6}) + 2r_{2,6,3}({n-8}) + 3r_{2,6,3}({n-10}) + r_{2,6,3}({n-14})$ for $n \geq 21$. Furthermore, and $r_{2,6,3}(n) \sim C_1 \alpha^{n} + C_2 (-\alpha)^n$, where $C_1 \doteq 11.58110542$, $C_2 \doteq 0.00264754$, and $\alpha \doteq 1.244528319539183$ is the largest real zero of $X^{14} - X^8 -2X^6 - 3X^4 - 1$. \end{theorem} \begin{proof} The minimal polynomial of the corresponding matrix is $$X^7 (X-2)(X^2 + 1)(X^{14} - X^8 -2X^6 - 3X^4 - 1)(X^{12} - X^{10} + X^8 - 2X^6 + X^2 - 1).$$ The minimal annihilator of the recurrence can be determined by using the ideas in Section~\ref{three}; it is $X^{14} - X^8 -2X^6 - 3X^4 - 1$. \end{proof} \begin{remark} The sequence $r_{2,6,3}$ is sequence \seqnum{A330131} in the OEIS. \end{remark} \subsection{Alphabet size 3} \begin{theorem} There are uncountably many uniformly recurrent words over $\Sigma_3$ containing no (nonempty) even palindromic factors and longest odd palindrome of length $\leq 3$. \end{theorem} \begin{figure}[H] \begin{center} \includegraphics[width=6.5in]{palm3-03.pdf} \end{center} \caption{Automaton for ternary words with no even palindromic factors and longest odd palindrome of length $3$} \label{palm3-03} \end{figure} \begin{proof} We construct the automaton for $R_{0,3} (\Sigma_3)$ as discussed in Section~\ref{two}. Before minimization it has 88 states. After minimization it has 34 states. State 16 is birecurrent, with cycles labeled $x_0 = 021210102$ and $x_1 = 1210102$. \end{proof} \begin{theorem} Let $(r_{3,0,3} (n))_{n \geq 0}$ denote the number of finite ternary words containing no (nonempty) even palindromic factors and longest odd palindrome of length $3$. Then $$r_{0,3} (n) = r_{0,3}({n-1}) + r_{0,3} ({n-3})$$ for $n \geq 7$. Furthermore, $r_{0,3} (n) \sim C \alpha^n$, where $C \doteq 5.37711043$ and $\alpha \doteq 1.465571231876768$ is the largest real zero of $X^3 - X^2 - 1$. \end{theorem} \begin{proof} The minimal polynomial of the corresponding matrix is $$X^4 (X-3) (X^2 - X + 1)(X^3 - X^2 - 1)(X^4 + 2X^3 + 2X^2 + X + 1).$$ The minimal annihilator of the recurrence can be determined by using the technique in Section~\ref{three}; it is $X^3 - X^2 - 1$. \end{proof} \begin{remark} The sequence is \seqnum{A330132} in the OEIS. $r_{0,3} (n) = 6 \cdot$ \seqnum{A000930}$(n-1)$ for $n \geq 5$, where \seqnum{A000930} is the Narayana cow sequence. \end{remark} The case of largest even palindrome of length $2$ and largest odd palindrome of length $1$ is already covered in Theorem~\ref{pal3len2-thm}. \section{Number of odd and even palindromes} Our final application is to infinite words containing a specified number of even and odd palindromes. We define $t_{k,\ell,m}(n)$ to be the number of length-$n$ words in $T_{\ell,m} (\Sigma_k)$. \subsection{Alphabet size 2} Here, instead of providing the details, we simply summarize our results in tabular form. The minimal annihilators for the sequences can be computed from the data we computed. The following cases have infinite words, but not aperiodic infinite words. \begin{table}[H] \centering \begin{tabular}{|c|c|c|c|c|c|c|} Max number of & Max number of & States & States & Example word \\ even palindromes & odd palindromes & (unminimized) & (minimized) & \\ \hline 3 & 9 & 10795 & 1468 & $01(00010111)^\omega$ \\ 3 & 8 & 3911 & 799 & $1(00010111)^\omega$ \\ 4 & 7 & 7505 & 1181 & $01(0001011)^\omega$ \\ 4 & 6 & 2413 & 530 & $1(0001011)^\omega$ \\ 5 & 5 & 1647 & 419 & $0 (001011)^\omega$ \\ 5 & 4 & 461 & 136 & $(001011)^\omega$ \\ 6 & 5 & 3141 & 604 & $(00001011)^\omega$ \\ 6 & 4 & 699 & 177 & $0(011001)^\omega$ \\ 7 & 4 & 1081 & 261 & $10(011001)^\omega$ \\ 8 & 4 & 1729 & 375 & $1101(001011)^\omega$ \\ \end{tabular} \end{table} The following cases have examples of aperiodic infinite words. \begin{table}[H] \centering \resizebox{\columnwidth}{!}{% \begin{tabular}{|c|c|c|c|c|c|c|} Max number of & Max number of & States & States & $x_0$ & $x_1$ & Birecurrent \\ even palindromes & odd palindromes & (unminimized) & (minimized) & & & state number \\ \hline 3 & 10 & 33685 & 3071 & 00011101 & 0100011101 & 1836 \\ 4 & 8 & 26937 & 2830 & 0010111 & 00010111 & 2364 \\ 5 & 6 & 7495 & 1269 & 001011 & 0001011 & 1035 \\ 7 & 5 & 6741 & 955 & 001011 & 00001011 & 904 \\ 9 & 4 & 2789 & 545 & 001011 & 0011001011 & 450 \\ \end{tabular}% } \end{table} \subsection{Alphabet size 3} The only interesting case is one even palindrome and five odd palindromes. Here the automaton has 6208 states (632 when minimized) and has a birecurrent state, corresponding to $x_0 = 01012$ and $x_1 = 012$. \section{Conclusions} We have reproved most of the theorems in \cite{Fici&Zamboni:2013} using a unified approach based on finite automata. This is evidence for the thesis, previously announced in \cite{Rajasekaran&Shallit&Smith:2019}, that long case-based arguments are good candidates for replacement by algorithms and logical decision procedures. All of the code referred to in this paper is available at \\ \centerline{\url{https://cs.uwaterloo.ca/~shallit/papers.html} \ .} \newcommand{\noopsort}[1]{} \newcommand{\singleletter}[1]{#1}

https://arxiv.org/abs/1512.06193

Ulrich Schur bundles on flag varieties

In this paper, we study equivariant vector bundles on partial flag varieties arising from Schur functors. We show that a partial flag variety with three or more steps does not admit an Ulrich bundle of this form with respect to the minimal ample class. We classify Ulrich bundles of this form on two-step flag varieties F(1,n-1;n), F(2,n-1;n), F(2,n-2;n), F(k,k+1;n) and F(k,k+2;n). We give a conjectural description of the two-step flag varieties which admit such Ulrich bundles.

\section{Introduction}\label{sec-intro} Let $V$ be an $n$-dimensional vector space and $0< k_1 < \cdots < k_r <n$ an increasing sequence of integers. For convenience, set $k_0 = 0$ and $k_{r+1} =n$. The $r$-step partial flag variety $F(k_1, \dots, k_r; n)$ parameterizes partial flags $$W_1 \subset W_2 \subset \cdots \subset W_r \subset V,$$ where $W_i$ is a $k_i$-dimensional subspace of $V$ for $1 \leq i \leq r$. The variety $F(k_1, \dots, k_r; n)$ is endowed with a collection of tautological subbundles $$0=T_0 \subset T_1 \subset T_2 \subset \cdots \subset T_r \subset T_{r+1}= \underline{V} = V \otimes \OO_{F(k_1, \dots, k_r; n)}, $$ where $T_i$ is a vector bundle of rank $k_i$ and $\underline{V}$ denotes the trivial bundle of rank $n$. Let $U_i = T_i/ T_{i-1}$. In this paper, we study the equivariant vector bundles $E_\lambda$ on $F(k_1, \dots, k_r; n)$ defined by tensor products of Schur functors of these tautological bundles $$E_\lambda = \mathbb{S}^{\lambda_1} U_1^* \otimes \mathbb{S}^{\lambda_2} U_2^* \otimes \cdots \otimes \mathbb{S}^{\lambda_{r+1}} U_{r+1}^*,$$ where the partition $\lambda = (\lambda_1|\cdots |\lambda_{r+1})$ is the concatenation of partitions $\lambda_i$ of length $k_i-k_{i-1}$. We call the bundles $E_\lambda$ on the flag variety \emph{Schur bundles}. The question we study in this paper is to determine which Schur bundles on partial flag varieties are Ulrich bundles with respect to the minimal ample class. The Picard group of $F(k_1, \dots, k_r; n)$ is generated by the classes of Schubert divisors. The sum of the Schubert divisors corresponds to a line bundle $\cO(1)$ that defines a projectively normal embedding of $F(k_1, \dots, k_r; n)$ \cite{Ramanathan}. Unless we specify otherwise, we will always consider the partial flag varieties in this embedding. Let $X \subset \PP^m$ be an arithmetically Cohen-Macaulay variety of dimension $d$. A vector bundle $\mathcal{E}$ of rank $r$ on $X$ is Ulrich if $H^i(X, \mathcal{E}(-i)) = 0$ for $i>0$ and $H^j(X, \mathcal{E}(-j-1))=0$ for $j < d$ (see \cite{BaHU}, \cite{BHU}, \cite{ESW}). These are the bundles whose Hilbert polynomials have $d$ zeros at the first $d$ negative integers. Their pushforwards to $\PP^m$ via the inclusion of $X$ has a minimal free resolution where all the syzygies are linear and the $i$-th Betti number is $\deg(X) r {m-d \choose i}$ \cite[Proposition 2.1]{ESW}. Ulrich bundles have important applications in liaison theory, singularity theory and Boij-S\"{o}derberg theory. For example, if $X$ admits an Ulrich bundle, then its cone of cohomology tables coincides with that of $\PP^n$ \cite{EisenbudSchreyer}. In view of their importance, Eisenbud, Schreyer and Weyman \cite{ESW} formulated the problem of determining which varieties admit Ulrich bundles. Many authors have considered this problem. We refer the reader to \cite{CasanellasHartshorne}, \cite{CKM}, \cite{CostaMiroRoig1}, \cite{CMP}, \cite{ESW}, \cite{Faenzi}, \cite{Miro}, \cite{MP1}, \cite{MP2} for references and further results. For example, Ulrich bundles exist on curves, linear determinantal varieties, hypersurfaces and more generally on complete intersections (see \cite{BaHU}, \cite{BHU}, \cite{ESW}, \cite{KP}). The second and fourth authors have initiated the study of homogeneous Ulrich bundles on homogeneous varieties. In \cite{CostaMiroRoig1}, they classified all Schur bundles which are Ulrich on Grassmannians under the Pl\"{u}cker embedding. In this paper, we extend this study to flag varieties. Our first main theorem is the following. \begin{theorem*}[Theorem \ref{thm-higherStep}] If $r\geq 3$, then no flag variety $F(k_1, \dots, k_r; n)$ admits a Schur bundle $E_\lambda$ which is Ulrich with respect to $\OO(1)$. \end{theorem*} In view of Theorem \ref{thm-higherStep}, we concentrate on two-step flag varieties. We refer the reader to \S \ref{sec-overview2step} for a detailed description of our results for two-step flag varieties. We may summarize our results as follows. \begin{theorem*}[Theorems \ref{thm-1n1}, \ref{thm-beta1intro}, \ref{thm-beta2intro}, \ref{thm-1n2intro}, \ref{thm-2n2intro}] We classify all the Schur bundles $E_\lambda$ which are Ulrich with respect to $\cO(1)$ on $$ F(1,n-1;n), \quad F(1, n-2; n), \quad F(2, n-2; n), \quad F(k, k+1; n), \quad \mbox{and} \quad F(k, k+2; n).$$ \end{theorem*} The Borel-Weil-Bott Theorem computes the cohomology of Schur bundles on flag varieties. Let $N$ denote the dimension of $F(k_1, \dots, k_r; n)$. In order to determine whether a bundle $E_\lambda$ is Ulrich, we need to check that the cohomology of $N$ consecutive twists of $E_\lambda$ vanishes. Using the Borel-Weil-Bott Theorem, this reduces to an intricate combinatorial problem. In \S \ref{sec-prelim}, we explain the algebraic geometry and representation theory background needed to turn the problem into a combinatorial problem. Then for the rest of the paper we study the combinatorial problem, which may be stated as follows. \begin{problem} Let $$P= (a_1^1, \dots, a_{l_1}^1 | a_1^2, \dots, a^2_{l_2}| \cdots | a_1^{r+1}, \dots, a_{l_{r+1}}^{r+1})$$ be a strictly decreasing sequence of integers divided into $r+1$ subsequences. Let $P(t)$ denote the sequence $$(a_1^1 - tr , \dots, a_{l_1}^1 -tr | a^2_1-t(r-1) , \dots, a^2_{l_2} - t (r-1) | \cdots | a_1^{r+1} , \dots, a^{r+1}_{l_{r+1}})$$ obtained by subtracting $t (r-i+1)$ from $a^i_j$. Classify the sequences $P$ for which the sequences $P(t)$ have exactly two equal entries for $1 \leq t \leq \sum_{1\leq i<j \leq r+1} l_i l_j$. \end{problem} Theorem \ref{thm-higherStep} is equivalent to the statement that such sequences do not exist if $r \geq 3$. The combinatorial problem is most interesting when $r=2$. In this case, there are infinite families of examples, which makes their classification challenging. We conjecture that there do not exist such sequences with both $l_1$ and $l_2$ at least three. In terms of geometry, this conjecture translates to the following. \begin{conjecture}\label{conj-main} The two-step flag variety $F(k_1, k_2; n)$ does not admit a Schur bundle which is Ulrich with respect to $\OO(1)$ if $k_1 \geq 3$ and $k_2-k_1 \geq 3$. \end{conjecture} If Conjecture \ref{conj-main} is true, then our results give a complete classification of all Schur bundles which are Ulrich with respect to $\OO(1)$ on partial flag varieties. In particular, such bundles are very rare. Hence, in order to construct Ulrich bundles on partial flag varieties, one has to look elsewhere. One may also consider Schur bundles which are Ulrich with respect to other embeddings of $F(k_1, \dots, k_r; n)$. Our constructions all scale. In particular, if $E_\lambda$ is Ulrich with respect to $\OO(1)$, then multiplying the integers in $\lambda$ by $m$ gives a bundle which is Ulrich with respect to $\OO(m)$. Our combinatorial techniques can in principle be applied to other polarizations, though we do not discuss any such results here. \subsection*{The organization of the paper} In \S \ref{sec-prelim}, we explain the necessary background from algebraic geometry and representation theory needed to turn the classification of Schur bundles which are Ulrich into a combinatorial problem. In \S \ref{sec-comb}, we explain the combinatorial problem. In \S \ref{sec-higherStep}, we solve the combinatorial problem for flag varieties of at least three steps and show that they do not admit a Schur bundle which is Ulrich for $\cO(1)$. In \S \ref{sec-overview2step}, we explain our results for two-step flag varieties and classify Ulrich bundles on flag varieties of the form $F(1,n-1;n)$. In sections \S \ref{sec-sumset}, \S \ref{sec-beta2}, \S \ref{sec-1n2} and \S \ref{sec-2n2}, we classify Ulrich bundles on flag varieties of the form $F(k, k+1; n)$, $F(k, k+2; n)$, $F(2, n-1; n)$ and $F(2, n-2; n)$, respectively. \section{Algebraic geometry background}\label{sec-prelim} \subsection*{Partial flag varieties} Let $0< k_1 < \cdots < k_r <n$ be an increasing sequence of integers. Set $k_0 = 0$ and $k_{r+1} = n$. The $r$-step partial flag variety $F(k_1, \dots, k_r; n)$ parameterizes partial flags $W_1 \subset W_2 \subset \cdots \subset W_r \subset V,$ where $\dim W_i = k_i$. Given any set of indices $1 \leq i_1 < i_2 < \cdots < i_s \leq r$, there is a natural projection $$\pi_{i_1, \dots, i_s} : F(k_1, \dots, k_r; n) \rightarrow F(k_{i_1}, \dots, k_{i_s}; n).$$ In particular, the partial flag variety can be realized as an iterated Grassmannian bundle $$F(k_1, \dots, k_r; n) \rightarrow F(k_2, \dots, k_r; n) \rightarrow \cdots \rightarrow G(k_r; n).$$ From this description, it is immediate to read that $$N:=\dim(F(k_1, \dots, k_r; n))= \sum_{i=1}^r k_i (k_{i+1} - k_i).$$ The partial flag variety has $r$ projections to Grassmannians $$\pi_i : F(k_1, \dots, k_r; n) \rightarrow G(k_i; n).$$ The Picard group of $F(k_1, \dots, k_r; n)$ is generated by $L_i= \pi_i^* \OO_{G(k_i; n)}(1)$. A line bundle $L$ is ample on $F(k_1, \dots, k_r; n)$ if and only if $L = L_1^{\otimes a_1} \otimes \cdots \otimes L_r^{\te a_r}$ with $a_i > 0$ for every $1 \leq i \leq r$. \subsection*{The degree of partial flag varieties} Let $X \subset \PP^n$ be an equivariant embedding of a projective homogeneous variety. The Weyl dimension formula implies that the Hilbert polynomial of $X$ factors as a product of linear polynomials. Consequently, one obtains a formula for the degree of $X$ \cite{GrossWallach}. Let $\psi$ denote the dominant weight defining the embedding of $X$ in $\PP^n$. Let $\rho$ be half the sum of the positive roots and let $N$ be the dimension of $X$. Then the degree of $X$ is given by \begin{equation}\label{eq-degree} N! \prod_{\alpha} \frac{\langle \psi, \check{\alpha} \rangle }{\langle \rho, \check{\alpha} \rangle }, \end{equation} where the product is over all positive roots $\alpha$ with $ \langle \psi, \check{\alpha} \rangle >0$ and $\check{\alpha}$ denotes the coroot \cite{GrossWallach}. Now we specialize to the case of partial flag varieties. Let $a_1, \dots, a_r$ be a sequence of positive integers. Set $$b_{ij} = \sum_{k=i}^j a_k.$$ We would like to calculate the degree of $F(k_1, \dots, k_r; n)$ under the embedding defined by $$L_{a_1, \dots, a_r} = L_1^{\otimes a_1} \otimes \cdots \otimes L_r^{\otimes a_r}.$$ The ample line bundle $L_{a_1, \dots, a_r}$ corresponds to the dominant weight $$ \psi_{a_1, \dots, a_r} = b_{1r} e_1 + \cdots + b_{1r}e_{k_1} + b_{2r} e_{k_1 + 1} + \cdots + b_{2r} e_{k_2} + \cdots + b_{rr} e_{k_r}.$$ We have that $$\rho = (n-1) e_1 + (n-2) e_2 + \cdots + e_{n-1}.$$ The coroots that have nonzero intersection with $\psi_{a_1, \dots, a_r}$ are precisely $e_i - e_j$, where $k_{s-1} < i \leq k_s$ for some $1 \leq s \leq r$ and $k_t < j \leq k_{t+1}$ for $s<t \leq r$. For such $e_i - e_j$, the numerator in Equation (\ref{eq-degree}) is $b_{st}$ and the denominator is $j-i$. We thus conclude the following proposition. \begin{proposition} The degree of the partial flag variety $F(k_1, \dots, k_r; n)$ embedded by the line bundle $L_{a_1, \dots, a_r}$ is \begin{equation}\label{eq:partialflagdegree} N! \frac{\prod_{1 \leq i \leq j \leq r} b_{ij}^{(k_i - k_{i-1})(k_{j+1} - k_j)}}{\prod_{s=1}^r \prod_{k_{s-1} < i \leq k_s} \frac{(n-i)!}{(k_s -i)!}}. \end{equation} \end{proposition} \begin{remark} We record a few special cases of the formula. The degree of $G(k,n)$ under the $a$th power of the Pl\"{u}cker embedding is $$(k(n-k))! a^{k(n-k)} \prod_{1 \leq i \leq k} \frac{(k-i)!}{(n-i)!}.$$ The degree of $F(k_1, \dots, k_r; n)$ under the line bundle $L_{1,\dots,1}$ is $$N ! \frac{\prod_{1 \leq i \leq j \leq r} (j-i+1)^{(k_i - k_{i-1})(k_{j+1} - k_j)}}{\prod_{s=1}^r \prod_{k_{s-1} < i \leq k_s} \frac{(n-i)!}{(k_s -i)!}}.$$ In particular, specializing to the case $F(1, n-1; n)$, we obtain $$(2n-3)! \frac{2}{(n-1)!(n-2)!}.$$ Specializing to the case $F(1, n-2, n)$, we obtain $$(3n-7)! \frac{4}{(n-1)!(n-2)!(n-3)!}.$$ \end{remark} \subsection*{Borel-Weil-Bott Theorem and dimension of cohomology groups} The partial flag variety $F(k_1, \dots, k_r; n)$ is endowed with a collection of tautological subbundles $$0=T_0 \subset T_1 \subset T_2 \subset \cdots \subset T_{r+1} = \underline{V}, $$ where $T_i$ is a vector bundle of rank $k_i$ and $\underline{V}$ is the trivial bundle of rank $n$. Let $U_i = T_i/ T_{i-1}$. Let $E_\lambda$ be a Schur bundle on $F(k_1, \dots, k_r; n)$. The partition $\lambda$ is a concatenation $\lambda = (\lambda_1|\cdots|\lambda_{r+1})$, where $\lambda_i$ has length $k_i - k_{i-1}$, and $$E_\lambda = \mathbb{S}^{\lambda_1} U_1^* \otimes \mathbb{S}^{\lambda_2} U_2^* \otimes \cdots \otimes \mathbb{S}^{\lambda_{r+1}} U_{r+1}^* ,$$ where $\mathbb{S}^{\lambda_i}$ denotes the Schur functor of type $\lambda_i$. Such bundles can be characterized as the pushforwards of line bundles on complete flag varieties. Since $U_1, \dots, U_{r+1}$ give a filtration of the trivial bundle $\underline{V}$, the tensor product of their determinants is trivial by the Whitney formula. Hence, adding a constant integer to all the entries of $\lambda$ corresponds to tensoring the bundle $E_{\lambda}$ with a power of the trivial line bundle and does not change its isomorphism class. We will say two partitions $\lambda,\lambda'$ are equivalent if they differ by adding a constant to all the entries; equivalent partitions give isomorphic bundles. There is also no harm in allowing negative integers in $\lambda$, since adding a large constant will make all entries positive. For studying Ulrich bundles, it is important to know the effect of tensoring $E_\lambda$ by an ample line bundle $L_{a_1,\ldots,a_r}$. By the Littlewood-Richardson rule, adding a constant sequence to $\lambda_i$ corresponds to tensoring the bundle by the determinant of $U_i$. Since the determinant of $U_i$ is $L_i \otimes L_{i-1}^{-1}$, we have $$E_\lambda \te L_{a_1,\ldots,a_r} \cong E_{\lambda + \psi_{a_1,\ldots,a_r}},$$ where $ \psi_{a_1,\ldots,a_r}$ is the dominant weight corresponding to $L_{a_1,\ldots,a_r}$. Let $\rho = (n-1, n-2, \dots, 1, 0)$. We say that $\lambda$ is {\em singular} if $\lambda + \rho$ has repeated entries. Otherwise, we say that $\lambda$ is {\em regular}. Let $s$ be the permutation in $\mathfrak{S}_n$ that lists the entries in $\lambda + \rho$ in decreasing order. Let $q(\lambda)$ denote the minimal length of the permutation $s$. The famous Borel-Weil-Bott Theorem computes the cohomology of line bundles on the complete flag variety, which allows us to calculate the cohomology of the Schur bundle $E_{\lambda}$ \cite{Weyman}. \begin{theorem}[Borel-Weil-Bott]\label{thm-BWB} Let $E_{\lambda}$ be a Schur bundle on $F(k_1, \dots, k_r; n)$. \begin{enumerate} \item If $\lambda$ is singular, then $H^i(E_{\lambda})=0$ for every $i$. \item If $\lambda$ is regular, then $H^i(E_{\lambda}) = 0$ for $i \not= q(\lambda)$ and $H^{q(\lambda)}(E_{\lambda}) = \mathbb{S}^{s(\lambda+\rho) - \rho} V^*$. \end{enumerate} \end{theorem} We recall how to compute the dimension of a representation $\mathbb{S}^{\mu} V$ of $GL(n)$. Let $Y(\mu)$ be the Young diagram corresponding to the partition $\mu$. The hook-length $\hook(i,j)$ is the number of boxes in the same row to the right of the box $(i,j)$ and in the same column below the box $(i,j)$ (including the box itself). The dimension of the representation $\mathbb{S}^{\mu} V$ is given by $$\prod_{(i,j) \in Y(\mu)} \frac{n+j -i}{\hook(i,j)},$$ which is always nonzero. In particular, the rank of the vector bundle $E_{\lambda}$ is given by $$\prod_{s=1}^{r+1} \prod_{(i,j) \in Y(\lambda_s)} \frac{k_s-k_{s-1} +j -i}{\hook (i,j)}. $$ \section{Ulrich bundles and the combinatorial problem}\label{sec-comb} In this section we recall basic facts on Ulrich bundles and phrase the condition that a Schur bundle $E_\lambda$ on the flag variety $F(k_1,\ldots,k_r;n)$ is Ulrich combinatorially. \subsection{Ulrich bundles in general}Let $(X, \OO_X(1))$ be a polarized projective variety. A vector bundle $E$ on $X$ is {\em arithmetically Cohen Macaulay} if it is locally Cohen-Macaulay and $H_*^i (X, E) =0$ for every $1 \leq i \leq \dim(X) -1$. A vector bundle on $X$ is {\em initialized} if $H^0(X, E) \not= 0$, but $H^0(X, E(-1)) = 0$. If $E$ is ACM, then there exists a twist of $E$ which is initialized. \begin{definition} A vector bundle $E$ is an {\em Ulrich bundle} if $E$ is ACM and the initialized twist $E_{\init}$ of $E$ satisfies $h^0(X, E_{\init}) = \rk(E) \deg(X)$. \end{definition} \begin{remark}\label{rem-UlrichCrit}For our purposes, there is another useful criterion for determining when a bundle is Ulrich. Let $X$ be a smooth, $N$-dimensional arithmetically Cohen-Macaulay projective variety and suppose $E$ is an initialized bundle on $X$. Then, by \cite[Proposition 2.1, Corollary 2.2]{ESW}, $E$ is Ulrich if and only if $H^i(E(-t))=0$ for all $i$ and $t\in [N] = \{1,\ldots,N\}$ (see \cite[Remark 2.6]{CostaMiroRoig1}). It is well-known that flag varieties are arithmetically Cohen-Macaulay when embedded by any ample complete linear system (see \cite[Theorem 5]{Ramanathan}). \end{remark} \begin{remark} Since for an initialized Ulrich bundle $h^0(E) = \rk(E) \deg(X)$, whenever there is a Schur bundle $E_\lambda$ which is Ulrich we obtain interesting combinatorial identities relating two hook-length formulas to the degree of the flag variety. In the other direction, one could potentially find Ulrich bundles by searching for such identities; however, this seems difficult. \end{remark} \subsection{The combinatorial problem} Let $E_\lambda$ be a Schur bundle on $F(k_1,\ldots,k_r;n)$ corresponding to the concatenated sequence $\lambda = (\lambda_1|\cdots|\lambda_{r+1})$ of partitions. The Borel-Weil-Bott Theorem \ref{thm-BWB} allows us to rephrase the condition that $E_\lambda$ is an initialized Ulrich bundle with respect to $\OO(1)$ purely in terms of the combinatorics of $\lambda$. Recall $\lambda_i$ is a partition of length $k_i-k_{i-1}$, so $\lambda$ has total length $n$. First, we write $P = \lambda +\rho$, where $\rho = (n-1,n-2,\ldots,1,0)$. For concreteness, we explicitly write out the partition $P$ as $$P = (a_1^1, \dots, a_{l_1}^1 | a_1^2, \dots, a^2_{l_2}| \cdots | a_1^{r+1}, \dots, a_{l_{r+1}}^{r+1});$$ here $l_i = k_i-k_{i-1}$. Observe that $H^0(E_\lambda)\neq 0$ if and only if the sequence of entries of $P$ is strictly decreasing. For any integer $t\geq 0$, we define a partition $$P(t) = (a_1^1 - tr,\ldots, a_{l_1}^1 - tr | a_1^2 - t(r-1),\ldots , a_{l_2}^2-t(r-1)|\cdots|a_1^{r+1},\ldots,a_{l_{r+1}}^{r+1})$$ by subtracting $t(r+1-i)$ from every entry in the $i$th part, and analogously define partitions $\lambda(t)$. It is also convenient to view the entries of $P(t)$ as being functions of $t$ and define $$a_k^i(t) = a_k^i-t(r+1-i),$$ so that $$P(t) = (a_1^1(t),\ldots,a_{l_1}^1(t)|\cdots|a_1^{r+1}(t),\ldots,a_{l_{r+1}}^{r+1}(t)).$$ Then $E_{\lambda}(-t) = E_{\lambda(t)}$ has $H^i(E_{\lambda}(-t))=0$ for all $i$ precisely when there is a repeated entry in $P(t)$. Therefore $E_\lambda$ is an initialized Ulrich bundle if and only if $P(t)$ has a repeated entry for all integers $t\in [N]$, where $N = \dim F(k_1,\ldots,k_r;n)$ is the dimension of the flag variety. We now make the central combinatorial definitions of the paper. \begin{definition} Let the partition $$P = (a_1^1,\ldots,a_{l_1}^1|a_1^2,\ldots,a_{l_2}^2|\cdots|a_1^{r+1},\ldots, a_{l_{r+1}}^{r+1})$$ consist of a decreasing sequence of integers arranged into $r+1$ blocks, with $l_i>0$ for all $i$. The \emph{type} of the partition is $(l_1,\ldots,l_{r+1})$. The \emph{dimension} of the partition is $N = \sum_{k<h} l_kl_h$. The partition $P$ is \emph{Ulrich} if for all $t\in [N]=\{1,\ldots,N\}$ the partition $P(t)$ has a repeated entry. Two partitions are \emph{equivalent} if they differ by adding a constant to all the entries. When we count Ulrich partitions or talk about uniqueness, it is understood that we always do so modulo equivalence. \end{definition} We make the correspondence between the algebro-geometric and combinatorial problems explicit. \begin{proposition} There is a bijective correspondence $\lambda \mapsto \lambda + \rho = P$ between equivalence classes of partitions $\lambda$ such that the Schur bundle $E_\lambda$ on $F(k_1,\ldots,k_r;n)$ is Ulrich with respect to $\OO(1)$ and equivalence classes of Ulrich partitions $P$ of type $(k_1,k_2-k_1,\ldots,k_r-k_{r-1},n-k_r)$. \end{proposition} The main question which we study from this point forward is to determine the possible types of Ulrich partitions, and, when there exists an Ulrich partition of a given type, to classify all such partitions up to equivalence. \begin{remark} The dimension $N$ of a partition $P$ is precisely the number of pairs of entries coming from different blocks. For there to be a repeated entry in $P(t)$ at every time $t\in [N]$, it is therefore necessary and sufficient that every pair of entries meets at some time $t\in [N]$ and that no two pairs of entries meet at the same time. \end{remark} \begin{remark} There is no harm in considering lists $P$ where some of the entries are negative since the notion of Ulrich partition is invariant under equivalence. It will at times also be convenient to change the definition of $P(t)$ slightly by subtracting a different consecutive range of integers from each part. For example, in the two-step flag variety case where $P$ has $3$ parts it will be convenient to subtract $1,0,-1$ from the three parts, so that the middle block is constant with time. We will make our choice of normalization clear in each section as needed. \end{remark} We now provide a simple example illustrating the combinatorial definitions and the conversion to the algebro-geometric language. \begin{example}\label{ex-combToAG} Consider the partition $P = (5|3,-1,-2,-4|{-5})= (a|b_1,b_2,b_3,b_4|c)$ of type $(1,4,1)$. For convenience, we normalize $P(t)$ by subtracting $1,0,-1$ at each time. The dimension is $N=9$, and this partition is Ulrich. To visualize the Ulrich condition, we draw a \emph{time evolution diagram} showing the positions of the entries of $P(t)$ in Figure \ref{fig-141}. Along the first row we place the entries $a,b,b,b,b,c$ in the corresponding positions; we provide numerical positions above some of the entries for reference. For each time $t\in [N]$ we similarly place the entries of $P(t)$ in the corresponding row. When two entries intersect, we put a box around them to highlight the intersection. The Ulrich condition means that for every time $t\in [N]$ there is exactly one box in the corresponding row. We also display $P(N+1)$ at time $N+1$, which is relevant to the computation of the \emph{dual partition} $P^*$; see \S \ref{ssec-symmetryDuality}. \begin{figure}[htbp] \input{1n1Ex.pstex_t} \caption{Time evolution diagram of the Ulrich partition $(5|3,-1,-2,-4|{-5})$ of type $(1,4,1)$. See Example \ref{ex-combToAG}.}\label{fig-141} \end{figure} Replacing $P$ by the equivalent partition $(11|9,5,4,2|1)$ and putting $\lambda = P - \rho$, we have $\lambda = (6|5,2,2,1|1) = (\lambda_1|\lambda_2|\lambda_3)$. The corresponding initialized Ulrich bundle on $F(1,5;6)$ is $$E_\lambda = \mathbb{S}^{\lambda_1} U_1^* \otimes \mathbb{S}^{\lambda_2} U_2^* \otimes \mathbb{S}^{\lambda_{3}} U_3^*.$$ A straightforward computation with the hook-length formula shows that $\rk(E_\lambda) = 70$ and $h^0(E_\lambda) = 17640$. This is consistent with the calculation $\deg F(1,5;6) = 252$ for the embedding of $F(1,5;6)$ by $\OO(1)$. \end{example} \subsection{Symmetry and duality}\label{ssec-symmetryDuality} There are certain symmetries satisfied by Ulrich partitions which we will exploit throughout the paper. Suppose $P = (a_1^1,\ldots,a_{l_1}^1|\cdots | a_1^{r+1},\ldots,a_{l_{r+1}}^{r+1})$ is a partition of type $(l_1,\ldots,l_{r+1})$. Multiplying all the entries of $P$ by $-1$ and reversing the order in which they are written, we obtain the \emph{symmetric} partition $$P^s=(-a_{l_{r+1}}^{r+1},\ldots,-a_{1}^{r+1}|\cdots| {-a_{l_1}^1},\ldots,-a_1^1)$$ of the reversed type $(l_{r+1},\ldots,l_1)$. It is clear that if $P$ is Ulrich then $P^s$ is also Ulrich. The time evolution diagram of $P^s$ is flipped about a vertical axis. It is often convenient to add a constant to all the entries of the symmetric partition to keep some entries fixed; there is no harm in this since we only care about Ulrich partitions up to equivalence. There is another slightly less trivial symmetry which we call \emph{duality}. Starting from an Ulrich partition $P$ we define the partition $$P^* = (a_1^{r+1}(N+1),\ldots,a_{l_{r+1}}^{r+1}(N+1)|\cdots | a_1^1(N+1),\ldots , a_{l_1}^1(N+1)).$$ The integers which appear in $P^*$ are precisely the integers in $P(N+1)$; we have just reordered them to be decreasing. In the time evolution diagram for $P$, $P^*$ can be read off from the final row below the horizontal line. Then $P^*$ is Ulrich of type $(l_{r+1},\ldots, l_1)$; its time evolution diagram is obtained from the diagram for $P$ by flipping about a horizontal axis $t = (N+1)/2$. Both the symmetries we have described reverse the type of the Ulrich partition, so change the type unless it is symmetric. The partition $(P^s)^*$ has the same type as $P$, and we call this partition the \emph{symmetric dual}. In particular, if there is a unique Ulrich partition of a given type then it is equal to its symmetric dual. \begin{example} The symmetric partition obtained from $(5|3,-1,-2,-4|{-5})$ is $(5|4,2,1,-3|{-5})$. Both of these partitions are self-dual. \end{example} \begin{remark} Suppose the flag variety $F(k_1,\ldots,k_r;n)$ carries an Ulrich Schur bundle $E_{\lambda}$, and let $P = \lambda +\rho$. Then $E_{P^s-\rho}$ and $E_{P^* -\rho}$ are Ulrich bundles on the dual flag variety $F(n-k_r,\ldots,n-k_1;n)$, while $E_{(P^{s})^*-\rho}$ is an Ulrich bundle on the original flag variety. \end{remark} \section{Ulrich bundles on flag varieties with three or more steps}\label{sec-higherStep} In this section, we prove the following theorem. \begin{theorem}\label{thm-higherStep} A flag variety of three or more steps does not have a Schur bundle $E_\lambda$ which is Ulrich with respect to the line bundle $\cO(1)$. \end{theorem} As a special case, we obtain the following corollary. \begin{corollary} The complete flag variety $F(1, 2, 3, \dots, n-1; n)$ does not admit a Schur bundle which is Ulrich for $n \geq 4$. \end{corollary} We will split the proof into three parts. We first analyze flag varieties with at least five steps. The three and four step flag varieties require separate (and more intricate) arguments. We begin by deriving simple congruences among the entries of an Ulrich partition. \begin{lemma}\label{lem-congruence} Let $(a^1_i|a^2_i|\cdots|a^{s}_i)$ be an Ulrich partition. Suppose $1 \leq j \leq k \leq s$. Then all the entries $a^j_i$ in the $j$-block are congruent to all the entries $a^k_i$ in the $k$-block modulo $k-j$. \end{lemma} \begin{proof} We can normalize the evolution of the Ulrich partition so that we keep all the entries $a^j_i$ in the $j$-block fixed and add $k-j$ to all the entries $a^k_i$ in the $k$-block. This preserves the congruence classes of the $a^j_i$ and $a^k_i$ modulo $k-j$. In order for them to collide at some time, they must all have the same congruence class. \end{proof} \subsection{Flag varieties with at least five steps} \begin{proposition}\label{prop-fiveStep} There are no Ulrich partitions with $6$ or more blocks. \end{proposition} \begin{proof} Suppose there is an Ulrich partition $(a^1_1,\ldots,a^1_{l_1}|\cdots|a^s_1,\ldots,a^s_{l_s})$ with $s \geq 6$. Since $s \geq 6$, we have that $\max\{k-1,s-k\} \geq 3$ for every $1 \leq k \leq s$. By Lemma \ref{lem-congruence}, we conclude that the entries $a^k_i$ in the $k$-block all have the same congruence class modulo 6. In particular, they all have the same parity. At time $t=1$, two adjacent blocks must meet; say $a_{l_k}^k(1) = a_{1}^{k+1}(1)$. This forces $a_{l_k}^k$ and $a_1^{k+1}$ to have opposite parities. If $k-j$ is even, then the entries $a^k_i(t)$ and $a^j_h(t)$ in the $k$- and $j$-blocks have the same parities at all times $t$. On the other hand, if $k-j$ is odd, then the entries in the $k$- and $j$-blocks have the same parity at odd times and opposite parities at even times $t$. Consequently, at even times the intersections must occur between entries in nonadjacent blocks. Suppose that at time $t=2$, $a^j_{l_j}$ and $a^{j+2}_1$ collide. Normalize the evolution of the sequence so that $1,0,-1$ is subtracted from the $j$-, $(j+1)$-, and $(j+2)$-blocks, respectively. Then, up to switching their order, at times $t=1,3$, the collisions must be $a^j_{l_j}a^{j+1}_1$ and $a^{j+1}_1a^{j+2}_1$ depending on the sequence at time $t=0$. We observe that the possible positions of the entries at time $t=0$ are $$ \cdots \times \times \ a^j_{l_j} \ a^{j+1}_1 \ \times \times \ a^{j+2}_1 \ \cdots \quad \mbox{or} \quad \cdots \times \times \ a^j_{l_j} \ \times \times \ a^{j+1}_1 \ a^{j+2}_1 \cdots,$$ where we denote the positions unoccupied by an entry by $\times$. In particular, after normalizing the positions, the triple $(a_{l_j}^j,a_1^{j+1},a_1^{j+2})$ either equals $(1,0,-3)$ or $(3,0,-1)$. Furthermore, there is a unique entry in the $(j+1)$-part: $l_{j+1}=1$. At time $t=4$, the collision again must be between entries from nonadjacent blocks. For $k \not= j$, if $a^k_{l_k}(4) = a^{k+2}_1(4)$, then $a^k_{l_k}$ passes through the entries in the $(k+1)$-block without intersecting them, contradicting that the partition is Ulrich. Hence, the only possible collisions at time $t=4$ are either $a^j_{l_j} a^{j+2}_2 $ or $a^j_{l_j-1} a^{j+2}_1$. However, these are both impossible since they would give a separation of only four between either $a^j_{l_j}$ and $a^{j}_{l_j-1}$ or $a^{j+2}_1$ and $a^{j+2}_2$. This contradicts the fact that entries in each block are congruent modulo 6. We conclude that there cannot be any Ulrich partitions with $s \geq 6$. \end{proof} \subsection{Three-step flag varieties} In this subsection, we analyze Ulrich bundles on three-step flag varieties. \begin{proposition}\label{prop-threeStep} There are no Ulrich partitions with $4$ blocks. \end{proposition} \begin{proof} Let $(a_1,\dots, a_j|b_1, \dots, b_k|c_1, \dots, c_l|d_1, \dots, d_m)$ be an Ulrich partition for a three-step flag variety. By Lemma \ref{lem-congruence}, the parities of $a_i(t)$ and $c_h(t)$ (respectively, $b_i(t)$ and $d_h(t)$) are equal at all times. On the other hand, by the same argument as in the proof of Proposition \ref{prop-fiveStep}, entries in adjacent blocks have the same parity at odd times $t$ and opposite parities at even times $t$. Consequently, when $t$ is even, the only collisions can be $ac$ or $bd$ collisions. Since the last intersection at time $t = N$ has to be between entries from adjacent blocks, we conclude that the dimension $N$ is odd. By symmetry (see \S \ref{ssec-symmetryDuality}), we can assume that at time $t=2$, the collision is between $a_j$ and $c_1$. We normalize the evolution of the partition so that we subtract $2,1,0,-1$, respectively, from the $A$-, $B$-, $C$-, and $D$-blocks. Consequently, the first three collisions are $$a_jb_1,\ a_j c_1, \ b_1 c_1 \quad \mbox{or} \quad b_1 c_1,\ a_j c_1, \ a_j b_1.$$ The corresponding sequences at time $t=0$ look like $$\cdots a_j \ b_1 \times \times \ c_1 \ \times \times \cdots \quad \mbox{respectively} \quad \cdots a_j \ \times \times \ b_1 \ c_1 \ \times \times \cdots $$ In either case, we conclude that the length $k$ of the $B$-block is one since by time $t=3$, $a_j$ must have passed through the entire $B$-block. From now on we denote the unique entry in the $B$-block by $b$. Similarly, $t=N-1$ is an even time. Hence, the only possible collisions are $ac$ or $bd$ collisions. If the collision is between $b$ and $d_m$, then the last three collisions at times $N-2, N-1$ and $N$ are $$ c_1 d_m,\ b d_m, \ b c_1 \quad \mbox{or} \quad b c_1, \ b d_m, \ d_m c_1.$$ In particular, we conclude that the length of the $C$-block is also one. Denote the lengths of the $A$- and $D$-blocks by $j$ and $m$, respectively. The dimension of the flag variety is $$N=2j + jm + 2m + 1.$$ Since the collisions at even times are $ac$ or $bd$ collisions, the dimension can be at most one more than twice the total number $j+m$ of possible $ac$ and $bd$ collisions. We, therefore, have the inequality $$2j + jm + 2m \leq 2(j+m).$$ This implies that $jm \leq 0,$ which is absurd. We conclude that at time $t= N-1$, the intersection must be between $a_1$ and $c_l$. Hence, the last three intersections at times $N-2,N-1,N$ are one of $$a_1 b,\ a_1 c_l, \ b c_l \quad \mbox{or} \quad b c_l,\ a_1 c_l, \ a_1 b.$$ Normalize the position of the sequence so that $b=y$ and $d_1=-y$. Then $c_1$ equals either $y-1$ or $y -3$. Furthermore, there cannot be any $c$'s in positions $-y+1$, $-y+2$ or $-y+3$ since $d_1$ is at these positions at times $1,2,3$ when there are other collisions. In either case, $d_1$ intersects $c_1$ at a time $t_0 \geq 2y -3$. On the other hand, $b$ intersects $c_l$ at some time $t_1< 2y-3$. Hence, the intersection $bc_l$ happens before the intersection $c_1d_1$. This is a contradiction, since $bc_l$ is one of the last three intersections and $c_1d_1$ is not. We conclude that there are no Ulrich partitions for three step flag varieties with respect to $\cO(1)$. \end{proof} \subsection{Four-step flag varieties} Finally, we analyze Ulrich bundles on four-step flag varieties. This is the most intricate case. \begin{proposition} \label{prop-fourStep} There are no Ulrich partitions with $5$ blocks. \end{proposition} \begin{proof} Let $(a_1, \dots, a_j|b_1, \dots, b_k|c_1, \dots, c_l|d_1, \dots, d_m|e_1, \dots, e_p)$ be an Ulrich partition for a four-step flag variety. We can normalize the evolution so that at each time we subtract $2,1,0,-1,-2$ from the entries in the $A$-, $B$, $C$, $D$, $E$-blocks, respectively. By parity, at time $t=2$ we must have an $ac$, $bd$ or $ce$ collision. Suppose we have a $bd$ collision. As before, the length $l$ of the $C$-block is $1$, so we denote its entry by $c$. Then the first three collisions are one of $$b_k c, \ \ b_k d_1, \ \ c d_1 \quad \mbox{or} \quad c d_1, \ \ b_k d_1, \ \ b_k c$$ corresponding to the following sequences at time $t=0$ $$ \cdots b_k \ c \ \times \ \times \ d_1 \cdots \quad \mbox{or} \quad \cdots b_k \ \times\ \times \ c \ d_1 \cdots.$$ By Lemma \ref{lem-congruence}, the entries in the $B$- (respectively, $D$-) block are the same modulo 6, so the distances between consecutive entries are at least 6. Similarly, the consecutive entries in the $A$- (respectively, $E$-) blocks are at least 12 apart. At time $t=4$, the collision cannot be a $bd$ collision since then the distance between two consecutive entries in either the $B$- or $D$- blocks would be only 4 instead of 6. Hence, the collision at time $t=4$ must be either $a_jc$ or $ce_1$. However, neither of these are possible. In the first case, $a_j$ never collides with $d_1$ and in the latter case $e_1$ never collides with $b_k$. We conclude that the first collision cannot be a $bd$ collision. The collision at time $t=2$ must be an $ac$ or $ce$ collision. Without loss of generality, we may assume it is $ac$. Therefore, there is only one $b$ and $b$ collides with $c_1$ at time $1$ or $3$. As in the case of three-step flag varieties, the dimension $N$ of the flag variety must be odd. Hence, at time $N-1$ we must have an $ac$, $bd$ or $ce$ collision. By symmetry, we see that the intersection at time $t=N-1$ cannot be a $bd$ collision, hence must be an $ac$ or $ce$ collision. If it were an $ac$ collision, then $b$ would have to intersect $c_l$ at time $N$ or $N-2$. Since the time it takes for $d_1$ to intersect $c_1$ is longer than the time it takes $b$ to intersect $c_l$ (as in the proof of Proposition \ref{prop-threeStep}), we conclude that the intersection at time $t = N-1$ must be $c_1e_p$ and there can be only one entry in the $D$-block. We have so far concluded that the sequence looks like $(a_1, \dots, a_j|b|c_1, \dots, c_l|d|e_1, \dots, e_p)$. The first three intersections are one of $$\mbox{Case I:} \ a_j b ,\ \ a_j c_1, \ \ b c_1 \quad \mbox{or} \quad \mbox{Case II:} \ b c_1, \ \ a_j c_1, \ \ a_j b$$ and the last three intersections are one of $$\mbox{Case A:} \ d e_p, \ \ c_1 e_p, \ \ c_1 d \quad \mbox{or} \quad \mbox{Case B:} \ c_1 d, \ \ c_1 e_p, \ \ d e_p.$$ For the rest of the argument, we normalize positions so that $b=y$ and $d=-y$. Then at time $y$ we have the $bd$ collision at position $0$. We first show that case IB is impossible. We compute $c_1 = y-3$, and the equality $$c_1 =c_1(N-2) = d(N-2)=d+N-2$$ gives $N = 2y-1$. Similarly, from $e_p(N) = d(N)$ we find $e_p = -3y+1$. As $a_j = y+1$, we conclude the intersection $a_je_p$ happens at time $y$. This contradicts that $b$ and $d$ already collide at time $y$. Symmetrically, case IIA is also impossible since the intersection $a_1e_1$ happens at time $y$. We conclude that we must be in Case IA or Case IIB. These cases are the same under duality (see \S\ref{ssec-symmetryDuality}). Consequently, it suffices to analyze Case IA. In Case IA, we first prove that at time $t=0$ the sequence must look like \begin{equation}\tag{$\ast$} \cdots a_j \ b \ \times \times \ c_1 \ \times \times \times \ c_2 \ \times \times \times \dots \times \times \times \times \times \ d \ \times \times \times \times \times \cdots ,\end{equation} where $\times$ denotes a position unoccupied by any entries $a,b,c,d,e$. At time $t=1,2,3$, the collisions are $a_j b, a_j c_1$ and $b c_1$. Since at these times $d$ is at positions $-y+1, -y+2, -y+3$, there cannot be any entries of the $C$-block in these positions. Furthermore, there cannot be any entries of the $C$-block in position $y-5$ because $a_j$ is at position $y-5$ at time $t=3$ while the collision $bc_1$ occurs. By Lemma \ref{lem-congruence}, the distance between the entries in the $C$-block are even. Hence, there cannot be any entries of the $C$-block at positions $y-2\alpha$ for any integer $\alpha$. At time $t=4$, a collision of the form $ac$, $bd$, or $ce$ must occur. Since $e_1$ has not passed through $d$ and $b(4)>a_j(4)>d(4)$, the only possible collision is $a_jc_2$. Hence, $c_2=y-7$. As in case IB, we now compute that the dimension of the flag variety is $N=2y-3$. At times $N-2$, $N-1$, and $N$, the collisions are specified and do not involve $b$. At these times $b$ is at positions $-y+5, -y+4, -y+3$. Hence, we conclude that there cannot be any entries of the $C$-block at positions $-y+1, \ldots, -y+5$. Since $e_p(N-2) = d(N-2)$ we have $e_p = -3y+5$. The $c_2d$ collision happens at time $2y-7$, and $e_p(2y-7) = y-9$, so there cannot be a $c$ at position $y-9$ either. Combining all these claims, we recover the claimed shape of the sequence at time $t=0$. Next consider the position of $e_1$. Since there are intersections for times $1\leq t \leq 4$, we find $e_1\leq -y-5$. To complete the claim that the entries of the partition are as in $(\ast)$, we only need to show that in fact $e_1< -y-5$. So suppose $e_1 = -y-5$. Since $c_2 = y-7$, the intersection $c_2e_1$ occurs at time $y-1$. However, since $a_j = y+1$ and $e_p = -3y+5$, the intersection $a_j e_p$ also occurs at time $y-1$. This contradiction shows $e_1< -y-5$, and establishes the pattern $(\ast)$ for the entries. We finally obtain a contradiction by showing there is no possible collision at time $t=5$. The collision cannot be a $de$ collision since $e_1<-y-5$. Since nothing has passed through the $d$ before time $5$, the time $t=5$ intersection cannot involve an $e$. Since there are no $c$ entries at position $-y+5$, the collision cannot be of the form $cd$. Since there are no $c$ entries at position $y-5$, the collision cannot be of the form $bc$. Since $b$ has not passed though $c_2$, the collision cannot be of the form $bd$. Since there are no $c$ entries in position $y-9$, the collision cannot be $a_j c$. Since the entries in the $A$-block are at least 12 apart, the collision cannot be $a_{j-1} c_1$. We conclude that the collision is not of the form $ac$. For the same reason, the collision cannot be of the form $a b$. This only leaves the possibility of the collision being between $a_1$ and $d$. This implies $y=7$, so $c_2=0$ and there is a triple intersection by $b$, $c_2$, and $d$ at time $7$. This concludes the proof that there cannot be an Ulrich partition for four-step flag varieties. \end{proof} Theorem \ref{thm-higherStep} follows immediately from Propositions \ref{prop-fiveStep}, \ref{prop-threeStep} and \ref{prop-fourStep}. \section{Results on two-step flag varieties}\label{sec-overview2step} The rest of the paper is entirely combinatorial and will focus on the classification of Ulrich partitions with three parts, corresponding to Ulrich bundles $E_\lambda$ on two-step flag varieties with respect to $\OO(1)$. In this section, we state our classification results, and outline the plan for the remainder of the paper. From now on, $F(p,q;v)$ denotes a two-step flag variety. Let $(\alpha,\beta,\gamma)= (p, q-p, v-q)$ be the type of a partition. We present two different flavors of classification results. We classify Ulrich partitions where $\beta \leq 2$ but $\alpha$ and $\gamma$ are arbitrary. We then also classify Ulrich partitions where $\alpha \leq 2$ and $\gamma \leq 2$ but $\beta$ is arbitrary. We begin by analyzing the simplest case of $(1,n,1)$ partitions, which correspond to Schur bundles on $F(1, n; n+1)$. \begin{theorem}\label{thm-1n1} There are $2^n$ Ulrich partitions of type $(1,n,1)$ for any $n\geq 1$ (up to equivalence). \end{theorem} \begin{proof}Indeed, suppose $(a|b_1,\ldots,b_n|c)$ is Ulrich. Since $a$ and $c$ have the same parity, we may normalize $a = y+1$, $c = -y-1$ for some integer $y\geq 0$. At time $y+1$, we have the $ac$ intersection. Before time $y+1$, every intersection is of type $ab$ or $bc$. Thus, for every position $p\in [y]$, exactly one of $p$ or $-p$ is in $B=\{b_1,\ldots,b_n\}$. We conclude $y = n$, and if $B'\subset [n]$ is any subset then there is a unique Ulrich partition $(n+1|B|{-n-1})$ with $B\cap [n] = B'$. Thus equivalence classes of Ulrich partitions of type $(1,n,1)$ are in bijective correspondence with subsets of $[n]$. \end{proof} \begin{example} The Ulrich partition $(5|3,-1,-2,-4|{-5})$ corresponding to $B' = \{3\}\subset [4]$ in the previous construction was already studied in Example \ref{ex-combToAG}. \end{example} We next describe the small $\beta$ case, where $\beta\leq 2$. There are new one- and two-parameter infinite families of examples; let $m_1,m_2\geq 0$ be nonnegative integers and define $$k_i = \sum_{j=0}^{m_i} 4^j = \frac{4^{m_i+1}-1}{3}.$$ These sums of powers of $4$ will somewhat surprisingly appear in many examples. We prove the next result in \S \ref{sec-sumset} by recasting it as a problem about sumsets. \begin{theorem}\label{thm-beta1intro} For every $m_1\geq 0$, there is a unique Ulrich partition of type $(2,1,k_1)$. Up to symmetry, any Ulrich partition of type $(\alpha,1,\gamma)$ is of this form. \end{theorem} The $\beta = 2$ case is more difficult, and occupies \S \ref{sec-beta2}. \begin{theorem}\label{thm-beta2intro} There are Ulrich partitions of the following types: \begin{enumerate} \item\label{case222} $(2,2,2)$, \item\label{case223} $(2,2,3)$, \item\label{case12k} $(1,2,k_1)$ for any $m_1\geq 0$, and \item\label{case12kk} $(1,2,k_1+k_2)$ for any $m_1,m_2\geq 0$. \end{enumerate} Unless the partition is of type $(1,2,2)$ or $(2,2,2)$, it is unique up to taking the symmetric dual. Up to symmetry, any Ulrich partition of type $(\alpha,2,\gamma)$ is one of the above examples. \end{theorem} We next turn our attention to cases where $\alpha$ and $\gamma$ are small and $\beta$ is arbitrary. We first classify all partitions of type $(1,n,2)$ in \S\ref{sec-1n2}. \begin{example}\label{example-1n2} There is a fundamental example of this type, given by the partition $$(a|b_1,\ldots,b_n|c_1,c_2) = (2n+1|n-1,n-2,\ldots,1,0|{-1},-2n-3).$$ Then \begin{itemize} \item for times $t\in [1,n]$, the $c_1$ entry meets the $B$-block. \item At time $t = n+1$, $a$ meets $c_1$. \item For times $t\in [n+2,2n+1]$, $a$ meets the $B$-block. \item When $t=2n+2$, $a$ meets $c_2$, and \item for $t\in [2n+3,3n+2]$, $c_2$ meets the $B$-block. \end{itemize} Therefore the partition is Ulrich. In Figure \ref{fig-132} we present the time evolution diagram of the Ulrich partition $(7|2,1,0|{-1},-9)$ of type $(1,3,2)$. \begin{figure}[htbp] \input{132Ex.pstex_t} \caption{Time evolution diagram of the Ulrich partition $(7|2,1,0|{-1},-9)$ of type $(1,3,2)$.}\label{fig-132} \end{figure} \end{example} \begin{theorem}\label{thm-1n2intro} There is a bijective correspondence between Ulrich partitions of type $(1,n,2)$ and decompositions $$n = 2km+m-1 \qquad (k\geq 0, m>0)$$ of the integer $n$. \end{theorem} The fundamental example of a type $(1,n,2)$ partition corresponds to the trivial decomposition of $n$ given by $k=0$ and $m=n+1$. For example, when $n=4$, $k=2$, and $m=1$ we will obtain a partition of type $(1,4,2)$ different from the fundamental example. Finally, we classify all partitions of type $(2,n,2)$ in \S \ref{sec-2n2}. \begin{theorem}\label{thm-2n2intro} If there is an Ulrich partition of type $(2,n,2)$, then $n$ is even. If $n$ is even, there are exactly two partitions of this type, and they are symmetric to each other. \end{theorem} \subsection{Conjectures} Our classification of Ulrich partitions for two-step flag varieties remains incomplete. However, we conjecture that we have found all the examples: \begin{conjecture}\label{conj-fullList} If $(\alpha,\beta,\gamma)$ is the type of an Ulrich partition, then up to symmetry this type is one of $$(1,n,1),\quad (1,n,2),\quad (2,2n,2),\quad (2,2,3),\quad (2,1,k_1),\quad (1,2,k_1),\quad\mbox{or}\quad (1,2,k_1+k_2)$$ where $n$ denotes a positive integer and $k_i$ is a number of the form $(4^{m_i+1}-1)/3$ for some $m_i\geq 0$. \end{conjecture} Our partial classification implies that Conjecture \ref{conj-fullList} is equivalent to the following a priori weaker statement. \begin{conjecture}\label{conj-simple} There are no Ulrich partitions of type $(\alpha,\beta,\gamma)$ if $\beta\geq 3$ and $\gamma \geq 3$. \end{conjecture} Indeed, suppose Conjecture \ref{conj-simple} is true. If $\beta\leq 2$, then Theorems \ref{thm-beta1intro} and \ref{thm-beta2intro} completely classify Ulrich partitions of type $(\alpha,\beta,\gamma)$. So suppose $\beta \geq 3$. If $\alpha \leq 2$ and $\gamma\leq 2$, then Theorems \ref{thm-1n1}, \ref{thm-1n2intro}, and \ref{thm-2n2intro} completely classify Ulrich partitions of type $(\alpha,\beta,\gamma)$. If instead either $\alpha \geq 3$ or $\gamma\geq 3$, then by symmetry we may assume $\gamma\geq 3$ and therefore there are no Ulrich partitions of type $(\alpha,\beta,\gamma)$. Therefore Conjecture \ref{conj-simple} implies Conjecture \ref{conj-fullList}. \begin{remark} As evidence for Conjecture \ref{conj-simple}, we performed a brute-force computer search to verify that if there is an Ulrich partition of type $(\alpha,\beta,\gamma)$ with $\beta\geq 3$ and $\gamma \geq 3$, then $\alpha + \beta +\gamma \geq 13$. We note that Conjecture \ref{conj-simple} is open even in the special case $\alpha=1$. \end{remark} \section{Sumsets and Ulrich partitions of type $(\alpha,1,\gamma)$}\label{sec-sumset} In this section we produce a new class of examples of Ulrich partitions and classify all the types $(\alpha,\beta,\gamma)$ of an Ulrich partition where $\beta = 1$. \begin{theorem}\label{thm-sumset} For every $m\geq 0$, there is a unique Ulrich partition of type $(2,1,k)$, where $k = (4^{m+1}-1)/3$. Up to symmetry, any Ulrich partition of type $(\alpha,1,\gamma)$ is of this form. \end{theorem} \begin{proof} We first rephrase our combinatorial problem to make it more tractable in this case. Given a partition $$(a_1,\ldots,a_\alpha|b|c_1,\ldots,c_\gamma)$$ we first normalize $b=0$ (so the $c_j$ are all negative). The entry $a_i$ meets $b=0$ at time $t= a_i$ and meets $c_j$ at time $\frac{1}{2}(a_i-c_j)$, while $c_j$ meets $b$ at time $t=-c_j$. In order for the $a$'s to meet the $c$'s at integral times, they must all have the same parity. Furthermore, they must all be odd, for otherwise at time $t=1$ no two entries would meet. Let $A =\{a_1,\ldots,a_\alpha\}$ and $C = \{-c_1,\ldots,-c_\gamma\}$. Then the Ulrich condition says precisely that the sumset $A+C:= \{a+c:a\in A,c\in C\}$ has size $\alpha\gamma$ and the set $[N]$ of times is a disjoint union $$ [N] = A\amalg C \amalg \tfrac{1}{2}(A+C).$$ Equivalently, by subtracting $1$ from every element of $A$ and $C$ and putting $N' = N-1$, we see that the existence of an Ulrich partition of type $(\alpha,1,\gamma)$ is equivalent to the existence of subsets of even numbers $A',C' \subset [0,N']$ of sizes $\alpha$ and $\gamma$ such that $[0,N']$ is a disjoint union $$[0,N'] = A'\amalg C' \amalg \tfrac{1}{2}(A'+C').$$ The numerics force $A'+C'$ to have size $\alpha\gamma$---no repetition can occur in the sumset. This decomposition is the problem we will study for the rest of the proof. We now claim that if $(\alpha,1,\gamma)$ is the type of an Ulrich partition then at least one of $\alpha,\gamma$ is $2$. Suppose $A',C'\subset [0,N']$ correspond to such an Ulrich partition, and put $S' = \frac{1}{2}(A'+C')$. Since $A'$ and $C'$ are disjoint, we have $0\notin S'$, and therefore either $0\in A'$ or $0\in C'$. By symmetry we may assume $0\in A'$. We must have $1\in S'$ since $A',C'$ consist of even numbers, and this forces $2\in C'$. The number $N'$ must be even since $A'$ and $C'$ consist of even numbers no larger than $N'$. Then $N'$ must actually be in $A'$ or $C'$. If $N'\in C'$ then since $N'-1$ must be in $S'$ we have to have $N'-2\in A'$. This is a contradiction since $0+N' = 2+(N'-2)$ gives a repetition in the sumset $A'+C'$. We conclude $N'\in A'$ and similarly $N'-2\in C'$. We next show by induction that if $k$ is an integer with $k\geq 0$ and $4k+2<N'$ then $4k+2\in C'$. We already know $2\in C'$. Suppose $4k+2\in C'$ for all $k\in [0,k_0)$. Then $4k+2 \notin A'$ for all $k\in [0,k_0)$. Furthermore, $4k+4\notin A'$ for all $k\in [0,k_0)$, for if $4k+4\in A'$ then $(4k+4)+(N'-2) = N'+(4k+2)$ is a repetition in the sumset. Thus $A'$ contains no numbers in the interval $(0,4k_0]$. The odd number $2k_0+1$ must be in $S'$, but since every element of $C'$ is $\geq 2$ while the only nonzero elements of $A'$ are at least $4k_0+2$ the only possibility is that $4k_0+2\in C'$. By induction, $C'$ must contain all numbers in $(0,N')$ which are $2 \pmod{4}$. This argument also shows that $A' = \{0,N'\}$, i.e. $\alpha=2$. Suppose $N'$ is not divisible by $4$. Reversing the argument in the previous paragraph, we see that $C'$ must consist precisely of all the even numbers in $(0,N')$. In particular, since $N'>4$ we have $4\in C'$, so $2\in S'$; this contradicts that $2\in C'$. Therefore $4$ divides $N'$. At this point we conclude that the sets $A',C'\subset [0,N']$ must have a recursive structure. Let $$N'' = N'/4 \qquad A'' = \{0,N''\} \qquad C''=\{c:4c\in C'\} \qquad C''' = \{c\in [0,N']:c\equiv 2 \pmod 4\},$$ so that $C' = 4C'' \cup C'''$. Then $$C''' \cup \tfrac{1}{2}(A'+C''') = \{d\in [0,N']:d\not\equiv 0 \pmod 4\};$$ furthermore, this union is disjoint and there is no repetition in the sumset. It then follows that the sets $A',C'$ correspond to an Ulrich partition if and only if $$[0,N''] = A'' \amalg C'' \amalg \tfrac{1}{2}(A''+C'').$$ That is, either $A'',C''$ correspond to an Ulrich partition or $C''$ is empty, $N'' = 1$, and $A'' = \{0,1\}$, in which case we originally had $A' = \{0,4\}$ and $C' = \{2\}$. Since $C''$ always has fewer elements than $C'$, we conclude by induction. We may assume $C''$ has $(4^m-1)/3$ elements for some $m\geq 1$. Then $$N'' +1=|A''|+|C''|+|A''||C''|=2+\frac{(4^{m}-1)}{3}+\frac{2(4^m-1)}{3}=4^m+1,$$ so $N'' = 4^m$ and $N' = 4^{m+1}$. From this we similarly conclude $|C'| = (4^{m+1}-1)/3$, completing the proof. \end{proof} \begin{example} By studying the proof of Theorem \ref{thm-sumset}, it is straightforward to recover the Ulrich partitions from the sumset construction. As an example, let us find the Ulrich partition of type $(2,1,5)$, corresponding to $m=1$. Then $N = 17$, so $N' = 16$, $A' = \{0,16\}$, and $C' = \{2,6,8,10,14\}$. Correspondingly, the Ulrich partition is $(17,1|0|{-3},-7,-9,-11,-15)$. We give the time evolution diagram of this partition in Figure \ref{fig-215}. \begin{figure}[htbp] \input{215Ex.pstex_t} \caption{Time evolution diagram of the Ulrich partition of type $(2,1,5)$.}\label{fig-215} \end{figure} \end{example} \section{Ulrich partitions of type $(\alpha,2,\gamma)$}\label{sec-beta2} In this section we completely classify Ulrich partitions of type $(\alpha,2,\gamma)$. There are many examples of such partitions, including a two-parameter infinite family and several sporadic examples. The existence of these examples complicates the classification. \subsection{The greedy algorithm} Most of our arguments on 3-part Ulrich partitions will involve studying the following algorithm for the construction of Ulrich partitions. We normalize the evolution of partitions by subtracting $1,0,-1$, respectively, from the three blocks. Fix a finite set $B\subset \Z$, which we view as being the list of entries in the $B$-block of a hypothetical Ulrich partition. Let $A\subset \Z$ and $C\subset \Z$ be two finite sets such that $a>b>c$ holds for any $a\in A$, $b\in B$, $c\in C$. We write $$A(t) = A -t \qquad B(t) = B \qquad C(t) = C+t,$$ where the right hand side is interpreted as a sumset, so that e.g. $A(t)$ gives the set of positions of the $A$-entries of $(A|B|C)$ at time $t$. \begin{definition} The partition $(A|B|C)$ is \emph{pre-Ulrich} if \begin{enumerate} \item all elements of $A$ and $C$ have the same parity, and \item the set $T\subset \N_{>0}$ of times $t$ where two of $A(t), B(t),C(t)$ have a common entry has size $N$, the dimension of $(A|B|C)$. \end{enumerate} \end{definition} \begin{remark} Write $(A'|B'|C')\subset (A|B|C)$ if $A'\subset A$, $B'\subset B$, and $C'\subset C$. Then if $(A|B|C)$ is pre-Ulrich, $(A'|B'|C')$ is pre-Ulrich. Clearly Ulrich partitions are pre-Ulrich; the Ulrich condition requires that in addition $T = [N]$ is a consecutive range of integers.\end{remark} Given any pre-Ulrich triple $(A|B|C)$, we introduce two operations for enlarging the $A$ or $C$ part of the triple. \begin{definition} Let $(A|B|C)$ be a pre-Ulrich partition. Let $t_0\in \N_{>0}$ be the smallest time not in $T$. \begin{enumerate} \item Put $a_0(t_0) := \max(B(t_0) \cup C(t_0))$ and $a_0 := a_0(t_0)+t_0$. The triple $(A\cup \{a_0\} |B|C)$ is \emph{obtained by adding a new $a$}. \item Put $c_0(t_0)=\min(A(t_0)\cup B(t_0))$ and $c_0 = c_0(t_0)-t_0$. The triple $(A|B|C\cup \{c_0\})$ is \emph{obtained by adding a new $c$}. \end{enumerate} \end{definition} \begin{warning} It is not generally the case that the triple obtained from a pre-Ulrich triple by adding a new $a$ or new $c$ is again pre-Ulrich, since the new entry may meet old entries at times after $t_0$ that already have intersections. See the proof of Lemma \ref{lem-BblockRestrict} for an example of this. \end{warning} The next result shows that an Ulrich partition $(A|B|C)$ can always be obtained from $(\emptyset|B|\emptyset)$ by greedily adding $a$'s and $c$'s. \begin{proposition} If $(A|B|C)$ is an Ulrich triple, then it can be obtained by repeatedly adding new $a$'s and $c$'s to the triple $(\emptyset|B|\emptyset)$. Furthermore, the sequence of $a$'s and $c$'s which are added is uniquely determined. \end{proposition} \begin{proof} The sequence of $a$'s and $c$'s to add to $(\emptyset|B|\emptyset)$ is readily determined by the pattern of intersections of the Ulrich partition. Let $(A'|B|C') \subset (A|B|C)$ be any triple. Since $(A|B|C)$ is Ulrich, $(A'|B|C')$ is pre-Ulrich. Let $t_0\in \N_{>0}$ be the smallest time where $(A'|B|C')$ does not have an intersection, and consider the intersection occurring in $(A|B|C)$ at time $t_0$. First suppose the intersection at time $t_0$ occurs between the $A$- and $B$- blocks, between an entry $a_0\in A$ and $b_0\in B$, so that $a_0(t_0)= b_0(t_0)$. Then $a_0\notin A'$, since otherwise $(A'|B|C')$ has an intersection at time $t_0$. No intersections between $a_0$ and an entry of $B$ or $C'$ have occurred before time $t_0$ since all these times already have intersections from $(A'|B|C')$. In order for the intersections between $a_0$ and the entries of $B$, $C'$ to occur at time $t_0$ or later, we must have $a_0(t_0) \geq b(t_0)$ and $a_0(t_0) \geq c(t_0)$ for all $b\in B$ and $c\in C$, and we conclude $a_0(t_0) = \max(B(t_0)\cup C(t_0))$. The triple obtained from $(A'|B|C')$ by adding a new $a$ is precisely $(A'\cup \{a_0\}|B|C')\subset (A|B|C)$. If instead the intersection at time $t_0$ occurs between the $B$- and $C$-blocks, a symmetric argument shows that the triple $(A',B,C'\cup \{c_0\})$ obtained by adding a new $c$ is contained in $(A,B,C)$. Finally suppose the intersection at time $t_0$ occurs between the $A$- and $C$-blocks, and let $a_0\in A$ and $c_0\in C$ be such that $a_0(t_0) = c_0(t_0)$. By the choice of $t_0$, it can't be the case that both $a_0\in A'$ and $c_0\in C'$; we claim that in fact exactly one of $a_0\in A'$ or $c_0\in C'$ holds. Suppose $a_0\notin A'$. As before, this implies that $a_0(t_0) \geq b(t_0)$ and $a_0(t_0) \geq c(t_0)$ for all $b\in B$ and $c\in C'$, so $a_0(t_0) = \max(B(t_0)\cup C(t_0))$. In fact, $a_0(t_0)>b(t_0)$ for all $b\in B$ since $a_0$ has not yet met the $B$-block at time $t_0$; fix some $b_0\in B$. Since $c_0(t_0)=a_0(t_0) >b_0(t_0)$, the intersection between $c_0$ and $b_0$ must have occurred at a time before $t_0$. This implies $c_0\in C'$. Thus if $a_0\notin A'$, we conclude that the triple obtained from $(A'|B|C')$ by adding a new $a$ is $(A'\cup \{a_0\}|B|C')\subset (A|B|C)$; similarly, if $c_0\notin C'$, then the triple obtained from $(A'|B|C')$ by adding a new $c$ is $(A'|B|C'\cup \{c_0\})\subset (A|B|C)$. Starting from $(\emptyset|B|\emptyset)$, we can now construct a chain $$(\emptyset|B|\emptyset) \subset (A_1|B|C_1)\subset \cdots \subset (A_n|B|C_n) = (A|B|C)$$ of pre-Ulrich partitions where each triple differs from the previous one by adding an $a$ or a $c$. Uniqueness is clear. \end{proof} \subsection{Duality and the trapezoid rule} The greedy algorithm will allow us to determine the structure of an Ulrich partition $(A|B|C)$ for early times $t$. Recall that the dual $$(A|B|C)^* := ( A^* |B^*|C^*) := (C(N+1),B(N+1),A(N+1))$$ is also an Ulrich partition (see \S\ref{ssec-symmetryDuality}); this fact will allow us to determine the structure of an Ulrich partition at late (i.e. close to $N$) times. The trapezoid rule is a simple observation which will allow us to combine information about a partition and its dual to say something about middle (i.e. close to $N/2$) times. This is our principal tool for classifying Ulrich partitions of type $(\alpha,\beta,\gamma)$ with $\beta = 2$. \begin{observation}[Trapezoid rule]\label{obsTrapezoid} Let $(A|B|C)$ be an Ulrich partition and $(A^*|B^*|C^*)$ its dual. If there exist $a\in A$, $a^*\in A^*$, $c\in C$, and $c^*\in C^*$ such that $a^*-a = c-c^*$, then $c(N+1) = a^*$ and $a(N+1) = c^*$. In particular, $N+1=a^*-c=a-c^*$. \end{observation} \begin{proof} The assumptions give that $c':=a^*-(N+1)\in C$ and $a':=c^*+(N+1)\in A$. Then $a$ meets $c'$ at time $\frac{1}{2}(a-c')=\frac{1}{2}(a-a^*+(N+1))$ and $a'$ meets $c$ at time $\frac{1}{2}(a'-c)=\frac{1}{2}(c^*-c+(N+1))$. The hypothesis is that these times are equal; thus since $(A|B|C)$ is Ulrich we must have $a=a'$ and $c=c'$. \end{proof} The name of the trapezoid rule comes from the following geometric interpretation using time evolution diagrams. Suppose we can find $a,a^*,c,c^*$ as in the trapezoid rule. View $a$ and $c$ as being entries of $(A|B|C)$ at time $0$, and view $a^*$ and $c^*$ as being entries of $(C|B|A)$ at time $N+1$. The plane trapezoid with vertices at $(a,0)$, $(c,0)$, $(c^*,N+1)$, and $(a^*,N+1)$ is horizontally symmetric by assumption. The conclusion of the trapezoid rule is that this trapezoid has diagonals which meet orthogonally. See Figure \ref{fig-trapezoid}. \begin{figure}[htbp] \input{trapezoid.pstex_t} \caption{Graphical depiction of the trapezoid rule. If $(A|B|C)$ is Ulrich and $a\in A,$ $a^*\in A^*,$ $c\in C,$ and $c^*\in C^*$ can be chosen such that the trapezoid displayed above is horizontally symmetric, then the diagonals meet orthogonally. Equivalently, $N+1 = a^*-c = a-c^*$.}\label{fig-trapezoid} \end{figure} See \S\ref{diff5} for the first simple application of this rule; more intricate applications will be discussed throughout the rest of the section. We also point out one trivial rule which excludes many combinations of configurations of intersections for the early and late times. \begin{observation}[Rectangle rule]\label{obsRectangle} Let $(A|B|C)$ be Ulrich, and let $(A^*|B^*|C^*)$ be its dual. It is not possible that $A$ and $A^*$ share (at least) two entries. \end{observation} \begin{proof} The hypotheses imply that there are two entries in $A$ with the same gap between them as two entries in $C$; there will necessarily be a multiple intersection at some time. \end{proof} Of course, by symmetry, an analogous version of the rectangle rule holds for $C$ and $C^*$. \subsection{Restricting the $B$-block} Suppose $(A|b_1,b_2|C)$ is an Ulrich partition. Our first result restricts the gap $b_1-b_2$ in the $B$-block. \begin{lemma}\label{lem-BblockRestrict} If $(A|b_1,b_2|C)$ is an Ulrich partition then $$b_1-b_2\in \{1,3,5\}.$$ \end{lemma} \begin{proof} First let us show that $b_1-b_2$ is odd. By way of contradiction, suppose $b_1=k$ and $b_2=-k$. We consider what happens when we add new $a$'s and $c$'s to the pre-Ulrich triple $(\emptyset|k,-k|\emptyset)$. Without loss of generality, we first add a new $a$, giving the triple $(k+1|k,-k|\emptyset)$. This triple has intersections at times $1$ and $2k+1$; in particular there is no intersection at time $2$. If we add a new $a$ to this triple we get $(k+2,k+1|k,-k|\emptyset)$, while if we add a new $c$ we get $(k+1|k,-k|{-k-2})$. Neither triple is pre-Ulrich, since they both violate the parity requirement. Next suppose that $b_1-b_2$ is odd and $\geq 7$. We consider adding $a$'s and $c$'s to $(\emptyset|k,-k-1|\emptyset)$. Without loss of generality, we first add an $a$ at time $1$ to get $(k+1|k,-k-1|\emptyset)$. The triple is no longer pre-Ulrich if we add an $a$ at time $2$, so we add a $c$ at time $2$ and get $(k+1|k,-k-1|{-k-3})$. At times $3$ and $4$ (which do not have intersections yet since $k\geq 3$) the parity condition implies we must add an $a$ and then another $c$, yielding $$(k+3,k+1|k,-k-1|{-k-3},-k-5).$$ This partition is no longer pre-Ulrich since both the $A$- and $C$-blocks have entries which are $2$ apart; there is a multiple intersection at time $k+3$. \end{proof} The classification varies considerably according to the value of $b_1-b_2$. We consider each case separately in the next several subsections. \subsection{Classification when $b_1-b_2=5$}\label{diff5} This is the easiest case to classify. For this subsection we fix $B = \{5,0\}$, let $(A|B|C)$ be an Ulrich partition, and assume that the intersection at time $t=1$ occurs between the $A$- and $B$-blocks, so $6\in A$. After these choices, we prove there is a unique such partition. \begin{example}[The (2,2,1) partition]\label{ex-221} The partition $(8,6|5,0|{-2})$ is Ulrich. It is obtained from $(\emptyset|5,0|\emptyset)$ by adding an $a$, then a $c$, then an $a$. See Figure \ref{fig-221}. \begin{figure}[htbp] \input{221Ex.pstex_t} \caption{Time evolution diagram of the Ulrich partition $(8,6|5,0|{-2})$ of type $(2,2,1)$.}\label{fig-221} \end{figure} \end{example} \begin{lemma}\label{diff5contain} The partition $(A|B|C)$ contains $(8,6|5,0|{-2})$. \end{lemma} \begin{proof} Consider the sequence of $a$'s and $c$'s which must be added to $(\emptyset|5,0|\emptyset)$ to produce $(A|B|C)$. By assumption, at time $t=1$ an $a$ must be added, giving $(6|5,0|\emptyset)$. By parity, at time $t=2$ a $c$ is added, giving $(6|5,0|{-2})$. Again by parity, at time $t=3$ an $a$ is added, yielding $(8,6|5,0|{-2})$ as required. \end{proof} The next proposition is now our first application of duality and the trapezoid rule. \begin{proposition} The partition $(A|B|C)$ equals $(8,6|5,0|{-2})$. \end{proposition} \begin{proof} By Lemma \ref{diff5contain}, we have $(8,6|5,0|{-2})\subset (A|5,0|C)$. The dual partition $(A^*|B^*|C^*)$ is Ulrich. Thus, by Lemma \ref{diff5contain}, it contains either $(8,6|5,0|{-2})$ or the symmetric partition $(7|5,0|{-1},-3)$, according to whether the time $t=1$ intersection occurs between the first two or last two blocks. The first possibility is ruled out by the rectangle rule (Observation \ref{obsRectangle}). Likewise, if it contains $(7|5,0|{-1},-3)$, then since $7-6 = -2-(-3)$ the trapezoid rule (Observation \ref{obsTrapezoid}) shows $N=8$. Since $(8,6|5,0|{-2})$ has dimension $8$, this implies $(A|B|C) = (8,6|5,0|{-2})$. \end{proof} \subsection{Classification when $b_1-b_2=3$} The classification here is substantially more complicated than when $b_1-b_2=5$. We normalize $B=\{3,0\}$, and assume the first intersection occurs between the $A$- and $B$-blocks, so $4\in A$. There are three such examples of Ulrich partitions. \begin{example}[The (1,2,1) partition]\label{ex121} The partition $(4|3,0|{-2})$ is Ulrich. It is obtained from $(\emptyset|B|\emptyset)$ by adding an $a$ and then a $c$. This is an example of a partition obtained from Theorem \ref{thm-1n1}. \end{example} \begin{example}[The (2,2,2) partition]\label{ex222} The partition $(12,4|3,0|{-2},-8)$ is Ulrich. It is obtained from $(\emptyset|3,0|\emptyset)$ by adding $a,c,c,a$. See Figure \ref{fig-222}. \begin{figure}[htbp] \input{222Ex.pstex_t} \caption{Time evolution diagram of the Ulrich partition $(12,4|3,0|{-2},-8)$ of type $(2,2,2)$.}\label{fig-222} \end{figure} \end{example} \begin{example}[The (3,2,2) partition]\label{ex322} The partition $(16,10,4|3,0|{-2},{-12})$ is Ulrich. It is obtained from $(\emptyset|3,0|\emptyset)$ by adding in order $a,c,a,c,a$. See Figure \ref{fig-322}. \begin{figure}[htbp] \input{322Ex.pstex_t} \caption{Time evolution diagram of the Ulrich partition $(16,10,4|3,0|{-2},-12)$ of type $(3,2,2)$.}\label{fig-322} \end{figure} \end{example} By comparison with the $b_1-b_2=5$ case, there is less forced structure to the early intersections for an Ulrich triple in this case. We let $\sigma\in \{a,c\}^*$ be the string of $a$'s and $c$'s which must be added to $(\emptyset|B|\emptyset)$ to yield the Ulrich triple $(A|B|C)$. We write $|\sigma|$ for the length of $\sigma$. \begin{lemma}\label{lemInitialSegment} If $\sigma \neq ac$ (i.e. if $(A|B|C)$ is not the type $(1,2,1)$ partition of Example \ref{ex121}), then $\sigma$ begins with one of the strings \begin{enumerate} \item $acaaa$, so that $(16,14,10,4|3,0|{-2})\subset (A|B|C)$, \item $acaca$, so that $(16,10,4|3,0|{-2},-12)\subset (A|B|C)$, \item $acca$, so that $(12,4|3,0|{-2},-8)\subset (A|B|C)$, or \item $acccc$, so that $(4|3,0|{-2},-8,-10,-14)\subset (A|B|C)$. \end{enumerate} \end{lemma} \begin{proof} By assumption the first letter of $\sigma$ is $a$, and by parity the second letter must be $c$. Observe that the pre-Ulrich partitions $(10,4|3,0|{-2})$ and $(4|3,0|{-2},-8)$ corresponding to the strings $aca$ and $acc$ are both not Ulrich, so $|\sigma|\geq 4$. To prove the lemma, we must show that unless $\sigma$ begins with $acca$ then $|\sigma|\geq 5$ and the fifth letter in $\sigma$ is determined by the first four. Suppose $\sigma$ begins with $acaa$. This gives $(14,10,4|3,0|{-2})\subset (A|B|C)$. The first time where $(14,10,4|3,0|{-2})$ has no intersection is time $t=9$, and this triple is not Ulrich so $|\sigma|\geq 5$. Adding a $c$ would yield $(14,10,4|3,0|{-2},-14)$, but this is not pre-Ulrich since there is a multiple intersection at time $t=14$. Thus $\sigma$ must begin with $acaaa$. If instead $\sigma$ begins with $acac$, then $(10,4|3,0|{-2},-12)\subset (A|B|C)$, and this triple is not Ulrich so $|\sigma|\geq 5$. Adding a $c$ would yield $(10,4|3,0|{-2},-12,-14)$, which is not pre-Ulrich since there is a multiple intersection at time $t=12$. So, $\sigma$ begins with $acaca$. Finally suppose $\sigma$ begins with $accc$, so $(4|3,0|{-2},-8,-10)\subset (A|B|C)$. This is not Ulrich, and adding an $a$ gives $(16,4|3,0|{-2},-8,-10)$. Considering time $13$ shows this is not pre-Ulrich, so $\sigma$ begins with $acccc$. \end{proof} We now treat each case of Lemma \ref{lemInitialSegment} in further detail. \begin{proposition}\label{propaccc} Case (4) of Lemma \ref{lemInitialSegment} never actually arises: if $(A|B|C)$ is an Ulrich partition normalized as in this subsection, then the word $\sigma$ cannot begin with $acccc$. \end{proposition} \begin{proof} If $\sigma$ begins with $acccc$ then $(4|3,0|{-2},-8,-10,-14)\subset (A|B|C)$, so $N \geq 17$ since there is an intersection at time $17$ in this subpartition. We first study the dual Ulrich partition $(A^*|B^*|C^*)$. By Lemma \ref{lemInitialSegment}, this partition contains one of the following 8 partitions, of which the first 7 are easily ruled out. \begin{enumerate} \item $(16,14,10,4|3,0|{-2})$. In this case the equality $4-4=-2-(-2)$ and the trapezoid rule implies $N=5$, which is absurd. \item $(5|3,0|{-1},-7,-11,-13)$. Here the equality $5-4=-10-(-11)$ and the trapezoid rule gives $N=14$, a contradiction. \item $(16,10,4|3,0|{-2},-12)$. The same logic as in (1) applies. \item $(15,5|3,0|{-1},-7,-13)$. This time the equality $15-4 = -2-(-13)$ and the trapezoid rule gives $N = 16$. \item $(12,4|3,0|{-2},-8)$. Same as (1). \item $(11,5|3,0|{-1},-9)$. The equality $5-4=-8-(-9)$ and the trapezoid rule gives $N=12$. \item $(4|3,0|{-2},-8,-10,-14)$. Same as (1). \item $(17,13,11,5|3,0|{-1}).$ \end{enumerate} We conclude that $(A^*|B^*|C^*)$ must contain $(17,13,11,5|3,0|{-1})$. Note that the sequence of $a$'s and $c$'s which must be added to $(\emptyset,B,\emptyset)$ to arrive at this subpartition is the sequence $caaaa$. As a first observation, we find $|A|\geq 2$, for if $|A|=1$ then $N=4$. Thus there is some $k\geq 4$ such that $\sigma$ begins with $ac^ka$. By way of contradiction, let $k\geq 4$ be minimal such that there is an Ulrich partition $(A|B|C)$ such that the corresponding word $\sigma$ begins with $ac^ka$. It follows from minimality and symmetry that the word $\sigma^*$ corresponding to the dual partition $(A^*|B^*|C^*)$ at least begins with $ca^k$. We now compute the partition $(4|3,0|C_k)$ ($k\geq 1$) obtained from $(\emptyset|B|\emptyset)$ by adding the letters of the word $ac^k$. We label the elements of $C_k$ in decreasing order as $C_k = \{c_1,\ldots,c_k\}$, so $C_{k+1} = C_k \cup \{c_{k+1}\}$. We have $c_1 = -2$. Let $T_k\subset \N_{>0}$ be the set of times which are \emph{not} intersection times for the partition $(4|3,0|C_k)$, so $T_1 = [6,\infty]$. We let $t_{k+1}=\min T_k$. For $k\geq 1$, when a $c$ is added to $(4|3,0|C_k)$ it is added at time $t_{k+1}$ and meets the $A$-entry at that time, so $c_{k+1}(t_{k+1}) = 4-t_{k+1}$ and $c_{k+1} = 4-2t_{k+1}$. Computing the sequence of sets $C_k$ is therefore equivalent to computing the sequence of times $\{t_k\}_k$. The computation of the sequence of times $\{t_k\}_k$ when new $c$'s are added is best explained in terms of a sieve. Adding a new $c$ at time $t_{k+1}$ means we include $c_{k+1} = 4-2t_{k+1}$ in $C_{k+1}$. Then $c_{k+1}(2t_{k+1}-4) = 0$ and $c_{k+1}(2t_{k+1}-1) = 3$, so $c_{k+1}$ meets the $B$-block at times $2t_{k+1}-4$ and $2t_{k+1}-1$. Therefore $T_{k+1} = T_{k} \setminus \{t_{k+1},2t_{k+1}-4,2t_{k+1}-1\}$. We now make this computation explicit. To get started, we have $t_2 = 6$, which sieves out times $8$ and $11$. We must then include $t_3 = 7$, which sieves times $10$ and $13$, etc. We include the result of this sieve computation for small times below; $\times$'s denote times which are sieved out. $$\begin{array}{ccccccccccccccccccccccc}6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & 16 & 17 & 18 & 19 & 20 & 21 & 22 & 23 & 24 & 25 & 26\\ t_2 & t_3 &\times& t_4 & \times & \times &t_5& \times &\times &t_6&t_7&\times&t_8&t_9&\times&t_{10}&t_{11}&\times&t_{12}&t_{13}&\times \\ \\ 27 & 28 & 29 & 30 & 31 & 32 & 33 & 34 & 35 & 36 & 37 & 38 & 39 & 40 & 41 & 42 & 43 & 44 & 45 & 46 & 47 \\ t_{14}&\times&\times&t_{15}&\times&\times&t_{16} &\times&\times & t_{17} &\times&\times& t_{18} & \times & \times& t_{19} & \times & \times & t_{20} & \times& \times \\ \\ 48 &49 & 50 & 51 & 52 & 53 & 54 & 55 & 56 & 57 & 58 & 59 & \cdots & 95 & 96 & 97 & 98 & 99 & 100 \\ t_{21} & \times& \times & t_{22} & t_{23} & \times & t_{24} & t_{25} & \times & t_{26} & t_{27} & \times & \cdots & \times & t_{52} & t_{53} & \times & t_{54} & \times \end{array}$$ The result of the computation is easy to describe. First, every time $t\geq 6$ with $t\equiv 0 \pmod{3}$ appears in $\{t_k\}$. No times with $t \equiv 2 \pmod{3}$ appear. The times congruent to $1 \pmod{3}$ may or may not appear, but the pattern with which they appear is simple. The first time $(7)$ congruent to $1$ appears, the next 2 times (10,13) do not, the next 4 times (16,19,22,25) do appear, the next 8 do not, the next 16 do, etc. This description is straightforward to prove, and completely specifies the sequence $\{t_k\}$. Now suppose that an $a$, call it $a_2$, is added to $( 4|3,0|C_k)$. Unless $k$ is very special, it turns out that the resulting partition is not pre-Ulrich. We have $a_2(t_{k+1}) = c_1(t_{k+1})$, so $a_2 = 2t_{k+1}-2$. Since $\{c_1,c_2,c_3\} = \{-2,-8,-10\}$, we find that $a_2$ will also meet the $C$-block at times $t_{k+1}+3$ and $t_{k+1}+4$. For the resulting partition to be pre-Ulrich, it is therefore necessary that these times are not sieved out. Additionally, if $t_{k+2} = t_{k+1} + 1$ then no intersection with $a_2$ will occur at time $t_{k+2}$, so it is necessary to add another $a$ or $c$ at this time; it must be a $c$ that is added, for otherwise we will have two $a$'s which are $2$ apart from one another, a contradiction since $c_2$ and $c_3$ are already $2$ apart from one another. But then the $c$ which is added at time $t_{k+2}$, call it $c_{k+1}$, will satisfy $c_{k+1} = 4-2t_{k+2} = 2-2t_{k+1}$. Then $a_2$, $0\in B$, and $c_{k+1}$ all coincide at time $2t_{k+1}-2$. We finally conclude that the times $t_{k+1}+1$ must be sieved out and $t_{k+1}+3$ and $t_{k+1}+4$ cannot be sieved out in order for the partition to have a chance of being pre-Ulrich. Inspecting the sieve, the only way for this to occur is for $t_{k+1}$ to be one of the times immediately before the $t_k$'s start occurring in pairs, e.g. $t_5 = 12$, $t_{21} = 48$, or $t_{85} = 192$. A straightforward computation shows that these times are precisely the numbers $t_{k+1}$ of the form $3\cdot 4^m$ with $m\geq 1$, and $k+1 = \sum_{i=0}^m 4^i$. Finally, suppose $t_{k+1} = 3\cdot 4^m$ and $a_2 = 2t_{k+1}-2$. We use the trapezoid rule to show that $(a_2,4|3,0|C_k)$ is not contained in any Ulrich partition $(A|B|C)$. Observe that a $c$ was added at time $\frac{1}{2}t_{k+1}+1$; this time is the last time that was added in the block of pairs of times preceding $t_{k+1}$. Thus $4-2(\frac{1}{2}t_{k+1}+1)=2-t_{k+1}\in C$. Since the dual partition $(A^*|B^*|C^*)$ has corresponding word $\sigma^*$ beginning with $ca^k$ we symmetrically have $-(2-t_{k+1})+3=t_{k+1}+1\in A^*$ and $-1\in C^*$. The equality $$(t_{k+1}+1)-(2t_{k+1}-2)=(2-t_{k+1})-(-1)$$ and the trapezoid rule prove $a_2(N+1) = -1$, from which it follows that $|A|=2$. We also have $N=a_2 = 2t_{k+1}-2$. At time $N$ we have $a_2(N) = 0\in B$, so at time $N-1$ the smallest entry $c_\gamma \in C$ has $c_\gamma(N-1) = 3$. Then $c_\gamma = 3-(N-1)= 4-2(t_{k+1}-1)$. This means that $c_\gamma$ had to be added at time $t_{k+1}-1$, contradicting that this time was sieved out since it is congruent to $2 \pmod{3}$. \end{proof} The next result has a proof that uses the exact same technique as the proof of Proposition \ref{propaccc}, so we omit it. \begin{proposition}\label{propacaa} Case (1) of Lemma \ref{lemInitialSegment} never actually arises: if $(A|B|C)$ is an Ulrich partition normalized as in this subsection, then the word $\sigma$ cannot begin with $acaaa$. \end{proposition} Fortunately, the positive classification results are easier. \begin{proposition}\label{prop322} If $\sigma$ begins with $acaca$ then $\sigma = acaca$ and $(A|B|C)$ is the $(3,2,2)$ partition of Example \ref{ex322}. \end{proposition} \begin{proof} Since $\sigma$ begins in $acaca$, $(A|B|C)$ contains $(16,10,4|3,0|{-2},-12)$, and therefore $N\geq 16$ with equality iff $\sigma = acaca$. By duality, the partition $(A^*|B^*|C^*)$ is Ulrich. By Lemma \ref{lemInitialSegment} and Propositions \ref{propaccc} and \ref{propacaa} it contains one of the following partitions: \begin{enumerate} \item $(16,10,4|3,0|{-2},-12)$; this is impossible by the rectangle rule. \item $(15,5|3,0|{-1},-7,-13)$; in this case, the equality $15-10=-2-(-7)$ and the trapezoid rule give $N=16$, so in fact $\sigma = acaca$. \item $(12,4|3,0|{-2},-8)$; the equality $4-4=-2-(-2)$ and the trapezoid rule give $N = 5$, a contradiction. \item $(11,5|3,0|{-1},-9)$; the equality $11-4=-2-(-9)$ and the trapezoid rule again gives $N=12$. \end{enumerate} Thus the only possibility is that $\sigma = acaca$. \end{proof} Our final result in this subsection completes the classification when $B = \{3,0\}$. \begin{proposition}\label{prop222} If $\sigma$ begins with $acca$ then $\sigma = acca$ and $(A|B|C)$ is the $(2,2,2)$ partition of Example \ref{ex222}. \end{proposition} \begin{proof} The Ulrich partition $(A|B|C)$ contains $(12,4|3,0|{-2},-8)$, so $N\geq 12$ with equality iff $\sigma = acca$. As in the proof of Proposition \ref{prop322}, the dual $(A^*|B^*|C^*)$ contains one of the following partitions: \begin{enumerate} \item $(16,10,4|3,0|{-2},-12)$; this is impossible by the trapezoid rule applied to the equality $4-4 = -2 - (-2)$. \item $(15,5|3,0|{-1},-7,-13)$; the equality $5-12= -8-(-1)$ and the trapezoid rule gives $N=12$ so $\sigma = acca$ (in fact this is also a contradiction, since the dual partition has the wrong structure.) \item $(12,4|3,0|{-2},-8)$; this contradicts the rectangle rule. \item $(11,5|3,0|{-1},-9)$; the equality $11-4=-2-(-9)$ and the trapezoid rule give $N=12$, so $\sigma = acca$. \end{enumerate} We conclude that $\sigma = acca$. \end{proof} \subsection{Classification when $b_1-b_2=1$} In this final case we normalize $B = \{1,0\}$. We first classify an infinite family of examples which serve as primitive building blocks for a further family of examples. \begin{proposition}\label{propn21} Let $(2|1,0|C_k)$ be the partition obtained from $(2|1,0|\emptyset)$ by adding $k$ $c$'s. This partition is Ulrich if and only if $k$ is a number of the form $k = \sum_{i=0}^m 4^i = \frac{1}{3}(4^{m+1}-1)$ for some $m\geq 0$. \end{proposition} \begin{proof} Write $C_k = \{c_1,\ldots,c_k\}$ with increasing entries. We prove the result by induction on $k$; clearly $(2|1,0|C_1)=(2|1,0|{-4})$ is Ulrich. Fix some $m_0\geq 0$ and let $k_0 = \frac{1}{3}(4^{m_0+1}-1)$, and assume $(2|1,0|C_{k_0})$ is Ulrich. The dimension of the flag variety corresponding to the type $(1,2,k_0)$ is $4^{m_0+1}+1$, so the set of times where $(2|1,0|C_{k_0})$ has an intersection is $[1,4^{m_0+1}+1]$. At time $4^{m_0+1}+2$ the $a$ entry is at position $-4^{m_0+1}$, so adding a new $c$ to $(2|1,0|C_{k_0})$ gives $c_{k_0+1}(4^{m_0+1}+2) = -4^{m_0+1}$, i.e. $c_{k_0+1} = -2\cdot 4^{m_0+1}-2.$ This entry intersects the $B$-block at times $2\cdot 4^{m_0+1}+2$ and $2\cdot 4^{m_0+1}+3$. Continuing inductively, if $\ell$ $c$'s are added to $(2|1,0|C_{k_0})$ and $\ell\leq 4^{m_0+1}$, then $c_{k_0+\ell}(4^{m_0+1}+1+\ell) = -4^{m_0+1}+1-\ell$ so $c_{k_0+\ell} = -2\cdot 4^{m_0+1} - 2\ell$, which meets the $B$-block at times $2\cdot 4^{m_0+1}+2\ell$ and $2\cdot 4^{m_0+1}+1+2\ell$. Then if $\ell <4^{m_0+1}$, the partition $(2|1,0|C_{k_0+\ell})$ is not Ulrich since there is no intersection at time $2\cdot 4^{m_0+1}+1$. When $\ell = 4^{m_0+1}$, there are intersections between some $c_{k_0+\ell}$ and the $a$ for all $t\in [4^{m_0+1}+2,2\cdot 4^{m_0+1}+1]$ and between some $c_{k_0+\ell}$ and the $B$-block for all $t\in [2\cdot 4^{m_0+1}+2,4^{m_0+2}+1]$. Therefore $(2|1,0|C_{k_0})$ is Ulrich. \end{proof} \begin{remark} Analyzing the proof of Proposition \ref{propn21}, the sets $C_k$ (equivalently, the numbers $c_k$) which appear are easily computable. Here are the first several terms: $$-4, \qquad -10,-12,-14,-16, \qquad -34,-36,\ldots,-64, \qquad -130,-132,\ldots,-256, \qquad \ldots$$ A decreasing block of $4^k$ even integers ending in $4^{k+1}$ is followed by a gap of size $4^{k+1} + 2$. In Figure \ref{fig-125} we give the time evolution diagram for the partition $(2|1,0|C_5)$ of type $(1,2,5)$. \begin{figure}[htbp] \input{125Ex.pstex_t} \caption{Time evolution diagram of the Ulrich partition $(2|1,0|C_5)$ of type $(1,2,5)$.}\label{fig-125} \end{figure} \end{remark} A more careful analysis proves the following result. \begin{lemma}\label{lem-stupid} Suppose $(A|1,0|C)$ is an Ulrich partition and that the word $\sigma\in \{a,c\}^*$ generating it from $(\emptyset|B|\emptyset)$ begins with $acc$. Then $|A|=1$ and $2\in A$, so $(A|B|C)$ is one of the examples of Proposition \ref{propn21}. \end{lemma} \begin{proof} We will be brief, as the proof is very similar to other arguments in this section. Consider the partition obtained from $(2|1,0|C_k)$ by adding an $a$, with $k\geq 2$. It is easy to show that unless $k$ is of the form $k = (4^{m+1}-1)/3$ (so $m\geq 1$ and the original partition is Ulrich) then the resulting partition is not even pre-Ulrich. On the other hand, if $k$ is of this form, then adding a new $a_2$ at time $t_0 = 4^{m+1}+2$ will give $a_2 = 2\cdot 4^{m+1}$. One shows that we will be forced to add new $c$'s at positions $-3\cdot 4^{m+1}$ and $-4^{m+2}$. Then the $c$ at position $-3\cdot 4^{m+1}$ meets $0\in B$ at the same time as $a_2$ meets the $c$ at position $-4^{m+2}$, a contradiction. \end{proof} Proposition \ref{propn21} and Lemma \ref{lem-stupid} allow us to focus on classifying Ulrich partitions $(A|1,0|C)$ with $2\in A$ and $|A|\geq 2$, since they completely classify Ulrich partitions where $A = \{2\}$. The next lemma explains the importance of the examples of Proposition \ref{propn21} to more general Ulrich partitions. \begin{lemma}\label{lemNoMorecs} Let $(A|1,0|C) $ be an Ulrich partition, and suppose $2\in A$. Let $c_1 = \max C$, so that $c_1$ meets the $B$-block at time $t_0:=-c_1$. Let $A'\subset A$ be those $a$'s which have met the $B$-block before time $t_0$. Then the partition $(A'|1,0|c_1)\subset (A|1,0|C)$ is Ulrich of dimension $t_0+1$, so the dual $(A'|1,0|c_1)^*$ is one of the Ulrich partitions of Proposition \ref{propn21}. \end{lemma} \begin{proof} Write $A = \{a_\alpha,\ldots, a_1\}$ and $C = \{c_1,\ldots, c_\gamma\}$ in decreasing order, and consider how $(A|1,0|C)$ is built from $(\emptyset|1,0|\emptyset)$ by adding $a$'s and $c$'s; let $\sigma\in \{a,c\}^*$ be the corresponding word. Since the time $t=1$ intersection is between $A$ and $B$, $\sigma$ begins with an $a$. We can then write $\sigma = a^k c \sigma'$ for some $k\geq 1$ and some word $\sigma'$. Then $A$ contains the first $k$ even integers $\{2,\ldots,2k\}$ and the intersections at times $t\in [1,2k-1]$ occur between the $A$- and $B$-blocks. At time $2k$ the entry $a_1$ is at position $2-2k$, so $c_1(2k) = -2k+2$ and $c_1 = -4k+2$. Suppose $c_2$ meets $a_1$ at time $t_1$. We claim $t_1>t_0$. Indeed, first notice that $c_1$ meets $a_i$ $(1\leq i\leq k)$ at time $2k-1+i$, so $t_0 \geq 3k$. Since $c_2$ meets $a_1$ at time $t_1$, it meets $a_i$ $(1\leq i\leq k)$ at time $t_1-1+i$. But then assuming $3k\leq t_1 \leq t_0=4k-2$, we find that $c_2$ meets some $a_i$ at the same time as $c_1$ meets $0\in B$. We conclude $t_1>t_0$. Therefore, if $\sigma$ contains at least $2$ $c$'s, then the second $c$ is added after time $t_0$. It follows that if $\sigma' = a^k c a^\ell$, where $\ell$ is the additional number of $\ell$'s which are added before time $t_0$, then $\sigma'$ is an initial segment of $\sigma$, the corresponding subpartition is $(A'|1,0|c_1)$, and this partition is Ulrich. The last intersection in this partition occurs between $c_1$ and $1\in B$, so its dual is $(2|1,0|A'^*)$ and Proposition \ref{propn21} applies. \end{proof} \begin{example}[A two-parameter family of Ulrich partitions]\label{ex2param} Let $m_1,m_2\geq 0$, and let $k_i = \frac{1}{3}(4^{m_i+1}-1)$. Let $(2|1,0|C_{k_1})$ be the Ulrich partition of Proposition \ref{propn21}, and let $(A_{k_2}|1,0|{-1})$ be the partition symmetric to $(2|1,0|C_{k_2})$. There is an Ulrich partition $(A|B|C)$ uniquely specified by the requirements that the type is $(k_1+k_2,2,1)$ and \begin{align*}(2|1,0|C_{k_1})^* &\subset (A|B|C) \\ (A_{k_2}|1,0|{-1})^* &\subset (A|B|C)^*.\end{align*} The dimension of the type $(k_1+k_2,2,1)$ is $N = 4^{m_1+1}+4^{m_2+1}$. Let $t_0 = 4^{m_1+1}$ be as in Lemma \ref{lemNoMorecs}. For times $t\in [1,t_0]$, the pattern of intersections in $(A|B|C)$ is the same as that of $(2|1,0|C_{k_1})^*$. Applying Lemma \ref{lemNoMorecs} to the dual $(A|B|C)^*$, the corresponding time is $t_0^* = 4^{m_2+1}$, and the pattern of intersections in $(A|B|C)$ for times $t\in [t_0+1,N]$ is the same as that of $(A_{k_2}|1,0|{-1})$ for times $t\in [1,t_0^*]$. Thus every time $t\in [1,N]$ has an intersection. Observe that if $m_1\neq m_2$ then the partitions corresponding to $(m_1,m_2)$ and $(m_2,m_1)$ are distinct but related by the symmetric dual. The partition corresponding to $(m_1,m_1)$ is its own symmetric dual. For example, consider the case $m_1=0$ and $m_2= 1$. Then we compute \begin{align*} N&= 20\\C_{k_1} &= \{-4\} \\ A_{k_2} &= \{17,15,13,11,5\}, \\(2|1,0|C_{k_1})^* &= (2|1,0|{-4})\\ (A_{k_2}|1,0|{-1})^*&= (17|1,0|{-1},-3,-5,-7,-13)\\ (A|B|C) &= (20,18,16,14,8,2|1,0|{-4}). \end{align*} See Figure \ref{fig-621} for the time evolution diagram. Note that the examples of Proposition \ref{propn21} can be regarded as (duals of) the degenerate case where $m_2 = -1$. \begin{figure}[htbp] \input{621Ex.pstex_t} \caption{Time evolution diagram of the partition $(20,18,16,14,8,2|1,0|{-4})$ of type $(6,2,1)$.}\label{fig-621} \end{figure} \end{example} \begin{theorem} If $(A|1,0|C)$ is an Ulrich partition with $2\in A$ and $|A|\geq 2$ then it is either the dual of one of the examples from Proposition \ref{propn21} or there exist $m_1,m_2\geq 0$ such that $(A|1,0|C)$ is the partition corresponding to $(m_1,m_2)$ in Example \ref{ex2param}. \end{theorem} \begin{proof}Applying Lemma \ref{lemNoMorecs} to $(A|B|C)$ and its dual, we find that there are $m_1,m_2\geq 0$ and $k_i = \frac{1}{3}(4^{m_i+1}-1)$ such that $(2|1,0|C_{k_1})^*\subset (A|B|C)$ and either $(2|1,0|C_{k_2})^*\subset (A|B|C)^*$ or $(A_{k_2}|1,0|{-1})^*\subset (A|B|C)^*$, according to whether $2\in A^*$ or $-1\in C^*$. Without loss of generality, assume $k_1\leq k_2$. If $2\in A^*$ then $(2|1,0|C_{k_2})^*\subset (A|B|C)^*$. Computing the dual of the Ulrich partition $(2|1,0|C_{k_2})$ of dimension $4^{m_2+1}+1$, we find $-4^{m_2+1}\in C^*$ and $4^{m_2+1}-2\in A^*$ (since $-4\in C_{k_2}$). Using the containment $(2|1,0|C_{k_1})^*\subset (A|B|C)$, we also have that $-4^{m_1+1}\in C$ and $4^{m_1+1}-2\in A$. Then the equality $$(4^{m_2+1}-2) - (4^{m_1+1}-2) = (-4^{m_1+1})-(-4^{m_2+1})$$ and the trapezoid rule give that $N = 4^{m_1+1}+4^{m_2+1}-1$ and $|C|=1$. This is impossible: the dimension of the type $(k_1+k_2,2,1)$ is $4^{m_1+1}+4^{m_2+1}$, so no type $(k_3,2,1)$ can have dimension $4^{m_1+1}+4^{m_2+1}-1$ by considering congruences mod $3$. Therefore it must be the case that $-1\in C^*$. We now know that $(2|1,0|C_{k_1})^*\subset (A|B|C)$ and $(A_{k_2}|1,0|{-1})^*\subset (A|B|C)^*$. Assume both of these containments are proper, since otherwise we are in the case of Proposition \ref{propn21}. Think of building $(A|B|C)$ from $(\emptyset|B|\emptyset)$ by adding $a$'s and $c$'s. After the word corresponding to $(2|1,0|C_{k_1})^*$ has been added, we must either add an $a$ or a $c$. \emph{Case 1:} Suppose the next letter which is added is an $a$. We claim that the trapezoid rule implies that $(A|B|C)$ is the partition of Example \ref{ex2param} corresponding to the integers $m_1,m_2$. The Ulrich subpartition $(2|1,0|C_{k_1})^*$ has dimension $4^{m_1+1}+1$, so the new $a$ is added at time $4^{m_1+1}+2$. The element of $C$ arising from the inclusion $(2|1,0|C_{k_1})^*\subset (A|B|C)$ is $-4^{m_1+1}\in C$, so the new $a$ is at position $2$ at time $4^{m_1+1}+2$, and thus $4^{m_1+1}+4\in A$. On the other hand, the inclusion $(A_{k_2}|1,0|{-1})^* \subset (A|B|C)^*$ gives $4^{m_2+1}+1 \in A^*$ and $5-(4^{m_2+1}+2)=-4^{m_2+1}+3\in C^*$. We have $$(4^{m_2+1}+1)-(4^{m_1+1}+4)=(-4^{m_1+1})-(-4^{m_2+1}+3)$$ so by the trapezoid rule $N = 4^{m_1+1}+4^{m_2+1}$ and $|C|=1$. Therefore $(A|B|C)$ is the Example \ref{ex2param}. \emph{Case 2:} Suppose the next letter which is added is a $c$. We claim that this is impossible: there are no such Ulrich partitions. We focus solely on the Ulrich partition $(A|B|C) := (2|1,0|C_{k_1})^*,$ as the contradiction arises from this initial segment and not from ``global'' considerations given by the trapezoid rule. If $m_1 = 0$ then Lemma \ref{lem-stupid} gives $|A| = 1$, a contradiction, so we assume $m_1\geq 1$. Let $(A|B|C\cup \{c_2\})$ be the partition obtained by adding $c_2$ at time $t_0:=4^{m_1+1} + 2$. Writing $A = \{a_1,\ldots,a_{k_1}\}$ in increasing order, we have $c_2(t_0) =a_1 ( t_0) = -4^{m_1+1}$ since $a_1 = 2$. For $1\leq i\leq 4^{m_1}$ we have $a_i = 2i$, so $c_2$ meets the $A$-block for all times $t\in [t_0,t_0+4^{m_1}-1]$. At time $t_0+4^{m_1}$ there is no intersection yet. It is not possible to add a new $c$ at this time. If we were to add some $c_3$ at time $t_0+4^{m_1}$ then this would provide intersections between $c_3$ and the $A$-block for the next $4^{m_1}$ times. However, $a_{4^{m_1}+1} = a_{4^{m_1}}+4^{m_1}+2$ meets $c_2$ at time $t_0+4^{m_1}+2\cdot 4^{m_1-1}$, which is a time excluded by the intersection of $c_3$ with the $A$-block. Thus, at time $t_0+4^{m_1}$ we must add a new $a$, call it $a_{k_1+1}$. It has $$a_{k_1+1}(t_0+4^{m_1}) = c_1(t_0+4^{m_1})$$ so $$a_{k_1+1} = -4^{m_1+1}+2(t_0+4^{m_1})=t_0+2\cdot 4^{m_1}+2.$$ By the same argument as in the last paragraph, no new $c$'s can be added before the time $t_1$ when $c_2$ and $a_{k_1}$ meet. At each time in $[t_0,t_1]$ where $c_2$ does not meet the $A$-block, a new $a$ \emph{must} be added. This statement amounts to the claim that $a_{k_1+1}$ does not meet the $B$-block before time $t_1$. Since $a_{k_1}(t_0)=-4$ and $c_2(t_0) = -4^{m_1+1}$, we have $t_1 =t_0+2\cdot 4^{m_1}-2$. Thus $a_{k_1+1}$ meets the $B$-block at times $t_1+3$ and $t_1+4$. Finally, we obtain our contradiction at time $t_1+1$. No intersection has been scheduled yet. We cannot add some $c_3$ at this time, since it would still meet the $A$-block at time $t_1+3$ when $a_{k_1+1}$ meets $1\in B$. Thus we must add an $a$, call it $a'$, at time $t_1+1$. We have $a'(t_1+1)=c_1(t_1+1)$, so $$a'(t_0) = a'(t_1+1)+(t_1+1-t_0)= -4^{m_1+1}+2(t_1+1)-t_0= 4^{m_1+1}$$ while $c_2(t_0) = -4^{m_1+1}$. Thus, at time $t_0+4^{m_1+1}$, all three of $a'$, $0\in B$, and $c_2$ coincide. Therefore, the partition $(A|B|C\cup \{c_2\})$ cannot be extended to an Ulrich partition by adding $a$'s and $c$'s. \end{proof} \section{Ulrich partitions of type $(2,n,1)$}\label{sec-1n2} In this section, we classify Ulrich partitions of type $(2,n,1)$. Throughout the section, we consider Ulrich partitions of the form $P = (a_1,a_2|b_1,\ldots,b_n|c)$, and normalize the evolution of the partition to subtract $1,0,-1$ from the blocks, as in \S\ref{sec-beta2}. We further normalize the positions $$a_2 = y, \quad a_1 = y+2m,\quad \mbox{and}\quad c=-y,$$ so that the intersection $a_2c$ happens at time $y$ and the gap between the $a$-entries is $2m$. In our classification we will view $m$ as being a fixed parameter, similarly to how $b_1-b_2$ was a fixed parameter in the classification of Ulrich partitions of type $(\alpha,2,\gamma)$. \begin{definition} For any $m\geq 2$, the {\em fundamental pattern $F_{m}$ of type $m$} is the partition of type $(2, m-1,1)$ given by $$F_{m}=(3m, m | m-1, m-2, \dots, 2, 1 | {-m}).$$ The fundamental pattern is Ulrich by Example \ref{example-1n2}. \end{definition} \begin{definition} The {\em elongation} of a partition $$P=(a_1, a_2|b_1, \dots, b_n|c) = (y+2m,y|b_1,\ldots,b_n|{-y})$$ of type $(2,n,1)$ is the partition $E(P)$ of type $(2, n + 2m, 1)$ obtained by adding two contiguous blocks of length $m$ at the beginning and end of the $b$ sequence and shifting the $a$ and $c$ entries as follows: $$(y + 5m, y + 3m | y + 3m-1, \dots, y + 2m, b_1, \dots, b_n,- y-m, \dots, -y-2m+1 |{-y- 3m}).$$ The $k$th elongation of $P$ is defined inductively by $E^k(P) := E(E^{k-1}(P))$ and $E^0(P)=P$; it has type $(2,n+2mk,1)$. \end{definition} \begin{example} The fundamental pattern of type 2 is the partition $F_2= (6,2|1|{-2})$. Its first and second elongations are $$E(F_2)= (12, 8|7, 6, 1, -4,-5|{-8}) \quad E^2(F_2) = (18, 14|13,12, 7,6,1,-4,-5, -10,-11 |{ -14}).$$ The fundamental pattern of type 3 is the partition $F_3 = ( 9,3|2,1|{-3})$. Its first elongation is $$E(F_3)= (18, 12|11,10,9,2, 1, -6,-7,-8|{-12}).$$ See Figure \ref{fig-251} for the time evolution diagram of $E(F_2)$. \begin{figure}[htbp] \input{251Ex.pstex_t} \caption{Time evolution diagram of the partition $E(F_2)$ of type $(2,5,1)$.}\label{fig-251} \end{figure} \end{example} \begin{remark} We will also need a degenerate case of the previous definitions. We define the fundamental pattern $F_1$ to be the partition $(3,1|\emptyset|{-1})$. Its elongation $E(F_1) = (6,4|3,-2|{-4})$ still makes sense. Observe that $E(F_1)$ is the Ulrich partition of Example \ref{ex-221}. To avoid discussing trivialities in the arguments that follow we generally focus on the $m\geq 2$ case and assure the reader that appropriate arguments work in the $m=1$ case. \end{remark} The main theorem in this section is the following. \begin{theorem}\label{thm-2n1} A partition $P$ of type $(2,n,1)$ is Ulrich if and only if there exists $k \geq 0$ and $m >0$ such that $n = 2mk + m -1$ and $P=E^k(F_{m})$. \end{theorem} The partitions $E^k(F_m)$ are clearly all distinct, so Theorem \ref{thm-2n1} implies Theorem \ref{thm-1n2intro}. First we observe that the partitions $E^k(F_m)$ are indeed Ulrich. \begin{lemma}\label{lem-2n1} If $P$ is an Ulrich partition of type $(2,n,1)$ and $P$ has dimension $2y+m-1$ then $E(P)$ is an Ulrich partition of dimension $2y'+m-1$, where $y' = y+3m$. In particular, the partition $E^k(F_m)$ is Ulrich of type $(2,2mk+m-1,1)$.\end{lemma} \begin{proof} The beginning and ending intersections in $E(P)$ all occur between $a$'s or $c$'s and the new contiguous blocks of $b$'s as follows. \begin{itemize} \item For $t\in [1,m]$, $a_2$ meets the left new $B$-block. \item For $t\in (m,2m]$, $c$ meets the right new $B$-block. \item For $t\in (2m,3m]$, $a_1$ meets the left new $B$-block. \item For $t\in [2y+4m,2y+5m)$, $a_2$ meets the right new $B$-block. \item For $t\in [2y+5m,2y+6m)$, $c$ meets the left new $B$-block. \item For $t\in [2y+6m,2y+7m)$, $a_1$ meets the right new $B$-block. \end{itemize} Note that $P$ can be obtained from $E(P)$ by shifting to the time $3m$ position $(E(P))(3m)$ and throwing out the two new $B$ blocks. Since $P$ is Ulrich of dimension $2y+m-1$, we conclude that there are unique intersections in $E(P)$ at all times $t\in (3m,2y+4m)$. Clearly $\dim E(P) = \dim P + 6m=2y+7m-1$, so $E(P)$ is Ulrich. For the second statement, it suffices to observe that the fundamental partition $F_m$ satisfies the equality $\dim F_m = 2y+m-1=3m-1$, which is clear. \end{proof} \begin{observation}\label{obs-2n1} By Lemma \ref{lem-2n1}, if $P=(y+2m,y|b_1,\ldots,b_n|{-y})$ is of the form $E^k(F_m)$ then it satisfies the formula $$\dim P = 2y+m-1.$$ For any $P$, we say that it satisfies the \emph{dimension formula} if the above equality holds. Theorem \ref{thm-2n1} in particular claims that the dimension formula holds for any Ulrich partition of type $(2,n,1)$. We also observe that Theorem \ref{thm-2n1} implies that for any Ulrich $P$ the sequence $b_1,\ldots,b_n$ begins with a contiguous block $y-1,\ldots,y-l$ of length exactly $l$, where $l$ is either $m$ or $m-1$ depending on whether $k>0$ or $k=0$ in the equality $P = E^k(F_m)$. \end{observation} The next lemma will form the base of an induction to prove Theorem \ref{thm-2n1}. \begin{lemma}\label{lem-2n1base} Let $P=(a_1,a_2|b_1,\ldots,b_n|c) = (y+2m,y|b_1,\ldots,b_n|{-y})$ be an Ulrich partition. If $y\leq 2m$ then $P = F_m$. \end{lemma} \begin{proof} The intersection $a_2c$ occurs before the intersection $a_1b_1$. It follows that the partition $(y|b_1,\ldots,b_n|{-y})$ is Ulrich of type $(1,n,1)$. Recalling the classification of such partitions, the only possibility is that $n=y-1$, $a_1$ meets $c$ at time $2y$, and $(b_1,\ldots,b_n) = (y-1,\ldots,1)$. \end{proof} On the other hand, if $P$ is too large to be treated by Lemma \ref{lem-2n1base} then we show that it is an elongation of a smaller partition. The next lemma completes the proof of Theorem \ref{thm-2n1}. \begin{lemma} Let $P = (a_1,a_2|b_1,\ldots,b_n|c) = (y+2m,y|b_1,\ldots,b_n|{-y})$ be an Ulrich partition. If $y>2m$ then there is some Ulrich partition $P'$ of type $(2,n',1)$ with $E(P')=P$. \end{lemma} \begin{proof} Inducting on $n$, by Lemma \ref{lem-2n1base} we may assume that any Ulrich partition of the form $$P'=(y'+2m,y'|b'_1,\ldots,b'_{n'}|{-y'})$$ with $n'<n$ is equal to $E^k(F_m)$ for some $k$. In particular, $P'$ satisfies the dimension formula $$\dim P' = 2y' + m-1$$ and the $(b')$'s begin with a contiguous block $y'-1,\ldots,y'-l$ of length exactly $m$ or $m-1$. \emph{Claim 1:} in the partition $P$ the time $t=1$\ intersection is $a_2b_1$, so $b_1 = y-1$. Suppose this is not the case. Then $b_n = -y+1$, and $a_1$ meets $b_n$ at time $2y+2m-1$. By time $2y$ the $a_2$ and $c_1$ entries have already passed through the $B$-block, so all intersections for times $t\in [2y,2y+2m)$ must occur between $a_1$ and the $B$-block. This gives that a contiguous block $B'=\{-y+2m,\ldots,-y+1\}\subset B$ of length $2m$ occurs in the $B$-block. Since $a_2$ meets $B'$ for times in $[2y-2m,2y)$ and $c$ meets $B'$ for times in $[1,2m]$, it follows that the partition $$P'=(y+m,y-m|B\setminus B'|{-y+m})$$ is Ulrich. By induction, $B\sm B'$ starts with a contiguous block $\{y-m-1,\ldots,y-m-l\}$ of length exactly $l \in \{m,m-1\}$. In $P'$ the intersection at time $l+1$ must be between $c$ and some $b_0\in B\setminus B'$. This is a contradiction, since in $P$ we find that $a_1$ meets $b_0$ at the same time as $a_2$ meets an entry of $B'$. Therefore $b_1=y-1$. Having established the claim, let $m_1\geq 1$ be the largest integer such that the contiguous block $B' = \{y-1,\ldots,y-m_1\}\subset B$. Clearly $m_1\leq 2m$, since otherwise $a_1$ and $a_2$ would both intersect $B'$ at the same time. An argument similar to the previous paragraph shows that in fact we must have $m_1<2m$. At time $m_1+1$ the intersection must be $b_nc$; let $m_2\geq 1$ be the largest integer such that the contiguous block $B''=\{-y+m_1+m_2,\ldots,-y+m_1+1\}\subset B$. Observe that $$\dim P = 2y+2m-m_1-1$$ since the last intersection is $a_1b_n$, so $P$ satisfies the dimension formula if and only if $m_1=m$; our eventual goal is to show that $m_1=m_2 = m$. \emph{Claim 2: $m_1+m_2 = 2m$.} Let $t\in (m_1,2m]$. When $a_1$ is at position $-y+t$, both $a_2$ and $c$ have finished intersecting the $B$-block. Thus $-y+t\in B''$ for all $t\in (m_1,2m]$, and so $m_1+m_2\geq 2m$. On the other hand, at time $t=2m+1$ we have an intersection $a_1b_1$, so $-y+2m+1\notin B''$. Thus $m_1+m_2= 2m$. \emph{Claim 3: $y> 2m+m_1$.} By assumption $y>2m$. If $t\in (2m,2m+m_1]$ then $a_1$ meets $B'$ at time $t$, so it is not possible for the intersection $a_2c$ to happen at such a time. Thus $y>2m+m_1$. \emph{Claim 4: $m_1=m_2=m$.} Consider the partition $$P' = (y-m_1,y-2m-m_1|B\sm(B'\cup B'')|{-y+2m+m_1})$$ obtained by looking at the time $2m+m_1$ evolution $P(2m+m_1)$ and removing the contiguous blocks $B',B''$. This makes sense since $y>2m+m_1$, and $P'$ is Ulrich because $P$ is Ulrich. By induction, $P'$ satisfies the dimension formula $$\dim P' = 2(y-2m-m_1)+m-1.$$ On the other hand, the type of $P'$ is $(2,n-2m,1)$, so $$\dim P = \dim P'+6m = 2y+3m-2m_1-1.$$ Comparing this with our earlier expression for $\dim P$ gives $m_1=m$ and so $m_2=m$ as well. This implies $P = E(P')$. \end{proof} \section{Ulrich partitions of type $(2,n,2)$}\label{sec-2n2} In this section, we classify Ulrich partitions of type $(2,n,2)$. Throughout the section, we consider Ulrich partitions of the form $(a_1, a_2| b_1, \dots, b_n| c_1 , c_2)$, and normalize the evolution of the partition to subtract $1, 0, -1$ from the blocks. By symmetry, we may as well assume $a_1-a_2>c_1-c_2$. We further normalize the positions $$a_1 = y + s + r, \quad a_2 = y, \quad c_1 = - y, \quad \mbox{and} \quad c_2 = -y -s.$$ The intersection $a_2 c_1$ occurs at time $y$, the gap between the $a$-entries is $s+r$, and the gap between the $c$-entries is $s$. \begin{definition} Let $P_{u}$ be the partition of type $(2, 2u,2)$ given by $$P_{u} := (6 u +5, 2u+1 | 2 u, 2 u -2, \dots, 4, 2, -1, -3, \dots, - 2 u +3, - 2u +1 | {- 2u -1}, - 6 u -3).$$ We observe that the subpartition $(a_2|B|c_1)$ is the Ulrich partition of type $(1,2u,1)$ corresponding to the subset of $[2u]$ consisting of even numbers in the proof of Theorem \ref{thm-1n1}. \end{definition} \begin{example} We have $$P_1 = (11, 3|2, -1|{-3}, -9),$$ $$P_2 = (17, 5|4,2, -1, -3|{-5}, -15).$$ The time evolution diagram of $P_1$ was given in Example \ref{ex222}. The time diagram for $P_2$ is Figure \ref{fig-242}. \begin{figure}[htbp] \input{242Ex.pstex_t} \caption{Time evolution diagram of the partition $P_2$ of type $(2,4,2)$.}\label{fig-242} \end{figure} \end{example} The main theorem of this section asserts these are the only examples. \begin{theorem}\label{thm-2n2} If $P$ is an Ulrich partition of type $(2, n, 2)$, then $n= 2 u$ is even and up to symmetry $P= P_{u}$. \end{theorem} We first show that these examples are in fact Ulrich. \begin{lemma} The partition $P_{u}$ is Ulrich. In particular, every flag variety $F(2, 2n+2; 2n+4)$ admits a Schur bundle which is Ulrich for $\cO(1)$. \end{lemma} \begin{proof} The dimension of $P_u$ is $N = 8u+4$. For times $t\in[1,4u+1]$, there are intersections from the Ulrich subpartition $(a_2|B|c_1)$ of type $(1,2u,1)$. At time $4u+2$ we have the intersection $a_2 c_2$. As $P_u$ is its own symmetric dual, times in $[4u+3,8u+4]$ also have intersections. \end{proof} The plan of the proof of Theorem \ref{thm-2n2} is similar to the classification of partitions of type $(2,n,1)$ in the previous section. We will first show that if $P$ is an Ulrich partition with $y\leq s$ then $P$ is a known example. We next show that if instead $y>s$ then $P$ can be obtained from a shorter example by a process of elongation. However, we will finally show that elongations of the known examples are never Ulrich; this final step is the primary difference from the strategy in the $(2,n,1)$ case, where such elongations were possible. Before beginning the proof in earnest, we establish a couple of lemmas which are true for arbitrary Ulrich partitions of type $(2,n,2)$. Let $P$ be an Ulrich partition of this type, normalized as in the first paragraph of the section. (In particular, recall that $s+r = a_1-a_2>c_1-c_2 = s$.) \begin{lemma}\label{lem-afirst} In the Ulrich partition $P$, the intersection at time $t=1$ is $a_2 b_1$. \end{lemma} \begin{proof} If not, then the intersection at time $t=1$ is of the form $b c_1$. Let $m\geq 1$ be the maximal number so that $\{-y+1,\ldots,-y+m\}\subset B$, so $-y+m+1\notin B$. This implies that $y-i\notin B$ for $i\in [m]$. Note that $m\leq s$. At the time $t_0 = 2y+s+r-m-1$ when $a_1(t_0) = -y+m+1$ we have $c_2(t_0)=y-m-1+r\geq y-m$, so $c_2(t_0)\notin B$. Furthermore, $c_1(t_0)=y+s+r-m-1>y$ and $a_1(t_0) = -y-s-r+m+1<-y$. We conclude that there is no intersection at time $t_0$ even though $a_2$ has not yet passed through the $B$-block. \end{proof} Lemma \ref{lem-afirst} applied to the symmetric dual implies that the last intersection must be $a_1 b_n$. From now on, we let $m\geq 1$ be the largest number such that $\{y-1,\ldots,y-m\}\subset B$. \begin{lemma}\label{lem-boundm} We have $m<\min\{r,y-1\}$. \end{lemma} \begin{proof} First we show that $m<y-1$. Clearly $m\leq y-1$ since the intersection $a_2c_1$ happens at time $y$. If $m= y-1$ then $B$ is the contiguous block $\{y-1,\ldots,1\}$. By Lemma \ref{lem-afirst} the last intersection in $P$ is between $a_1$ and $1\in B$, so $\dim P = s+r+y-1 = 4y$. At time $2y$ we have $a_2(2y) = -y$ and $c_1(2y) = y$, so the intersection must be of the form $ac$; since $a_2-c_2< a_1-c_2$, it must be $a_2c_2$. Thus $s=2y$, $r = y+1$, and $$P = (4y+1,y|y-1,\ldots,1|{-y},-3y).$$ But then $a_1$ and $c_2$ meet at position $(y+1)/2$, which is in $B$ if $y\geq 3$. Parity is violated if $y=2$, so we conclude $m<y-1$. Next we show $m<r$. If $m>r$ then the last intersection in $P$ is $b_1c_2$, contradicting Lemma \ref{lem-afirst}. Suppose $m=r$. At time $m+1$ the intersection is one of $bc_1$, $a_1b$, or $a_2c_1$. If it is $a_2c_1$ then $m=y-1$ and we are done by the previous paragraph. If the intersection is $bc_1$ then $-y+r+1\in B$ and $c_2$ meets $y-1\in B$ when $a_1$ meets $-y+r+1\in B$, a contradiction. Finally, if the intersection is $a_1b$ then again the last intersection is of type $bc$, violating Lemma \ref{lem-afirst}. \end{proof} \begin{observation}\label{obs-dim2n2} Combining Lemma \ref{lem-afirst} and $\ref{lem-boundm}$, we have $b_n = -y+m+1$ for any Ulrich partition $P$ of type $(2,n,2)$, and thus we have the dimension formula $$\dim P =2y+s+r-m-1.$$ \end{observation} Next we establish the fact that if the intersection $a_2c_1$ happens before $a_1$ or $c_2$ meet the $B$-block then $P$ is a known example. \begin{lemma}\label{lem-fundamental2n2} Let $P$ be an Ulrich partition of type $(2,n,2)$ normalized as in this section. If $y\leq s+m$ then $P = P_{\frac{s-2}{4}}$. \end{lemma} \begin{proof} If $y \leq s+m$, then every intersection until time $y$ is of the form $a_2 b$ or $b c_1$. Hence, there must be a $b$ either at position $p$ or $-p$ for every $1 \leq p < y$. Consequently, the total number of $b$ entries is $y-1$ and for $t \in (y, 2y)$ every intersection is also of the form $a_2 b$ or $b c_1$. By the dimension formula, $4y = 2y+s+r-m-1$, so $2y = s+r-m-1$. At time $t=2y$, $a_2$ and $c_1$ are at positions $-y$ and $y$, respectively. They cannot intersect a $b$. The intersection $a_2 c_2$ happens before $a_1 c_1$, so at time $t = 2 y$ the intersection is $a_2 c_2$ and $s=2y$. At time $2y+1$, $c_2$ cannot intersect a $b$, hence the intersection must be $a_1 c_1$. We conclude that $r=2$, so $m = 1$. We now inductively determine the $B$-block. We have $y-1,-y+2\in B$. When $a_2$ is at position $-y+k$ with $3\leq k<y$ the entry $c_1$ is at position $y-k+2$, while $c_2$ and $a_1$ have already intersected all other entries. By induction we may assume $y-k+2\in B$ iff $k$ is odd; therefore $-y+k\notin B$ iff $k$ is odd and $y-k\in B$ iff $k$ is odd. Finally, note that $y$ is odd. Equivalently, $2\in B$. Otherwise, at time $t = 3y$, $c_2$ is at position $0$ and $a_2$ is at position $2$ and there is no intersection. Therefore $s\equiv 2 \pmod 4$, and $P = P_{\frac{s-2}{4}}$. \end{proof} We next argue that any Ulrich partition of type $(2,n,2)$ must be obtained from some $P_u$ by a process of elongation. Finally, we will see that $P_u$ cannot be elongated. Given a partition $P$ at time $t=0$, we will view it as three sequences $A B C$, where $A$ is the sequence of $a$'s and blank spaces $a_1 \times \cdots \times a_2$, $C$ is the sequence of $c$'s and blank spaces $c_1 \times \cdots \times \ c_2$ and $B$ is the sequence of $b$'s and blank spaces in between $a_2$ and $c_1$. The partition $P$ is a concatenation of these three. Given a contiguous pattern $\Psi$ of $b$'s and blank spaces at positive positions, let $\Psi^c$ be the complementary pattern which has a $b$ at position $p$ if $\Psi$ does not have a $b$ at position $-p$ and vice versa. Let $\ell(\Psi)$ be the length of $\Psi$. Let $X_\ell$ denote a contiguous sequence of blank spaces of length $\ell$. \begin{example} If $\Psi = b \times b \ b \ \times$, then $\Psi^c = b \times \times \ b \ \times $. Also, $X_3 = \times \times \times$. \end{example} \begin{definition} Let $P=ABC$ be a partition and $\Psi$ a pattern of $b$'s and blanks. The \emph{$\Psi$-elongation of $P$} is the partition corresponding to the concatenation $$A \Psi X_{\ell(\Psi)} B X_{\ell(\Psi)} \Psi^c C$$ of patterns. \end{definition} \begin{example} Continuing the previous example, if $$P = a_1 \times \times \times \ a_2 \ b_1 \ b_2 \ \times \times \ b_3 \times \times \ c_1 \times c_2,$$ then the $\Psi$-elongation of $P$ is $$ a_1 \times \times \times a_2 \ b \times b\ b \times \times \times \times \times \times \ b_1 \ b_2 \times \times \ b_3 \times \times \times \times \times \times \times b \times \times \ b \times c_1 \times c_2.$$ \end{example} For integers $q,r>0$, we let $K(q, r,m)$ be the pattern of $b$'s consisting of $q$ iterations of a contiguous block of $m$ $b$'s followed by $r-m$ blanks. \begin{example} $K(2, 6, 2) = b \ b \times \times \times \times \ b \ b \times \times \times \times $. \end{example} Our final lemma easily implies Theorem \ref{thm-2n2}. \begin{lemma}\label{lem-mainMeat} Let $P$ be an Ulrich partition of type $(2,n,2)$, normalized as above. If $y > s+m$, then $r$ divides $s+m$; let the quotient be $q$. Then $P$ is the $K(q, r,m)$-elongation of an Ulrich partition $P'$ of type $(2, n- s-m, 2)$. Furthermore, the initial block of $b$'s in $P'$ also has length $m$. \end{lemma} \begin{proof}[Proof of Theorem \ref{thm-2n2} assuming Lemma \ref{lem-mainMeat}] We have already classified the Ulrich partitions of type $(2,2,2)$ in \S \ref{sec-beta2}. By induction on $n$, suppose that for $n < n_0$ the only Ulrich partitions of type $(2, n, 2)$ are $P_{u}$ with $n=2u$. Let $P$ be an Ulrich partition of type $(2,n_0,2)$. By Lemma \ref{lem-fundamental2n2} we may assume $y>s+m$. By Lemma \ref{lem-mainMeat}, $P$ is the $K(q, r,m)$-elongation of a smaller Ulrich partition $P'$. By induction and Lemma \ref{lem-mainMeat}, we must have $r=2$ and $m=1$. However, $s$ is even, so $r=2$ does not divide $s+m= s+1$. This contradiction proves the theorem. \end{proof} \begin{proof}[Proof of Lemma \ref{lem-mainMeat}] We can determine the pattern of intersections inductively until the time $s+m<y$ immediately before $c_2$ first meets the $B$-block. Until this time, all intersections are of the form $a_1b$ or $bc_2$. For the base of the induction, we first recall $\{y-1,\ldots,y-m\}\subset B$. We know $-y+m+1\in B$ by Lemma \ref{lem-boundm}, and $m<r$. Suppose $v\in (m,r]$. The only possible intersection at time $t_0$ when $a_1(t_0) = -y+v$ is $a_1b$ since $a_2(t_0)=-y+v-r-s < -y$ and $c_2(t_0)=y-v+r\geq y$. Therefore $-y+v\in B$ for $v\in (m,r]$. We now continue by induction until we reach time $s+m$. The positions $y - (h-1)r -v$ have a $b$ for $1 \leq v \leq m$ by induction. Consequently, the positions $-y+ hr + v$ cannot have $b$'s for $1 \leq v \leq m$. Otherwise, when $c_2$ is at position $y- (h-1)r -v$, $a_1$ would be at position $-y + hr + v$ giving a coincident intersection. Thus there is a contiguous $B$-block of length $m$ at positions $y - hr - v$ for $1 \leq v \leq m$. Similarly, there are no $b$ entries in positions $y - (h-1)r - v$ for $m+1 \leq v \leq r$ by induction. When $c_2$ is at these positions $a_1$ is at the positions $-y + hr + v$. Since $a_1b$ is the only possible intersection, there must be a contiguous $B$-block of length $r-m$ at these positions. \emph{Claim: $r|s+m$.} Write $s+m = qr+j$ with $0\leq j<r$ as in the division algorithm. What we have shown so far is that the pattern of $b$'s and blanks in the interval $B \cap (y,y-s-m]$ is the truncation of $K(q,r,m)$ to a sequence of length $s+m$. Similarly, the pattern of $b$'s and blanks in $B\cap [-y+s+m,-y)$ is the truncation of $K(q,r,m)^c$ to a sequence of length $s+m$. As $\ell(K(q,r,m)) = qr$, the claim is that no truncation actually takes place. There are two cases to consider depending on the remainder $j$. \emph{Case 1: $1\leq j\leq m$}. In this case, $y-s-m\in B$. When $c_2$ is at position $y-s-m$, we find that $c_1$ is at position $y-m\in B$, a contradiction. \emph{Case 2: $m< j <r$}. Consider the time $t_0$ when $a_1(t_0) = -y+s+m+1$. Then $c_2(t_0) = y-s-m-1+r\notin B$, and $c_1(t_0)\geq y$, $a_2(t_0)\leq -y$ hold. Furthermore, $-y+s+m+1\notin B$, since $c_1$ is at this position when $c_2$ is at $-y+m+1\in B$. Thus there is no intersection at time $t_0$. Therefore $r|s+m$. Let us analyze the known intersections. For times $t\in [1,s+m]$, the intersections are of type $a_2b$ or $bc_1$. When $t\in (s+m,2s+2m]$, the intersections are all $a_1b$ or $bc_2$. This gives $y-t,-y+t\notin B$ for all such times and $y>2s+2m$. Dually, a symmetric description holds for the last $2s+2(r-m)$ times. Evolving $P$ to time $2s+2m$ and throwing out all the $b$'s which have already met $a$'s and $c$'s, we arrive at an Ulrich partition $$P'=(y-s+r-2m,y-2s-2m|B'|{-y+2s+2m},-y+s+2m)$$ such that $P$ is the $K(q,r,m)$-elongation of $P'$; it is easy to see that $B'$ is nonempty. Finally, we analyze the length $m'$ of the initial block of $b$'s in $P'$. The dimension formula Observation \ref{obs-dim2n2} gives equalities \begin{align*} \dim P' &= 2(y-2s-2m)+r+s-m'-1\\ \dim P &= 2y+r+s-m-1\\ \dim P &= \dim P' + 4(s+m), \end{align*} from which $m=m'$ follows immediately. \end{proof} \bibliographystyle{plain}

https://arxiv.org/abs/1502.03733

Approximation error estimates and inverse inequalities for B-splines of maximum smoothness

In this paper, we develop approximation error estimates as well as corresponding inverse inequalities for B-splines of maximum smoothness, where both the function to be approximated and the approximation error are measured in standard Sobolev norms and semi-norms. The presented approximation error estimates do not depend on the polynomial degree of the splines but only on the grid size.We will see that the approximation lives in a subspace of the classical B-spline space. We show that for this subspace, there is an inverse inequality which is also independent of the polynomial degree. As the approximation error estimate and the inverse inequality show complementary behavior, the results shown in this paper can be used to construct fast iterative methods for solving problems arising from isogeometric discretizations of partial differential equations.

\section{Introduction}\label{sec:intro} The objective of this paper is to prove approximation error estimates as well as corresponding inverse estimates for B-splines of maximum smoothness. The presented approximation error estimates do not depend on the degree of the splines but only on the grid size. All bounds are given in terms of classical Sobolev norms and semi-norms. In approximation theory, B-splines have been studied for a long time and many properties are already well known. We do not go into the details of the existing results but present the results of importance for our study throughout this paper. The emergence of Isogeometric Analysis, introduced in~\cite{Hughes:2005}, sparked new interest in the theoretical properties of B-splines. Since isogeometric Galerkin methods are aimed at solving variational formulations of differential equations, approximation properties measured in Sobolev norms need to be studied. The results presented in this paper improve the results given in \cite{Schumaker:1981,devore:1993,Bazilevs:2006} by explicitly studying the dependence on the polynomial degree~$p$. Such an analysis was done in~\cite{daVeiga:2011}. However, the results there do not cover (for~$p\geq 2$) the most important case of B-splines of maximum smoothness~$k=p-1$. It turns out that the methods established in~\cite{daVeiga:2011} for proving those bounds are not suitable in that case. Therefore, we develop a framework based on Fourier analysis to prove rigorous bounds for $k=p-1$, which has the limitation that it is only applicable for uniform grids. Unlike the aforementioned papers we only consider approximation with B-splines in the parameter domain within the framework of Isogeometric Analysis. A generalization of the results to NURBS as well as the introduction of a geometry mapping, as presented in \cite{Bazilevs:2006}, is straightforward and does not lead to any additional insight. Note that a detailed study of direct and inverse estimates may lead to a deeper understanding of isogeometric multigrid methods and give insight to suitable preconditioning methods. We refer to \cite{Garoni:2014,Donatelli:2015}, where similar techniques were used. \subsection{The main results} We now go through the main results of this paper. For simplicity, we consider the case of one dimension first, where $\Omega = (a,b)$ with $a<b$ is the open \emph{parameter domain}. For this domain we can introduce a \emph{uniform grid} by subdividing $\Omega$ into \emph{elements} (subintervals) of length $\hn$. The setup of a uniform grid is only possible if \begin{equation*} \nh := \hn^{-1}(b-a)\in \mathbb{N}, \end{equation*} where $\mathbb{N}:= \{1,2,3,\ldots\}$. In other words, the grid size $h$ has to be chosen such that $\nh$, the number of subintervals, is an integer. We will assume this implicitly throughout the paper. On these grids we can introduce spaces of spline functions. \begin{definition} The space of spline functions on the domain $\Omega$ of degree $p\in \mathbb{N}_0:=\{0,1,2,\ldots\}$ and continuity $k\in\{-1,0,1,2,\ldots\}$ over the uniform grid of size $\hn$ is given by \begin{equation*} S_{p,k,h}(\Omega) := \left\{ u \in H^k(\Omega): \; u |_{(a+hj,a+h(j+1)]} \in \mathbb{P}^p \mbox{ for all } j=0,\ldots,\nh-1 \right\}, \end{equation*} where $\mathbb{P}^p$ is the space of polynomials of degree $p$. \end{definition} Here and in what follows, $L^2(\Omega)$ and $H^r(\Omega)$ denote the standard Lebesque and Sobolev spaces with norms $\|\cdot\|_{L^2(\Omega)}$, $\|\cdot\|_{H^r(\Omega)}$ and semi-norms $|\cdot|_{H^r(\Omega)}$. Moreover, let $(\cdot,\cdot)_{L^2(\Omega)}$ be the standard scalar product for $L^2(\Omega)$ and \begin{equation*} (u,v)_{H^r(\Omega)} := \left(\frac{\partial^r}{\partial x^r} u,\frac{\partial^r}{\partial x^r} v\right)_{L^2(\Omega)} \end{equation*} be the scalar product for $H^r(\Omega)$, where $\frac{\partial^r}{\partial x^r}$ denotes the $r$-th derivative. We then have $|u|^2_{H^r(\Omega)} := (u,u)_{H^r(\Omega)}$ as well as \begin{equation*} \|u\|^2_{H^r(\Omega)} := \|u\|^2_{L^2(\Omega)} + \sum^r_{s=1} |u|^2_{H^s(\Omega)} \end{equation*} for all $r\in\mathbb{N}_0:=\{0,1,2,\ldots\}$. Using standard trace theorems, we obtain that for $k>0$ the space $S_{p,k,h}(\Omega)$ is the space of all $k-1$ times continuously differentiable functions ($C^{k-1}(\Omega)$-functions), which are polynomials of degree $p$ on each element of the uniform grid on $\Omega$. For $k=0$, there is no continuity condition, i.e., the space $S_{p,0,h}(\Omega)$ is the space of piecewise polynomials of degree $p$. For $k>p$, the spline spaces reduce to spaces of global polynomials. So, the largest possible choice for $k$ without having this effect is $k=p$. Therefore we call B-splines with $k=p$ B-splines of \emph{maximum smoothness}. As we are mostly interested in this case, here and in what follows, we will use $S_{p,h}(\Omega):=S_{p,p,h}(\Omega)$. The main result of this paper is the following. \newcommand{\citeThrmApprox}{1} \begin{theorem}\label{thrm:approx} For all $u\in H^1(\Omega)$, all grid sizes $h$ and each degree $p\in\mathbb{N}$ with ${h\,p < |\Omega| = b-a}$, there is a spline approximation $u_{p,h}\in S_{p,h}(\Omega)$ such that \begin{equation}\label{eq:thrm:approx} \|u-u_{p,h}\|_{L^2(\Omega)} \le \sqrt{2}\; \hn |u|_{H^1(\Omega)} \end{equation} is satisfied. \end{theorem} Note that, in contrast to the existing results presented in the next subsection, this theorem achieves two goals, it covers the case of maximum smoothness and gives a uniform estimate for all polynomial degrees~$p$. \begin{remark} Obviously $S_{p,k,h}(\Omega) \supseteq S_{p,h}(\Omega)$ for all $0\le k < p$. So, Theorem~\ref{thrm:approx} is also valid in that case. However, for this case there might be better estimates for these larger B-spline spaces. Moreover, Theorem~\ref{thrm:approx} is also satisfied in the case of having repeated knots, as this is just a local reduction of the continuity (which enlarges the corresponding space of spline functions). \end{remark} In Section~\ref{sec:reduced}, we will introduce a subspace $\widetilde{S}_{p,h}(\Omega) \subseteq S_{p,h}(\Omega)$ (cf. Definition~\ref{defi:Ssymm}) and show that the spline approximation is even in that subspace (cf. Corollary~\ref{cor:approx:nonper}). Moreover, we show also a corresponding \emph{inverse inequality} for $\widetilde{S}_{p,h}(\Omega)$ (cf. Theorem~\ref{thrm:inverse} in Section~\ref{sec:inverse}), i.e., we will show that \begin{equation*} |u_{p,h}|_{H^1(\Omega)} \le 2 \sqrt{3} \hn^{-1} \|u_{p,h}\|_{L^2(\Omega)} \end{equation*} is satisfied for all grid sizes $h$, each $p\in \mathbb{N}$ and all $u_{p,h}\in \widetilde{S}_{p,h}(\Omega)$. We will moreover show that both the approximation error estimate and the inverse inequality are \emph{sharp up to constants} (Corollaries~\ref{corr:sharp1} and~\ref{corr:sharp2}). \begin{remark}\label{rem:counterexample} This inverse inequality does not extend to the whole space $S_{p,h}(\Omega)$. Here it is easy to find a counterexample: Let $\Omega = (0,1)$. The function $u_{p,h}$, given by \begin{equation*} u_{p,h}(x) = \left\{ \begin{array}{ll} (1-x/\hn)^p & \mbox{\qquad for $x \in [0,\hn)$}\\ 0 & \mbox{\qquad for $x\in [\hn,1]$}, \end{array} \right. \end{equation*} is a member of the space $S_{p,h}(0,1)$. Straight-forward computations yield \begin{equation*} \frac{|u_{p,h}|_{H^1(0,1)}}{\|u_{p,h}\|_{L^2(0,1)}} = \sqrt{\frac{2p+1}{2p-1}} \;p \;\hn^{-1}, \end{equation*} which cannot be bounded from above by a constant times $\hn^{-1}$ uniformly in $p$. Using a standard scaling argument, this counterexample can be extended to any $\Omega=(a,b)$. \end{remark} The approximation error estimate and the inverse inequality are extended to higher Sobolev indices in Section~\ref{sec:sobolev}. Corresponding results for two and more dimensions are given in Section~\ref{sec:dim}. There, also the extension to Isogeometric Analysis is discussed. \subsection{Known approximation error estimates} Before proving the main theorems, we start with recalling two important pre-existing approximation error estimates. The first result is well-known in literature, cf.~\cite{Schumaker:1981}, Theorem~6.25 or \cite{devore:1993}, Theorem~7.3. In the framework of Isogeometric Analysis, such results have been used, e.g., in \cite{Bazilevs:2006}, Lemma~3.3. \begin{theorem}\label{thrm:known} For each $r\in\mathbb{N}_0$, each $k\in\mathbb{N}$, each $q\in\mathbb{N}$ and each $p\in\mathbb{N}$, with $0\le r\le q\le p+1$ and $r\le k \le p$, there is a constant $C(p,k,r,q)$ such that the following approximation error estimate holds. For all $u\in H^q(\Omega)$ and all grid sizes $h$, there is a spline approximation $u_{p,k,h} \in S_{p,k,h}(\Omega)$ such that \begin{equation*} |u-u_{p,k,h}|_{H^r(\Omega)} \le C(p,k,r,q) \hn^{q-r} |u|_{H^q(\Omega)} \end{equation*} is satisfied. \end{theorem} This theorem is valid for tensor-product spaces in any dimension and gives a local bound for locally quasi-uniform knot vectors. However, the dependence of the constant on the polynomial degree has not been derived. A major step towards estimates with explicit $p$-dependence was presented in \cite{daVeiga:2011}, Theorem~2, where an estimate with an explicit dependence on $p$, $k$, $r$ and $q$ was given. However, there the continuity $k$ is limited by the upper bound $\tfrac12(p+1)$. In our notation, the theorem reads as follows. \begin{theorem}\label{thrm:known:2} There is a constant $C>0$ such that for each $r\in\mathbb{N}_0$, each $k\in\mathbb{N}$, each $q\in\mathbb{N}$ and each $p\in\mathbb{N}$ with $0\le r\le k\le q\le p+1$ and $k \le \tfrac12(p+1)$ and all grid sizes $h$, the following approximation error estimate holds. For all $u\in H^q(\Omega)$, there is a spline approximation $u_{p,k,h}\in S_{p,k,h}(\Omega)$ such that \begin{equation}\nonumber |u-u_{p,k,h}|_{H^r(\Omega)} \le C \hn^{q-r} (p-k+1)^{-(q-r)} |u|_{H^q(\Omega)} \end{equation} is satisfied. \end{theorem} Again, the original result was stated for locally quasi-uniform knots. For any $p\geq2$ the relevant case $k=p$, which we consider, is not covered by this theorem. Similar results to Theorem~\ref{thrm:approx} are known in approximation theory, cf.~\cite{Korneichuk:1991}. There, however, different norms have been discussed. Hence we do not go into the details. In~\cite{Evans:2009}, it was suggested and confirmed by numerical experiments that Theorem~\ref{thrm:approx} is satisfied. A proof was however not given. \subsection{Organization of this paper} This paper is organized as follows. In Section~\ref{sec:prelim}, we present the main steps of the proof of Theorem~\ref{thrm:approx} and give some preliminaries. In the following two sections, the details of the proof are worked out. In Section~\ref{sec:reduced}, we introduce the reduced spline space $\widetilde{S}_{p,h}(\Omega)$, discuss its properties and extend Theorem~\ref{thrm:approx} to that space. In the following section, Section~\ref{sec:inverse}, we give an inverse inequality for $\widetilde{S}_{p,h}(\Omega)$. In the remainder of the paper, we generalize those results: In Section~\ref{sec:sobolev} we consider higher Sobolev indices and in Section~\ref{sec:dim}, the results are generalized to two or more dimensions. \section{Concept of the proof of Theorem~\citeThrmApprox{} and Preliminaries}\label{sec:prelim} The proof of Theorem~\ref{thrm:approx} is based on an estimate for periodic splines, which is formulated as Lemma~\ref{lem:approx:per}. The proof of Lemma~\ref{lem:approx:per} is based on a telescoping argument based on a hierarchy of grids. For the proof, we require \begin{itemize} \item an estimate for the difference of the spline approximations of a given function on two consecutive grids, cf.~\eqref{eq:whattolfa}, and \item an estimate for the difference between the spline approximation on some finest grid and the given function, cf.~Lemma~\ref{lem:non:robust}. \end{itemize} As the size of the finest grid approaches $0$, the constant in Lemma~\ref{lem:non:robust} or its dependence on the spline degree $p$ does not matter, whereas the constant in~\eqref{eq:whattolfa} directly affects the constant in the final result. The estimate~\eqref{eq:whattolfa} is shown in Section~\ref{sec:twogrid}, cf. Lemma~\ref{lem:lfa}. There, the proof is done by means of Fourier analysis, which causes the restriction of the analysis to equidistant grids. The Fourier analysis follows a classical line: first, a matrix-vector formulation is introduced, cf. Lemma~\ref{lemma:decomp}, then the symbols of the involved matrices are derived, cf. Subsections~\ref{subsec:symbols} and~\ref{subsec:symbols2}. A closed form for the symbol of the mass matrix is not available, so some statements on that matrix are derived (Lemmas~\ref{lem:mass} and~\ref{lem:mass:estim}), which are used in the proof of Lemma~\ref{lem:lfa}. Having the result for two consecutive grids in the periodic case, we use the aforementioned telescoping argument to give an approximation error estimate for apprximating a general periodic $H^1$-function. The extension to the non-periodic case is done by means of a periodic extension. \subsection{Periodic splines} To establish the theory within this paper, we need to introduce spaces of periodic splines, which we define as follows. \begin{definition}\label{defi:Speriodic} Given a spline space $S_{p,h}(\Omega)$ over $\Omega=(a,b)$, the \emph{periodic spline space} $\widehat{S}_{p,h}(\Omega)$ contains all functions $u_{p,h}\in S_{p,h}(\Omega)$ that satisfy the linear periodicity condition \begin{equation}\label{eq:sym:cond} \frac{\partial^{l}}{\partial x^{l}} u_{p,h}(a)=\frac{\partial^{l}}{\partial x^{l}} u_{p,h}(b) \mbox{ for all } l \in \mathbb{N}_0 \mbox{ with } l < p. \end{equation} \end{definition} The next step is to introduce a B-spline-like basis for this space. First, we introduce the cardinal B-splines. On $\mathbb{R}$, the cardinal B-splines are defined as follows, cf.~\cite{Schumaker:1981}, (4.22). \begin{definition} The cardinal B-splines of degree $p=0$, $\psi^{(i)}_{0}: \;\mathbb{R}\rightarrow\mathbb{R}$ coincide with the characteristic function, i.e., \begin{equation*} \psi^{(i)}_{0}(x) := \left\{ \begin{array}{ll} 1 & \mbox{ for } x \in ( i , i+1 ],\\ 0 & \mbox{ else,} \end{array} \right. \end{equation*} where $i \in \mathbb{Z}$. The cardinal B-splines $\psi^{(i)}_{p}: \;\mathbb{R}\rightarrow\mathbb{R}$ of degree $p\in\mathbb{N}$ are given by the recurrence formula \begin{equation}\label{eq:recur:bspline} \psi^{(i)}_{p}(x) := \frac{x-i}{p} \psi^{(i)}_{p-1}(x) + \frac{(p+i+1)-x}{p} \psi^{(i+1)}_{p-1}(x), \end{equation} where $i \in\mathbb{Z}$. \end{definition} From the cardinal B-splines $\psi^{(i)}_{p}$, we derive the B-splines $\varphi_{p,h}^{(i)}$ on $\Omega$ over a uniform grid of size $\hn$ by a suitable scaling and shifting. \begin{definition} For $i\in\mathbb{Z}$ the uniform B-spline $\varphi_{p,h}^{(i)}: \;\Omega= (a,b)\rightarrow\mathbb{R}$ of degree $p\in\mathbb{N}_0$ and grid size $\hn$ is given by \begin{equation} \varphi_{p,h}^{(i)}(x) := \psi^{(i)}_{p}\left(\frac{x-a}{h}\right). \end{equation} \end{definition} We obtain by construction that $\mbox{supp}(\varphi_{p,h}^{(i)}) \subset [i \hn+a, (i+p+1) \hn+a ]$. Hence, $-p$ and $\nh-1$ with $\nh = \hn^{-1}(b-a)$ are the first and last indices of the B-splines supported in $\Omega$, respectively, i.e. $\textnormal{supp} (\varphi_{p,h}^{(i)}) \cap \Omega \neq \emptyset$ is equivalent to $-p \leq i \leq \nh-1$. Moreover, $\{\varphi_{p,h}^{(i)}\}^{\nh-1}_{i=-p}$ forms a basis for $S_{p,h}$, see, e.g., \cite{Schumaker:1981}. Note that both $\nh$ and the basis functions depend implicitly on the choice of $\Omega$, i.e., the values $a$ and $b$. Throughout the paper, it is clear from the context which $\Omega$ is chosen. For the construction of the basis for the periodic spline space~$\widehat{S}_{p,h}(\Omega)$, we assume that \begin{equation}\label{eq:condition-grid-size} hp < |\Omega| = b-a, \end{equation} i.e., that the grid is fine enough not to have basis functions that are non-zero at both end points of the grid, cf.~\cite{Schumaker:1981}. \begin{definition}\label{defi:basis-per} For $\widehat{S}_{p,h}(\Omega)$, the \emph{B-spline-like basis} $\{\widehat{\varphi}_{p,h}^{(i)}\}^{\nh-1}_{i=0}$ is given by \begin{align*} & \widehat{\varphi}_{p,h}^{(i)}:= \varphi_{p,h}^{(i)}&&\mbox{ if } i<\nh-p, \mbox{ and} \\ & \widehat{\varphi}_{p,h}^{(i)}:= \varphi_{p,h}^{(i)}+\varphi_{p,h}^{(i-\nh)}&&\mbox{ if } i \geq \nh-p. \end{align*} \end{definition} Up to indexing, this definition coincides with~(8.6) and (8.7) in~\cite{Schumaker:1981}. Theorem~8.2 in~\cite{Schumaker:1981} states that~\eqref{eq:basis:varphi} is actually a basis. As $\varphi_{p,h}^{(i)}$ vanishes on $\Omega$ for all $i\not\in \{-p,\ldots,\nh-1\}$, we have \begin{equation}\label{eq:basis:varphi} \widehat{\varphi}_{p,h}^{(i)}=\sum_{j\in\mathbb{Z}} \varphi_{p,h}^{(i+j \nh)}, \end{equation} where $\mathbb{Z}$ is the set of integers, for all $i=0,\ldots, \nh-1$. Using this definition, we directly obtain that also $\widehat{\varphi}_{p,h}^{(i)} = \widehat{\varphi}_{p,h}^{(i+j\nh)}$ for any $j\in \mathbb{Z}$, which we will use for ease of notation throughout this paper. We call this basis B-spline-like, as each function is a non-negative linear combination of B-splines and it forms a partition of unity on $\Omega$. \subsection{A non-robust approximation error estimate in the periodic case} We can extend Theorem~\ref{thrm:known} for $k=p-1$ to the following Lemma \ref{lem:non:robust} stating that the approximation error estimate is still satisfied if we approximate periodic functions with periodic splines. First, we introduce the spaces of periodic functions as follows. \begin{definition} For $\Omega=(a,b)$, the space $\widehat{H}^q(\Omega)$ is the space of all $u\in H^q(\Omega)$ that satisfy the periodicity condition \begin{equation}\label{eq:pc} \frac{\partial^{l}}{\partial x^{l}} u(a)=\frac{\partial^{l}}{\partial x^{l}} u(b) \mbox{ for all } l \in \mathbb{N}_0 \mbox{ with } l < q. \end{equation} \end{definition} Note that standard trace theorems guarantee that the periodicity condition~\eqref{eq:pc} is well-defined. For this space, the following lemma holds. \begin{lemma}\label{lem:non:robust} For each $r\in\mathbb{N}_0$, each $q\in\mathbb{N}$ and each $p\in\mathbb{N}$ with $0\le r\le q\le p+1$, there is a constant $C(p,r,q)$ such that the following approximation error estimate holds. For all $u \in \widehat{H}^{q}(\Omega)$ and all grid sizes $h$, there is a spline approximation $u_{p,h} \in \widehat{S}_{p,h}(\Omega)$ such that \begin{equation*} |u-u_{p,h} |_{H^r(\Omega)} \le C(p,r,q) \hn^{q-r} |u|_{H^q(\Omega)} \end{equation*} is satisfied. \end{lemma} \begin{proof} In the following, we assume without loss of generality that $\Omega=(0,1)$. The extension to any other $\Omega=(a,b)$, follows using a standard scaling argument. Let $w$ be the periodic extension of the function $u$ to $\mathbb{R}$, i.e., $w(x):=u(x-\lfloor x \rfloor)$. Note that the restriction of $w$ to any finite interval is again a function in the Sobolev space~$H^q$. The following of the proof is based on the proof in \S~6.4 in~\cite{Schumaker:1981}. We make use of the fact that the proof uses local projections. Let $Q_{p,h}: H^q(\mathbb{R}) \rightarrow S_{p,h}(\mathbb{R})$ be the projection operator, as introduced in (6.40) in~\cite{Schumaker:1981}. The value of the approximation $Q_{p,h}w$ of a function $w$ at a certain subinterval $I_i:=(i\;h,(i+1)\;h)\subseteq \Omega$ only depends on the values of the function to be approximated in a certain neighborhood $\widetilde{I}_i:=((i-p)\;h,(i+p+1)\;h)$. So, from the periodicity of $w$, the periodicity of $Q_{p,h}w$ follows immediately. Hence its restriction to $(0,1)$ is a periodic spline, i.e. $Q_{p,h}w|_{(0,1)}\in\widehat{S}_{p,h}(0,1)$. We define $u_{p,h}$ to be the restriction of $Q_{p,h}w$ to $(0,1$). Due to \cite{Schumaker:1981}, Theorem~6.24, the local estimate \begin{equation}\nonumbe |w-Q_{p,h}w|_{H^r(I_i)} \le \widetilde{C}(p,r,q) \hn^{q-r} |w|_{H^q(\widetilde{I}_i)}. \end{equation} is satisfied for the projector $Q_{p,h}$ and a constant $\widetilde{C}(p,r,q)$, which is independent of~$\hn$. By summing over all elements, we obtain \begin{align*} &|u-u_{p,h} |_{H^r(0,1)}^2 = |w-Q_{p,h}w |_{H^r(0,1)}^2 = \sum_{i=0}^{\nh-1} |w-Q_{p,h}w|_{H^r(I_i)}^2 \\ & \quad \le \widetilde{C}^2(p,r,q) \hn^{2(q-r)} \sum_{i=0}^{\nh-1} |w|_{H^q(\widetilde{I}_i)}^2 =\widetilde{C}^2(p,r,q) \hn^{2(q-r)} \sum_{i=0}^{\nh-1}\sum_{j=-p}^{p} |w|_{H^q(I_{i+j})}^2. \end{align*} Using the periodicity of $w$, we can express the last term using~$|u|_{H^q(I_{l})}$ for~$l\in\{0,\ldots,\nh-~1\}$ only. By counting the occurrences of the summands~$|u|_{H^q(I_{l})}$, we obtain \begin{equation}\nonumber \sum_{i=0}^{\nh-1}\sum_{j=-p}^{p} |w|_{H^q(I_{i+j})}^2 = (2p+1) \sum_{i=0}^{\nh-1} |u|_{H^q(I_{i})}^2 = (2p+1) |u|_{H^q(0,1)}^2, \end{equation} which finishes the proof for $C(p,r,q) = (2p+1)^{1/2} \widetilde{C}(p,r,q)$.\qed \end{proof} \section{A robust approximation error estimate for two consecutive grids in the periodic case}\label{sec:twogrid} In this section we analyze the case of approximating a periodic spline function on a fine grid by a periodic spline function on a coarser grid. In the next section, we extend these results to the approximation of general functions and to the non-periodic case. The extension to the non-periodic case is done by extending functions in $H^1(0,1)$ to $(-1,1)$ by reflecting them on the $y$-axis. So, without loss of generality, we will restrict ourselves to $\Omega=(-1,1)$ throughout this section. Moreover, for the construction of~\eqref{eq:our:mass:formula}, we will need that $hp<1$, which is stronger than the requirement $hp<b-a$, cf. Theorem~\ref{thrm:approx}. So, throughout this section, we will use the following assumptions. \begin{assumption} The domain is given by $\Omega=(-1,1)$ and the grid size is small enough such that $hp<1$ holds. \end{assumption} In the next section, we will make use of a telescoping argument. For this purpose, we have to analyze a fixed interpolation operator. So, within this section, we will show that \begin{equation}\label{eq:whattolfa} \|(I- \widehat{\Pi}_{p,\hn} ) u_{p,h}\|_{L^2(-1,1)} \le \frac{1}{\sqrt{2}}\; \hn |u_{p,h}|_{H^1(-1,1)} \end{equation} holds for all $u_{p,h}\in \widehat{S}_{p,\frac{h}{2}}(-1,1)$, where $I$ is the identity and $\widehat{\Pi}_{p,\hn}$ is the $H^1$-orthogonal projection operator, given by the following definition. \begin{definition}\label{def:H1projection} The projection $\widehat{\Pi}_{p,\hn}:\widehat{H}^1(-1,1)\rightarrow \widehat{S}_{p,h}(-1,1)$ maps every $u\in \widehat{H}^1(-1,1)$ to the function $u_{p,h} \in \widehat{S}_{p,h}(-1,1)$ satisfying \begin{equation}\label{eq:def:projection} (u_{p,h}, v_{p,h})_{H^1_{\circ}(-1,1)} = (u, v_{p,h})_{H^1_{\circ}(-1,1)} \end{equation} for all $v_{p,h}\in \widehat{S}_{p,h}(-1,1)$, where \begin{equation*} (u,v)_{H^1_{\circ}(-1,1)}:=(u,v)_{H^1(-1,1)} + \left(\int_{-1}^1 u(x)\textnormal{d} x\right)\left(\int_{-1}^1 v(x)\textnormal{d} x\right). \end{equation*} \end{definition} Within the next subsections, we will prove~\eqref{eq:whattolfa}. This will be done by a rigorous version of Fourier analysis. Fourier analysis is a well-known tool for analyzing convergence properties of numerical methods, cf. the work by A. Brandt, like~\cite{Brandt:1977}, and many others. It provides a framework to determine sharp bounds for the convergence rates of multigrid methods and other iterative solvers for problems arising from partial differential equations. This is different to classical analysis, which typically yields qualitative statements only. For a detailed introduction into Fourier analysis, see, e.g.,~\cite{Trottenberg:2001}. Recently, it has also been applied in the area of Isogeometric Analysis, cf.~\cite{Garoni:2014}. Typically, Fourier analysis is done under simplifying assumptions, like assuming uniform grids and neglecting the boundary. In this case, one refers to \emph{local} Fourier analysis (or local mode analysis). This analysis can be understood as a heuristic method to study methods of interest. In a recent work, cf.~\cite{Garoni:2014}, it was understood also as a rigorous statement for a limit case. We, however, are interested in a completely rigorous analysis. As we restrict ourselves to periodic spline spaces, the Fourier modes are the exact eigenvectors of the matrices of interest, which will allow us to diagonalize these matrices using a similarity transformation. Based on such a diagonalization, we will be able to prove~\eqref{eq:whattolfa}. As a first step, we introduce a matrix-vector formulation of~\eqref{eq:whattolfa}. \subsection{A matrix-vector formulation of the estimate} Having fixed the B-spline like basis $\{\widehat{\varphi}_{p,h}^{(i)}\}_{i=0}^{\nh-1}$, we can write any function $u_{p,h}\in \widehat{S}_{p,h}(-1,1)$ as a linear combination of these basis functions: \begin{equation*} u_{p,h} = \sum_{i=0}^{\nh-1} u_{p,h}^{(i)} \widehat{\varphi}_{p,h}^{(i)}. \end{equation*} The coefficients $u_{p,h}^{(i)}$ can be collected in a coefficient vector: We define $\ul{u}_{p,h}:=(u_{p,h}^{(i)})_{i=0}^{\nh-1}$. So, the vector $\ul{u}_{p,h}$ is the representation of the function $u_{p,h}$ with respect to the B-spline like basis. Here and in what follows, we will always assume underlined quantities to be the basis representation of the corresponding function with respect to the basis $\{\widehat{\varphi}_{p,h}^{(i)}\}_{i=0}^{\nh}$. By plugging such a decomposition into the standard $L^2$-scalar product $(\cdot,\cdot)_{L^2(-1,1)}$, we obtain \begin{equation*} (u_{p,h},v_{p,h})_{L^2(-1,1)} = \sum_{i=0}^{\nh-1}\sum_{j=0}^{\nh-1}u_{p,h}^{(i)}\;v_{p,h}^{(j)} \; (\widehat{\varphi}_{p,h}^{(i)},\widehat{\varphi}_{p,h}^{(j)})_{L^2(-1,1)}. \end{equation*} As the grid is equidistant and the splines are periodic, we obtain that for all $i$ and $j$ the relation $(\widehat{\varphi}_{p,h}^{(i)},\widehat{\varphi}_{p,h}^{(j)})_{L^2(-1,1)}=m_{p,h}^{(i-j)}$ holds with coefficients $m_{p,h}^{(i)}:=(\widehat{\varphi}_{p,h}^{(i)},\widehat{\varphi}_{p,h}^{(0)})_{L^2(-1,1)}$. Those coefficients form a circulant matrix $M_{p,h}:=(m_{p,h}^{(i-j)})_{i=0,\ldots,\nh-1}^{j=0,\ldots,\nh-1}$, which is called the mass matrix. We immediately obtain \begin{equation*} (u_{p,h},v_{p,h})_{L^2(-1,1)} = (\ul{u}_{p,h},\ul{v}_{p,h})_{M_{p,h}} := \ul{v}_{p,h}^T M_{p,h}\ul{u}_{p,h} \end{equation*} and \begin{equation*} \|u_{p,h}\|_{L^2(-1,1)}^2 = \|\ul{u}_{p,h}\|_{M_{p,h}}^2 := \ul{u}_{p,h}^T M_{p,h} \ul{u}_{p,h}. \end{equation*} Having a look onto the support of the functions $\widehat{\varphi}_{p,h}^{(0)}$, we obtain that the bandwidth of the mass matrix is $2p+1$, i.e. $m_{p,h}^{(i-j)} = 0$ for all $i,j$ with $|i-j|>p$. Analogously to the definition of the mass matrix, we can introduce the stiffness matrix, representing the $H^1_{\circ}$-scalar product. The stiffness matrix is given by $K_{p,h}:=(k_{p,h}^{(i-j)})_{i=0,\ldots,\nh-1}^{j=0,\ldots,\nh-1}$, where the coefficients are given by \begin{equation*} k_{p,h}^{(i)} := \left( \widehat{\varphi}_{p,h}^{(i)}, \widehat{\varphi}_{p,h}^{(0)} \right)_{H^1_{\circ}(-1,1)}. \end{equation*} Since the basis functions $\widehat{\varphi}_{p,h}^{(i)}$ form a partition of unity on $\Omega=(-1,1)$, $\int_{-1}^1 \widehat{\varphi}_{p,h}^{(i)}(x) \textnormal{d} x=\hn$ and further \begin{equation}\label{eq:def:stiff} k_{p,h}^{(i)} = \left( \widehat{\varphi}_{p,h}^{(i)}, \widehat{\varphi}_{p,h}^{(0)} \right)_{H^1(-1,1)} + \hn^2. \end{equation} Note that for uniform knot vectors the identity \begin{equation*} \frac{\partial}{\partial x} \varphi_{p,h}^{(j)}(x) = \frac{1}{\hn}\left(\varphi_{p-1,h}^{(j-1)}(x)- \varphi_{p-1,h}^{(j)}(x)\right) \end{equation*} holds, see e.g. (5.36) in \cite{Schumaker:1981}. This statement directly carries over to the periodic splines using relation \eqref{eq:basis:varphi}, i.e., \begin{equation*} \frac{\partial}{\partial x} \widehat{\varphi}_{p,h}^{(j)}(x) = \frac{1}{\hn}\left(\widehat{\varphi}_{p-1,h}^{(j-1)}(x)- \widehat{\varphi}_{p-1,h}^{(j)}(x)\right) \end{equation*} also holds. By plugging this into~\eqref{eq:def:stiff}, the entries of the stiffness matrix can be derived directly using the entries of the mass matrix for splines of order $p-1$. Straight-forward calculations show that \begin{equation}\label{eq:k:decomp} K_{p,h} = D_{h} M_{p-1,h} D_{h}^T + E_{h}, \end{equation} where the gradient matrix $D_{h}:=(d_{h}^{(i-j)})_{i=0,\ldots,\nh-1}^{j=0,\ldots,\nh-1}$ is given by the coefficients \begin{equation*} d_{h}^{(i)} := \frac{1}{\hn} \left\{ \begin{array}{ll} 1 & \mbox{ for } i\in \nh \,\mathbb{Z} \\ -1 & \mbox{ for } i\in \nh \,\mathbb{Z}-1 \\ 0 & \mbox{ else} \end{array} \right., \end{equation*} the rank-one matrix $E_h$ is given by $E_h := \hn^2 \underline{\bf{1}}_{h} \underline{\bf{1}}_{h}^T$, where $\underline{\bf{1}}_{h}:=(1,\ldots,1)^T\in\mathbb{R}^{\nh}$ is a vector consisting only of ones, representing the constant function. Note that $D_h$, $E_h$ and, consequently, $K_h$ are also circulant matrices. To derive a matrix-vector formulation of \eqref{eq:whattolfa}, we have to introduce a matrix that represents the canonical embedding from $\widehat{S}_{p,h}(-1,1)$ into $\widehat{S}_{p,\frac{h}{2}}(-1,1)$. The following lemma is rather well-known in literature, cf. \cite{Chui:1992} equation (4.3.4), and can be easily shown by induction in $p$. \begin{lemma} For all $p\in\mathbb{N}$, all grid sizes $h$ and all $x\in\mathbb{R}$, \begin{equation}\nonumber \varphi_{p,h}^{(j)}(x) = 2^{-p} \sum_{l=0}^{p+1} \left(\begin{array}{c}p+1\\l\end{array}\right) \varphi_{p,\tfrac{h}{2}}^{(2j+l)}(x) \end{equation} is satisfied for all $j=-p,\ldots,\nh-p-1$. \end{lemma} This directly carries over to the periodic splines, i.e., we obtain \begin{equation}\label{eq:intergrid:per} \widehat{\varphi}_{p,h}^{(j)}(x) = 2^{-p} \sum_{l=0}^{p+1} \left(\begin{array}{c}p+1\\l\end{array}\right) \widehat{\varphi}_{p,\tfrac{h}{2}}^{(2j+l)}(x) = \sum_{i\in\mathbb{Z}} \underbrace{2^{-p}\left(\begin{array}{c}p+1\\ i-2j \end{array}\right)} _{\displaystyle p_{p,\tfrac{h}{2}}^{(i,j)}:=} \widehat{\varphi}_{p,\tfrac{h}{2}}^{(i)}(x). \end{equation} Here, we use equation \eqref{eq:basis:varphi} and that the binomial coefficient $\left(\begin{array}{c}a\\b \end{array}\right)$ vanishes for $b\not\in \{0,\ldots,a\}$. Again, we define the matrix $P_{p,\tfrac{h}{2}}:=(p_{p,\tfrac{h}{2}}^{(i,j)})_{i=0,\ldots, 2\nh-1}^{j=0,\ldots,\nh-1}$. Here and in what follows, we make use of $n_{\frac{h}{2}} = 2 \nh$. \begin{lemma}\label{lemma:decomp} The inequality~\eqref{eq:whattolfa} is equivalent to \begin{equation}\label{eq:whattolfa2} \|M_{p,\frac{\hn}{2}}^{1/2} (I-P_{p,\frac{\hn}{2}} K_{p,\hn}^{-1} P_{p,\frac{\hn}{2}}^T K_{p,\frac{\hn}{2}}) K_{p,\frac{\hn}{2}}^{-1/2}\| \le \frac{1}{\sqrt{2}} h , \end{equation} which is a consequence of the combination of \begin{align} & \|M_{p,\frac{\hn}{2}}^{1/2}M_{p-1,\frac{\hn}{2}}^{-1/2}\|\le 1\qquad \mbox{and}\label{eq:whattolfa3}\\ & \|M_{p-1,\frac{\hn}{2}}^{1/2} (I-P_{p,\frac{\hn}{2}} K_{p,\hn}^{-1} P_{p,\frac{\hn}{2}}^T K_{p,\frac{\hn}{2}}) K_{p,\frac{\hn}{2}}^{-1/2}\| \le \frac{1}{\sqrt{2}} h. \label{eq:whattolfa4} \end{align} \end{lemma} Here and in what follows, $\|\cdot\|$ is the Euclidean norm and the square root $A^{1/2}$ of a symmetric and positive definite matrix $A$ is that symmetric and positive definite matrix that satisfies $A^{1/2}A^{1/2} = A$. \begin{proof}{\em of Lemma~\ref{lemma:decomp}} Using the introduced matrices $K_{p,h}$ and $P_{p,\tfrac{h}{2}}$, we can rewrite~\eqref{eq:def:projection} for the choice $u := u_{p,\frac{h}{2}} \in \widehat{S}_{p,\frac{h}{2}}$ in matrix-vector form as \begin{equation}\nonumber (P_{p,\frac{\hn}{2}} \ul{u}_{p,h}, P_{p,\frac{\hn}{2}} \ul{v}_{p,h})_{K_{p,\frac{\hn}{2}}} = (\ul{u}_{p,\frac{h}{2}}, P_{p,\frac{\hn}{2}}\ul{v}_{p,h})_{K_{p,\frac{\hn}{2}}}, \end{equation} which is equivalent to \begin{equation}\nonumber P_{p,\frac{\hn}{2}}^TK_{p,\frac{\hn}{2}}P_{p,\frac{\hn}{2}} \ul{u}_{p,h} = P_{p,\frac{\hn}{2}}^T K_{p,\frac{\hn}{2}} \ul{u}_{p,\frac{h}{2}}. \end{equation} This yields, using the Galerkin principle ($P_{p,\frac{\hn}{2}}^TK_{p,\frac{\hn}{2}}P_{p,\frac{\hn}{2}}=K_{p,\hn}$), that the coarse-grid approximation $\ul{u}_{p,h}$ is given by \begin{equation}\nonumber \ul{u}_{p,h} = K_{p,\hn}^{-1} P_{p,\frac{\hn}{2}}^T K_{p,\frac{\hn}{2}} \ul{u}_{p,\frac{h}{2}}. \end{equation} By plugging this into~\eqref{eq:whattolfa}, we see that we have to show \begin{equation*} \|(I-P_{p,\frac{\hn}{2}} K_{p,\hn}^{-1} P_{p,\frac{\hn}{2}}^T K_{p,\frac{\hn}{2}}) \ul{u}_{p,\frac{h}{2}}\|_{M_{p,\frac{\hn}{2}}} \le \frac{1}{\sqrt{2}} \hn \|\ul{u}_{p,\frac{h}{2}}\|_{K_{p,\frac{\hn}{2}}} \end{equation*} for all $\ul{u}_{p,\frac{h}{2}} \in \mathbb{R}^{2\nh}$. By rewriting this using a standard matrix norm, we obtain \eqref{eq:whattolfa2}. Using the semi-multiplicativity of matrix norms, we obtain that~\eqref{eq:whattolfa2} is a consequence of \eqref{eq:whattolfa3} and \eqref{eq:whattolfa4}. \qed\end{proof} Note that the stiffness matrix for some degree $p$ depends implicitly on the mass matrix for the degree $p-1$. So, analyzing~\eqref{eq:whattolfa4} is more convenient than analyzing~\eqref{eq:whattolfa2} as the inequality~\eqref{eq:whattolfa4} depends just on the one mass matrix $M_{p-1,\frac{\hn}{2}}$, whereas~\eqref{eq:whattolfa2} depends on two mass matrices: $M_{p-1,\frac{\hn}{2}}$ and $M_{p,\frac{\hn}{2}}$. We will show~\eqref{eq:whattolfa3} in the next subsection and~\eqref{eq:whattolfa4} in the remainder of this section. \subsection{A lemma relating the mass matrices for different polynomial degrees} The estimate~\eqref{eq:whattolfa3} is a direct consequence of the following lemma. \begin{lemma}\label{lem:mass} For all $p\in \mathbb{N}$, grid sizes $h$ and vectors $\underline{u}_h \in \mathbb{R}^{\nh}$, the inequality \begin{equation}\nonumbe \|\underline{u}_h\|_{M_{p,h}} \le \|\underline{u}_h\|_{M_{p-1,h}} \end{equation} is satisfied. \end{lemma} \begin{proof} First we observe that the convolution formula for cardinal B-splines, cf. equation~(13) in~\cite{Garoni:2014}, can be carried over to the functions $\widehat{\varphi}_{p,h}^{(i)}$, i.e., that \begin{equation}\label{eq:rel1} \widehat{\varphi}_{p,h}^{(i)}(x) = h^{-1} \int_0^{h} \widehat{\varphi}_{p-1,h}^{(i)}(x-t) \textnormal{d} t \end{equation} holds. Let $\underline{u}_h=(u_h^{(i)})_{i=0}^{\nh-1}$. Then, using~\eqref{eq:rel1}, we have that \begin{align*} \|\underline{u}_h\|_{M_{p,h}}^2 & = \int_{-1}^1 \left( \sum_{i=0}^{\nh-1} u_h^{(i)} \widehat{\varphi}_{p,h}^{(i)}(x) \right)^2 \textnormal{d} x \\ & = \int_{-1}^1 \left( \sum_{i=0}^{\nh-1} u_h^{(i)} h^{-1} \int_0^h \widehat{\varphi}_{p-1,h}^{(i)}(x-t) \textnormal{d} t \right)^2 \textnormal{d} x \\ & = h^{-2} \int_{-1}^1 \left( \int_0^{h} \left( \sum_{i=0}^{\nh-1} u_h^{(i)} \widehat{\varphi}_{p-1,h}^{(i)}(x-t) \right) \textnormal{d} t \right)^2 \textnormal{d} x \\ & = h^{-2} \int_{-1}^1 \left( \int_0^{h} 1 \, s(x-t) \textnormal{d} t \right)^2 \textnormal{d} x \end{align*} holds, where $s(x):= \sum_{i=0}^{\nh-1} u_h^{(i)} \widehat{\varphi}_{p-1,h}^{(i)}(x-t)$. Now, we apply the Cauchy-Schwarz inequality to the inner integral and obtain \begin{align*} \|\underline{u}_h\|_{M_{p,h}}^2 & \leq h^{-2} \int_{-1}^1 \left( \int_0^{h} 1^2 \textnormal{d} t \right)\left( \int_0^{h}s^2(x-t) \textnormal{d} t \right) \textnormal{d} x \\ & = h^{-1} \int_{-1}^1 \int_0^{h} s^2(x-t) \textnormal{d} t\,\textnormal{d} x = h^{-1} \int_0^{h} \int_{-1}^1 s^2(x-t) \textnormal{d} x\,\textnormal{d} t. \end{align*} Observe that due to periodicity, $\int_{-1}^1 s^2(x-t) \textnormal{d} x = \int_{-1}^1 s^2(\xi) \textnormal{d} \xi$ for all $t\in[0,h]$, which implies \begin{align*} \|\underline{u}_h\|_{M_{p,h}}^2 & \le h^{-1} \int_0^{h} \int_{-1}^1 s^2(\xi) \textnormal{d} \xi\,\textnormal{d} t = h^{-1} \left(\int_{0}^h 1\textnormal{d} t\right)\left( \int_{-1}^1 s^2(\xi) \textnormal{d} \xi \right)\\ & = \int_{-1}^1 \left( \sum_{i=0}^{\nh-1} u_h^{(i)} \widehat{\varphi}_{p-1,h}^{(i)}(\xi) \right)^2 \textnormal{d} \xi = \|\underline{u}_h\|_{M_{p-1,h}}^2, \end{align*} which finishes the proof. \qed\end{proof} \subsection{Symbols of mass matrix and stiffness matrix}\label{subsec:symbols} As the matrices $M_{p,h}$ and $K_{p,h}$ are circulant matrices, we can analyze them using Fourier analysis. So, we consider the Fourier vectors \begin{equation*} \underline{f}_{h,j}:=(\textnormal{\bf e}^{ 2ij h\pi \textnormal{\bf i}})_{i=0}^{\nh-1} \qquad \mbox{ for } j = 0,\ldots, \nh-1, \end{equation*} where $\textnormal{\bf i}$ is the imaginary unit. We observe (using that the bandwith of the mass matrix is $2p+1$) that \begin{align*} ( M_{p,h} \underline{f}_{h,j} )_i &= \sum_{l=-p}^p m_{p,h}^{(l)} \textnormal{\bf e}^{2(i+l)jh\pi \textnormal{\bf i} } = \sum_{l=-p}^p m_{p,h}^{(l)} \textnormal{\bf e}^{2ljh\pi \textnormal{\bf i} } \textnormal{\bf e}^{2ijh\pi \textnormal{\bf i} }\\ &= \underbr{\sum_{l=-p}^p m_{p,h}^{(l)} \textnormal{\bf e}^{2ljh\pi \textnormal{\bf i} }}{ \widehat{m}_{p,h}^{(j)}:= } ( \underline{f}_{h,j} )_i \end{align*} for all $i=0,\ldots,\nh-1$ and $j=0,\ldots,\nh-1$ and consequently \begin{equation}\nonumbe M_{p,h} \underline{f}_{h,j} = \widehat{m}_{p,h}^{(j)} \underline{f}_{h,j} \end{equation} is satisfied for all $j=0,\ldots,\nh-1$, i.e., that $\underline{f}_{h,j}$ is an eigenvector of $M_{p,h}$ with corresponding eigenvalue $\widehat{m}_{p,h}^{(j)}$. As we have identified $\nh$ different eigenvalues, the corresponding eigenvectors define a basis of $\mathbb{R}^{\nh}$. Therefore, the matrix $\mathbb{F}_{h}$, obtained by collecting the vectors $\underline{f}_{h,j}$, i.e., \begin{equation*} \mathbb{F}_{h} := \left( \begin{array}{cccc} \underline{f}_{h,0} & \underline{f}_{h,1} & \cdots & \underline{f}_{h,\nh-1} \end{array}\right) = (\textnormal{\bf e}^{ 2ij h\pi \textnormal{\bf i}})_{i=0,\ldots,\nh-1}^{j=0,\ldots,\nh-1}, \end{equation*} is a non-singular matrix. As~$\mathbb{F}_{h}$ is the matrix built from the eigenvectors, it diagonalizes the matrix $M_{p,h}$, i.e., \begin{equation}\label{eq:mhat} \mathbb{F}_{h}^{-1} M_{p,h} \mathbb{F}_{h} = \widehat{M}_{p,h}, \end{equation} where $\widehat{M}_{p,h}:=\textnormal{diag}(\widehat{m}_{p,h}^{(0)},\ldots,\widehat{m}_{p,h}^{(\nh-1)})$. Analogously, we obtain \begin{equation}\label{eq:dhat} \mathbb{F}_{h}^{-1} D_{h} \mathbb{F}_{h} = \widehat{D}_{h}, \end{equation} where $\widehat{D}_{h}:=\textnormal{diag}(\widehat{d}_{h}^{(0)},\ldots,\widehat{d}_{h}^{(\nh-1)})$ with \begin{equation}\label{eq:dhatcoef} \widehat{d}_{h}^{(j)}:=\hn^{-1}(1-\textnormal{\bf e}^{2jh\pi\textnormal{\bf i}}). \end{equation} Using the same construction we obtain that further \begin{equation}\label{eq:dhat:star} \mathbb{F}_{h}^{-1} D_{h}^T \mathbb{F}_{h} = \widehat{D}_{h}^*. \end{equation} With $\widehat{D}_{h}^*$ we denote the adjoint (the conjugate transpose) of the matrix $\widehat{D}_{h}$. Note that $E_h=\hn^2 \underline{\bf{1}}_h \underline{\bf{1}}_h^T$ is a circulant matrix with rank $1$. The only non-zero eigenvalue is $\hn$, with corresponding eigenvector $\underline{\bf{1}}_h = \underline{f}_{h, 0}$. So, we obtain \begin{equation}\label{eq:ehat} \mathbb{F}_{h}^{-1} E_h \mathbb{F}_{h} = \widehat{E}_h \end{equation} where $\widehat{E}_h:=\textnormal{diag}(\widehat{e}_{h}^{(0)},\ldots,\widehat{e}_{h}^{(\nh-1)})$ with \begin{equation}\label{eq:ehatcoef} \widehat{e}_{h}^{(j)}:=\left\{\begin{array}{ll}\hn &\mbox{ for } j=0\\0&\mbox{ otherwise.}\end{array}\right. \end{equation} So, we can determine, $\widehat{K}_{h}$, the symbol of the stiffness matrix. Using~\eqref{eq:k:decomp}, \eqref{eq:mhat}, \eqref{eq:dhat}, \eqref{eq:dhat:star} and~\eqref{eq:ehat}, we obtain that \begin{equation}\label{eq:khat} \mathbb{F}_{h}^{-1} K_{p,h} \mathbb{F}_{h}= \widehat{K}_{h}, \end{equation} where $\widehat{K}_{h}:=\textnormal{diag}(\widehat{k}_{p,h}^{(0)},\ldots,\widehat{k}_{p,h}^{(\nh-1)})$ with \begin{equation}\label{eq:khatcoef} \widehat{k}_{p,h}^{(j)} := \widehat{d}_{h}^{(j)}\widehat{m}_{p-1,h}^{(j)} (\widehat{d}_{h}^{(j)})^*+\widehat{e}_{h}^{(j)}. \end{equation} \subsection{Symbol of the intergrid transfer}\label{subsec:symbols2} The following lemma characterizes the symbol of the intergrid transfer. \begin{lemma}\label{lem:phat} We have \begin{equation}\label{eq:phat} \mathbb{F}_{\frac{\hn}{2}}^{-1} P_{p,\frac{\hn}{2}}\mathbb{F}_{\hn} = \widehat{P}_{p,\frac{\hn}{2}}, \end{equation} where $\widehat{P}_{p,\frac{\hn}{2}}:=(\widehat{p}_{p,\frac{\hn}{2}}^{(i,j)})_{i=0,\ldots,2\nh-1}^{j=0,\ldots,\nh-1}$ with \begin{equation}\label{eq:phatcoef} \widehat{p}_{p,\frac{\hn}{2}}^{(i,j)} := 2^{-p-1}\left\{ \begin{array}{ll} \left(1+\textnormal{\bf e}^{- 2i\frac{h}{2}\pi\textnormal{\bf i}} \right)^{p+1} & \mbox{ for } i - j \in \{0,\nh\}\\ 0 & \mbox{ otherwise } \end{array} \right. \end{equation} for all $i=0,\ldots,2\nh-1$ and all $j=0,\ldots,\nh-1$. \end{lemma} \begin{proof} The equation~\eqref{eq:phat} is equivalent to $P_{\frac{\hn}{2}}\mathbb{F}_{\hn} = \mathbb{F}_{\frac{\hn}{2}} \widehat{P}_{\frac{\hn}{2}}$. We obtain using~\eqref{eq:intergrid:per} and the definition of $\mathbb{F}_{\hn}$ for any unit vector $\underline{\textbf{I}}_{\hn}^{(j)}$ with $j=0,\ldots,\nh-1$ that \begin{align*} &P_{\frac{\hn}{2}}\mathbb{F}_{\hn}\underline{\textbf{I}}_{\hn}^{(j)} = P_{\frac{\hn}{2}}\underline{f}_{\hn,j} = 2^{-p} \left( \sum_{r\in \mathbb{Z} } \left(\begin{array}{c}p+1\\ i-2r \end{array}\right) \textnormal{\bf e}^{ 2jr h \pi \textnormal{\bf i}} \right)_{i=0}^{2\nh-1}. \end{align*} Because $\tfrac12(1+\textnormal{\bf e}^{ t \pi \textnormal{\bf i}})$ takes the value $0$ for $t$ being odd and $1$ for $t$ being even, we can substitute $r$ by $2t$ and obtain \begin{align*} &P_{\frac{\hn}{2}}\mathbb{F}_{\hn}\underline{\textbf{I}}_{\hn}^{(j)} = 2^{-p-1} \left( \sum_{t\in \mathbb{Z} } \left(\begin{array}{c}p+1\\ i-t \end{array}\right) \textnormal{\bf e}^{ 2 jt \frac{h}{2} \pi \textnormal{\bf i}} ( 1+\textnormal{\bf e}^{ t \pi \textnormal{\bf i}})\right)_{i=0}^{2\nh-1} \\ &\quad = 2^{-p-1} \left( \sum_{k\in \mathbb{Z} } \left(\begin{array}{c}p+1\\ k \end{array}\right) \textnormal{\bf e}^{ 2 j(i-k) \frac{h}{2} \pi \textnormal{\bf i}} ( 1+\textnormal{\bf e}^{ (i-k) \pi \textnormal{\bf i}})\right)_{i=0}^{2\nh-1} \\ &\quad = 2^{-p-1} \sum_{k\in \mathbb{Z} } \left(\begin{array}{c}p+1\\k\end{array}\right) \left( \textnormal{\bf e}^{- 2 jk \frac{h}{2} \pi\textnormal{\bf i}} \underline{f}_{\frac{\hn}{2},j} + \textnormal{\bf e}^{-2(j+\nh)k\frac{h}{2} \pi\textnormal{\bf i}} \underline{f}_{\frac{\hn}{2},j+\nh} \right) \\ &\quad = 2^{-p-1}\left(1+\textnormal{\bf e}^{- 2j\frac{h}{2}\pi\textnormal{\bf i}} \right)^{p+1} \underline{f}_{\frac{\hn}{2},j} + 2^{-p-1}\left(1+\textnormal{\bf e}^{-2(j+\nh)\frac{h}{2}\pi\textnormal{\bf i}} \right)^{p+1} \underline{f}_{\frac{\hn}{2},j+\nh}. \end{align*} This shows that the $j$-th column of $P_{\frac{\hn}{2}}\mathbb{F}_{\hn}$ is just the combination of two columns of $\mathbb{F}_{\frac{\hn}{2}}$. Therefore, the matrix $\widehat{P}_{\frac{\hn}{2}}$ has just two non-zero entries, in the $j$-th row: those which we have claimed in~\eqref{eq:phatcoef}. \qed\end{proof} For determining the symbol of $P_{p,\frac{h}{2}}^T$, we observe as follows. As the Fourier modes $\underline{f}_{\hn,j}$ are pairwise orthogonal, and $\underline{f}_{\hn,j}^*\underline{f}_{\hn,j} = \nh$, we immediately obtain $\mathbb{F}_{\hn}^*\mathbb{F}_{\hn} = \nh I$ and, consequently, $\mathbb{F}_{\hn}^{-1} = \hn \mathbb{F}_{\hn}^*$. So, we obtain using~\eqref{eq:phat} that \begin{equation}\label{eq:phattranspose} \mathbb{F}_{h}^{-1} P_{p,\frac{h}{2}}^T \mathbb{F}_{\frac{h}{2}} = ( \mathbb{F}_{\frac{h}{2}}^* P_{p,\frac{h}{2}} \mathbb{F}_{h}^{-*} )^* = ( 2 \mathbb{F}_{\frac{h}{2}}^{-1} P_{p,\frac{h}{2}} \mathbb{F}_{h} )^* = 2 \widehat{P}_{p,\frac{h}{2}}^*. \end{equation} \subsection{Some statements on the symbol of the mass matrix} A closed form for the symbol of the mass matrix is not known. Within this subsection we will show a few statements characterizing the symbol, which we will need later on. Due to \cite{Chui:1992,Wang:2010}, we have \begin{equation}\label{eq:our:mass:formula} m_{p,\hn}^{(j)} = \hn \frac{ E(2 p + 1, p + j)}{(2 p + 1)!}, \end{equation} where $j\in\{-p,\ldots,p\}$. Here, $E(n,k)$ are the Eulerian numbers, which satisfy the recurrence relation \begin{equation*} E(n, k) = (n - k) E(n - 1, k - 1) + (k + 1) E(n - 1, k) \end{equation*} and the initial condition \begin{equation*} E(0, j) =\left\{ \begin{array}{ll} 1 & \mbox{for } j= 0\\ 0 & \mbox{for } j\not= 0 \end{array} \right.. \end{equation*} A similar result was also stated in~\cite{Garoni:2014}. There, the entries of the mass matrix, i.e., the $L^2$-products of two B-splines of order $p$ have been shown to be equal to the function value of one B-spline of order $p+1$. Using the recurrence relation~\eqref{eq:recur:bspline}, one obtains that the result in~\cite{Garoni:2014} is equivalent to~\eqref{eq:our:mass:formula}. As $m_{p,\hn}^{(j)}=m_{p,\hn}^{(-j)}$ and $\textnormal{\bf e}^{\theta \textnormal{\bf i}} +\textnormal{\bf e}^{-\theta \textnormal{\bf i}} = 2 \cos\theta$, we obtain \begin{equation}\nonumbe \widehat{m}_{p,\hn}^{(j)} = \hn \sum_{l=-p}^p\frac{ E(2 p + 1, p + l)}{(2p + 1)!}\cos(2ljh\pi). \end{equation} The symbol is better characterized by the following lemma. \begin{lemma}\label{lem:mass:estim} The following two statements hold: \begin{itemize} \item $\widehat{m}_{p,\hn}^{(j)}> 0$ for all $j=0,\ldots,\nh-1$ and \item $\widehat{m}_{p,\hn}^{(j)}\le \widehat{m}_{p,\hn}^{(k)}$ for all $j,k=0,\ldots,\nh-1$ with $\cos(2jh\pi) \le \cos(2kh\pi)$. \end{itemize} \end{lemma} \begin{proof} For $c\in[0,2]$, we define \begin{equation*} g_{p}(c) := \sum_{l=-p}^p\frac{ E(2 p + 1, p + l)}{(2p + 1)!}\cos(l\arccos (c-1)) \end{equation*} and observe $g_{p}(c) = \hn^{-1} \widehat{m}_{p,\hn}^{(\eta(c))}$, where $\eta(c):=\frac{1}{2h\pi}\arccos (c-1)$. The statement of the lemma is now equivalent to the combination of the following two statements: \begin{itemize} \item $\hn^{-1}\widehat{m}_{p,\hn}^{(\eta(0))}=g_{p}(0)>0$ and \item $\hn^{-1}\widehat{m}_{p,\hn}^{(\eta(c))}=g_{p}(c)$ is monotonically increasing for $c>0$. \end{itemize} Since we can express $\cos(l\arccos (c-1))$ as the $l$-th Chebyshev polynomial, $g_{p}$ is a polynomial function in $c$. Using the recurrence relation for the Eulerian numbers, we can derive the following recurrence formula for $g_{p}$: \begin{align*} g_{p}(c)=\frac{1+c p}{1+2 p} g_{p-1}(c)+\frac{(2-c) (1+c (2 p-1))}{p (1+2 p)} g_{p-1}'(c)+\frac{(c-2)^2 c}{p (1+2 p)} g_{p-1}''(c). \end{align*} We can make an ansatz \begin{equation*} g_{p}(c) = \sum_{j=0}^p a_{p,j} c^j, \end{equation*} where we use $0^0 = 1$, and derive the recurrence formula \begin{equation*} a_{p,j}=\underbr{\frac{(1-j+p)^2}{p+2 p^2}}{A_{p,j}:=} a_{p-1,j-1} +\underbr{\frac{4j (p-j)+j+p}{p+2 p^2}}{B_{p,j}:=} a_{p-1,j} +\underbr{\frac{2+6 j+4 j^2}{p+2 p^2}}{C_{p,j}:=} a_{p-1,j+1} \end{equation*} for the coefficients $a_{p,j}$. For $p=1$, we obtain \begin{equation*} a_{1,j} = \left\{ \begin{array}{ll} \tfrac13 & \mbox{ for $j\in\{0,1\}$ } \\ 0 & \mbox{ otherwise.} \end{array} \right. \end{equation*} As $A_{p,j}>0$, $B_{p,j}>0$ and $C_{p,j}>0$ for $0\le j \le p$, one can show using induction in $p$ that for all $p\ge 1$: \begin{equation*} \left\{ \begin{array}{ll} a_{p,j} > 0 & \mbox{ for $j\in\{0,1,\ldots,p\}$ } \\ a_{p,j} = 0 & \mbox{ otherwise.} \end{array} \right. \end{equation*} This immediately implies that $g_{p}(0)>0$ and that $g_{p}(c)$ is monotonically increasing for $c>0$, which concludes the proof.\qed \end{proof} \subsection{An estimate for the projection operator} Now, we are able to prove the following lemma. \begin{lemma}\label{lem:lfa0} The inequality~\eqref{eq:whattolfa4} holds. \end{lemma} \begin{proof} The inequality~\eqref{eq:whattolfa4} is equivalent to \begin{equation*} \underbr{\hn^{-1} \|M_{p-1,\frac{h}{2}}^{1/2}(I-P_{p,\frac{h}{2}} K_{p,h}^{-1} P_{p,\frac{h}{2}}^T K_{p,\frac{h}{2}}) K_{p,\frac{h}{2}}^{-1/2}\|}{q:=} \le \frac{1}{\sqrt{2}} . \end{equation*} Using Galerkin orthogonality, we obtain $K_{p,h} = P_{p,\frac{h}{2}}^T K_{p,\frac{h}{2}}P_{p,\frac{h}{2}}$. Note that $\mathcal{H}:=I-P_{p,\frac{h}{2}} K_{p,h}^{-1} P_{p,\frac{h}{2}}^T K_{p,\frac{h}{2}}$ is a projection operator, so $\mathcal{H}\mathcal{H} = \mathcal{H}$. Moreover, observe that $\mathcal{H} K_{p,\frac{h}{2}}^{-1} = K_{p,\frac{h}{2}}^{-1}\mathcal{H}^T$. Using these identities and $\|W\|^2 = \rho(WW^T)$, where $\rho$ denotes the spectral radius, we obtain \begin{align*}\nonumber q^2 &= \hn^{-2}\rho( M_{p-1,\frac{h}{2}}^{-1/2}\mathcal{H} K_{p,\frac{h}{2}}^{-1}\mathcal{H}^T M_{p-1,\frac{h}{2}}^{-1/2}) = \hn^{-2}\rho( K_{p,\frac{h}{2}}^{-1}M_{p-1,\frac{h}{2}}^{-1}\mathcal{H} ) \\&= \hn^{-2}\rho( K_{p,\frac{h}{2}}^{-1}M_{p-1,\frac{h}{2}} (I-P_{p,\frac{h}{2}} (P_{p,\frac{h}{2}}^TK_{p,\frac{h}{2}}P_{p,\frac{h}{2}})^{-1} P_{p,\frac{h}{2}}^T K_{p,\frac{h}{2}}) ). \end{align*} Using~\eqref{eq:mhat}, \eqref{eq:dhat}, \eqref{eq:dhat:star}, \eqref{eq:khat}, \eqref{eq:phat} and~\eqref{eq:phattranspose}, we obtain further \begin{align*}\nonumber q^2 = \hn^{-2}\rho( \underbr{\widehat{K}_{p,\frac{h}{2}}^{-1}\widehat{M}_{p-1,\frac{h}{2}} (I-2 \widehat{P}_{p,\frac{h}{2}} (2\widehat{P}_{p,\frac{h}{2}}^*\widehat{K}_{p,\frac{h}{2}}\widehat{P}_{p,\frac{h}{2}})^{-1} \widehat{P}_{p,\frac{h}{2}}^* \widehat{K}_{p,\frac{h}{2}}) }{ \widehat{T}_{p,\frac{h}{2}}:=} ). \end{align*} Lemma~\ref{lem:mass:estim} states that all digonal entries of $\widehat{M}_{p-1,\frac{h}{2}}$ are non-zero. It is straight-forward to see that also the diagonal entries of $\widehat{K}_{p,\frac{h}{2}}$ and $\widehat{K}_{p,h}=\widehat{P}_{p,\frac{h}{2}}^*\widehat{K}_{p,\frac{h}{2}}\widehat{P}_{p,\frac{h}{2}}$ are non-zero. So, $\widehat{T}_{p,\frac{h}{2}}$ is well-defined. Recall that Lemma~\ref{lem:phat} states that the matrix $\widehat{P}_{p,\frac{h}{2}}=(\widehat{p}_{p,\frac{h}{2}}^{(i,j)})_{i=0,\ldots,2\nh-1}^{j=0,\ldots,\nh-1}$ has a block-structure, given by \begin{equation*} \widehat{p}_{p,\frac{h}{2}}^{(i,j)} = 0 \mbox{ for all } i-j \not\in\{0,\nh\}. \end{equation*} Therefore and because the matrices $\widehat{M}_{p-1,\frac{h}{2}}$ and $\widehat{K}_{p,\frac{h}{2}}$ are diagonal, the matrix $\widehat{T}_{p,\frac{h}{2}} = (\widehat{t}_{p,\frac{h}{2}}^{(i,j)})_{i=0,\ldots,2\nh-1}^{j=0,\ldots,2\nh-1}$ has a block-structure, given by \begin{equation*} \widehat{t}_{p,\frac{h}{2}}^{(i,j)} = 0 \mbox{ for all } i-j \not\in\{-\nh,0,\nh\}. \end{equation*} By reordering the coefficients of the matrix $\widehat{T}_{p,\frac{h}{2}}$, we obtain a block-diagonal matrix with blocks \begin{equation*} \mathcal{T}_{p,\frac{h}{2}}^{(l)} = \left(\begin{array}{cc} \widehat{t}_{p,\frac{h}{2}}^{(l,l)} & \widehat{t}_{p,\frac{h}{2}}^{(l,l+\nh)} \\ \widehat{t}_{p,\frac{h}{2}}^{(l+\nh,l)} & \widehat{t}_{p,\frac{h}{2}}^{(l+\nh,l+\nh)} \end{array}\right). \end{equation*} As this block-diagonal matrix is spectrally equivalent to $\widehat{T}_{p,\frac{h}{2}}$ and the spectral radius of a block-diagonal matrix is just the maximum over the spectral radii of the blocks, we obtain \begin{equation*} q^2 = \rho(\widehat{T}_{p,\frac{h}{2}}) = \max_{l=0,\ldots,\nh-1} \rho( \mathcal{T}_{p,\frac{h}{2}}^{(l)} ). \end{equation*} So, in the following, we derive the spectral radius of $\mathcal{T}_{p,\frac{h}{2}}^{(l)}$ for any particular $l$. Straight-forward computation yields that for $l\in\{0,\ldots,\nh-1\}$, $i\in\{l,l+\nh\}$ and $j\in\{l,l+\nh\}$, we have \begin{equation}\label{eq:xxxy} \widehat{t}_{p,\frac{h}{2}}^{(i,j)} = \frac{\widehat{m}_{p-1,\frac{h}{2}}^{(i)}}{\widehat{k}_{p,\frac{h}{2}}^{(i)}} \left (\delta_{i,j} - \frac{\widehat{p}_{p,\frac{h}{2}}^{(i,l)}(\widehat{p}_{p,\frac{h}{2}}^{(j,l)})^* }{\sum_{r=0}^1 % (\widehat{p}_{p,\frac{h}{2}}^{(l+r\nh,l)})^* \widehat{k}_{p,\frac{h}{2}}^{(l+r\nh)} \widehat{p}_{p,\frac{h}{2}}^{(l+r\nh,l)} % } \widehat{k}_{p,\frac{h}{2}}^{(j)} \right), \end{equation} where $\delta_{i,j}$ is the Kronecker-delta, i.e., $\delta_{i,j}=1$ for $i=j$ and $\delta_{i,j}=0$ for $i\not=j$. Now, consider \emph{case A}: $l\in\{1,\ldots,\nh-1\}$. Here, we plug the values of $\widehat{k}_{p,\frac{h}{2}}^{(j)}$, $\widehat{d}_{\frac{h}{2}}^{(j)}$, $\widehat{e}_{\frac{h}{2}}^{(j)}$ (which takes the value $0$ for $j\in\{l,l+\nh\}$), $\widehat{p}_{p,\frac{h}{2}}^{(i,j)}$, as given by \eqref{eq:khatcoef}, \eqref{eq:dhatcoef}, \eqref{eq:ehatcoef} and~\eqref{eq:phatcoef}, into~\eqref{eq:xxxy} and substitute $\widehat{m}_{p-1,\frac{h}{2}}^{(l+n_h)}$ by $\xi \widehat{m}_{p-1,\frac{h}{2}}^{(l)}$. Doing so, the term $\widehat{m}_{p-1,\frac{h}{2}}^{(l)}$ cancels out and we obtain by straight-forward computation \begin{equation}\nonumber \mathcal{T}_{p,\frac{h}{2}}^{(l)} = \frac{1}{\delta} \left( \begin{array}{c} -z(1 - z)^{p-3} \xi \\ z(1 + z)^{p-3} \end{array} \right) \left( \begin{array}{c} (-1)^{p} (1 - z)^{p+1} \\ (1 + z)^{p+1} \end{array} \right)^T, \end{equation} where $\delta:=(1 + z)^{2 p} + (-1)^p (1 - z)^{2 p} \xi$ and $z:=\textnormal{\bf e}^{ 2 l \frac{h}{2} \pi\textnormal{\bf i}}$. Note that the computations are not a problem, as none of the symbols (except $\widehat{e}_{\frac{h}{2}}^{(j)}$) takes the value $0$ for case A. Moreover, for case~A we have that $z\not\in\{-1,1\}$. Observe that $\mathcal{T}_{p,\frac{h}{2}}^{(l)}$ has rank $1$. Therefore, its spectral radius equals its trace, so we obtain by straight-forward computations that \begin{align*} \rho( \mathcal{T}_{p,\frac{h}{2}}^{(l)} ) &= \frac{ z (1+z)^{2p-2} - (-1)^p z (1-z)^{2p-2} \xi }{ (1+z)^{2p} + (-1)^p (1-z)^{2p} \xi } \\ &= \frac{ z^{-p+1} (1+2z+z^2)^{p-1} - (-1)^p z^{-p+1} (1-2z+z^2)^{p-1} \xi }{ z^{-p} (1+2z+z^2)^{p} + (-1)^p z^{-p} (1-2z+z^2)^{p} \xi } \\ &= \frac{ (z^{-1}+2+z)^{p-1} - (-1)^p (z^{-1}-2+z)^{p-1} \xi }{ (z^{-1}+2+z)^{p} + (-1)^p (z^{-1}-2+z)^{p} \xi } \\ &= \frac{ (2+2 c)^{p-1} - (-1)^p (-2+2 c)^{p-1} \xi }{ (2+2 c)^{p} + (-1)^p (-2+2 c)^{p} \xi } = \underbr{\frac{ (1+ c)^{p-1} + (1- c)^{p-1} \xi }{ 2 ( (1+ c)^{p} + (1- c)^{p} \xi ) }}{\Psi_p(c,\xi):=} \end{align*} holds, where $c:=\cos(2 l \frac{h}{2} \pi)$ and, as defined above, $\xi=\widehat{m}_{p-1,\frac{h}{2}}^{(l+\nh)}/\widehat{m}_{p-1,\frac{h}{2}}^{(l)}$. Note that $c\in(-1,1)$ holds as we have restricted ourselves to $l\in\{1,\ldots,\nh-1\}$. Observe that Lemma~\ref{lem:mass:estim} implies that $\xi>0$. Now, consider two cases: \begin{itemize} \item If $c=\cos(2l\frac{h}{2}\pi)> 0$, then $\cos(2(l+\nh)\frac{h}{2}\pi)=\cos(2l\frac{h}{2}\pi+\pi)\le 0$. For this case Lemma~\ref{lem:mass:estim} states that $\widehat{m}_{p-1,\frac{h}{2}}^{(l+\nh)}\le\widehat{m}_{p-1,\frac{h}{2}}^{(l)}$, so $\xi\le 1$ holds. \item Analogously, $\xi\ge 1$ holds if $c\le0$. \end{itemize} To finalize the proof of case~A, we need to show \begin{equation*} \Psi_p\left(\cos\left(2l\frac{h}{2}\pi \right),\frac{\widehat{m}_{p-1,\frac{h}{2}}^{(l+\nh)}}{\widehat{m}_{p-1,\frac{h}{2}}^{(l)}} \right) \le \frac{1}{2} \end{equation*} for all $l=1,\ldots,\nh-1$. It suffices to show \begin{equation}\label{eq:cond1} \Psi_p(c,\xi) \le \frac{1}{2} \end{equation} for all $(c,\xi) \in [0,1)\times(0,1]\cup (-1,0] \times[1,\infty)$ and all $p \in \mathbb{N}$, i.e., to show the inequality for the whole range of $c$ and $\xi$, ignoring their dependence on $l$. As a next step, we observe that $\Psi_p(c,\xi)= \Psi_p(-c,\xi^{-1})$, which indicates that it suffices to show~\eqref{eq:cond1} for all $(c,\xi) \in [0,1)\times(0,1]$ and all $p \in \mathbb{N}$. We observe that \begin{equation}\nonumber \Psi_p(c,\xi) = \frac{ 1 + \left(\frac{1-c}{1+ c}\right)^{p-1} \xi }{ 2 \left( (1+ c) + (1-c)\left(\frac{1-c}{1+ c}\right)^{p-1} \xi \right) } \end{equation} and $ \omega :=\left(\frac{1-c}{1+c}\right)^{p-1} \in [0,1]$ for $c\in[0,1]$. So, it suffices to show that \begin{equation}\label{eq:cond2} \frac{ 1 + \omega \xi }{ 2 ( (1+ c) + (1-c) \omega \xi ) } \le \frac12 \end{equation} for all $(c,\xi,\omega)\in [0,1)\times(0,1]\times[0,1]$ and all $p \in \mathbb{N}$, again ignoring the dependence of $\omega$ on $p$ and $c$. As the denominator is always positive,~\eqref{eq:cond2} is equivalent to \begin{equation*} 1+ \omega \xi \le 1+\omega\xi + c (1- \omega \xi), \end{equation*} which is obviously true for all $(c,\xi,\omega)\in [0,1)\times(0,1]\times[0,1]$. Now, we consider \emph{case B}: $l = 0$. Here, we have to use that $\widehat{e}_{p,\frac{h}{2}}^{(0)}\not=0$ and obtain -- by straight-forward computation -- that \begin{equation*} \mathcal{T}_{p,\frac{h}{2}}^{(0)} = \left( \begin{array}{cc} 0&0\\0&\tfrac14\end{array} \right) \end{equation*} and consequently $\rho( \mathcal{T}_{p,\frac{h}{2}}^{(0)} ) = \tfrac14$. Also this is bounded from above by $\tfrac12$, which finishes the proof. \qed \end{proof} \subsection{The approximation error estimate} Now, we are able to show the approximation error estimate~\eqref{eq:whattolfa}. \begin{lemma}\label{lem:lfa} The inequality~\eqref{eq:whattolfa}, i.e., \begin{equation}\nonumber \|(I- \widehat{\Pi}_{p,\hn} ) u_{p,\frac{h}{2}}\|_{L^2(-1,1)} \le \frac{1}{\sqrt{2}}\; \hn |u_{p,\frac{h}{2}}|_{H^1(-1,1)}, \end{equation} holds for all $u_{p,\frac{h}{2}}\in \widehat{S}_{p,\frac{h}{2}}(-1,1)$. \end{lemma} \begin{proof} Lemma~\ref{lemma:decomp} states that~\eqref{eq:whattolfa} is a consequence of \eqref{eq:whattolfa3} and \eqref{eq:whattolfa4}. As Lemma~\ref{lem:mass} shows~\eqref{eq:whattolfa3} and Lemma~\ref{lem:lfa0} shows~\eqref{eq:whattolfa4}, this finishes the proof.\qed \end{proof} \section{The proof of Theorem~\citeThrmApprox{}}\label{sec:thrm1} In the previous section, we have given a proof for the approximation error of discretized functions between two consecutive grids. Using a telescoping argument, we can extend this result to an approximation error estimate for general functions. As in the last section, we first consider the periodic case. \begin{lemma}\label{lem:approx:per} For all $u \in \widehat{H}^{1}(-1,1)$, all grid sizes $h$ and each $p\in \mathbb{N}$, with $hp<1$, \begin{equation*} \|(I-\widehat{\Pi}_{p,h})u\|_{L^2(-1,1)} \le \sqrt{2}\; \hn |u|_{H^1(-1,1)} \end{equation*} is satisfied, where $\widehat{\Pi}_{p,h}$ is given as in Definition~\ref{def:H1projection}. \end{lemma} \begin{proof} Using a telescoping argument, i.e. iteratively applying the triangular inequality, and the relation $\widehat{\Pi}_{p,2h} \widehat{\Pi}_{p,h} = \widehat{\Pi}_{p,2h}$ for the projectors, we obtain for any $q\in\mathbb{N}$ \begin{align*} \|(I-\widehat{\Pi}_{p,h})u\|_{L^2(-1,1)} & \le \|(I-\widehat{\Pi}_{p,2^{-q}h})u \|_{L^2(-1,1)} \\ &\qquad+ \sum_{l=0}^{q-1} \|(I-\widehat{\Pi}_{p,2^{-l}h} )\widehat{\Pi}_{p,2^{-l-1}h} u \|_{L^2(-1,1)}. \end{align*} We use Lemma~\ref{lem:non:robust} and a standard Aubin-Nitsche duality argument to estimate $\|(I-\widehat{\Pi}_{p,2^{-q}h})u\|_{L^2(-1,1)}$ from above. Using \cite{Braess:1997}, Lemma~7.6, and Lemma~\ref{lem:non:robust} for $r=1$ and $q=2$, we immediately obtain \begin{equation}\label{eq:aubin} \|(I-\widehat{\Pi}_{p,2^{-q}h})u\|_{L^2(-1,1)} \le \widetilde{C}(p) 2^{-q} h \|u\|_{H^1(-1,1)}, \end{equation} where $\widetilde{C}(p)$ is independent of the grid size. Using~\eqref{eq:aubin} and Lemma~\ref{lem:lfa}, we obtain \begin{align*} \|(I-\widehat{\Pi}_{p,h})u \|_{L^2(-1,1)} &\le \widetilde{C}(p) \; 2^{-q}\hn \|u \|_{H^1(-1,1)} \\ &\qquad + \sum_{l=0}^{q-1} \frac{1}{\sqrt{2}} \;2^{-l} \hn |\widehat{\Pi}_{p,2^{-l-1}h} u|_{H^1(-1,1)}. \end{align*} Because $\widehat{\Pi}_{p,h}$ is $H^1$-orthogonal, we obtain $|\widehat{\Pi}_{p,2^{-l-1}h} u|_{H^1(-1,1)} \leq |u |_{H^1(-1,1)}$ and further \begin{equation*} \|(I-\widehat{\Pi}_{p,h})u \|_{L^2(-1,1)} \le \widetilde{C}(p) \; 2^{-q} \hn \|u \|_{H^1(-1,1)} + \sum_{l=0}^{q-1} \frac{1}{\sqrt{2}} \;2^{-l} \hn |u |_{H^1(-1,1)}. \end{equation*} The summation formula for the infinite geometric series gives \begin{equation*} \|(I-\widehat{\Pi}_{p,h})u \|_{L^2(-1,1)} \le \widetilde{C}(p) \; 2^{-q} \hn \|u \|_{H^1(-1,1)} + 2 \frac{1}{\sqrt{2}} \hn |u|_{H^1(-1,1)}. \end{equation*} As this is true for all $q\in \mathbb{N}$, we can take the limit $q\rightarrow \infty$ and obtain the desired result.\qed \end{proof} Having this result, we note that Theorem~\ref{thrm:approx} is just the extension of Lemma~\ref{lem:approx:per} to the non-periodic case. So, we can easily prove Theorem~\ref{thrm:approx}. \begin{proof}{\em of Theorem~\ref{thrm:approx}} In the following, we assume without loss of generality that $\Omega=(0,1)$. The extension to any other $\Omega=(a,b)$, follows using a standard scaling argument. Observe that any $u\in H^{1}(0,1)$ can be extended to a $w\in \widehat{H}^1(-1,1)$ by defining $w(x):=u(|x|)$. The assumption $hp<1$ in Theorem~\ref{thrm:approx} guarantees that Lemma~\ref{lem:approx:per} can be applied. We set $w_{p,h}:= \widehat{\Pi}_{p,h} w \in \widehat{S}_{p,h}(-1,1)$ as in Lemma~\ref{lem:approx:per}, such that \begin{equation*} \|w-w_{p,h}\|_{L^2(-1,1)} \le \sqrt{2}\; \hn |w|_{H^1(-1,1)}. \end{equation*} The function $w_{p,h}$ is symmetric, i.e., $w_{p,h}(x)=w_{p,h}(-x)$ holds. This can be seen by the following argument: As $w$ satisfies $w(x)=w(-x)$, we have for $\widetilde{w}_{p,h}(x):=w_{p,h}(-x)$ \begin{equation*} \|w-w_{p,h}\|_{L^2(-1,1)} = \|w-\widetilde{w}_{p,h}\|_{L^2(-1,1)} \end{equation*} and as $w_{p,h}$ was a unique minimizer, consequently $w_{p,h}(x) = \widetilde{w}_{p,h}(x)=w_{p,h}(-x)$ holds. By restricting $w_{p,h}$ to $(0,1)$, we obtain a function $u_{p,h}\in S_{p,h}(0,1)$. This function satisfies the desired approximation error estimate since $|w|_{H^1(-1,1)} = \sqrt{2} |u|_{H^1(0,1)}$ and $\|w-w_{p,h}\|_{L^2(-1,1)} = \sqrt{2} \|u-~u_{p,h}\|_{L^2(0,1)}$ hold due to the symmetry of $w$.\qed \end{proof} \section{Approximation error estimate for a reduced spline space}\label{sec:reduced} In the proof of Theorem~\ref{thrm:approx} we have defined $u_{p,h}$ to be the restriction of a symmetric and periodic spline $w_{p,h} \in \widehat{S}_{p,h}(-1,1)$ to $(0,1)$. So, we know more about $u_{p,h}$ than just $u_{p,h}\inS_{p,h}(0,1)$. Throughout this Section we again assume $hp<|\Omega|$. As we have shown in the proof of Theorem~\ref{thrm:approx} the spline~$w_{p,h}$ is symmetric, i.e., $w_{p,h}(x)=w_{p,h}(-x)$, so we have \begin{equation*} \frac{\partial^{l}}{\partial x^{l}} w_{p,h}(x)= (-1)^l \frac{\partial^{l}}{\partial x^{l}} w_{p,h}(-x) \mbox{ for all } l \in \mathbb{N}_0 . \end{equation*} By plugging~$x=0$ into this relation, we obtain that all odd derivatives vanish for~$x=0$. By plugging~$x=1$ into the relation, we obtain together with~\eqref{eq:sym:cond} that also for~$x=1$ all odd derivatives vanish. So, we have shown that the approximation error estimate~\eqref{eq:thrm:approx} is still satisfied if we restrict the approximating spline $u_{p,h}$ to be in the space $\widetilde{S}_{p,h}(0,1)$, defined as follows. \begin{definition}\label{defi:Ssymm} Given a spline space $S_{p,h}(\Omega)$ over $\Omega=(a,b)$, the \emph{space of splines with vanishing odd derivatives} $\widetilde{S}_{p,h}(\Omega)$ is the space of all $u_{p,h}\in S_{p,h}(\Omega)$ that satisfy the following condition: \begin{equation*} \frac{\partial^{2l+1}}{\partial x^{2l+1}} u_{p,h}(a)=\frac{\partial^{2l+1}}{\partial x^{2l+1}} u_{p,h}(b) = 0 \mbox{ for all } l \in \mathbb{N}_0 \mbox{ with } 2l+1 < p. \end{equation*} \end{definition} Using a standard scaling argument, we can again extend the result for $\Omega=(0,1)$ to any $\Omega=(a,b)$ and obtain the following Corollary. \begin{corollary}\label{cor:approx:nonper} For all $u\in H^1(\Omega)$, all grid sizes $h$ and all $p\in\mathbb{N}$, with $hp<|\Omega|$, there is a spline approximation $u_{p,h}\in \widetilde{S}_{p,h}(\Omega)$ such that \begin{equation*} \|u-u_{p,h}\|_{L^2(\Omega)} \le \sqrt{2}\; \hn |u|_{H^1(\Omega)} \end{equation*} is satisfied. \end{corollary} In the Appendix, we will introduce a basis for the space~$\widetilde{S}_{p,h}(\Omega)$. Based on the bases of those spaces, we obtain that their dimensions are as given in Table~\ref{tab:dof}. \begin{table} \begin{center} \begin{tabular}{cccc} \hline\noalign{\smallskip} & dim $S_{p,h}(0,1)$ & dim $\widehat{S}_{p,h}(0,1)$ & dim $\widetilde{S}_{p,h}(0,1)$ \\ \noalign{\smallskip}\hline\noalign{\smallskip} $p$ even & $n+p$ & $n$ & $n$ \\ $p$ odd & $n+p$ & $n$ & $n+1$ \\ \noalign{\smallskip}\hline \end{tabular} \caption{Degrees of freedom, where $n$ is the number of elements in $(0,1)$.} \label{tab:dof} \end{center} \end{table} \section{An inverse inequality for the reduced spline space}\label{sec:inverse} For the space~$\widetilde{S}_{p,h}(\Omega)$, a robust inverse inequality holds. Note that an extension to~$S_{p,h}(\Omega)$ is not possible (cf. Remark~\ref{rem:counterexample}). \begin{theorem}\label{thrm:inverse} For all grid sizes $h$ and each $p\in \mathbb{N}$, \begin{equation}\label{eq:inv2} |u_{p,h}|_{H^1(\Omega)} \le 2 \sqrt{3} \hn^{-1} \|u_{p,h}\|_{L^2(\Omega)} \end{equation} is satisfied for all $u_{p,h}\in \widetilde{S}_{p,h}(\Omega)$. \end{theorem} \begin{proof} In the following, we assume without loss of generality that $\Omega=(0,1)$. The extension to any other $\Omega=(a,b)$, follows directly using a standard scaling argument. We can extend every $u_{p,h}\in\widetilde{S}_{p,h}(0,1)$ to $(-1,1)$ by defining $w_{p,h}(x):=u_{p,h}(|x|)$ and obtain $w_{p,h} \in \widehat{S}_{p,h}(-1,1)$. \eqref{eq:inv2} is equivalent to \begin{equation}\label{eq:inv2aa} |w_{p,h}|_{H^1(-1,1)} \le 2 \sqrt{3} \hn^{-1} \|w_{p,h}\|_{L^2(-1,1)}. \end{equation} This is shown using induction in $p$ for all $u\in \widetilde{S}_{p,h}(-1,1)$. For $p=1$, \eqref{eq:inv2aa} is known, cf.~\cite{schwab:1998}, Theorem~3.91. Now, we show that the constant does not increase for larger $p$. So assume $p>1$ to be fixed. Due to the periodicity and due to the Cauchy-Schwarz inequality, \begin{align*} |w_{p,h}|_{H^1(-1,1)}^2 &= \int_{-1}^1 (w_{p,h}')^2 dx = -\int_{-1}^1 w_{p,h}'' w_{p,h} dx \\ &\le \|w_{p,h}''\|_{L^2(-1,1)} \|w_{p,h}\|_{L^2(-1,1)}= |w_{p,h}'|_{H^1(-1,1)} \|w_{p,h}\|_{L^2(-1,1)} \end{align*} is satisfied. Using the induction assumption (and~$w_{p,h}'\in \widehat{S}_{p-1,h}(-1,1)$, cf. \cite{Schumaker:1981}, Theorem~5.9), we know that \begin{equation*} |w_{p,h}'|_{H^1(-1,1)} \le 2 \sqrt{3} \hn^{-1} \|w_{p,h}'\|_{L^2(-1,1)} = 2 \sqrt{3} \hn^{-1} |w_{p,h}|_{H^1(-1,1)}. \end{equation*} Combining these results, we obtain \begin{equation*} |w_{p,h}|_{H^1(-1,1)}^2 \le 2 \sqrt{3} \hn^{-1} |w_{p,h}|_{H^1(-1,1)}\|w_{p,h}\|_{L^2(-1,1)} \end{equation*} and further \begin{equation*} |w_{p,h}|_{H^1(-1,1)} \le 2 \sqrt{3} \hn^{-1} \|w_{p,h}\|_{L^2(-1,1)}. \end{equation*} This shows \eqref{eq:inv2aa}, which concludes the proof.\qed \end{proof} \begin{remark} Neither Theorem~3.91 in~\cite{schwab:1998}, nor any of the arguments in the proof of Theorem~\ref{thrm:inverse} requires the grid to be equidistant. So, also having a general grid, estimate \begin{equation}\nonumber |u_{p,h}|_{H^1(\Omega)} \le 2 \sqrt{3} \; h_{\min}^{-1} \|u_{p,h}\|_{L^2(\Omega)} \end{equation} is satisfied for all splines $u_{p,h}$ on $\Omega=(a,b)$ with vanishing odd derivatives at the boundary. Here, as in any standard inverse inequality, $h_{\min}$ is the size of the \emph{smallest} element. \end{remark} As we have proven both an approximation error estimate and a corresponding inverse inequality, both of them are sharp (up to constants independent of $p$ and $\hn$). First, we show that there is a lower bound for the approximation error. As~\eqref{eq:corr:sharp1} is obviously true for constant functions, we show that there also exist other functions satisfying this inequality. \begin{corollary}\label{corr:sharp1} For all grid sizes $h$ and each $p\in\mathbb{N}$, there is a non-constant function $u\in H^1(\Omega)$ such that \begin{equation}\label{eq:corr:sharp1} \inf_{u_{p,h}\in \widetilde{S}_{p,h}(\Omega)} \|u-u_{p,h}\|_{L^2(\Omega)} \ge \frac{1}{4\sqrt{3}}\; \hn |u|_{H^1(\Omega)}. \end{equation} \end{corollary} \begin{proof} Let $u\in S_{p,\frac{\hn}{2}}(\Omega)\backslash\{0\}$ be such that $(u,\tilde{u}_{p,\hn})_{L^2(\Omega)}=0$ for all $\tilde{u}_{p,\hn}\in S_{p,h}(\Omega)$. As the constant functions are in $S_{p,h}(\Omega)$, this orthogonality implies that $u$ is non-constant. Using this orthogonality, we know that the infimum in~\eqref{eq:corr:sharp1} is taken for $u_{p,h}=0$. So, we obtain using Theorem~\ref{thrm:inverse} $\inf_{u_{p,h}\in \widetilde{S}_{p,h}(\Omega)} \|u-u_{p,h}\|_{L^2(\Omega)} = \|u\|_{L^2(\Omega)} \ge \frac{1}{2\sqrt{3}}\; \frac{\hn}{2} |u|_{H^1(\Omega)}$, which finishes the proof. \qed\end{proof} Similarly, we can give an lower bound for the inverse inequality. Again, we show the existence of a non-trivial function. \begin{corollary}\label{corr:sharp2} For all grid sizes $h$ with $2hp<|\Omega|$ and each $p\in\mathbb{N}$, there is a non-constant function $u_{p,h}\in \widetilde{S}_{p,h}(\Omega)$ such that \begin{equation}\nonumber |u_{p,h}|_{H^1(\Omega)} \ge \frac{1}{2\sqrt{2}}\; \hn^{-1} \|u_{p,h}\|_{L^2(\Omega)}. \end{equation} \end{corollary} \begin{proof} Let $u_{p,h}\in S_{p,h}(\Omega)\backslash\{0\}$ be such that $(u_{p,h},\tilde{u}_{p,2\hn})_{L^2(\Omega)}=0$ for all $\tilde{u}_{p,2\hn}\in S_{p,2\hn}(\Omega)$. As the constant functions are in $S_{p,2\hn}(\Omega)$, this orthogonality implies that $u_{p,h}$ is non-constant. Using this orthogonality and Theorem~\ref{thrm:approx}, we obtain $\|u_{p,h}\|_{L^2(\Omega)} = \inf_{u_{p,2\hn}\in \widetilde{S}_{p,n-1}(\Omega)} \|u_{p,h}-u_{p,2\hn}\|_{L^2(\Omega)} \le \sqrt{2} (2\hn) |u_{p,h}|_{H^1(\Omega)}$, which finishes the proof. \qed\end{proof} \section{An extension to higher Sobolev indices}\label{sec:sobolev} We can easily lift the statement of Theorem~\ref{thrm:approx} (and also Corollary~\ref{cor:approx:nonper}) up to higher Sobolev indices. \begin{theorem}\label{thrm:approx:sob} For all grid sizes $h$, each $q \in \mathbb{N}$ and each $p\in \mathbb{N}$ with $0< q\le p+1$ and with $h(p-q+1)<|\Omega|$, there is for each $u\in H^q(\Omega)$, a spline approximation $u_{p,h}\in \widetilde{S}_{p,h}^{(q)}(\Omega)$ such that \begin{equation*} |u-u_{p,h}|_{H^{q-1}(\Omega)} \le \sqrt{2} \; \hn |u|_{H^q(\Omega)}, \end{equation*} where $\widetilde{S}_{p,h}^{(q)}(\Omega)$ is the space of all $u_{p,h} \in S_{p,h}(\Omega)$ that satisfy the following symmetry condition: \begin{equation*} \frac{\partial^{2l+q}}{\partial x^{2l+q}} u_{p,h}(a)=\frac{\partial^{2l+q}}{\partial x^{2l+q}} u_{p,h}(b) = 0 \mbox{ for all } l \in \mathbb{N}_0 \mbox{ with } 2l+q < p. \end{equation*} \end{theorem} \begin{proof} Let again $\Omega=(0,1)$ without loss of generality. The proof is done by induction. From Corollary~\ref{cor:approx:nonper}, we know the estimate for $q=1$ (as $\widetilde{S}_{p,h}^{(1)}(0,1)=\widetilde{S}_{p,h}(0,1)$) and all $p> q-1=0$. For $q=1$ and $p=q-1=0$, the estimate is a well-known result, cf. \cite{Schumaker:1981}, Theorem~6.1,~(6.7), where (in our notation) $|u-u_{0,h}|_{L^2(0,1)} \le \hn |u|_{H^1(0,1)}$ has been shown. So, now we assume to know the estimate for some $q-1$ and show it for $q$. As $u\in H^q(0,1)$, we know that $u'\in H^{q-1}(0,1)$, so we can apply the induction hypothesis and obtain that there is some $u_{p-1,n}\in \widetilde{S}_{p-1,n}^{(q-1)}(0,1)$ with \begin{equation*} |u'-u_{p-1,n}|_{H^{q-2}(0,1)} \le \sqrt{2} \; \hn |u'|_{H^{q-1}(0,1)}. \end{equation*} Define \begin{equation}\label{eq:thrm:approx:sob} u_{p,h}(x):=c+\int_0^x u_{p-1,n}(\xi)d\xi. \end{equation} Note that $u_{p,h}\in S_{p,h}(0,1)$ as integrating increases both the polynomial degree and the differentiability by $1$, cf.~\cite{Schumaker:1981}, Theorem~5.16. After integrating, the boundary conditions on the $l$-th derivative become conditions on the $l+1$-st derivative, therefore we further have $u_{p,h}\in \widetilde{S}_{p,h}^{(q)}(0,1)$. Therefore, we have \begin{equation*} |u'-u_{p,h}'|_{H^{q-2}(0,1)} \le \sqrt{2} \; \hn |u'|_{H^{q-1}(0,1)}, \end{equation*} which is the same as \begin{equation*} |u-u_{p,h}|_{H^{q-1}(0,1)} \le \sqrt{2} \; \hn |u|_{H^{q}(0,1)}. \end{equation*} The bound on the grid size with respect to the degree, i.e. $h(p-q+1)<|\Omega|$ is sufficient, as the degree of $\partial^{q-1}/\partial x^{q-1} u$ is equal to $p-q+1$. This finishes the proof.\qed \end{proof} \begin{remark} The integration constant (integration constants for $q>2$) in~\eqref{eq:thrm:approx:sob} can be used to guarantee that \begin{equation*} \int_{\Omega} \frac{\partial^l}{\partial x^l}(u(x)-u_{p,h}(x)) \textnormal{d} x= 0 \end{equation*} for all $l\in\{0,1,\ldots,q-1\}$. \end{remark} For the spaces $\widetilde{S}_{p,h}^{(q)}(\Omega)$ there is again an inverse inequality. \begin{theorem}\label{thrm:inverse:sob} For all grid sizes $h$, each $q\in \mathbb{N}$ and each $p\in \mathbb{N}$ with $0< q \le p+1$, \begin{equation}\label{eq:inv2:sob} |u_{p,h}|_{H^q(\Omega)} \le 2 \sqrt{3} \hn^{-1} |u_{p,h}|_{H^{q-1}(\Omega)} \end{equation} is satisfied for all $u_{p,h}\in \widetilde{S}_{p,h}^{(q)}(\Omega)$, where $\widetilde{S}_{p,h}^{(q)}(\Omega)$ is as defined in Theorem~\ref{thrm:approx:sob}. \end{theorem} \begin{proof} First note that~\eqref{eq:inv2:sob} is equivalent to \begin{equation}\label{eq:inv2:sob:2} \left|\frac{\partial^{q-1}}{\partial x^{q-1}} u_{p,h}\right|_{H^1(\Omega)} \le 2 \sqrt{3} \hn^{-1} \left\|\frac{\partial^{q-1}}{\partial x^{q-1}} u_{p,h}\right\|_{L^2(\Omega)}. \end{equation} As $\frac{\partial^{q-1}}{\partial x^{q-1}}u_{p,h}\in \widetilde{S}_{p-q+1,n}^{(1)}(\Omega) = \widetilde{S}_{p-q+1,n}(\Omega)$, cf. \cite{Schumaker:1981}, Theorem~5.9, the estimate~\eqref{eq:inv2:sob:2} follows directly from Theorem~\ref{thrm:inverse}.\qed \end{proof} Again, as we have both an approximation error estimate and an inverse inequality, we know that both of them are sharp (cf. Corollaries~\ref{corr:sharp1} and~\ref{corr:sharp2}). The following theorem is directly obtained from telescoping. \begin{theorem}\label{thrm:approx:sob:2} For all grid sizes $h$, each $q\in\mathbb{N}_0$, each $p\in\mathbb{N}$, each $r\in\mathbb{N}$ with $0\le r\le q\le p+1$ and $h(p-r)<|\Omega|$, there is for each $u\in H^q(\Omega)$ a spline approximation $u_{p,h}\in S_{p,h}(\Omega)$ such that \begin{equation*} |u-u_{p,h}|_{H^r(\Omega)} \le (\sqrt{2}\; \hn)^{q-r} |u|_{H^q(\Omega)} \end{equation*} is satisfied. \end{theorem} \begin{proof} Theorem~\ref{thrm:approx:sob} states the desired result for $r=q-1$. For $r<q-1$, the statement is shown by induction in $r$. So, we assume to know the desired result for some $r$, i.e., there is a spline approximation $w_{p,h}\in S_{p,h}(\Omega)$ such that \begin{equation}\label{eq:thrm:0:ia} |u-w_{p,h}|_{H^r(\Omega)} \le (\sqrt{2}\; \hn)^{q-r} |u|_{H^q(\Omega)}. \end{equation} Now, we show that there is some $u_{p,h}\in S_{p,h}(\Omega)$ such that \begin{equation}\label{eq:thrm:0:ih} |u-u_{p,h}|_{H^{r-1}(\Omega)} \le (\sqrt{2}\; \hn)^{q-(r-1)} |u|_{H^q(\Omega)}. \end{equation} As $u-w_{p,h}\in H^r(\Omega)$, Theorem~\ref{thrm:approx:sob} states that there is a function $u_{p,h}\in S_{p,h}(0,1)$ such that \begin{equation*} |u-u_{p,h}|_{H^{r-1}(\Omega)} \le \sqrt{2}\; \hn |u-w_{p,h}|_{H^r(\Omega)}, \end{equation*} which shows together with the induction assumption~\eqref{eq:thrm:0:ia} the induction hypothesis~\eqref{eq:thrm:0:ih}. Again, the bound on the grid size $h(p-r)<|\Omega|$ follows directly from the bounds in Theorem~\ref{thrm:approx:sob}. \qed \end{proof} Here, it is not known to the authors how to choose a proper subspace of~$S_{p,h}(\Omega)$ such that a complementary inverse inequality can be shown. \section{Extension to two and more dimensions and application in Isogeometric Analysis}\label{sec:dim} Without loss of generality and to simplify the notation, we restrict ourselves to $\Omega:=(0,1)^d$ throughout this section. We can extend Theorem~\ref{thrm:approx} (and also Corollary~\ref{cor:approx:nonper}) to the following theorem for a tensor-product structured grid on $\Omega$. Here, we can introduce $\widetilde{W}_{p,h}(\Omega) = \otimes_{l=1}^d \widetilde{S}_{p,h}(0,1)$. Let $n=\nh$, for even $p$, and $n=\nh+1$ for odd $p$. Assuming that $(\widetilde{\varphi}_{p,h}^{(0)},\ldots , \widetilde{\varphi}_{p,h}^{(n-1)})$ is a basis of $\widetilde{S}_{p,h}(0,1)$, the space $\widetilde{W}_{p,h}(\Omega)$ is given by \begin{equation*} \widetilde{W}_{p,h}(\Omega)=\left\{w\,:\,w(x_1,\ldots,x_d)=\hspace{-2mm}\sum_{i_1,\ldots,i_d=0}^{n-1}\hspace{-2mm} w_{i_1,\ldots,i_d} \widetilde{\varphi}_{p,h}^{(i_1)}(x_1) \cdots \widetilde{\varphi}_{p,h}^{(i_d)}(x_d) \right\}. \end{equation*} \begin{theorem}\label{eq:approx2d} For all $u\in H^1(\Omega)$, all grid sizes $h$ and each $p\in\mathbb{N}_0$, with $hp<1$, there is a spline approximation $w_{p,h}\in \widetilde{W}_{p,n}(\Omega)$ such that \begin{equation*} \|u-w_{p,h}\|_{L^2(\Omega)} \le \sqrt{2d}\; \hn |u|_{H^1(\Omega)} \end{equation*} is satisfied. \end{theorem} The proof is similar to the proof in~\cite{Beirao:2012}, Section 4, for the two dimensional case. To keep the paper self-contained we give a proof of this theorem. \begin{proof}{\em of Theorem~\ref{eq:approx2d}} For sake of simplicity, we restrict ourselves to $d=2$. The extension to more dimensions is completely analogous. Here \begin{equation*} \widetilde{W}_{p,h}(\Omega)=\widetilde{S}_{p,h}(0,1)\otimes\widetilde{S}_{p,h}(0,1)=\left\{w\;:\;w(x,y)=\sum_{i,j=0}^{n-1} w_{i,j} \widetilde{\varphi}_{p,h}^{(i)}(x) \widetilde{\varphi}_{p,h}^{(j)}(y) \right\}. \end{equation*} We assume $u\in C^\infty(\Omega)$ and show the desired result using a standard density argument. Using Corollary~\ref{cor:approx:nonper}, we can introduce for each $x\in(0,1)$ a function $v(x,\cdot)\in \widetilde{S}_{p,h}(0,1)$ with \begin{equation*} \|u(x,\cdot)-v(x,\cdot)\|_{L^2(0,1)} \le \sqrt{2}\; \hn |u(x,\cdot)|_{H^1(0,1)}. \end{equation*} By squaring and taking the integral over $x$, we obtain \begin{equation}\label{eq:2d:1} \|u-v\|_{L^2(\Omega)} \le \sqrt{2}\; \hn \left\|\frac{\partial}{\partial y} u\right\|_{L^2(\Omega)}. \end{equation} By choosing $v(x,\cdot)$ to be the $L^2$-orthogonal projection, we also have \begin{equation*} \|v(x,\cdot)\|_{L^2(0,1)} \le \|u(x,\cdot)\|_{L^2(0,1)} \end{equation*} for all $x\in(0,1)$ and consequently \begin{equation}\label{eq:2d:x} \left\|\frac{\partial}{\partial x}v(x,\cdot)\right\|_{L^2(0,1)} \le \left\|\frac{\partial}{\partial x}u(x,\cdot)\right\|_{L^2(0,1)}. \end{equation} As $v(x,\cdot)\in \widetilde{S}_{p,h}(0,1)$, there are coefficients $v_j(x)$ such that \begin{equation*} v(x,y) = \sum_{j=0}^{n-1} v_j(x) \widetilde{\varphi}_{p,h}^{(j)}(y). \end{equation*} Using Corollary~\ref{cor:approx:nonper}, we can introduce for each $j\in\{0,\ldots,N\}$ a function $w_j\in \widetilde{S}_{p,h}(0,1)$ with \begin{equation}\label{eq:2d:2a} \|v_j-w_j\|_{L^2(0,1)} \le \sqrt{2}\; \hn |v_j|_{H^1(0,1)}. \end{equation} Next, we introduce a function $w$ by defining \begin{equation*} w(x,y):=\sum_{j=0}^{n-1} w_{j}(x) \widetilde{\varphi}_{p,h}^{(j)}(y), \end{equation*} which is obviously a member of the space $\widetilde{W}_{p,n}(\Omega)$. By squaring \eqref{eq:2d:2a}, multiplying it with $\widetilde{\varphi}_{p,h}^{(j)}(y)^2$, summing over $j$ and taking the integral, we obtain \begin{equation}\nonumber \int_0^1 \sum_{j=0}^{n-1}\|v_j-w_j\|_{L^2(0,1)}^2\widetilde{\varphi}_{p,h}^{(j)}(y)^2\textnormal{d} y \le 2\; \hn^2 \int_0^1\sum_{j=0}^{n-1}|v_j|_{H^1(0,1)}^2 \widetilde{\varphi}_{p,h}^{(j)}(y)^2\textnormal{d} y. \end{equation} Using the definition of the norms, we obtain \begin{equation}\nonumber \int_0^1\int_0^1 \sum_{j=0}^{n-1} (v_j(x)-w_j(x))^2\widetilde{\varphi}_{p,h}^{(j)}(y)^2\textnormal{d} x\textnormal{d} y \le 2\; \hn^2 \int_0^1\int_0^1\sum_{j=0}^{n-1} v_j'(x)^2 \widetilde{\varphi}_{p,h}^{(j)}(y)^2\textnormal{d} x\textnormal{d} y \end{equation} and further \begin{equation}\nonumber \|v-w\|_{L^2(\Omega)} \le \sqrt{2}\; \hn \left\|\frac{\partial}{\partial x} v\right\|_{L^2(\Omega)}. \end{equation} Using~\eqref{eq:2d:x}, we obtain \begin{equation}\label{eq:2d:2} \|v-w\|_{L^2(\Omega)} \le \sqrt{2}\; \hn \left\|\frac{\partial}{\partial y} u\right\|_{L^2(\Omega)}. \end{equation} Using~\eqref{eq:2d:1} and~\eqref{eq:2d:2}, we obtain \begin{align} \|u-w\|_{L^2(\Omega)} &\le \|u-v\|_{L^2(\Omega)} + \|v-w\|_{L^2(\Omega)} \nonumber \\ & \le \sqrt{2}\; \hn \left\|\frac{\partial}{\partial y} u\right\|_{L^2(\Omega)} + \sqrt{2}\; \hn \left\|\frac{\partial}{\partial x} u\right\|_{L^2(\Omega)}\label{eq:anisotropic-estimate}\\ & \le 2\; \hn |u|_{H^1(\Omega)},\nonumber \end{align} which finishes the proof.\qed \end{proof} The extension of Theorem~\ref{thrm:inverse} to two or more dimensions is rather easy. \begin{theorem} For all grid sizes $h$ and each $p\in \mathbb{N}$, the inequality \begin{equation}\nonumber |u_{p,h}|_{H^1(\Omega)} \le 2 \;\sqrt{3d} \;\hn^{-1}\;\|u_{p,h}\|_{L^2(\Omega)} \end{equation} is satisfied for all $u_{p,h}\in \widetilde{W}_{p,h}(\Omega)$. \end{theorem} \begin{proof} For sake of simplicity, we restrict ourselves to $d=2$. The generalization to more dimensions is completely analogous. We have obviously \begin{align*}\nonumber |u_{p,h}|_{H^1(\Omega)}^2 &= \left\|\frac{\partial}{\partial x} u_{p,h}\right\|_{L^2(\Omega)}^2 + \left\|\frac{\partial}{\partial y} u_{p,h}\right\|_{L^2(\Omega)}^2\\ & = \int_0^1 |u_{p,h}(\cdot,y)|_{H^1(0,1)}^2 \textnormal{d} y + \int_0^1 |u_{p,h}(x,\cdot)|_{H^1(0,1)}^2 \textnormal{d} x \end{align*} This can be bounded from above using Theorem~\ref{thrm:inverse} via \begin{align*}\nonumber |u_{p,h}|_{H^1(\Omega)}^2 \leq& 12 \hn^{-2} \left( \int_0^1 \|u_{p,h}(\cdot,y)\|_{L^2(0,1)}^2 \textnormal{d} y + \int_0^1 \|u_{p,h}(x,\cdot)\|_{L^2(0,1)}^2 \textnormal{d} x\right) \\ =& 24 \hn^{-2} \|u_{p,h}\|_{L^2(\Omega)}^2, \end{align*} which finishes the proof.\qed \end{proof} The extension to isogeometric spaces can be done following the approach presented in \cite{Bazilevs:2006}, Section 3.3. In Isogeometric Analysis, we have a geometry parameterization $\mathbf{F}:(0,1)^d \rightarrow \hat\Omega$. An isogeometric function on $\hat\Omega$ is then given as the composition of a B-spline on $(0,1)^d$ with the inverse of $\mathbf{F}$. The following result can be shown using a standard chain rule argument. There exists a constant $C=C(\mathbf{F},q)$ such that \begin{equation}\label{eq:norm:equivalence} C^{-1} \left\| f \right\|_{H^{q}(\hat\Omega)} \leq \left\| f \circ \mathbf{F} \right\|_{H^{q}((0,1)^d)} \leq C \left\| f \right\|_{H^{q}(\hat\Omega)} \end{equation} for all $f\in H^{q}(\hat\Omega)$. See \cite{Bazilevs:2006}, Lemma 3.5, or \cite{Beirao:2012}, Corollary 5.1, for related results. In both papers the statements are slightly more general, \cite{Bazilevs:2006} gives a more detailed dependence on the parameterization $\mathbf{F}$ whereas \cite{Beirao:2012} establishes bounds for anisotropic grids. Obviously, an extension to anisotropic grids can be achieved directly using the estimate \eqref{eq:anisotropic-estimate}. Note that the degree and the grid size are not necessarily equal in each parameter direction. Using this equivalence of norms, we can transfer all results from the parameter domain $(0,1)^d$ to the physical domain $\hat\Omega$. However, we need to point out that this equivalence is not valid for seminorms. Hence, in Theorem \ref{thrm:approx} (and follow-up Theorems \ref{thrm:approx:sob}, \ref{thrm:approx:sob:2} and \ref{eq:approx2d}) the seminorms on the right hand side of the equations need to be replaced by the full norms. Moreover, the bounds depend on the geometry parameterization via the constant $C$ in \eqref{eq:norm:equivalence}. A similar strategy can be followed when extending the results to NURBS. We do not go into the details here but refer to \cite{Bazilevs:2006,Beirao:2012} for a more detailed study. In the case of NURBS the seminorms again have to be replaced by the full norms due to the quotient rule of differentiation. In that case the constants of the bounds additionally depend on the given denominator of the NURBS parameterization. \section*{Acknowledgements} The authors want to thank Walter Zulehner for his suggestions, which helped to improve the presentation of the results in this paper. \section*{Appendix} At this point, we want to give a basis for $\widetilde{S}_{p,h}(\Omega)$ to make the reader more familiar with that space and to demonstrate that it is possible to work with it. The basis, which we introduce, is directly related to the (scaled) cardinal B-splines $\{\varphi_{p,h}^{(i)}\}^{\nh-1}_{i=-p}$. \begin{lemma}\label{lem:basis-tilde} The set $\{ \widetilde{\varphi}_{p,h}^{(i)} \}^{}_{i=-\left\lceil\frac p2\right\rceil , \ldots, \nh-\left\lfloor\frac p2\right\rfloor-1}$ with \begin{equation}\label{eq:basis-tilde} \widetilde{\varphi}_{p,h}^{(i)} := \sum_{l \in \{-i-p-1,i,2\nh -i -p -1\}} \varphi_{p,h}^{(l)} \end{equation} is a basis of $\widetilde{S}_{p,h}(\Omega)$. \end{lemma} Before we prove Lemma~\ref{lem:basis-tilde} we give a more practical representation of the basis functions by removing all contributions vanishing in $\Omega$. We obtain for odd $p$ that \begin{align*} &\widetilde{\varphi}_{p,h}^{(i)} = \varphi_{p,h}^{(i)} && i = -(p+1)/2\\ &\widetilde{\varphi}_{p,h}^{(i)} = \varphi_{p,h}^{(i)} + \varphi_{p,h}^{(-i-p-1)} && i = -(p-1)/2,\ldots,-1 \\ &\widetilde{\varphi}_{p,h}^{(i)} = \varphi_{p,h}^{(i)} && i=0,\ldots,\nh -p \\ &\widetilde{\varphi}_{p,h}^{(i)} = \varphi_{p,h}^{(i)} + \varphi_{p,h}^{(2\nh-i-p-1)} && i=\nh-p+1,\ldots, \nh-(p+1)/2 \\ &\widetilde{\varphi}_{p,h}^{(i)} = \varphi_{p,h}^{(i)} && i=\nh-(p-1)/2 \end{align*} and for even $p$ that \begin{align*} &\widetilde{\varphi}_{p,h}^{(i)} = \varphi_{p,h}^{(i)} + \varphi_{p,h}^{(-i-p-1)} && i = -p/2,\ldots,-1 \\ &\widetilde{\varphi}_{p,h}^{(i)} = \varphi_{p,h}^{(i)} && i=0,\ldots,\nh -p-1 \\ &\widetilde{\varphi}_{p,h}^{(i)} = \varphi_{p,h}^{(i)} + \varphi_{p,h}^{(2\nh-i-p-1)} && i=\nh-p,\ldots, \nh-p/2-1. \end{align*} Note that here we need that $0 \leq \nh-p-1$, which is equivalent to $hp<1$. \begin{proof}{\em of Lemma~\ref{lem:basis-tilde}} For the sake of simplicity, we consider the case $\Omega=(0,1)$ only. We show first that every function in~\eqref{eq:basis-tilde} is in $\widetilde{S}_{p,h}(0,1)$. Note that we have constructed $\widetilde{S}_{p,h}(0,1)$ such that the restriction of any symmetric function in $\widehat{S}_{p,h}(-1,1)$ to $(0,1)$ is a member of $\widetilde{S}_{p,h}(0,1)$. Let $n=1/h$. So, consider the functions $\{\widehat{\varphi}_{p,h}^{(j)}\}_{j=-n}^{n-1}$, forming a basis for $\widehat{S}_{p,h}(-1,1)$. Here we consider a different indexing with $j=i-n$. Defining $$s_{j}(x) := \widehat{\varphi}_{p,h}^{(j)}(x)+\widehat{\varphi}_{p,h}^{(j)}(-x)= \widehat{\varphi}_{p,h}^{(j)}(x)+\widehat{\varphi}_{p,h}^{(-j-p-1)}(x),$$ for $j=-n,\ldots,n-1$, we obtain symmetric functions in $\widehat{S}_{p,h}(-1,1)$. Using the relation $$\widehat{\varphi}_{p,h}^{(j)}|_{(0,1)} = \sum_{k\in\mathbb{Z}} \varphi_{p,h}^{(j+2nk)},$$ we obtain that the restriction of $s_j$ to $(0,1)$ fulfills $$s_j|_{(0,1)} = \sum_{k\in\mathbb{Z}} \varphi_{p,h}^{(j+2nk)} + \sum_{k\in\mathbb{Z}} \varphi_{p,h}^{(-j-p-1+2nk)} = \varphi_{p,h}^{(j)} + \sum_{k\in\mathbb{Z}} \varphi_{p,h}^{(-j-p-1+2nk)},$$ which is \begin{align*} &s_j|_{(0,1)} = \varphi_{p,h}^{(j)} + \varphi_{p,h}^{(-j-p-1)} &&\mbox{ for } j\in \{ -n,\ldots,-1\},\\ &s_j|_{(0,1)} = \varphi_{p,h}^{(j)} &&\mbox{ for } j\in \{ 0,\ldots,n-p-1\}, \mbox{ or}\\ &s_j|_{(0,1)} = \varphi_{p,h}^{(j)} + \varphi_{p,h}^{(-j-p-1+2n)} &&\mbox{ for } j\in \{ n-p,\ldots,n-1\}. \end{align*} In all three cases $s_j$ equals $\widetilde{\varphi}_{p,h}^{(j)}$ or $2\widetilde{\varphi}_{p,h}^{(j)}$. This shows that $\widetilde{\varphi}_{p,h}^{(i)}\in \widetilde{S}_{p,h}(0,1)$. It is easy to see that the functions in~\eqref{eq:basis-tilde} are linear independent for $i=-\left\lceil\frac p2\right\rceil , \ldots, n-\left\lfloor\frac p2\right\rfloor-1$. So, it remains to show that every function $u_{p,h} \in \widetilde{S}_{p,h}(0,1)$ can be expressed as a linear combination of the functions in~\eqref{eq:basis-tilde}. As we have already noticed, by construction the function $u_{p,h}$ can be extended to $(-1,1)$, by defining $w_{p,h}(x):=u_{p,h}(|x|)$. Note that $w_{p,h} \in \widehat{S}_{p,h}(-1,1)$. So, we can express it as a linear combination of basis functions of the basis given in~\eqref{eq:basis:varphi} via \begin{equation*} w_{p,h} = \sum_{j=-n}^{n-1} w_j \widehat{\varphi}_{p,h}^{(j)}. \end{equation*} By construction, $w_{p,h}(x)=w_{p,h}(-x)$, so we obtain \begin{align*} w_{p,h}(x) & = \frac12 (w_{p,h}(x)+w_{p,h}(-x)) = \frac12 \sum_{j=-n}^{n-1} w_j (\widehat{\varphi}_{p,h}^{(j)}(x)+\widehat{\varphi}_{p,h}^{(j)}(-x) ) \\ & = \frac12 \sum_{j=-n}^{n-1} w_j (\widehat{\varphi}_{p,h}^{(j)}(x)+\widehat{\varphi}_{p,h}^{(-j-p-1)}(x) )\\ & = \frac12 \sum_{j=-n}^{n-1}\sum_{k\in\mathbb{Z}} w_j (\varphi_{p,h}^{(j+2nk)}(x)+\varphi_{p,h}^{(-j-p-1+2nk)}(x) )\\ & = \frac12 \sum_{j=-n}^{n-1}w_j (\varphi_{p,h}^{(-j-p-1)}(x)+\varphi_{p,h}^{(j)}(x)+\varphi_{p,h}^{(2n-j-p-1)}(x) ). \end{align*} Again, it can be checked easily, that for all $j,n\in\mathbb{Z}$ the term $$ \varphi_{p,h}^{(-j-p-1)}(x)+\varphi_{p,h}^{(j)}(x)+\varphi_{p,h}^{(2n-j-p-1)}(x) $$ is in the span of $\{ \widetilde{\varphi}_{p,h}^{(i)} \}^{}_{i=-\left\lceil\frac p2\right\rceil , \ldots, n-\left\lfloor\frac p2\right\rfloor-1}$, which concludes the proof. \qed\end{proof} We observe that the basis forms a partition of unity. Moreover, all basis functions are obviously non-negative linear combinations of B-splines. Hence we call it a \emph{B-spline-like basis}. Fig.~\ref{fig:p2} and~\ref{fig:p4} depict the B-spline basis functions that span $\widetilde{S}_{p,h}(0,1)$. Here, the basis functions that have an influence at the boundary are plotted with solid lines. The basis functions that have zero derivatives up to order $p-1$ at the boundary coincide with standard B-spline functions. They are plotted with dashed lines. If we compare the pictures of the B-spline basis functions in $\widetilde{S}_{p,h}(0,1)$ (Fig.~\ref{fig:p2} and~\ref{fig:p4}) with the standard B-spline basis functions for $S_{p,h}(0,1)$ (Fig.~\ref{fig:p2a} and~\ref{fig:p4a}) obtained from a classical open knot vector, we see that the latter ones have more basis functions that are associated with the boundary. This can also be seen by counting the number of degrees of freedom, cf. Table~\ref{tab:dof}. \begin{figure} \begin{center} \includegraphics[scale=.55]{spline1.pdf} \includegraphics[scale=.55]{spline2.pdf} \caption{B-spline-like basis functions for $\widetilde{S}_{1,h}(0,1)$ and $\widetilde{S}_{2,h}(0,1)$} \label{fig:p2} \end{center} \begin{center} \includegraphics[scale=.55]{spline3.pdf} \includegraphics[scale=.55]{spline4.pdf} \caption{B-spline-like basis functions for $\widetilde{S}_{3,h}(0,1)$ and $\widetilde{S}_{4,h}(0,1)$} \label{fig:p4} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[scale=.55]{spline1a.pdf} \includegraphics[scale=.55]{spline2a.pdf} \caption{B-spline basis functions for ${S}_{1,h}(0,1)$ and ${S}_{2,h}(0,1)$} \label{fig:p2a} \end{center} \begin{center} \includegraphics[scale=.55]{spline3a.pdf} \includegraphics[scale=.55]{spline4a.pdf} \caption{B-spline basis functions for ${S}_{3,h}(0,1)$ and ${S}_{4,h}(0,1)$} \label{fig:p4a} \end{center} \end{figure} \bibliographystyle{amsplain}

https://arxiv.org/abs/1904.05322

Convex envelopes on Trees

We introduce two notions of convexity for an infinite regular tree. For these two notions we show that given a continuous boundary datum there exists a unique convex envelope on the tree and characterize the equation that this envelope satisfies. We also relate the equation with two versions of the Laplacian on the tree. Moreover, for a function defined on the tree, the convex envelope turns out to be the solution to the obstacle problem for this equation.

\section{Introduction} In mathematics, a real-valued function defined on an interval is called convex if the line segment between any two points on the graph of the function lies above or on the graph. Equivalently, a function is convex if its epigraph (the set of points on or above the graph of the function) is a convex set. For a twice differentiable function of a single variable, if the second derivative is always greater than or equal to zero in the entire interval then the function is convex. Convex functions play an important role in many areas of mathematics. They are especially important in the study of optimization problems where they are distinguished by a number of convenient properties. For instance, a (strictly) convex function has no more than one minimum. Even in infinite-dimensional spaces, under suitable additional hypotheses, convex functions continue to satisfy such properties and as a result, they are the most well-understood functionals in the calculus of variations. In probability theory, a convex function applied to the expected value of a random variable is always less than or equal to the expected value of the convex function of the random variable. On the other hand, linear and nonlinear equations (coming mainly from mean value properties) on trees are models that are close (and related to) to linear and nonlinear PDEs in the unit ball of $\mathbb{R}^N$, therefore, it seems natural to look for convex functions on trees. This is our main goal in this paper. For other analytical issues on discrete structures (including graphs such as trees) we refer to \cite{ary,BBGS,DPMR1,DPMR2,DPMR3,KLW,KW,Ober,s-tree,s-tree1} and references therein. Let us begin by making precise the well-known notion of convexity in $\mathbb{R}^N.$ We fix a bounded smooth domain $\Omega \subset {\mathbb{R}}^N$. A function $u\colon\Omega \to {\mathbb{R}}$ is said to be a convex function if for any two points $x,y \in \Omega$ such that the segment $[x,y]$ is contained in $\Omega$, it holds that \begin{equation} \label{convexo-usual} u(tx+(1-t)y) \leq tu(x) + (1-t) u(y), \qquad \forall t \in (0,1). \end{equation} With this definition one can define the convex envelope of a boundary datum $g \colon \partial \Omega \to {\mathbb{R}}$ as \begin{equation} \label{convex-envelope-usual} u^* (x) \coloneqq \sup \left\{u(x)\colon u \text{ is convex and verifies } u|_{\partial \Omega} \leq g \right\}. \end{equation} Here by $u|_{\partial \Omega} \leq g,$ we understand \[ \limsup_{\Omega \ni x\to z}u(z)\le g(z) \quad\forall z\in\partial\Omega. \] This convex envelope turns out to be the largest solution to \begin{equation} \label{convex-envelope-usual-eq} \begin{array}{ll} \lambda_1 (D^2 u) (x) = 0 \qquad &x\in \Omega, \end{array} \end{equation} (the equation has to be interpreted in viscosity sense) with \[ u(x) \leq g (x) \qquad x\in \partial \Omega. \] Here $\lambda_1\leq \lambda_2\leq\cdots\leq\lambda_N$ are the ordered eigenvalues of the Hessian matrix, $D^2u$. We refer to \cite{BlancRossi,HL1,OS,Ober33}. Notice that the equation \[ \lambda_1 (D^2 u) (x) = 0 \] is equivalent to \[ \min \left\{\langle D^2 u(x) v, v \rangle\colon v\in\mathbb{R}^N \text{ such that } \|v\|=1\right\} =0. \] This says that the equation that governs the convex envelope is just the minimum among all possible directions of the second derivative of the function at $x$ equal to zero. Here we notice that we have existence of a continuous up to the boundary convex envelope for every continuous data if and only if the domain is strictly convex, see \cite{BlancRossi,HL1,OS}. In this paper, our main goal is to develop similar ideas and concepts on a tree. When one wants to expend the notion of convexity to an ambient space beyond the Euclidean setting the two key ideas are to introduce what is a ``{}segment"{} in our space and once this is done, to understand what is a ``{}midpoint"{} in the segment. We introduce two different definitions of segments and midpoints and study the associated notion of convexity linked to each of them. In particular, for both definitions we are able to characterize the related equation that governs the convex envelope of a given boundary datum. Closely related to our results is \cite{BKND} where the authors considered convex functions on finite trees and showed that some relevant functions that are naturally related to spectral problems on the tree are convex. Before starting with our main goal, we need to introduce our ambient space. Given $m\in\mathbb{N}_{\ge2},$ a tree $\mathbb{T}_m$ with regular $m-$branching is an infinite graph that consists of the empty set $\emptyset$ and all finite sequences $(a_1,a_2,\dots,a_k)$ with $k\in{\mathbb N},$ whose coordinates $a_i$ are chosen from $\{0,1,\dots,m-1\}.$ \begin{center} \pgfkeys{/pgf/inner sep=0.19em} \begin{forest} [$\emptyset$, [0 [0 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] [1 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] [2 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] ] [1 [0 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] [1 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] [2 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] ] [2 [0 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] [1 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] [2 [0 [,edge=dotted]] [1 [,edge=dotted]] [2 [,edge=dotted]] ] ] ] \end{forest} A tree with $3-$branching. \end{center} The elements in $\mathbb{T}_m$ are called vertices. Each vertex $x$ has $m$ successors, obtained by adding another coordinate. We will denote by \[ \S(x)\coloneqq\{(x,i)\colon i\in\{0,1,\dots,m-1\}\} \] the set of successors of the vertex $x.$ If $x$ is not the root then $x$ has a only an immediate predecessor, which we will denote $\hat{x}.$ The segment connecting a vertex $x$ with $\hat{x}$ is called an edge and denoted by $(\hat{x},x).$ A vertex $x\in\mathbb{T}_m$ has level $k\in\mathbb{N}$ if $x=(a_1,a_2,\dots,a_k)$. The level of $x$ is denoted by $|x|$ and the set of all $k-$level vertices is denoted by $\mathbb{T}_m\!\!\!^k.$ We say that the edge $e=(\hat{x},x)$ has $k-$level if $x\in \mathbb{T}_m\!\!\!^k.$ A branch of $\mathbb{T}_m$ is an infinite sequence of vertices, where each one of them is followed by one of its immediate successors. The collection of all branches forms the boundary of $\mathbb{T}_m$, denoted by $\partial\mathbb{T}_m$. Observe that the mapping $\psi:\partial\mathbb{T}_m\to[0,1]$ defined as \[ \psi(\pi)\coloneqq\sum_{k=1}^{+\infty} \frac{a_k}{m^{k}} \] is surjective, where $\pi=(a_1,\dots, a_k,\dots)\in\partial\mathbb{T}_m$ and $a_k\in\{0,1,\dots,m-1\}$ for all $k\in\mathbb{N}.$ Whenever $x=(a_1,\dots,a_k)$ is a vertex, we set \[ \psi(x)\coloneqq\psi(a_1,\dots,a_k,0,\dots,0,\dots). \] At this point, we just have to recall that as we have mentioned, the definition of a convex function depends on what we understand by a segment and how to define the midpoint of the segment. A path from a vertex $x$ to a vertex $y$ in $\mathbb{T}_m$ is a sequence of vertices $x_0, x_1, x_2, \dots, x_k$ such that $x_0= x$, $x_k=y_0$ and for any $i= 1, 2, \dots, k$ we have that either $\hat{x}_{i-1}=x_i$ or $x_i\in\mathcal{S}({x}_{i-1})$, that is, $x_i$ and $x_{i+1}$ are adjacent (connected) in the graph $\mathbb{T}_m$. A path is called a minimal path if and only if all the vertices on the path are different. Observe that for any $x,y\in\mathbb{T}_m$ there is a unique minimal path from $x$ to $y.$ This minimal path is denoted by $[x,y].$ This is our first idea of what is a segment of $\mathbb{T}_m$. \begin{center} \pgfkeys{/pgf/inner sep=0.19em} \begin{forest} [$\emptyset$,circle,fill=yellow,draw=blue, [0,circle,fill=yellow,draw=blue[0[0][1][2]][1[0][1][2]] [2,circle,fill=yellow,draw=blue[0] [$x$,circle,fill=yellow,draw=blue][2]]] [1,circle,fill=yellow,draw=blue[0,circle,fill=yellow,draw=blue[$y$,circle,fill=yellow,draw=blue][1][2]][1[0][1][2]][2[0][1][2]]] [2[0[0][1][2]][1[0][1][2]][2[0][1][2]]] ] \end{forest} A path from a vertex $x$ to a vertex $y$. \end{center} In a slight abuse of notation, we say that an edge $e$ belongs to a path $\gamma$ if there is a vertex $x\in\gamma$ such that $e=(\hat{x},x).$ We now define the length of an edge $e$ as follows: \[ \text{length} (e) \coloneqq \frac{1}{m^k} \qquad \text{ if } e \text{ has level } k. \] With this length we can define the total length of a path $\gamma$ as the sum of the lengths of the edges involved in $\gamma$, that is, \[ \text{lenght} (\gamma) \coloneqq \sum_{e\in\gamma} \text{length} (e). \] Now, with this notion of length of an edge and then of a path, let us consider the distance between nodes given by the length of the minimal path. That is, given two nodes $x$, $y$ we let \[ d(x,y) \coloneqq \text{lenght}([x,y]). \] Remark that $d$ is a genuine metric since $d(x,y)\ge0,$ $d(x,y)=0$ iff $x=y$ and the triangular inequality holds (since the infimum of the lengths of the paths that joins $x$ with $y$ is less or equal than the infimum of the length of the paths that joins $x$ with $z$ plus the infimum of the length of the paths that joins $z$ with $y$). We are now ready to introduce our first notion of convex function. We say that a function $u:\mathbb{T}_m \mapsto \mathbb{R}$ is convex if for any $x,y,z\in \mathbb{T}_m$ with $z\in[x,y],$ it holds that \[ u(z) \leq \frac{d(y,z)}{d(x,y)} u(x) + \frac{d(x,z)}{d(x,y)} u(y). \] Using this definition, we can look for the convex envelope of a function defined on $\partial \mathbb{T}_m$. Given a function $g\colon[0,1]\to\mathbb{R},$ we define the convex envelope of $g$ on $\mathbb{T}_m$ as follows \begin{equation} \label{convex-envelope-arbol} u^*_g (x) \coloneqq \sup \Big\{u(x) \colon u\in\mathcal{C}(g) \Big\}, \end{equation} where \[ \mathcal{C}(g)\coloneqq \left\{ u\colon\mathbb{T}_m\to\mathbb{R}\colon u \text{ is a convex function} \text{ and } \limsup_{x\to \pi\in \partial \mathbb{T}_m} u(x)\leq g (\psi(\pi)) \right\}. \] The convex envelope is unique (this follows easily since the maximum of two convex functions is also convex). Moreover, associated to this convex envelope we have a nonlinear equation that is verified on the whole tree. \begin{teo} \label{teo.convex.arbol} The convex envelope of a continuous function $g\colon[0,1]\to\mathbb{R}$ is the largest solution to \begin{equation} \label{eq.tree.convex} u (x) = \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u(\hat{x}) + m \, u(y)}{m+1} \right\}\quad\text{on }\mathbb{T}_m, \end{equation} that verifies \begin{equation}\label{eq:bordecond} \limsup_{x\to \pi\in \partial \mathbb{T}_m}u(x) \leq g (\psi(\pi)). \end{equation} \end{teo} Let us clarify that in the case $x=\emptyset$, relation \eqref{eq.tree.convex} is understood as \[ u (x)= \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2, \] since $\emptyset$ does not have a predecessor because it is the root of $\mathbb{T}_m.$ Notice that \eqref{eq.tree.convex} is a nonlinear mean value property at the nodes of the tree. Once we have characterized the convex envelope by means of being the largest solution to \eqref{eq.tree.convex} that is below the datum on $\partial \mathbb{T}_m$, we want to study the associated Dirichlet problem, that is, given a datum $g$ on $\partial \mathbb{T}_m$ we want to solve the equation in $\mathbb{T}_m$ and find a solution that attains continuously the datum in the sense that \begin{equation}\label{eq:bordecond.77} \lim_{x\to \pi\in \partial \mathbb{T}_m}u(x) = g (\psi(\pi)). \end{equation} For this Dirichlet problem, we can show existence and uniqueness for continuous data. \begin{teo} \label{teo.convex.arbol.eq} Given a continuos boundary datum $g$, there is a unique solution to \eqref{eq.tree.convex} that verifies \eqref{eq:bordecond.77}. \end{teo} Therefore, from Theorems \ref{teo.convex.arbol} and \ref{teo.convex.arbol.eq}, we have that for every continuous datum on $\partial \mathbb{T}_m$ the unique convex envelope attains this datum with continuity, that is, \eqref{eq:bordecond.77} holds. Recall that in the Euclidean case we have to ask the domain $\Omega$ to be strictly convex for the validity of this continuity up to the boundary of the convex envelope. Notice that the equation \eqref{eq.tree.convex} can be rewritten as \begin{equation} \label{eq.tree.convex.derivadassegundas} 0 = \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) + u(z) -2u(x)}2 ; \min_{y\in\mathcal{S}(x)} \frac{u(\hat{x}) + m u(y) -(m+1) u(x)}{m+1}, \right\} \end{equation} and hence we identify one possible analogous to the eigenvalues of the Hessian for the Euclidean case but in the case of the tree \begin{equation} \label{uu} \left\{ \frac{u(x,i) + u(x,j) -2u(x)}2 \right\}_{i < j} \text{ and } \left\{ \frac{u(\hat{x}) + m u(y) -(m+1) u(x)}{m+1} \right\}_{y\in\S(x)}. \end{equation} Recall that for the convex envelope in the Euclidean space the associated equation reads as \[ \min_{1\le j\le N} \lambda_j (D^2 u)=0, \] and compare it with \eqref{eq.tree.convex.derivadassegundas}. Then, recalling that in the Euclidean setting when we add the eigenvalues of the Hessian we obtain the Laplacian, we obtain the following versions of a Laplacian on the tree adding the expressions found in \eqref{uu}, \begin{equation} \label{eq.Laplaciano.tree} u(x)= \frac{2}{(m+1)^2} u(\hat{x} ) + \frac{ m^2 +2m -1}{(m+1)^2} \frac1{m}\sum_{y\in\S(x)} u(y). \end{equation} Notice that this is a special case of the equations (given by mean value properties) that we previously studied in \cite{DPFR-DtoN} showing existence and uniqueness for the Dirichlet problem. Finally, we also study the convex envelope of a function defined on $\mathbb{T}_m$. That is, given a bounded function $f:\mathbb{T}_m \mapsto \mathbb{R}$ (not necessarily convex), we look for its convex envelope that is given by \begin{equation} \label{convex-envelope-arbol.f} u^*_f (x) \coloneqq \sup \big\{u(x) \colon u\in\mathcal{C}(f) \big\}, \end{equation} where \[ \mathcal{C}(f)\coloneqq \Big\{ u\colon\mathbb{T}_m\to\mathbb{R}\colon u \text{ is a convex function and } u(x)\leq f (x) \ \forall x\in \mathbb{T}_m \Big\}. \] The convex envelope is unique (this follows easily since the maximum of two convex functions is also convex). Notice that when $f$ attains a minimum this minimum coincides with the minimum of the convex envelope (and it is attained at the same vertices). This convex envelope turns out to be the solution to the obstacle problem for the equation \eqref{eq.tree.convex}. For the analogous property in the Euclidean setting, we refer to \cite{Ober33}. Recall that for the obstacle problem one important set is the coincidence set, that is given by the set of points where $f$ and its convex envelope $u^*_f$ coincide, \[ CS(f) =\left\{ x\in \mathbb{T}_m \, : \, f(x) = u^*_f (x) \right\}. \] \begin{teo} \label{teo.convex.arbol.f} The convex envelope of a function $f\colon \mathbb{T}_m \to\mathbb{R}$ is the solution to the obstacle problem for the equation \eqref{eq.tree.convex}. That is, $u^*_f$ is the largest solution to \begin{equation} \label{eq.tree.convex.ffff} u (x) \leq \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u(\hat{x}) + m u(y)}{m+1} \right\}\quad\text{on }\mathbb{T}_m, \end{equation} that verifies \[ u(x) \leq f(x) \qquad \forall x \in \mathbb{T}_m. \] In the coincidence set, the function $f$ verifies an inequality \begin{equation} \label{eq.tree.convex.ffff.latengo} f (x) \leq \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{f(y) +f(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ f(\hat{x}) + m f(y)}{m+1} \right\}\quad\text{on } CS(f), \end{equation} while outside the coincidence set the convex envelope, $u^*_f$, is a solution of the equation, i.e., it holds \begin{equation} \label{eq.tree.convex.ffff.latengo.mmmm} u^*_f (x) = \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u^*_f(y) +u^*_f(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u^*_f(\hat{x}) + m u^*_f(y)}{m+1} \right\}\quad\text{on }\mathbb{T}_m\setminus CS(f). \end{equation} \end{teo} On the other hand, in the arborescence (directed) tree, i.e., the tree defined as before but adding a direction to the edges in such a way that any two edges are not directed to the same vertex and the root is the unique vertex that has no edge directed into it), the Laplacian is defined as the mean value of the successors minus the value at the vertex, that is, \[ \Delta u (x) \coloneqq \frac1m \sum_{y\in\S(x)} u(y) - u(x) \quad\forall x\in\mathbb{T}_m, \] see \cite{KLW}. Now, to obtain an interpretation of this Laplacian as the sum of eigenvalues of the Hessian as we did before, we just have to change the notion of convexity. Now we need to introduce extra notations. Given $x\in\mathbb{T}_m,$ $\mathbb{T}_{2}^{x}$ denotes the set of all subgraphs $\mathbb{B}$ that are formed from a finite subset of the vertices of $\mathbb{T}_m$ and such that $x\in\mathbb{B},$ $\S(x)\cap\mathbb{B}$ has two elements and for any $y\in\mathbb{B}\setminus\{x\},$ \begin{itemize} \item $|y|>|x|;$ \item either $\S(y)\cap\mathbb{B}=\emptyset$ or $\S(y)\cap\mathbb{B}$ has exactly two elements. \end{itemize} We say that $y\in \mathbb{B}$ is an endpoint of $\mathbb{B}$ if $\S(y)\cap\mathbb{B}=\emptyset.$ The set of all endpoints of $\mathbb{B}$ is denoted $\mathcal{E}(\mathbb{B}).$ That is, $\mathbb{B}$ is just a finite binary subtree of $\mathbb{T}_m$. \begin{center} \pgfkeys{/pgf/inner sep=0.18em} \begin{forest} [$\emptyset$, [$x$, circle,fill=green,draw =orange [0[0][1][2]] [1, circle,fill=green,draw =orange [0][1][2]] [2, circle,fill=green,draw =orange [0, circle,fill=green,draw =orange] [1, circle,fill=green,draw =orange][2]]] [1[0[0][1][2]][1[0][1][2]][2[0][1][2]]] [2[0[0][1][2]][1[0][1][2]][2[0][1][2]]] ] \end{forest} An element of $\mathbb{T}_{2}^{x}.$ Root: $x$. Endpoints: $(x,1),$ $(x,2,0),$ and $(x,2,1).$ \end{center} Our second notion of convexity is the following: a function $u:\mathbb{T}_m \to \mathbb{R}$ is called binary convex if for any $x\in\mathbb{T}_m$ \begin{equation}\label{convex.II} u(x) \leq \sum_{y\in\mathcal{E}(\mathbb{B})}\dfrac1{2^{|y|-|x|}} u(y) \quad \forall \mathbb{B}\in\mathbb{T}_{2}^{x}. \end{equation} In this notion of convexity, a segment is $\mathbb{B}$, a finite binary subgraph of $\mathbb{T}_m$; a midpoint is the root of this subgraph $\mathbb{B}$ and the convexity property just says that the value of the function at the midpoint is less or equal than the mean value of the values at the endpoints. Associated to this new version of convexity, we have a convex envelope of a bounded boundary datum $g$ that is, defined as in \eqref{convex-envelope-arbol}, that is we define the binary convex envelope of $g$ on $\mathbb{T}_m$ as follows \[ \tilde{u}_g(x) \coloneqq \sup \left\{u(x) \colon u\in\mathcal{B}(g) \right\} \] where \[ \mathcal{B}(g)\coloneqq \left\{ u\colon \mathbb{T}_m\to\mathbb{R}\colon \text{ is a binary convex function on } \mathbb{T}_m, \limsup_{x\to \pi\in \partial \mathbb{T}_m} u(x)\leq g (\psi(\pi)) \right\}. \] For this notion of binary convex envelope, we also have an equation (simpler than with the previous notion). \begin{teo} \label{teo.convex.arbol.II} The binary convex envelope of a bounded boundary datum $g$ is the largest solution to \begin{equation} \label{eq.tree.convex.II} u(x) = \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) + u(z)}2 \quad\text{on }\mathbb{T}_m, \end{equation} that verifies \eqref{eq:bordecond}. \end{teo} Again, when $g$ is continuous we have a unique solution to the equation that attains the boundary datum continuously. \begin{teo} \label{teo.convex.arbol.eq.II} Given a continuous boundary datum $g$, there exists a unique \eqref{eq.tree.convex.II} that verifies \eqref{eq:bordecond.77}. \end{teo} In this case, written \eqref{eq.tree.convex.II} as \[ 0 = \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \left\{ \frac12 u(y) + \frac12 u(z) - u(x) \right\}, \] we can identify the analogous to the eigenvalues of the Hessian that for this case are given by, \begin{equation} \label{uu.II} \left\{ \frac12 u(x,i) + \frac12 u(x,j) - u(x) \right\}_{i < j} . \end{equation} Then, adding the eigenvalues in \eqref{uu.II}, we obtain \begin{equation} \label{eq.Laplaciano.tree.II} 0= \frac1{m}\sum_{y\in \S(x)} u(y) - u(x). \end{equation} Notice that this is the usual Laplacian in the arborescence tree studied in \cite{KLW}. For this notion of convexity we can also deal with the problem of the convex envelope of a function $f:\mathbb{T}_m \mapsto \mathbb{R}$ defined as in \eqref{convex-envelope-arbol.f}. In this case we also find that this convex envelope can be characterized as the solution to the obstacle problem for the associated equation, \eqref{eq.tree.convex.II}, and prove a theorem analogous to Theorem \ref{teo.convex.arbol.f} for this case. Once we have proved Theorem \ref{teo.convex.arbol.II}, the proof of this result is similar to the one of Theorem \ref{teo.convex.arbol.f} and hence we leave the statement and the proof to the reader. To end this introduction, let us give a natural generalization of the notion of binary convexity. Given $k\in\{2,\dots,m-2\}$ and $x\in\mathbb{T}_m,$ $\mathbb{T}_{2}^{x k}$ denotes the set of all subgraphs $\mathbb{B}$ that are formed from a finite subset of vertices in $\mathbb{T}_m$ and such that, $x\in\mathbb{B},$ $\S(x)\cap\mathbb{B}$ has exactly $k$ elements and for any $y\in\mathbb{B}\setminus\{x\},$ \begin{itemize} \item $|y|>|x|;$ \item either $\S(y)\cap\mathbb{B}=\emptyset$ or $\S(y)\cap\mathbb{B}$ has exactly $k$ elements. \end{itemize} Observe that $\mathbb{T}_{2}^{x 2}=\mathbb{T}_{2}^{x}.$ As before we denote by $\mathcal{E}(\mathbb{B})$ (the set of endpoints) the set of points $y\in \mathbb{B}$ such that $\S(y)\cap\mathbb{B}=\emptyset.$ A function $u:\mathbb{T}_m \to \mathbb{R}$ is called $k-$convex if for any $x\in\mathbb{T}_m$ \[ u(x) \leq \sum_{y\in\mathcal{E}(\mathbb{B})}\dfrac1{k^{|y|-|x|}} u(y) \quad \forall \mathbb{B}\in\mathbb{T}_{2}^{x k}. \] As before, associated to this version of convexity, we have a convex envelope of a bounded boundary datum $g$ that we will call the $k-$convex envelope of $g.$ Following the proof of Theorems \ref{teo.convex.arbol.II} and \ref{teo.convex.arbol.eq.II}, we can show that the $k-$convex envelope of $g$ is the largest solution to \begin{equation}\label{eq.tree.convex.III} u(x) = \min_{\substack{x_1,\dots,x_k\in\mathcal{S}(x)\\ x_i\neq x_j}} \frac1k\sum_{i=1}^k u(x_i) \quad\text{on }\mathbb{T}_m, \end{equation} that verifies \eqref{eq:bordecond}. Moreover, if $g$ is a continuous function then the $k-$convex envelope of $g$ is the a unique solution to \eqref{eq.tree.convex.III} that verifies \eqref{eq:bordecond.77}. In this case, written \eqref{eq.tree.convex.III} as \[ 0 = \min_{\substack{x_1,\dots,x_k\in\mathcal{S}(x)\\ x_i\neq x_j}} \left\{ \frac1k\sum_{i=1}^k u(x_i) - u(x) \right\}, \] we identify the analogous to the eigenvalues of the Hessian, \begin{equation} \label{uu.III} \left\{ \frac1k\sum_{i=1}^k u(x,j_i) - u(x), \right\}_{j_i < j_{i+1}}. \end{equation} Adding the eigenvalues in \eqref{uu.III}, we obtain again \eqref{eq.Laplaciano.tree.II}, the usual Laplacian on the arborescence tree. \bigskip {\bf Organization of the paper.} In Section \ref{sect-convex}, we will prove general results for convex functions; In Section \ref{sect-convex-envelopes}, we prove our main result for the convex envelope; In Sections \ref{section.biconvfunction} and \ref{sect-biconvex-envelopes}, we extend the results of previous sections to the notion of binary convexity. \section{Convex functions}\label{sect-convex} We begin this section showing a different characterization of convex functions. \begin{lem}\label{lema:convexeq} A function $u$ on the tree is convex if and only if $u$ satisfies \begin{equation} \label{subsol1} u (x) \le \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u(\hat{x}) + m u(y)}{m+1} \right\} \qquad \forall x \in \mathbb{T}_m. \end{equation} \end{lem} \begin{proof} Let $u$ be a convex function. Our goal is to show that \eqref{subsol1} holds. Given $x$ for any $y,z\in\S(x)$ with $y\neq z$ we have that \[ u(x)\le\dfrac12 u(y)+\dfrac12 u(z) \] due to the fact that $u$ is a convex function (just take $x\in[y,z],$ $d(y,z)=\tfrac2{m^{|x|+1}},$ and $d(y,x)=d(z,x)=\tfrac1{m^{|x|+1}}$ in the definition of convexity). Then, we get \[ u (x) \le \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 \] for any $x\in\mathbb{T}_m.$ Now, given $x\in \mathbb{T}_m$ for any $y\in \S(x)$ \[ u(x)\le\dfrac{m}{m+1} u(y)+\dfrac{1}{m+1} u(\hat{x}) \] again due to the fact that $u$ is a convex function (in this case, take $x\in[\hat{x},y],$ $d(\hat{x},y)=\tfrac{m+1}{m^{|x|+1}},$ $d(\hat{x},x)=\tfrac{1}{m^{|x|}},$ and $d(y,x)=\tfrac1{m^{|x|+1}}$). Thus \[ u (x) \le \min_{y\in\mathcal{S}(x)} \frac{ u(\hat{x}) + m u(y)}{m+1} \] for any $x\in\mathbb{T}_m.$ Therefore, we have that if $u$ is a convex function then $u$ satisfies \eqref{subsol1}. To see the converse, let $u$ be a function defined on the tree that verifies \eqref{subsol1} at every node and $x,y\in \mathbb{T}_m.$ We begin by analyzing the case that $[x,y]$ is a ``straight line", that is the vertices $x_0,\dots,x_N$ of $[x,y]$ are such that $x_0=x,$ $x_N=y,$ $\hat{x}_i=x_{i+1}$ for any $i\in\{0,\dots,N-1\}.$ More precisely, first we show that if $[x,y]$ is a ``straight line" then \begin{equation} \label{eq:auxenvol1} u(x_i)\le \dfrac{d(x_i,x_0)}{d(x_N,x_0)} u(x_N) + \dfrac{d(x_i,x_N)}{d(x_N,x_0)} u(x_0) \quad \forall i\in\{0,\dots,N\}. \end{equation} We proceed by induction in $N$. When $N=2,$ by \eqref{subsol1}, we have \[ u(x_1)\le\dfrac{u(x_2)+mu(x_0)}{m+1}= \dfrac{d(x_1,x_0)}{d(x_2,x_0)} u(x_2) + \dfrac{d(x_1,x_2)}{d(x_2,x_0)} u(x_0) \] since $d(x_1,x_2)=\tfrac{1}{m^{|x_1|}},$ $d(x_1,x_0)=\tfrac{1}{m^{|x_0|}}=\tfrac{1}{m^{|x_1|+1}}$ and $d(x_2,x_0)=\tfrac{m+1}{m^{|x_1|+1}}.$ Suppose now that \eqref{eq:auxenvol1} is true for all straight line that have $N-1$ vertices, where $N>2.$ Then, \[ u(x_1)\le \dfrac{d(x_1,x_{N-1})}{d(x_{N-1},x_0)} u(x_0) +\dfrac{d(x_1,x_{0})}{d(x_{N-1},x_0)} u(x_{N-1}) \] and \[ u(x_{N-1})\le \dfrac{d(x_1,x_{N-1})}{d(x_{N},x_1)} u(x_N) +\dfrac{d(x_{N},x_{N-1})}{d(x_{N},x_1)} u(x_{1}). \] Thus, \begin{align*} u(x_1)\le \dfrac{d(x_1,x_{N-1})}{d(x_{N-1},x_0)}u(x_0) &+\dfrac{d(x_1,x_{0})}{d(x_{N-1},x_0)} \dfrac{d(x_1,x_{N-1})}{d(x_{N},x_1)} u(x_N)\\ & \qquad + \dfrac{d(x_1,x_{0})}{d(x_{N-1},x_0)} \dfrac{d(x_{N},x_{N-1})}{d(x_{N},x_1)} u(x_{1}). \end{align*} Therefore, \begin{align*} &d(x_1,x_{N-1})\left[d(x_1,x_{N})u(x_0) +d(x_1,x_0)u(x_N)\right]\\[10pt] & \ge \left[d(x_{N-1},x_0)d(x_1,x_N)-d(x_1,x_0)d(x_{N-1},x_N) \right]u(x_1)\\[10pt] &\ge\left[ \left\{ d(x_{N},x_0)-d(x_N,x_{N-1}) \right\} \left\{ d(x_N,x_0)-d(x_1,x_0) \right\} -d(x_1,x_0)d(x_{N-1},x_N) \right]u(x_1)\\[10pt] &\ge d(x_{N},x_0) \left[ d(x_{N},x_0) -d(x_N,x_{N-1})-d(x_1,x_0) \right]u(x_1)\\[10pt] &\ge d(x_{N},x_0)d(x_{1},x_{N-1})u(x_1). \end{align*} Then \[ u(x_1)\le \dfrac{d(x_1,x_0)}{d(x_N,x_0)} u(x_N) + \dfrac{d(x_1,x_N)}{d(x_N,x_0)} u(x_N). \] In similar manner, we get \begin{equation} \label{eq:auxenvol2} u(x_{N-1})\le \dfrac{d(x_{N-1},x_0)}{d(x_{N},x_0)} u(x_N) +\dfrac{d(x_{N-1},x_N)}{d(x_N,x_0)} u(x_N). \end{equation} If $z\in [x,y]\setminus\{x_0,x_1,x_{N-1},x_N\},$ by the inductive hypothesis and \eqref{eq:auxenvol2}, we have \begin{align*} u(z)&\le\dfrac{d(z,x_{N-1})}{d(x_{N-1},x_0)} u(x_0) +\dfrac{d(z,x_{0})}{d(x_{N-1},x_0)} u(x_{N-1})\\[10pt] &\le \dfrac{d(z,x_0)}{d(x_N,x_0)}u(x_N) +\dfrac{d(z,x_{N-1})d(x_N,x_0)+d(z,x_0)d(x_{N-1},x_N)} {d(x_{N-1},x_0)d(x_N,x_0)}u(x_0)\\[10pt] &\le \dfrac{d(z,x_0)}{d(x_N,x_0)}u(x_N)\\[10pt] &\quad +\dfrac{\left[d(z,x_{N})-d(x_{N-1},x_N)\right]d(x_N,x_0)+ \left[d(x_N,x_0)-d(z,x_N)\right]d(x_{N-1},x_N)} {d(x_{N-1},x_0)d(x_N,x_0)}u(x_0)\\[10pt] &\le \dfrac{d(z,x_0)}{d(x_N,x_0)}u(x_N) +d(z,x_N) \dfrac{d(x_N,x_0)-d(x_{N-1},x_N)} {d(x_{N-1},x_0)d(x_N,x_0)}u(x_0)\\[10pt] &\le \dfrac{d(z,x_0)}{d(x_N,x_0)}u(x_N) +\dfrac{d(z,x_{N})}{d(x_N,x_0)}u(x_0), \end{align*} showing that indeed \eqref{eq:auxenvol1} holds when $[x,y]$ is a straight line. When $[x,y]$ is not a straight line, there is a $z\in[x,y]$ such that $[x,z]$ and $[z,y]$ are straight lines. Observe that $[x,y]=[x,z]\cup[z,y]$ and $\S(z)\cap[x,y]=\{w_1,w_2\}.$ We can assume that $w_1\in [x,z]$ and $w_2\in[z,y].$ Thus, from \eqref{eq:auxenvol1}, we have \begin{align*} 2u(z)&\le u(w_1)+ u(w_2)\\ &\le \left[\dfrac{d(w_1,x)}{d(x,z)} +\dfrac{d(w_2,y)}{d(y,z)} \right]u(z)+\dfrac{d(w_1,z)}{d(x,z)}u(x) +\dfrac{d(w_2,z)}{d(y,z)} u(y). \end{align*} Then, \[ \dfrac{2d(x,z)d(y,z)-d(w_1,x)d(z,y)-d(w_2,y)d(z,x)} {d(z,x)d(z,y)}u(z)\le \dfrac{d(w_1,z)}{d(x,z)}u(x) +\dfrac{d(w_2,z)}{d(y,z)} u(y). \] Since $d(w_1,z)=d(w_2,z),$ we get \begin{align*} d(w_1,z)&\left[d(y,z)u(x)+d(x,z) u(y)\right]\\[10pt] &\ge \left[ 2d(x,z)d(y,z)-d(w_1,x)d(z,y)-d(w_2,y)d(z,x) \right]u(z)\\[10pt] &\ge \left\{ 2d(x,z)d(y,z)-\left[d(x,z)-d(w_1,z)\right]d(z,y)- \left[d(y,z)-d(w_2,z)\right]d(z,x) \right\}u(z)\\[10pt] &\ge d(w_1,z)\left[d(z,y)+d(z,x)\right] u(z)\\ &\ge d(w_1,z)d(x,y) u(z). \end{align*} Therefore, we obtain \[ u(z)\le \dfrac{d(x,z)}{d(x,y)}u(y)+\dfrac{d(y,z)}{d(x,y)}u(x). \] If $w\in[x,y]\setminus\{x,z,y\}$ then $w\in[x,z]$ or $w\in[z,y].$ In the case that $w\in[x,z],$ since $[x,z]$ is a straight line we have \begin{align*} u(w)&\le \dfrac{d(x,w)}{d(x,z)}u(z)+\dfrac{d(z,w)}{d(x,z)}u(x)\\[10pt] &\le \dfrac{d(x,w)}{d(x,y)}u(y)+ \left[\dfrac{d(x,w)d(y,z)}{d(x,z)d(x,y)}+\dfrac{d(z,w)}{d(x,z)} \right]u(x)\\[10pt] &\le \dfrac{d(x,w)}{d(x,y)}u(y)+ \dfrac{d(x,w)d(y,z)+d(z,w)d(x,y)}{d(x,z)d(x,y)}u(x)\\[10pt] &\le \dfrac{d(x,w)}{d(x,y)}u(y)+ \dfrac{\left[d(x,y)-d(y,w)\right]d(y,z)+ d(z,w)d(x,y)}{d(x,z)d(x,y)}u(x)\\[10pt] &\le \dfrac{d(x,w)}{d(x,y)}u(y)+ \dfrac{\left[d(y,z)+d(z,w)\right]d(x,y)- d(y,w)d(y,z)}{d(x,z)d(x,y)}u(x)\\[10pt] &\le \dfrac{d(x,w)}{d(x,y)}u(y)+ \dfrac{d(y,w)\left[d(x,y)-d(y,z)\right]}{d(x,z)d(x,y)}u(x)\\[10pt] &\le \dfrac{d(x,w)}{d(x,y)}u(y)+ \dfrac{d(y,w)}{d(x,y)}u(x). \end{align*} In the case that $w\in [z,y]$ the proof is similar. Therefore, we conclude that a function $u$ that verifies \eqref{subsol1} is a convex function in $\mathbb{T}_m$. \end{proof} Our second result show that the sum of convex function is also a convex function. \begin{co}\label{co:sumconv} Let $u,v\colon \mathbb{T}_m\to\mathbb{R}$ be convex functions. Then $u+v$ is a convex function. \end{co} \begin{proof} Since $u,v$ are convex function, by Lemma \ref{lema:convexeq}, for any $x\in\mathbb{T}_m$ we have that \begin{align*} &u (x) + v(x) \le \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u(\hat{x}) + m u(y)}{m+1} \right\}\\ &\hspace{5cm}+\min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{v(y) +v(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ v(\hat{x}) + m v(y)}{m+1} \right\}\\ &\le\min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2+\frac{v(y) +v(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u(\hat{x}) + m u(y)}{m+1}+ \frac{ v(\hat{x}) + m v(y)}{m+1} \right\}. \end{align*} Therefore, by Lemma \ref{lema:convexeq}, $u+v$ is a convex function. \end{proof} It is immediate to check that the constant function $u=1$ is a convex function such that \[ \lim_{x\to\pi}u(x)=\mbox{\Large$\chi$}_{[0,1]}(\psi(\pi)) \quad \forall\pi\in\partial\mathbb{T}_m. \] We now show that for any $x_0\in\mathbb{T}_m\setminus\{\emptyset\}$ there is a convex function $u$ such that \[ \limsup_{x\to\pi}u(x)\le \mbox{\Large$\chi$}_{I_{x_0}} (\psi(\pi)) \quad\pi\in\partial\mathbb{T}_m. \] Here $I_{x_0}$ is the interval associated to the vertex $x_0$ of length $\tfrac{1}{m^{|x_0|}}$ given by \[ I_{x_0}\coloneqq\left[\psi(x_0),\psi(x_0)+\frac1{m^{|x_0|}}\right]. \] Observe that for $x_0\in \mathbb{T}_m$, $I_{x_0} \cap \partial\mathbb{T}_m$ is the subset of $\partial\mathbb{T}_m$ consisting of all branches that pass through $x_0$. To find such a convex function we introduce the following set: given $x_0\in\mathbb{T}_m$, let us consider \[ \mathbb{T}_m^{x_{0}}\coloneqq \{x\in\mathbb{T}_m \colon |x|\ge |x_0|,I_x\subset I_{x_0}\}. \] \begin{lem}\label{lema:caract} Let $x_0\in\mathbb{T}_m\setminus\{\emptyset\}.$ Then the function $u_{x_0}\colon\mathbb{T}_m\to\mathbb{R}$ \[ u_{x_0}(x)\coloneqq \frac{m-1}m \begin{cases} 0 &\text{if } x\not\in\mathbb{T}_m^{x_0},\\ \displaystyle\sum_{i=0}^{|x|-|x_0|}\frac1{m^i} &\text{if } x\in\mathbb{T}_m^{x_0}, \end{cases} \] is a convex function such that \[ \limsup_{x\to\pi}u_{x_0}(x) \le \mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)) \quad\forall \pi\in\partial\mathbb{T}_m. \] Moreover, \[ u_{x_0}(x)=\min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u_{x_0}(y) +u_{x_0}(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u_{x_0}(\hat{x}) + m u_{x_0}(y)}{m+1} \right\} \quad\forall x\in\mathbb{T}_m , \] and for any $\pi\in\mathbb{T}_m$ such that $\psi(\pi)$ is not one of the two endpoints of $I_{x_0}$ we have \[ \lim_{x\to\pi}u_{x_0}(x) =\mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)). \] \end{lem} \begin{proof} Let us start by showing that the function $u_{x_0}$ is convex. By Lemma \eqref{lema:convexeq}, it is enough to show that $u_{x_0}$ satisfies \eqref{subsol1}. If $x\in \mathbb{T}_m\setminus \mathbb{T}_m^{x_0}$ then there exist $y,z\in \S(x)$ such that $y\neq z,$ $u_{x_0}(y)=u_{x_0}(z)=0$. So, we have \begin{equation} \label{eq:caract1} \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u_{x_0}(y) +u_{x_0}(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u_{x_0}(\hat{x}) + m u_{x_0}(y)}{m+1} \right\}=0=u_{x_0}(x). \end{equation} If $x=x_0$, then $u_{x_0}(\hat{x}_0)=0$ and $u_{x_0}(y)=\tfrac{m-1}m(1+\tfrac1m)$ for any $y\in\S(x_0).$ Therefore \begin{equation} \label{eq:caract2} \begin{aligned} & \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u_{x_0}(y) +u_{x_0}(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ u_{x_0}(\hat{x}) + m u_{x_0}(y)}{m+1} \right\} \\[10pt] & \qquad =\frac{m-1}m\min\left\{ 1+\frac1m; 1\right\}\\[10pt] &\qquad =\frac{m-1}m\\[10pt] &\qquad =u_{x_0}(x_0). \end{aligned} \end{equation} Now, suppose that $x\in\mathbb{T}_m\setminus\{x_0\}$, and so, \[ u_{x_0}(\hat{x})=\frac{m-1}m \sum_{i=0}^{|x|-1-|x_0|}\dfrac1{m^i} \] and \[ u_{x_0}(y)=\frac{m-1}m \sum_{i=0}^{|x|+1-|x_0|}\dfrac1{m^i}. \] Hence, we obtain \begin{equation} \label{eq:caract3} \begin{aligned} \min \Bigg\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u_{x_0}(y) +u_{x_0}(z)}2 &; \min_{y\in\mathcal{S}(x_0)} \frac{ u_{x_0}(\hat{x}) + m u_{x_0}(y)}{m+1} \Bigg\}\\[10pt] &=\frac{m-1}m\min\left\{ \sum_{i=0}^{|x|+1-|x_0|}\dfrac1{m^i}; \sum_{i=0}^{|x|-|x_0|}\dfrac1{m^i}\right\}\\[10pt] &=\frac{m-1}m \sum_{i=0}^{|x|-|x_0|}\dfrac1{m^i}\\[10pt] &=u_{x_0}(x). \end{aligned} \end{equation} Therefore, by \eqref{eq:caract1}, \eqref{eq:caract2} and \eqref{eq:caract3} we get that \[ u_{x_0}(x)=\min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u_{x_0}(y) +u_{x_0}(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u_{x_0}(\hat{x}) + m u_{x_0}(y)}{m+1} \right\} \quad\forall x\in\mathbb{T}_m. \] Thus, $u_{x_0}$ is a convex function. Finally, we have to show that \[ \limsup_{x\to\pi}u_{x_0}(x) \le \mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)) \quad\forall \pi\in\partial\mathbb{T}_m. \] {\it Case 1}. If $\pi\in\partial\mathbb{T}_m,$ $\psi(\pi)\in I_{x_0}$ and $\psi(\pi)$ is not an endpoint of $I_{x_0}$ then for any sequence $\{x_k\}_{k\in\mathbb{N}}$ in $\mathbb{T}_m$ such that $\pi=(x_1,\dots,x_k,\dots)$, there is $k_0\in\mathbb{N}$ such that $x_k\in \mathbb{T}_m^{x_0}$ for all $k\ge k_0.$ Then \[ u_{x_0}(x_k)=\frac{m-1}m \sum_{i=0}^{|x_k|-|x_0|}\frac1{m^i}\quad \forall k\le k_0. \] Thus, as $k\to\infty$ we have \[ u_{x_0}(x_k)\to 1=\mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)). \] {\it Case 2}. Similarly, if $\pi\in\partial\mathbb{T}_m,$ $\psi(\pi)\not\in I_{x_0}$ then for any sequence $\{x_k\}_{k\in\mathbb{N}}$ on $\mathbb{T}_m$ such that $\pi=(x_1,\dots,x_k,\dots)$, we get \[ u_{x_0}(x_k)\to 0=\mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)) \] as $k\to\infty.$ {\it Case 3}. Finally suppose that $\pi\in\partial\mathbb{T}_m,$ $\psi(\pi)\in I_{x_0}$ and $\psi(\pi)$ is an endpoint of $I_{x_0}.$ In the case that $\psi(\pi)=0$ or $\psi(\pi)=1$ for any sequence $\{x_k\}_{k\in\mathbb{N}}$ on $\mathbb{T}_m$ such that $\pi=(x_1,\dots,x_k,\dots)$, there is $k_0\in\mathbb{N}$ such that $x_k\in \mathbb{T}_m^{x_0}$ for all $k\ge k_0.$ Therefore \[ u_{x_0}(x_k)\to 1=\mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)) \] as $k\to\infty.$ In the case that $\psi(\pi)\not\in\{0,1\}$ there exists sequences $\{x_k\}_{k\in\mathbb{N}},$ $\{y_k\}_{k\in\mathbb{N}}$ on $\mathbb{T}_m$ and $k_0\in\mathbb{N}$ such that $\pi=(x_1,\dots,x_k,\dots),$ $\pi=(y_1,\dots,y_k,\dots),$ for all $k\ge k_0$ we get $x_k\in \mathbb{T}_m^{x_0}$ and $y_k\in \mathbb{T}_m^{x_0}.$ Therefore, \begin{align*} &u_{x_0}(x_k)\to 1=\mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)),\\ &u_{x_0}(y_k)\to 0\le\mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)). \end{align*} This fact, together with the previous cases 1 and 2, completes the proof. \end{proof} \section{Convex envelopes}\label{sect-convex-envelopes} \setcounter{equation}{0} In this section we deal with convex functions on the tree. Let us start by showing that the convex envelop $u^*_g$ of function $g\colon[0,1]\to\mathbb{R}$, defined in \eqref{convex-envelope-arbol}, is a convex function. \begin{lem} \label{lema:convenv1} For any function $g\colon[0,1]\to\mathbb{R},$ the convex envelop $u_g^*$ is a convex function. \end{lem} \begin{proof} This follows easily from the fact that the supremum of convex functions is also convex. Given $g\colon[0,1]\to\mathbb{R},$ for every function $u\in\mathcal{C}(g)$ it holds that \[ u(z) \leq \frac{d(y,z)}{d(x,y)} u(x) + \frac{d(x,z)}{d(x,y)} u(y)\le \frac{d(y,z)}{d(x,y)} u^*_g(x) + \frac{d(x,z)}{d(x,y)} u^*_g(y) \] for any $x,y,z\in \mathbb{T}_m$ with $z\in[x,y].$ Hence we get \[ u^*_g(z) \le \frac{d(y,z)}{d(x,y)} u^*_g(x) + \frac{d(x,z)}{d(x,y)} u^*_g(y) \] for any $x,y,z\in \mathbb{T}_m$ with $z\in[x,y].$ Thus $u^*_g$ is a convex function. \end{proof} Our second aim is to show that if $g$ is a continuous function then \begin{equation} \label{eq:limb} \lim_{x\to \pi\in \partial \mathbb{T}_m} u_g^* (x) = g(\psi(\pi))\quad\forall\pi\in\partial\mathbb{T}_m. \end{equation} To prove this property, we need to show a comparison principle. \begin{lem} \label{lema.compar.convexas} Let $u$ and $v$ satisfy \begin{align} \label{eq.envolvente.44.compar.u} u(x) &\ge \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ u(\hat{x}_0) + m u(y)}{m+1} \right\} \quad\forall x\in\mathbb{T}_m,\\ \label{eq.envolvente.44.compar.v} v(x) &\le \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{v(y) +v(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ v(\hat{x}_0) + m v(y)}{m+1} \right\}\quad\forall x\in\mathbb{T}_m, \end{align} with \[ \lim_{x\to \pi\in \partial \mathbb{T}_m} u (x) \geq \lim_{x\to \pi\in \partial \mathbb{T}_m} v (x) , \] for every $\pi\in \partial \mathbb{T}_m$. Then, \[ u(x) \geq v(x) \qquad \forall x\in \mathbb{T}_m. \] \end{lem} \begin{proof} Adding a positive constant $c$ to $u,$ we may assume that \begin{equation} \label{estric} \lim_{x\to \pi\in \partial \mathbb{T}_m} u (x) > \lim_{x\to z\in \partial \mathbb{T}_m} v (x) . \end{equation} Our goal is to show that in this case we have \[ u(x) \geq v(x) \quad \forall x\in\mathbb{T}_m \] (and then we obtain the result just by letting $c\to 0$). We argue by contradiction and assume that \[ M = \max_{x\in \mathbb{T}_m} ( v(x)-u(x) ) >0. \] Notice that the maximum is attained thanks to \eqref{estric}. Also thanks to \eqref{estric}, we have that $M$ is attained only in a finite set of nodes. Let ${x}$ be one of such nodes. From \eqref{eq.envolvente.44.compar.u} and \eqref{eq.envolvente.44.compar.v} we obtain \begin{align*} M= v(x) - u(x) \le& \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{v(y) +v(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ v(\hat{x}_0) + m v(y)}{m+1} \right\}\\ &-\min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ u(\hat{x}_0) + m u(y)}{m+1} \right\}.\\ \end{align*} From this inequality, using that $$ M\geq \left(\frac{v(y) +v(z)}2 \right) - \left( \frac{u(y) +u(z)}2 \right), \qquad \forall y,z\in\mathcal{S}(x_0)\, y\neq z, $$ and $$ M\geq \left(\frac{ v(\hat{x}_0) + m v(y)}{m+1} \right) - \left( \frac{ u(\hat{x}_0) + m u(y)}{m+1} \right), \qquad \forall y\in\mathcal{S}(x_0), $$ we get \begin{align*} M \le& \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{v(y) +v(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ v(\hat{x}_0) + m v(y)}{m+1} \right\}\\ &-\min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ u(\hat{x}_0) + m u(y)}{m+1} \right\} \leq M.\\ \end{align*} Hence, we obtain that there are two nodes $x_1$ and $x_2$ connected with $x$ (one of them can be the predecessor) such that \[ v(x_1)-u(x_1) = M,\qquad \text{and} \qquad v(x_2)-u(x_2) = M. \] Since this happens for every $x$ in the set of maximums of $v-u$ and this set is finite, we obtain a contradiction that shows that \[ u(x) \geq v(x), \] and proves the result. \end{proof} Now we will prove \eqref{eq:limb}. \begin{teo}\label{teo:limb} Let $g\colon[0,1]\to\mathbb{R}$ be a continuous function. Then \[ \lim_{x\to \pi\in \partial \mathbb{T}_m} u_g^* (x) = g(\psi(\pi)) \] for any $\pi\in\partial\mathbb{T}_m.$ \end{teo} \begin{proof} Let us start by observing that, for any constant $c$, $u\in\mathcal{C}(g)$ if only if $u+c\in\mathcal{C}(g+c).$ Therefore, without loss of generality, we may assume that $g$ is a nonnegative function. Let $\pi_0=(y_1,\dots,y_k,\dots)\in\partial\mathbb{T}_m.$ For any $n\in\mathbb{T}_m,$ there exist $z_n\in\mathbb{T}_m^n$ and $k_0$ such that $\psi(y_k)\in I_{z_n}$ for all $k\ge k_0.$ Now taking $c=\min\{g(t)\colon t\in I_{z_n}\}$ and $w_{n}=cu_{z_n}$ where $u_{z_n}$ is given by Lemma \ref{lema:caract}, we have that $w_{n}$ is a convex function such that \[ \lim_{x\to\pi}w_{n}(x)\le g(\psi(\pi)), \qquad \forall\pi \in\partial\mathbb{T}_m. \] Here, we are using that $g\ge 0$. Then, $w_{n}\in\mathcal{C}(g),$ and therefore $w_{n}(x)\le u^*_g(x)$ for any $x\in\mathbb{T}_m.$ In particular, $w_{n}(y_k)\le u^*_g(y_k)$ for any $k.$ Therefore, \[ \min\{g(t)\colon t\in I_{z_n}\} = \lim_{k\to\infty} w_{n}(y_k)\le \liminf_{k\to\infty}u_g^*(y_k). \] Taking the limit as $n\to \infty,$ we have \[ g(\psi(\pi_0))\le \liminf_{k\to\infty}u^*(y_k) \] since $g$ is a continuous function. Moreover, taking \[w^{n}(x)=a(1-u_{z_n})+bu_{z_n}=a+(b-a)w_n\] where $a=2\max\{g(t)\colon t\in[0,1]\}$ and $b=\max\{g(t)\colon t\in I_{z_n}\},$ we have that \begin{align*} &w^{n}(x)=a+(b-a)w_n(x)\\ &=a+(b-a)\min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{w_n(y) +w_n(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ w_n(\hat{x}_0) + m w_n(y)}{m+1} \right\}\\ &=\max \left\{ \max_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} a+\frac{(b-a)(w_n(y) +w_n(z))}2 ; \max_{y\in\mathcal{S}(x_0)} a+\frac{(b-a) (w_n(\hat{x}_0) + m w_n(y))}{m+1} \right\}\\ &=\max \left\{ \max_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{w^n(y) +w^n(z)}2 ; \max_{y\in\mathcal{S}(x_0)} \frac{w^n(\hat{x}) + m w^n(y)}{m+1} \right\}\\ &\ge\min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{w^n(y) +w^n(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{w^n(\hat{x}) + m w^n(y)}{m+1} \right\} \end{align*} for any $x\in\mathbb{T}_m$ and \[ g(\psi(\pi))\le \liminf_{z\to\pi}w^{n}(x_k) , \qquad \forall\pi\in\partial\mathbb{T}_m. \] Thus, by Lemma \ref{lema.compar.convexas}, for any $u\in\mathcal{C}(g)$ we have that $u(x)\le w^{n}(x)$ for any $x\in\mathbb{T}_m.$ Therefore $u_g^*(x)\le w^{n}(x)$ for any $x\in\mathbb{T}_m.$ In particular, $u_g^*(y_k)\le w^{n}(y_k) $ for any $k.$ Then \[ \limsup_{k\to\infty}u_g^*(y_k)\le \lim w_{n}(y_k) =\max\{g(t)\colon t\in I_{z_n}\}. \] Again, taking the limit as $n\to \infty,$ we have \[ \limsup_{k\to\infty}u_g^*(y_k)\le g(\psi(\pi_0)). \] Therefore, we conclude that \[ \lim_{k\to\infty}u^*(y_k)= g(\psi(\pi_0)). \] As $\pi_0\in\partial\mathbb{T}_m$ was arbitrary, we conclude \[ \lim_{x\to\pi_0}u^*(x)= g(\psi(\pi_0)) \] for any $\pi_0\in\partial\mathbb{T}_m.$ \end{proof} Now our next goal is to find the equation that $u^*_g$ verifies on $\mathbb{T}_m$. \begin{teo}\label{teo:largesol} Let $g\colon[0,1]\to\mathbb{R}$ be a continuous functions. The convex envelope $u^*_g$ is characterized as the largest solution to \begin{equation} \label{eq.envolvente} u (x) = \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u(\hat{x}) + m u(y)}{m+1} \right\}\quad\text{on }\mathbb{T}_m, \end{equation} that verifies \[ \limsup_{x\to \pi\in \partial \mathbb{T}_m} u (x) \leq g(\psi(\pi)). \] \end{teo} \begin{proof} Given $g\colon[0,1]\to\mathbb{R},$ by Lemmas \ref{lema:convenv1} and \ref{lema:convexeq} we get that $u^*_g$ verifies \eqref{subsol1}. Now, to see that we have an equality, we argue by contradiction. Assume that at some node $x_0\in\mathbb{T}_m,$ we have \[ u_g^* (x_0) < \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u_g^*(y) +u_g^*(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ u_g^*(\hat{x}_0) + m u_g^*(y)}{m+1} \right\} \] Taking $\delta>0$ such that \[ u_g^* (x_0) + \delta < \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u_g^*(y) +u_g^*(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u_g^*(\hat{x}) + m u_g^*(y)}{m+1} \right\} \] and consider \[ v(x) = \begin{cases} u^*_g (x) &\text{if }x\neq x_0,\\ u_g^* (x_0) + \delta&\text{if }x= x_0. \end{cases} \] Observe that $v$ verifies \eqref{subsol1}. Thus, by Lemma \ref{lema:convexeq}, $v$ is convex. In addition, we have that $v\in{\mathbb{C}}(g).$ Therefore \[ v(x)\le u_g^*(x)\quad\forall x\in \mathbb{T}_m, \] leading to a contradiction. This proves that $u^*_g$ is a solution to \eqref{eq.envolvente}. Finally, to see that $u^*_g$ is the largest solution to \eqref{eq.envolvente} that verifies \[ \limsup_{x\to \pi\in \partial \mathbb{T}_m} u^*_g (x) \leq g(\psi(\pi)), \] it is enough to define \[ \overline{u} (x) = \sup \left\{ u(x)\colon u\text{ verifies \eqref{eq.envolvente} and } \limsup_{x\to \pi\in \partial \mathbb{T}_m} u (x) \leq g(\psi(\pi)) \right\}. \] This function $\overline{u}$ trivially verifies \[ \overline{u} (x) \geq u^*_g (x) \quad x \in \mathbb{T}_m, \] just notice that $u^*_g$ belongs to the set defining $\overline{u}$. On the other hand, since $\overline{u}$ is a solution to \eqref{eq.envolvente}, by Lemma \ref{lema:convexeq}, we have that $\overline{u}$ is convex and therefore $\overline{u}\in\mathcal{C}(g).$ Then \[ \overline{u} (x) \leq u^*_g (x) \quad\forall x \in \mathbb{T}_m. \] We conclude that \[ u^*_g(x) = \overline{u} (x) = \sup \left\{ v(x) \colon u \text{ verifies \eqref{eq.envolvente} and } \limsup_{x\to \pi\in \partial \mathbb{T}_m} u (x) \leq g(\psi(\pi)) \right\}. \] \end{proof} Observe that by Lemma \ref{lema.compar.convexas} and Theorem \ref{teo:largesol}, for any continuous function $g\colon [0,1] \mapsto \mathbb{R}$, the equation defining the convex envelope has a unique solution that attains the datum $g$ continuously. \begin{teo} Let $g\colon[0,1] \mapsto \mathbb{R}$ be a continuous function. There exists a unique solution to \eqref{eq.envolvente} such that \[ \lim_{x\to \pi \in \partial \mathbb{T}_m} u (x) = g(\psi(\pi)). \] for any $\pi\in\mathbb{T}_m.$ \end{teo} To end this section we prove Theorem \ref{teo.convex.arbol.f} that deals with the convex envelope of a function $f:\mathbb{T}_m \mapsto \mathbb{R}$ given by \eqref{convex-envelope-arbol.f}. \begin{teo} \label{teo.convex.arbol.f.sec} The convex envelope of a function $f\colon \mathbb{T}_m \to\mathbb{R}$ is the solution to the obstacle problem for the equation \eqref{eq.tree.convex}. \end{teo} \begin{proof} Let us call $v^*$ to the largest solution to \begin{equation} \label{eq.tree.convex.ffffhhh} u (x) \leq \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u(\hat{x}) + m u(y)}{m+1} \right\}\quad\text{on }\mathbb{T}_m, \end{equation} that verifies \[ u(x) \leq f(x) \qquad \forall x \in \mathbb{T}_m. \] We have to prove that the convex enevolpe of $f$, $u^*_f$, verifies \[ u^*_f (x) = v^* (x), \qquad \forall x \in \mathbb{T}_m. \] Since $u^*_f$ is convex from Lemma \ref{lema:convexeq} we obtain that it is a solution to \eqref{eq.tree.convex.ffffhhh} that verifies $u^*_f \leq f$ on $\mathbb{T}_m$ and then we obtain \[ u^*_f (x) \leq v^* (x), \qquad \forall x \in \mathbb{T}_m. \] We have also from Lemma \ref{lema:convexeq} that $v^*$ being a solution to \eqref{eq.tree.convex.ffffhhh} is a convex function and it verifies $v^* \leq f$ on $\mathbb{T}_m$. Hence, \[ v^* (x) \leq u^*_f (x), \qquad \forall x \in \mathbb{T}_m. \] We conclude that \[ u^*_f (x) = v^* (x), \qquad \forall x \in \mathbb{T}_m. \] In the coincidence set, the function $f$ verifies an inequality. From the fact that $u^*_f$ is convex and smaller than $f$ we obtain for $x \in CS(f)$, \begin{align} \label{eq.tree.convex.ffff.latengo.kkk} f (x) & = u^*_f(x) \\ & \leq \displaystyle \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u^*_f(y) + u^*_f(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ u^*_f(\hat{x}) + m \, u^*_f (y)}{m+1} \right\} \\[10pt] & \displaystyle \leq \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{f(y) + f(z)}2 ; \min_{y\in\mathcal{S}(x)} \frac{ f(\hat{x}) + m f (y)}{m+1} \right\}. \end{align} Finally, outside of the coincidence set the convex envelope, $u^*_f$, is a solution to the equation. In fact, arguing by contradiction, assume that for some $x_0 \not\in CS(f)$ it holds \begin{equation} \label{eq.tree.convex.ffff.latengo.mmmm.llll} u^*_f (x_0) < \min \left\{ \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u^*_f(y) +u^*_f(z)}2 ; \min_{y\in\mathcal{S}(x_0)} \frac{ u^*_f(\hat{x_0}) + m u^*_f(y)}{m+1} \right\}. \end{equation} Then, since $x_0 \not\in CS(f)$ and we have a strict inequality in \eqref{eq.tree.convex.ffff.latengo.mmmm.llll} there exists $\delta>0$ such that the function \[ v(x) = \begin{cases} u^*_f (x) &\text{if }x\neq x_0,\\ u_f^* (x_0) + \delta&\text{if }x= x_0 \end{cases} \] is convex and still verifies $v\leq f$ on $\mathbb{T}_m$ contradicting the maximality of the convex envelope $u^*_f$. This contradiction shows that we have an equality in \eqref{eq.tree.convex.ffff.latengo.mmmm.llll}. \end{proof} \section{Binary convex functions} \label{section.biconvfunction} As in Section \ref{sect-convex}, we begin showing a different characterization of binary convex functions. \begin{lem}\label{lema:biconvexeq} A function $u$ on the tree is binary convex if and only if $u$ satisfies \begin{equation} \label{subsol2} u (x) \le \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 \qquad \forall x\in \mathbb{T}_m. \end{equation} \end{lem} \begin{proof} Let us start the proof observing that if $x\in \mathbb{T}_m,$ $y,z\in \S(x)$ and $y\neq z$ then $[y,z]\in\mathbb{T}_{2}^{x}$ and $\mathcal{E}([y,z])=\{y,z\}.$ Therefore if $u$ is a binary convex function, $x\in\mathbb{T}_m$ and $y,z\in \S(x)$ are such taht $y\neq z$ then \[ u(x)\le \dfrac{u(y)+u(z)}{2}. \] Thus, $u$ satisfies \eqref{subsol2} in $\mathbb{T}_m.$ Now assume that $u$ satisfies \eqref{subsol2}. Our aim is to prove that $u$ is a binary convex function, that is, we aim to show that \begin{equation}\label{eq:biconvdef} u(x)\le \sum_{y\in\mathcal{E}(\mathbb{B})} \dfrac{u(y)}{2^{|y|-|x|}}\quad \forall \mathbb{B}\in\mathbb{T}_{2}^{x}. \end{equation} Fix $x\in \mathbb{T}_m.$ Given $\mathbb{B}\in\mathbb{T}_{2}^{x},$ we define \[ |\mathbb{B}|\coloneqq \max \big\{|z|-|x|\colon z\in\mathcal{E}(\mathbb{B})\big\}\in\mathbb{N} \] and \[ \mathbb{T}_{2}^{x n} \coloneqq \big\{\mathbb{B}\in\mathbb{T}_{2}^{x}\colon |\mathbb{B}|=n\big\}\subset\mathbb{T}_{2}^{x}. \] The proof of \eqref{eq:biconvdef} runs by induction in $n.$ Observe that in the case $|\mathbb{B}|=1$ there exist $y,z\in\S(x)$ such that $\mathbb{B}=[y,z]$ and obviously $\mathcal{E}(\mathbb{B})=\{y,z\}.$ Then, since $u$ satisfies \eqref{subsol2}, we get \[ u(x)\le\dfrac{u(y)+u(z)}2= \sum_{y\in\mathcal{E}(\mathbb{B})} \dfrac{u(y)}{2^{|y|-|x|}}. \] That is \eqref{eq:biconvdef} holds for any $\mathbb{B}\in\mathbb{T}_{2}^{x 1}.$ Now we assume that \eqref{eq:biconvdef} holds for any $\mathbb{B}\in\mathbb{T}_{2}^{x n},$ and we will show that it also holds for any $\mathbb{B}\in\mathbb{T}_{2}^{x (n+1}.$ If $\mathbb{B}\in\mathbb{T}_{2}^{x (n+1)}$ then $\mathbb{B}'=\mathbb{B}\setminus \{y\in\mathcal{E}(\mathbb{B})\colon |y|-|x|=n+1\}\in\mathbb{T}_{2}^{x n}.$ Then, by the inductive hypothesis, we get \begin{equation}\label{eq:biconaux1} u(x)\le\sum_{y\in\mathcal{E}(\mathbb{B}')} \dfrac{u(y)}{2^{|y|-|x|}}. \end{equation} On the other hand, for any $y\in\mathcal{E}(\mathbb{B})$ we have that $y\in\mathcal{E}(\mathbb{B}')$ or there are $w\in \mathcal{E}(\mathbb{B}')$ and $z\in\mathcal{E}(\mathbb{B})\setminus\{y\}$ such that $y,z\in \S(w).$ Thus, since $u$ satisfies \eqref{subsol2}, from \eqref{eq:biconaux1}, we have that \[ u(x)\le\sum_{y\in\mathcal{E}(\mathbb{B})} \dfrac{u(y)}{2^{|y|-|x|}}. \] Finally, since $x$ is arbitrary, we conclude that $u$ is a binary convex function. \end{proof} \begin{re}\label{re:cfisbcf} Now, by Lemmas \ref{lema:convexeq} and \ref{lema:biconvexeq}, it is easy to check that a convex function is also a binary convex function. \end{re} Proceeding as in the proof of Corollary \ref{co:sumconv} we can prove the following result. \begin{co}\label{co:sumbiconv} Let $u,v\colon \mathbb{T}_m\to\mathbb{R}$ be binary convex functions. Then $u+v$ is a binary convex function. \end{co} Now, we obtain the following result, whose proof is similar to Lemma \ref{lema:caract}. \begin{lem}\label{lema:bicaract} Let $x_0\in\mathbb{T}_m\setminus\{\emptyset\}.$ Then the function $u_{x_0}\colon\mathbb{T}_m\to\mathbb{R}$ defined by \[ u_{x_0}(x)\coloneqq \begin{cases} 0 &\text{if } x\not\in\mathbb{T}_m^{x_0},\\ 1&\text{if } x\in\mathbb{T}_m^{x_0}, \end{cases} \] is a binary convex function such that \[ \limsup_{x\to\pi}u_{x_0}(x) \le \mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)) \quad\forall \pi\in\partial\mathbb{T}_m. \] Moreover, \[ u_{x_0}(x)= \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u_{x_0}(y) +u_{x_0}(z)}2 \quad\forall x\in\mathbb{T}_m, \] and for any $\pi\in\mathbb{T}_m$ such that $\psi(\pi)$ is not an endpoint of $I_{x_0}$ we have \[ \lim_{x\to\pi}u_{x_0}(x) =\mbox{\Large$\chi$}_{I_{x_0}}(\psi(\pi)). \] \end{lem} \section{Binary Convex envelopes} \label{sect-biconvex-envelopes} Proceeding as in the proof of Lemma \ref{lema:convenv1} we can show that the binary convex envelop is a binary convex function. \begin{lem} \label{lema:biconvenv1} For any function $g\colon[0,1]\to\mathbb{R},$ the binary convex envelop $\tilde{u}_g$ is a binary convex function. \end{lem} In a similar way to Section \ref{sect-convex-envelopes}, we will show that if $g$ is a continuous function, then \begin{equation} \label{eq:bilimb} \lim_{x\to \pi\in \partial \mathbb{T}_m} \tilde{u}_g (x) = g(\psi(\pi))\quad\forall\pi\in\partial\mathbb{T}_m. \end{equation} As before, to prove this claim we need a comparison principle. \begin{lem} \label{lema.bicompar.convexas} Let $u$ and $v$ satisfy \begin{align} \label{eq.bienvolvente.44.compar.u} u(x) &\ge \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}}\frac{u(y) +u(z)}2 \quad\forall x\in\mathbb{T}_m,\\ \label{eq.bienvolvente.44.compar.v} v(x) &\le \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{v(y) +v(z)}2 , \quad\forall x\in\mathbb{T}_m, \end{align} with \[ \lim_{x\to \pi\in \partial \mathbb{T}_m} u (x) \geq \lim_{x\to \pi\in \partial \mathbb{T}_m} v (x) , \] for every $\pi\in \partial \mathbb{T}_m$. Then, \[ u(x) \geq v(x) \qquad \forall x\in \mathbb{T}_m. \] \end{lem} \begin{proof} Adding a positive constant $c$ to $u,$ we may assume that \begin{equation} \label{biestric} \lim_{x\to \pi\in \partial \mathbb{T}_m} u (x) > \lim_{x\to z\in \partial \mathbb{T}_m} v (x) . \end{equation} We argue by contradiction, so, assume that \[ M = \max_{x\in \mathbb{T}_m} (v(x)-u(x)) >0. \] Notice that the maximum is attained thanks to \eqref{biestric}. Also by \eqref{biestric}, we have that $M$ is attained only in a finite set of vertices. Let ${x}$ be one of such vertices. From \eqref{eq.bienvolvente.44.compar.u} and \eqref{eq.bienvolvente.44.compar.v} we obtain \[ M= v(x) - u(x) \le \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}}\frac{v(y) +v(z)}2 -\min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u(y) +u(z)}2. \] Now, using that $$ M\geq \left(\frac{v(y) +v(z)}2 \right) - \left( \frac{u(y) +u(z)}2 \right), \qquad \forall y,z\in\mathcal{S}(x_0)\, y\neq z, $$ we get \[ M \le \min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}}\frac{v(y) +v(z)}2 -\min_{\substack{y,z\in\mathcal{S}(x_0)\\ y\neq z}} \frac{u(y) +u(z)}2 \leq M. \] Then, there exist two nodes $x_1, x_2\in\S(x)$ such that \[ v(x_1)-u(x_1) = M,\qquad \text{and} \qquad v(x_2)-u(x_2) = M. \] Since this happens for every $x$ in the set of maximums of $v-u$ and this set is finite, we obtain a contradiction that shows that \[ u(x) \geq v(x) \] and proves the result. \end{proof} Now we will show \eqref{eq:bilimb}. \begin{teo}\label{teo:bilimb} Let $g\colon[0,1]\to\mathbb{R}$ be a continuous function. Then, for any $\pi\in\partial\mathbb{T}_m$ \[ \lim_{x\to \pi\in \partial \mathbb{T}_m} \tilde{u}_g (x) = g(\psi(\pi)). \] \end{teo} \begin{proof} Let us start by observing that, for any constant $c$, $u\in\mathcal{B}(g)$ if only if $u+c\in\mathcal{B}(g+c).$ Therefore, without loss of generality, we may assume that $g$ is nonnegative. Consequently, by Remark \ref{re:cfisbcf} and Theorem \ref{teo:limb}, we have \[ \liminf_{x\to \pi\in \partial \mathbb{T}_m} \tilde{u}_g (x) \ge \lim_{x\to \pi\in \partial \mathbb{T}_m} u^*_g (x)= g(\psi(\pi)) \] for any $\pi\in\partial\mathbb{T}_m.$ To complete the proof, we proceed as in the end of the proof of Theorem \ref{teo:limb}, using Lemmas \ref{lema:bicaract} and \ref{lema.bicompar.convexas} instead of Lemmas \ref{lema:caract} and \ref{lema.compar.convexas}. \end{proof} Finally, with analogous arguments of the Section \ref{sect-convex-envelopes}, we get the following results. \begin{teo}\label{teo:bilargesol} Let $g\colon[0,1]\to\mathbb{R}$ be a continuous functions. The binary convex envelope $\tilde{u}_g$ is characterized as the largest solution to \begin{equation} \label{eq.bienvolvente} u (x) = \min_{\substack{y,z\in\mathcal{S}(x)\\ y\neq z}} \frac{u(y) +u(z)}2 \quad\text{on }\mathbb{T}_m, \end{equation} that verifies \[ \limsup_{x\to \pi\in \partial \mathbb{T}_m} u (x) \leq g(\psi(\pi)). \] \end{teo} \begin{teo} Let $g\colon [0,1] \mapsto \mathbb{R}$ be a continuous function. Then, there exists a unique solution to \eqref{eq.bienvolvente} such that for any $\pi\in\mathbb{T}_m$ \[ \lim_{x\to \pi \in \partial \mathbb{T}_m} u (x) = g(\psi(\pi)). \] \end{teo} \medskip {\bf Acknowledgements.} \ Supported by CONICET grant PIP GI 11220150100036CO (Argentina), by UBACyT grant 20020160100155BA (Argentina) and by MINECO MTM2015-70227-P (Spain).

https://arxiv.org/abs/1601.05525

On Drury's solution of Bhatia \& Kittaneh's question

Let $A, B$ be $n\times n$ positive semidefinite matrices. Bhatia and Kittaneh asked whether it is true $$ \sqrt{\sigma_j(AB)}\le \frac{1}{2} \lambda_j(A+B), \qquad j=1, \ldots, n$$ where $\sigma_j(\cdot)$, $\lambda_j(\cdot)$, are the $j$-th largest singular value, eigenvalue, respectively. The question was recently solved by Drury in the affirmative. This article revisits Drury's solution. In particular, we simplify the proof for a key auxiliary result in his solution.

\section{Introduction} Bhatia has made many fundamental contributions to Matrix Analysis \cite{Bha97}. One of his favorite topics is matrix inequalities. Roughly speaking, matrix inequalities are noncommutative versions of the corresponding scalar inequalities. To get a glimpse of this topic, let us start with a simple example. The simplest AM-GM inequality says that $$a, b>0 \implies \frac{a+b}{2}\ge \sqrt{ab}.$$ Now it is known that \cite[p. 107]{Bha07} its most ``direct" noncommutative version is \begin{eqnarray}\label{am-gm} A, B ~ \hbox{are $n\times n$ positive definite matrices} \implies \frac{A+B}{2}\ge A\sharp B, \end{eqnarray} where $A\sharp B:=A^{\frac{1}{2}}(A^{-\frac{1}{2}}BA^{-\frac{1}{2}})^{\frac{1}{2}}A^{\frac{1}{2}}$ is called the geometric mean of $A$ and $B$. For two Hermitian matrices $A$ and $B$ of the same size, in this article, we write $A\ge B$ (or $B\le A$) to mean that $A-B$ is positive semidefinite. If we denote $S:=A\sharp B$, then $B=SA^{-1}S$. Thus a variant of (\ref{am-gm}) is the following \begin{eqnarray}\label{am-gm1} A, S ~ \hbox{are $n\times n$ positive definite matrices} \implies A+SA^{-1}S\ge 2S. \end{eqnarray} There is a long tradition in matrix analysis of comparing eigenvalues or singular values. To proceed, let us fix some notation. The $j$-th largest singular value of a complex matrix $A$ is denoted by $\sigma_j(A)$. If all the eigenvalues of $A$ are real, then we denote its $j$-th largest one by $\lambda_j(A)$. By Weyl's Monotonicity Theorem \cite[p. 63]{Bha97}, (\ref{am-gm}) readily implies \begin{eqnarray*} \lambda_j(A+B)\ge 2 \lambda_j(A\sharp B), \qquad j=1, \ldots, n. \end{eqnarray*} As far as the eigenvalues or singular values are considered, there are other versions of ``geometric mean". Bhatia and Kittaneh studied this kind of inequalities over a twenty year period \cite{BK90, BK00, BK08}. Their elegant results include the following: If $A, B$ are $n\times n$ positive semidefinite matrices, then \begin{eqnarray}\label{bk1} && \lambda_j(A+B)\ge 2\sqrt{\lambda_j(AB)}=2\sigma_j(A^{\frac{1}{2}}B^{\frac{1}{2}}); \\&& \label{bk2} \lambda_j(A+B)\ge 2\lambda_j(A^{\frac{1}{2}}B^{\frac{1}{2}}) \end{eqnarray} for $j=1, \ldots, n$. To complete the picture in (\ref{bk1})-(\ref{bk2}), they asked whether it is true \begin{eqnarray*}\lambda_j(A+B)\ge 2\sqrt{\sigma_j(AB)}, \qquad j=1, \ldots, n? \end{eqnarray*} This question was recently answered in the affirmative by Drury in his very brilliant work \cite{Dru12}. The purpose of this expository article is to revisit Drury's solution. Hopefully, some of our arguments would shed new insights into the beautiful result, which is now a theorm. \begin{thm}\cite{Dru12} If $A, B$ are $n\times n$ positive definite semidefinite matrices, then \begin{eqnarray} \label{bkd} \lambda_j(A+B)\ge 2\sqrt{\sigma_j(AB)}, \qquad j=1, \ldots, n. \end{eqnarray} \end{thm} \section{Drury's reduction in proving (\ref{bkd})} Our presentation here is just slightly different from that in \cite{Dru12}. Assume without loss of generality that $A, B$ are positive definite (the general case is by a standard purturbation argument). Fix $r$ in the range $1\le r\le n$ and normalize so that $\sigma_r(AB)=1$. Our goal is to show that $\lambda_r(A+B)\ge 2$. Note that $\sigma_r(AB)=1$ is the same as $\lambda_r(AB^2A)=1$. Consider the spectral decomposition $$AB^2A=\sum_{k=1}^n\lambda_k(AB^2A)P_k,$$ where $P_1, P_2, \ldots, P_n$, are orthogonal projections. Then $\lambda_k(AB^2A)\ge 1$ for $k=1, \ldots, r$. Define a positive semidefinite $$B_1:=\left(A^{-1}\left(\sum_{k=1}^r P_k\right)A^{-1}\right)^{1/2}.$$ It is easy to see (indeed, from $B^2\ge B_1^2$) that $$B=\left(A^{-1}\left(\sum_{k=1}^r\lambda_k(AB^2A) P_k\right)A^{-1}\right)^{1/2}\ge B_1.$$ So we are done if we can show \begin{eqnarray}\label{reduction1} \lambda_r(A+B_1)\ge 2. \end{eqnarray} As $B_1$ has rank $r$, split the underlying space as the direct sum of image and kernel of $B_1$, we may partition comformally $B_1$ and $A$ in the following form $$B_1=\begin{pmatrix} X & 0 \\0& 0 \end{pmatrix}, \ A=\begin{pmatrix} A_{11} & A_{12}\\ A_{12}^*& A_{22} \end{pmatrix}.$$ Note $AB_1^2A$ is an orthogonal projection of rank $r$, the same is true for $B_1A^2B_1$. Therefore, $$B_1A^2B_1=\begin{pmatrix} X(A_{11}^2+A_{12}A_{12}^*)X & 0\\ 0& 0 \end{pmatrix}\implies X(A_{11}^2+A_{12}A_{12}^*)X=I_r$$ where $I_r$ is the $r\times r$ identity matrix. Finally, observe that $$A\ge A_1:=\begin{pmatrix} A_{11} & A_{12}\\ A_{12}^*& A_{12}^*A_{11}^{-1}A_{12} \end{pmatrix}.$$ Therefore, (\ref{reduction1}) would follow from \begin{eqnarray}\label{reduction2} \lambda_r(A_1+B_1)\ge 2. \end{eqnarray} Thus, the remaining effort is made to show (\ref{reduction2}), which we formulate as a proposition. \begin{proposition}\label{p1} Let $A_{11}$ and $X$ be $r\times r$ positive definite matrices and $A_{12}$ is an $(n-r)\times (n-r)$ matrix such that $X(A_{11}^2+A_{12}A_{12}^*)X=I_r$. Then \begin{eqnarray}\label{reduction3} \lambda_r\begin{pmatrix} A_{11}+X & A_{12}\\ A_{12}^*& A_{12}^*A_{11}^{-1}A_{12} \end{pmatrix}\ge 2. \end{eqnarray} \end{proposition} \section{The mystified part} In order to prove (\ref{reduction3}), Drury made the following key observations. \begin{proposition}\label{p2}\cite[Proposition 2]{Dru12} Let $M$ and $N$ be $r\times r$ positive definite matrices. Then \begin{eqnarray*} \lambda_r\begin{pmatrix} M &(M\sharp N)^{-1}\\ (M\sharp N)^{-1}& N \end{pmatrix}\ge 2. \end{eqnarray*} \end{proposition} \begin{proposition}\label{p3}\cite[Theorem 7]{Dru12} Let $L$ and $M$ be $r\times r$ positive definite matrices, and let $Z$ be an $r\times r$ matrix such that $ML(I+ZZ^*)LM=I_r$. Then \begin{eqnarray} \label{reduction4} \lambda_r\begin{pmatrix} L+M & LZ\\ Z^*L& Z^*LZ \end{pmatrix}\ge 2. \end{eqnarray} \end{proposition} The way that Drury proved (\ref{reduction4}) is by showing that $T:=\begin{pmatrix} L+M & LZ\\ Z^*L& Z^*LZ \end{pmatrix}$ and $R:=\begin{pmatrix} M &(M\sharp N)^{-1}\\ (M\sharp N)^{-1}& N \end{pmatrix}$ have the same characteristic polynomial, and so the eigenvalues of $R$ and $T$ coincide. As explained in \cite{Dru12a}, this connection (between $R$ and $T$) is mystified. Formally, the mystified part also comes from $R$ and $T$ themselves, indeed, $T$ is always positive semidefinite while $R$ is not! In order to apply Proposition \ref{p3} to Proposition \ref{p1}, Drury discussed three possible relations between the size $n$ and $r$. Our proof of Proposition \ref{p1} in the next section allows us to skip this discussion on the size. \section{Proof of Proposition \ref{p1}} The following lemma slightly generalizes Proposition \ref{p2} in form. \begin{lemma} \label{lem1} Let $X$ be a $r\times r$ positive definite matrix and let $S$ be a $r\times r$ nonsingular matrix. Then \begin{eqnarray*} \lambda_r\begin{pmatrix} SX^{-1}S^*& (S^{-1})^*\\S^{-1}& X \end{pmatrix}\ge 2. \end{eqnarray*} \end{lemma} \begin{proof} Consider the polar decomposition of $S$, $S=U|S|$, where $U$ is unitary and $|S|=(S^*S)^{\frac{1}{2}}$. The matrix $\begin{pmatrix} SX^{-1}S^*& (S^{-1})^*\\S^{-1}& X \end{pmatrix}$ is unitarily similar to $$\begin{pmatrix}U^* SX^{-1}S^*U& U^*(S^{-1})^*\\S^{-1}U& X \end{pmatrix}=\begin{pmatrix} |S|X^{-1}|S|& |S|^{-1}\\|S|^{-1}& X \end{pmatrix}.$$ As $P:=\frac{1}{2}\begin{pmatrix} I_r\\ I_r \end{pmatrix}$ is a partial isometry, \begin{eqnarray*} \lambda_r\begin{pmatrix} SX^{-1}S^*& (S^{-1})^*\\S^{-1}& X \end{pmatrix}&=&\lambda_r\begin{pmatrix} |S|X^{-1}|S|& |S|^{-1}\\|S|^{-1}& X \end{pmatrix}\\&\ge& \lambda_r \left(P^*\begin{pmatrix} |S|X^{-1}|S|& |S|^{-1}\\|S|^{-1}& X \end{pmatrix}P\right)\\&=&\lambda_r\left(\frac{X+|S|X^{-1}|S|}{2}+|S|^{-1}\right)\\&\ge& \lambda_r(|S|+|S|^{-1}) \ge 2. \qquad \hbox{by (\ref{am-gm1})} \end{eqnarray*} The required result follows. \end{proof} Now we are ready to give a simpler proof of Proposition \ref{p1}. ~ \noindent {\it Proof. } Consider the factorization $$\begin{pmatrix} A_{11}+X & A_{12}\\ A_{12}^*& A_{12}^*A_{11}^{-1}A_{12} \end{pmatrix}=\begin{pmatrix} A_{11}^{\frac{1}{2}}&X^{\frac{1}{2}}\\ A_{12}^*A_{11}^{-\frac{1}{2}}& 0 \end{pmatrix}\begin{pmatrix} A_{11}^{\frac{1}{2}}&X^{\frac{1}{2}}\\ A_{12}^*A_{11}^{-\frac{1}{2}}& 0 \end{pmatrix}^*.$$ Clearly, $\begin{pmatrix} A_{11}+X & A_{12}\\ A_{12}^*& A_{12}^*A_{11}^{-1}A_{12} \end{pmatrix}$ is unitarily similar to \begin{eqnarray*} \begin{pmatrix} A_{11}^{\frac{1}{2}}&X^{\frac{1}{2}}\\ A_{12}^*A_{11}^{-\frac{1}{2}}& 0 \end{pmatrix}^*\begin{pmatrix} A_{11}^{\frac{1}{2}}&X^{\frac{1}{2}}\\ A_{12}^*A_{11}^{-\frac{1}{2}}& 0 \end{pmatrix}&=&\begin{pmatrix} A_{11}+A_{11}^{-\frac{1}{2}}A_{12}A_{12}^*A_{11}^{-\frac{1}{2}}& A_{11}^{\frac{1}{2}}X^{\frac{1}{2}}\\ X_{11}^{\frac{1}{2}}A^{\frac{1}{2}}& X \end{pmatrix}\\&=&\begin{pmatrix} A_{11}^{-\frac{1}{2}}X^{-2}A_{11}^{-\frac{1}{2}}& A_{11}^{\frac{1}{2}}X^{\frac{1}{2}}\\ X_{11}^{\frac{1}{2}}A^{\frac{1}{2}}& X \end{pmatrix}. \end{eqnarray*} Now setting $S=A_{11}^{-\frac{1}{2}}X^{-\frac{1}{2}}$ in Lemma \ref{lem1} yields the desired result. \qed \section{A conjecture} A weighted version of (\ref{bk1}) is known. That is, if $A, B$ are $n\times n$ positive semidefinite matrices, then for any $t\in [0, 1]$ and $j=1, \ldots, n$ \begin{eqnarray}\label{ando} && \lambda_j((1-t)A+tB)\ge \sqrt{\lambda_j(A^{2(1-t)}B^{2t})}=\sigma_j(A^{1-t}B^{t}). \end{eqnarray} Inequality (\ref{ando}) is due to Ando \cite{Ando95}. With \ref{ando}), it is not hard to present a weighted version of (\ref{bk2}). \begin{proposition}\label{p4} If $A, B$ are $n\times n$ positive semidefinite matrices, then for any $t\in [0, 1]$ and $j=1, \ldots, n$ \begin{eqnarray}\label{lin} \lambda_j((1-t)A+tB)\ge \lambda_j(A^{1-t}B^{t}). \end{eqnarray} \end{proposition} \begin{proof} By (\ref{ando}) and the matrix convexity of the square function, \begin{eqnarray*} \lambda_j(A^{1-t}B^{t})&=&\sigma_j^2(A^{(1-t)/2}B^{t/2})\\&\le&\lambda_j((1-t)A^{1/2}+tB^{1/2})^2 \\ &\le& \lambda_j((1-t)A+tB). \end{eqnarray*} \end{proof} We conclude the paper with the following conjecture \begin{conj} If $A, B$ are $n\times n$ positive definite semidefinite matrices, then for any $t\in [0, 1]$ \begin{eqnarray*} \lambda_j((1-t)A+tB)\ge \sqrt{\sigma_j(A^{2(1-t)}B^{2t})}, \qquad j=1, \ldots, n. \end{eqnarray*} \end{conj} The present method of proof does not seem to lead to a solution of this conjecture. \section*{Acknowledgement} The author thanks some helpful conversations with T. Ando and P. van den Driessche.

https://arxiv.org/abs/math/0508396

Group actions in number theory

Students having had a semester course in abstract algebra are exposed to the elegant way in which finite group theory leads to proofs of familiar facts in elementary number theory. In this note we offer two examples of such group theoretical proofs using the action of a group on a set. The first is Fermat's little theorem and the second concerns a well known identity involving the famous Euler phi function. The tools that we use to establish both results are sometimes seen in a second semester algebra course in which group actions are studied. Specifically, we will use the class equation of a group action and Burnside's theorem.

\section{Introduction} Students having had a semester course in abstract algebra are exposed to the elegant way in which finite group theory leads to proofs of familiar facts in elementary number theory. In this note we offer two examples of such group theoretical proofs using the action of a group on a set. The first is Fermat's little theorem and the second concerns a well known identity involving the famous Euler phi function. The tools that we use to establish both results are sometimes seen in a second semester algebra course in which group actions are studied. Specifically, we will use the class equation of a group action and Burnside's theorem. \section{Fermat's Little Theorem} A well known consequence of the class equation of a group action asserts that if $G$ is a $p$-group (that is, $G$ is a finite group of order $p^n$ for some integer $n\ge 1$ and a prime integer $p$), and $G$ acts on a finite set $S$, then the number of elements in $S$ is congruent to the number of fixed points of the action modulo $p$. Recall that an element $s\in S$ is a fixed point of the action if $gs=s$ for all $g\in G$. The set of fixed points is usually denoted by $S^G$, and so using this notation, the aforementioned theorem asserts that \begin{equation} \label{fact} |S|\equiv |S^G|\pmod p. \end{equation} This seemingly obscure result appears to have limitless utility in group theory! A line of argumentation due to R.~J.~Nunke (c.f~\cite{Fraleigh,Hungerford}) uses (\ref{fact}) repeatedly to establish the three famous Sylow theorems in elementary group theory. In this section, we use (\ref{fact}) to obtain a new proof of the following well known result in number theory. \begin{theorem}[Fermat's little theorem]\label{fermat} If $a\ge 1$ is any integer and $p$ is a prime, then $a^p\equiv a \pmod p$. \end{theorem} {\bf Proof.} Let $a\ge 1$ be an integer and let $A=\{1,\cdots,a\}$. Let $S=A^p$ and let a cyclic group $G$ of order $p$ act on $S$ by cyclic permutation of the entries of an element in $S$. Easily, this is a well defined action and $(a_1,\cdots,a_p)\in S$ is a fixed point if and only if $a_1=\cdots = a_p$. Therefore $|S^G|=a$. Since $|S|=a^p$, an application of (\ref{fact}) completes the proof. {\hfill\vrule height 5pt width 5pt depth 0pt} Of course, Fermat's little theorem holds for all integers $a$, but the construction in the our proof is not valid for $a\le 0$. The case $a=0$ is trivial, and if $a\le -1$, then $-a\ge 1$ and what we have proved so far shows that $(-a)^p\equiv (-a)\pmod p$. But $(-a)^p=-a^p$ and hence $a^p\equiv a\pmod p$. We remark here that we do not believe the above proof of Theorem~\ref{fermat} is better than the standard group theoretic proof. Indeed, the ideas in it are seldom encountered in a first semester undergraduate level course in algebra. However, the non-zero elements of $\mathbb {Z}_{p^j}$ do not form a group under multiplication if $j>1$, so that the standard argument does not generalize to the case of a power of a prime. On the other hand, our method immediately gives a proof of this case as well. \begin{theorem} If $a\ge 1$ is any integer and $p$ is a prime, then $a^{p^j}\equiv a \pmod p$ for all $j\ge 1$. \end{theorem} {\bf Proof.} We note that a cyclic group $G$ of order $p^j$ acts on the set $A^{p^j}$ cyclically, and there are still precisely $a$ fixed points. This time, however, there are $a^{p^j}$ total elements in the set. {\hfill\vrule height 5pt width 5pt depth 0pt} \section{The Euler Phi Function} If $n\ge 1$ is an integer, we denote by $\varphi(n)$ the number of elements in the set $\{1,\dots, n\}$ that are relatively prime to $n$. The function $\varphi$ is called the {\it Euler phi function}. In this section, we use group actions to give a proof of the following well known result. \begin{theorem}\label{second} If $n\ge 1$ is an integer, then $\displaystyle \sum_{d|n}\varphi(d)=n$. \end{theorem} In contrast to Fermat's little theorem, this result is not typically discussed in algebra classes. The usual proof given in number theory exploits the fact that the function $\varphi$ is {\it multiplicative} (that is, satisfies $\varphi(ab)=\varphi(a)\varphi(b)$ when ever $a$ and $b$ are relatively prime integers). Our method here will employ Burnside's theorem as well as the structure of the lattice of subgroups of a finite cyclic group. If a finite group $G$ acts on a finite set $S$, then for each $g\in G$, we let $S^g=\{s\in S : gs=s\}$ denote the set of elements in $S$ left fixed by $g$. If $r$ denotes the number of orbits in $S$ under the action of $G$, Burnside's theorem states that \begin{equation} \label{burnside} r\cdot |G|=\sum_{g\in G} |S^g|. \end{equation} We refer the reader to \cite{Fraleigh} for an excellent account of the details. To establish Theorem~\ref{second}, we consider the problem of counting the number of distinguishable ways of coloring the edges of a regular $n$-gon ($n\ge 3$) with $q$ colors ($q\ge 1$). The dihedral group $D_n$ of order $2n$ has a natural action on a regular $n$-gon as the group of symmetries. Under this action, two colorings of the $n$-gon are indistinguishable if and only if they belong to the same orbit under the action. Therefore the solution to our counting problem is the number of distinct orbits under this action. For notation, we let $D_n=\langle a,b | a^n=1, b^2=1, ba=a^{-1}b\rangle$. We refer to an element of the cyclic subgroup $\langle a\rangle$ as a {\it rotation} and an element of the coset $b\langle a\rangle$ as a {\it flip}. To use Burnside's theorem, we must compute $|S^g|$ for all $g\in D_n$ where $S$ is the set of all $q^n$ possible colorings of the $n$-gon. If $g$ is a flip and $n$ is odd, then the line of reflection for $g$ must pass through a vertex of the $n$-gon and the midpoint of the edge opposite this vertex. If a coloring $s$ is fixed under $g$, this opposite edge may be colored any one of the $q$ colors, but the remaining $(n-1)/2$ edges must be colored the same as their image under $g$. Therefore there are $qq^{(n-1)/2}=q^{(n+1)/2}$ colorings fixed by $g$. It follows that if $n$ is odd, then \[\sum_{g\in b\langle a\rangle} |S^g|=n\left (q^{(n+1)/2}\right).\] If $n$ is even, then there are $q^{n/2}$ colorings fixed by $g$ if $g$ is a flip in a line through opposite vertices and there are $q^{(n+2)/2}$ colorings fixed by $g$ if $g$ is a flip in a line through the midpoints of opposite edges. Since there are exactly $n/2$ of each of these types of flips, if $n$ is even we have \[\sum_{g\in b\langle a\rangle} |S^g|=\frac{n}{2}\left (q^{n/2}\right)+\frac{n}{2}q^{(n+2)/2}=\frac{n}{2}q^{n/2}(q+1).\] Now we turn our attention toward the rotations. For every positive divisor $d$ of $n=|\langle a\rangle|$, there is a unique subgroup of $\langle a\rangle$ of order $d$, and this subgroup has precisely $\varphi(d)$ generators. For each of these generators $g$, if we choose an edge of the $n$-gon, each of the images of this edge under the $d$ distinct powers of $g$ must be colored the same color if $g$ leaves the coloring fixed. Therefore there are $q^{n/d}$ colorings left fixed by each of the $\varphi(d)$ elements of order $d$ and hence we have \[\sum_{g\in \langle a\rangle} |S^g|=\sum_{d|n}\varphi(d)q^{n/d}.\] Combining this with our results for the flips and Burnside's theorem, we have shown that the number of orbits in $S$ under the action of $D_n$ is given by \begin{equation*} r=\left\{\begin{array}{ll} \frac{1}{2n}\left (nq^{(n+1)/2}+{\displaystyle\sum_{d|n}\varphi(d)q^{n/d}}\right ) & \mbox {if $n$ is odd,}\\ \frac{1}{2n}\left (\frac{n}{2}q^{n/2}(q+1)+{\displaystyle\sum_{d|n}\varphi(d)q^{n/d}}\right ) & \mbox {if $n$ is even.}\\ \end{array}\right. \end{equation*} Now, Theorem~\ref{second} is easily verified for $n=1$ and $n=2$. If $n\ge 3$, then setting $q=1$ above and noting that we must have $r=1$, we see (using $n$ even or odd) we have \[n+\sum_{d|n}\varphi(d)=2n,\] which completes the proof.

https://arxiv.org/abs/math/0703504

Ubiquity of simplices in subsets of vector spaces over finite fields

We prove that a sufficiently large subset of the $d$-dimensional vector space over a finite field with $q$ elements, $ {\Bbb F}_q^d$, contains a copy of every $k$-simplex. Fourier analytic methods, Kloosterman sums, and bootstrapping play an important role.

\section{Introduction} \vskip.125in Many problems in combinatorial geometry ask, in one form or another, whether a certain structure must be present in a set of sufficiently large size. Perhaps the most celebrated result of this type is Szemeredi's theorem (\cite{SO75}) which says that if a subset of the integers has positive density, then it contains an arbitrary large arithmetic progression. The conclusion has recently been extended to the subsets of prime numbers by Green and Tao (\cite{GT07}). In Euclidean space, a result due to Katznelson and Weiss (\cite{FKB90}) says that a subset of Euclidean space of positive Lebesgue upper density contains every sufficiently large distance. A subsequent result by Bourgain (\cite{BK86}), says that a subset of $\Bbb R^k$ of positive Lebesgue upper density contains an isometric copy of all large dilates of a set of $k$ points spanning a $(k-1)$-dimensional hyperplane. Ergodic theory has been used to show that positive upper density implies that the set contains a copy of a sufficiently large dilate of every convex polygon with finitely many sides. See, for example, a recent survey by Bryna Kra (\cite{K06}). Let $\Bbb F_q^d$ be a $d$-dimensional vector space over a finite field $\Bbb F_q$ of odd characteristic. A plausible analogy to Bourgain's result (\cite {BK86}) in this context would be to consider whether a subset of positive density contains a isometric copy of a set of $k$ points spanning a $(k-1)$-dimensional hyperplane. It turns out however, that the positive density condition is much too strong in the context of vector spaces over finite fields and the same conclusion follows from a much weaker assumption on the size of the underlying set. \begin{definition} Let a $k$-simplex be a set of $k+1$ points in general position, which means that no $n + 1$ of these points, $n \leq k$, lie in a $(n - 1)$-dimensional sub-space of ${\Bbb F}_q^d$. \end{definition} \begin{definition} We say that a linear transformation $T$ on ${\Bbb F}_q^d$ is an isometry if $$ ||Tx||=||x||,$$ where $$ ||x||=x_1^2+x_2^2+\dots+x_d^2,$$ an element of ${\Bbb F}_q$. \end{definition} The question we ask in this paper is how large does $E \subset {\Bbb F}_q^d$ need to be in order to be sure that it contains a copy of every $k$-simplex. Our main result is the following. \begin{theorem} \label{main} Let $E \subset {\Bbb F}_q^d$, $d>{k+1 \choose 2}$, such that $|E| \ge C q^{\tfrac{k}{k+1}d}q^{\tfrac{k}{2}}$ with a sufficiently large constant $C>0$. Then $E$ contains an isometric copy of every $k$-simplex. \end{theorem} Note that we obtain non-trivial results only when $k<<\sqrt{d}$. Nevertheless, in that range we are able to dip considerably below the positive density condition on the underlying set $E$. The method of proof relies on the fact that orthogonal transformations on ${\Bbb F}_q^d$ are isometries. A "distance representation" of a simplex is then used to reduce Theorem \ref{main} to an appropriate weighted incidence theorem for spheres and points. Weil's estimate (\cite{We48}) for classical Kloosterman sums is used to control the size of the Fourier transform of spheres of non-zero radius. The key idea in the proof is to show at each step of an inductive argument that a collection of distances among vertices of a given simplex can not only be realized, but actually occur a "statistically correct" number of times. \vskip.125in \section{Preliminaries and Definitions} Let $\Bbb F _q^d$ be the $d$-dimensional vector space over the finite field ${\Bbb F}_q$. The Fourier transform of a function $$f: \Bbb F_q^d \rightarrow \Bbb C$$ is given by \begin{equation*} \widehat{f}(m) := q^{-d} \sum_{x \in \Bbb F_q^d} f(x) \chi(-x \cdot m), \end{equation*} where $\chi$ is an additive character on $\Bbb F_q$. The orthogonality property of the Fourier Transform says that $$ q^{-d}\sum_{x \in \Bbb F_q^d} \chi(-x \cdot m)=1$$ for $m=(0, \dots, 0)$ and $0$ otherwise yields many standard properties of the Fourier Transform. We summarize some of the properties of the Fourier Transform as follows. \begin{lemma}[The Fourier Transform] Let $$f,g:\Bbb F_q^d \rightarrow \Bbb C.$$ \begin{align*} & \hat f (0, \dots, 0) =q^{-d} \sum_{x \in \Bbb F_q^d} f(x), \\ & (\text{Plancherel})\ q^{-d} \sum_{x \in \Bbb F_q^d} f(x) \overline{g (x)} =\sum_{m \in \Bbb F_q^d} \hat{f}(m) \overline{\hat{g}(m)}, \\ & (\text{Inversion})\ \ \ f(x) =\sum_{m \in \Bbb F_q^d} \hat{f}(m) \chi(x \cdot m). \end{align*} \end{lemma} \subsection{Notation} Throughout the paper $X \lesssim Y$ means that there exists $C>0$ such that $X\leq CY$, $X \gtrsim Y$ means $Y \lesssim X$, and $X \approx Y$ if both $X\lesssim Y$ and $X\gtrsim Y$. Along the same lines, $X \ll Y $ means that $\tfrac{X} {Y}\rightarrow 0$, as $q \to \infty$ $X \gg Y$ means $Y\ll X$, and $X \sim Y$ if $\tfrac{X}{Y}\rightarrow 1$ as $q \rightarrow \infty$. \vskip.125in \section{Proof of the main result} Even though a finite field with $q$ elements, ${\Bbb F}_q$, is not a metric space, we define the "distance" between two points $x$ and $y$ in ${\Bbb F}_q^d$ by the formula $$ ||x-y||={(x_1-y_1)}^2+{(x_2-y_2)}^2+\dots+{(x_d-y_d)}^2.$$ The same notion of "distance" was used by Bourgain, Katz and Tao(\cite {BKT04}), and Iosevich and Rudnev(\cite{IR07}) in their study of the Erd\H os distance problem in vector spaces over finite fields. As we noted above, a geometric justification of this notion of distance is that an orthogonal transformation on ${\Bbb F}_q^d$, a matrix $O$ such that $O^t \cdot O=I$, preserves this notion of a "distance". Represent a $k$-simplex in a subset $E \subset \Bbb F_q^d$ on $k+1$ points recursively by setting $$ {\cal T}_{l_k}=\{(x_0,\dots,x_{k-1},x_k) \in {\cal T}_{l_{k-1}} \times E: ||x_0-x_k||=t_{1,k}, ||x_1-x_k||=t_{2,k},\dots,||x_{k-1}-x_k||=t_{k,k}\},$$ for $l_k=l_{k-1}\cup_\{t_{1,k},\dots t_{k,k}\},t_{i,j}\in \mathbb F_q^*$ where $$ {\cal T}_{l_1}=\{(x_0,x_1) \in E^2: ||x_0-x_1||=t_{1,1}\}.$$ This representation does not, in general, always embody a simplex as $ {\cal T}_{l_k}^k$ is not guaranteed to be in general position. However, as we show below, "legitimate" $k$-simplices are equivalent up to an orthogonal transformation. \begin{theorem} \label{simplexlemma} Let $E \subset {\Bbb F}_q^d$, $d> {k+1 \choose 2}$, such that $|E| \ge Cq^{\tfrac{k}{k+1}d}q^{\tfrac{k}{2}}$, with a sufficiently large constant $C$. Then for every side length set $l_k$, $l_k\in (\Bbb F_q^*)^{k+1 \choose 2}$ we have that $|{\cal T}_{l_k}|>0$. Furthermore, $$ |{\cal T}_{l_k}| \sim |E|^{k+1}q^{-{k+1 \choose 2}}.$$ \end{theorem} Using this theorem we recover the main result of the paper using the following linear algebraic observation. \begin{lemma}\label{uptocrap} Let $P$ be a simplex with vertices $V_0, V_1, \dots, V_k$, $V_j \in {\Bbb F}_q^d$. Let $P'$ be another simplex with vertices $V'_0, V'_1, \dots, V'_k$. Suppose that \begin{equation} \label{equalnorm} ||V_i-V_j||=||V'_i-V'_j|| \end{equation} for all $i,j$. Then there exists an orthogonal, affine transformation $O$ on ${\Bbb F}_q^d$ such that $O(P)=P'$. \end{lemma} \vskip.125in \subsection{Proof of Theorem \ref{simplexlemma}-the main result reformulated in terms of "distances"} The proof proceeds by induction. The first step is the case $k=2$. For a set $E$ we define the characteristic or indicator function to be $E(x)$. Now define the sphere of radius $t_{1,1} \in \Bbb F_q^*$ to be $$S_{t_{1,1}}=\{x \in {\Bbb F}_q^d: ||x||=t_{1,1}\},$$ then \begin{align*} |{\cal T}_{l_1}|&=|\{(x_0,x_1) \in E \times E: ||x_0-x_1||=t_{1,1}\}| \\ &=\sum_{x_0,x_1} E(x_0)E(x_1)S_{t_{1,1}}(x_0-x_1). \end{align*} In order to obtain information from this quantity the behavior of incidences of spheres and points in $E$ will be critical. The following classical fact, whose proof will be given in a subsequent section, states that the sphere has optimal Fourier decay away from the origin. \begin{lemma} \label{sphere} Let $S_{t}$, $t \in \Bbb F_q^* $ be defined as above. If $m \not=(0, \dots, 0)$ then $$ |\widehat{S}_{t}(m)| \lesssim q^{-\frac{d+1}{2}}, $$ and $$ \widehat{S}_{t}(0, \dots, 0)=q^{-d} |S_{t}| \approx q^{-1}.$$ \end{lemma} Applying Fourier inversion to the sphere, \begin{align*} |{\cal T}_{l_1}|=&q^{2d}\sum_{x_0,x_1} E(x_0)E(x_1)\sum_m \widehat{S}_{t_{1,1}}(m)\chi(m\cdot(x_0-x_1)) \\ &=q^{2d} \sum_m {|\widehat{E}(m)|}^2 \widehat{S}_{t_{1,1}}(m) \\ &={|E|}^2 \cdot q^{-d} \cdot |S_{t_{1,1}}|+q^{2d} \sum_{m \not=(0, \dots, 0)} {|\widehat{E}(m)|}^2 \widehat{S}_{t_{1,1}}(m) \\ &=M+R.\, \end{align*} By Lemma \ref{sphere}, $$ M \approx \frac{{|E|}^2}{q},$$ and using Lemma \ref{sphere} once again, \begin{align*} |R| &\lesssim q^{2d} \cdot q^{-\frac{d+1}{2}} \cdot \sum_m {|\widehat {E}(m)|}^2 \\ &=q^{\frac{d-1}{2}} \cdot |E|, \end{align*} which is smaller than $M$ if $|E| \ge Cq^{\frac{d+1}{2}}$ with a sufficiently large constant $C$ and thus ${\cal T}_{l_1}$ is non-empty. Moreover, if $|E| \gg q^{\frac{d+1} {2}}$, we get the "statistically expected" number of distances, $$|{\cal T}_{l_1}| \sim \frac{{|E|}^2}{q}.$$ Assuming the $(k-1)$st case, we count the number of $k$-simplices in $E$ as an extension of the $(k-1)$-simplices in $E$. $$|{\cal T}_{l_{k}}|= \sum_{x_0,\dots,x_k} {\cal T}_{l_{k-1}}(x_0,\dots,x_{k-1})E(x_k)S_{t_{1,k}}(x_0-x_k)\dots S_{t_{k,k}}(x_{k-1}-x_k).$$ By Fourier inversion, the expression equals $$ \sum_{x_0,\dots,x_k} \sum_{m_0,\dots,m_{k-1}} \prod_{i=1}^{k}\chi ((x_{i-1}-x_k) \cdot m_{i-1})\widehat{S}_{t_{i,k}}(m_{i-1}) {\cal T}_{l_{k-1}}(x_0,\dots,x_{k-1})E(x_k)$$ $$=q^{(k+1)d} \sum_{m_0,\dots,m_{k-1}} \widehat{{\cal T}}_{l_{k-1}} (-m_0,\dots,-m_{k-1}) \widehat{E}(m_0+\dots+m_{k-1}) \widehat{S}_ {t_{1,k}}(m_0)\dots\widehat{S}_{t_{k,k}}(m_{k-1}),$$ where the Fourier transform of ${\cal T}_{l_{k-1}}$ is actually the Fourier transform on ${\Bbb F}_q^d \times\dots\times {\Bbb F}_q^d$, $k$ times. Extracting the zero term and breaking the remaining sum into pieces on which we may apply Lemma \ref{sphere}, this expression equals $$q^{(k+1)d} \cdot q^{-(k+1)d} \cdot |{\cal T}_{l_{k-1}}| \cdot |E| \cdot |S_{t_{1,k}}| \cdot q^{-d}\cdot\dots \cdot |S_{t_{k,k}}| \cdot q^{-d}$$ $$+q^{(k+1)d} \sum_{\substack{\mathcal{I} \cup \mathcal{I}'= \{ 0, \dots ,k-1 \} \\ m_i=0\ (i\in \mathcal{I})\\ m_i\neq 0\ (i \notin \mathcal{I})}} \widehat{{\cal T}}_{l_{k-1}}(-m_0,\dots,-m_{k-1}) \widehat{E}(m_0+\dots+m_{k-1})\widehat{S}_{t_{1,k}}(m_0)\dots\widehat{S}_{t_{k,k}}(m_{k-1})$$ $$=M+R,$$ where the sum defining $R$ runs over all the partitions of $\{0,\dots,k-1\}$ with the case $\mathcal{I}'=\emptyset$ extracted and used as the main term $M$ above. By Lemma \ref{sphere} and the induction hypothesis, $$ M \sim {|E|}^{k+1}q^{-{k+1 \choose 2}}.$$ By Lemma \ref{sphere} we have that $$|R|\lesssim q^{(k+1)d}\sum_{\substack{\mathcal{I} \cup \mathcal{I}'= \{ 0, \dots ,k-1 \} \\ m_i=0\ (i\in \mathcal{I})\\ m_i\neq 0\ (i \notin \mathcal{I})}} q^{-|\mathcal{I}'|(d+1)/2-|\mathcal{I}|} |\widehat{{\cal T}}_{l_{k-1}}(-m_0,\dots,-m_{k-1})| |\widehat{E}(m_0+\dots+m_{k-1})|.$$ Then for each term in the sum corresponding to a partition $\mathcal{I} \cup \mathcal{I}'$ we apply Cauchy-Schwarz, $$\sum_{\substack{m_i=0\ (i\in \mathcal{I})\\ m_i\neq 0\ (i \notin \mathcal{I})}} |\widehat{{\cal T}}_{l_{k-1}}(-m_0,\dots,-m_{k-1})| |\widehat{E}(m_0+\dots+m_{k-1})| \lesssim A^{1/2}B^{1/2}.$$ Applying Plancherel and the induction hypothesis, $$A\leq\sum_{m_0,\dots,m_{k-1}} |\widehat{{\cal T}}_{l_{k-1}}(-m_0, \dots,-m_{k-1})|^2 =q^{-kd}|{\cal T}_{l_{k-1}}|\sim q^{-kd}q^{-{k \choose 2}} {|E|}^{k}.$$ Now $$B=\sum_{m_i(i\in \mathcal{I}')} \left| \widehat{E}\left(\sum_{i\in \mathcal{I}'}m_i \right)\right|^2=q^{|\mathcal{I}'|d}q^{-2d}|E|.$$ This implies that $$|R|\lesssim q^{\tfrac{kd}{2}}q^{-\tfrac{k(k-1)}{4}} {|E|}^{\tfrac{k+1}{2}}\sum_{\mathcal{I} \cup \mathcal{I}'= \{ 0, \dots ,k-1 \}} q^{-|\mathcal{I}'|(d+1)/2-|\mathcal{I}|} q^{|\mathcal{I}'|d}.$$ The largest term in the sum occurs when $\mathcal{I}=\emptyset$. We conclude that $$|R| \lesssim q^{\tfrac{kd}{2}} q^{-\tfrac{k(k+1)}{4}} {|E|}^ {\tfrac{k+1}{2}}.$$ The term $R$ is smaller than, say, $\frac{M}{2}$ if $$q^{\tfrac{kd}{2}} q^{-\tfrac{k(k+1)}{4}} {|E|}^{\tfrac{k+1}{2}} \leq C {|E|}^{k+1}q^{-{k+1 \choose 2}},$$ with a sufficiently large constant $C$, which happens if $$ |E| \ge C'q^{\tfrac{k}{k+1}d}q^{\tfrac{k}{2}},$$ with a sufficiently large constant $C'$ depending on the constants implicit in the estimates above. This completes the proof. \vskip.125in \section{Proof of Lemma \ref{uptocrap}} To prove Lemma \ref{uptocrap}, let $\pi_r(x)$ denote the $r$th coordinate of $x$. There is no harm in assuming that $V_0=(0, \dots, 0)$. We may also assume that $V_1, \dots, V_k$ are contained in ${\Bbb F}_q^k$. The condition (\ref{equalnorm}) implies that \begin{equation} \label{dotproduct} \sum_{r=1}^k \pi_r(V_i) \pi_r(V_j)= \sum_{r=1}^k \pi_r(W_i) \pi_r(W_j). \end{equation} Let $T$ be the linear transformation uniquely determined by the condition $$ T(V_i)=V'_i.$$ In order to prove that $T$ is orthogonal, it suffices to show that $$ ||Tx||=||x||$$ for any $x \not=(0, \dots, 0)$. Since $V_j$s form a basis, by assumption, we have $$ x=\sum_i t_i V_i, $$ so it suffices to show that $$ ||x||=\sum_r \sum_{i,j} t_i t_j \pi_r(V_i) \pi_r(V_j)$$ $$=\sum_r \sum_{i,j} t_i t_j \pi_r(V'_i) \pi_r(V'_j)=||Tx||,$$ which follows immediately from (\ref{dotproduct}). Observe that we used the fact that orthogonality of $T$, the condition that $T^t \cdot T=I$ is equivalent to the condition that $||Tx||=||x||$. To see this observe that to show that $T^t \cdot T=I$ it suffices to show that $T^tTx=x$ for all non-zero $x$. This, in turn, is equivalent to the statement that $$ <T^tTx,x>=||x||,$$ where $$ <x,y>=\sum_{i=1}^k x_iy_i.$$ Now, $$ <T^tTx,x>=<Tx, Tx>$$ by definition of the transpose, so the stated equivalence is established. This completes the proof of Lemma \ref{uptocrap}. \vskip.125in \section{Estimation of the Fourier transform of the sphere: proof of Lemma \ref{sphere}} The proof of Lemma \ref{sphere} is fairly standard, but we outline the argument for reader's convenience. For any $m\in{\mathbb F}^d_q$, we have \begin{equation} \label{sphereparade} \begin{array}{llllll} \widehat{S}_t(m)&=& q^{-d} \sum_{x \in {\mathbb F}^d_q} q^{-1} \sum_{j \in {\mathbb F}_q} \chi( j(\|x\|-t)) \chi( - x \cdot m)\\ \hfill \\&=&q^{-1}\delta(m) + q^{-d-1} \sum_{j \in {\mathbb F}^{*}_q} \chi(-jt) \sum_{x} \chi( j\|x\|) \chi(- x \cdot m)\\ \hfill \\&=&q^{-1}\delta(m)+ Q^d q^{-\frac{d+2}{2}} \sum_{j \in {\mathbb F}^{*}_q} \chi\left(\frac{\|m\|}{4j}+jt\right)\eta^d(-j),\end{array}\end{equation} where the notation $\delta(m)=1$ if $m=(0\ldots,0)$ and $\delta(m)=0$ otherwise. In the last line we have completed the square, changed $j$ to $-j$, and used $d$ times the Gauss sum equality \begin{equation} \sum_{c\in {\mathbb F}_q} \chi(jc^2) = \eta(j)\sum_{c\in{\mathbb F}_q}\eta(c)\chi(c)=\eta(j)\sum_{c\in{\mathbb F}_q^*}\eta(c)\chi(c) = Q\sqrt{q}\,\eta(j), \label{gauss}\end{equation} where the constant $Q$ equals $\pm1$ or $\pm i$, depending on $q$, and $\eta$ is the quadratic multiplicative character (or the Legendre symbol) of ${\mathbb F}_q^*$. The conclusion now follows from the following classical estimate due to A. Weil (\cite{We48}). \begin{theorem} \label{kloosterman} Let $$ K(a)=\sum_{s \not=0} \chi(as+s^{-1}) \psi(s), $$ where, once again, $\psi$ is a multiplicative character on ${\Bbb F}_q^{*}$. Then $$ |K(a)| \leq 2 \sqrt{q}$$ if $a \not=0$. \end{theorem} \vskip.125in \newpage

https://arxiv.org/abs/2010.05415

Unweighted linear congruences with distinct coordinates and the Varshamov--Tenengolts codes

In this paper, we first give explicit formulas for the number of solutions of unweighted linear congruences with distinct coordinates. Our main tools are properties of Ramanujan sums and of the discrete Fourier transform of arithmetic functions. Then, as an application, we derive an explicit formula for the number of codewords in the Varshamov--Tenengolts code $VT_b(n)$ with Hamming weight $k$, that is, with exactly $k$ $1$'s. The Varshamov--Tenengolts codes are an important class of codes that are capable of correcting asymmetric errors on a $Z$-channel. As another application, we derive Ginzburg's formula for the number of codewords in $VT_b(n)$, that is, $|VT_b(n)|$. We even go further and discuss connections to several other combinatorial problems, some of which have appeared in seemingly unrelated contexts. This provides a general framework and gives new insight into all these problems which might lead to further work.

\section{Introduction}\label{Sec_1} A \textit{$Z$-channel} (also called a \textit{binary asymmetric channel}) is a channel with binary input and binary output where a transmitted $0$ is always received correctly but a transmitted $1$ may be received as either $1$ or $0$. These channels have many applications, for example, some data storage systems and optical communication systems can be modelled using these channels. In 1965, Varshamov and Tenengolts \cite{VATE} introduced an important class of codes, known as the Varshamov--Tenengolts codes or VT-codes, that are capable of correcting asymmetric errors on a $Z$-channel (see also \cite{STYO, VAR}). Levenshtein \cite{LEV1, LEV2}, by giving an elegant decoding algorithm, showed that these codes could also be used for correcting a single deletion or insertion. Using the Varshamov--Tenengolts codes, Gevorkyan and Kabatiansky \cite{GEKA} constructed a class of binary codes of a specific length correcting single localized errors whose cardinality attains the ordinary Hamming bound. \begin{definition} Let $n$ be a positive integer and $0\leq b\leq n$ be a fixed integer. The Varshamov--Tenengolts code $VT_b(n)$ is the set of all binary vectors $\langle y_1,\ldots,y_n \rangle$ such that $$ \sum_{i=1}^{n}iy_i \equiv b \pmod{n+1}. $$ \end{definition} For example, $VT_0(5)=\lbrace 00000, 10001, 01010, 11100, 00111, 11011 \rbrace$, where we have shown vectors as strings. So, $|VT_0(5)|=6$. The \textit{Hamming weight} of a string over an alphabet is defined as the number of non-zero symbols in the string. Equivalently, the Hamming weight of a string is the Hamming distance between that string and the all-zero string of the same length. For example, the Hamming weight of $01010$ is $2$, and the number of codewords in $VT_0(5)$ with Hamming weight $2$ is $2$. Varshamov in his fundamental paper ``On an arithmetic function with an application in the theory of coding" (\cite{VAR2}) proved that the maximum number of codewords in the Varshamov--Tenengolts code $VT_b(n)$ is achieved when $b=0$, that is, $|VT_0(n)| \geq |VT_b(n)|$ for all $b$. Several natural questions arise: What is the number of codewords in the Varshamov--Tenengolts code $VT_b(n)$, that is, $|VT_b(n)|$? Given a positive integer $k$, what is the number of codewords in $VT_b(n)$ with Hamming weight $k$, that is, with exactly $k$ $1$'s? Ginzburg \cite{GIN} in 1967 considered the first question and proved an explicit formula for $|VT_b(n)|$. In this paper, we deal with both questions and obtain explicit formulas for them via a novel approach, namely, \textit{connecting the Varshamov--Tenengolts codes to linear congruences with distinct coordinates}. We even go further and show that the number of solutions of these congruences is related to several other combinatorial problems, some of which have appeared in seemingly unrelated contexts. (For example, as we will discuss in Section~\ref{Sec_4}, Razen, Seberry, and Wehrhahn \cite{RSW} considered two special cases of a function considered in this paper and gave an application in coding theory in finding the complete weight enumerator of a code generated by a circulant matrix.) This provides a general framework and gives new insight into all these problems which might lead to further work. Let us now describe these congruences. Throughout the paper, we use $(a_1,\ldots,a_k)$ to denote the greatest common divisor (gcd) of the integers $a_1,\ldots,a_k$, and write $\langle a_1,\ldots,a_k\rangle$ for an ordered $k$-tuple of integers. Let $a_1,\ldots,a_k,b,n\in \mathbb{Z}$, $n\geq 1$. A linear congruence in $k$ unknowns $x_1,\ldots,x_k$ is of the form \begin{align} \label{cong form} a_1x_1+\cdots +a_kx_k\equiv b \pmod{n}. \end{align} By a solution of (\ref{cong form}), we mean an $\mathbf{x}=\langle x_1,\ldots,x_k \rangle \in \mathbb{Z}_n^k$ that satisfies (\ref{cong form}). The following result, proved by D. N. Lehmer \cite{LEH2}, gives the number of solutions of the above linear congruence: \begin{proposition}\label{Prop: lin cong} Let $a_1,\ldots,a_k,b,n\in \mathbb{Z}$, $n\geq 1$. The linear congruence $a_1x_1+\cdots +a_kx_k\equiv b \pmod{n}$ has a solution $\langle x_1,\ldots,x_k \rangle \in \mathbb{Z}_{n}^k$ if and only if $\ell \mid b$, where $\ell=(a_1, \ldots, a_k, n)$. Furthermore, if this condition is satisfied, then there are $\ell n^{k-1}$ solutions. \end{proposition} Counting the number of solutions of the above congruence with some restrictions on the solutions is also a problem of great interest. As an important example, one can mention the restrictions $(x_i,n)=t_i$ ($1\leq i\leq k$), where $t_1,\ldots,t_k$ are given positive divisors of $n$. The number of solutions of the linear congruences with the above restrictions, which we called {\it restricted linear congruences} in \cite{BKSTT}, was first considered by Rademacher \cite{Rad1925} in 1925 and Brauer \cite{Bra1926} in 1926, in the special case of $a_i=t_i=1$ $(1\leq i \leq k)$. Since then, this problem has been studied, in several other special cases, in many papers (very recently, we studied it in its `most general case' in \cite{BKSTT}) and has found very interesting applications in number theory, combinatorics, geometry, physics, computer science, and cryptography; see \cite{BKS2, BKSTT, BKSTT2, JAWILL} for a detailed discussion about this problem and a comprehensive list of references. Another restriction of potential interest is imposing the condition that all $x_i$ are {\it distinct} modulo $n$. Unlike the first problem, there seems to be very little published on the second problem. Recently, Grynkiewicz et al. \cite{GPP}, using tools from additive combinatorics and group theory, proved necessary and sufficient conditions under which the linear congruence $a_1x_1+\cdots +a_kx_k\equiv b \pmod{n}$, where $a_1,\ldots,a_k,b,n$ ($n\geq 1$) are arbitrary integers, has a solution $\langle x_1,\ldots,x_k \rangle \in \mathbb{Z}_{n}^k$ with all $x_i$ distinct modulo $n$; see also \cite{ADP, GPP} for connections to zero-sum theory. So, it would be an interesting problem to give an explicit formula for the number of such solutions. Quite surprisingly, this problem was first considered, in a special case, by Sch\"{o}nemann \cite{SCH} almost two centuries ago(!) but his result seems to have been forgotten. Sch\"{o}nemann \cite{SCH} proved an explicit formula for the number of such solutions when $b=0$, $n=p$ a prime, and $\sum_{i=1}^k a_i \equiv 0 \pmod{p}$ but $\sum_{i \in I} a_i \not\equiv 0 \pmod{p}$ for all $I\varsubsetneq \lbrace 1, \ldots, k\rbrace$. Very recently, the authors \cite{BKS6} generalized Sch\"{o}nemann's theorem using Proposition~\ref{Prop: lin cong} and a result on graph enumeration recently obtained by Ardila et al. \cite{ACH}. Specifically, we obtained an explicit formula for the number of solutions of the linear congruence $a_1x_1+\cdots +a_kx_k\equiv b \pmod{n}$, with all $x_i$ distinct modulo $n$, when $(\sum_{i \in I} a_i, n)=1$ for all $I\varsubsetneq \lbrace 1, \ldots, k\rbrace$, where $a_1,\ldots,a_k,b,n$ $(n\geq 1)$ are arbitrary integers. Clearly, this result does not resolve the problem in its full generality; for example, it does not cover the important case of $a_i=1$ ($1\leq i\leq k$) and this is what we consider in this paper with an entirely different approach. Specifically, we give an explicit formula for the number $N_n(k,b)$ of such solutions when $a_i=1$ ($1\leq i\leq k$), and do the same when in addition all $x_i$ are \textit{positive} modulo $n$. Our main tools in this paper are properties of Ramanujan sums and of the discrete Fourier transform of arithmetic functions which are reviewed in the next section. In Section~\ref{Sec_3}, we derive the explicit formulas, and discuss applications to the Varshamov--Tenengolts codes. In Section~\ref{Sec_4}, we discuss connections to several other combinatorial contexts. \section{Ramanujan sums and discrete Fourier transform} \label{Sec_2} Let $e(x)=\exp(2\pi ix)$ be the complex exponential with period 1. For integers $m$ and $n$ with $n \geq 1$ the quantity \begin{align}\label{def1} c_n(m) = \mathlarger{\sum}_{\substack{j=1 \\ (j,n)=1}}^{n} e\!\left(\frac{jm}{n}\right) \end{align} is called a {\it Ramanujan sum}. It is the sum of the $m$-th powers of the primitive $n$-th roots of unity, and is also denoted by $c(m,n)$ in the literature. From (\ref{def1}), it is clear that $c_n(-m) = c_n(m)$. Clearly, $c_n(0)=\varphi (n)$, where $\varphi (n)$ is {\it Euler's totient function}. Also, $c_n(1)=\mu (n)$, where $\mu (n)$ is the {\it M\"{o}bius function}. The following theorem, attributed to Kluyver~\cite{KLU}, gives an explicit formula for $c_n(m)$: \begin{theorem} \label{thm:Ram Mob} For integers $m$ and $n$, with $n \geq 1$, \begin{align}\label{for:Ram Mob} c_n(m) = \mathlarger{\sum}_{d\, \mid\, (m,n)} \mu \!\left(\frac{n}{d}\right)d. \end{align} \end{theorem} By applying the M\"{o}bius inversion formula, Theorem~\ref{thm:Ram Mob} yields the following property: For $m,n\geq 1$, \begin{align} \label{Orth1 for cons} \sum_{d\, \mid\, n} c_{d}(m)&= \begin{cases} n, & \text{if $n\mid m$},\\ 0, & \text{if $n\nmid m$}. \end{cases} \end{align} The {\it von Sterneck number} (\cite{von}) is defined by \begin{align}\label{def3} \Phi(m,n)=\frac{\varphi (n)}{\varphi \left(\frac{n}{\left(m,n\right)}\right)}\mu \!\left(\frac{n}{\left(m,n\right)} \right). \end{align} A crucial fact in studying Ramanujan sums and their applications is that they coincide with the von Sterneck number. This result is attributed to Kluyver~\cite{KLU}: \begin{theorem} \label{thm:von rama} For integers $m$ and $n$, with $n \geq 1$, we have \begin{align}\label{von rama for} \Phi(m,n)=c_n(m). \end{align} \end{theorem} A function $f:\mathbb{Z} \to \mathbb{C}$ is called {\it periodic} with period $n$ (also called {\it $n$-periodic} or {\it periodic} modulo $n$) if $f(m + n) = f(m)$, for every $m\in \mathbb{Z}$. In this case $f$ is determined by the finite vector $(f(1),\ldots,f(n))$. From (\ref{def1}) it is clear that $c_n(m)$ is a periodic function of $m$ with period $n$. We define the {\it discrete Fourier transform} (DFT) of an $n$-periodic function $f$ as the function $\widehat{f}={\cal F}(f)$, given by \begin{align}\label{FFT1} \widehat{f}(b)=\mathlarger{\sum}_{j=1}^{n}f(j)e\! \left(\frac{-bj}{n}\right)\quad (b\in \mathbb{Z}). \end{align} The standard representation of $f$ is obtained from the Fourier representation $\widehat{f}$ by \begin{align}\label{FFT2} f(b)=\frac1{n} \mathlarger{\sum}_{j=1}^{n}\widehat{f}(j)e\!\left(\frac{bj}{n}\right) \quad (b\in \mathbb{Z}), \end{align} which is the {\it inverse discrete Fourier transform} (IDFT); see, e.g., \cite[p.\ 109]{MOVA}. \section{Solutions with distinct coordinates}\label{Sec_3} In this section, we obtain an explicit formula for the number of solutions $\langle x_1,\ldots,x_k \rangle \in \mathbb{Z}_{n}^k$ of the linear congruence $x_1+\cdots +x_k\equiv b \pmod{n}$, with all $x_i$ distinct modulo $n$. First, we need some preliminary results. \begin{lemma}\label{lem: cyclo 1} Let $n$ be a positive integer and $m$ be a non-negative integer. We have $$ \mathlarger{\prod}_{j=1}^{n}\left(1-ze^{2\pi ijm/n}\right)=(1-z^{\frac{n}{d}})^d, $$ where $d=(m,n)$. \end{lemma} \begin{proof} It is well-known that (see, e.g., \cite[p. 167]{STAN}) $$ 1-z^n=\mathlarger{\prod}_{j=1}^{n}\left(1-ze^{2\pi ij/n}\right). $$ Now, letting $d=(m,n)$, we obtain \begin{align*} \mathlarger{\prod}_{j=1}^{n}\left(1-ze^{2\pi ijm/n}\right) &= \mathlarger{\prod}_{j=1}^{n} \left(1-ze^{2\pi ij\frac{m/d}{n/d}}\right)\\ &= \left(\mathlarger{\prod}_{j=1}^{n/d}\left(1-ze^{2\pi ij\frac{m/d}{n/d}}\right)\right)^d\\ &{\stackrel{(\frac{m}{d},\frac{n}{d})=1}{=}} \left(\mathlarger{\prod}_{j=1}^{n/d}\left(1-ze^{\frac{2\pi ij}{n/d}}\right)\right)^d=(1-z^{\frac{n}{d}})^d. \end{align*} \end{proof} Similarly, we can prove that: \begin{lemma}\label{lem: cyclo 2} Let $n$ be a positive integer and $m$ be a non-negative integer. We have $$ \mathlarger{\prod}_{j=1}^{n}\left(z-e^{2\pi ijm/n}\right)=(z^{\frac{n}{d}}-1)^d, $$ where $d=(m,n)$. \end{lemma} Now, we simply get: \begin{corollary}\label{cor: cyclo 1} Let $n$ be a positive integer and $m$, $k$ be non-negative integers. The coefficient of $z^k$ in $$ \mathlarger{\prod}_{j=1}^{n}\left(1+ze^{2\pi ijm/n}\right), $$ is $(-1)^{k+\frac{kd}{n}}\binom{d}{\frac{kd}{n}}$, where $d=(m,n)$. Note that the binomial coefficient $\binom{d}{\frac{kd}{n}}$ equals zero if $\frac{kd}{n}$ is not an integer. \end{corollary} Now, we are ready to obtain an explicit formula for the number of solutions of the linear congruence. \begin{theorem} \label{main thm dist ai=1} Let $n$ be a positive integer and $b \in \mathbb{Z}_n$. The number $N_n(k,b)$ of solutions $\langle x_1,\ldots,x_k \rangle \in \mathbb{Z}_{n}^k$ of the linear congruence $x_1+\cdots +x_k\equiv b \pmod{n}$, with all $x_i$ distinct modulo $n$, is \begin{align} \label{main thm dist ai=1: for} N_n(k,b)=\frac{(-1)^k k!}{n}\mathlarger{\sum}_{d\, \mid \, (n,\;k)}(-1)^{\frac{k}{d}}c_{d}(b)\binom{\frac{n}{d}}{\frac{k}{d}}. \end{align} \end{theorem} \begin{proof} It is well-known that (see, e.g., \cite[pp. 3-4]{GUP}) the number of partitions of $b$ into exactly $k$ \textit{distinct} parts each taken from the given set $A$, is the coefficient of $q^bz^k$ in $$ \mathlarger{\prod}_{j \in A}\left(1+zq^j\right). $$ Now, take $A=\mathbb{Z}_n$ and $q=e^{2\pi im/n}$, where $m$ is a non-negative integer. Then, the number $P_n(k,b)$ of partitions of $b$ into exactly $k$ \textit{distinct} parts each taken from $\mathbb{Z}_n$ (that is, the number of solutions of the above linear congruence, with all $x_i$ distinct modulo $n$, if order does not matter), is the coefficient of $e^{2\pi ibm/n}z^k$ in $$ \mathlarger{\prod}_{j=1}^{n}\left(1+ze^{2\pi ijm/n}\right). $$ This in turn implies that $$ \mathlarger{\sum}_{b=1}^{n}P_n(k,b)e^{2\pi ibm/n} = \text{the coefficient of $z^k$ in $\mathlarger{\prod}_{j=1}^{n}\left(1+ze^{2\pi ijm/n}\right)$}. $$ Let $e(x)=\exp(2\pi ix)$. Note that $N_n(k,b)=k!P_n(k,b)$. Now, using Corollary~\ref{cor: cyclo 1}, we get $$ \mathlarger{\sum}_{b=1}^{n}N_n(k,b)e\left(\frac{bm}{n}\right) = (-1)^{k+\frac{kd}{n}}k!\binom{d}{\frac{kd}{n}}, $$ where $d=(m,n)$. Now, by (\ref{FFT1}) and (\ref{FFT2}), we obtain \begin{align*} N_n(k,b) &= \frac{(-1)^{k}k!}{n}\mathlarger{\sum}_{m=1}^{n}(-1)^{\frac{kd}{n}}e\left(\frac{-bm}{n}\right)\binom{d}{\frac{kd}{n}}\\ &= \frac{(-1)^{k}k!}{n}\mathlarger{\sum}_{d\, \mid \, n}\mathlarger{\sum}_{\substack{m=1 \\ (m,\;n)=d}}^{n}(-1)^{\frac{kd}{n}}e\left(\frac{-bm}{n}\right)\binom{d}{\frac{kd}{n}}\\ &{\stackrel{m'=m/d}{=}} \;\; \frac{(-1)^{k}k!}{n}\mathlarger{\sum}_{d\, \mid \, n}\mathlarger{\sum}_{\substack{m'=1 \\ (m',\;n/d)=1}}^{n/d}(-1)^{\frac{kd}{n}}e\left(\frac{-bm'}{n/d}\right)\binom{d}{\frac{kd}{n}}\\ &= \frac{(-1)^{k}k!}{n}\mathlarger{\sum}_{d\, \mid \, n}(-1)^{\frac{kd}{n}}c_{n/d}(-b)\binom{d}{\frac{kd}{n}}\\ &= \frac{(-1)^{k}k!}{n}\mathlarger{\sum}_{d\, \mid \, n}(-1)^{\frac{kd}{n}}c_{n/d}(b)\binom{d}{\frac{kd}{n}}\\ &= \frac{(-1)^{k}k!}{n}\mathlarger{\sum}_{d\, \mid \, n}(-1)^{\frac{k}{d}}c_{d}(b)\binom{\frac{n}{d}}{\frac{k}{d}}\\ &= \frac{(-1)^{k}k!}{n}\mathlarger{\sum}_{d\, \mid \, (n,\;k)}(-1)^{\frac{k}{d}}c_{d}(b)\binom{\frac{n}{d}}{\frac{k}{d}}. \end{align*} \end{proof} \begin{corollary} \label{special cases: b=0,1} If $n$ or $k$ is odd then from (\ref{main thm dist ai=1: for}) we obtain the following important special cases of the function $P_n(k,b)=\frac{1}{k!}N_n(k,b)$: \begin{align} \label{special case: b=0} P_n(k,0)= \frac{1}{n}\mathlarger{\sum}_{d\, \mid \, (n,\;k)}\varphi(d)\binom{\frac{n}{d}}{\frac{k}{d}}, \end{align} \begin{align} \label{special case: b=1} P_n(k,1)= \frac{1}{n}\mathlarger{\sum}_{d\, \mid \, (n,\;k)}\mu(d)\binom{\frac{n}{d}}{\frac{k}{d}}. \end{align} \end{corollary} \begin{corollary} If $(n,k)=1$ then (\ref{main thm dist ai=1: for}) is independent of $b$ and simplifies as $$N_n(k)=\frac{k!}{n}\binom{n}{k}.$$ (Of course, this can also be proved directly.) If in addition we have $n=2k+1$ then $$ P_n(k)=\frac{1}{k!}N_n(k)=\frac{1}{2k+1}\binom{2k+1}{k}=\frac{1}{k+1}\binom{2k}{k}, $$ which is the Catalan number. \end{corollary} \begin{rema} Using (\ref{Orth1 for cons}), it is easy to see that (\ref{main thm dist ai=1: for}) also works when $k=0$. \end{rema} Now, we introduce the important function $T_n(b)$ which is the sum of $P_n(k,b)$ over $k$. There are several interpretations for the function $T_n(b)$, for example, $T_n(b)$ can be interpreted as the number of subsets of the set $\lbrace 1, 2, \ldots, n \rbrace$ which sum to $b$ modulo $n$. \begin{corollary}\label{nice function} Let $T_n(b):=\sum_{k=0}^{n}\frac{1}{k!}N_n(k,b)=\sum_{k=0}^{n}P_n(k,b)$. Then we have \begin{align} \label{nice function: for} T_n(b)=\frac{1}{n}\mathlarger{\sum}_{\substack{d\, \mid \, n \\ d \; \textnormal{odd}}}c_{d}(b)2^{\frac{n}{d}}. \end{align} \end{corollary} \begin{proof} We have \begin{align*} T_n(b)&= \mathlarger{\sum}_{k=0}^{n} \frac{(-1)^{k}}{n}\mathlarger{\sum}_{d\, \mid \, (n,\;k)}(-1)^{\frac{k}{d}}\binom{\frac{n}{d}}{\frac{k}{d}}c_{d}(b)\\ &= \frac{1}{n}\mathlarger{\sum}_{d\, \mid \, n}c_{d}(b)\mathlarger{\sum}_{\substack{k=0 \\ d\, \mid \,k}}^{n}(-1)^{k+\frac{k}{d}}\binom{\frac{n}{d}}{\frac{k}{d}}\\ &= \frac{1}{n}\mathlarger{\sum}_{\substack{d\, \mid \, n \\ d \; \textnormal{odd}}}c_{d}(b)\mathlarger{\sum}_{\substack{k=0 \\ d\, \mid \,k}}^{n}(-1)^{k+\frac{k}{d}}\binom{\frac{n}{d}}{\frac{k}{d}} +\frac{1}{n}\mathlarger{\sum}_{\substack{d\, \mid \, n \\ d \; \textnormal{even}}}c_{d}(b)\mathlarger{\sum}_{\substack{k=0 \\ d\, \mid \,k}}^{n}(-1)^{k+\frac{k}{d}}\binom{\frac{n}{d}}{\frac{k}{d}} \\ &=\frac{1}{n}\mathlarger{\sum}_{\substack{d\, \mid \, n \\ d \; \textnormal{odd}}}c_{d}(b)2^{\frac{n}{d}}. \end{align*} Note that in the last equality we have used the fact that if $d \mid n$ and $d$ is even then $$ \mathlarger{\sum}_{\substack{k=0 \\ d\, \mid \,k}}^{n}(-1)^{k+\frac{k}{d}}\binom{\frac{n}{d}}{\frac{k}{d}} = \mathlarger{\sum}_{\substack{k=0 \\ d\, \mid \,k}}^{n}(-1)^{\frac{k}{d}}\binom{\frac{n}{d}}{\frac{k}{d}}= 0. $$ \end{proof} What is the number of subsets of the set $\lbrace 1, 2, \ldots, n-1 \rbrace$ which sum to $b$ modulo $n$? Using Corollary~\ref{nice function}, we can obtain an explicit formula for the number of such subsets (see also \cite{MAZ}). \begin{corollary}\label{nice function 2} The number $T'_n(b)$ of subsets of the set $\lbrace 1, 2, \ldots, n-1 \rbrace$ which sum to $b$ modulo $n$ is \begin{align} \label{nice function 2: for} T'_n(b)=\frac{1}{2}T_n(b)=\frac{1}{2n}\mathlarger{\sum}_{\substack{d\, \mid \, n \\ d \; \textnormal{odd}}}c_{d}(b)2^{\frac{n}{d}}. \end{align} \end{corollary} \begin{proof} Let $A$ be a subset of the set $\lbrace 1, 2, \ldots, n-1 \rbrace$ which sum to $b$ modulo $n$. Then $A$ and $A\cup \lbrace n \rbrace$ are both subsets of the set $\lbrace 1, 2, \ldots, n \rbrace$ and both sum to $b$ modulo $n$. Therefore, $T'_n(b)=\frac{1}{2}T_n(b)$. \end{proof} Ginzburg \cite{GIN} in 1967 proved an explicit formula for the number of codewords in the $q$-ary Varshamov--Tenengolts codes, where $q$ is an arbitrary positive integer. This result was later rediscovered by Stanley and Yoder \cite{STYO} in 1973, and in the binary case (that is, when $q=2$) by Sloane \cite{SLO} in 2002. Now, we give a short proof for the binary case which we derive as a consequence of our results. \begin{corollary}\label{VT exa tot} The number $|VT_b(n)|$ of codewords in the Varshamov--Tenengolts code $VT_b(n)$ is \begin{align} \label{VT exa tot: for} |VT_b(n)|=\frac{1}{2(n+1)}\mathlarger{\sum}_{\substack{d\, \mid \, n+1 \\ d \; \textnormal{odd}}}c_{d}(b)2^{\frac{n+1}{d}}. \end{align} \end{corollary} \begin{proof} Let $\langle y_1,\ldots,y_n \rangle$ be a codeword in $VT_b(n)$. Note that $\sum_{i=1}^{n}iy_i$ is just the sum of some elements of the set $\lbrace 1, 2, \ldots, n \rbrace$. Therefore, finding the number of codewords in $VT_b(n)$ boils down to finding the number of subsets of the set $\lbrace 1, 2, \ldots, n \rbrace$ which sum to $b$ modulo $n+1$. The result now follows by a direct application of Corollary~\ref{nice function 2}. \end{proof} In some applications (for example, in coding theory) we also need to consider the case that all $x_i$ are \textit{positive} and \textit{distinct} modulo $n$. Now, we obtain an explicit formula for the number of such solutions. \begin{theorem} \label{main thm dist pos ai=1} Let $n$ be a positive integer and $b \in \mathbb{Z}_n$. The number $N_n^{>0}(k,b)$ of solutions $\langle x_1,\ldots,x_k \rangle \in \mathbb{Z}_{n}^k$ of the linear congruence $x_1+\cdots +x_k\equiv b \pmod{n}$, with all $x_i$ positive and distinct modulo $n$, is \begin{align} \label{main thm dist pos ai=1: for} N_n^{>0}(k,b)=\frac{(-1)^k k!}{n}\mathlarger{\sum}_{d\, \mid \, n}(-1)^{\lfloor\frac{k}{d}\rfloor}c_{d}(b)\binom{\frac{n}{d}-1}{\lfloor\frac{k}{d}\rfloor}. \end{align} \end{theorem} \begin{proof} Clearly, $N_n^{>0}(k,b)=N_n(k,b)-N_n^{0}(k,b)$, where $N_n^{0}(k,b)$ denotes the number of solutions $\langle x_1,\ldots,x_k \rangle \in \mathbb{Z}_{n}^k$ with all $x_i$ distinct modulo $n$ and one of $x_i$ is zero modulo $n$. Also, clearly, $N_n^{0}(k,b)=kN_n^{>0}(k-1,b)$. Thus, \begin{align}\label{relation bet N and N>0} N_n(k,b)=N_n^{>0}(k,b)+kN_n^{>0}(k-1,b). \end{align} Now, we use Theorem~\ref{main thm dist ai=1}. We have \begin{align*} N_n(k,b)&= \frac{(-1)^k k!}{n}\mathlarger{\sum}_{d\, \mid \, (n,\;k)}(-1)^{\frac{k}{d}}c_{d}(b)\binom{\frac{n}{d}}{\frac{k}{d}}\\ &= \frac{(-1)^k k!}{n}\mathlarger{\sum}_{d\, \mid \, (n,\;k)}(-1)^{\frac{k}{d}}c_{d}(b)\left(\binom{\frac{n}{d}-1}{\frac{k}{d}}+\binom{\frac{n}{d}-1}{\frac{k}{d}-1}\right)\\ &= \frac{(-1)^k k!}{n}\mathlarger{\sum}_{d\, \mid \, n}c_{d}(b)\left((-1)^{\frac{k}{d}}\binom{\frac{n}{d}-1}{\frac{k}{d}}-(-1)^{\frac{k}{d}-1}\binom{\frac{n}{d}-1}{\frac{k}{d}-1}\right)\\ &= \frac{(-1)^k k!}{n}\mathlarger{\sum}_{d\, \mid \, n}c_{d}(b)\left((-1)^{\lfloor\frac{k}{d}\rfloor}\binom{\frac{n}{d}-1}{\lfloor\frac{k}{d}\rfloor}-(-1)^{\lfloor\frac{k-1}{d}\rfloor}\binom{\frac{n}{d}-1}{\lfloor\frac{k-1}{d}\rfloor}\right)\\ &= \frac{(-1)^k k!}{n}\mathlarger{\sum}_{d\, \mid \, n}(-1)^{\lfloor\frac{k}{d}\rfloor}c_{d}(b)\binom{\frac{n}{d}-1}{\lfloor\frac{k}{d}\rfloor}\\ &+k\frac{(-1)^{k-1} (k-1)!}{n}\mathlarger{\sum}_{d\, \mid \, n}(-1)^{\lfloor\frac{k-1}{d}\rfloor}c_{d}(b)\binom{\frac{n}{d}-1}{\lfloor\frac{k-1}{d}\rfloor}. \end{align*} Note that in the fourth equality above we have used the fact that $\lfloor\frac{k}{d}\rfloor=\lfloor\frac{k-1}{d}\rfloor+1$ if $d \mid k$, and $\lfloor\frac{k}{d}\rfloor=\lfloor\frac{k-1}{d}\rfloor$ if $d\nmid k$. Now, recalling (\ref{relation bet N and N>0}) we obtain the desired result. \end{proof} We believe that Theorem~\ref{main thm dist pos ai=1} is also a strong tool and might lead to interesting applications. Denote by $VT_b^{1,k}(n)$ the set of codewords in the Varshamov--Tenengolts code $VT_b(n)$ with Hamming weight $k$. Theorem~\ref{main thm dist pos ai=1} immediately gives an explicit formula for the number of such codewords. This result is useful in the study of a class of binary codes that are immune to single repetitions \cite{DOAN}. \begin{corollary}\label{VT exa k1s} The number $|VT_b^{1,k}(n)|$ of codewords in the Varshamov--Tenengolts code $VT_b(n)$ with Hamming weight $k$ is \begin{align} \label{VT exa k1s: for} |VT_b^{1,k}(n)|=\frac{(-1)^k}{n+1}\mathlarger{\sum}_{d\, \mid \, n+1}(-1)^{\lfloor\frac{k}{d}\rfloor}c_{d}(b)\binom{\frac{n+1}{d}-1}{\lfloor\frac{k}{d}\rfloor}. \end{align} \end{corollary} \begin{proof} Let $\langle y_1,\ldots,y_n \rangle$ be a codeword in $VT_b(n)$ with Hamming weight $k$, that is, with exactly $k$ $1$'s. Denote by $x_j$ the position of the $j$th one. Note that $1\leq j \leq k$ and $1 \leq x_1 < x_2 < \cdots < x_k \leq n$. Now, we have $$ \sum_{i=1}^{n}iy_i \equiv b \pmod{n+1} \Longleftrightarrow x_1+\cdots +x_k\equiv b \pmod{n+1}. $$ Therefore, finding the number of codewords in $VT_b(n)$ with Hamming weight $k$ boils down to finding the number of solutions $\langle x_1,\ldots,x_k \rangle \in \mathbb{Z}_{n+1}^k$ of the linear congruence $x_1+\cdots +x_k\equiv b \pmod{n+1}$, with all $x_j$ positive and distinct modulo $n+1$, and with disregarding the order of the coordinates. The result now follows by a direct application of Theorem~\ref{main thm dist pos ai=1}. \end{proof} \begin{rema} There is an earlier interesting result of Dolecek and Anantharam \cite{DOAN} which gives the formula (\ref{VT exa k1s: for}) in a special case where the Hamming weight is dependent on the modulus, but here we give a more general treatment where the Hamming weight is \textit{arbitrary}. Of course, the expression (3.7) in their paper is exactly the same as our formula (\ref{main thm dist ai=1: for}), so it is an interesting problem to prove a 1-1 correspondence between these two results. \end{rema} \section{More connections}\label{Sec_4} Interestingly, some special cases of the functions $P_n(k,b)$, $N_n(k,b)$, $T_n(b)$, and $T'_n(b)$ that we studied in this paper have appeared in a wide range of combinatorial problems, sometimes in seemingly unrelated contexts. Here we briefly mention some of these connections. It would be interesting to prove 1-1 correspondences between these interpretations. \bigskip \textbf{Ordered partitions acted upon by cyclic permutations.} Consider the set of all ordered partitions of a positive integer $n$ into $k$ parts acted upon by the cyclic permutation $(1 2 \ldots k)$. Razen, Seberry, and Wehrhahn \cite{RSW} obtained explicit formulas for the cardinality of the resulting family of orbits and for the number of orbits in this family having exactly $k$ elements. These formulas coincide with the expressions for $P_n(k,0)$ and $P_n(k,1)$, respectively, when $n$ or $k$ is odd (see Corollary~\ref{special cases: b=0,1}). Razen et al. \cite{RSW} also discussed an application in coding theory in finding the complete weight enumerator of a code generated by a circulant matrix. \bigskip \textbf{Permutations with given cycle structure and descent set.} Gessel and Reutenauer \cite{GERE} counted permutations in the symmetric group $S_n$ with a given cycle structure and descent set. One of their results gives an explicit formula for the number of $n$-cycles with descent set $\lbrace k \rbrace$, which coincides with the expression for $P_n(k,1)$ when $n$ or $k$ is odd. \bigskip \textbf{Fixed-density necklaces and Lyndon words.} If $n$ or $k$ is odd then the expressions for $P_n(k,0)$ and $P_n(k,1)$ give, respectively, the number of fixed-density binary necklaces and fixed-density binary Lyndon words of length $n$ and density $k$, as described by Gilbert and Riordan \cite{GIRI}, and Ruskey and Sawada \cite{RUSA}. \bigskip \textbf{Necklace polynomial.} The function $T_n(b)$ is closely related to the polynomial $$ M(q, n)= \frac{1}{n}\sum_{d\, \mid \, n}\mu(d)q^{\frac{n}{d}}, $$ which is called the \textit{necklace polynomial} of degree $n$ (it is easy to see that $M(q, n)$ is integer-valued for all $q \in \mathbb{Z}$). In fact, if $n$ is odd then $M(2, n)=T_n(1)$. The necklace polynomials turn up in various contexts in combinatorics and algebra. \bigskip \textbf{Quasi-necklace polynomial.} The function $T'_n(b)$ is also closely related to the polynomial $$ M'(q, n)= \frac{1}{2n}\sum_{d\, \mid \, n}\mu(d)q^{\frac{n}{d}}, $$ that we call the \textit{quasi-necklace polynomial} of degree $n$. In fact, if $n$ is odd then $M'(2, n)=T'_n(1)$. The quasi-necklace polynomials also turn up in various contexts in combinatorics. For example, they appear as: \begin{itemize} \item the number of transitive unimodal cyclic permutations obtained by Weiss and Rogers \cite{WERO} (motivated by problems related to the structure of the set of periodic orbits of one-dimensional dynamical systems) using methods related to the work of Milnor and Thurston \cite{MITH}. See also \cite{THIB} which gives a generating function for the number of unimodal permutations with a given cycle structure; \item the number of periodic patterns of the tent map \cite{AREL}. \end{itemize} \section*{Acknowledgements} The authors would like to thank the anonymous referees for a careful reading of the paper and helpful suggestions. During the preparation of this work the first author was supported by a Fellowship from the University of Victoria (UVic Fellowship).

https://arxiv.org/abs/1910.08182

Approximate solutions of one dimensional systems with fractional derivative

The fractional calculus is useful to model non-local phenomena. We construct a method to evaluate the fractional Caputo derivative by means of a simple explicit quadratic segmentary interpolation. This method yields to numerical resolution of ordinary fractional differential equations. Due to the non-locality of the fractional derivative, we may establish an equivalence between fractional oscillators and ordinary oscillators with a dissipative term.

\section{Introduction} The study of fractional derivatives for its application in classical and quantum physics has lately received a lot of attention \cite{KST,HER,UCH}. Needless to say that one of the simplest and most studied of those systems is the one dimensional harmonic oscillator. Thus, it would be a good point of departure in the study of systems with fractional derivative, a task which has been carried out in \cite{MAI}. Damped oscillator with fractional derivative has been also the objective of some studies, see \cite{OO}. Some extensions of the theory to other classical systems have been proposed, see for instance \cite{OO1} and references therein, or in \cite{CEY}. In many of these papers, it was noted an analogy between a fractional oscillator and a classical oscillator with a damping term. This could be an idea to be exploited in order to model quantum systems with dissipation, in which the second derivative of the wave function in the Schr\"odinger equation be replaced by a fractional derivative, see \cite{OO2}. The present work has been inspired by the article by Narahari et. al. \cite{AHEC}, where in addition to the study of the one dimensional harmonic oscillator with fractional derivative, they give a comparison with an equivalent dissipative oscillator described on the phase plane and analyze the stability of the solution. Along the present manuscript, we show that it may be possible the determination of a time interval in which the solution of a fractional one dimensional oscillator may be approximated by the solution of a one dimensional ordinary equation with a dissipative term. The idea could be described by using a very simple example. Let us consider the Caputo derivative $D^\alpha_0$, defined in \eqref{1} below, and the fractional differential equation $D^\alpha_0\,x(t)=0$, with $1<\alpha\le 2$ and initial conditions $x(0)=0$ and $\dot x(0)=-1$. The solution is $x(t)=-t$. Then, let us consider the equation $\ddot z(t)=-p\dot z(t)$, $p>0$. The goal is the determination of a value of $p$ such that the solution of this equation with initial conditions $z(0)=0$ and $\dot z(0)=-1$, i.e., the same initial conditions imposed to the fractional equation, be approximated by the solution of the fractional equation over a finite time interval. This is clear, since the solution of the equation on $z(t)$ is $$ z(t)=\frac 1p (-1+e^{-pt})\,. $$ Therefore, on a time interval $0<t<\tau$, with $\tau=1/p$, we have $z(t)=-t+o(t^2)$. In this sense of having similar approximate solutions on a finite interval, we say that the fractional and the dissipative equations are {\it equivalent}. Here, we want to extend this idea. Observe that in our notion of equivalence, we have discarded the asymptotic regime. This is essentially due to two reasons: i.) for large values of time, the fractional oscillator does not show oscillations; ii.) the behaviour of the oscillator from a strictly physical point of view, whether linear or non-linear but particularly the latter, has interest for finite times only. Its asymptotic behaviour is not measurable and has a mathematical interest only, and it is not the object of our study. The present paper is organized as follows: On Section 2, we construct a method to obtain approximate solutions of fractional differential equations, with fractional Caputo derivative to be defined there, based on segmentary interpolation. This kind of interpolation has been used successfully to obtain approximate solutions of ordinary differential equations \cite{GL}. On Section 3, we apply this method to the fractional linear oscillator and to some other simple examples and make estimations on its precision. We compare results with those obtained replacing the fractional oscillator by the ordinary oscillator with a dissipative term. We present a similar analysis by replacing the equation of the oscillator by the van der Pol equation on Section 4. We close this paper with some concluding remarks. \section{Caputo fractional derivative and its evaluation by segmentary interpolation} Let $\alpha$ be a real positive number and denote by $n=\lceil \alpha \rceil$ the smaller integer bigger than $\alpha$. Let us define the Caputo fractional derivative, $D^\alpha_a$, of a $n$ times differentiable function of real variable, $x(t)$, as \cite{DIE} \begin{equation}\label{1} D^\alpha_a \,x(t) = \frac 1{\Gamma(n-\alpha)} \int_a^t \frac{x^{(n)}(s)}{(t-s)^{\alpha-n+1}} \,ds\,, \end{equation} where $x^{(n)}(s)$ means the $n$-th derivative of the function $x(s)$. Our objective, as mentioned at the header of the present section, is an evaluation of \eqref{1} using segmentary interpolation. Here, we consider that $0<\alpha<1$, so that the only choice for $n$ is $n=1$, and this will be the case for some of our applications. Segmentary interpolation is a standard tool of wide use in the approximation of solutions of differential equations \cite{HNW}. Let us sketch the method here for completeness, using an approach that has been used in previous articles by our group \cite{GL,FLS}. Let $[a,b]$ be a compact interval in the real axis $\mathbb R$. At regular intervals, we select $n$ nodes, $a= t_0<t_1<\dots<t_n=b$, with $t_k-t_{k-1}=h$, for all $k=1,2,\dots,n$, so that $kn=b-a$. Let $x(t):[a,b] \longmapsto\mathbb C$ be a continuous function and use the notation $x_k:=x(t_k)$ and $I_k:=[t_{k-1},t_k]$\,, for all $k=1,2,\dots,n$. Then, a quadratic segmentary interpolator $S(t)$ for the function $x(t)$, is a continuous function $S(t):[a,b]\longmapsto \mathbb C$, with first continuous derivative, such that 1.- On each interval of the form $I_k=[t_{k-1},t_k]$, $k=1,2,\dots,n$, we have that $S(t)\equiv P_k(t)$, where $P_k(t)$ is a polynomial of order two, depending on the given interval. 2.- The function $S(t)$ interpolates $x(t)$, in the sense that for any of the nodes $\{t_k\}$, one has that \begin{equation}\label{2} P_k(t_{k-1})=x_{k-1}\,,\qquad P_k(t_k)=x_k\,, \qquad k=1,2,\dots,n\,. \end{equation} The condition on the continuity of the derivative $S'(t)$ implies that \begin{equation}\label{3} P'_k(t_k) =P'_{k+1}(t_k)\,, \qquad k=1,2,\dots,n-1\,. \end{equation} Thus, the construction of the segmentary interpolator $S(t)$ relies in the construction of the interpolating polynomials $P_k(t)$. We propose the following form for the interpolating polynomials: For each of the intervals $I_k$, let us define, \begin{equation}\label{4} P_k(t) =p_k(t)+a_k(t-t_{k-1})(t-t_k)\,, \end{equation} with \begin{equation}\label{5} p_k(t) =\frac{t-t_{k-1}}{h}\,x_k-\frac{t-t_k}{h}\,x_{k-1}\,, \end{equation} and the complex coefficients $a_k$ are given by \begin{equation}\label{6} a_k= \sum_{j=0}^n c_{k,j}\,x_j\,. \end{equation} We still need to determine the values of the $c_{k,j}$, which are \begin{equation}\label{7} c_{j,k} = \left\{ \begin{array}{ll} \frac{(-1)^k}{h^2}\,\eta_1\,, & {\rm if}\; j=0\,, \\[2ex] \frac{(-1)^{k+1}}{h^2}\,(2\eta_1+\eta_2)\,, & {\rm if}\; j=1\,, \\[2ex] \frac{(-1)^{k+j}}{h^2}\,(\eta_{j-1}+2\eta_j+\eta_{j+1}) \,, & {\rm if}\; 1<j<n-1\,, \\[2ex] \frac{(-1)^{k+n-1}}{h^2}\,(2\eta_{n-2}+\eta_{n-1}) \,, & {\rm if}\; j=n-1\,, \\[2ex] \frac{(-1)^{k+n}}{h^2}\, \eta_{n-1} \,, & {\rm if}\; j=n \,, \end{array} \right. \end{equation} where $\eta_j=j/n$ if $j\le k-1$ and $\eta_j=j/n-1$ if $j>k-1$. Taking into account that $S(t)$ is an approximation of $x(t)$, on each of the nodes $t_k$ the Caputo fractional derivative \eqref{1} is approximated by \begin{equation}\label{8} D^\alpha_{a} x(t_k) \approx \frac{1}{\Gamma(1-\alpha)} \sum_{j=1}^k \int_{t_{j-1}}^{t_j} \frac{P'_j(s)}{(t_k-s)^\alpha}\,ds\,,\qquad k=1,\dots,n\,. \end{equation} Then, on each of the intervals $I_k$, we have that $P'_k(t)=\alpha_k\,t+\beta_k$, with \begin{equation}\label{9} \alpha_k=2a_k\,,\qquad \beta_k= \frac{x_k-x_{k-1}}{h} - a_k(2t_{k-1}+h)\,, \end{equation} and, consequently, equation \eqref{8}, takes the following form: \begin{equation}\label{10} D^\alpha_{a} x(t_k) \approx \frac{1}{\Gamma(1-\alpha)} \sum_{j=1}^k \widetilde c_{k,j}\,\alpha_j + \widetilde d_{k,j}\,\beta_j\,. \end{equation} The new coefficients $\widetilde c_{k,j}$ and $\widetilde d_{k,j}$ are given by \begin{equation}\label{11} \widetilde c_{k,j} = \int_{t_{j-1}}^{t_j} \frac{s\,ds}{(t_k-s)^\alpha}\,,\qquad \widetilde d_{k,j} = \int_{t_{j-1}}^{t_j} \frac{ds}{(t_k-s)^\alpha}\,, \end{equation} which obviously depend solely of the partition. It is customary to choose $a=x_0=0$, which obviously does not restrict generality. Since the integrals in \eqref{11} are easily solvable and we know expressions for $\alpha_j$ and $\beta_j$, we can write the right hand side in \eqref{10} as \begin{equation}\label{12} D^\alpha_{a} x(t_k) \approx \frac{1}{(-1+\alpha)(-2+\alpha)\,\Gamma(1-\alpha)} \, \sum_{j=1}^k \gamma_{k,j}\,\alpha_j\,, \end{equation} where, \begin{eqnarray}\label{13} \gamma_{k,j} = [h(-j+k)]^{-\alpha} \,[h(1-j+k)]^{-\alpha} \, \Big\{ h(-1+j-k) \,[h(-j+k)]^\alpha \nonumber\\[2ex] \times [-2h(-2+j+k) +h (-3+2j)\alpha-2(-2+\alpha)t_{j-1}] \nonumber\\[2ex] -h(j-k)\,[h(1-j+k)]^\alpha \, [-2h(-1+j+k)+h(-1+2j)\alpha-2(-2+\alpha) t_{j-1}] \Big\}\,. \end{eqnarray} This is a quite simple and workable receipt to obtain, once $x(t)$ is given, the values of its Caputo fractional derivative at the nodes $t_k$, so that we have an estimation of this fractional derivative. It is interesting to remark that, due to the linear dependence on $\{x_n\}$ of the coefficients $\alpha_j=2a_j$ given in \eqref{6}, then, the derivative $D^\alpha_{a} x(t_k)$ in \eqref{12} can be explicitly determined from $x(t)$. \subsection{A type of differential equations with fractional derivative} Let $x(t):[a,b]\longmapsto \mathbb R^m$ be a differentiable real function of the real variable $t$ and $f(t,x):[a,b]\times\mathbb R^m\longmapsto \mathbb R^m$. In addition, we assume that $x(t^*)=x^*$, where $t^*$ is one of the nodes $\{t_k\}$, $a\le t^*\le b$, and $0<\alpha<1$. Then, let us consider the following fractional differential equation: \begin{equation}\label{14} D^\alpha_a x(t) = f(t,x(t))\,. \end{equation} The objective is to obtain an approximation for the solution of equation \eqref{14} under the condition $x(t^*)=x^*$. We already know how to obtain the identity \eqref{14} in the nodes $t_k$. Take these nodes with the exception of $t^*$. Then, \eqref{14} provides of an algebraic system of equations where the indeterminates are $\{x_{j,k}\}_{j=1,k=0}^{m,n}$ with $x_{j,k}:=x_j(t_k)$ and $x_{j,k}\ne x^*_j=x_j(t^*)$. This algebraic system may or may not be linear depending on the form of $f(t,x(t))$ and is of order $(mn)\times(mn)$. A numerical solution of this system could be obtained by whatever method, which gives a segmentary solution $S(t)$, which is given once one has obtained the coefficients $a_k$ defined in \eqref{6}. In the particular case in which $f(t,x(t))$ contains an eigenvalue $\lambda$ and $f$ be linear with respect to $(\lambda,x)$ this algebraic system is linear and homogeneous. The eigenvalue is determined in the usual way. As the reader may easily understand, this method is more general than the usual way to obtain a solution knowing an initial value, since now $t^*$ could be any node. In particular, the restriction to the solution that replaces the initial value condition could be imposed at $t^*$, and this represents a great advantage when compared with the shooting method worked out in \cite{DIE1,DB}. \section{The fractional linear oscillator} A simple example of an equation of the type \eqref{14} is the linear oscillator with the fractional derivative, which is defined as \begin{equation}\label{15} D_0^\alpha\,x(t) =-\omega^2\,x(t)\,. \end{equation} As in the standard harmonic oscillator the constant $\omega^2=k/m$, where $m$ is the oscillator mass and $k$ a constant, $\alpha$ being the order of derivation that in the present case we assume to be $1<\alpha \le 2$. Using the definition \eqref{1}, taking into account that for some differentiable function $f(t)$ (in our case $f(t)=x(t)$ or $f(t)=\dot x(t)$, where the dot means first derivative), we have that \begin{equation}\label{16} \lim_{\alpha\to 0^+} \frac1{\Gamma(\alpha+1)}\, t^\alpha\,f(0)=f(0)\,, \end{equation} and that $n$ is either 2 or 3, we may integrate by parts \eqref{15} using \eqref{1}, which gives the following integral version of \eqref{2}: \begin{equation}\label{17} x(t) = x(0)+\dot x(0) -\frac{\omega^2}{\Gamma(-\alpha)}\int_0^t (t-s)^{-\alpha-1}\,x(s)\,ds\,. \end{equation} The general solution has the form \begin{equation}\label{18} x(t) = c_1\, E_{\alpha,1}(-\omega^2 t^{\alpha}) + c_2\,t\, E_{\alpha,2} (-\omega^2 t^{\alpha})\,, \end{equation} where $E_{\alpha,\beta}(z)$ is the so called Mittag-Leffler function \begin{equation}\label{19} E_{\alpha,\beta}(z) = \sum_{k=0}^\infty \frac{z^k}{\Gamma(\alpha k+\beta)}\,. \end{equation} Thus, in order to obtain a particular solution, we have to impose some initial conditions. For instance, if we choose $x(0)=1$ and $\dot x(0)=0$, the solution to \eqref{17} with these initial conditions is given by \begin{equation}\label{20} x(t)= E_{\alpha,1} (-\omega^2 t^{\alpha})\,. \end{equation} Let us find a particular numerical solution to the fractional linear oscillator, using the method introduced in Section 2.1. We have to choose a particular value for $\omega$ and the simplest possibility is $\omega=1$. This is developed in the forthcoming subsection. \subsection{Some numerical estimations.} First of all, it is not the objective here to give explicit expressions for the approximate solutions for the studied examples. It is not difficult to plot these solutions for different values of $n$. Let us start with equation \eqref{15} with $\omega=1$ on the interval $0\le t\le 1$, with $0<\alpha<1$ and initial condition $x(0)=1$. As seen above, this equation has exact solution given by $x_{\rm exact}(t)=E_{\alpha,1}(-t^\alpha)$ \cite{DIE}. The objective is now an estimation on the precision of the proposed method. As customary, this precision is measured by using the following parameter: \begin{equation}\label{21} e_n(\alpha)= \int_0^1 [x_{\rm exact}(t)-x_n(t)]^2\,dt\,, \end{equation} Here, $n$ is the number of sub-intervals $I_n$ in which we partite $[0,1]$, the number of nodes being $n+1$. The dependence of this parameter on $\alpha$ shows that the smaller is the value of $\alpha$, or equivalently the closer is $\alpha$ to zero, the lower is the precision and, therefore, the slower is the convergence to the exact value. However, we do not observe significative variations on the precision when we increase the value of $n$, i.e., as we make the sub-intervals smaller and smaller. \vskip1cm \centerline{$ \begin{array} [c]{cccc} n & e_{n}(0.1) & e_{n}(0.5) & e_{n}(0.9)\\[2ex] 5 & 7.4\text{ }10^{-3} & 1.7\text{ }10^{-3} & 2.0\text{ }10^{-4}\\ 10 & 3.0\text{ }10^{-3} & 4.9\text{ }10^{-4} & 2.8\text{ }10^{-5}\\ 20 & 1.1\text{ }10^{-3} & 1.4\text{ }10^{-4} & 5.9\text{ }10^{-6}\\ 40 & 3.0\text{ }10^{-4} & 3.7\text{ }10^{-5} & 1.3\text{ }10^{-6} \end{array} $} \medskip \centerline{Table 1.- Values of the precision in terms of $n$ and $\alpha$.} \vskip1cm This can be seen in Table 1, where we have chosen values of $n$ ranging from 5 to 40. The values of $\alpha$ studied are 0.1, 0.5 and 0.9. Let us study the precision of the method with another example different from the fractional oscillator, yet an equation of the form \eqref{14}. Here, we have chosen, \begin{equation}\label{22} D^{1/2}_0 x(t) =\sin x(t)\,, \end{equation} on the interval $[0,1]$, with the initial condition $x(1)=5/2$, which was already studied in \cite{DIE1}, where the integration was performed by means of the iterative shot method. Contrarily to the previous example, here we do not know an exact solution. The way out is to define the precision as \begin{equation}\label{23} e_n= \int_0^1 [D^{1/2}_0 x_n(t) -\sin x_n(t)]^2\, dx\,, \end{equation} where $n$ is again the number of intervals and $x_n(t)$ is the interpolating function for the studied case. After integration and using the boundary condition, we obtain $x(0)$. Along with \eqref{23}, we introduce another parameter that measures the convergence and that we denote as $e_r\%$. It represents the relative variation between the value of $y(0)$, obtained for a given value of $n$, and the value given for the precedent value of $n$ as listed on Table 2. Table 2 is just a sample of numerous numerical examples we have performed. This sample is significative, as it manifest an obvious convergence and shows that the result obtained for a small number of nodes is satisfactory. \vskip1cm \centerline{$ \begin{array} [c]{cccc} n & y(0) & e_{r}\% & e_{n}\\[2ex] 5 & 1.74895 & -- & 1.8\text{ }10^{-2}\\ 10 & 1.73812 & 0.5 & 7.4\text{ }10^{-3}\\ 20 & 1.73326 & 0.3 & 2.6\text{ }10^{-3}\\ 30 & 1.73166 & 0.09 & 1.1\text{ }10^{-3}\\ 40 & 1.73085 & 0.05 & 5.6\text{ }10^{-4} \end{array} $} \medskip \centerline{Table 2: Values of $y(0)$, $e_r\%$ and $e_n$ for a given value of $n$} \vskip1cm Finally, we have performed another reliability test, which was the use of the value of $y(0)$ obtained numerically as initial value and evaluate the value of $y(1)$. In all cases, we have recovered the value $y(1)=5/2$. \subsection{Damped oscillator with entire derivative.} As is well known, the damped oscillator with entire derivative is given by \begin{equation}\label{24} m \ddot y(t)+ p\dot y(t)+k y(y)=0\,, \end{equation} where $m$, $p$ and $k$ are constants. Here, we assume that $p>0$. For $p=0$, the limit for $\alpha\longmapsto 2^-$ in \eqref{15} should give the solution $y(t)$ for \eqref{24}, which we denote here as $\lim_{\alpha\to 2^-}x(t)=y(t)$. The solution $x(t)$ is damped oscillatory on the transitory regime only \cite{AHEC,DIE,KST}. Based on these notions, we propose the following Ansatz: {\it For each given $1<\alpha\le 2$, there exists $p>0$ such that the solution $x(t)$ of \eqref{15} with $\alpha$ is a good approximation of the solution $y(t)$ of \eqref{24} with $p$, in the transitory regime}. Using this Ansatz, let us obtain an approximate solution $y(t)$ for \eqref{24} such that this and its corresponding solution $x(t)$ for \eqref{15} fulfil the conditions $y(0)=x(0)$ and $\dot y(0)=\dot x(0)$. This is: \begin{equation}\label{25} y(t) = \exp\left( -\frac pm\,t \right)\, (c_-\exp(-\Delta t) +c_+ \exp(\Delta t))\,, \end{equation} where, \begin{equation}\label{26} \Delta= \sqrt{(p/m)^2-4\omega^2}\,,\quad \lambda_\pm =\frac{-p\pm\Delta}{2}\,,\quad c_\pm =\pm \frac{\lambda_\mp}{\Delta}\,, \end{equation} where $\omega^2$ was given in \eqref{15}. Then, the point is the determination of the value of $p$ being given the value of $\alpha$, or equivalently the determination of a function $p=p(\alpha)$, in application to our Ansatz. This is an optimal control problem. We have to find the optimal solution, which minimizes the following functional: \begin{equation}\label{27} E(\alpha):= \frac 1T \int_0^T [x(t)-y(t)]^2\,dt\,, \end{equation} where $T$ is some time scale in which the amplitude of the oscillations are reduced by a factor of $1/T$. On the interval $[0,T]$, the transitory regime, we compare the solutions of the fractional derivative $x(t)$ and of the damped oscillator $y(t)$. The functional $E(\alpha)$ measures the deviation between $x(t)$ and $y(t)$. Then, go back to \eqref{20} and note that the function $E_{\alpha,1} (-\omega^2 t^{\alpha})$ is not asymptotically oscillating as $t\longmapsto 0$. This permits us to choose a value of $T$, although not small, not very high either. Numerical experiments have shown that the choice $T=20$ is appropriate. Let us give some numerical results. On Table 3, we give the dependence between values of $\alpha$, $p(\alpha)$ and $E(\alpha)$ for the values $k=m=1$ and $n=50$. \vskip1cm \centerline{$ \begin{array} [c]{ccc} \alpha & p & E\\[2ex] 1.10 & 1.140 & 6.4\,10^{-3}\\ 1.30 & 0.891 & 5.7\,10^{-3}\\ 1.50 & 0.668 & 4.7\,10^{-3}\\ 1.70 & 0.433 & 3.3\,10^{-3}\\ 1.90 & 0.152 & 1.1\, 10{-3}\\ 1.95 & 0.754 & 4.3\, 10^{-4}\\ 2 & 0 & 0 \end{array} $} \medskip \centerline{Table 3: Comparison between the values of $\alpha$, $p$ and $E$,} \centerline{for $T=20$, $k=m=1$ and $n=50$.} \vskip1cm An explicit expression of the function $p(\alpha)$ may be obtained by the least-square method and this gives $p(\alpha)= 1.49409 +0.056127 \,\alpha - 0.401446\,\alpha^2$. This is depicted on Figure 2. On Figure 1, we represent the usual behaviour in the transitory regime for $x(t)$ and $y(t)$, when we choose $\alpha=1.7$ and $n=50$. It is interesting to remark that numerical experiments show that $p(\alpha)$ does not depend on any choice of initial values. \subsection{Non-linear oscillator} Following the discussion on the damped oscillator, we present a similar problem given by the following non-linear oscillator: \begin{equation}\label{28} D_0^\alpha\,x(t) =y(t)\,, \qquad D_0^\alpha\,y(t) =-\sin x(t)\,, \end{equation} with $0<\alpha\le 1$. Let us choose the initial values given by $x(0)=1$ and $y(0)=0$. Clearly, for small oscillations equation \eqref{28} becomes equation \eqref{15} with the replacement $\alpha \rightarrow 2\alpha$. Again, this is a non-linear problem having no analytic solution for $\alpha$ non-integer. Then, we proceed by analogy with the damped oscillator. In this case, we consider the following system involving entire derivatives only: \begin{equation}\label{29} \dot z(t)=w(t)\,,\qquad \dot w(t)=-p w(t)-\sin z(t)\,,\qquad p>0\,. \end{equation} On the transitory regime, we compare the solutions of systems \eqref{28} and \eqref{29} under the conditions $z(0)=x(0)=1$ and $w(0)=y(0)=0$. To do this, we need the previous determination of $p(\alpha)$, which we assume that minimizes the following quadratic dispersion: \begin{equation}\label{30} E(\alpha)=\frac 1T \int_0^T \{ [x_n(t)-z(t)]^2 +[y_n(t)-w(t)]^2 \}\,dt\,. \end{equation} Obviously, this expression generalizes \eqref{27}. Again, we adjust the value of $T$ by numerical experiments, which show that $T=20$ is, again, a convenient choice. On Table 4, we give some values for the dependence between $\alpha$, $p(\alpha)$ and $E(\alpha)$ after the choice $T=20$ and $n=50$. \vskip1cm \centerline{$ \begin{array} [c]{ccc} \alpha & p & E\\[2ex] 0.50 & 1.203 & 5.5\,10^{-3}\\ 0.70 & 0.757 & 3.9\,10^{-3}\\ 0.90 & 0.294 & 2.1\,10^{-3}\\ 0.95 & 0.148 & 1.4\,10^{-3}\\ 1.00 & 0 & 0 \end{array} $} \medskip \centerline{Table 4: Comparison between the values of $\alpha$, $p$ and $E$,} \centerline{for the values $T=20$ and $n=50$.} \vskip1cm The above examples manifest an analogous behaviour between a fractional linear oscillator and a damped or even non-linear oscillator on some time interval. The solutions between the fractional and the integer equation are very similar on some time scale. This could be a rather general situation, so that in many practical cases and inside a time interval, we conjecture that a fractional operator might be replaced by the frictional additional term on the classical oscillator. The behaviour of the solutions is similar to that shown in Figure 1. \begin{figure} \centering \includegraphics[width=0.4\textwidth]{OLNL} \caption{\small { The continuous line represents the solution, $x(t)$ of the fractional equation \eqref{15}, which is given by \eqref{25}. The dashed line gives the solution, $y(t)$, of the equation with ordinary derivative \eqref{25}.} } \label{} \end{figure} \begin{figure} \centering \includegraphics[width=0.4\textwidth]{PALF} \caption{\small Function $p(\alpha)$. } \label{} \end{figure} \section{On the fractional van der Pol equation} The van der Pol equation is a second order non-linear ordinary differential equation \cite{VDP,FAR}. It has the following form: \begin{equation}\label{31} \ddot x(t) - \mu(1-x^2(t))\,\dot x(t) + x(t)=0\,, \end{equation} where $\mu\ge 0$ is a constant. When $\mu=0$, \eqref{31} is the equation of the ordinary harmonic oscillator. The van der Pol equation may be written in terms of a first order system as \begin{equation}\label{32} \dot x(t):= z(t)\,,\qquad \dot z(t)=-\mu (x^2(t)-1) z(t) -x(t)\,, \qquad \mu\ge 0\,. \end{equation} This equation has a unique limit cycle for $\mu\ne 0$, after the Li\'enard theorem \cite{STR}. This suggested us the interest of considering the possible existence of a limit cycle for the fractional system analogous to \eqref{32} given by \begin{equation}\label{33} D_0^\alpha\, x(t)=z(t)\,,\qquad D_0^\alpha \,z(t)= -\mu z(t) (x^2(t)-1)-x(t)\,,\qquad 0<\alpha\le 1. \end{equation} One fractional van der Pol equation of the type \begin{equation}\label{34} D_0^{\alpha+1}\, x(t) +\mu (x^2(t)-1) D_0^\alpha \,x(t) +x(t)=0\,, \end{equation} has been studied in \cite{GUO}, where a relation between the parameters $\alpha$ and $\mu$ was given as a sufficient condition for the existence of a limit cycle, using the balance harmonic method \cite{DEU}. We have studied the system \eqref{33} through numerical as well as analytic methods. We have performed a big amount of numerical experiments, which have shown the existence of a value of the parameter $\mu$, here called $\mu_c$, where the subindex $c$ stands for {\it critical}, which depends on $\alpha$ and $\mu_c(\alpha)>0$, such that \begin{itemize} \item{For values of $\mu$ with $0<\mu<\mu_c$, there is a fixed point $(x^*,z^*)=(0,0)$, which remains stable at the limit $t\longmapsto\infty$, $\lim_{t\to\infty}(x(t),z(t))=(0,0)$. Therefore, there is no stable limit cycle. In addition, there is no evidence of the existence of an unstable limit cycle.} \item{For values $\mu>\mu_c$, the fixed point $(x^*,z^*)=(0,0)$ is unstable. We found a unique stable limit cycle. In this case a Hopf bifurcation emerges with $\mu_c$ as critical parameter.} \item{As shown in Figure 3, $\mu_c(\alpha)$ decreases with $\alpha$ and $\lim_{\alpha \to1}\mu_c(\alpha)=0$.} \end{itemize} The point here is to show the existence of the critical value for the parameter $\mu$ for a given value of $0<\alpha<1$, $\mu_c(\alpha)$. This existence has been manifested by the numerical estimations above. Nevertheless, this existence may be also shown analytically. To this end, we use the following result \cite{AEE}: Let us consider the following system, where $D_0^\alpha$ represents the Caputo fractional derivative: \begin{equation}\label{35} D^\alpha_0 \, x(t)=f(x,z)\,,\qquad D^\alpha_0\,z(t)= g(x,z)\,,\qquad 0<\alpha<1\,. \end{equation} A solution $(x^*(t),z^*(t))$ is in equilibrium if $f(x^*(t),z^*(t))=g(x^*(t),z^*(t))=0$. It is asymptotically stable if the eigenvalues, $\lambda$, of the Jacobian matrix \begin{equation}\label{36} J(x,z):= \left(\begin{array}{cc} \partial f/\partial x & \partial f/\partial z \\[2ex] \partial g/\partial x & \partial g/\partial z \end{array} \right)\,, \end{equation} when evaluated at the equilibrium point satisfies \begin{equation}\label{37} |\arg(\lambda)|>\alpha\,\frac \pi 2\,. \end{equation} A comparison between \eqref{33} and \eqref{35} gives the precise form of $f(x,z)$ and $g(x,z)$ for our particular case. This gives the precise form of \eqref{36} as \begin{equation}\label{38} J(x,z):= \left(\begin{array}{cc} 0 & 1 \\[2ex] -1-2\mu\,z(t)x(t) & -\mu(x^2(t)-1) \end{array} \right)\,. \end{equation} Taking into account that the fixed point is located at $(x^*,z^*)=(0,0)$, we have that \begin{equation}\label{39} J(0,0)=\left(\begin{array}{cc} 0 & 1\\[2ex] -1 & \mu \end{array} \right)\,, \end{equation} which has the following eigenvalues: \begin{equation}\label{40} \lambda_\pm= \frac 12 (\mu\pm \sqrt{\mu^2-4})\,. \end{equation} Obviously, if $\mu \ge 2$, then $\arg(\lambda_\pm)=0$. On the other hand, if $\mu<2$, one has that, \begin{equation}\label{41} \arg(\lambda_\pm)= \arctan \left(\pm\sqrt{\left( \frac 2\mu\right)^2-1} \right)\,. \end{equation} From \eqref{37}, the critical value, $\mu_c$, of $\mu$ should obey the following relation: \begin{equation}\label{42} \arctan \left(\pm\sqrt{\left( \frac 2\mu_c\right)^2-1} \right) = \alpha\,\frac \pi 2\,, \end{equation} which gives \begin{equation}\label{43} \mu_c= \frac{2}{\displaystyle \sqrt{1+\tan^2\left( \frac{\alpha\pi}{2}\right)}}\,. \end{equation} \begin{figure} \centering \includegraphics[width=0.4\textwidth]{MUC} \caption{\small Function $\mu_c(\alpha)$. } \label{} \end{figure} This is to say, if we fix $\alpha$ and start from $\mu\approx 0$, as we increase $\mu$, we go from a situation with a asymptotically stable fixed point to an unstable point. This happens when $\mu>\mu_c$. The transition from the stability to the unstability drives to the emergency of a limit cycle. We may qualitatively interpret the limit cycle loss as follows: let us consider $\mu\approx 0$ in \eqref{33}, which may then be approximated by \begin{equation}\label{44} D_0^\alpha \,x(t) = z(t)\,,\qquad D_0^\alpha \,z(t)=-x(t)\,. \end{equation} This is the same than equation \eqref{28} with the paraxial approximation $\sin y(x)\approx y(x)$. Note that \eqref{44} does not show a limit cycle and, further, the trivial solution $(0,0)$ is an attractor. The second equation in \eqref{33} contains the term $-\mu\,z(x)(y^2(x)-1)$, which in the case of $\mu>\mu_c$ is not negligible. This fact outbalances the dissipation and this is precisely which makes it possible the existence of a limit cycle. On Figure 4 and on the phase plane, the continuous and slashed curves represent the solution with entire and fractional derivative, respectively. Both trajectories are determined with same initial values and same parameter $\mu$. In all cases, the fractional limit circle is enclosed by the trajectory of the limit cycle with entire derivative. On Figure 3, we show the relation $\mu_c=\mu_c(\alpha)$. In the region above the curve, there exists a stable cycle limit and, furthermore, the fixed point $(0,0)$ is unstable. Below the curve the limit cycle does not exist and the fixed point is asymptotically stable. There is an obvious difference with the results obtained in \cite{GUO}, which is due to the fact that the fractional equations \eqref{33} and \eqref{34} are {\it not} equivalent. \subsection{Equivalence between the fractional van der Pol equation and the same equation with entire derivative and dissipation.} On the previous section, we have compared the approximate solutions of a dissipative oscillator with entire derivative with those of the linear oscillator with fractional derivative. Now, we want to carry out a similar analysis with the fractional van der Pol equation and a van der Pol equation with entire derivative and a dissipative term of the form $\beta z(t)$, $\beta>0$. This system has the form, \begin{equation}\label{45} x'(t)=z(t)\,,\qquad z'(t)=-z(t)(\beta+\mu(x^2(t)-1))-z(t)\,, \qquad \mu\,,\beta>0\,. \end{equation} Here, the fixed point is $(x^*,z^*)=(0,0)$. To check its stability, we consider again equation \eqref{37}, which in the present case gives at the fixed point the following expression \begin{equation}\label{46} J(0,0)= \left(\begin{array}{cc} 0 & 1 \\[2ex] -1 & \mu-\beta \end{array}\right)\,, \end{equation} which has the eigenvalues \begin{equation}\label{47} \lambda_\pm = \frac{(\mu-\beta) \pm \sqrt{(\mu-\beta)^2-4}}2\,. \end{equation} Therefore, the fixed point is stable if Re$(\lambda_\pm)<0$ and unstable if Re$(\lambda_\pm)>0$, or equivalently, if $\mu-\beta<0$ and $\mu-\beta>0$, respectively. Then, for each $\mu$, there exists a $\beta_c=\mu$, where the Hopf bifurcation of the fixed point appears and, consequently, the destruction of the limit cycle. In any case, according to the Li\'enard theorem \cite{STR,FAR}, we may show that there exists a unique stable limit cycle if $\mu-\beta<0$. In consequence, the van der Pol equations with entire derivative and dissipation and fractional have qualitatively the same properties. We have checked numerically a qualitative equivalence, in the sense of having approximately the same solution, through a substantial number of numerical experiments, between equations \eqref{44} (with fractional derivative) and \eqref{45} (with entire derivative). Thus by trial and error, we have determined a value of $\beta$ giving the same cycle in both cases. For instance, if we give the values $\alpha=0.9$ and $\mu=0.1$, we obtained $\beta\cong 0.315$. In an analogous manner, we trial with values for which $\mu<\mu_c$ and obtained similar conclusions. On Figure 5, we represent limit cycles for the fractional and dissipative entire van der Pol equations. \begin{figure} \centering \includegraphics[width=0.4\textwidth]{CLEF} \caption{\small Comparison between the entire (continuous curve) and the fractional (slashed curve) van der Pol solutions, with the same value of $\mu$. } \label{} \end{figure} \begin{figure} \centering \includegraphics[width=0.4\textwidth]{CERF} \caption{\small Comparison between the damped (continuous curve) and the fractional (slashed curve) van der Pol solutions. } \label{} \end{figure} \section{Concluding remarks} We have applied a quadratic spline method in order to obtain functions that approximate the result of applying a fractional derivative to a given function. This is suitable to obtain solutions to some differential equations with initial values or mixed conditions of potential interest in Physics or Engineering. We have tested our method with the fractional linear oscillator, where exact solutions are known and checked its degree of precision. Results of numerous numerical experiments show that for values of $\alpha$ in the range $0<\alpha<1$, the higher is $\alpha$, the better is the precision of our method. Here, $\alpha$ is the order of the fractional derivation, $D^\alpha$. However, there are not substantial differences when we increase the number of nodes on the interval under our consideration. Similar results have been obtained for non-linear oscillators. Based on previous studies on the approximation of solutions of the fractional linear oscillator by solutions of a damped oscillator, we have used our method to confirm these results. We have shown that there exists a time interval for which the solutions of both equations are similar with a high degree of accuracy, under the condition that a relation is given between the coefficient $p$ of the dissipative term of the damped oscillator and the order of the fractional derivation, $\alpha$. A similar study compares a fractional and a damped van der Pol equations, written as a system of two equations on phase space, with similar results. In addition, we have considered the behaviour of limit cycles and fixed points in terms of $\alpha$ and a parameter $\mu$ characteristic of the van der Pol equation. Using analytic as well as numerical arguments, we show the existence of a critical value for the parameter $\mu$, $\mu_c$, such that if $\mu<\mu_c$ the origin of phase space is stable and if $\mu>\mu_c$ is unestable. This limit value $\mu_c$ depends on $\alpha$ and we give the exact relation. \section*{Acknowledgements} This research has been financed by the Projects No. ING 19/i 402 and ING 80020180100064 of the Universidad Nacional de Rosario, the Spanish MINECO (Project No. MTM2014-57129) and the Junta de Castilla y Le\'on (Project Nos. BU229P18 and VA137G18).

https://arxiv.org/abs/1402.2456

Equal Sum Sequences and Imbalance Sets of Tournaments

Reid conjectured that any finite set of non-negative integers is the score set of some tournament and Yao gave a non-constructive proof of Reid's conjecture using arithmetic arguments. No constructive proof has been found since. In this paper, we investigate a related problem, namely, which sets of integers are imbalance sets of tournaments. We completely solve the tournament imbalance set problem (TIS) and also estimate the minimal order of a tournament realizing an imbalance set. Our proofs are constructive and provide a pseudo-polynomial time algorithm to realize any imbalance set. Along the way, we generalize the well-known equal sum subsets problem (ESS) to define the equal sum sequences problem (ESSeq) and show it to be NP-complete. We then prove that ESSeq reduces to TIS and so, due to the pseudo-polynomial time complexity, TIS is weakly NP-complete.

\section{Introduction} A tournament is an orientation of a complete simple graph. In a tournament, the \emph{score} $s_{i}$ of a vertex $v_{i}$ is the number of arcs directed away from that vertex, that is, the outdegree of $v_{i}$. The \emph{score sequence} of a tournament is formed by listing the scores in nondecreasing order. Let us write $[x_{i}]_{1}^{n}$ to denote a sequence with $n$ terms. Landau \cite{landau1} gave a simple characterization of the score sequences of tournaments. \begin{theorem}\label{landau} A sequence $[s_{i}]_{1}^{n}$ of non-negative integers in nondecreasing order is the score sequence of a tournament if and only if for every $I\subseteq\{1,2,\ldots,n\}$, \begin{equation}\label{land} \sum_{i\in I}{s_{i}}\geq{\left|I\right|\choose 2}, \end{equation} \noindent with equality when $\left|I\right|=n$, where $\left|I\right|$ is the cardinality of the set $I$. \end{theorem} Several proofs of Landau's theorem have appeared over the years \cite{bang1, brualdi1, griggs1, landau1, thomassen1} and it continues to play a central role in the theory of tournaments and their generalizations. Brualdi and Shen \cite{brualdi2} strengthened Landau's theorem by deriving a set of inequalities that are individually stronger than inequalities (\ref{land}) but are collectively equivalent to these inequalities. The set of scores of vertices in a tournament is called the \textit{score set} of the tournament. Reid \cite{reid1} conjectured that any finite nonempty set $S$ of non-negative integers is the score set of some tournament. He gave a constructive proof of the conjecture for the cases $\left|S\right|=1,2,3$, while Hager \cite{hager1} settled the cases $\left|S\right|=4,5$. In 1986, Yao announced a nonconstructive proof of Reid's theorem by arithmetic arguments \cite{yao1}. Pirzada and Naikoo \cite{pirzada1} obtained the construction of a tournament with a given score set in the special case when the score increments are increasing. However, so far no constructive proof has been found for Reid's theorem in general. In a digraph, the \emph{imbalance} of a vertex $v_{i}$ is defined as $t_{i} = d_{i}^{+} -d_{i}^{-}$, where $d_{i}^{+}$ and $d_{i}^{-}$ are respectively the outdegree and indegree of $v_{i}$. The \emph{imbalace sequence} of a digraph is formed by listing the vertex imbalances in nonincreasing order. If $T$ is a tournament with imbalance sequence $[t_{i}]_{1}^{n}$, we say that $T$ \emph{realizes} $[t_{i}]_{1}^{n}$. Mubayi, Will and West \cite{west1} gave necessary and sufficient conditions for a sequence of integers to be the imbalance sequence of a simple digraph. \begin{theorem}\label{west} A sequence of integers $[t_{i}]_{1}^{n}$ with $t_{1}\geq\cdots\geq t_{n}$ is an imbalance sequence of a simple digraph if and only if $\sum_{i=1}^{j}t_i\leq j(n-j)$, for $1\leq j\leq n$ with equality when $j=n$. \end{theorem} On rearranging the imbalances in nondecreasing order, we obtain the equivalent inequalities $\sum_{i=1}^{j}t_i\geq j(j-n)$, for $1\leq j\leq n$ with equality when $j=n$. Koh and Ree \cite{koh1} showed that if an additional parity condition is satisfied the sequence $[t_{i}]_{1}^{n}$ can be realized by a tournament. In fact, they proved the result in the more general setting of hypertournaments. The following corollary of Theorem 6 in \cite{koh1} provides a characterization of imbalance sequences of tournaments. \begin{theorem}\label{seq} A nonincreasing sequence $[t_{i}]_{1}^{n}$ of integers is the imbalance sequence of a tournament if and only if $n-1, t_{1}, \ldots, t_{n}$ have the same parity and \begin{equation}\label{charac} \sum_{i=1}^{j}{t_{i}} \leq j(n-j), \end{equation} \noindent for $j = 1, \ldots , n$ with equality when $j = n$. \end{theorem} In a digraph, the set of imbalances of the vertices is called its \textit{imbalance set} \cite{pirzada2}. In \cite{pirzada2} the following result regarding the imbalance sets of oriented graphs is proved. \begin{theorem}\label{pir} Let $P = \{p_1 , \ldots, p_m \}$ and $Q = \{-q_1 , \ldots, -q_n \}$, where $p_1 < \cdots < p_m$ and $q_1 < \cdots < q_n$ are positive integers. Then there exists an oriented graph with imbalance set $P \cup Q$. \end{theorem} Due to the interest in Reid's score set theorem, it is natural to ask if a similar result holds for imbalacne sets of tournaments. Furthermore, since a constructive proof of Reid's theorem has not yet been found, it would be interesting to look for an algorithm that generates a tournament from its imbalance set. In this paper we address both questions. We study the following decision problem and its search version. \begin{definition}[\textbf{Tournament Imbalance Set Problem (TIS)}]\label{TIS} Given a set $Z$ of integers, decide if $Z$ is the imbalance set of a tournament. \end{definition} In Section \ref{odd case}, we first show that the obvious necessary conditions for the existence of tournament imbalance sets are not sufficient. We then completely characterize the sets of odd integers that are imbalance sets of tounraments. In Section \ref{even case}, we treat the case of even integers, which is more involved. We show that any set of even integers that contains at least one positive and at least one negative integer or only consists of a single element 0, is the imbalnce set of a partial tournament in which each vertex is joined to every other vertex except one. However, not all such sets are imbalance sets of tournaments. This is followed by necessary and sufficient conditions for a set of even integers to be a tournament imbalance set. In Section \ref{algorithm}, we define a new variant of the \textit{equal sum subsets problem} (ESS) called \textit{equal sum sequences problem} (ESSeq). We show that ESSeq is NP-hard and ESSeq reduces to TIS in polynomial time. Furthermore, we propose a pseudo-polynomial time algorithm that determines if a set of integers is a tournament imbalance set and, in addition, generates a tournmanet realizing any such set. Thus TIS is shown to be weakly NP-complete. We also consider extremal cases and determine upper bounds for the minimal order of a tournament realizing an imbalance set. \section{Characterizing odd imbalance sets} \label{odd case} Consider a tournament of order (number of vertices) $n$. Let $v_{i}$ be a vertex with score $s_{i}$ and imbalance $t_{i}$ then $t_{i}=d_{i}^{+}-d_{i}^{-}=s_{i}-(n-1-s_{i})=2s_{i}-(n-1)$, or by rearranging $s_{i} = \frac{n-1+t_{i}}{2}$. Converesly, assume that $v_{i}$ is a vertex of a tournament with $n$ vertices and $t_{i}$ is the imbalance of $v_{i}$. Let $s_{i} = \frac{n-1+t_{i}}{2}$ then $s_{i}$ is the score of $v_{i}$. Thus we have \begin{lemma}\label{aux} Let $t_{i}$ be the imbalance of a vertex $v_{i}$ in a tournament. Then $s_{i}$ is the score of $v_{i}$ if and only if \begin{equation}\label{inter1} s_{i} = \frac{n-1+t_{i}}{2}, \end{equation} \noindent where $n$ is the order of the tournament. \end{lemma} A tournament is said to be \emph{regular} if all the vertices have the same score \cite{chartrand1}. Clearly, there exists a regular tournament on $n$ vertices with score $s_{i}$ if and only if $n$ is odd and $s_{i}=\frac{n-1}{2}$. Therefore, the imbalance of any vertex $v_{i}$ of a regular tournament is $t_{i}=2s_{i}-(n-1)=0$ and the imbalance set of any regular tournament is $\{0\}$. The following is a set of obvious necessary conditions for an imbalance set of a tournament. \begin{theorem}\label{necessary} If a finite nonempty set $Z$ of integers is the imbalance set of a tournament of order $n$ then all the elements of $Z$ have the same parity as $n-1$ and it either contains at least one positive and at least one negative integer or contains only a single element 0. \end{theorem} \begin{proof} If $Z$ is the imbalance set of a tournament with $n$ vertices then by Theorem \ref{seq}, the elements of $Z$ must have the same parity as $n-1$. Furthermore, either the tournament is regular and $Z=\{0\}$ or $Z$ must contain at least one positive and at least one negative integer so the corresponding imbalance sequence sums to zero. \end{proof} The question is whether these conditions are also sufficient. The answer is `no' as can be seen from the following example. \begin{example} Let $Z=\{6, -10\}$. Then $Z$ satisfies the necessary conditions given in Theorem \ref{necessary} and it can potentially be the imbalance set of a tournament with an odd number of vertices. However, any sequence with elements chosen from $Z$ can sum to zero only if it consists of an even number of elements (e.g., $6$, $6$, $6$, $6$, $6$, $-10$, $-10$, $-10$ ). Thus by Theorem \ref{seq}, we cannot construct a tournament imbalance sequence from $Z$ and so $Z$ is not a tournament imbalance set. \end{example} Although the conditions given in Theorem \ref{necessary} are not sufficient in general, they are sufficient if $Z$ consists of odd integers. We first show that \begin{theorem}\label{odd} Let $X=\{x_{1}, \ldots, x_{l}\}$ and $Y=\{-y_{1},\ldots,-y_{m}\}$ be disjoint nonempty sets of odd integers, where $x_{1}>\cdots> x_{l}$ are positive odd integers and $y_{1}<\cdots< y_{m}$ are also positive odd integers. Let $L=\sum_{i=1}^{l}{x_{i}}$, $M=\sum_{i=1}^{m}{y_{i}}$ and $n=lM+mL$. Then there exists a tournament of order $n$ with imbalance set $X\cup Y$. \end{theorem} \begin{proof} We observe that $n$ is even and all the elements of $X\cup Y$ have the same parity as $n-1$. Let $x^{(p)}$ denote that $x$ is appearing in $p$ consecutive terms of a sequence. We use Theorem \ref{seq} to prove that the $n$-term sequence \[ [t_{i}]_{1}^{n}={x_{1}}^{(M)},\ldots,{x_{l}}^{(M)},{-y_{1}}^{(L)},\ldots,{-y_{m}}^{(L)} \] \noindent is the imbalance sequence of a tournament arranged in nonincreasing order. First note that \[ \sum_{i=1}^{M}{t_{i}}= Mt_{1}=Mx_{1}\leq M((l-1)M+mL)=M(n-M), \] \[ \sum_{i=1}^{2M}{t_{i}}= M(t_{1}+t_{2})=M(x_{1}+x_{2})\leq 2M((l-2)M+mL)=2M(n-2M), \] \[\ldots \ \ \ \ \ldots \ \ \ \ \ldots \ \ \ \ \ldots \ \ \ \ \ldots \] \[ \sum_{i=1}^{lM}{t_{i}}= M\sum_{i=1}^{l}x_{i}=LM\leq lM(mL)=lM(n-lM), \] \[ \sum_{i=1}^{lM+L}{t_{i}}= LM-Ly_{1}\leq (lM+L)(m-1)L=(lM+L)(n-lM-L), \] \[ \sum_{i=1}^{lM+2L}{t_{i}}= LM-L(y_{1}+y_{2})\leq (lM+L)(m-2)L=(lM+2L)(n-lM-2L), \] \[ \ldots \ \ \ \ \ldots \ \ \ \ \ldots \ \ \ \ \ldots \ \ \ \ \ldots \] and \[ \sum_{i=1}^{n}{t_{i}}=M\sum_{i=1}^{l}{x_{i}}-L\sum_{i=1}^{m}{y_{i}}=0=n(n-lM-mL). \] So inequality (\ref{charac}) holds for $j=M,2M,\ldots,lM,lM+L,lM+2L,\ldots,lM+mL(=n)$ with equality when $j=n$. Now suppose that for some other value $j=j_{0}$ we have $\sum_{i=1}^{j_{0}}{t_{i}} > j_{0}(n-j_{0})$ and $j_{0}$ is the smallest such integer. But then $t_{j_{0}}>n-2j_{0}+1$ and as $j_{0}\neq M,2M,\ldots,lM,lM+L,lM+2L,\ldots,n$, we have $t_{j_{0}+1}= t_{j_{0}}>n-2j_{0}+1>n-2j_{0}-1=n-2(j_{0}+1)+1$. Thus $\sum_{i=1}^{j_{0}+1}{t_{i}} > (j_{0}+1)(n-j_{0}-1)$, showing that $j_{0}+1\neq M,2M,\ldots,lM,lM+L,lM+2L,\ldots,n$. Continuing in this way leads to a contradiction as we must reach one of $M,2M,\ldots,lM,lM+L,lM+2L,\ldots,n$ in finitely many steps. \end{proof} Together Theorems \ref{necessary} and \ref{odd} immediately give the following necessary and sufficient conditions for odd tournament imbalance sets. \begin{theorem}\label{oddsuff} A finite nonempty set of odd integers is the imbalance set of a tournament if and only if it contains at least one positive and at least one negative integer. \end{theorem} \section{The case of even imbalances} \label{even case} Mubayi, Will and West \cite{west1} considered simple digraphs with maximum number of arcs that realize imbalance sequences. \begin{lemma}\cite{west1}\label{aux2} Let $D$ be a simple digraph with maximum number of arcs realizing the imbalance sequence $[t_{i}]_{1}^{n}$. Then any vertex in $D$ has at most one non-neighbour and the number of arcs in $D$ equals $\sum_{i=1}^{n}\left\lfloor \frac{n-1+t_{i}}{2}\right\rfloor$. \end{lemma} A \textit{partial tournament} is a simple digraph obtained by removing one or more arcs from a tournament \cite{brualdi1}. We say that a partial tournament of order $n$ is a \emph{near tournament} if each vertex is joined to all the other vertices except exactly one. Clearly, every near tournament has even order. In this section, we characterize the sets of even integers that are imbalance sets of tournaments. Recall that $\{0\}$ is the imbalance set of every regular tournament. Therefore, in the remainder of this section we focus on nonzero sets of even integers. Example 1 shows that not every set of even integers that satisfies the necessary conditions of Theorem \ref{necessary} is the imbalance set of a tournament. We can nevertheless prove that any such set is the imbalance set of a near tournament. \begin{theorem}\label{even} Let $X=\{x_{1}, \ldots, x_{l}\}$ and $Y=\{-y_{1},\ldots,-y_{m}\}$ be disjoint nonempty sets of even integers, where $x_{1}>\cdots> x_{l}$ are non-negative even integers and $y_{1}<\cdots< y_{m}$ are positive even integers. Suppose that $X\cup Y\neq \{0\}$. Let $L=\sum_{i=1}^{l}{x_{i}}$, $M=\sum_{i=1}^{m}{y_{i}}$ and $n=lM+mL$. Then there exists a near tournament of order $n$ with imbalance set $X\cup Y$. \end{theorem} \begin{proof} Since $L$ and $M$ are even, $n$ is even and so we cannot construct a tournament of order $n$ with imbalance set $X\cup Y$. By mirroring the proof of Case 1 of Theorem \ref{odd}, we can show that the $n$-term sequence \[ [t_{i}]_{1}^{n}={x_{1}}^{(M)},\ldots,{x_{l}}^{(M)},{-y_{1}}^{(L)},\ldots,{-y_{m}}^{(L)} \] \noindent is the imbalance sequence of a simple digraph. Let $D$ be a realization of $[t_{i}]_{1}^{n}$ with maximum number of arcs. Since all the imbalances $t_{i}$ are even while $n-1$ is odd, by Lemma \ref{aux2} the number of arcs in $D$ is \[ \sum_{i=1}^{n}\left\lfloor \frac{n-1+t_{i}}{2}\right\rfloor =\sum_{i=1}^{n}\frac{n-2+t_{i}}{2}=\frac{n(n-2)}{2}, \] \noindent which is $\frac{n}{2}$ less than the number of arcs of a tournament of order $n$. Therefore, Lemma \ref{aux2} implies that every vertex of $D$ has exactly one non-neighbour and $D$ must be a near tournament. \end{proof} The following result shows that under certain conditions we can transform $D$ into a tournament by adding a suitable number of vertices. \begin{theorem}\label{evensuff} Let $X$, $Y$, $l$, $m$, $L$, $M$ and $n$ be as defined in Theorem \ref{even}. \item (i) If $0\in X\cup Y$ then there exists a tournament of order $n+1$ with imbalance set $X\cup Y$. \item (ii) If there exists an $x_{p}\in X$ and (not necessarily distinct) $-y_{q},-y_{r}\in Y$ such that $x_{p}=y_{q}+y_{r}$ then there exists a tournament of order $n+3$ with imbalance set $X\cup Y$. \item (iii) If there exists a $-y_{p}\in Y$ and (not necessarily distinct) $x_{q},x_{r}\in X$ such that $y_{p}=x_{q}+x_{r}$ then there exists a tournament of order $n+3$ with imbalance set $X\cup Y$. \end{theorem} \begin{proof} Let $T$ be a near tournament realizing the imbalance sequence $[t_{i}]_{1}^{n}$ as defined in the proof of Theorem \ref{even}. For a vertex $v_{i}$ in $T$ let $v'_{i}$ denote the unique non-neighbour of $v_{i}$. Also let $(u,v)$ denote an arc directed from vertex $u$ to vertex $v$. In the three cases we can transform $T$ into a tournament as follows. \noindent \textit{(i)} Add a vertex $v$ to $T$ in such a way that for every pair of non-adjacent vertices $v_{i}$ and $v'_{i}$ we insert the arcs $(v_{i},v'_{i})$, $(v'_{i},v)$ and $(v,v_{i})$. Thus the imbalance of all the vertices of $T$ is preserved and the new vertex $v$ has imbalance 0. Since every vertex of $T$ has been linked with every other vertex of $T$ as well as the new vertex $v$, the resulting digraph is a tournament. \begin{figure}[htb] \centering \includegraphics[scale=1.3]{3-3-2-ink.pdf} \caption{Construction of tournament in case (ii). Figure (a) represents step 1-2 and figure (b) step 3.} \label{tournamentex} \end{figure} \noindent \textit{(ii)} Add three new vertices $u_{1}$, $u_{2}$ and $u_{3}$ to $T$ and insert arcs in the following manner: \begin{enumerate} \item Insert $(u_{1},u_{2})$, $(u_{2},u_{3})$ and $(u_{3},u_{1})$. \item Choose any $\frac{x_{p}}{2}=\frac{y_{q}+y_{r}}{2}$ pairs $\{v_{i},v'_{i}\}$ of non-neighbouring vertices in $T$ and insert $(v_{i},v'_{i})$. Out of these choose any $\frac{y_{q}}{2}$ pairs. For each of these pairs insert the arcs $(u_{1},v_{i})$, $(u_{1},v'_{i})$, $(v_{i},u_{2})$, $(v'_{i},u_{2})$, $(v'_{i},u_{3})$ and $(u_{3},v_{i})$. For the other $\frac{y_{r}}{2}$ pairs insert the arcs $(u_{1},v_{i})$, $(u_{1},v'_{i})$, $(v_{i},u_{3})$, $(v'_{i},u_{3})$, $(v'_{i},u_{2})$ and $(u_{2},v_{i})$. \item For the remaining $\frac{n-x_{p}}{2}$ pairs $\{v_{i},v'_{i}\}$ of non-neighbours, insert the arcs $(u_{1},v_{i})$, $(v'_{i},u_{1})$, $(v_{i},v'_{i})$, $(v_{i},u_{2})$, $(u_{2},v'_{i})$, $(u_{3},v_{i})$ and $(v'_{i},u_{3})$. \end{enumerate} Since every vertex is joined with every other vertex, the resulting digraph is a tournament. Furthermore, the imbalance of each vertex of $T$ is preserved, while the new vertices $u_{1}$, $u_{2}$ and $u_{3}$ have imbalances $x_{p}$, $-y_{q}$ and $-y_{r}$ respectively. \noindent \textit{(iii)} The proof is essentially the same as that of case (ii). \end{proof} Theorem \ref{evensuff} can be generalized and it is the generalized version that is of interest to us as it leads to the characterization of even imbalance sets. However, we stated and proved Theorem \ref{evensuff} to provide the reader a concrete perspective of what is happening in the more abstract setting of Theorem \ref{evensuffgen}. \begin{theorem}\label{evensuffgen} Let $X$, $Y$, $l$, $m$, $L$, $M$ and $n$ be as defined in Theorem \ref{even}. The set $X\cup Y$ is the imbalance set of a tournament if any one of the following conditions is satisfied: \item (i) $0\in X\cup Y$, \item(ii) there exist an odd number of (not necessarily distinct) $x_{p_1},\ldots,x_{p_{2r+1}}\in X$ and an even number of (not necessarily distinct) $-y_{q_1},\ldots,-y_{q_{2s}}\in Y$ such that $\sum_{j=1}^{2r+1}x_{p_j}=\sum_{j=1}^{2s}y_{q_j}$, \item(iii) there exist an odd number of (not necessarily distinct) $-y_{p_1},\ldots,-y_{p_{2r+1}}\in Y$ and an even number of (not necessarily distinct) $x_{q_1},\ldots,x_{q_{2s}}\in X$ such that $\sum_{j=1}^{2r+1}y_{p_j}=\sum_{j=1}^{2s}x_{q_j}$. \end{theorem} \begin{proof} Let $T$ be a near tournament realizing the imbalance sequence $[t_{i}]_{1}^{n}$ as defined in the proof of Theorem \ref{even}. \noindent \textit{(i)} The proof is exactly the same as part (i) of Theorem \ref{evensuff}. \noindent \textit{(ii)} Add $2r+2s+1$ new vertices labelled $u_{p_1},\ldots,u_{p_{2r+1}}, u_{q_1}, \ldots,u_{q_{2s}}$ to $T$. Note that in the construction that follows we will relabel them in different ways, such as $u_{1},\ldots,u_{2r+2s+1}$, for the sake of convenience. We insert arcs in $T$ using the following procedure. \begin{figure}[htb] \centering \includegraphics[scale=1.2]{3-4-1-ink.pdf} \caption{Figure (a) represents steps 1-4 and figure (b) step 5 of \textsc{Add Arcs}.} \label{tournamentex} \end{figure} \vspace{2mm} \textsc{Add Arcs}: \begin{enumerate} \item Insert arcs so that the newly added vertices induce a regular tournament of order $2r+2s+1$. \item Choose any $\frac{\sum_{j=1}^{2r+1}x_{p_j}}{2}=\frac{\sum_{j=1}^{2s}y_{q_j}}{2}$ pairs $\{v_{i},v'_{i}\}$ of non-neighbouring vertices in $T$ and order them arbitrarily. For each $i=1,\ldots,\frac{\sum_{j=1}^{2r+1}x_{p_j}}{2}$ insert the arc $(v_{i}, v'_{i})$. Therefore, the imbalances of $v_{i}$ and $v'_{i}$ change by $+1$ and $-1$ respectively, for all $i$. \item For each $j=1,\ldots,2r+1$ choose $i=\frac{\sum_{h=1}^{p_{_{j-1}}}x_h}{2}+1,\ldots,\frac{\sum_{h=1}^{p_{j}}x_h}{2}$ and insert $x_{p_{j}}$ arcs $(u_{p_j},v_{i})$ and $(u_{p_j},v'_{i})$. This gives $u_{p_{j}}$ the imbalance $x_{p_{j}}$. \item For each $j=1,\ldots,2s$ choose $i=\frac{\sum_{h=1}^{q_{_{j-1}}}y_h}{2}+1,\ldots,\frac{\sum_{h=1}^{q_{j}}y_h}{2}$ and insert the arcs $(v_{i},u_{q_j})$, $(v'_{i},u_{q_j})$. Thus the imbalance of $u_{q_{j}}$ is $-y_{q_{j}}$. But the imbalances of $v_{i}$ and $v'_{i}$ are still perturbed by $+1$ and $-1$ respectively, for $i=1,\ldots,\frac{\sum_{j=1}^{2r+1}x_{p_j}}{2}$. \item For every $i=1,\ldots,\frac{\sum_{j=1}^{2r+1}x_{p_j}}{2}$ list the $u$'s that are not already linked with $v_{i}$ and $v'_{i}$. There are exactly $2r+2s-1$ such $u$'s, for each $i$. Label them arbitrarily from $1,\ldots,2r+2s-1$. For $j=1,\ldots,\frac{2r+2s-2}{2}$ insert the arcs $(u_{j},v_{i})$ and $(v'_{i},u_{j})$. For $j=\frac{2r+2s-2}{2}+1,\ldots,2r+2s-2$ insert the arcs $(u_{j},v_{i})$ and $(v'_{i},u_{j})$. Finally, insert the arcs $(u_{2r+2s-1},v_{i})$ and $(v'_{i}, u_{2r+2s-1})$. This preserves all the imbalances. \item For the remaining $\frac{n-\sum_{j=1}^{2r+1}x_{p_j}}{2}$ pairs $\{v_{i}, v'_{i}\}$ of non-neighbours insert the arc $(v_{i},v'_{i})$. Label all the $u$'s arbitrarily from $1,\ldots,2r+2s+1$. For $j=1,\ldots,\frac{2r+2s}{2}$ insert the arcs $(u_{j},v_{i})$ and $(v'_{i},u_{j})$. For $j=\frac{2r+2s}{2}+1,\ldots,2r+2s$ insert the arcs $(u_{j},v_{i})$ and $(v'_{i},u_{j})$. Finally, insert the arcs $(u_{2r+2s+1},v_{i})$ and $(v'_{i}, u_{2r+2s+1})$. This preserves all the imbalances. \begin{figure}[htb] \centering \includegraphics[scale=1.5]{3-4-2-ink.pdf} \caption{Inserting the remaining arcs in step 6 of \textsc{Add Arcs}.} \label{tournamentex} \end{figure} \end{enumerate} \vspace{2mm} Since every vertex is joined with every other vertex, the resulting digraph is a tournament. Furthermore, the imbalance of each vertex of $T$ is preserved, while the new vertices $u_{p_1},\ldots,u_{p_{2r+1}}, u_{q_1},\ldots,u_{q_{2s}}$ have imbalances $x_{p_1},\ldots,x_{p_{2r+1}}$, $-y_{q_1},\ldots,-y_{q_{2s}}$ respectively. \noindent \textit{(iii)} The proof is essentially the same as that of case (ii). \end{proof} The reader can easily draw parallels between the proofs of Theorems \ref{evensuff}(ii) and \ref{evensuffgen}(ii). For instance, the first and the last steps of both proofs are essentially achieveing the same target while steps 2-5 of the later are similar to but more complicated than step 2 of the former. Analyzing the above proof leads to a couple of simpler sufficient conditions for tournament imbalance sets. The first is one is a fairly straightforward consequence of Theorem \ref{evensuff} (i). \begin{corr}\label{zero} If $Z$ is the empty set or it contains at least one positive and at least one negative even integer then $Z\cup \{0\}$ is the imbalance set of a tournament. \end{corr} The second condition is not as obvious and is more of an arithmetic result than a combinatorial one. First, we note that for any positive integer $p\geq 1$, the set $\{2^{p}, -2^{p}\}$ is not a tournament imbalance set as any zero sum sequence formed by the elements of this set necessarily consists of an even number of elements. However, the following sufficient condition shows that any other set of positive and negative even integers containing a power of $2$ is a tournament imbalance set. \begin{corr}\label{arithmetic} Let $Z$ be a finite nonempty set of even integers containing at least one positive and at least one negative integer. Suppose $Z$ contains an element of the form $2^{p}$ or $-2^{p}$, for some positive integer $p\geq 1$, and $Z\neq \{2^{p}, -2^{p}\}$. Then $Z$ is the imbalance set of a tournament. \end{corr} \begin{proof} Let us assume that for some positive integer $p\geq 1$, $2^{p}$ is an element of $Z$. Choose any negative element $-y\in Z$. Then $y=r2^{q}$, where $q\geq 1$ is a positive integer and $r\geq 1$ is an odd positive integer such that if $r=1$ then $q\neq p$. (If this is not possible, we can start with $-2^{p}\in Z$ and choose an $x=r2^{q}$ from $Z$ with $q\neq p$.) Without loss of generality, let $p=\max\{p,q\}$. We have \[\underbrace{2^{p}+\cdots+2^{p}}_{r \textnormal{ terms}}=\underbrace{y+\cdots+y}_{2^{p-q} \textnormal{ terms}}, \] \noindent and by Theorem \ref{evensuffgen}, $Z$ is a tournament imbalance set. \end{proof} After deriving a number of sufficient conditions for tournament imbalance sets of even integers, the natural question is whether the sufficient conditions given in Theorem \ref{evensuffgen} are also necessary. The answer is positive as seen from the following result. \begin{theorem}\label{evennecess} Let $Z=X\cup Y$ be a finite nonempty set of even integers, where $X$ is the set of non-negative integers and $Y$ is the set of negative integers in $Z$. Then $Z$ is the imbalance set of a tournament if and only if either $Z=\{0\}$ or both $X$ and $Y$ are nonempty and satisfy one of the conditions (i), (ii) or (iii) of Theorem \ref{evensuffgen}. \end{theorem} \begin{proof} The sufficiency follows from Theorem \ref{evensuffgen}. To prove the necessity, suppose that $0\notin X\cup Y$ and let $X\cup Y$ be the imbalance set of a tournament of order $k$. This implies that we can form a sequence $[t_{i}]_{1}^{k}$ consisting of an odd number of not necessarily distinct terms from the elements of $X\cup Y$ that sums to zero. Since $k$ is odd, therefore either the number of terms from $X$ is odd or the number of terms from $Y$ is odd, but not both. Thus we have an odd (respectively even) number of terms $x\in X$ and an even (respectively odd) number of terms $-y\in Y$ such that $\sum{x}=\sum{y}$. \end{proof} \section{Algorithmic aspects} \label{algorithm} The aim of this section is to study the \textit{tournament imbalance set problem} (TIS) and present an algorithm that generates a tournament realizing any tournament imbalance set. We begin by proving a theorem on the lengths of equal sum sequences chosen from two set of non-negative integers that will play a crucial role in developing the algorithm. \begin{theorem}\label{zerosumsequence} Let $X$, $Y$, $l$, $m$, $L$, $M$ and $n$ be as defined in Theorem \ref{even}. If $k=p+q$ is the minimum odd number such that there exists a $p$-term sequence from $X$ and a $q$-term sequence from $-Y=\{y:-y\in Y\}$ having the same sum, then $k<n$. \end{theorem} \begin{proof} We observe that $k$ equals the minimal length of a zero sum sequence from $X\cup Y$. We prove the result by induction on $n$. Note that according to the conditions of Theorem \ref{even}, $X\cup Y\neq \{0\}$ and so the minimum possible value of $n$ is 4 that only corresponds to the sets $X=\{2\}$ and $Y=\{-2\}$. Since it is not possible to form a zero sum sequence with odd number of terms from $\{2, -2\}$. The next smallest value of $n$ is $6$ that corresponds to the sets $\{2,0,-2\}$, $\{2,-4\}$ and $\{4,-2\}$. Each of these sets admits a zero sum sequence of length $k=3$ and so $k<n$. Now we aim to show that the result holds for any $n>6$ by assuming that it holds for all values less than $n$. Let $X$ and $Y$ be any two sets of integers corresponding to $n$ and let $k\geq 3$ be the minimum odd number such that there exists a $k$-term zero sum sequence $a_{1},\ldots, a_{p},-b_{1},\ldots,-b_{q}$, where $k=p+q$, $a_{1}>\cdots>a_{p}\in X$ and $-b_{1}>\cdots> -b_{q}\in Y$. Assume that $k>n$, then $k\geq 5$. The sequence $a_{1}+a_{2},\ldots,a_{p},-b_{1}\ldots,-b_{q-1}-b_{q}$ is a zero sum sequence of minimal odd length $k'=k-2$ corresponding to the sets $X'=X-\{a_{1},a_{2}\}\cup \{a_{1}+a_{2}\}$ and $Y'=Y-\{-b_{q-1},-b_{q}\}\cup\{-b_{q-1}-b_{q}\}$. For the sets $X'$ and $Y'$ we have $n'<n-2$ and so $k'>n'$, contradicting the induction hypothesis. \end{proof} Given a set of non-negative integers, we call the search problem of finding two disjoint nonempty subsets that have identical sums the \textit{equal sum subsets problem} (ESS). Several authors \cite{bazgan1,woeginger1} have considered the corresponding decision and optimization problems. Additionally, many variants of ESS have been studied in literature. For instance, if we require subsets to be found from two different sets of positive integers the problem is called \textit{equal sum subsets from two sets} (ESST). It is known that ESS and ESST are weakly NP-hard as they admit pseudo-polynomial time algorithms \cite{cieliebak1,woeginger1}. The best known algorithm for the ESS is the dynamic programming procedure by Bazgan, Santha and Tuza \cite{bazgan1} that runs in $O(\left|I\right|\times{Sum}^{2})$ time, and determines all possible solutions of an ESS instance. Here $\left|I\right|$ and $Sum$ respectively denote the number of elements and the sum of elements of the input set. This procedure can be easily adapted to solve ESST \cite{cieliebak1}. Here we are interested in the following variation of ESS. \begin{definition}[\textbf{Equal Sum Sequences Problem (ESSeq)}]\label{ESSeq} Given two sets $X$ and $Y$ of non-negative integers and a positive integer $k$, find two nonempty finite sequences $[x]$ and $[y]$ consisting of elements from $X$ and $Y$ respectively, with each element allowed to repeat at the most $k$ times, such that $\sum x = \sum y$. \end{definition} We now study the complexity of ESSeq. Clearly, ESSeq is in class NP. Furthermore, ESST corresponds to the special case $k=1$ of ESSeq. Thus ESSeq is NP-hard. In fact, ESSeq is weakly NP-hard as we can solve any instance $ESSeq(X,Y,k)$ by using the multisets $X^{(k)}$ and $Y^{(k)}$, in which each element is repeated $k$ times, as input for the pseudo-polynomial ESST algorithm. Let us call the resulting algorithm, which finds all possible solutions to an ESSeq instance, \textsc{Equal Seq}. We have shown the following. \begin{theorem}\label{esseq np} The ESSeq decision (search) problem is weakly NP-complete (weakly NP-hard). \end{theorem} On the other hand, any algorithm that solves the even case of TIS must be able to check the existence of equal sum sequences $[x]_{1}^{a}$ and $[y]_{1}^{b}$ from any given nonempty finite sets $X$ and $Y$ of even integers such that $a$ and $b$ have different parity. Let $hX$ denote the set obtained by multiplying every element of a set $X$ by $h$ and $X+\{h\}$ denote the set obtained by adding a number $h$ to every element of $X$. We can solve an instance $ESSeq(X,Y,k)$ of the ESSeq decision problem by using any TIS algorithm to solve $\left|X\right|+1$ even instances $TIS(2X,2Y)$, $TIS(2(X+\{x\}),2Y)_{x\in X}$ of TIS. \begin{theorem}\label{np} The odd case of TIS can be solved in linear time. On the other hand, the even case of TIS is NP-complete. Hence in general TIS is NP-complete. \end{theorem} We now present a pseudo-polynomial time algorithm that not only solves TIS but also generates a tournaments realizing any tournament imbalance set. Our algorithm is based on the proofs of Theorems \ref{odd}, \ref{evensuffgen} and \ref{zerosumsequence}. First we form a suitable $n$-term imbalance sequence and then realize it as a tournament. Lemma \ref{aux2} can be used to construct a simple digraph, with maximum possible number of arcs, realizing an imbalance sequence $[t_{i}]_{1}^{n}$. The idea is to start with an arbitrary vertex $v$ having imbalance $t_{i}$ and attach it to $\left\lfloor \frac{n-1+t_{i}}{2}\right\rfloor$ vertices by arcs directed away from $v$. If $t_{i}$ has the same parity as $n-1$ then it is joined with $n-1-\left\lfloor \frac{n-1+t_{i}}{2}\right\rfloor$ other vertices by arcs directed towards $v$. Otherwise, it is joined with $n-2-\left\lfloor \frac{n-1+t_{i}}{2}\right\rfloor$ other vertices by arcs directed towards $v$. Thus $v$ is joined to every vertex except possibly one. These steps are then repeated for every vertex without attaching any new arcs to the preprocessed vertices. We name this $O(n^{2})$ procedure \textsc{Max Realization}. Now suppose that $Z$ is a finite nonempty set of integers arranged in decreasing order. Form the sets $X=\{z\in Z:z\geq 0\}=\{x_{1}, \ldots, x_{l}\}$ and $Y=\{z\in Z:z<0\}=\{-y_{1}, \ldots, -y_{m}\}$ arranged in decreasing order. Let $L=\sum_{i=1}^{l}{x_{i}}$, $M=\sum_{i=1}^{m}{y_{i}}$ and $n=lM+mL$ as in the earlier proofs. The following algorithm outputs a tournament that realizes $Z$, whenever such a tournament exists. \begin{algo}[\textsc{Imbalance Set}]\label{generate} \begin{enumerate} \item If either $X$ or $Y$ is empty, then $Z$ is not a tournament imbalance set. Stop. \item If elements of $Z$ have different parity, $Z$ is not a tournament imbalance set. Stop. \item Form the sequence $[t_{i}]_{1}^{n}={x_{1}}^{(M)},\ldots,{x_{l}}^{(M)},{-y_{1}}^{(L)},\ldots,{-y_{m}}^{(L)}.$ \item Call the procedure \textsc{Max Realization} to realize $[t_{i}]_{1}^{n}$ as a simple digraph $D$ with maximum number of arcs. \item If elements of $Z$ have odd parity, output $D$. End. \item If elements of $Z$ have even parity, call \textsc{Equal Seq} with the input $(X^{(n)}, (-Y)^{(n)}, n)$ to find sequences $[x]_{1}^{a}$ and $[y]_{1}^{b}$, with $a$ and $b$ having different parity and $\sum x=\sum y$. If no such sequences exist then $Z$ is not a tournament imbalance set. End. \item Add $2a+2b+1$ isolated vertices to $D$. \item Call \textsc{Add Arcs} to add $a+b$ vertices and arcs to $D$ to form a tournament $T$. Return $T$. \end{enumerate} \end{algo} The following result shows that the procedure \textsc{Imbalance Set} runs in pseudo-polynomial time and hence TIS is weakly NP-hard. \begin{theorem}\label{correct} Algorithm \ref{generate} is correct and runs in pseudo-polynomial time. \end{theorem} \begin{proof} The correctness follows immediately from Theorems \ref{odd}, \ref{even}, \ref{evensuffgen} and \ref{zerosumsequence}. In particular, Theorem \ref{zerosumsequence} guarantees that in the case when $Z$ is a set of even integers, step 6 of Algorithm \ref{generate} necessarily finds the required sequences if they exist. Now note that the computational complexity of Algorithm \ref{generate} is dominated by steps 4 and 6. Step 4 can be performed in $O(n^{2})=O((lM+mL)^{2})$ time, whereas step 6 takes $O(( n\left| X \right| + n\left| Y \right|)\times(n\sum_{x\in X}x + n\sum_{-y\in Y}y)^{2})=O(n^{3}(l+m)\times(L+M)^{2})=O(n^{3}\left|Z\right|\times (L+M)^{2})$ time. The overall complexity is therefore $O(\left|Z\right|\times n^{5})$, which is pseudo-polynomial since $n$ depends on the numeric value of the input. \end{proof} Thus we can use Algorithm \ref{generate} to check if a given set of integers is the imbalance set of a tournament and moreover, to construct a tournament realizing the set, if it exists. We now illustrate Algorithm \ref{generate} by showing how it generates a tournament realizing the imbalance set $\{4,2,-2\}$. \begin{figure}[htb] \centering \includegraphics[scale=0.7]{tournament1.pdf} \caption{A tournament realizing the imbalance set $\{4, 2, -2\}$ obtained from Algorithm 1.} \label{tournamentex} \end{figure} \begin{example} Consider the set $Z=\{4, 2, -2\}$. Since $Z$ satisfies the conditions in the first two steps of Algorithm 1, the algorithm goes to step 3 and forms the sequence $4, 4, 2, 2, -2, -2, -2, -2, -2, -2$. Step 4 calls the procedure \textsc{Max Realization} to output a simple digraph of order 10 realizing $Z$. However, this simple digraph is only a near tournament and not a tournament (see the digraph induced by the black vertices in Figure \ref{tournamentex}). Since the elements of $Z$ have even parity, the algorithm proceeds to step 6 and finds the the sequences $[x_{i}]_{1}^{1}=4$ and $[y_{i}]_{1}^{2}=-2,-2$ with odd and even number of terms respectively, such that $\sum{x_{i}} = -\sum{y_{i}}$. Then step 7 adds 3 new vertices (colored white in Figure \ref{tournamentex}) to the near tournament obtained in step 4. In the end, step 8 adds the arcs (dashed arcs in Figure \ref{tournamentex}) necessary to form a tournament of order 13 in such a way that the imbalances of the old vertices are preserved and the new vertices have imbalances $4$, $-2$, and $-2$. The output of Algorithm 1 is a tournament with imbalance sequence $4, 4, 4, 2, 2, -2, -2, -2, -2, -2, -2, -2, -2$ as shown in Figure \ref{tournamentex}. \end{example} The results presented in Sections 2, 3 and 4 can be used to estimate the order of a tournament realizing an imbalance set. Let us denote by $ord(Z)$ the minimal order of a tournament realizing an imbalance set $Z$. \begin{theorem}\label{extremal} Let $Z$ be a tournament imbalance set (i.e., it satisfies the conditions of Theorem 2.4 or Theorem 3.6). Define $X$, $Y$, $l$, $m$, $L$, $M$ and $n$ as in the earlier results. \item(i) If $Z$ consists of odd integers then $ord(Z)\leq n=lM+mL$. \item(ii) If $Z$ consists of even integers and $0\in Z$ then $ord(Z)\leq n+1=lM+mL+1$. \item(iii) If $Z$ consists of even integers and $0\notin Z$ then $ord(Z)< 2n=2(lM+mL)$. \end{theorem} \begin{proof} The proof of (i) and (ii) follows from Theorem \ref{odd}, while (iii) follows from Theorem \ref{evensuff}(i). For (iv), observe that the order of the tournament constructed in the proof of Theorem \ref{evensuffgen} is $n+2r+2s+1$. From Theorem \ref{zerosumsequence}, $2r+2s+1< n$. As a result, the constructed tournament can be at the most of order $2n-1$. \end{proof} \bigskip \noindent \textbf{Acknowledgements} \noindent The author would like to thank Professor K\'{a}roly Bezdek for having many useful discussions on the topic and helping in impproving this manuscript. The author is also thankful to Dr. Shariefuddin Pirzada for drawing his attention to the imbalance set problem. This research was performed at the Center for Computational and Discrete Geometry, Department of Mathematics and Statistics, University of Calgary.

https://arxiv.org/abs/2009.07768

Duality Mapping for Schatten Matrix Norms

In this paper, we fully characterize the duality mapping over the space of matrices that are equipped with Schatten norms. Our approach is based on the analysis of the saturation of the Hölder inequality for Schatten norms. We prove in our main result that, for $p\in (1,\infty)$, the duality mapping over the space of real-valued matrices with Schatten-$p$ norm is a continuous and single-valued function and provide an explicit form for its computation. For the special case $p = 1$, the mapping is set-valued; by adding a rank constraint, we show that it can be reduced to a Borel-measurable single-valued function for which we also provide a closed-form expression.

\section{Introduction} \label{Sec:intro} In linear algebra and matrix analysis, Schatten norms are a family of spectral matrix norms that are defined via the singular-value decomposition \cite{bhatia2013matrix}. They have appeared in many applications such as image reconstruction \cite{lefki2013Poisson,lefki2013HS}, image denoising \cite{xie2016weighted}, and tensor decomposition \cite{gao2020robust}, to name a few. Generally, the Schatten-$p$ norm of a matrix is the $\ell_p$ norm of its singular values. The family contains some well-known matrix norms: The Frobenius and the spectral (operator) norms are special cases in the family, with $p=2$ and $p=\infty$, respectively. The case $p=1$ (trace or nuclear norm) is of particular interest for applications as it can be used to recover low-rank matrices \cite{davenport2016overview}. This is the current paradigm in matrix completion, where the goal is to recover an unknown matrix given some of its entries \cite{candes2009exact}. Prominent examples of applications that can be reduced to low-rank matrix-recovery problems are phase retrieval \cite{candes2015phase}, sensor-array processing \cite{davies2012rank}, system identification \cite{fazel2013hankel}, and index coding \cite{asadi2017fast,esfahanizadeh2014matrix}. In addition to their many applications in data science, Schatten norms have been extensively studied from a theoretical point of view. Various inequalities concerning Schatten norms have been proven \cite{kittaneh1985inequalities,kittaneh1987inequalities2,kittaneh1986inequalities3,kittaneh1986inequalities4,kittaneh1987inequalities5,bourin2006matrix,Hirzallah2010Schatten,moslehian2011schatten,conde2016norm}; sharp bounds for commutators in Schatten spaces have been given \cite{wenzel2010impressions,cheng2015schatten}; moreover, facial structure \cite{so1990facial}, Fr\'echet differentiablity \cite{potapov2014frechet}, and various other aspects \cite{kittaneh1989continuity,bhatia2000cartesian} have been studied already. Our objective in this paper is to investigate the duality mapping in spaces of matrices that are equipped with Schatten norms. The duality mapping is a powerful tool to understand the topological structure of Banach spaces \cite{beurling1962theorem,cioranescu2012geometry}. It has been used to derive powerful characterizations of the solution of variational problems in function spaces \cite{de1976best,unser2019unifying} and also to determine generalized linear inverse operators \cite{liu2007best}. Here, we prove that the duality mapping over Schatten-$p$ spaces with $p\in (1,+\infty)$ is a single-valued and continuous function which, in fact, highlights the strict convexity of these spaces. For the special case $p=1$, the mapping is set-valued. However, we prove that, by adding a rank constraint, it reduces to a single-valued Borel-measurable function. In both cases, we also derive closed-form expressions that allow one to compute them explicitly. The paper is organized as follows: In Section \ref{Sec:Prelim}, we present relevant mathematical tools and concepts that are used in this paper. We study the duality mapping of Schatten spaces and propose our main result in Section \ref{Sec:DualMap}. We provide further discussions regarding the introduced mappings in Section \ref{Sec:discuss}. \section{Preliminaries} \label{Sec:Prelim} \subsection{Dual Norms, H\"older Inequality, and Duality Mapping} \label{Sec:Holder} Let $V$ be a finite-dimensional vector space that is equipped with an inner-product $\langle \cdot,\cdot \rangle:V \times V \rightarrow \mathbb{R}$ and let $\|\cdot\|_{X}:V\rightarrow \mathbb{R}_{\geq 0}$ be an arbitrary norm on $V$. We then denote by $X$ the space $V$ equipped with $\|\cdot\|_X$. Clearly, $X$ is a Banach space, because all finite-dimensional normed spaces are complete. The dual norm of $X$, denoted by $\|\cdot\|_{X'}: V\rightarrow \mathbb{R}_{\geq 0}$, is defined as \begin{equation}\label{Eq:DualNorm} \|{\bf v}\|_{X'} = \sup_{{\bf u} \in V\backslash \{\boldsymbol{0}\} } \frac{\langle {\bf v} , {\bf u} \rangle}{\|{\bf u}\|_{X}}, \end{equation} for any ${\bf v} \in V$. Following this definition, one would directly obtain the generic duality bound \begin{equation}\label{Eq:DualityBound} \langle {\bf v}, {\bf u} \rangle \leq \|{\bf v}\|_{X'} \|{\bf u}\|_{X}, \end{equation} for any ${\bf v}, {\bf u} \in V$. Saturation of Inequality \eqref{Eq:DualityBound} is the key concept of dual conjugates that is formulated in the following definition. \begin{definition}\label{Def:DualConj} Let $V$ be a finite-dimensional vector space and let $(\|\cdot\|_X,\|\cdot\|_{X'})$ be a pair of dual norms that are defined over $V$. The pair $({\bf u},{\bf v}) \in V\times V$ is said to be a $(X,X')$-conjugate, if \begin{itemize} \item $\langle {\bf v}, {\bf u}\rangle = \|{\bf v}\|_{X'} \|{\bf u}\|_{X}$, \item $\|{\bf v}\|_{X'} = \|{\bf u}\|_{X}$. \end{itemize} For any ${\bf u}\in V$, the set of all elements $ {\bf v} \in V$ such that $({\bf u},{\bf v})$ forms an $(X,X')$-conjugate is denoted by $\mathcal{J}_{X}({\bf u})\subseteq V$. We refer to the set-valued mapping $\mathcal{J}_{X}: V \rightarrow 2^V$ as the duality mapping. If, for all ${\bf u}\in V$, the set $\mathcal{J}_{X}({\bf u})$ is a singleton, then we indicate the duality mapping for the $X$-norm via the single-valued function ${\rm J}_X:V\rightarrow V$ with $\mathcal{J}_{X}({\bf u}) = \{{\rm J}_X({\bf u})\}$. \end{definition} It is worth mentioning that, for any ${\bf u}\in V$, the set $\mathcal{J}_{X}({\bf u})$ is nonempty . In fact, the closed ball $B= \{ {\bf v}\in V : \|{\bf v} \|_{X'}= \|{\bf u} \|_{X} \}$ is compact and, hence, the function ${\bf v}\mapsto \langle {\bf v} , {\bf u}\rangle$ attains its maximum value at some ${\bf v}^*\in B$. Now, following Definition \ref{Def:DualConj}, one readily verifies that $({\bf u},{\bf v}^*)$ is an $({X},{X}')$-conjugate. We conclude this part by providing a classical and illustrative example. Let $V= \mathbb{R}^n$ for some $n\in\mathbb{N}$. For any $p \in [1,+\infty]$, the $\ell_p$-norm of a vector ${\bf u} =(u_i)\in \mathbb{R}^n$ is defined as \begin{equation}\label{Eq:lp} \|{\bf u}\|_{p} = \begin{cases} \left(\sum_{i=1}^n |u_i|^p\right)^{\frac{1}{p}}, & p< +\infty \\ \max_{i} |u_i|, & p= +\infty. \end{cases} \end{equation} It is widely known that the dual norm of $\ell_p$ is the $\ell_q$-norm, where $(p,q)$ are H\"older conjugates ({\it i.e.}, $1/p+1/q=1$) \cite{rudin1991functional}. This stems from the H\"older inequality which states that \begin{equation}\label{Eq:HolderVec} \langle {\bf v}, {\bf u} \rangle \leq \| {\bf u}\|_{p} \|{\bf v}\|_q, \end{equation} for all ${\bf u}=(u_i),{\bf v}=(v_i) \in \mathbb{R}^n$. In the sequel, we exclude the trivial cases ${\bf u}=\boldsymbol{0}$ and ${\bf v}=\boldsymbol{0}$ to avoid unnecessary complexities in our statements. When $1<p<+\infty$, Inequality \eqref{Eq:HolderVec} is saturated if and only if $u_i v_i \geq 0$ for $i=1,\ldots,n$ and there exists a constant $c >0$ such that $|{\bf u}|^p = c |{\bf v}|^q$, where $|{\bf u}|^p = (|u_i|^p)$. This ensures that the duality mapping is single-valued and also yields the map \begin{equation}\label{Eq:DualMapLp} \mathrm{J}_p({\bf u})= {\rm sign}({\bf u}) \frac{|{\bf u}|^{p-1}}{\|{\bf u}\|_p^{p-2}}. \end{equation} For $p=1$, one can verify that the equality happens if and only if, for any index $i=1,\ldots,n$ with $u_i \neq 0$, one has that \begin{equation}\label{Eq:HolderEq1} v_i = {\rm sign} (u_i)\|{\bf v}\|_{\infty}. \end{equation} In other words, the vector ${\bf v}$ should attain its extreme values at places where ${\bf u}$ has nonzero values, with the sign being determined by the corresponding element in ${\bf u}$. Due to \eqref{Eq:HolderEq1}, the set $\mathcal{J}_1({\bf u})$ is not necessarily a singleton. However, if we add an additional sparsity constraint, then the mapping becomes single-valued. This leads us to introduce the new notion of {\it sparse duality mapping} in Definition \ref{Def:SparseDual}. \begin{definition}\label{Def:SparseDual} Let $V$ be a finite-dimensional vector space and let $s_0: V \rightarrow \mathbb{N}$ be an integer-valued function that acts as a sparsity measure. Assuming a pair $(\|\cdot\|_X,\|\cdot\|_{X'})$ of dual norms over $V$, we call the pair $({\bf u},{\bf v})\in V\times V$ a sparse $(X,X')$-conjugate if \begin{itemize} \item $({\bf u},{\bf v})$ forms an $(X,X')$-conjugate pair. In other words, ${\bf v} \in \mathcal{J}({\bf u})$. \item The quantity $s_0 ( {\bf v})$ attains its minimal value over the set $\mathcal{J}({\bf u})$. \end{itemize} We denote the set of sparse conjugates of ${\bf u}$ by $\mathcal{J}_{X,s_0}({\bf u})$. Whenever $\mathcal{J}_{X,s_0}({\bf u})$ is a singleton for any ${\bf u}\in V$, we refer to the single-valued function ${\rm J}_{X,s_0}: V \rightarrow V$ with $\mathcal{J}_{X,s_0}({\bf u})=\{{\rm J}_{X,s_0}({\bf u})\}$ as the sparse duality mapping. \end{definition} Following Definition \ref{Def:SparseDual}, if we use the $\ell_0$-norm as the sparsity measure, that is $s_0 ({\bf u}) = \|{\bf u}\|_0={\rm Card}\left(\{i: u_i\neq 0\}\right)$\footnote{Although this functional does not satisfy the homogeneity property of a norm, it has been widely referred to as the $\ell_0$-norm.}, then we have the sparse duality mapping \begin{align}\label{Eq:DualMapSparseL1} &{\rm J}_{1,0}:\mathbb{R}^n\rightarrow\mathbb{R}^n:{\bf u}=(u_i) \mapsto {\bf v}=(v_i) ={\rm J}_{1,0}({\bf u}), \nonumber \\ &v_i = \begin{cases} {\rm sign}(u_i) \|{\bf u}\|_1, & u_i\neq 0 \\ 0, & u_i=0.\end{cases} \end{align} Finally, we mention that, for $p= +\infty$, the reduced set $\mathcal{J}_{\infty,0}$ is not single-valued. Indeed, let us define $I_{\max}({\bf u})=\{i: |u_i|=\|{\bf u}\|_{\infty}\}\subseteq \{1,\ldots,n\}$. We readily deduce from \eqref{Eq:HolderEq1} that ${\bf v}=(v_1,\ldots,v_n) \in \mathcal{J}_{\infty}({\bf u})$ if and only if $v_i = 0$ whenever $i\not \in I_{\max}({\bf u})$ and ${\rm sign}(v_i) = {\rm sign}(u_i)$ for $i \in I_{\max}({\bf u})$ with $\sum_{i\in I_{\max}({\bf u})} | v_i|= \|{\bf u}\|_{\infty}$. This shows that $\mathcal{J}_{\infty}({\bf u})$ is a convex set with $\mathcal{J}_{\infty,0}({\bf u})$ being its extreme points, where $\mathcal{J}_{\infty,0}({\bf u})=\{u_i {\bf e}_i: i\in I_{\max}({\bf u})\}$. \subsection{Schatten $p$-Norm} It is widely known that any matrix ${\bf A} \in \mathbb{R}^{m\times n}$ can be decomposed as \begin{equation}\label{Eq:SVD} {\bf A}= {\bf U} {\bf S} {\bf V}^T, \end{equation} where ${\bf U}\in\mathbb{R}^{m\times m}$ and ${\bf V}\in\mathbb{R}^{n\times n}$ are orthogonal matrices and ${\bf S}$ is an $m$ by $n$ rectangular diagonal matrix with nonnegative real entries $\sigma_1 \geq \sigma_2 \geq \cdots \geq \sigma_{\min(m,n)} \geq 0$ sorted in descending order. In the literature, \eqref{Eq:SVD} is known as the singular-value decomposition (SVD) and the entries $\sigma_i$ are the singular values of ${\bf A}$. In general, the SVD of a matrix $\bf A$ is not unique. However, the diagonal matrix ${\bf S}$ and, consequently, its entries, are fully determined from ${\bf A}$. In other words, the values of $\sigma_i$ are invariant to a specific choice of decomposition. This is why one can refer to the diagonal entries of ${\bf S}$ as the ``singular values'' of ${\bf A}$. When ${\bf A}$ is not full rank, one can obtain a reduced version of \eqref{Eq:SVD}. Indeed, if we denote the rank of ${\bf A}$ by $r$, then we have that \begin{equation}\label{Eq:ReducedSVD} {\bf A} = {\bf U}_r {\bf S}_r {\bf V}_r^T, \end{equation} where $ {\bf U}_r \in \mathbb{R}^{m\times r}$ and $ {\bf V}_r \in \mathbb{R}^{n\times r}$ are (sub)-orthogonal matrices such that $ {\bf U}_r^T {\bf U}_r = {\bf V}_r^T {\bf V}_r = {\bf I}_r$ and ${\bf S}_r ={\rm diag}(\boldsymbol{\sigma})$ is a diagonal matrix that contains positive singular values $\boldsymbol{\sigma}=(\sigma_1,\ldots,\sigma_r)\in \mathbb{R}^{r}$ of ${\bf A}$. Finally, for any $p\in [1,+\infty]$, the Schatten-$p$ norm of ${\bf A}$ is defined as \begin{equation} \label{Eq:SchattenNorm} \|{\bf A}\|_{S_p} = \begin{cases} \left(\sum_{i=1}^r \sigma_i^p\right)^{\frac{1}{p}}, & p< +\infty \\ \sigma_1, & p= +\infty. \end{cases} \end{equation} \section{Duality Mapping in Schatten Spaces} \label{Sec:DualMap} The dual of the Schatten-$p$ norm is the Schatten-$q$ norm, where $q\in [1,\infty]$ is such that $\frac{1}{p}+\frac{1}{q}=1$ \cite{bhatia2013matrix}. This is due to the generalized version of H\"older's inequality for Schatten norms, as stated in Proposition \ref{Prop:Holder}. While this is a known result (see, for example, \cite{Lefki2015StructureTensor}), it is also the basis for the present work, which is the reason why we provide a proof in \ref{App:Holder}. \begin{proposition}\label{Prop:Holder} For any pair $(p,q)\in [1,+\infty]^2$ of H\"older conjugates with $\frac{1}{p}+\frac{1}{q}=1$ and any pair of matrices ${\bf A},{\bf B}\in \mathbb{R}^{m\times n}$, we have that \begin{equation}\label{Eq:Holder} \langle {\bf A}, {\bf B}\rangle = \mathrm{Tr}\left( {\bf A}^T {\bf B} \right) \leq \|{\bf A}\|_{S_p} \|{\bf B}\|_{S_q}. \end{equation} \end{proposition} In Proposition \ref{Prop:HolderSat}, we investigate the case where the H\"older inequality is saturated, in the sense that \begin{equation}\label{Eq:HolderSat} {\rm Tr}\left( {\bf A}^T {\bf B} \right) = \|{\bf A}\|_{S_p} \|{\bf B}\|_{S_q}. \end{equation} This saturation is central to our work, as it is tightly linked to the notion of duality mapping. \begin{proposition}\label{Prop:HolderSat} Let $(p,q)$ be a pair of H\"older conjugates and let ${\bf A},{\bf B}\in \mathbb{R}^{m\times n}$ be a pair of nonzero matrices with reduced SVDs of the form \begin{equation}\label{Eq:RedSVD} {\bf A}= {\bf U}_r {\rm diag}(\boldsymbol{\sigma}) {\bf V}_r^T, \quad {\bf B}= \tilde{\bf U}_{\tilde{r}} {\rm diag}(\boldsymbol{\tilde{\sigma}})\tilde{\bf V}_{\tilde{r}}^T. \end{equation} \begin{itemize} \item If $p\in (1,\infty)$, then the H\"older inequality is saturated if and only if we have that \begin{equation}\label{Eq:FormBgenP} {\bf B} = c {\bf U}_r{\rm diag}({\rm J}_p(\boldsymbol{\sigma})){\bf V}_r^T \end{equation} or, equivalently, \begin{equation}\label{Eq:FormAgenP} {\bf A} = c^{-1} {\bf \tilde{U}}_{\tilde{r}}{\rm diag}({\rm J}_q(\tilde{\boldsymbol{\sigma}})){\bf \tilde{V}}_{\tilde{r}}^T, \end{equation} where $c= \frac{\|{\bf B}\|_{S_q}}{\|{\bf A}\|_{S_p}}$ and ${\rm J}_p(\cdot)$ and ${\rm J}_q(\cdot)$ are the duality mappings for the $\ell_p$ and $\ell_q$ norms, respectively (see \eqref{Eq:DualMapLp}). \item If $p=1$, then a necessary condition for the saturation of the H\"older inequality is that \begin{equation} {\rm rank}({\bf A}) \leq r_1 \leq {\rm rank}({\bf B}), \end{equation} where $r_1= {\rm Card}\left(\{i: \tilde{\sigma}_i = \tilde{\sigma}_1\}\right)$ is the multiplicity of the first singular value of ${\bf B}$. Moreover, if we denote the first $r_1$ singular vectors of ${\bf B}$ in \eqref{Eq:RedSVD} by ${\bf \tilde{U}}_1\in \mathbb{R}^{m\times r_1}$ and ${\bf \tilde{V}}_1\in\mathbb{R}^{n\times r_1}$, then the saturation of the H\"older inequality is equivalent to the existence of a symmetric matrix ${\bf X}\in \mathbb{R}^{r_1\times r_1}$ such that \begin{align}\label{Eq:FormA1} {\bf A} = {\bf \tilde{U}}_1 {\bf X} {\bf \tilde{V}}_1^T. \end{align} Finally in the rank-equality case ${\rm rank}({\bf A}) ={\rm rank}({\bf B})$, we have saturation if and only if \begin{equation}\label{Eq:FormBrankEq} {\bf B} = c {\bf U}_r{\bf V}_r^T, \end{equation} where $c= \|{\bf B}\|_{S_{\infty}}$ and the matrices ${\bf U}_r$ and ${\bf V}_r$ are defined in \eqref{Eq:RedSVD}. \end{itemize} \end{proposition} \begin{remark}\label{Remark} The reduced SVD is not unique; there are multiple choices for the sub-orthogonal matrices in \eqref{Eq:RedSVD}. However, the parametric forms given in Proposition \ref{Prop:HolderSat} do not depend on a specific decomposition. \end{remark} The proof of Proposition \ref{Prop:HolderSat} can be found in \ref{App:HolderSat}. We observe that, in the case $p\in(1,\infty)$, the saturation of H\"older inequality provides a very tight link between the two matrices: If we know one of them, then the other lies in a one-dimensional ray that is parameterized by the constant $c>0$. However, in the special case $p=1$, the identification is not as simple. There again, for a given matrix ${\bf B}$, one can fully characterize the set of admissible matrices ${\bf A}$. However, for the reverse direction, an additional rank-equality constraint is essential to reduce the set of admissible matrices ${\bf B}$ to just one ray. Inspired from Proposition \ref{Prop:HolderSat}, we now propose our main result in Theorem \ref{Thm:p}, where we explicitly characterize the duality mapping for the Schatten $p$-norms. The proof of Theorem \ref{Thm:p} can be found in \ref{App:p}. \begin{theorem}\label{Thm:p} Let $p,q\in [1,+\infty]$ be a pair of H\"older conjugates with $\frac{1}{p}+\frac{1}{q}=1$ and ${\bf A}\in\mathbb{R}^{m\times n}$ a matrix whose reduced SVD is specified in \eqref{Eq:ReducedSVD}. \begin{itemize} \item If $1<p<+\infty$, then the single-valued duality mapping ${\rm J}_{S_p}: \mathbb{R}^{m\times n} \rightarrow \mathbb{R}^{m\times n}$ is well-defined and can be expressed as \begin{equation} {\rm J}_{S_p}:{\bf A} = {\bf U }_r {\rm diag}(\boldsymbol{\sigma} ) {\bf V}_r^T \mapsto {\bf A}^*= {\bf U}_r{\rm diag}({\rm J}_p(\boldsymbol{\sigma} )){\bf V}_r^T. \end{equation} \item If $p=1$ and if we consider the rank function as the sparsity measure in Definition \ref{Def:SparseDual}, then the sparse duality mapping ${\rm J}_{S_1,{\rm rank}}: \mathbb{R}^{m\times n} \rightarrow \mathbb{R}^{m\times n}$ is well-defined (singleton) and is given as \begin{equation} {\rm J}_{S_1,{\rm rank}}:{\bf A} = {\bf U }_r {\rm diag}(\boldsymbol{\sigma}) {\bf V}_r^T \mapsto {\bf A}^*= \|\boldsymbol{\sigma}\|_{1} {\bf U} _r{\bf V}_r^T. \end{equation} \item If $p=+\infty$, then the set-valued mapping $\mathcal{J}_{S_\infty}(\cdot)$ can be described as \begin{equation}\label{Eq:Jinf} \mathcal{J}_{S_\infty}({\bf A})= \left\{ \sigma_1 {\bf U}_1 {\bf X} {\bf V}_1^T: {\bf X}\in\mathbb{R}^{r_1\times r_1} \text{ is symmetric and } \|{\bf X}\|_{S_1} =1 \right\}, \end{equation} where $r_1$ denotes the multiplicity of the first singular value $\sigma_1$ of ${\bf A}$ and ${\bf U}_1,{\bf V}_1$ are singular vectors that correspond to $\sigma_1$ in \eqref{Eq:ReducedSVD}. It is a convex set whose extreme points are ${\bf E}_{i,j} = \frac{\sigma_1}{2}({\bf u}_i {\bf v}_j^T +{\bf v}_i {\bf u}_j^T )$ for $1\leq i\leq j\leq r_1$. Finally, the set of sparse dual conjugates is the collection of rank-1 elements of $\mathcal{J}_{S_\infty}({\bf A})$ which can be characterized as \begin{equation} \mathcal{J}_{S_\infty, {\rm rank}}({\bf A}) = \{ \sigma_1 {\bf U}_1 {\bf p}{\bf p}^T {\bf V}_1^T: {\bf p}\in \mathbb{R}^{r_1} , \|{\bf p}\|_2=1\}. \end{equation} \end{itemize} \end{theorem} \section{Discussion}\label{Sec:discuss} Theorem \ref{Thm:p} provides an interesting characterization of the duality mapping in three scenarios: The first case is $1<p<+\infty$ which is the most straightforward one. Theorem \ref{Thm:p} tells us that the mapping is single-valued and also gives a formula to compute the dual conjugate ${\bf A}^*$ of any matrix ${\bf A}\in\mathbb{R}^{m\times n}$. We use this result to deduce the continuity of the duality mapping as well as the strict convexity of the Schatten space in this case (see Corollary \ref{Cor:strict}). In the second case, with $p=1$, the mapping is not single-valued. However, there is a unique element in the set of dual conjugates with the minimal rank (that is equal to the rank of ${\bf A}$) and, hence, we can construct a single-valued sparse duality mapping. Finally, we showed in the third case, characterized by $p=+\infty$, that neither the set of dual conjugates nor the ones with the minimal rank are unique. However, we observe in \eqref{Eq:Jinf} that the entries of ${\bf X}$ can be independently chosen, up to symmetry and normalization assumptions. This suggests that the dimension of $\mathcal{J}_{S_\infty}({\bf A})$ is $d=\left( \frac{r_1(r_1+1)}{2}-1\right)$. Moreover, we show that this convex set has exactly $(d+1)$ extreme points, which is the minimal number for a convex set of dimension $d$. We also observe that the extreme points of $\mathcal{J}_{S_\infty}({\bf A})$ are low-rank. They are indeed a collection of rank-1 and rank-2 matrices. In Corollary \ref{Cor:strict}, we highlight some consequences of Theorem \ref{Thm:p} concerning the strict convexity of Schatten spaces and the continuity of the duality mapping. \begin{corollary}\label{Cor:strict} The Banach space of $m$ by $n$ matrices equipped with the Schatten-$p$ norm is strictly convex, if and only if $p\in (1,+\infty)$. In this case, the function ${\rm J}_{S_p} : \mathbb{R}^{m\times n} \rightarrow \mathbb{R}^{m\times n}$ is continuous. \end{corollary} \begin{proof} For $p\in (1,+\infty)$, we know from Theorem \ref{Thm:p} that the duality mapping ${\rm J}_{S_p}$ is bijective. Moreover, it is known that all finite-dimensional Banach spaces are reflexive. Now, following \cite{petryshyn1970characterization}, we deduce the strict convexity of the space of $m$ by $n$ matrices with Schatten-$p$ norm. For $p=1$ and $p=+\infty$, we can readily verify that \begin{align*} &\left \|\alpha \begin{pmatrix}1 & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & 0 \end{pmatrix} +(1-\alpha)\begin{pmatrix}0 & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & 1 \end{pmatrix} \right\|_{S_1} = \left \|\begin{pmatrix} \alpha & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & (1-\alpha) \end{pmatrix} \right\|_{S_1}=1, \\ &\left\|\alpha \begin{pmatrix}1 & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & 0 \end{pmatrix} +(1-\alpha)\begin{pmatrix}1 & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & 1 \end{pmatrix} \right\|_{S_\infty} = \left\| \begin{pmatrix}1 & \cdots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \cdots & (1-\alpha) \end{pmatrix} \right\|_{S_\infty}=1, \end{align*} for all $\alpha \in (0,1)$, which shows that the Schatten space is not strictly convex for $p=1,+\infty$. Finally, the Schatten-$p$ norm is known to be Fr\'echet differentiable for $p\in(1,+\infty)$ \cite{potapov2014frechet}. Moreover, the duality mapping of any Banach space with Fr\'echet-differentiable norms is guaranteed to be continuous \cite{giles1978geometrical,contreras1994upper}. Combining the two statements, we deduce the continuity of the duality mapping in this case. \end{proof} By contrast, the sparse duality mapping ${\rm J}_{S_1,{\rm rank}}(\cdot)$ is not continuous. This is best explained by providing a counterexample. Specifically, let us consider the sequence of 2 by 2 matrices $${\bf S}_k=\begin{pmatrix} 1 & 0\\ 0 & \frac{1}{k}\end{pmatrix}, \quad k\in\mathbb{N}.$$ It is clear that ${\bf S}_k \rightarrow {\bf S}_{\infty} =\begin{pmatrix} 1 & 0\\ 0 & 0\end{pmatrix}$. However, we have that $$\forall k\in\mathbb{N}: {\rm J}_{S_1,{\rm rank}}({\bf S}_k) = \begin{pmatrix} 1 & 0\\ 0 & 1\end{pmatrix}, \quad \text{while} \quad {\rm J}_{S_1,{\rm rank}}({\bf S}_{\infty}) = \begin{pmatrix} 1 & 0\\ 0 & 0\end{pmatrix},$$ which shows the discontinuity of ${\rm J}_{S_1,{\rm rank}}$ in the space of 2 by 2 matrices. This can be generalized to space of matrices with arbitrary dimensions $m,n\in\mathbb{N}$. Although ${\rm J}_{S_1,{\rm rank}}$ is not continuous, we now show that it is Borel-measurable and, hence, that it can be approximated with arbitrary precision by a continuous mapping due to Lusin's theorem \cite{rudin1991functional}. \begin{proposition}\label{Prop:measure} For any $m,n\in\mathbb{N}$, the sparse duality mapping ${\rm J}_{S_1,{\rm rank}}$ is a Borel-measurable matrix-valued function over the space of $m$ by $n$ matrices. \end{proposition} Before going into the proof of Proposition \ref{Prop:measure}, we present a preliminary result. \begin{lemma}\label{Lem} The set $\mathcal{R}_{ r}\subseteq \mathbb{R}^{m\times n}$ of $m$ by $n$ matrices of rank $r$ is Borel-measurable. \end{lemma} \begin{proof} First note that $$\mathcal{R}_{1}=\{ {\bf u}{\bf v}^T: {\bf u}\in\mathbb{R}^m, {\bf v}\in \mathbb{R}^n\}. $$ The set $\mathcal{R}_{1}$ is the image of the continuous mapping $\mathbb{R}^{m}\times \mathbb{R}^n \rightarrow \mathbb{R}^{m\times n}:({\bf u}, {\bf v}) \mapsto {\bf u}{\bf v}^T$ and, hence, is Borel-measurable. Now, denote by $\mathcal{R}_{\leq r}\subseteq \mathbb{R}^{m\times n}$, the set of matrices with rank no more than $r$. Using the identity $$\mathcal{R}_{\leq r} = \mathcal{R}_{1} + \cdots + \mathcal{R}_{1} ,\quad \text{(r times)}, $$ we deduce that $\mathcal{R}_{\leq r}$ and, consequently, $\mathcal{R}_{r} = \mathcal{R}_{\leq r} \backslash \mathcal{R}_{\leq (r-1)}$ are also Borel-measurable sets. \end{proof} \begin{proof}[Proof of Proposition \ref{Prop:measure}] Consider a Borel-measurable set $\mathcal{B}\subseteq \mathbb{R}^{m\times n}$. We show that $\mathcal{B}_{\rm inv}= {\rm J}_{S_1,{\rm rank}}^{-1}(\mathcal{B})$ is also Borel-measurable. By defining $\mathcal{B}_{{\rm inv},r} = \mathcal{B}_{\rm inv}\cap \mathcal{R}_{r}$, we can partition $ \mathcal{B}_{\rm inv}$ as $$\mathcal{B}_{\rm inv}= \bigcup_{r=1}^{\min (m,n)} \mathcal{B}_{\rm inv}\cap \mathcal{R}_{r}.$$ Hence, it is sufficient to show that each partition $\mathcal{B}_{\rm inv}\cap \mathcal{R}_{r}$ is Borel-measurable. Define the set $\mathcal{P}_{r} \subseteq \mathcal{R}_{r}^2$ as $$\mathcal{P}_{r} = \{ ({\bf A},{\bf B}) \in \mathcal{R}_{r} \times \mathcal{B}: {\rm Tr}({\bf A}^T{\bf B}) = \|{\bf A}\|_{S_1} \|{\bf B}\|_{S_\infty}, \quad \|{\bf A}\|_{S_1}= \|{\bf B}\|_{S_\infty}\}.$$ The set $\mathcal{P}_{r}$ introduces a relation over $ \mathcal{R}_{r} $ whose domain is $\mathcal{B}_{\rm inv}\cap \mathcal{R}_{r}$. Since the trace and norm are continuous (and, consequently, Borel-measurable) functions and $\mathcal{R}_{r}\times \mathcal{B}$ is a Borel-measurable set (using Lemma \ref{Lem}), we deduce that the relation induced from $\mathcal{P}_{r}$ is Borel-measurable as well. Finally, we use \cite[Proposition 2.1]{himmelberg1975measurable} to show that its domain is Borel-measurable. \end{proof} \section{Conclusion} In this paper, we studied the duality mapping in finite-dimensional Schatten spaces. Based on a careful investigation of the cases where the H\"older inequality saturates, we provided an explicit form for this mapping when $p\in (1,+\infty)$. Furthermore, by adding a rank constraint, we proved that the mapping becomes single-valued for the special case $p=1$. As for $p=+\infty$, we showed that the mapping yields a convex set whose extreme points are low-rank matrices. Finally, we discussed our theorem and studied the continuity of the introduced mappings as well as the strict convexity of the Schatten spaces. A possible future direction of research is to extend the results of this paper to infinite-dimensional Schatten spaces and even, in full generality, to linear operators over Hilbert spaces.

https://arxiv.org/abs/1311.5240

PENLAB: A MATLAB solver for nonlinear semidefinite optimization

PENLAB is an open source software package for nonlinear optimization, linear and nonlinear semidefinite optimization and any combination of these. It is written entirely in MATLAB. PENLAB is a young brother of our code PENNON \cite{pennon} and of a new implementation from NAG \cite{naglib}: it can solve the same classes of problems and uses the same algorithm. Unlike PENNON, PENLAB is open source and allows the user not only to solve problems but to modify various parts of the algorithm. As such, PENLAB is particularly suitable for teaching and research purposes and for testing new algorithmic ideas.In this article, after a brief presentation of the underlying algorithm, we focus on practical use of the solver, both for general problem classes and for specific practical problems.

\section{Introduction} Many problems in various scientific disciplines, as well as many industrial problems lead to (or can be advantageously formulated) as nonlinear optimization problems with semidefinite constraints. These problems were, until recently, considered numerically unsolvable, and researchers were looking for other formulations of their problem that often lead only to approximation (good or bad) of the true solution. This was our main motivation for the development of PENNON \cite{pennon}, a code for nonlinear optimization problems with matrix variables and matrix inequality constraints. Apart from PENNON, other concepts for the solution of nonlinear semidefinite programs are suggested in literature; see \cite{sun-sun-zhang} for a discussion on the classic augmented Lagrangian method applied to nonlinear semidefinite programs, \cite{correa2004global,fares,freund-jarre-vogelbusch} for sequential semidefinite programming algorithms and \cite{kanzow-nagel-newt} for a smoothing type algorithm. However, to our best knowledge, none of these algorithmic concepts lead to a publicly available code yet. In this article, we present PENLAB, a younger brother of PENNON and a new implementation from NAG. PENLAB can solve the same classes of problems, uses the same algorithm and its behaviour is very similar. However, its performance is relatively limited in comparison to \cite{pennon} and \cite{naglib}, due to MATLAB implementation. On the other hand, PENLAB is open source and allows the user not only to solve problems but to modify various parts of the algorithm. As such, PENLAB is particularly suitable for teaching and research purposes and for testing new algorithmic ideas. After a brief presentation of the underlying algorithm, we focus on practical use of the solver, both for general problem classes and for specific practical problems, namely, the nearest correlation matrix problem with constraints on condition number, the truss topology problem with global stability constraint and the static output feedback problem. More applications of nonlinear semidefinite programming problems can be found, for instance, in \cite{annad,kanno,leibfritz-volkwein}. PENLAB is distributed under GNU GPL license and can be downloaded from {\tt http://web.mat.bham.ac.uk/kocvara/penlab}. We use standard notation: Matrices are denoted by capital letters ($A,B,X,\ldots$) and their elements by the corresponding small-case letters ($a_{ij}, b_{ij}, x_{ij},\ldots$). For vectors $x,y\in\RR^n$, $\langle x,y\rangle:=\sum_{i=1}^n x_iy_i$ denotes the inner product. $\SS^{m}$ is the space of real symmetric matrices of dimension $m\times m$. The inner product on $\SS^{m}$ is defined by $\langle A, B\rangle_{\SS^{m}} := \Tr (AB)$. When the dimensions of $A$ and $B$ are known, we will often use notation $\langle A, B\rangle$, same as for the vector inner product. Notation $A\preccurlyeq B$ for $A,B\in\SS^{m}$ means that the matrix $B-A$ is positive semidefinite. If $A$ is an $m\times n$ matrix and $a_j$ its $j$-th column, then $\Vec A$ is the $mn\times 1$ vector $$ \Vec A = \begin{pmatrix} a_1^T\ \ a_2^T\ \ \cdots\ \ a_n^T\end{pmatrix}^T\,. $$ Finally, for $\Phi:\SS^m\to\SS^m$ and $X,Y\in \SS^m$, $D\Phi(X;Y)$ denotes the directional derivative of $\Phi$ with respect to $X$ in direction $Y$. \vfill \section{The problem} We intend to solve optimization problems with a nonlinear objective subject to nonlinear inequality and equality constraints and nonlinear matrix inequalities (NLP-SDP): \begin{align} & \min_{x\in \RR^n, Y_1\in\SS^{p_1},\ldots,Y_k\in\SS^{p_k}} f(x,Y)\label{eq:nlpsdp}\\ & \begin{aligned} \mbox{subject to}\quad &g_i(x,Y) \leq 0, \qquad &&i=1,\ldots,m_g\\ &h_i(x,Y) = 0, \qquad &&i=1,\ldots,m_h \\ &{ {\cal A}_i(x,Y)\preceq 0,} \qquad &&{ i=1,\ldots,m_A}\\ &\underline{\lambda}_i I \preceq Y_i\preceq \overline{\lambda}_i I, \qquad&&i=1,\ldots,k\,. \end{aligned}\nonumber \end{align} Here \begin{itemize} \item $x\in\RR^n$ is the vector variable; \item $Y_1\in\SS^{p_1},\ldots,Y_k\in\SS^{p_k}$ are the matrix variables, $k$ symmetric matrices of dimensions $p_1\times p_1,\ldots,p_k\times p_k$; \item we denote $Y=(Y_1,\ldots,Y_k)$; \item $f$, $g_i$ and $h_i$ are $C^2$ functions from $\RR^n\times \SS^{p_1}\times\ldots\times\SS^{p_k}$ to $\RR$; \item $\underline{\lambda}_i$ and $\overline{\lambda}_i$ are the lower and upper bounds, respectively, on the eigenvalues of~ $Y_i$, $i=1,\ldots,k$; \item ${\cal A}_i(x,Y)$ are twice continuously differentiable nonlinear matrix operators from $\RR^n\times \SS^{p_1}\times\ldots\times\SS^{p_k}$ to $\SS^{p_{A_i}}$ where ${p_{A_i}}$, $i=1,\ldots,m_A$, are positive integers. \end{itemize} \section{The algorithm} The basic algorithm used in this article is based on the nonlinear rescaling method of Roman~Polyak \cite{polyak} and was described in detail in \cite{pennon} and \cite{stingl}. Here we briefly recall it and stress points that will be needed in the rest of the paper. The algorithm is based on a choice of penalty/barrier functions $\varphi:\RR\to\RR$ that penalize the inequality constraints and $\Phi:\SS^p\to\SS^p$ penalizing the matrix inequalities. These functions satisfy a number of properties (see \cite{pennon,stingl}) that guarantee that for any $\pi>0$ and $\Pi > 0$, we have $$ z(x) \le 0 \ \Longleftrightarrow \ \pi\varphi(z(x)/\pi) \le 0, \quad z\in C^2(\RR^n\to\RR) $$ and $$ Z \preceq 0 \ \Longleftrightarrow \ \Pi \Phi(Z/\Pi) \preceq 0, \quad Z\in\SS^p \,. $$ This means that, for any $\pi>0$, $\Pi>0$, problem (\ref{eq:nlpsdp}) has the same solution as the following ``augmented" problem \begin{align} & \min_{x\in\RR^n, Y_1\in\SS^{p_1},\ldots,Y_k\in\SS^{p_k}} f(x,Y) \label{eq:nlpsdp_phi}\\ & \begin{aligned} \mbox{subject to}\quad &\varphi_{\pi}(g_i(x,Y)) \leq 0, \qquad &&i=1,\ldots,m_g\nonumber\\ &\Phi_{\Pi}({\cal A}_i(x,Y))\preceq 0,\qquad&&i=1,\ldots,m_A\nonumber\\ &\Phi_{\Pi}(\underline{\lambda}_i I - Y_i)\preceq 0, \qquad&&i=1,\ldots,k\nonumber\\ & \Phi_{\Pi}(Y_i-\overline{\lambda}_i I)\preceq 0, \qquad&&i=1,\ldots,k\nonumber\\ &h_i(x,Y) = 0, \qquad &&i=1,\ldots,m_h\,, \nonumber\\ \end{aligned} \end{align} where we have used the abbreviations $\varphi_{\pi} = \pi \varphi (\cdot / \pi)$ and $\Phi_{\Pi} = \Pi \Phi (\cdot / \Pi)$. \medskip The Lagrangian of (\ref{eq:nlpsdp_phi}) can be viewed as a (generalized) augmented Lagrangian of (\ref{eq:nlpsdp}): \begin{multline}\label{eq:lagr} F(x,Y,u,\Xi,\underline{U},\overline{U},v,\pi,\Pi)\\ =f(x,Y) + \sum_{i=1}^{m_g} u_i \varphi_{\pi}(g_i(x,Y)) + \sum_{i=1}^{m_A}\langle \Xi_i,\Phi_\Pi({\cal A}_i(x,Y))\rangle\\ + \sum_{i=1}^{k}\langle\underline{U}_i, \Phi_\Pi(\underline{\lambda}_i I - Y_i)\rangle + \sum_{i=1}^{k}\langle\overline{U}_i,\Phi_\Pi(Y_i-\overline{\lambda}_i I)\rangle +v^\top h(x, Y) \,; \end{multline} here $u\in\RR^{m_g}$, $\Xi = (\Xi_1,\ldots,\Xi_{m_A}), \Xi_i\in\SS^{p_{A_i}}$, and $\underline{U}=(\underline{U}_1,\ldots,\underline{U}_k), \overline{U}=(\overline{U}_1, \ldots, \overline{U}_k)$, $\underline{U}_i,\overline{U}_i\in\SS^{p_i}$, are Lagrange multipliers associated with the standard and the matrix inequality constraints, respectively, and $v\in \RR^{m_h}$ is the vector of Lagrangian multipliers associated with the equality constraints. The algorithm combines ideas of the (exterior) penalty and (interior) barrier methods with the augmented Lagrangian method. \begin{Algorithm}\label{algo:1} Let $x^1, Y^1$ and $u^1, \Xi^1, \underline{U}^1, \overline{U}^1, v^1$ be given. Let $\pi^1>0$, $\Pi^1>0$ and $\alpha^1>0$. For $\ell=1,2,\ldots$ repeat till a stopping criterium is reached: \begin{align*} (i)\qquad &\mbox{Find $x^{\ell+1}$, $Y^{\ell+1}$ and $v^{\ell+1}$ such that}\\ \qquad &\qquad\|\nabla_{x,Y} F(x^{\ell+1},Y^{\ell+1},u^{\ell},\Xi^{\ell},\underline{U}^{\ell}, \overline{U}^{\ell},v^{\ell+1},\pi^{\ell},\Pi^{\ell})\| \leq \alpha^{\ell}\\ \qquad &\qquad\| h(x^{\ell+1},Y^{\ell+1}) \| \leq \alpha^{\ell}\\ (ii)\qquad &u_i^{\ell+1} = u_i^{\ell}\varphi_{\pi^{\ell}}'(g_i(x^{\ell+1},Y^{\ell+1})),\quad i=1,\,\ldots,m_g\\ &\Xi_i^{\ell+1} = D_{\cal A} \Phi_{\Pi^{\ell}}({\cal A}_i(x^{\ell+1},Y^{\ell+1}); \Xi_i^{\ell}),\quad i=1,\,\ldots,m_A\\ &\underline{U}_i^{\ell+1} = D_{\cal A} \Phi_{\Pi^{\ell}}((\underline{\lambda}_i I - Y_i^{\ell+1}); \underline{U}_i^{\ell}),\quad i=1,\,\ldots,k\\ &\overline{U}_i^{\ell+1} = D_{\cal A} \Phi_{\Pi^{\ell}}(( Y_i^{\ell+1}- \overline{\lambda}_i I ); \overline{U}_i^{\ell}),\quad i=1,\,\ldots,k\\ (iii)\qquad &\pi^{\ell+1} < \pi^{\ell},\quad \Pi^{\ell+1} < \Pi^{\ell},\quad \alpha^{\ell+1} < \alpha^{\ell}\,. \end{align*} \end{Algorithm} In Step~(i) we attempt to find an approximate solution of the following system (in $x, Y$ and $v$): \begin{equation}\label{eq:KKT-eq} \begin{aligned} \qquad \nabla_{x,Y} {F} (x,Y,u,\Xi,\underline{U},\overline{U},v,\pi,\Pi) &= 0 \\ h(x,Y) &= 0\,, \end{aligned} \end{equation} where the penalty parameters $\pi, \Pi$, as well as the multipliers $u,\Xi,\underline{U},\overline{U}$ are fixed. In order to solve it, we apply the damped Newton method. Descent directions are calculated utilizing the MATLAB command {\tt ldl} that is based on the factorization routine MA57, in combination with an inertia correction strategy described in \cite{stingl}. In the forthcoming release of PENLAB, we will also apply iterative methods, as described in \cite{pen-iter}. The step length is derived using an augmented Lagrangian merit function defined as $$ {F} (x,Y,u,\Xi,\underline{U},\overline{U},v,\pi,\Pi) + \frac{1}{2\mu}\|h(x,Y)\|_2^2 $$ along with an Armijo rule. If there are no equality constraints in the problems, the unconstrained minimization in Step~(i) is performed by the modified Newton method with line-search (for details, see \cite{pennon}). The multipliers calculated in Step~(ii) are restricted in order to satisfy: $$ \mu < \frac{u_i^{\ell+1}}{u_i^{\ell}} < \frac1{\mu} $$ with some positive ${\mu} \leq 1$; by default, ${\mu} = 0.3$. A similar restriction procedure can be applied to the matrix multipliers $\underline{U}^{\ell+1}, \overline{U}^{\ell+1}$ and $\Xi$; see again \cite{pennon} for details. The penalty parameters $\pi, \Pi$ in Step~(iii) are updated by some constant factor dependent on the initial penalty parameters $\pi^1, \Pi^1$. The update process is stopped when $\pi_{eps}$ (by default $10^{-6}$) is reached. Algorithm~\ref{algo:1} is stopped when a criterion based on the KKT error is satisfied and both of the inequalities holds: \begin{eqnarray*} \frac{|f(x^{\ell},Y^{\ell}) - F(x^{\ell},Y^{\ell},u^{\ell},\Xi^{\ell},\underline{U}^{\ell},\overline{U}^{\ell},v^{\ell},\pi^{\ell},\Pi^{\ell})|} {1+|f(x^{\ell},Y^{\ell})|} &<& \epsilon\\ \frac{|f(x^{\ell},Y^{\ell}) - f(x^{\ell-1},Y^{\ell-1})|}{1+|f(x^{\ell},Y^{\ell})|} &<& \epsilon\,, \end{eqnarray*} where $\epsilon$ is by default $10^{-6}$. \subsection{{Choice of $\varphi$ and $\Phi$}}\label{sec:hess} To treat the standard NLP constraints, we use the penalty/barrier function proposed by Ben-Tal and Zibulevsky \cite{ben-tal-zibulevsky}: \begin{equation} \varphi_{\bar{\tau}} (\tau) = \left \{ \begin{aligned} &\tau + \frac{1}{2} \, \tau^2 &\mbox{if~}& \tau \geq \bar{\tau} \\ &- (1+ \bar{\tau})^2 \log \left ( \frac{1+ 2 \bar{\tau} -\tau} {1 + \bar{\tau}} \right) + \bar{\tau} + \frac{1}{2} \bar{\tau}^2 \ &\mbox{if~}& \tau < \bar{\tau} \,; \end{aligned} \right . \label{eq:phi} \end{equation} by default, $\bar{\tau} = - \frac{1}{2}$. The penalty function $\Phi_\Pi$ of our choice is defined as follows (here, for simplicity, we omit the variable $Y$): \begin{equation}\label{eq:pen} \Phi_\Pi({\cal A}(x)) = -\Pi^2({\cal A}(x) - \Pi I)^{-1} - \Pi I \,. \end{equation} The advantage of this choice is that it gives closed formulas for the first and second derivatives of $\Phi_\Pi$. Defining \begin{equation}\label{eq:Z} {\cal Z}(x) = -({\cal A}(x) - \Pi I)^{-1} \end{equation} we have (see \cite{pennon}): \begin{align*} \frac{\partial}{\partial x_i} \Phi_\Pi({\cal A}(x))& = \Pi^2{\cal Z}(x) \frac{\partial{\cal A}(x)}{\partial x_i} {\cal Z}(x) \label{eq:der1} \\ \frac{\partial^2}{\partial x_i\partial x_j} \Phi_\Pi({\cal A}(x)) & = \Pi^2{\cal Z}(x) \left(\frac{\partial{\cal A}(x)}{\partial x_i} {\cal Z}(x) \frac{\partial{\cal A}(x)}{\partial x_j} + \frac{\partial^2{\cal A}(x)}{\partial x_i\partial x_j} \right.\nonumber\\ &\left.\phantom{\Pi^2{\cal Z}(x)}\qquad\ \ + \frac{\partial{\cal A}(x)}{\partial x_j} {\cal Z}(x) \frac{\partial{\cal A}(x)}{\partial x_i}\right){\cal Z}(x)\,.&\nonumber \end{align*} \subsection{Strictly feasible constraints} In certain applications, some of the bound constraints must remain strictly feasible for all iterations because, for instance, the objective function may be undefined at infeasible points (see examples in Section~\ref{ex:truss}). To be able to solve such problems, we treat these inequalities by a classic barrier function. In case of matrix variable inequalities, we split $Y$ in non-strictly feasible matrix variables $Y_1$ and strictly feasible matrix variables $Y_2$, respectively, and define the augmented Lagrangian \begin{equation}\label{eq:lagr2} \widetilde{F}(x,Y_1,Y_2,u,\Xi,\underline{U},\overline{U},v,\pi,\Pi,\kappa) = F(x,Y_1,u,\Xi,\underline{U},\overline{U},v,\pi,\Pi) + \kappa \Phi_{\rm bar}(Y_2), \end{equation} where $\Phi_{\rm bar}$ can be defined, for example for the constraint $Y_2\succeq 0$, by $$\Phi_{\rm bar}(Y_2) = -\log\det(Y_2).$$ Strictly feasible variables $x$ are treated in a similar manner. Note that, while the penalty parameter $\pi$ may be constant from a certain index $\bar{\ell}$ (see again \cite{stingl} for details), the barrier parameter $\kappa$ is required to tend to zero with increasing $\ell$. \section{The code} PENLAB is a free open-source MATLAB implementation of the algorithm described above. The main attention was given to clarity of the code rather than tweaks to improve its performance. This should allow users to better understand the code and encourage them to edit and develop the algorithm further. The code is written entirely in MATLAB with an exception of two mex-functions that handles the computationally most intense task of evaluating the second derivative of the Augmented Lagrangian and a sum of multiple sparse matrices (a slower non-mex alternative is provided as well). The solver is implemented as a MATLAB handle class and thus it should be supported on all MATLAB versions starting from R2008a. PENLAB is distributed under GNU GPL license and can be downloaded from {\tt http://web.mat.bham.ac.uk/kocvara/penlab}. The distribution package includes the full source code and precompiled mex-functions, PENLAB User's Guide and also an internal (programmer's) documentation which can be generated from the source code. Many examples provided in the package show various ways of calling PENLAB and handling NLP-SDP problems. \subsection{Usage} The source code is divided between a class \verb|penlab| which implements Algorithm 1 and handles generic NLP-SDP problems similar to formulation (\ref{eq:nlpsdp}) and interface routines providing various specialized inputs to the solver. Some of these are described in Section~\ref{sec:modules}. The user needs to prepare a MATLAB structure (here called \verb|penm|) which describes the problem parameters, such as number of variables, number of constraints, lower and upper bounds, etc. Some of the fields are shown in Table~\ref{tab:7}, for a complete list see the PENLAB User's Guide. The structure is passed to \verb|penlab| which returns the initialized problem instance: \begin{verbatim} >> problem = penlab(penm); \end{verbatim} The solver might be invoked and results retrieved, for example, by calling \begin{verbatim} >> problem.solve() >> problem.x \end{verbatim} The point \verb|x| or option settings might be changed and the solver invoked again. The whole object can be cleared from the memory using \begin{verbatim} >> clear problem; \end{verbatim} \begin{table}[htbp] \caption{Selection of fields of the MATLAB structure {\tt penm} used to initialize PENLAB object. Full list is available in PENLAB User's Guide.} \begin{tabular*}{\hsize}{@{\extracolsep{\fill}}ll} \hline field name & meaning \\ \hline Nx & dimension of vector $x$ \\ NY & number of matrix variables $Y$\\ Y & cell array of length NY with a nonzero pattern of each of the matrix variables\\ lbY & NY lower bounds on matrix variables (in spectral sense) \\ ubY & NY upper bounds on matrix variables (in spectral sense) \\ NANLN & number of nonlinear matrix constraints \\ NALIN & number of linear matrix constraints \\ lbA & lower bounds on all matrix constraints\\ ubA & upper bounds on all matrix constraints\\ \hline \end{tabular*} \label{tab:7} \end{table} \subsection{Callback functions} The principal philosophy of the code is similar to many other optimization codes---we use callback functions (provided by the user) to compute function values and derivatives of all involved functions. For a generic problem, the user must define nine MATLAB callback functions: {\tt objfun}, {\tt objgrad}, {\tt objhess}, {\tt confun}, {\tt congrad}, {\tt conhess}, {\tt mconfun}, {\tt mcongrad}, {\tt mconhess} for function value, gradient, and Hessian of the objective function, (standard) constraints and matrix constraint. If one constraint type is not present, the corresponding callbacks need not be defined. Let us just show the parameters of the most complex callbacks for the matrix constraints: \begin{verbatim} function [Ak, userdata] = mconfun(x,Y,k,userdata) function [dAki,userdata] = mcongrad(x,Y,k,i,userdata) function [ddAkij, userdata] = mconhess(x,Y,k,i,j,userdata) \end{verbatim} Here $x,Y$ are the current values of the (vector and matrix) variables. Parameter $k$ stands for the constraint number. Because every element of the gradient and the Hessian of a matrix function is a matrix, we compute them (the gradient and the Hessian) element-wise (parameters $i,j$). The outputs {\tt Ak,dAki,ddAkij} are symmetric matrices saved in sparse MATLAB format. Finally, {\tt userdata} is a MATLAB structure passed through all callbacks for user's convenience and may contain any additional data needed for the evaluations. It is unchanged by the algorithm itself but it can be modified in the callbacks by user. For instance, some time-consuming computation that depends on $x,Y,k$ but is independent of $i$ can be performed only for $i=1$, the result stored in {\tt userdata} and recalled for any $i>1$ (see, e.g., Section~\ref{ex:truss}, example Truss Design with Buckling Constraint). \subsection{Mex files} Despite our intentions to use only pure Matlab code, two routines were identified to cause a significant slow-down and therefore their m-files were substituted with equivalent mex-files. The first one computes linear combination of a set of sparse matrices, e.g., when evaluating ${\cal A}_i(x)$ for polynomial matrix inequalities, and is based on ideas from \cite{davis}. The second one evaluates matrix inequality contributions to the Hessian of the augmented Lagrangian (\ref{eq:lagr}) when using penalty function (\ref{eq:pen}). The latter case reduces to computing $z_{\ell} = \langle TA_kU,\, A_{\ell}\rangle$ for $\ell=k,\ldots,n$ where $T, U \in \SS^m$ are dense and $A_{\ell} \in \SS^m$ are sparse with potentially highly varying densities. Such expressions soon become challenging for nontrivial $m$ and can easily dominate the whole Algorithm~\ref{algo:1}. Note that the problem is common even in primal-dual interior point methods for SDPs and have been studied in \cite{fujisawa-kojima-nakata}. We developed a relatively simple strategy which can be viewed as an evolution of the three computational formulae presented in \cite{fujisawa-kojima-nakata} and offers a minimal number of multiplications while keeping very modest memory requirements. We refer to it as a \emph{look-ahead strategy with caching}. It can be described as follows: \begin{Algorithm}\label{algo:trace} Precompute a set ${\cal J}$ of all nonempty columns across all $A_{\ell}, {\ell}=k,\ldots,n$ and a set ${\cal I}$ of nonempty rows of $A_k$ \emph{(look-ahead)}. Reset flag vector $c\leftarrow 0$, set $z=0$ and $v=w=0$. For each $j \in {\cal J}$ perform: \begin{enumerate} \item compute selected elements of the $j$-th column of $A_kU$, i.e.,\\ $v_i = \sum_{\alpha=1}^m(A_k)_{i \alpha} U_{\alpha j}$ for $i \in {\cal I}$, \item for each $A_{\ell}$ with nonempty $j$-th column go through its nonzero elements $(A_{\ell})_{ij}$ and \begin{enumerate} \item if $c_i<j$ compute $w_i = \sum_{\alpha \in {\cal I}} T_{i\alpha}v_{\alpha}$ and set $c_i \leftarrow j$ \emph{(caching)}, \item update trace, i.e., $z_{\ell} = z_{\ell} + w_i(A_{\ell})_{ij}$. \end{enumerate} \end{enumerate} \end{Algorithm} \section{Gradients and Hessians of matrix valued functions} There are several concepts of derivatives of matrix functions; they, however, only differ in the ordering of the elements of the resulting ``differential''. In PENLAB, we use the following definitions of the gradient and Hessian of matrix valued functions. \begin{definition}\label{def:1} Let $F$ be a differentiable $m\times n$ real matrix function of an $p\times q$ matrix of real variables $X$. The $(i,j)$-th element of the \emph{gradient} of $F$ at $X$ is the $m\times n$ matrix \begin{equation}\label{eq:a041} \left[\nabla F(X)\right]_{ij} := \frac{\partial F(X)}{\partial x_{ij}}, \qquad i=1,\ldots,p,\ j=1,\ldots,q \,. \end{equation} \end{definition} \begin{definition}\label{def:2} Let $F$ be a twice differentiable $m\times n$ real matrix function of an $p\times q$ matrix of real variables $X$. The $(ij,k\ell)$-th element of the \emph{Hessian} of $F$ at $X$ is the $m\times n$ matrix \begin{equation}\label{eq:a040} \left[\nabla^2 F(X)\right]_{ij,k\ell} := \frac{\partial^2 F(X)}{\partial x_{ij}\partial x_{kl}} , \qquad i,k=1,\ldots,p,\ j,\ell=1,\ldots,q \,. \end{equation} \end{definition} In other words, for every pair of variables $x_{ij},\ x_{k\ell}$, elements of $X$, the second partial derivative of $F(X)$ with respect to these variables is the $m\times n$ matrix $\frac{\partial^2 F(X)}{\partial x_{ij}\partial x_{k\ell}}$. How to compute these derivatives, i.e., how to define the callback functions? In Appendix A, we summarize basic formulas for the computation of derivatives of scalar and matrix valued functions of matrices. For low-dimensional problems, the user can utilize MATLAB's Symbolic Toolbox. For instance, for $F(X)=XX$, the commands \begin{verbatim} >> A=sym('X',[2,2]); >> J=jacobian(X*X,X(:)); >> H=jacobian(J,X(:)); \end{verbatim} generate arrays $J$ and $H$ such that the $i$-th column of $J$ is the vectorized $i$-th element of the gradient of $F(X)$; similarly, the $k$-th column of $H$, $k=(i-1)n^2+j$ for $i,j=1,\ldots,n^2$ is the vectorized $(i,j)$-th element of the Hessian of $F(X)$. Clearly, the dimension of the matrix variable is fixed and for a different dimension we have to generate new formulas. Unfortunately, this approach is useless for higher dimensional matrices (the user is invited to use the above commands for $F(X)=X^{-1}$ with $X\in\SS^5$ to see the difficulties). However, one can always use symbolic computation to check validity of general dimension independent formulas on small dimensional problems. \section{Pre-programmed interfaces}\label{sec:modules} PENLAB distribution contains several pre-programmed interfaces for standard optimization problems with standard inputs. For these problems, the user does not have to create the \verb|penm| object, nor the callback functions. \subsection{Nonlinear optimization with AMPL input} PENLAB can read optimization problems that are defined in and processed by AMPL \cite{ampl}. AMPL contains routines for automatic differentiation, hence the gradients and Hessians in the callbacks reduce to calls to appropriate AMPL routines. Assume that nonlinear optimization problem is processed by AMPL, so that we have the corresponding \verb|.nl| file, for instance \verb|chain.nl|, stored in directory \verb|datafiles|. All the user has to do to solve the problem is to call the following three commands: \begin{verbatim} >> penm = nlp_define('datafiles/chain100.nl'); >> problem = penlab(penm); >> problem.solve(); \end{verbatim} \subsection{Linear semidefinite programming} Assume that the data of a linear SDP problem is stored in a MATLAB structure \verb|sdpdata|. Alternatively, such a structure can be created by the user from SDPA input file \cite{sdpa}. For instance, to read problem \verb|arch0.dat-s| stored in directory \verb|datafiles|, call \begin{verbatim} >> sdpdata = readsdpa('datafiles/control1.dat-s'); \end{verbatim} To solve the problem by PENLAB, the user just has to call the following sequence of commands: \begin{verbatim} >> penm = sdp_define(sdpdata); >> problem = penlab(penm); >> problem.solve(); \end{verbatim} \subsection{Bilinear matrix inequalities} We want to solve an optimization problem with quadratic objective and constraints in the form of bilinear matrix inequalities: \begin{align}\label{eq:bmiproblem} &\min_{x\in\RR^n} \frac{1}{2} x^T H x + c^T x \\ & \begin{aligned} \mbox{subject to}\quad &b_{\rm low}\leq Bx\leq {b_{\rm up}}\nonumber\\ &Q^i_0 + \sum_{k=1}^{n} x_k Q^i_k + \sum_{k=1}^{n}\sum_{\ell=1}^{n} x_k x_\ell Q^i_{k\ell} \succcurlyeq 0, \quad i=1,\ldots,m\,. \nonumber \end{aligned} \end{align} The problem data should be stored in a simple format explained in PENLAB User's Guide. All the user has to do to solve the problem is to call the following sequence of commands: \begin{verbatim} >> load datafiles/bmi_example; >> penm = bmi_define(bmidata); >> problem = penlab(penm); >> problem.solve(); \end{verbatim} \subsection{Polynomial matrix inequalities} We want to solve an optimization problem with constraints in the form of polynomial matrix inequalities: \begin{align}\label{eq:pmiproblem} &\min_{x\in\RR^n} \frac{1}{2} x^T H x + c^T x \\ & \begin{aligned} \mbox{subject to}\quad &b_{\rm low}\leq Bx\leq {b_{\rm up}}\nonumber\\ &{\cal A}_i(x)\succcurlyeq 0,\quad i=1,\ldots,m \nonumber \end{aligned} \end{align} with $$ {\cal A}_i(x) = \sum_j x^{(\kappa^i(j))} Q^i_j $$ where $\kappa^i(j)$ is a multi-index of the $i$-th constraint with possibly repeated entries and $x^{(\kappa^i(j))}$ is a product of elements with indices in $\kappa^i(j)$. For example, for $${\cal A}(x) = Q_1 + x_1 x_3 Q_2 + x_2 x_4^3 Q_3$$ the multi-indices are $\kappa(1) = \{0\}$ ($Q_1$ is an absolute term), $\kappa(2) = \{1,3\}$ and $\kappa(3) = \{2,4,4,4\}$. Assuming now that the problem is stored in a structure \verb|pmidata| (as explained in PENLAB User's Guide), the user just has to call the following sequence of commands: \begin{verbatim} >> load datafiles/pmi_example; >> penm = pmi_define(pmidata); >> problem = penlab(penm); >> problem.solve(); \end{verbatim} \section{Examples} All MATLAB programs and data related to the examples in this section can be found in directories \verb|examples| and \verb|applications| of the PENLAB distribution. \subsection{Correlation matrix with the constrained condition number}\label{ex:cond} We consider the problem of finding the nearest correlation matrix (\cite{higham}): \begin{align} &\min_X \sum_{i,j=1}^n (X_{ij}-H_{ij})^2\label{corr1}\\ &\mbox{subject to}\nonumber\\ &\qquad X_{ii} = 1,\quad i=1,\ldots,n\nonumber\\ &\qquad X\succeq 0\,.\nonumber \end{align} In addition to this standard setting of the problem, let us bound the condition number of the nearest correlation matrix by adding the constraint $$ \mbox{cond}(X) = \kappa \,. $$ We can formulate this constraint as \begin{align} I\preceq\widetilde{X}\preceq \kappa I \end{align} the variable transformation $$ \widetilde{X} = \zeta X\,. $$ After the change of variables, and with the new constraint, the problem of finding the nearest correlation matrix with a given condition number reads as follows: \begin{align} &\min_{\zeta,\widetilde{X}} \sum_{i,j=1}^n (\frac{1}{\zeta}\widetilde{X}_{ij}-H_{ij})^2\label{corr_cond}\\ &\mbox{subject to}\nonumber\\ &\qquad \widetilde{X}_{ii} -\zeta = 0,\quad i=1,\ldots,n\nonumber\\ &\qquad I\preceq\widetilde{X}\preceq \kappa I\nonumber \end{align} The new problem now has the NLP-SDP structure of (\ref{eq:nlpsdp}). We will consider an example based on a practical application from finances; see \cite{werner-schoettle}. Assume that we are given a $5\times 5$ correlation matrix. We now add a new asset class, that means, we add one row and column to this matrix. The new data is based on a different frequency than the original part of the matrix, which means that the new matrix is no longer positive definite: $$ H_{\rm ext} = \begin{pmatrix}1 &-0.44& -0.20 &0.81& -0.46& -0.05\\ -0.44& 1 &0.87& -0.38& 0.81 & -0.58\\ -0.20 &.87 &1& -0.17& 0.65& -0.56\\ 0.81 &-0.38& -0.17& 1& -0.37& -0.15\\ -0.46& 0.81& 0.65& -0.37& 1& -0.08\\ -0.05&-0.58&-0.56&-0.15&0.08&1 \end{pmatrix}\,. $$ When solving problem (\ref{corr_cond}) by PENLAB with $\kappa=10$, we get the solution after 11 outer and 37 inner iterations. The optimal value of $\zeta$ is $3.4886$ and, after the back substitution $X = \frac{1}{\zeta}\widetilde{X}$, we get the nearest correlation matrix \begin{verbatim} X = 1.0000 -0.3775 -0.2230 0.7098 -0.4272 -0.0704 -0.3775 1.0000 0.6930 -0.3155 0.5998 -0.4218 -0.2230 0.6930 1.0000 -0.1546 0.5523 -0.4914 0.7098 -0.3155 -0.1546 1.0000 -0.3857 -0.1294 -0.4272 0.5998 0.5523 -0.3857 1.0000 -0.0576 -0.0704 -0.4218 -0.4914 -0.1294 -0.0576 1.0000 \end{verbatim} with eigenvalues \begin{verbatim} eigenvals = 0.2866 0.2866 0.2867 0.6717 1.6019 2.8664 \end{verbatim} and the condition number equal to 10, indeed. \paragraph{Gradients and Hessians} What are the first and second partial derivatives of functions involved in problem (\ref{corr_cond})? The constraint is linear, so the answer is trivial here, and we can only concentrate on the objective function \begin{equation}\label{eq:f} f(z,\widetilde{X}):=\sum_{i,j=1}^n (z\widetilde{X}_{ij}-H_{ij})^2 = \langle z\widetilde{X}-H, z\widetilde{X}-H \rangle\,, \end{equation} where, for convenience, we introduced a variable $z=\frac{1}{\zeta}$. \begin{theorem} Let $x_{ij}$ and $h_{ij}$, $i,j=1,\ldots,n$ be elements of $\widetilde{X}$ and $H$, respectively. For the function $f$ defined in (\ref{eq:f}) we have the following partial derivatives: \begin{itemize} \item[(i)] $\nabla_{\!z}\, f(z,\widetilde{X}) = 2\langle \widetilde{X}, z\widetilde{X}-H\rangle$ \item[(ii)] $\left[\nabla_{\!\widetilde{X}}f(z,\widetilde{X})\right]_{ij} = 2z(zx_{ij}-h_{ij})$,\quad $i,j=1,\ldots ,n$ \item[(iii)] $\nabla^2_{\!z,z}\, f(z,\widetilde{X}) = 2\langle \widetilde{X},\widetilde{X}\rangle$ \item[(iv)] $\left[\nabla^2_{\!z,\widetilde{X}}\, f(z,\widetilde{X})\right]_{ij} = \left[\nabla^2_{\!\widetilde{X},z}\, f(z,\widetilde{X})\right]_{ij} = 4zx_{ij} - 2h_{ij}$,\quad $i,j=1,\ldots ,n$ \item[(v)] $\left[\nabla^2_{\!\widetilde{X},\widetilde{X}}\, f(z,\widetilde{X})\right]_{ij,k\ell} = 2z^2$\quad for $i=k,\ j=\ell$ and zero otherwise ($i,j,k,\ell=1,\ldots ,n$)\,. \end{itemize} \end{theorem} The proof follows directly from formulas in Appendix~A. \paragraph{PENLAB distribution} This problem is stored in directory \verb|applications/CorrMat| of the PENLAB distribution. To solve the above example and to see the resulting eigenvalues of $X$, run in its directory \begin{verbatim} >> penm = corr_define; >> problem = penlab(penm); >> problem.solve(); >> eig(problem.Y{1}*problem.x) \end{verbatim} \subsection{Truss topology optimization with stability constraints}\label{ex:truss} In truss optimization we want to design a pin-jointed framework consisting of $m$ slender bars of constant mechanical properties characterized by their Young's modulus $E$. We will consider trusses in a $d$-dimensional space, where $d=2$ or $d=3$. The bars are jointed at $\tilde{n}$ nodes. The system is under load, i.e., forces $f_j\in\RR^{d}$ are acting at some nodes $j$. They are aggregated in a vector $f$, where we put $f_j=0$ for nodes that are not under load. This external load is transmitted along the bars causing displacements of the nodes that make up the displacement vector $u$. Let $p$ be the number of fixed nodal coordinates, i.e., the number of components with prescribed discrete homogeneous Dirichlet boundary condition. We omit these fixed components from the problem formulation reducing thus the dimension of $u$ to $$ n=d\,\cdot\,\tilde{n} - p . $$ Analogously, the external load $f$ is considered as a vector in $\RR^n$. The design variables in the system are the bar volumes $x_1,\ldots,x_m$. Typically, we want to minimize the weight of the truss. We assume to have a unique material (and thus density) for all bars, so this is equivalent to minimizing the volume of the truss, i.e., $\sum_{i=1}^m x_i$. The optimal truss should satisfy mechanical equilibrium conditions: \begin{equation} K(x)u=f \,; \label{eq:3b1} \end{equation} here \begin{equation} K(x):= \sum\limits_{i=1}^m x_iK_i,\quad K_i=\frac{E_i}{\ell_i^2}\gamma_i\gamma_i^{\top} \label{KO5eq:1} \end{equation} is the so-called stiffness matrix, $E_i$ the Young modulus of the $i$th bar, $\ell_i$ its length and $\gamma_i$ the $n-$vector of direction cosines. We further introduce the compliance of the truss $f^{\top}u$ that indirectly measures the stiffness of the structure under the force $f$ and impose the constraints $$ f^{\top}u \leq \gamma\,. $$ This constraint, together with the equilibrium conditions, can be formulated as a single linear matrix inequality (\cite{buck}) $$ \begin{pmatrix} K(x) & f\\ f^T &\gamma\end{pmatrix} \succeq 0\,. $$ The minimum volume single-load truss topology optimization problem can then be formulated as a linear semidefinite program: \begin{align} &\min_{x\in\RR^m} \sum_{i=1}^m x_i\label{minvolc}\\ &\mbox{subject to}\nonumber\\ &\qquad \begin{pmatrix} K(x) & f\\ f^T &\gamma\end{pmatrix} \succeq 0 \nonumber\\ &\qquad x_i\geq 0,\quad i=1,\ldots,m\,.\nonumber \end{align} We further consider the constraint on the global stability of the truss. The meaning of the constraint is to avoid global buckling of the optimal structure. We consider the simplest formulation of the buckling constraint based on the so-called linear buckling assumption \cite{buck}. As in the case of free vibrations, we need to constrain eigenvalues of the generalized eigenvalue problem \begin{equation}\label{eq:buckEVP} K(x) w = \lambda {G}(x) w \,, \end{equation} in particular, we require that all eigenvalues of (\ref{eq:buckEVP}) lie outside the interval [0,1]. The so-called geometry stiffness matrix ${G}(x)$ depends, this time, nonlinearly on the design variable $x$: \begin{equation}\label{eq:G} {G}(x) = \sum_{i=1}^m {G}_i(x), \qquad {G}_i(x) = \frac{E x_i}{\ell_i^d} (\gamma_i^{\top} K(x)^{-1}f) (\delta_i\delta_i^{\top}+\eta_i\eta_i^{\top}). \end{equation} Vectors $\delta,\eta$ are chosen so that $\gamma,\delta,\eta$ are mutually orthogonal. (The presented formula is for $d=3$. In the two-dimensional setting the vector $\eta$ is not present.) To simplify the notation, we denote $$\Delta_i = \delta_i\delta^T_i + \eta_i\eta^T_i\,.$$ It was shown in \cite{buck} that the eigenvalue constraint can be equivalently written as a nonlinear matrix inequality \begin{equation} \label{eq:truss_buck} K(x)+{G}(x) \succcurlyeq 0 \end{equation} that is now to be added to (\ref{minvolc}) to get the following nonlinear semidefinite programming problem. Note that $x_i$ are requested to be strictly feasible. \begin{align} &\min_{x\in\RR^m} \sum_{i=1}^m x_i\label{eq:truss}\\ &\mbox{subject to}\nonumber\\ &\qquad \begin{pmatrix} K(x) & f\\ f^T &\gamma\end{pmatrix} \succeq 0 \nonumber\\ &\qquad K(x)+{G}(x) \succcurlyeq 0 \nonumber\\ &\qquad x_i > 0,\quad i=1,\ldots,m\,\nonumber \end{align} \paragraph{Gradients and Hessians} Let $M: \RR^m \to \RR^{n\times n}$ be a matrix valued function assigning each vector $\xi$ a matrix $M(\xi)$. We denote by $\nabla\!_k M$ the partial derivative of $M(\xi)$ with respect to the $k$-th component of vector $\xi$. \begin{lemma}[based on \cite{magnus1988matrix}] Let $M: \RR^m \to \RR^{n\times n}$ be a symmetric matrix valued function assigning each $\xi\in\RR^m$ a nonsingular $(n\times n)$ matrix $M(\xi)$. Then (for convenience we omit the variable $\xi$) $$ \nabla\!_k M^{-1} = -M^{-1} (\nabla\!_k M) M^{-1}\,. $$ If $M$ is a linear function of $\xi$, i.e., $M(\xi) = \sum_{i=1}^m \xi_i M_i$ with symmetric positive semidefinite $M_i, i=1,\ldots,m,$ then the above formula simplifies to $$ \nabla\!_k M^{-1} = -M^{-1} M_k M^{-1}\,. $$ \end{lemma} \begin{theorem}[\cite{buck}] Let $G(x)$ be given as in (\ref{eq:G}). Then $$ [\nabla {G}\,]_k = {E\over\ell_k^3}\gamma_k^T K^{-1}f\Delta_k - \sum_{j=1}^m {Et_j\over\ell_j^3}\gamma_j^T K^{-1}K_k K^{-1} f \Delta_j $$ and \begin{multline*} \displaystyle [\nabla^2 {G}\,]_{\!k\ell} = \displaystyle -{E\over\ell_k^3}\gamma_k^T K^{-1}K_\ell K^{-1}f\Delta_k -{E\over\ell_\ell^3}\gamma_\ell^T K^{-1}K_k K^{-1}f\Delta_\ell\\ \displaystyle -\sum_{j=1}^m {Et_j\over\ell_j^3}\gamma_j^T K^{-1}K_\ell K^{-1}K_k K^{-1} f \Delta_j\\ -\sum_{j=1}^m {Et_j\over\ell_j^3}\gamma_j^T K^{-1}K_k K^{-1}K_\ell K^{-1} f \Delta_j. \end{multline*} \end{theorem} \paragraph{Example} Consider the standard example of a laced column under axial loading (example \verb|tim| in the PENLAB collection). Due to symmetry, we only consider one half of the column, as shown in Figure~\ref{fig:tr1}(top-peft); it has 19 nodes and 42 potential bars, so $n=34$ and $m=42$. The column dimensions are $8.5\times 1$, the two nodes on the left-hand side are fixed and the ``axial'' load applied at the column tip is $(0,-10)$. The upper bound on the compliance is chosen as $\gamma=1$. Assume first that $x_i=0.425, i=1,\ldots,m$, i.e., the volumes of all bars are equal and the total volume is 17.85. The values of $x_i$ were chosen such that the truss satisfies the compliance constraint: $f^{\top}u =0.9923\leq \gamma$. For this truss, the smallest nonnegative eigenvalue of (\ref{eq:buckEVP}) is equal to 0.7079 and the buckling constraint (\ref{eq:truss_buck}) is not satisfied. Figure \ref{fig:tr1}(top-right) shows the corresponding the buckling mode (eigenvector associated with this eigenvalue). \begin{figure}[h] \begin{center} \resizebox{0.39\hsize}{!} {\includegraphics{pic/tim_fig1.eps}}\qquad \resizebox{0.39\hsize}{!} {\includegraphics{pic/tim_fig2.eps}}\\[1.5em] \resizebox{0.39\hsize}{!} {\includegraphics{pic/tim_fig3.eps}}\qquad \resizebox{0.39\hsize}{!} {\includegraphics{pic/tim_fig4.eps}} \end{center} \caption{Truss optimization with stability problem: initial truss (top-left); its buckling mode (top-right); optimal truss without stability constraint (bottom-left); and optimal stable truss (bottom-right)}\label{fig:tr1} \end{figure} Let us now solve the truss optimization problem \emph{without} the stability constraint (\ref{eq:truss}). We obtain the design shown in Figure~\ref{fig:tr1}(bottom-left). This truss is much lighter than the original one ($\sum\limits_{i=1}^m x_i = 9.388$), it is, however, extremely unstable under the given load, as (\ref{eq:buckEVP}) has a zero eigenvalue. When solving the truss optimization problem \emph{with} the stability constraint (\ref{eq:truss}) by PENLAB, we obtain the design shown in Figure~\ref{fig:tr1}(bottom-right). This truss is still significantly lighter than the original one ($\sum\limits_{i=1}^m x_i = 12.087$), but it is now stable under the given load. To solve the nonlinear SDP problem, PENLAB needed 18 global and 245 Newton iterations and 212 seconds of CPU time, 185 of which were spent in the Hessian evaluation routines. \paragraph{PENLAB distribution} Directories \verb|applications/TTO| and \verb|applications/TTObuckling| of the PENLAB distribution contain the problem formulation and many examples of trusses. To solve the above example with the buckling constraint, run \begin{verbatim} >> solve_ttob('GEO/tim.geo') \end{verbatim} in directory \verb|TTObuckling|. \subsection{Static output feedback} Given a linear system with $A\in\RR^{n\times n}, B\in\RR^{n\times m}, C\in\RR^{p\times n}$ \begin{align*} \dot{x} &= Ax + Bu\\ y& = Cx \end{align*} we want to stabilize it by static output feedback $ u = Ky \,. $ That is, we want to find a matrix $K\in\RR^{m\times p}$ such that the eigenvalues of the closed-loop system $A+BKC$ belong to the left half-plane. The standard way how to treat this problem is based on the Lyapunov stability theory. It says that $A+BKC$ has all its eigenvalues in the open left half-plane if and only if there exists a symmetric positive definite matrix $P$ such that \begin{equation}\label{eq:sofbmi} (A+BKC)^T P+P(A+BKC) \succ 0\,. \end{equation} Hence, by introducing the new variable, the Lyapunov matrix $P$, we can formulate the SOF problem as a feasibility problem for the bilinear matrix inequality (\ref{eq:sofbmi}) in variables $K$ and $P$. As typically $n>p,m$ (often $n\gg p,m$), the Lyapunov variable dominates here, although it is just an auxiliary variable and we do not need to know its value at the feasible point. Hence a natural question arises whether we can avoid the Lyapunov variable in the formulation of the problem. The answer was given in \cite{SOF2005} and lies in the formulation of the problem using polynomial matrix inequalities. Let $k=\Vec K$. Define the characteristic polynomial of $A+BKC$: $$ q(s,k) = \det(sI-A-BKC) = \sum_{i=0}^n q_i(k)s^i\,, $$ where $q_i(k) = \sum_\alpha q_{i\alpha}k^\alpha$ and $\alpha\in\NN^{mp}$ are all monomial powers. The \emph{Hermite stability criterion} says that the roots of $q(s,k)$ belong to the stability region {\cal D} (in our case the left half-plane) if and only if $$ H(q) = \sum_{i=0}^n\sum_{j=0}^n q_i(k)q_j(k) H_{ij} \succ 0 \,. $$ Here the coefficients $H_{ij}$ depend on the stability region only (see, e.g., \cite{ecc}). For instance, for $n=3$, we have $$ H(q) = \begin{pmatrix} 2q_0q_1 & 0 & 2q_0q_3\\ 0 & 2q_1q_2-2q_0q_3 & 0\\ 2q_0q_3 & 0 & 2q_2q_3 \end{pmatrix}\,. $$ The Hermite matrix $H(q)=H(k)$ depends polynomially on $k$: \begin{equation}\label{eq:sofpmi} H(k) = \sum_\alpha H_\alpha k^\alpha \succ 0 \end{equation} where $H_\alpha = H_\alpha^T\in\RR^{n\times n}$ and $\alpha\in\NN^{mp}$ describes all monomial powers. \begin{theorem}[\cite{SOF2005}] Matrix $K$ solves the static output feedback problem if and only if $k=\Vec K$ satisfies the polynomial matrix inequality (\ref{eq:sofpmi}). \end{theorem} In order to solve the strict feasibility problem (\ref{eq:sofpmi}), we can solve the following optimization problem with a polynomial matrix inequality \begin{align}\label{eq:sofpmi1} &\max_{k\in\RR^{mp},\,\lambda\in\RR} \lambda-\mu\|k\|^2 \\ & \mbox{subject to}\quad H(k)\succcurlyeq \lambda I\,. \nonumber \end{align} Here $\mu>0$ is a parameter that allows us to trade off between feasibility of the PMI and a moderate norm of the matrix $K$, which is generally desired in practice. \paragraph{COMPlib examples} In order to use PENLAB for the solution of SOF problems (\ref{eq:sofpmi1}), we have developed an interface to the problem library COMPlib \cite{complib}\footnote{The authors would like to thank Didier Henrion, LAAS-CNRS Toulouse, for developing a substantial part of this interface.}. Table~\ref{tab:11} presents the results of our numerical tests. We have only solved COMPlib problems of small size, with $n<10$ and $mp<20$. The reason for this is that our MATLAB implementation of the interface (building the matrix $H(k)$ from COMPlib data) is very time-consuming. For each COMPlib problem, the table shows the degree of the matrix polynomial, problem dimensions $n$ and $mp$, the optimal $\lambda$ (the negative largest eigenvalue of the matrix $K$), the CPU time and number of Newton iterations/linesearch steps of PENLAB. The final column contains information about the solution quality. ``F'' means failure of PENLAB to converge to an optimal solution. The plus sign ``+'' means that PENLAB converged to a solution which does not stabilize the system and "0" is used when PENLAB converged to a solution that is on the boundary of the feasible domain and thus not useful for stabilization. \begin{table}[htbp] \caption{mmm} \begin{tabular*}{\hsize}{@{\extracolsep{\fill}}lcccrrrc} \hline Problem & degree & $n$ & $mp$ & $\lambda_{\rm opt}$& CPU (sec)&iter &remark \\ \hline AC1 & 5 & 5 & 9 & $-0.871\cdot 10^{0}$ & 2.2 & 27/30& \\ AC2 & 5 & 5 & 9 & $-0.871\cdot 10^{0}$ & 2.3 & 27/30& \\ AC3 & 4 & 5 & 8 & $-0.586\cdot 10^{0}$ & 1.8 & 37/48& \\ AC4 & 2 & 4 & 2 & $0.245\cdot 10^{-2}$ & 1.9 & 160/209& + \\ AC6 & 4 & 7 & 8 & $-0.114\cdot 10^{4}$ & 1.2 & 22/68& \\ AC7 & 2 & 9 & 2 & $-0.102\cdot 10^{3}$& 0.9 & 26/91 &\\ AC8 & 2 & 9 & 5 & $0.116\cdot 10^{0}$ & 3.9 & 346/1276 & F \\ AC11 & 4 & 5 & 8 & $-0.171\cdot 10^{5}$ & 2.3 & 65/66& \\ AC12 & 6 & 4 & 12 & $0.479\cdot 10^{0}$ & 12.3 & 62/73& + \\ AC15 & 4 & 4 & 6 & $-0.248\cdot 10^{-1}$ & 1.2 & 25/28 & \\ AC16 & 4 & 4 & 8 & $-0.248\cdot 10^{-1}$ & 1.2 & 23/26 & \\ AC17 & 2 & 4 & 2 & $-0.115\cdot 10^{2}$ & 1.0 & 19/38 & \\ HE1 & 2 & 4 & 2 & $-0.686\cdot 10^{2}$ & 1.0 & 22/22 & \\ HE2 & 4 & 4 & 4 & $-0.268\cdot 10^{0}$ & 1.6 & 84/109 & \\ HE5 & 4 & 8 & 8 & $0.131\cdot 10^{2}$ & 1.9 & 32/37 & + \\ REA1 & 4 & 4 & 6 & $-0.726\cdot 10^{2}$ & 1.4 & 33/35 & \\ REA2 & 4 & 4 & 4 &$-0.603\cdot 10^{2}$ & 1.3 & 34/58 & \\ DIS1 & 8 & 8 & 16 & $-0.117\cdot 10^{2}$ & 137.6 & 30/55 & \\ DIS2 & 4 & 3 & 4 & $-0.640\cdot 10^{1}$ & 1.6 & 59/84 & \\ DIS3 & 8 & 6 & 16 & $-0.168\cdot 10^{2}$ & 642.3 & 66/102 & \\ MFP & 3 & 4 & 6 & $-0.370\cdot 10^{-1}$ & 1.0 & 20/21 & \\ TF1 & 4 & 7 & 8 & $-0.847\cdot 10^{-8}$ & 1.7 & 27/31 & 0 \\ TF2 & 4 & 7 & 6 & $-0.949\cdot 10^{-7}$ & 1.3 & 19/23 & 0 \\ TF3 & 4 & 7 & 6 & $-0.847\cdot 10^{-8}$ & 1.6 & 28/38 & 0 \\ PSM & 4 & 7 & 6 & $-0.731\cdot 10^{2}$ & 1.1 & 17/39 & \\ NN1 & 2 & 3 & 2 & $-0.131\cdot 10^{0}$ & 1.2 & 32/34 & 0\\ NN3 & 2 & 4 & 1 & $0.263\cdot 10^{2}$ & 1.0 & 31/36 & +\\ NN4 & 4 & 4 & 6 & $-0.187\cdot 10^{2}$ & 1.2 & 33/47 & \\ NN5 & 2 & 7 & 2 & $0.137\cdot 10^{2}$ & 1.5 & 108/118 & +\\ NN8 & 3 & 3 & 4 & $-0.103\cdot 10^{1}$ & 1.0 & 19/29 & \\ NN9 & 4 & 5 & 6 & $0.312\cdot 10^{1}$ & 1.6 & 64/97 & +\\ NN10 & 6 & 8 & 9 & $0.409\cdot 10^{4}$ & 18.3 &300/543 & F\\ NN12 & 4 & 6 & 4 & $0.473\cdot 10^{1}$ & 1.4 & 47/58 & + \\ NN13 & 4 & 6 & 4 & $0.279\cdot 10^{12}$ & 2.2 & 200/382 &F\\ NN14 & 4 & 6 & 4 & $0.277\cdot 10^{12}$ & 2.3 & 200/382 &F\\ NN15 & 3 & 3 & 4 & $-0.226\cdot 10^{0}$ & 1.0 & 15/14 & \\ NN16 & 7 & 8 & 16 & $-0.623\cdot 10^{3}$ &613.3 &111/191 & \\ NN17 & 2 & 3 & 2 & $0.931\cdot 10^{-1}$ & 1.0 & 25/26 & +\\ \hline \end{tabular*} \label{tab:11} \end{table} The reader can see that PENLAB can solve all problems apart from AC7, NN10, NN13 and NN14; these problems are, however, known to be very ill-conditioned and could not be solved via the Lyapunov matrix approach either (see \cite{SOF2004}). Notice that the largest problems with polynomials of degree up to 8 did not cause any major difficulties to the algorithm. \paragraph{PENLAB distribution} The related MATLAB programs are stored in directory \verb|applications/SOF| of the PENLAB distribution. To solve, for instance, example AC1, run \begin{verbatim} >> sof('AC1'); \end{verbatim} COMPlib program and library must be installed on user's computer. \section{PENLAB versus PENNON (MATLAB versus C)}\label{sec:comparison} The obvious concern of any user will be, how fast (or better, how slow) is the MATLAB implementation and if it can solve any problems of non-trivial size. The purpose of this section is to give a very rough comparison of PENLAB and PENNON, i.e., the MATLAB and C implementation of the same algorithm. The reader should, however, not make any serious conclusion from the tables below, for the following reasons: \begin{itemize} \item Both implementations slightly differ. This can be seen on the different numbers of iterations needed to solve single examples. \item The difference in CPU timing very much depends on the type of the problem. For instance, some problems require multiplications of sparse matrices with dense ones---in this case, the C implementation will be much faster. On the other hand, for some problems most of the CPU time is spent in the dense Cholesky factorization which, in both implementations, relies on LAPACK routines and thus the running time may be comparable. \item The problems were solved using an Intel i7 processor with two cores. The MATLAB implementation used both cores to perform \emph{some} commands, while the C implementation only used one core. This is clearly seen, e.g., example lame\_emd10 in Table~\ref{tab:12}. \item For certain problems (such as mater2 in Table~\ref{tab:14}), most of the CPU time of PENLAB is spent in the user defined routine for gradient evaluation. For linear SDP, this only amounts to reading the data matrices, in our implementation elements of a two-dimensional cell array, from memory. Clearly, a more sophisticated implementation would improve the timing. \end{itemize} For all calculations, we have used a notebook running Windows 7 (32 bit) on Intel Core i7 CPU [email protected] with 4GB memory and MATLAB 7.7.0. \subsection{Nonlinear programming problems} We first solved selected examples from the COPS collection \cite{cops} using AMPL interface. These are medium size examples mostly coming from finite element discretization of optimization problems with PDE constraints. Table~\ref{tab:12} presents the results. \begin{table}[htbp] \caption{Selected COPS examples. CPU time is given in seconds. Iteration count gives the number of the global iterations in Algorithm~\ref{algo:1} and the total number of steps of the Newton method.} \begin{tabular*}{\hsize}{@{\extracolsep{\fill}}crrrrrrr} \hline problem & vars& constr. & constraint & \multicolumn{2}{c}{PENNON} & \multicolumn{2}{c}{PENLAB} \\ & & & type & CPU & iter. & CPU & iter. \\\hline elec200 & 600 & 200 & $=$ & 40 & 81/224 & 31 & 43/135 \\ chain800 & 3199 & 2400 & $=$ & 1 & 14/23 & 6 & 24/56\\ pinene400 & 8000 & 7995 & $=$ & 1 & 7/7 & 11 & 17/17\\ channel800 & 6398 & 6398 & $=$ & 3 & 3/3 & 1 & 3/3\\ torsion100 & 5000 & 10000 & $\leq$ & 1 & 17/17 & 17 & 26/26 \\ bearing100 & 5000 & 5000 & $\leq$ & 1 & 17/17 & 13 & 36/36 \\ lane\_emd10& 4811 & 21 & $\leq$ & 217 & 30/86 & 64 & 25/49\\ dirichlet10& 4491 & 21 & $\leq$ & 151 & 33/71 & 73 & 32/68 \\ henon10 & 2701 & 21 & $\leq$ & 57 & 49/128 & 63 & 76/158 \\ minsurf100 & 5000 & 5000 & box & 1 & 20/20 & 97 & 203/203 \\ gasoil400 & 4001 & 3998 & $=$ \& box & 3 & 34/34 & 13 & 59/71\\ duct15 & 2895 & 8601 & $=$ \& $\leq$& 6 & 19/19 & 9 & 11/11\\ tri\_turtle& 3578 & 3968 &$\leq$ \& box& 3 & 49/49 & 4 & 17/17\\ marine400 & 6415 & 6392 &$\leq$ \& box& 2 & 39/39 & 22 & 35/35 \\ steering800& 3999 & 3200 &$\leq$ \& box& 1 & 9/9 & 7 & 19/40 \\ methanol400& 4802 & 4797 &$\leq$ \& box& 2 & 24/24 & 16 & 47/67 \\ catmix400 & 4398 & 3198 &$\leq$ \& box& 2 & 59/61 & 15 & 44/44 \\ \hline \end{tabular*} \label{tab:12} \end{table} \subsection{Linear semidefinite programming problems} We solved selected problems from the SDPLIB collection (Table~\ref{tab:13}) and Topology Optimization collection (Table~\ref{tab:14}); see \cite{sdplib,topo}. The data of all problems were stored in SDPA input files \cite{sdpa}. Instead of PENNON, we have used its clone PENSDP that directly reads the SDPA files and thus avoid repeated calls of the call back functions. The difference between PENNON and PENSDP (in favour of PENSDP) would only be significant in the mater2 example with many small matrix constraints. \begin{table}[htbp] \caption{Selected SDPLIB examples. CPU time is given in seconds. Iteration count gives the number of the global iterations in Algorithm~\ref{algo:1} and the total number of steps of the Newton method.} \begin{tabular*}{\hsize}{@{\extracolsep{\fill}}crrrrrrr} \hline problem & vars& constr. & constr. & \multicolumn{2}{c}{PENSDP} & \multicolumn{2}{c}{PENLAB} \\ & & & size & CPU & iter. & CPU & iter. \\\hline control3 & 136 & 2 & 30 & 1 & 19/103 & 20 & 22/315 \\ maxG11 & 800 & 1& 1600 & 18 & 22/41 & 186 & 18/61 \\ qpG11 & 800 & 1& 1600 & 43 & 22/43 & 602 & 18/64 \\ ss30 & 132 & 1& 294 & 20 & 23/112 & 17 & 12/63 \\ theta3 & 1106 & 1& 150 & 11 & 15/52 & 61 & 14/48 \\ \hline \end{tabular*} \label{tab:13} \end{table} \begin{table}[htbp] \caption{Selected TOPO examples. CPU time is given in seconds. Iteration count gives the number of the global iterations in Algorithm~\ref{algo:1} and the total number of steps of the Newton method.} \begin{tabular*}{\hsize}{@{\extracolsep{\fill}}crrrrrrr} \hline problem & vars& constr. & constr. & \multicolumn{2}{c}{PENSDP} & \multicolumn{2}{c}{PENLAB} \\ & & & size & CPU & iter. & CPU & iter. \\\hline buck2 & 144 & 2 & 97 & 2 & 23/74 & 22 & 18/184 \\ vibra2 & 144 & 2& 97 & 2 & 34/132 & 35 & 20/304 \\ shmup2 & 200 & 2& 441 & 65 & 24/99 & 172 & 26/179 \\ mater2 & 423 & 94 & 11 & 2 & 20/89 & 70 & 12/179 \\ \hline \end{tabular*} \label{tab:14} \end{table}

https://arxiv.org/abs/2003.11673

Explicit expanders of every degree and size

An $(n,d,\lambda)$-graph is a $d$ regular graph on $n$ vertices in which the absolute value of any nontrivial eigenvalue is at most $\lambda$. For any constant $d \geq 3$, $\epsilon>0$ and all sufficiently large $n$ we show that there is a deterministic poly(n) time algorithm that outputs an $(n,d, \lambda)$-graph (on exactly $n$ vertices) with $\lambda \leq 2 \sqrt{d-1}+\epsilon$. For any $d=p+2$ with $p \equiv 1 \bmod 4$ prime and all sufficiently large $n$, we describe a strongly explicit construction of an $(n,d, \lambda)$-graph (on exactly $n$ vertices) with $\lambda \leq \sqrt {2(d-1)} + \sqrt{d-2} +o(1) (< (1+\sqrt 2) \sqrt {d-1}+o(1))$, with the $o(1)$ term tending to $0$ as $n$ tends to infinity. For every $\epsilon >0$, $d>d_0(\epsilon)$ and $n>n_0(d,\epsilon)$ we present a strongly explicit construction of an $(m,d,\lambda)$-graph with $\lambda < (2+\epsilon) \sqrt d$ and $m=n+o(n)$. All constructions are obtained by starting with known ones of Ramanujan or nearly Ramanujan graphs, modifying or packing them in an appropriate way. The spectral analysis relies on the delocalization of eigenvectors of regular graphs in cycle-free neighborhoods.

\section{Introduction} An $(n,d,\lambda)$-graph is a $d$-regular graph on $n$ vertices in which the absolute value of every nontrivial eigenvalue is at most $\lambda$. This notation was introduced by the author in the early 90s motivated by the fact that such graphs in which $\lambda$ is much smaller than $d$ exhibit strong expansion and quasi-random properties, see \cite{Al1}, \cite{AC}, \cite{KS}. It is well known (see \cite{Al1}, \cite{Ni}, \cite{Fr1}) that if an $(n,d,\lambda)$-graph exists then $\lambda \geq 2\sqrt {d-1} -O(1/\log^2 n)$. An $(n,d,\lambda)$-graph is called (two-sided) {\em Ramanujan} if $\lambda \leq 2\sqrt {d-1}$. Lubotzky, Phillips and Sarnak \cite{LPS}, and independently Margulis \cite{Ma} proved that for every prime $p$ which is $1$ modulo $4$ there are infinite families of $d$-regular Ramanujan graphs. Friedman \cite{Fr2} (see also \cite{Bo} for a simpler proof) proved the existence of near Ramanujan graphs of every degree and every (large) admissible size. Indeed, establishing a conjecture of the present author he proved that a random $d$-regular graph on $n$ vertices is, with high probability, an $(n,d,\lambda)$-graph for $\lambda=2\sqrt{d-1}+o(1)$, where the $o(1)$-term tends to zero as $n$ tends to infinity. For applications, however, (see, e.g., \cite{HLW} and its references for many of those) it is desirable to have explicit constructions of such graphs. It is also sometime desirable to have explicit constructions with specified degrees and number of vertices, (see, for example, \cite{MRSV} for a recent example). A construction is called {\em explicit} if there is a deterministic polynomial time algorithm that, given $n$ and $d$, produces an $(n,d,\lambda)$-graph (or an $(n(1+o(1)), d, \lambda)$-graph). It is {\em strongly explicit } if the adjacency list of any given vertex can be produced in time $polylog(n)$. The construction of \cite{LPS}, and that of \cite{Ma} are strongly explicit \footnote {Though they require to find a large prime in a prescribed range. This can be done efficiently using randomization, but can also be avoided. More details appear in Section 2}, providing Cayley graphs of $SL(2,F_q)$, but work only for degrees that are $p+1$ for primes $p \equiv 1 \bmod 4$ and for numbers of vertices that are of the form $q(q^2-1)/2$ for primes $q$ which are $1$ modulo $4$ so that $p$ is a quadratic residue modulo $p$. Morgenstern \cite{Mo} gave a strongly explicit construction for every degree which is a prime power plus 1, but the possible numbers of vertices obtained are sparser. An observation in \cite{CM} provides strongly explicit families of $(n,d,\lambda)$-graphs with $\lambda \leq O(d^{0.525})$ for infinitely many values of $n$ (but not for every $n$). Similarly, the method in \cite{RVW} and its improvement in \cite{BaTs} provide strongly explicit families with $\lambda \leq O(d^{1/2+o(1)})$ (for infinitely many, but not for all $n$). The results of \cite{MSS} together with those of \cite{Co} and an observation of Srivastava (cf. \cite{MOP}) give explicit, but not strongly explicit $(n,d,\lambda)$-graphs for all admissible $d$ and $n$ with $\lambda \leq 4 \sqrt {d-1}$. In a recent work of Mohanty, O'Donnell and Paredes \cite{MOP} the authors describe an explicit (but not strongly explicit) construction of $(n,d,\lambda)$-graphs for every $d$, where $\lambda = 2\sqrt {d-1}+o(1)$ and the $o(1)$-term tends to $0$ as $n$ tends to infinity. This, again, works for infinitely many values of $n$, but not for all $n$. In the present short paper we describe improved explicit and strongly explicit constructions of near Ramanujan graphs of all degrees and (large) number of vertices. The first result is a (slightly improved version of an) observation I mentioned in several lectures in the 90s that, as far as I know, has never appeared in print. Although it is very simple, the parameters it provides are far better than the ones obtained from the constructions in \cite{CM}, \cite{RVW}, \cite{BaTs}, and I therefore decided to include it here. \begin{prop} \label{p11} For every degree $d$ there is a strongly explicit constructions of $(n,d,\lambda)$-graphs where $\lambda \leq (2+o_d(1))\sqrt {d}$, the $o_d(1)$-term tends to zero as $d$ tends to infinity, and the possible values of $n$ form a sequence in which the ratio between consecutive terms tends to $1$. \end{prop} Note that this means that for every desired number of vertices $n$ and any desired degree $d$, there is a strongly explicit construction of an $(n(1+o_n(1)),d, \lambda)$-graph with $\lambda \leq (2+o_d(1)) \sqrt d$. Here the term $o_n(1)$ tends to zero as $n$ tends to infinity and the $o_d(1)$-term tends to zero as $d$ tends to infinity. The next result provides strongly explicit constructions of $(n,d,\lambda)$ graphs for degrees $d=p+2$ with $p$ being a prime congruent to $1$ modulo $4$, for any desired (large) number of vertices. \begin{theo} \label{t12} For any prime $p \equiv 1 \bmod 4$ and every sufficiently large $n$ there is a strongly explicit construction of an $(n,d,\lambda)$-graph (on exactly $n$ vertices), where $d=p+2$ and $\lambda \leq \sqrt{2(d-1)}+\sqrt{d-1} +o(1) < (1+\sqrt 2) \sqrt {d-1} +o(1)$, and the $o(1)$-term tends to zero as $n$ tends to infinity. \end{theo} It is worth noting that here we allow to have at most one loop in every vertex, with the convention that a loop adds one to the degree (otherwise we must have an even number of vertices as the degree of regularity is odd). For even $n$ we can replace the loops by a matching with no loss in the spectral estimate. If an explicit, rather than strongly explicit construction suffices, we can combine a variant of our method with the new result of \cite{MOP} to get the following. \begin{theo} \label{t13} For every degree $d$, every $\epsilon$ and all sufficiently large $n \geq n_0(d,\epsilon)$, where $nd$ is even, there is an explicit construction of an $(n,d,\lambda)$-graph with $\lambda \leq 2\sqrt{d-1} +\epsilon$. \end{theo} The construction in the proof of Proposition \ref{p11} is a simple packing of known Ramanujan graphs on the same set of vertices. A crucial point is that these constructions are all Cayley graphs of the same group, so one can simply take a union of the corresponding generating sets. The proofs of Theorem \ref{t12} and \ref{t13} require more work. Here too the idea is to start from a known Ramanujan or nearly Ramanujan graph and modify it in an appropriate way. In the proof of Theorem \ref{t12} we add vertices connected to arbitrary disjoint sets of neighbors, adding loops (or a matching) to keep the graph regular. The eigenvalues are then estimated by their variational definition. In the construction for Theorem \ref{t13} we omit carefully chosen vertices from a given near-Ramanujan graph and add a matching between their neighbors to maintain regularity. A crucial point in the spectral analysis here is the delocalization of the eigenvectors of the graphs obtained, which is based on the absence of short cycles in the neighborhoods of the omitted vertices. The rest of this paper is organized as follows. In Section 2 we describe the strongly explicit constructions, including the proofs of Proposition \ref{p11} and Theorem \ref{t12}. In Section 3 we present the proof of Theorem \ref{t13}. The final Section 4 contains some concluding remarks and open problems. \section{Strongly explicit constructions} The basic construction we describe here requires the ability to find efficiently a large prime in a prescribed range. It is well known that this can be done efficiently by a randomized algorithm, and can also be done deterministically assuming some standard (open) number-theoretic conjectures about the gap between consecutive primes. Since this is the only non-deterministic part of the construction, we call it a $p$-strongly explicit construction (where $p$ stands for prime). This construction is described in the first subsection. We then show how it can be replaced by a totally strongly explicit construction. To do so, we first include a subsection presenting the (known) description of the construction of \cite{LPS}, \cite{Ma} as Cayley graphs of Quaternions over $Z_m$. We proceed with a proof of Theorem \ref{t12} with a $p$-strongly explicit construction, followed by its modification to a strongly explicit one. \subsection{The basic construction} \label{basic} We start with the simple proof of Proposition \ref{p11}, with a $p$-strongly explicit construction. It is based on the fact that if $G_i=(V,E_i)$, $i \in I$, are graphs on the same set of vertices $V$, where $G_i$ is an $(n,d_i,\lambda_i)$-graph, then their union $G=(V, \cup_i E_i)$ (considered as a multigraph in case the sets $E_i$ are not pairwise disjoint), is an $(n, \sum_i d_i, \sum_i \lambda_i)$ graph. This is a simple consequence of the variational definition of the eigenvalues. The Ramanujan graphs in \cite{LPS} or \cite{Ma} are Cayley graphs of the group $SL(2,F_q)$ of the two by two matrices with determinant $1$ over the finite field $F_q$, modulo its normal subgroup consisting of the identity $I$ and $-I$. The degree can be $1$ plus any prime $p$ congruent to $1$ modulo $4$, where $q$ is also a prime congruent to $1$ modulo $4$, and $p$ is a quadratic residue modulo $q$. Note that by quadratic reciprocity this is equivalent to $q$ being a quadratic residue modulo $p$. Given a desired degree $d=d_1$, let $p_1 $ be the largest prime congruent to $1$ modulo $4$ and satisfying $p_1+1 \leq d_1$. Put $d_2=d_1-p_1-1$. If $d_2>4$ let $p_2$ be the largest prime congruent to $1$ modulo $4$ which satisfies $p_2+1 \leq d_2$ and put $d_3=d_2-p_2-1$. Continuing in this manner we get primes $p_1, \ldots , p_s$ as above so that $(p_1+1)+(p_2+1)+ \cdots +(p_s+1) \leq d$ where $y=d-((p_1+1)+(p_2+1)+ \cdots +(p_s+1)) \leq 4$. Let $q$ be a prime congruent to $1$ modulo $4$ which is a quadratic residue modulo each $p_i$ (for example, any $q$ which is $1$ modulo each $p_i$ will do). Let $V$ be the set of elements of $SL(2,F_q)$. For each $i$ let $G_i$ be the $(p_i+1)$-regular Ramanujan Cayley graph of $SL(2,F_q)$ described in \cite{LPS}, and let $X_i$ be its (symmetric) set of generators. Let $G'$ be the Cayley graph of $SL(2,F_q)$ whose set of generators consists of the union of all sets $X_i$. Then $G'$ is $(d-y)$-regular, where $0 \leq y \leq 4$. If $y=0$ let $G$ be $G'$. If $y=1$ add to the set of generators the matrix $M$ with rows $(0,1)$ and $(-1,0)$ (which is of order $2$). If $y=2$ add an arbitrary generator and its inverse, if $y=3$ add such a generator, its inverse and $M$, and if $y=4$ add an arbitrary set of two generators and their inverses. In each of these cases the resulting graph $G$ is a $d$-regular Cayley graph of $SL(2,F_q)$. By the known results about the distribution of primes in arithmetic progressions each prime $p_i$ is much smaller than $p_{i-1}$ as long as $p_{i-1}$ is large. In fact, by \cite{BHP} it follows that $p_{i}=O(p_{i-1}^{0.525})$. Therefore, the resulting graph $G$ is an $(n,d,\lambda)$-graph for $n=q(q^2-1/2$ with $\lambda\leq (2+o_d(1))\sqrt d$, where the $o(1)$-term tends to zero as $d$ tends to infinity. Note that it is not difficult to ensure, if so desired, that the graph $G$ is simple: we just have to ensure the chosen primes are distinct. This is automatically the case whenever $d_i$ is still large, and if needed we can stop when $d_i$ becomes small and add arbitrary additional generators and their inverses, together with $M$ if $d$ is odd. Alternatively, if we have to repeat the same prime several times, we can take the corresponding generating set for this prime and conjugate it to get an isomorphic graph with different generators. The known results about the distribution of primes in arithmetic progressions imply also that for each choice of the primes $p_i$ the possible choices for the prime $q$ suffice to ensure that the sequence of possible values for the number of vertices $n$ of the graph is one in which the ratio between consecutive terms tends to $1$ as $n$ tends to infinity. This completes the proof of the proposition (with a $p$-strongly explicit construction resulting from the need to find the required large prime $q$). \hfill $\Box$ \vspace{0.2cm} \subsection{Ramanujan graphs as Cayley graphs of quaternions} \label{quaternions} In this subsection we present the known description of the LPS Ramanujan graphs as Cayley graphs of quaternions. The proof these are Ramanujan graphs appears (somewhat implicitly) in \cite{Lu}. Let $p$ be a prime congruent to $1$ modulo $4$, and let $A=A(p)$ be the set of all integral solutions $(a_0,a_1,a_2,a_3)$ of the equation $a_0^2+a_1^2+a_2^2+a_3^2=p$ where $a_0$ is positive odd, and all other $a_i$ are even. By a well known result of Jacobi there are exactly $p+1$ such vectors. Let $m$ be odd, relatively prime to $p$, and assume further that $p$ is a square in $Z_m^*$. let $Q(m)$ be the factor group of the multiplicative group of the quaternions over $Z_m$ whose norm is a square in $Z_m^*$, modulo its normal subgroup consisting of the scalars $Z_m^*$. Thus the elements of $Q(m)$ are all quaternions $x_0+x_1 i+x_2 j+x_3 k$ where $x_0^2+x_1^2+x_2^2+x_3^2 \in (Z_m^*)^2$ and two such elements are identified if one is a multiple of the other by a scalar. Finally, let $H=H(p,m)$ be the Cayley graph of $Q(m)$ with the generating set $$\{a_0+a_1i+a_2j+a_3k: (a_0,a_1,a_2,a_3) \in A(p)\}.$$ The following result is proved (somewhat implicitly) in \cite{Lu}, see pages 95-97. \begin{theo}[\cite{Lu}] \label{t91} For every $p$ and $m$ as above $H=H(p,m)$ is a non-bipartite $(p+1)$-regular Ramanujan graph, that is, the absolute value of each of its eigenvalues besides the top one is at most $2 \sqrt p$. \end{theo} \subsection{The proof of Proposition \ref{p11}} In the construction here we will start with the graphs $Q(p,m)$ with $p \equiv 1 \bmod 4$ a prime and $m=q_1^s q_2^t$, where $s,t \geq 1$ and $q_1,q_2 $ are distinct primes, each being $1 \bmod 4p$. For each fixed $p$ as above, the known results about the Linnik problem (see \cite{HB}) imply that there are $q_1,q_2$ as above, each being at most a polynomial in $p$. It is not difficult to check, using Hensel's Lemma and the Chinese Remainder Theorem, that the number of vertices of $H(p,q_1^s q_2^t)$ is $$ Q(q_1,q_2,s,t)=q_1^{3(s-1)} q_2^{3(t-1)} \frac{q_1(q_1-1)(q_1+1)}{2} \frac{q_2(q_2-1)(q_2+1)}{2}. $$ Indeed, by Hensel's Lemma, for elements $x_0,x_1,x_2,x_3$ of $Z_m$ the norm $x_0^2+x_1^2+x_2^2+x_3^2$ is a square in $Z^*_m$ if and only if it is a square in $Z^*_{q_1}$ and in $Z^*_{q_2}$. Since each $q_i$ is $1 \bmod 4$, $-1$ is a quadratic residue implying that the number of solutions of $y_1^2+y_2^2=0$ in $Z_{q_i}$ is $2q_i-1$. For each nonzero $b$ in $Z_{q_i}$ the number of solutions of $y_1^2+y_2^2=b$ (in $Z_{q_i}$) is the same as the number of solutions of $y^2-z^2(=(y-z)(y+z))=b$ , which is $q_i-1$. This shows that the number of solutions of $x_0^2+x_1^2+x_2^2+x_3^2=b$ for any nonzero $b \in Z_{q_i}$ is $$ 2 (2q_i-1)(q_i-1)+(q_i-2)(q_i-1)^2=(q_i-1)q_i(q_i+1). $$ (These include $(2q_i-1)(q_i-1)$ solutions with $x_0^2+x_1^2=0$ and $x_2^2+x_3^2=b$, $(2q_i-1)(q_i-1)$ ones with $x_0^2+x_1^2=b$ and $x_2^2+x_3^2=0$, and $(q_i-1)^2$ solutions for each of the $q_i-2$ possibilities $x_0^2+x_1^2=b_1$ and $x_2^2+x_3^2=b_2$ with $b_1+b_2=b$ and $b_1,b_2 \not \in \{0,b\}$.) Therefore, the number of elements over $Z_{q_i}$ whose norm is a nonzero square in $Z_{q_i}$ is $$ \frac{q_i-1}{2} (q_i-1)q_i(q_i+1)=\frac{q_i(q_i-1)^2(q_i+1)}{2}. $$ By the Chinese remainder Theorem there are $$ \frac{q_1(q_1-1)^2(q_1+1)}{2} \frac{q_2(q_2-1)^2(q_2+1)}{2}. $$ elements $(x_0,x_1,x_2,x_3)$ in $Z_{q_1q_2}$ so that $x_0^2+x_1^2+x_2^2+x_3^2$ is a square in $Z^*_{q_1q_2}$, and by Hensel's Lemma each of them provides $q_1^{4(s-1)}q_2^{4(t-1)}$ quaternions over $Z_{q_1^sq_2^t}$ whose norm is a square in $Z^*_{q_1^sq_2^t}$. To get the number of vertices of the graph we just have to divide by the cardinality of $Z^*_{q_1^sq_2^t}$ which is $q_1^{s-1}q_2^{t-1}(q_1-1)(q_2-1)$, obtaining the required number of vertices. Note also that by this description it is easy to number the vertices of the graph. (For our application here it is in fact enough to number a constant fraction of them. For fixed $q_1,q_2$ this can be done, for example, by numbering all vectors $(1,x_1,x_2,x_3) \in Z_{q_1^sq_2^t}$ with $x_1,x_2,x_3$ divisible by $q_1q_2$, lexicographically). We next show that for every fixed distinct primes $q_1,q_2$, the ratio between consecutive elements in the set of integers $\{Q(q_1,q_2,s,t): s,t \geq 1 \}$ tends to $1$ as the elements grow. \begin{lemma} \label{l32} Let $q_1,q_2$ be distinct primes. Then for every large integer $n$ there are positive integers $s,t$ so that $n \leq Q(q_1,q_2,s,t) \leq n+o(n)$. \end{lemma} \noindent {\bf Proof:\,} The constant $\alpha=\frac{\log {q_1}}{\log {q_2}}$ is irrational. Therefore, by the equidistribtion theorem (in fact, by a special case that follows easily from the pigeonhole principle), for every $\delta>0$ there is an integer $k_1=k_1(\alpha)$ so that $0<k_1 \alpha \bmod 1 < \delta$. It follows that for every $\mu>0$ there are integers $k_1,k_2$ such that $$ 1 \leq \frac{q_1^{k_1}}{q_2^{k_2}} \leq q_2^{\delta} \leq (1+\mu). $$ This implies that for every $s,t \geq \max\{k_1,k_2\}$ the ratio between $Q(q_1,q_2,s,t)$ and $Q(q_1,q_2,s-k_1,t-k_2)$ is between $1$ and $(1+\mu)^3$, implying the desired result. \hfill $\Box$ \vspace{0.2cm} \noindent The proof of Proposition \ref{p11} now proceeds exactly as in subsection \ref{basic}, using the description of the LPS graphs serving as the building blocks as given in subsection \ref{quaternions}. Since here $q_1,q_2$ are constants, there is no need to find any large primes for the construction, providing a strongly explicit construction for every fixed degree. \subsection{The proof of Theorem \ref{t12}} We first describe a $p$-strongly-explicit construction, starting, again, with the graphs of \cite{LPS}. Recall that the vertex sets of these graphs is the set of matrices in $SL(2,F_q)$ where each matrix $A$ is identified with $-A$. It is easy to number the vertices starting with the matrices $(a_{ij})$ with $a_{11} \neq 0$ and ordering them according to the lexicographic order of the elements $(a_{11},a_{12},a_{13})$ where $a_{14}$ is chosen to ensure that the determinant is $1$ (which is always possible as $a_{11} \neq 0$). Here $1 \leq a_{11} \leq (q-1)/2$, as we identify each matrix $A$ with $-A$. The first matrices are the $q^2$ matrices with $a_{11}=1$, then those with $a_{11}=2$, and so on. (The remaining $q(q-1)/2$ matrices with $a_{11}=0$ can appear last in our order according to the lexicographic order of $(a_{12},a_{24})$, but this will play no real role in the construction.) Given the desired number $n$ of vertices, and given the degree $d=p+2$ with $p$ as in the theorem, let $q$ be the largest prime which is $1$ modulo $4$, is a quadratic residue modulo $p$ and satisfies $|SL(2,F_q)|=m_q=q(q^2-1)/2 \leq n$. Put $m=m_q$, let $H$ be the Ramanujan $(p+1)=(d-1)$-regular graph of \cite{LPS} whose vertex set $V$ is the set of elements of $SL(2,F_q)$ numbered as described above. By the known results about the distribution of primes in progressions $n-m=o(m)$. Put $r=n-m$ and let $R$ be a set of $r$ additional vertices $u_1,u_2, \ldots, u_r$. Connect each vertex $u_i$ to the vertices numbered $(i-1)d+1,(i-1)d+2, \ldots ,id$ of $H$. Finally add a loop to each remaining vertex of $H$ to make the graph regular. This is the desired graph $G$. It is clearly $d=p+2$-regular. (If $n$ is even and we do not want loops we can replace them by a matching between consecutive pairs of vertices, saturating all non-neighbors of the $r$ new vertices). It is clear that the construction above is strongly explicit. To complete the proof it remains to show that the absolute value of any nontrivial eigenvalue of $G$ is at most $\sqrt{2(p+1)}+ \sqrt{p}+o(1)$. We proceed with a proof of this fact. By the variational definition of the nontrivial eigenvalues of $G$ this is equivalent to showing that for every real function $f(u)$ on the set of vertices $U=V \cup R$ of $G$ satisfying $\|f \|_2^2=1$ and $\sum_{u \in U} f(u)=0$ \begin{equation} \label{e31} |f^t A_G f| \leq \sqrt{2(p+1)}+ \sqrt{p}+o(1) \end{equation} where $A_G$ is the adjacency matrix of $G$. Let $W \subset V$ denote the set of all $(p+2)r$ neighbors of $R$, put $L=V-W$, and let $E_R$ denote the set of all edges between $R$ and $W$. Thus $E_R$ is a collection of pairwise vertex disjoint stars, each having $(p+2)$ leaves. The adjacency matrix of $G$ can be written as a sum $A_G=A_H+A_R+A_L$, where $A_H$ is the adjacency matrix of the Ramanujan graph $H$ (with the additional isolated vertices of $R$), $A_R$ is the adjacency matrix of the graph $(U,E_R)$, and $A_L$ is the adjacency matrix of the graph on $U$ whose edges are the loops on the vertices of $L$ (or the added matching on them, if we have chosen not to add loops). Therefore \begin{equation} \label{e32} f^t A_G f = f^t A_H f+f^t A_R f+f^t A_L f. \end{equation} We proceed to bound each of these terms. By Cauchy-Schwarz $$ |\sum_{u \in R} f(u)|^2 \leq |R| \sum_{u \in R} f^2(u) \leq |R|=o(m). $$ Since $\sum_{u \in U} f(u)=0$, this implies that $|\sum_{u \in V} f(u)|^2= |\sum_{u \in R} f(u)|^2 =o(m)$ Let $g$ be the trivial normalized eigenvector of $H$, that is, the vector given by $g(v)=1/\sqrt m$ for all $v \in V$. Expressing the restriction $f'$ of $f$ to $V$ as a linear combination of $g$ and a unit vector $h$ orthogonal to it, we get $f'=bg+ch$, where $\sum_{v \in V} h(v)=0$, $b^2+c^2=1$ and $b^2=|\sum_{u \in V} f(u)|^2/m=o(1)$. Since $H$ is a Ramanujan graph, $|h^t A_H h| \leq 2 \sqrt p.$ Therefore \begin{equation} \label{e33} |f^t A_H f| = |(f')^t A_H f'| \leq b^2 (p+1) +c^2 2 \sqrt p \leq 2 \sqrt p +o(1). \end{equation} Clearly \begin{equation} \label{e34} |f^t A_L f| \leq \sum_{v \in L} f^2(v). \end{equation} Indeed this is an equality if there are loops and an inequality in case a matching has been added. For bounding the absolute value of $f^t A_R f$ observe that for every positive $x$ $$ |f^t A_R f| =2 |\sum_{uv \in E_R} f(u) f(v) | $$ \begin{equation} \label{e35} \leq \sum_{u \in R, v \in W, uv \in E_R} (\frac{f^2(u)}{x} + x f^2(v)) =\frac{p+2}{x} \sum_{u \in R} f^2 (u) + x \sum_{v \in w} f^2 (v). \end{equation} Combining (\ref{e32}),(\ref{e33}),(\ref{e34}) and (\ref{e35}) we conclude that for every positive real $x$ \begin{equation} \label{e36} |f^t A_G f| \leq (2\sqrt p+1) \sum_{v \in L} f^2 (v) + (2 \sqrt p+x)\sum_{v \in W} f^2 (v) + \frac{p+2}{x} \sum_{v \in R} f^2(v) + o(1). \end{equation} Choosing $x=\sqrt{2p+2}-\sqrt{p}$ (which is at least $1$) and substituting in (\ref{e36}) we finally get $$ |f^t A_G f| \leq (\sqrt{2p+2}+\sqrt p) \sum_{u \in U} f^2 (u)+o(1) =(\sqrt{2p+2}+\sqrt p) +o(1). $$ This establishes (\ref{e31}) and completes the proof (with a $p$-strongly explicit construction). The conversion to a strongly explicit construction proceeds just as in the proof of Proposition \ref{p11}, based on the results in subsection \ref{quaternions}. Note that as mentioned in that subsection the description there provides a simple efficient way to number enough vertices of each graph $H(p,q_1^sq_2^t)$ and by Lemma \ref{l32} we can start by finding efficiently appropriate $s,t$ using binary search. \hfill $\Box$ \section{Explicit constructions} In this section we present the proof of Theorem \ref{t13}. We start with some preliminary lemmas. \begin{lemma} \label{l31} Let $G=(V,E)$ be a $d$-regular graph on $n$ vertices, where $d \geq 3$, and suppose that the $2r+4$-neighborhood of any vertex in it contains at most one cycle, where $r \leq \log_{d-1} n$. Then there is a subset $U \subset V$ of vertices of $G$ satisfying the following. \begin{enumerate} \item $|U| \geq \frac{n}{2d^{2r+3}}$ \item The $(r+1)$-neighborhood of any vertex in $U$ contains no cycle. \item The distance between any two vertices in $U$ is at least $2r+3$. \end{enumerate} Such a set $U$ can be found in polynomial time. \end{lemma} \vspace{0.1cm} \noindent {\bf Proof:}\, Let ${\cal C}$ denote the collection of all cycles of length at most $2r+4$ in $G$. Note that the distance between any two members $C_1,C_2$ of ${\cal C}$ is larger than $2r+4$, since otherwise there is a vertex $v$ within distance at most $r+2$ of both cycles $C_i$, and then its $2r+4$-neighborhood contains both cycles, contradiction. The $r+2$ neighborhood of each cycle $C \in {\cal C}$ contains no other cycle besides $C$, as it is contained in the $2r+4$ neighborhood of any vertex on the cycle. Thus the number of edges spanned by each such neighborhood is at most the number of vertices in it. As the neighborhoods are vertex disjoint, the total number of edges in all these neighborhoods together is at most the number of vertices of $G$ which is $n$. It follows that by omitting all vertices in the $(r+1)$-neighborhoods of all members of ${\cal C}$, at most $n$ edges are omitted, and as $G$ has $nd/2$ edges and $d \geq 3$ at least $n/2 $ edges, and hence at least $n/2d$ vertices have not been omitted. Let $Z$ be the set of non-omitted vertices. Note that the $(r+1)$-neighborhood of any vertex in $Z$ contains no cycle (as if it contains a cycle, it contains a cycle of length at most $2r+3<2r+4$ but the vertex is not within distance $r+1$ of any such cycle.) Starting with $U=\emptyset$ let $v_1$ be an arbitrary vertex of $Z$, add it to $U$ and remove all vertices of $Z$ within distance $2r+2$ of $v_1$. Clearly at most $d^{2r+2}$ vertices have been deleted. Let $v_2$ be an arbitrary vertex left in $Z$, add it to $U$ and remove all vertices of $U$ within distance $2r+2$ of $v_2$. Continuing in this manner we get a set $U$ of at least $\frac{n}{2d^{2r+3}}$ vertices. It is clear that this set satisfies all the conclusions of the lemma. It is also clear that $U$ can be computed in polynomial time. \hfill $\Box$ The next lemma about the delocalization of eigenvectors of regular graphs in cycle-free neighborhood can be proved using the method of Kahale in \cite{Ka} (see also \cite{AGS} for a recent application of this technique). Here we present a simple self contained proof. \begin{lemma} \label{l42} Let $G=(V,E)$ be a $d$-regular graph where $d \geq 3$, let $uv$ be an edge of $G$ and suppose that the $r$-neighborhood of $uv$ contains no cycle. For each $i$ satisfying $0 \leq i \leq r$ let $N_i$ denote the set of all vertices of distance exactly $i$ from $\{u,v\}$. (In particular, $N_0=\{u,v\}$). Let $f$ be a nonzero eigenvector of $G$ with eigenvalue $\mu \geq 2 \sqrt{d-1}$. Then for every $1 \leq i \leq r$ \begin{equation} \label{e41} \sum_{w \in N_i} f^2(w) \geq \sum_{w \in N_{i-1}} f^2 (w). \end{equation} \end{lemma} \vspace{0.1cm} \noindent {\bf Proof:}\, We apply induction on $i$. Note that by assumption the induced subgraph of $G$ on the $r$-neighborhood of $uv$ is a $d$-regular tree. Therefore $|N_i|=2(d-1)^i$ for all $i \leq r$. Let $u_1,u_2, \ldots ,u_{d-1}$ denote the neighbors of $u$ besides $v$, and let $v_1,v_2, \ldots ,v_{d-1}$ denote the neighbors of $v$ besides $u$. Then $$ f(v)+\sum_{i=1}^{d-1} f(u_i)=\mu f(u) $$ and $$ f(u)+\sum_{i=1}^{d-1} f(v_i)=\mu f(v). $$ By Cauchy-Schwarz, $$ f^2(v)+\sum_{i=1}^{d-1} f^2(u_i) \geq \frac{\mu^2 f^2(u)}{d} \geq \frac{4d-4}{d} f^2 (u) $$ and similarly $$ f^2(u)+\sum_{i=1}^{d-1} f^2(v_i) \geq \frac{4d-4}{d} f^2 (v). $$ Summing, we conclude that $$ f^2(u)+f^2(v)+\sum_{w \in N_1}f^2(w) \geq \frac{4d-4}{d} (f^2(u)+f^2 (v)), $$ implying that $$ \sum_{w \in N_1}f^2(w) \geq \frac{3d-4}{d} (f^2(u)+f^2 (v)) \geq f^2(u)+f^2(v) =\sum_{w \in N_0} f^2(w). $$ This proves (\ref{e41}) for $i=1$. Assuming it holds for $i-1$ we prove it for $i$. For each vertex $w \in N_{i-1}$ let $w'$ be its unique parent in $N_{i-2}$ and let $x_1,x_2, \ldots ,x_{d-1}$ be its neighbors in $N_i$. Then $$ f(w')+\sum_{i=1}^{d-1}f(x_i) = \mu f(w). $$ Since $f(w')=(d-1) \cdot \frac{f(w')}{d-1}$, we get, by Cauchy-Schwarz, $$ \frac{f^2(w')}{d-1}+\sum_{i=1}^{d-1}f^2(x_i) \geq \frac{\mu^2 f^2(w)}{2d-2} \geq 2 f^2(w). $$ Summing the above inequality for all $w$ in $N_{i-1}$, each vertex $w'$ appears in the LHS exactly $d-1$ times, yielding $$ \sum_{w' \in N_{i-2}} f^2(w') + \sum_{x \in N_{i}} f^2(x) \geq 2 \sum_{w \in N_{i-1}} f^2 (w). $$ This gives $$ \sum_{x \in N_{i}} f^2(x) \geq 2 \sum_{w \in N_{i-1}} f^2 (w) - \sum_{w' \in N_{i-2}} f^2(w') \geq \sum_{w \in N_{i-1}} f^2 (w), $$ where the last inequality follows from the induction hypothesis. This completes the proof of the induction step, establishing the assertion of the lemma. \hfill $\Box$ Finally, we need the main result of Mohanty, O'Donnell and Paredes in \cite{MOP}, which is the following. \begin{theo}[\cite{MOP}] \label{t33} For every $d$, $\epsilon>0$ and (large) $n$ there is an explicit construction of an $(n+o(n),d,\lambda)$-graph with $\lambda \leq 2 \sqrt{d-1}+\epsilon$ so that the $s$ neighborhood of any vertex contains at most one cycle, where $s \geq (\log \log n)^2 $. \end{theo} We note that the result is stated in \cite{MOP} without the conclusion about the cycles, but the version above follows from the proof as presented there. We are now ready to prove Theorem \ref{t13}. \vspace{0.2cm} \noindent {\bf Proof of Theorem \ref{t13}:}\, Put $r=\lceil 2/\epsilon \rceil$ and $s=2r+4$. Let $H=(V,E)$ be an explicit $(n+u,d,\lambda)$-graph with $u=o(n)$ and $\lambda \leq 2\sqrt{d-1}+\epsilon/2$ in which the $s$-neighborhood of any vertex contains at most one cycle. Such an $H$ exists by Theorem \ref{t33}. By Lemma \ref{l31} we can find efficiently a set $U$ of $u$ vertices in $H$ satisfying the assertion of the lemma. Omit these vertices from the graph to get a graph $H'$ and add a matching $M$ between their neighbors retaining the degree of regularity $d$. Let $G$ denote the resulting graph. Clearly it is $d$-regular and has $n$ vertices. Note that the $r$-neighborhood of any edge $uv$ of the added matching $M$ contains no cycle. In order to complete the proof it remains to show that every nontrivial eigenvalue of $G$ has absolute value at most $2 \sqrt {d-1}+\epsilon$. Let $A_G$ be the adjacency matrix of $G$, $A_{H'}$ the adjacency matrix of $H'$ (on the set of vertices $V$) and $A_M$ the adjacency matrix of the graph on the set of vertices $V $ whose edges are those of the matching $M$. Note that $A_G=A_{H'}+A_M$. Let $\lambda$ be a nontrivial eigenvalue of $G$ and let $f$ be a corresponding eigenvector satisfying $\sum_{v \in V} f^2(v)=1$. Then \begin{equation} \label{e91} \lambda=f^t A_G f=f^t A_{H'} f+ f^t A_M f. \end{equation} Since $H'$ is an induced subgraph of $H$ and all nontrivial eigenvalues of $H$ have absolute value at most $2 \sqrt{d-1}+\epsilon/2$ it follows, by eigenvalue interlacing, that \begin{equation} \label{e92} |f^t A_{H'} f| \leq 2 \sqrt{d-1}+\epsilon/2. \end{equation} It is also clear that \begin{equation} \label{e93} |f^t A_{M} f| = |2\sum_{uv \in M} f(u)f(v)| \leq \sum_{uv \in M} f^2(u)+f^2(v). \end{equation} If $|\lambda| \leq 2 \sqrt{d-1}$ there is nothing to prove, we thus assume that $\lambda \geq 2 \sqrt{d-1}$. Since the $r$-neighborhood of any edge of $M$ contains no cycle, Lemma \ref{l32} implies that for every such edge $uv$, $$ f^2(u)+f^2(v) \leq \frac{1}{r} \sum_{w \in N(u,v,r)} f^2(w), $$ where $N(u,v,r)$ denotes the $r$-neighborhood of $uv$. Since all these neighborhoods are pairwise disjoint it follows that \begin{equation} \label{e94} \sum_{uv \in M} f^2(u)+f^2(v) \leq \frac{1}{r} \sum_{w \in V} f^2 (v) \leq \frac{\epsilon}{2}. \end{equation} The desired result follows by plugging (\ref{e92}) and (\ref{e94}) in (\ref{e91}) (using (\ref{e93})). \hfill $\Box$ \section{Concluding remarks} \begin{itemize} \item Morgenstern \cite{Mo} gave a strongly explicit construction of Ramanujan graphs for every degree which is a prime power plus $1$, but we cannot apply his construction in the proof of Proposition \ref{p11} since his construction provides Cayley graphs of different groups (and different sizes) for different degrees and hence one cannot pack the graphs corresponding to several degrees. Similarly, we cannot use his construction in the proof of Theorem \ref{t12} since for every fixed degree the sequence of possible numbers of vertices in his construction for this degree is too sparse. \item The proof of Theorem \ref{t13} can be applied directly to high girth Ramanujan graphs like those in \cite{LPS}, \cite{Ma} in case the required degree is $p+1$ for a prime $p$ congruent to $1 \bmod 4$ to obtain near Ramanujan graphs of this degree with any required (large) number of vertices. \item The problem of obtaining strongly explicit (two-sided) Ramanujan (and not nearly Ramanujan) graphs for any degree and number of vertices remains open. \end{itemize} \noindent {\bf Acknowledgment} I thank L\'aszl\'o Babai, Oded Goldreich, Ryan O'Donnell, Ori Parzanchevski and Peter Sarnak for helpful discussions.

https://arxiv.org/abs/2009.09410

Tiling by translates of a function: results and open problems

We say that a function $f \in L^1(\mathbb{R})$ tiles at level $w$ by a discrete translation set $\Lambda \subset \mathbb{R}$, if we have $\sum_{\lambda \in \Lambda} f(x-\lambda)=w$ a.e. In this paper we survey the main results, and prove several new ones, on the structure of tilings of $\mathbb{R}$ by translates of a function. The phenomena discussed include tilings of bounded and of unbounded density, uniform distribution of the translates, periodic and non-periodic tilings, and tilings at level zero. Fourier analysis plays an important role in the proofs. Some open problems are also given.

\section{Introduction} \label{secI1} Let $f$ be a function in $L^1(\R)$ and let $\Lam \sbt \R$ be a discrete set. We say that \emph{$f$ tiles $\R$ at level $w$} with the translation set $\Lam$, or that \emph{$f+\Lam$ is a tiling of $\R$ at level $w$} (where $w$ is a constant), if we have \begin{equation} \label{eqI1.1} \sum_{\lambda\in\Lambda}f(x-\lambda)=w\quad\text{a.e.} \end{equation} and the series in \eqref{eqI1.1} converges absolutely a.e. In the same way one can define tiling by translates of an $L^1$ function on $\R^d$, or more generally, on any locally compact abelian group. The finite abelian groups, and in particular the cyclic ones, are an important class being often considered. If $f = \1_\Omega$ is the indicator function of a set $\Omega$, and $f + \Lam$ is a tiling at level one, then this means that the translated copies $\Omega+\lam$, $\lam\in\Lam$, fill the whole space without overlaps up to measure zero. To the contrary, for tilings by a general real or complex-valued function $f$, the translated copies may have overlapping supports and a wider variety of phenomena may (and do) occur. In dimension one, translational tilings exhibit a stronger structure than in higher dimensions, and there are interesting questions as to how rigid this structure must be, e.g.\ how close the translation set $\Lam$ is to being periodic, or to being constructed out of periodic sets, or, at an even more basic level, to being uniformly distributed in $\R$. The subject has been studied by several authors, see, in particular, \cite{LM91}, \cite{KL96}, \cite{Kol04}, \cite{KL16}, \cite{Liu18}, \cite{Lev20}. The aim of this paper is to survey the results obtained in earlier works, as well as to prove several new results. At the end of the paper, some open problems are also given. In this section we survey the main results on the structure of tilings by translates of a function on $\R$, and we state the new results that will be proved in this paper. \subsection{Tiling and density} We say that a set $\Lambda \sbt \R$ has \emph{bounded density} if \begin{equation} \label{eqI10.1} \sup_{x\in\mathbb R} \#(\Lambda\cap[x,x+1))<+\infty. \end{equation} The set $\Lambda$ is said to be \emph{uniformly distributed} if there is a number $D(\Lam)$ satisfying \begin{equation} \label{eqI10.2} \#(\Lambda\cap[x,x+r)) = D(\Lam) \cdot r + o(r), \quad r \to +\infty \end{equation} uniformly with respect to $x \in \R$. In this case, $D(\Lam)$ is called the \emph{uniform density} of $\Lam$. The following result establishes a connection between tiling and density: \begin{thm}[{\cite{KL96}}] \label{thmKL1} Let $f+\Lam$ be a tiling at some nonzero level $w$, where $f \in L^1(\R)$ and where $\Lam \sbt \R$ is a set of bounded density. Then $f$ has nonzero integral, and $\Lam$ has a uniform density given by $D(\Lam) = w \cdot (\int f)^{-1}$. \end{thm} This was proved in \cite[Lemma 2.3]{KL96} for a weaker notion of density\footnote{In \cite{KL96} the density $d(\Lam)$ of a set $\Lam \sbt \R$ is defined by $d(\Lam) := \lim_{r \to +\infty} \#(\Lambda\cap(-r,r)) / (2r)$.}, but a minor adjustment to the proof in fact yields the stronger statement above for the uniform density. A similar result is true also in $\R^d$. \subsection{Tiling at level zero} \label{secTLZ} It is not known whether \thmref{thmKL1} has an analog for tilings of bounded density \emph{at level zero}. It was conjectured in \cite[p.\ 660]{KL96} that if $f+\Lam$ is such a tiling, then $f$ must have zero integral. In \cite[Lemma 2.4]{KL96} this was proved under the extra assumption that $f$ has compact support. Here we will prove that the conclusion is true in the general case: \begin{thm} \label{thmA5} Let $f+\Lam$ be a tiling at level zero, where $f \in L^1(\R)$ and where $\Lam \sbt \R$ is a nonempty set of bounded density. Then $f$ must have zero integral. \end{thm} Do there exist tilings $f+\Lam$ at level zero such that the set $\Lam$ has density zero? \thmref{thmKL1} does not exclude such a possibility. In fact, in dimensions two and higher it is easy to exhibit tilings of this kind. For instance, in $\R^2$ one may take $f(x,y) = \varphi(x)\psi(y)$, $\Lam = \Gam \times \{0\}$, where $\varphi, \psi \in L^1(\R)$ and $\varphi + \Gam$ is a tiling of $\R$ at level zero. We will show, however, that this is \emph{not} the case in dimension one. To state the result, we recall that a set $\Lam \sbt \R$ is said to be \emph{relatively dense} if there is $r>0$ such that any interval $[x,x+r)$ contains at least one point from $\Lam$. We will prove the following: \begin{thm} \label{thmA6} Let $f+\Lam$ be a tiling at level zero, where $f \in L^1(\R)$ is nonzero and $\Lam \sbt \R$ is a nonempty set of bounded density. Then $\Lam$ must be a relatively dense set. \end{thm} In particular this implies that $\Lam$ cannot have density zero. Does it follow from the assumptions in \thmref{thmA6} that $\Lam$ has a uniform (positive) density $D(\Lam)$\,? The answer to this question is not known. The problem is nontrivial due to the existence of translation sets $\Lam$ that admit only tilings at level zero, so that \thmref{thmKL1} does not apply to these sets: \begin{thm}[{\cite{Lev20}}] \label{thmLev20.5} There exists a nonempty set $\Lam \sbt \R$ of bounded density which admits tilings $f+\Lam$ with nonzero $f \in L^1(\R)$, but any such a tiling is necessarily a tiling at level zero. \end{thm} \subsection{Periodic tilings} We say that a set $\Lam \sbt \R$ has a \emph{periodic structure} if it can be represented as a disjoint union of finitely many arithmetic progressions, namely \begin{equation} \label{eqI2.1} \Lam = \biguplus_{j=1}^{N} (a_j \Z + b_j) \end{equation} where $a_j, b_j$ are real numbers and $a_j>0$. The sets $\Lam$ with this structure constitute the basic examples of translation sets for tilings of $\R$. Indeed, one can check that if \begin{equation} f = \1_{[0,a_1]} \ast \1_{[0,a_2]} \ast \dots \ast \1_{[0,a_N]} \end{equation} and $\Lam$ is given by \eqref{eqI2.1}, then $f + \Lam$ is a tiling at some positive level $w$. The last example shows that any set $\Lam$ of the form \eqref{eqI2.1} admits a tiling by a \emph{compactly supported} function $f \in L^1(\R)$. A result first proved in \cite{LM91} and rediscovered in \cite{KL96} asserts that there are no other translation sets $\Lam$ with this property: \begin{thm}[{\cite{LM91}, \cite{KL96}}] \label{thmLM91} Let $f \in L^1(\R)$ be nonzero and have compact support. If $f$ tiles at some level $w$ with a translation set $\Lambda$ of bounded density, then $\Lambda$ has a periodic structure, namely, it must be of the form \eqref{eqI2.1}. \end{thm} The proof of this result is based on Cohen's characterization of idempotent measures in locally compact abelian groups. The group on which Cohen's theorem is used in the proof is the \emph{Bohr compactification} of the real line (see also \cite[p.\ 25]{Mey70}). A discrete set $\Lambda\subset\mathbb R$ is said to have \emph{finite local complexity} if $\Lambda$ can be enumerated as a sequence $\{\lambda_n\}$, $n\in\Z$, such that $\lambda_n<\lambda_{n+1}$ and the successive differences $\lambda_{n+1}-\lambda_n$ take only finitely many different values. The following result establishes that tilings of finite local complexity must be periodic, even if $f$ does not have compact support: \begin{thm}[{\cite{IK13}, \cite{KL16}}] \label{thmKL16FLC} Let $\Lambda \sbt \R$ have finite local complexity. If $f \in L^1(\R)$ is nonzero and $f+\Lambda$ is a tiling at some level $w$, then $\Lambda$ must be a periodic set, namely, it has the form $\Lam = a\Z + \{b_1, \dots, b_N\}$. \end{thm} \subsection{Non-periodic tilings} The papers \cite{LM91}, \cite{KL96} leave the following question open: Does there exist any set $\Lambda \sbt \R$ \emph{not} of periodic structure, which can tile with some function $f\in L^1(\R)$ of unbounded support? Such a set $\Lam$ cannot have finite local complexity by \thmref{thmKL16FLC}. We settled this question affirmatively in \cite{KL16}: \begin{thm}[{\cite{KL16}}] \label{thmKL16} There exists a tiling $f+\Lambda$ at level one, where $f\in L^1(\R)$ and $\Lambda \sbt \R$ has bounded density, but such that $\Lam$ has no periodic structure, i.e.\ the set $\Lam$ is not of the form \eqref{eqI2.1}. \end{thm} The proof was based on the implicit function method due to Kargaev \cite{Kar82}, and it yields a set $\Lam$ which is a small perturbation of the integers. The proof moreover allows to choose the function $f$ in the Schwartz class. However it yields a function $f$ satisfying $\operatorname{supp}(f) = \R$, where $\operatorname{supp}(f)$ is the closed support of $f$ (the smallest closed set such that $f$ vanishes a.e.\ on its complement). In this paper we will prove a stronger version of Theorem \ref{thmKL16}, which establishes the existence of non-periodic tilings $f+\Lam$ of bounded density, such that $f$ has ``sparse'' support. Precisely, we will show that the support (which must be unbounded, due to \thmref{thmLM91}) can be localized inside any given set $\Om \sbt \R$ which contains arbitrarily long intervals: \begin{thm} \label{thmA1} There is a discrete set $\Lambda \sbt \R$ of bounded density, but which is not of the form \eqref{eqI2.1}, with the following property: given any scalar $w$, and any set $\Om \sbt \R$ which contains arbitrarily long intervals, one can find a nonzero $f \in L^1(\mathbb R)$, $\operatorname{supp}(f) \sbt \Om$, such that $f+\Lambda$ is a tiling at level $w$. \end{thm} We note that the set $\Om$ in this theorem may be chosen very sparse, for example, one may take $\Om = \bigcup_{j=1}^{\infty} [\tau_j, \tau_j + j]$, where the numbers $\tau_j$ grow arbitrarily fast. \thmref{thmA1} implies in particular the existence of non-periodic tilings by a function $f$ such that $\operatorname{supp}(f) \sbt [0, +\infty)$, i.e.\ the support is \emph{bounded from below} (while it cannot be bounded from both above and below, due to \thmref{thmLM91}). The proof of \thmref{thmA1} relies on a recent result due to Kurasov and Sarnak \cite{KS20}, who constructed a new type of \emph{crystalline measures} on $\R$. These are pure point measures with discrete closed support, whose Fourier transform is a measure of the same type. \subsection{Tilings of unbounded density} \label{secTUB} It seems that very little attention has been paid to tilings which are \emph{not} of bounded density. In fact, we are not aware of any example of such a tiling in the literature. In \cite[Example 7.1]{KL96} some examples are given in a more general context, where the points of the translation set $\Lam$ are endowed with nonnegative integer multiplicities (so that $\Lam$ is actually not a set, but a multi-set). We will prove that there exist tilings of unbounded density in the proper sense. Let us say that a set $\Lam \sbt \R$ has \emph{tempered growth} if there is $N$ such that \begin{equation} \label{eqA3.16} \# (\Lam \cap (-r,r)) = O(r^N), \quad r \to +\infty. \end{equation} \begin{thm} \label{thmA3} There is a set $\Lambda \sbt \R$ which is not of bounded density but has tempered growth, with the following property: given any scalar $w$ one can find a nonzero function $f$ in the Schwartz class, such that $f+\Lambda$ is a tiling at level $w$. \end{thm} Moreover, in our example the set $\Lam$ is contained in a small neighborhood of the integers. The construction is done using a variant of Kargaev's implicit function method. It follows from \cite[Lemma 2.1]{KL96} that the function $f$ in \thmref{thmA3} must change sign, i.e.\ $f$ cannot be chosen nonnegative. This indicates that cancellations are playing a decisive role in the tiling. We can even prove a stronger claim: \begin{thm} \label{thmA9} Let $f+\Lam$ be a tiling at some level $w$, where the set $\Lambda \sbt \R$ is not of bounded density but has tempered growth, and where $f$ is a function in the Schwartz class. Then $f$ must have zero integral. \end{thm} It may seem counter-intuitive at first glance that one can tile $\R$ at a nonzero level $w$ by translates of a Schwartz function $f$ whose integral is zero. \subsection{Organization of the paper} The rest of the paper is organized as follows. In \secref{secF1} some preliminary background is given, and Fourier analytic conditions for tiling are discussed. In \secref{secL1} we prove that if $f+\Lam$ is a tiling at level zero, where $f \in L^1(\R)$ is a nonzero function and $\Lam \sbt \R$ is a nonempty set of bounded density, then $f$ has zero integral (\thmref{thmA5}) and $\Lam$ is relatively dense (\thmref{thmA6}). In \secref{secSM1} we establish the existence of non-periodic tilings $f+\Lam$ of bounded density, such that the function $f$ has ``sparse'' support (\thmref{thmA1}). In \secref{secUB1} we construct tilings $f+\Lam$ of unbounded density, such that $\Lam$ has tempered growth and $f$ is in the Schwartz class (\thmref{thmA3}). We also show that in any such a tiling, the function $f$ must have zero integral (\thmref{thmA9}). In the last section, \secref{secOP1}, we pose some open problems. \section{Preliminaries. Fourier analytic conditions for tiling.} \label{secF1} It is well-known that in the study of translational tilings, Fourier analysis plays an important role, see e.g.\ \cite{Kol04}. If $f \in L^1(\R)$ then its Fourier transform is defined by \begin{equation} \hat{f}(t) = \int_{\R} f(x) \, e^{-2\pi i t x} \, dx. \end{equation} If $\alpha$ is a tempered distribution on $\R$, then $\alpha(\varphi)$ denotes the action of $\alpha$ on a Schwartz function $\varphi$. The Fourier transform $\ft\alpha$ is defined by $\ft{\alpha}(\varphi) = \alpha(\ft{\varphi})$. If $\alpha$ is a tempered distribution on $\R$, and if $\varphi$ is a Schwartz function, then the convolution $\alpha \ast \varphi$ is a tempered distribution whose Fourier transform is $\ft{\alpha} \cdot \ft{\varphi}$. We use $\operatorname{supp}(\alpha)$ to denote the closed support of a tempered distribution $\alpha$. If $f \in L^1(\R)$ then $\operatorname{supp}(f)$ is the smallest closed set such that $f$ vanishes a.e.\ on its complement. This set coincides with the support of $f$ in the distributional sense. For more details on distribution theory we refer the reader to \cite{Rud91}. \subsection{Necessary conditions for tiling} For a discrete set $\Lambda \sbt \mathbb{R}$ we define the measure \begin{equation} \label{eqI5.4} \delta_\Lambda:=\sum_{\lambda\in\Lambda}\delta_\lambda. \end{equation} The tiling condition \eqref{eqI1.1} can then be restated as \begin{equation} \label{P2.2.11} f \ast \delta_\Lam = w \quad \text{a.e.} \end{equation} If $\Lambda$ has bounded density then the measure $\delta_\Lambda$ is a tempered distribution on $\R$. So, at least formally, taking the Fourier transform of both sides of \eqref{P2.2.11} yields \begin{equation} \label{P2.2.12} \ft{f} \cdot \ft{\delta}_\Lam = w \cdot \delta_0. \end{equation} If $f$ is a Schwartz function, then condition \eqref{P2.2.12} makes sense and it is equivalent to \eqref{P2.2.11}. To the contrary, for an arbitrary function $f \in L^1(\R)$ (not assumed to be in the Schwartz class) the product $\ft{f} \cdot \ft{\delta}_\Lam$ is not well-defined, and the condition \eqref{P2.2.12} can only be interpreted as a heuristic principle. The following result, inspired by the heuristic condition \eqref{P2.2.12}, provides a necessary condition for tiling which holds for any $f \in L^1(\R)$. \begin{thm}[{\cite{KL16}}] \label{thm4.5.1} Let $f\in L^1(\mathbb R)$, and $\Lambda\sbt\R$ be a discrete set of bounded density. If $f+\Lambda$ is a tiling at some level $w$, then \begin{equation} \label{eqI3.1} \operatorname{supp}(\hat\delta_\Lambda) \setminus \{0\} \subset \{\hat f=0\}. \end{equation} \end{thm} The proof of this result is based on Wiener's tauberian theorem. In the earlier works \cite{KL96}, \cite{Kol00a}, \cite{Kol00b} the result was proved under various extra assumptions. \subsection{Sufficient conditions for tiling} One may ask whether \thmref{thm4.5.1} admits a converse, i.e.\ if the condition \eqref{eqI3.1} implies that $f+\Lambda$ is a tiling at some level $w$. If the distribution $\ft{\delta}_\Lam$ happens to be a \emph{measure} on $\R$, then the product $\ft{f} \cdot \ft{\delta}_\Lam$ is a well-defined measure and the condition \eqref{P2.2.12} makes sense. In this case, the two conditions \eqref{P2.2.12} and \eqref{eqI3.1} are equivalent, and can be shown to imply that $f+\Lam$ is a tiling: \begin{thm}\label{thm4.5.2} Let $\Lambda \sbt \R$ have bounded density and suppose that $\ft{\delta}_\Lambda$ is a locally finite measure. If $f\in L^1(\R)$ satisfies \eqref{eqI3.1} then $f+\Lambda$ is a tiling at level $w = \ft{\delta}_\Lam(\{0\}) \cdot \int f$. \end{thm} A proof of this result is given below. In \cite{KL96} the result was proved under the extra assumption that $\ft{f}$ is a smooth function. As an example, \thmref{thm4.5.2} applies if the set $\Lambda$ has a periodic structure, i.e.\ if it has the form \eqref{eqI2.1}. In this case $\ft{\delta}_\Lambda$ is a (pure point) measure by Poisson's summation formula, and the condition \eqref{eqI3.1} is both necessary and sufficient for a function $f\in L^1(\R)$ to tile at some level $w$ with the translation set $\Lam$. \thmref{thm4.5.2} leaves open, though, the question as to whether the condition \eqref{eqI3.1} is sufficient for tiling in the general case, i.e.\ for an arbitrary set $\Lam \sbt \R$ of bounded density. This question was addressed recently in \cite{Lev20} where it was answered in the negative: \begin{thm}[{\cite{Lev20}}] \label{thmI8.10} There is a set $\Lambda \subset \R$ of bounded density and a function $f \in L^1(\mathbb{R})$, such that \eqref{eqI3.1} is satisfied however $f + \Lam$ is not a tiling at any level. \end{thm} The proof of this result is based on the relation of the problem to Malliavin's \emph{non-spectral synthesis} example. \thmref{thmI8.10} implies that a converse to \thmref{thm4.5.1} can only be valid under certain extra assumptions. One example of such a converse is given by \thmref{thm4.5.2}. Another example is the following result: \begin{thm}\label{thm4.5.8} Let $f\in L^1(\mathbb R)$, and let $\Lambda \sbt \R$ be a discrete set of bounded density. If the set $\operatorname{supp}(\hat\delta_\Lambda) \setminus \{0\}$ is closed and is disjoint from $\operatorname{supp}(\ft{f})$, then $f+\Lambda$ is a tiling at some level $w$. \end{thm} We will not use this result in the paper and we do not include its proof. For other results of similar type see \cite[Theorem 3]{Kol00a}, \cite[Theorem 5]{Kol00b}. \subsection{Translation-bounded measures} A measure $\mu$ on $\R$ satisfying the condition \begin{equation} \sup_{x \in \R} |\mu|([x,x+1)) < \infty \end{equation} is said to be \emph{translation-bounded}. If a measure $\mu$ is translation-bounded, then it is a tempered distribution. If $\mu$ is a translation-bounded measure on $\R$, and if $\nu$ is a finite measure on $\R$, then the convolution $\mu \ast \nu$ is a translation-bounded measure. \begin{thm} \label{thmC2} Let $\mu$ be a translation-bounded measure on $\R$, and suppose that $\ft{\mu}$ is a locally finite measure. If $\nu$ is a finite measure on $\R$, then the Fourier transform of the convolution $\mu \ast \nu$ is the measure $\ft{\mu} \cdot \ft{\nu}$. \end{thm} In particular, let $\Lambda \sbt \R$ be a discrete set of bounded density, and let $f \in L^1(\R)$. Then the measure $\delta_\Lam$ is translation-bounded, and the convolution $f \ast \del_\Lam$ is the sum of the series $\sum_{\lambda\in\Lambda}f(x-\lambda)$ which converges absolutely a.e., see \cite[Lemma 2.2]{KL96}. If the distribution $\ft{\delta}_\Lambda$ is assumed to be a locally finite measure, then using \thmref{thmC2} we obtain that the three conditions \eqref{P2.2.11}, \eqref{P2.2.12} and \eqref{eqI3.1} are equivalent. We thus see that \thmref{thm4.5.2} is a consequence of \thmref{thmC2}. \thmref{thmC2} should be known to experts but we could not find a proof in the literature, so we include one below for completeness. \begin{proof}[Proof of \thmref{thmC2}] We will use $\sig$ to denote the measure $\ft{\mu}$. The assertion of the theorem means that if $\psi$ is a Schwartz function with compact support then \begin{equation} \label{eqP1.10} \int_{\R} \ft{\psi(x)} \, d(\mu \ast \nu)(x) = \int_{\R} \psi(t) \, \ft{\nu}(t) \, d\sig(t). \end{equation} Let $\chi$ be a Schwartz function whose Fourier transform $\ft{\chi}$ is non-negative, has compact support, $\int \ft{\chi}(t) dt =1$, and for each $\eps > 0$ let $\chi_\eps(x) := \chi( \eps x)$. Let $h_\eps := (\psi \cdot \ft{\nu}) \ast \ft{\chi}_\eps$, then $h_\eps$ is an infinitely smooth function with compact support. As $\eps \to 0$, the function $h_\eps$ remains supported on a certain interval $I=[a,b]$ that does not depend on $\eps$, and $h_\eps$ converges to $\psi \cdot \ft{\nu}$ uniformly on $I$. Hence \begin{equation} \label{eqP1.3} \lim_{\eps \to 0} \ft{\mu}(h_\eps) = \lim_{\eps \to 0} \int_{\R} h_\eps(t) d\sig(t) = \int_{\R} \psi(t) \, \ft{\nu}(t) \, d\sig(t). \end{equation} The function $\psi$ is the Fourier transform of some function $\varphi$ in the Schwartz class. Let $g_\eps := (\varphi \ast \nu) \cdot \chi_\eps$, then $g_\eps$ is a smooth function in $L^1(\R)$ and we have $\ft{g}_\eps = h_\eps$. Since $h_\eps$ belongs to the Schwartz space, the same is true for $g_\eps$, and it follows that \begin{equation} \label{eqP1.7} \ft{\mu}(h_\eps) = \mu(\ft{h}_\eps) = \int_{\R} {g_\eps}(-x) \, d\mu(x) = \int_{\R} (\varphi \ast \nu)(-x) \, \chi_\eps(-x)\, d\mu(x). \end{equation} We observe that $|\chi_\eps(-x)| \leq 1$ and $\chi_\eps(-x) \to 1$ pointwise as $\eps \to 0$. We now wish to apply the dominated convergence theorem to \eqref{eqP1.7}. Using Fubini's theorem we obtain \begin{equation} \label{eqP1.23} \int_{\R} (|\varphi| \ast |\nu|)(-x) \, |d\mu|(x) = \int_{\R} |\varphi(-x)| \, d(|\mu| \ast |\nu|)(x). \end{equation} The function $\varphi$ has fast decay being in the Schwartz class, while $|\mu| \ast |\nu|$ is a translation-bounded measure, hence the integral in \eqref{eqP1.23} converges. We may therefore apply to \eqref{eqP1.7} the dominated convergence theorem, which yields \begin{equation} \label{eqP1.8} \lim_{\eps \to 0} \ft{\mu}(h_\eps) = \int_{\R} (\varphi \ast \nu)(-x) \, d\mu(x) = \int_{\R} \varphi (-x) \, d(\mu \ast \nu)(x). \end{equation} But $\varphi(-x) = \ft{\psi}(x)$, so we see that \eqref{eqP1.10} follows from \eqref{eqP1.3} and \eqref{eqP1.8} as needed. \end{proof} \section{Tiling at level zero} \label{secL1} In this section we prove that if $f+\Lam$ is a tiling at level zero, where $f \in L^1(\R)$ is nonzero and $\Lam \sbt \R$ is nonempty and has bounded density, then $f$ has zero integral (\thmref{thmA5}) and $\Lam$ is a relatively dense set (\thmref{thmA6}). \subsection{The function \texorpdfstring{$f$}{f} has zero integral} \label{subsecZI} We begin with \thmref{thmA5}, for which we give two proofs. The first proof is Fourier analytic, and is based on the following result: \begin{thm}[{\cite{KL16}}] \label{thm5.1} Let $f\in L^1(\mathbb R)$ have nonzero integral. If $\Lambda \sbt \R$ is a set of bounded density and $f+\Lambda$ is a tiling at some level $w$, then there is $a>0$ such that $\hat\delta_\Lambda=c\cdot\delta_0$ in $(-a,a)$, where $c=w\cdot(\int f)^{-1}$. \end{thm} This result follows from the proof of \cite[Theorem 4.1]{KL16} and it is stated in that paper as a remark on p.\ 4598. The proof of \thmref{thm5.1} is based on Wiener's tauberian theorem. \begin{proof}[First proof of \thmref{thmA5}] Let $f+\Lambda$ be a tiling at level $w=0$, and suppose to the contrary that $f$ has nonzero integral. Then by \thmref{thm5.1} there is $a>0$ such that $\hat\delta_\Lambda=c\cdot\delta_0$ in $(-a,a)$, where $c=w\cdot(\int f)^{-1}$. Since $w=0$ this means that $\ft{\delta}_\Lam$ vanishes in a neighborhood $(-a,a)$ of the origin. We will show that this cannot happen. Indeed, choose a Schwartz function $\varphi$ such that $\operatorname{supp}(\varphi)\subset (-a,a)$ and $\hat\varphi>0$. Then we have $\ft{\del}_\Lam(\varphi) = 0$ since $\ft{\delta}_\Lam$ vanishes in a neighborhood of $\operatorname{supp}(\varphi)$. On the other hand, \[ \ft{\del}_\Lam(\varphi) = \del_\Lam(\ft{\varphi}) = \sum_{\lam\in\Lam} \ft{\varphi}(\lam) > 0, \] since $\Lam$ is nonempty and $\hat\varphi$ is everywhere positive. We thus arrive at a contradiction. \end{proof} As a remark, we record here the following observation: \begin{lem} \label{lem5.2} Let $\Lambda \sbt \R$ be a nonempty set of bounded density, and suppose that there is $a>0$ such that $\hat\delta_\Lambda=c\cdot\delta_0$ in $(-a,a)$. Then $c>0$, and $\Lam$ has a uniform density given by $D(\Lam) = c$. \end{lem} \begin{proof} Let $\varphi$ be a Schwartz function such that $\varphi > 0$, $\int \varphi = 1$ and $\operatorname{supp}(\ft{\varphi})\subset (-a,a)$. Then we have $\ft{\varphi} \cdot \ft{\del}_\Lam = c \cdot \del_0$ and hence $\varphi + \Lam$ is a tiling at level $c$. Since $\varphi$ is a positive function and $\Lambda$ is nonempty, the tiling level $c$ must also be positive. Finally, we obtain from \thmref{thmKL1} that the set $\Lam$ has a uniform density given by $D(\Lam) = c$. \end{proof} Next we give our second proof of \thmref{thmA5}. This proof, which does not involve Fourier analysis, is based on the following result due to Ruzsa and Sz\'{e}kely \cite{RS83}. \begin{thm}[{\cite{RS83}}] \label{thmRS83} Let $f \in L^1(\R)$ be real-valued with $\int f > 0$. Then there is a nonnegative (nonzero) $g \in L^1(\R)$ such that the convolution $f \ast g$ is nonnegative a.e. \end{thm} In fact this is proved in \cite{RS83} not only for functions on $\R$, but on a wider class of locally compact abelian groups, which in particular includes also $\R^d$ for every $d \geq 1$. \begin{proof}[Second proof of \thmref{thmA5}] Let $f+\Lambda$ be a tiling at level zero, and suppose to the contrary that $\int f$ is nonzero. By replacing $f(x)$ with the function $\Re\big\{(\int f)^{-1} f(x)\big\}$, we may assume that $f$ is real-valued and that $\int f >0$. So we can use \thmref{thmRS83} to find a nonnegative, nonzero $g \in L^1(\R)$ such that $f \ast g$ is nonnegative a.e. Notice that also the function $f \ast g$ is nonzero, since we have $\int (f \ast g) = (\int f)(\int g) > 0$. On the other hand we have $f \ast \del_\Lam = 0$ a.e., and the measure $\delta_\Lam$ is translation-bounded since $\Lam$ has bounded density. Using Fubini's theorem this implies that \[ (f \ast g) \ast \del_\Lam = g \ast (f \ast \del_\Lam) = g \ast 0 = 0 \quad \text{a.e.,} \] hence $(f \ast g) + \Lam$ is a tiling at level zero. But this is a contradiction, since $f \ast g$ is a nonnegative, nonzero function in $L^1(\R)$, and $\Lam$ is a nonempty set. \end{proof} \subsection{The set \texorpdfstring{$\Lambda$}{Lambda} is relatively dense} We now move on to the proof of \thmref{thmA6}. First we note that this theorem \emph{cannot} be deduced based on \lemref{lem5.2} above, since there exist tilings $f+\Lam$ at level zero such that $\ft{\del}_\Lam$ is \emph{not} a scalar multiple of $\del_0$ in any neighborhood of the origin; see \cite[Section 5]{Lev20} for a construction of such tilings. We fix the following terminology: given a sequence of measures $\{\mu_n\}$ on $\R$, we say that the sequence is \emph{uniformly translation-bounded} if there is a constant $M>0$ not depending on $n$, such that $\sup_{x \in \R} |\mu_n|([x,x+1)) \leq M$ for every $n$. If $\{\mu_n\}$ is a uniformly translation-bounded sequence of measures on $\R$, then we say that $\mu_n$ \emph{converges vaguely} to a measure $\mu$ if for every continuous, compactly supported function $\varphi$ we have $\int \varphi \, d\mu_n \to \int \varphi \, d\mu$ (see, for instance, \cite[Section 7.3]{Fol99}). In this case, the vague limit $\mu$ must also be a translation-bounded measure. For a uniformly translation-bounded sequence of measures $\{\mu_n\}$ to converge vaguely, it is necessary and sufficient that $\{\mu_n\}$ converge in the space of tempered distributions. From any uniformly translation-bounded sequence of measures $\{\mu_n\}$ one can extract a vaguely convergent subsequence $\{\mu_{n_j}\}$. \begin{proof}[Proof of \thmref{thmA6}] Let $f+\Lam$ be a tiling at level zero, and suppose to the contrary that $\Lam$ is not relatively dense. Then for each $r>0$ one can find an open interval $I(r)$ of length $r$ which is disjoint from $\Lam$. By translating the interval $I(r)$ we may assume that one of its endpoints lies in $\Lam$ (since the set $\Lam$ is nonempty). Then for arbitrarily large values of $r$ the right endpoint of $I(r)$ is in $\Lam$, or for arbitrarily large values of $r$ the left endpoint is in $\Lam$. We will assume that the former is the case (the latter case can be treated similarly). It follows that there exist two sequences $r_j \to +\infty$ and $\lam_j \in \Lam$, such that for each $j$ the interval $I_j := (\lam_j - r_j, \lam_j)$ does not intersect $\Lam$. Define $\Lam_j := \Lam - \lam_j$. Since $\Lam$ has bounded density, the measures $\delta_{\Lam_j}$ are uniformly translation-bounded. By passing to a subsequence if necessary, we may assume that $\delta_{\Lam_j}$ converges vaguely to some (also translation-bounded) measure $\mu$ on $\R$. The vague limit $\mu$ is supported on $[0, +\infty)$, since $\delta_{\Lam_j}$ vanishes on $(-r_j, 0)$. Moreover, each measure $\delta_{\Lam_j}$ is positive and has a unit mass at the origin, which implies that $\mu$ is also positive and has an atom at the origin, of mass at least one. In particular the measure $\mu$ is nonzero. On the other hand, since $f+\Lam$ is a tiling (at level zero), then by \thmref{thm4.5.1} the support of $\ft{\delta}_\Lambda$ is contained in $\{\hat f=0\} \cup \{0\}$. Since $\ft{f}$ is a nonzero continuous function, it follows that there is an open interval $(a,b)$ on which $\ft{\delta}_\Lam$ vanishes. In turn this implies that $\ft{\delta}_{\Lam_j}$ also vanishes on $(a,b)$ for each $j$. But as the sequence $\ft{\delta}_{\Lam_j}$ converges to $\ft{\mu}$ in the distributional sense, we obtain that $\ft{\mu}$ vanishes on $(a,b)$ as well. We conclude that $\mu$ is a nonzero, translation-bounded, positive measure on $\R$, supported on $[0, +\infty)$ and whose Fourier transform $\ft{\mu}$ vanishes on some open interval $(a,b)$. But this contradicts classical results on boundary values of functions analytic in the upper half-plane. To be more concrete: choose two Schwartz functions $\varphi, \psi$ such that $\varphi > 0$, $\operatorname{supp}(\ft{\varphi}) \sbt (-\del,\del)$, $\psi$ is nonnegative, $\psi$ is supported on $[0,+\infty)$ and $\int \psi = 1$. Then $h := (\mu \cdot \varphi) \ast \psi$ is a nonzero function belonging to $L^1(\R)$, $\operatorname{supp}(h) \sbt [0,+\infty)$, and the Fourier transform $\ft{h} = (\ft{\mu} \ast \ft{\varphi}) \cdot \ft{\psi}$ is also in $L^1(\R)$ and $\ft{h}$ vanishes on some nonempty open interval provided that $\delta$ is small enough. This contradicts the uniqueness theorem for functions in the Hardy space $H^1$, see e.g.\ \cite[Chapter II]{Gar07}. \end{proof} \section{Non-periodic tilings by functions with sparse support} \label{secSM1} In this section we establish the existence of non-periodic tilings $f+\Lam$ of bounded density, such that the function $f$ has ``sparse'' support (\thmref{thmA1}). Recall that the first example of a tiling $f+\Lam$ such that $f\in L^1(\R)$, $\Lambda \sbt \R$ has bounded density, but the set $\Lam$ does \emph{not} have a periodic structure, was given in \cite{KL16}. The proof was based on the implicit function method due to Kargaev \cite{Kar82}, and it yields a set $\Lam$ which is a small perturbation of the integers. However, in this example the function $f$ is \emph{analytic}, which implies that $\operatorname{supp}(f) = \R$. In order to prove \thmref{thmA1} we will use a different approach, which is based on two main ingredients. The first one is a recent construction from \cite{KS20} of a new type of \emph{crystalline measures} on $\R$. The second main ingredient is a result concerning interpolation of discrete functions by continuous ones with a sparse spectrum. \subsection{Crystalline measures} A tempered distribution $\mu$ on $\R$ satisfying \begin{equation} \label{eqI1.10} \mu = \sum_{\lam\in \Lam} a_\lam \del_\lam, \quad \ft{\mu} = \sum_{s \in S} b_s \del_s \end{equation} (i.e.\ both $\mu$ and $\ft{\mu}$ are pure point measures), where $\Lam$ and $S$ are discrete, closed sets in $\R$, is called a \emph{crystalline measure} \cite{Mey16}. This notion plays a role in the mathematical theory of quasicrystals, i.e.\ atomic arrangements having a discrete diffraction pattern. The classical example of a crystalline measure is $\mu = \delta_\Z$, which satisfies $\ft{\mu} = \mu$ by Poisson's summation formula. More generally, the measure $\delta_\Lam$ is crystalline for any set $\Lam$ of the form \eqref{eqI2.1}, i.e.\ any set $\Lam \sbt \R$ having a periodic structure. There exist also examples of crystalline measures on $\R$, whose support is \emph{not} contained in any set $\Lam$ with a periodic structure. Constructions of such examples, using different approaches, were given in the papers \cite{LO16}, \cite{Kol16}, \cite{Mey16}, \cite{Mey17}, \cite{RV19}. Recently, new progress was achieved by Kurasov and Sarnak \cite{KS20} who constructed examples of crystalline measures on $\R$ enjoying some remarkable properties, answering several questions left open by the above mentioned papers. To state the result, we recall the terminology introduced in \secref{secTUB} above: \begin{definition} \label{defTG} We say that a set $S \sbt \R$ has \emph{tempered growth} if there is $N$ such that \begin{equation} \label{eqI2.5} \# (S \cap (-r,r)) = O(r^N), \quad r \to +\infty. \end{equation} \end{definition} We note that if a set $S \sbt \R$ has tempered growth then the measure $\del_S$ is a tempered distribution, and also the converse is true. \begin{thm}[{\cite{KS20}}] \label{thmI1} There exist crystalline measures $\mu$ of the form \eqref{eqI1.10} with the following properties: \begin{enumerate-num} \item \label{thmI1:i} $\Lambda$ is a set of bounded density; \item \label{thmI1:ii} $a_\lam=1$ for all $\lam\in\Lam$, i.e.\ we have $\mu = \del_\Lam$; \item \label{thmI1:iii} $\Lam$ does not have a periodic structure, i.e.\ it is not of the form \eqref{eqI2.1}; \item \label{thmI1:iv} $S$ is a set of tempered growth. \end{enumerate-num} \end{thm} Actually, the crystalline measures constructed in \cite{KS20} have even stronger properties than stated above -- we only mentioned the properties that will be used in this paper. In the more recent works \cite{Mey20}, \cite{OU20}, alternative approaches to the construction of crystalline measures with these properties can be found. \subsection{Interpolation by functions with a sparse spectrum} We now turn to the second main ingredient in our proof of \thmref{thmA1}. It can be described in the context of the classical uncertainty principle, which states that a nonzero function $f$ and its Fourier transform $\ft{f}$ cannot be ``too small'' at the same time. This general principle has many concrete versions, see \cite{HJ94}. In particular, using complex analysis one can show that if $f \in L^1(\R)$ has compact support, and if $\ft{f}$ vanishes on a set $S \sbt \R$ satisfying \begin{equation} \label{eqI1.6} \limsup_{r \to +\infty} \frac{\# (S \cap (-r,r))}{r} = + \infty, \end{equation} then $f = 0$ a.e. This fact was used in \cite{LM91}, \cite{KL96} in order to prove that if a nonzero function $f \in L^1(\R)$ has compact support, and if $f$ tiles at some level $w$ by a translation set $\Lambda$ of bounded density, then $\Lambda$ must have a periodic structure (\thmref{thmLM91}). On the other hand, suppose that $\Om \sbt \R$ is a set which contains arbitrarily long intervals (and so $\Om$ must be unbounded, but can be very sparse). Then given any discrete set $S \sbt \R$ of tempered growth, there exists a nonzero function $f \in L^1(\R)$, $\operatorname{supp}(f) \sbt \Om$, such that $\ft{f}$ vanishes on $S$. This is a consequence of the following result: \begin{thm} \label{thmB2} Let $\Om \sbt \R$ be a set which contains arbitrarily long intervals, and let $S \sbt \R$ be a set of tempered growth. Then given any values $\{c(s)\} \in \ell^1(S)$, one can find a smooth function $f \in L^1(\mathbb R)$, $\operatorname{supp}(f) \sbt \Om$, such that $\ft{f}(s)=c(s)$ for all $s \in S$. \end{thm} If we call $\operatorname{supp}(f)$ the ``spectrum'' of the function $\ft{f}$, then the result means that every discrete function in $\ell^1(S)$ can be interpolated by a continuous function (the Fourier transform of an $L^1$ function) with spectrum in $\Om$. We refer the reader to \cite{OU16} where the problem of interpolation by functions with a given spectrum is discussed in detail. The approach that we use to prove \thmref{thmB2} is inspired by \cite[Section 3]{OU09}. \subsection{Proof of \thmref{thmB2}} We begin with a simple lemma. \begin{lem} \label{lemC1} Let $S \sbt \R$ be a discrete set of tempered growth, and let $N$ be a sufficiently large number such that \eqref{eqI2.5} holds. Then for any $t \in \R \setminus S$ and any $p>N$ we have \begin{equation} \label{eqC1.1} \sum_{s \in S} |s - t|^{-p} < +\infty. \end{equation} \end{lem} \begin{proof} The condition \eqref{eqI2.5} remains valid if we replace $S$ with $S - t$, so we may assume that $t=0$. The function $n(r) := \# (S \cap (-r,r))$ is then a step function vanishing near the origin. For each $R>0$ which is not a jump discontinuity point of $n(r)$, we have \begin{equation} \label{eqC1.2} \sum_{|s| \leq R} |s|^{-p} = \int_{0}^{R} \frac{dn(r)}{r^{p}} = \frac{n(R)}{R^p} + p \int_{0}^{R} \frac{n(r)}{r^{p+1}} dr \end{equation} which follows from integration by parts. But as $n(r)=O(r^N)$ and $p>N$, the right-hand side of \eqref{eqC1.2} remains bounded as $R \to \infty$. The condition \eqref{eqC1.1} is thus established. \end{proof} \begin{proof}[Proof of \thmref{thmB2}] We suppose that $\Om \sbt \R$ is a set which contains arbitrarily long intervals, and that $S \sbt \R$ is a discrete set of tempered growth. We also choose and fix some enumeration $\{s_j\}$ of the set $S$. Let $\Phi$ be an infinitely smooth function on $\R$ supported on the interval $[-1,1]$, and such that $\int \Phi = 1$. For each $r>0$ we denote $\Phi_r(x) := r^{-1} \Phi(r^{-1}x)$. Define \[ f_j(x) := \Phi_{r_j} (x - \tau_j) \, e^{2 \pi i s_j x} \] where the numbers $r_j > 0$ and $\tau_j \in \R$ will be determined later on. Then we have \[ \ft{f}_j(t) = \ft{\Phi} (r_j(t-s_j)) \, e^{-2 \pi i \tau_j (t-s_j)}. \] In particular, $\ft{f}_j(s_j) = \ft{\Phi} (0) = 1$. Let $N$ be a sufficiently large number such that \eqref{eqI2.5} holds, and let $p > N$. Since $\ft{\Phi}$ is a Schwartz function, there is a constant $C=C(\Phi,p)>0$ such that $|\ft{\Phi}(t)| \leq C |t|^{-p}$ for every nonzero $t$. For $j$ fixed, we have \begin{equation} \label{eqC1.4} \sum_{k \neq j} |\ft{f}_j(s_k)| = \sum_{k \neq j} |\ft{\Phi}(r_j(s_k-s_j))| \leq C r_j^{-p} \sum_{k \neq j} |s_k-s_j|^{-p}. \end{equation} If we use \lemref{lemC1} with the set $\{s_k : k \neq j\}$ and $t = s_j$, the lemma yields that the sum on the right-hand side of \eqref{eqC1.4} converges. We thus see that given any $\eps > 0$, we can choose $r_j = r_j(S,\Phi,p,\eps) >0$ large enough such that the right-hand side of \eqref{eqC1.4} does not exceed $\eps$. It follows that we can choose the numbers $r_j$ in such a way that, no matter how the numbers $\tau_j$ are chosen, we have \begin{equation} \label{eqC1.6} M := \sup_{j} \sum_{k \neq j} |\ft{f}_j(s_k)| \leq \eps. \end{equation} To any sequence $\mathbf{b} = \{b_j\}$ belonging to $\ell^1$, we associate a corresponding sequence $\mathbf{c} = \{c_k\}$ defined by \[ c_k = \sum_{j} \ft{f}_j(s_k) b_j = b_k + \sum_{j \neq k} \ft{f}_j(s_k) b_j. \] It follows from \eqref{eqC1.6} that the mapping $\mathbf{b} \mapsto \mathbf{c}$ defines a bounded linear operator $A: \ell^1 \to \ell^1$ such that $\|A - I\| = M \leq \eps$, where $I$ is the identity operator. If we choose $\eps < 1$ then this implies that $A$ is invertible in $\ell^1$. Now suppose that we are given a sequence $\mathbf{c} = \{c_k\} \in \ell^1$. Let $\mathbf{b} = \{b_j\} \in \ell^1$ be the solution of the equation $A \mathbf{b} = \mathbf{c}$, and define \begin{equation} \label{eqC1.8} f(x) := \sum_{j} b_j f_j(x). \end{equation} We observe that \[ \|f_j\|_{L^1(\R)} = \|\Phi\|_{L^1(\R)}, \quad \sum |b_j| < \infty, \] which implies that the series \eqref{eqC1.8} converges in $L^1(\R)$. It follows that \begin{equation} \label{eqC1.9} \ft{f}(t) = \sum_{j} b_j \ft{f}_j(t), \end{equation} where the series \eqref{eqC1.9} converges uniformly on $\R$. In particular we have \begin{equation} \label{eqC1.10} \ft{f}(s_k) = \sum_{j} b_j \ft{f}_j(s_k) = c_k \end{equation} for each $k$. Finally we observe that $f_j$ is supported on the interval $I_j := [\tau_j - r_j, \tau_j + r_j]$. Since the $r_j$'s were chosen in a way that does not depend on the values of the $\tau_j$'s, we can use the assumption that $\Om$ contains arbitrarily long intervals in order to choose each $\tau_j$ such that the interval $I_j$ lies in $\Om$. Moreover, by choosing these intervals e.g.\ such that $\operatorname{dist}(I_j, I_k) \geq 1$ $(j \neq k)$ this implies that $\operatorname{supp}(f) \sbt \Om$ and ensures that $f$ is infinitely smooth. Thus $f$ has all the required properties, and \thmref{thmB2} is proved. \end{proof} \subsection{Proof of \thmref{thmA1}} Now we can finish the proof of \thmref{thmA1} based on the two results above, namely, \thmref{thmI1} and \thmref{thmB2}. Let $\mu = \del_\Lam$ be the crystalline measure given by \thmref{thmI1}, that is, $\Lambda \sbt \R$ is a set of bounded density but not of a periodic structure, such that $\ft{\del}_\Lam$ is a pure point measure, $\ft{\del}_\Lam = \sum_{s \in S} b_s \del_s$, where $S \sbt \R$ is a set of tempered growth. Since the set $S$ is discrete and closed, it follows that there is $a>0$ such that we have $\hat\delta_\Lambda=c\cdot\delta_0$ in $(-a,a)$. Moreover, we must have $c>0$ due to \lemref{lem5.2}. By rescaling the set $\Lam$ if needed, we may assume that $\hat\delta_\Lambda=\delta_0$ in $(-a,a)$. In particular, this means that the set $S$ contains the origin. Now suppose that we are given a scalar $w$, and a set $\Om \sbt \R$ which contains arbitrarily long intervals. Let $\xi$ be any real number, $\xi \notin S$, and define $S' := S \cup \{\xi\}$. Then also the set $S'$ has tempered growth. We prescribe the values $c(\xi) = 1$, $c(0) = w$, and $c(s) = 0$ for all $s \in S \setminus \{0\}$, then these values belong to $\ell^1(S')$. Hence using \thmref{thmB2} we can find a smooth function $f \in L^1(\mathbb R)$, $\operatorname{supp}(f) \sbt \Om$, such that $\ft{f}(s)=c(s)$ for all $s \in S'$. This means that we have $\ft{f}(0)=w$ and $\ft{f}(s)=0$ for all $s \in S \setminus \{0\}$. It also implies that $\ft{f}(\xi) \neq 0$, which ensures that the function $f$ is nonzero. We conclude that $\ft{f}$ vanishes on the set $\operatorname{supp}(\hat\delta_\Lambda) \setminus \{0\}$, i.e.\ the condition \eqref{eqI3.1} is satisfied, and moreover we have $\ft{\delta}_\Lam(\{0\}) \cdot \int f = w$. Due to \thmref{thm4.5.2} this implies that $f+\Lambda$ is a tiling at level $w$, and the proof is thus complete. \qed \section{Tilings of unbounded density} \label{secUB1} In this section we construct examples of tilings $f + \Lam$ such that the set $\Lam$ does \emph{not} have bounded density (\thmref{thmA3}). We are not aware of any previous example of such a tiling in the literature. Moreover, in our example the set $\Lam$ has tempered growth (see \defref{defTG}) and the function $f$ is in the Schwartz class. We will also show that in any such a tiling, the function $f$ must have zero integral, no matter what the level $w$ of the tiling is (\thmref{thmA9}). \subsection{Translation sets with unbounded density} We will construct tilings $f + \Lam$ such that the translation set $\Lam \sbt \R$ has the form \begin{equation} \label{eqR7.30.12} \Lam = \bigcup_{n \in \Z} \Lam_n, \quad \Lam_n \sbt (n-\eps, n+\eps), \quad \# \Lam_n = n^2+1, \end{equation} where $\eps >0$ is an arbitrarily small number. Then the condition \eqref{eqI10.1} is not satisfied and $\Lam$ does not have bounded density. On the other hand it follows from \eqref{eqR7.30.12} that $\#(\Lambda\cap(-r,r)) = O(r^3)$ as $r\to +\infty$, so the set $\Lam$ has tempered growth and $\delta_\Lambda$ is a tempered distribution on $\R$. We will prove the following theorem: \begin{thm} \label{thmR7.43} Given any $\eps >0$ there is a set $\Lam \sbt \R$ of the form \eqref{eqR7.30.12}, such that \begin{equation} \label{eqR7.30.9} \ft{\delta}_\Lam = \delta_0 - \frac{\delta''_0}{4\pi^2} \quad \text{in $(-\tfrac{1}{2}, \tfrac{1}{2})$.} \end{equation} \end{thm} In order to better understand what is behind condition \eqref{eqR7.30.9}, consider the measure $\mu$ on $\R$ which assigns the mass $n^2+1$ to each point $n \in \Z$. Using Poisson's summation formula one can check that the Fourier transform of $\mu$ is given by $\ft{\mu} = \sum_{k\in \Z} (\delta_k - (4\pi^2)^{-1} \delta''_k)$. \thmref{thmR7.43} now says that one can ``redistribute'' the mass $n^2+1$ assigned at each point $n \in \Z$ into equal unit masses at $n^2+1$ distinct points of a set $\Lam_n$ contained in a small neighborhood of $n$, in such a way that the Fourier transform of the new measure $\del_\Lam$ thus obtained remains unchanged on the interval $(-\tfrac{1}{2}, \tfrac{1}{2})$. The role of condition \eqref{eqR7.30.9} in the tiling problem is clarified by the following lemma: \begin{lem} \label{lemR7.30} Let $\Lam \sbt \R$ be a set of tempered growth, and suppose that there is $a>0$ such that \begin{equation} \label{eqR7.33} \operatorname{supp}(\hat\delta_\Lambda) \cap (-a,a) = \{0\}. \end{equation} Then given any scalar $w$ one can find a nonzero Schwartz function $f$, such that $f+\Lambda$ is a tiling at level $w$. \end{lem} \begin{proof} It is well-known that a distribution supported by the origin is a finite linear combination of derivatives of $\delta_0$ (see \cite[Theorem 6.25]{Rud91}). Hence \eqref{eqR7.33} implies that \begin{equation} \label{eqR7.41} \ft{\delta}_\Lam = \sum_{j=0}^{n} c_j \delta_0^{(j)} \quad \text{in $(-a, a),$} \end{equation} and we may assume that the highest order coefficient $c_n$ is nonzero. It follows that \begin{equation} \label{eqR7.44} t^n \, \ft{\del}_\Lam(t) = c_n n! (-1)^n \cdot \del_0(t) \quad \text{in $(-a, a)$} \end{equation} (see e.g.\ \cite[Section 9.1, Exercise 3]{Fol99}). Let $\varphi$ be a nonzero Schwartz function, $\operatorname{supp}(\ft{\varphi}) \sbt (-a, a)$, $\ft{\varphi}(0)=w$. Then \begin{equation} \label{eqR7.42} f(x) := \frac{\varphi^{(n)}(x)}{c_n n!(-2\pi i)^n} \end{equation} is also a nonzero Schwartz function. We claim that $f+ \Lam$ is a tiling at level $w$. Indeed, \begin{equation} \label{eqR7.46} \ft{f}(t) \, \ft{\delta}_\Lam(t) = \frac{\ft{\varphi}(t) t^n}{c_n n! (-1)^n} \cdot \ft{\delta}_\Lam(t) = \ft{\varphi}(t) \, \delta_0(t) = w \cdot \delta_0(t), \end{equation} where in the second equality we used \eqref{eqR7.44}. This implies that $f \ast \delta_\Lam=w$ as needed. \end{proof} Since condition \eqref{eqR7.30.9} implies \eqref{eqR7.33}, we see that \thmref{thmA3} is a consequence of \thmref{thmR7.43} and \lemref{lemR7.30}. It therefore remains to prove \thmref{thmR7.43}. \subsection{Kargaev's implicit function method} In order to prove \thmref{thmR7.43} we will use a variant of Kargaev's implicit function method \cite{Kar82}. See also \cite{KL16}, \cite{Lev20} for applications of the method in the construction of translational tiling examples. The proof will be done in several steps. \subsubsection{} For each $k=1,2,3,\dots$ let $\chi_k$ be the function on $\R$ defined by \begin{equation} \label{eqR20.1} \chi_k(x) = k-j+1, \quad x \in \big[\tfrac{2(j-1)}{k(k+1)}, \, \tfrac{2j}{k(k+1)}\big), \quad 1 \leq j \leq k, \end{equation} and $\chi_k(x) = 0$ outside the interval $[0, \frac{2}{k+1})$. Let $\{\alpha_n\}$, $n\in\Z$, be a bounded sequence of real numbers. To such a sequence we associate a function $F$ on the real line, defined by \begin{equation} \label{eqR21.4} F(x):=\sum_{n \in \Z} F_n(x), \quad x\in\R, \end{equation} where we let \begin{equation} \label{eqR21.5} F_n(x) := \operatorname{sign}(\alpha_n) \cdot \chi_{n^2+1}(\tfrac{x-n}{\alpha_n}) \end{equation} for each $n \in \Z$ such that $\alpha_n \neq 0$, while if $\alpha_n=0$ then we let $F_n(x):=0$. (The notation $\operatorname{sign}(\alpha_n)$ means $+1$ or $-1$ depending on whether $\alpha_n>0$ or $\alpha_n<0$.) One can verify, using the assumption that the sequence $\{\alpha_n\}$ is bounded, that the series \eqref{eqR21.4} converges in the space of tempered distributions. For instance, this follows from the fact that $(1+x^2)^{-1} \sum |F_n(x)|$ is a bounded function. \begin{thm} \label{thmR7.2} Let $0<\eps<1$. Then for any sufficiently small $r>0$ one can find a real sequence $\alpha = \{\alpha_n\}$, $n\in\Z$, satisfying \begin{equation} \label{eqR7.2.5} (1-\eps) r \le |\alpha_n| \le (1+\eps) r, \quad n \in \Z, \end{equation} and such that $\ft{F}=0$ in $(-\tfrac{1}{2},\tfrac{1}{2})$, where $F$ is the function defined by \eqref{eqR21.4}. \end{thm} We will first prove \thmref{thmR7.2} and then use it to deduce \thmref{thmR7.43}. \subsubsection{} We will need the following lemma. \begin{lem} \label{lemR20.90} The function $\chi_k$ has the following properties: \begin{enumerate-roman} \item \label{lemR20.90.1} $\int \chi_k(x) dx = 1$; \item \label{lemR20.90.2} $\int x \chi_k(x) dx \leq C k^{-1}$; \item \label{lemR20.90.3} $| \ft{\chi}_k(-s) - 1 | \leq C |s| k^{-1}$; \item \label{lemR20.90.4} $\big| v \big( \ft{\chi}_{k}(- v) - 1\big) - u \big( \ft{\chi}_{k}(- u) - 1\big) \big| \leq C k^{-1} |v-u| \cdot \max\{|u|, |v|\}$ \end{enumerate-roman} for every $s,u,v \in \R$, where $C>0$ is an absolute constant. \end{lem} \begin{proof} The property \ref{lemR20.90.1} can be checked directly. In order to establish \ref{lemR20.90.2} we can estimate the left hand side by $k \int_0^{2/(k+1)} x dx = 2k/(k+1)^2$. Together with the estimate \begin{equation} \label{eqR20.4.2} \textstyle | \ft{\chi}_k(-s) - 1 | = \Big| \int \chi_k(x) (e^{2 \pi i sx}-1) dx \Big| \leq 2 \pi |s| \int x \chi_k(x) dx \end{equation} this yields \ref{lemR20.90.3}. Finally, to establish \ref{lemR20.90.4} consider the function $\varphi(x) := x ( e^{2 \pi i x} - 1)$. We may suppose that $u<v$, then the left hand side of \ref{lemR20.90.4} is equal to \begin{align} \label{eqR20.5.7} & \Big| \int \chi_k(x) \frac{\varphi(vx) - \varphi(ux)}{x} dx \Big| = \Big| \int \chi_k(x) \int_u^v \varphi'(sx) ds dx \Big| \\ \label{eqR20.5.8} &\qquad\qquad\leq |v-u| \cdot \max_{s \in [u,v]} \int \chi_k(x) |\varphi'(sx)| dx. \end{align} If we use the estimate $|\varphi'(sx)| \leq C|sx|$, then \eqref{eqR20.5.8} and \ref{lemR20.90.2} imply \ref{lemR20.90.4}. \end{proof} \subsubsection{} Let $X$ be the space of all bounded sequences of real numbers $\alpha = \{\alpha_n\}$, $n \in \Z$, endowed with the norm \[ \|\alpha\|_X := \sup_{n \in \Z} |\alpha_n| \] that makes $X$ into a real Banach space. Denote $I := [-\tfrac{1}{2}, \tfrac{1}{2}]$. Let $Y$ be the space of all continuous functions $\psi:I \to \CC$ satisfying $\psi(-t)=\overline{\psi(t)}$ for all $t\in I$. If we endow $Y$ with the norm $\|\psi\|_Y=\sup|\psi(t)|$, $t\in I$, then also $Y$ is a real Banach space. Let $\alpha = \{\alpha_n\}$, $n \in \Z$, be a sequence in $X$. Define \begin{equation} \label{eqR2.3} (R \alpha)(t):= \sum_{n \in \Z} e^{2\pi int}\cdot \alpha_n \big( \ft{\chi}_{n^2+1}(-\alpha_n t) - 1\big). \end{equation} We notice that the terms of the series \eqref{eqR2.3} are elements of $Y$. By \lemref{lemR20.90}\ref{lemR20.90.3}, the $n$'th term of the series is bounded by $C |\alpha_n|^2 (n^2+1)^{-1} $ uniformly on $I$, where $C>0$ does not depend on $\alpha$ or $n$. Hence the series \eqref{eqR2.3} converges uniformly on $I$ to an element of $Y$, and we have \begin{equation} \label{eqR2.4.1} \|R \alpha\|_Y \leq C \|\alpha\|^2_X, \end{equation} where the constant $C$ does not depend on $\alpha$. We note that the mapping $R : X \to Y$ defined by \eqref{eqR2.3} is \emph{nonlinear}. \subsubsection{} For each $r>0$ let $U_r$ denote the closed ball of radius $r$ around the origin in $X$: \begin{equation} \label{eqR3.1.1} U_r := \{\alpha \in X : \|\alpha\|_X \leq r\}. \end{equation} \begin{lem} \label{lemR3.1} Given any $\rho > 0$ there is $r>0$ such that \begin{equation} \label{eqR3.1.2} \|R\beta - R\alpha\|_Y \leq \rho \|\beta-\alpha\|_X, \quad \alpha,\beta \in U_r. \end{equation} In particular, if $r$ is small enough then $R$ is a contractive (nonlinear) mapping on $U_r$. \end{lem} \begin{proof} Let $\alpha,\beta \in U_r$. Then using \eqref{eqR2.3} we have \begin{equation} \label{eqR3.2} (R \beta - R \alpha)(t) = \sum_{n \in \Z} e^{2\pi int}\cdot \Big[ \beta_n \big( \ft{\chi}_{n^2+1}(-\beta_n t) - 1\big) - \alpha_n \big( \ft{\chi}_{n^2+1}(-\alpha_n t) - 1\big) \Big]. \end{equation} By \lemref{lemR20.90}\ref{lemR20.90.4}, the $n$'th term of the series is bounded by $C r (n^2+1)^{-1} |\beta_n - \alpha_n|$ uniformly on $I$, where $C>0$ does not depend on $r$, $\alpha$, $\beta$ or $n$. Hence the series converges uniformly on $I$ and $\|R\beta - R\alpha\|_Y \leq Cr \|\beta-\alpha\|_X$, where $C$ is a constant not depending on $r$, $\alpha$ or $\beta$. It thus suffices to choose $r$ small enough so that $C r \leq \rho$. \end{proof} \subsubsection{} For each element $\psi \in Y$ we denote by $\F(\psi)$ the sequence \begin{equation} \label{eqR1.1.10} \ft{\psi}(n) = \int_{I} \psi(t) e^{-2\pi i n t} dt, \quad n \in \Z, \end{equation} of Fourier coefficients of $\psi$. Since the Fourier coefficients $\ft{\psi}(n)$ are real and bounded, we have a linear mapping $\F: Y \to X$ satisfying $\|\F(\psi)\|_X \leq \|\psi\|_Y$. \begin{lem} \label{lemR4.2} Given any $\eps>0$ there is $\delta>0$ with the following property: Let $\beta \in X$, $\|\beta\|_X \le \delta$. Then one can find an element $\alpha \in X$, $\|\alpha - \beta\|_X \le \eps \|\beta\|_X$, which solves the equation $\alpha + \F(R \alpha) = \beta$. \end{lem} \begin{proof} Fix $\beta \in X$ such that $\|\beta\|_X \le \delta$, and let \[ B = B(\beta,\eps) :=\{ \alpha \in X : \|\alpha-\beta\|_X \le \eps \|\beta\|_X\}. \] We observe that if $\alpha \in B$ then $\|\alpha\|_X \le (1+\eps) \|\beta\|_X$. Define a map $H: B \to X$ by \[ H(\alpha) := \beta - \F(R\alpha), \quad \alpha \in B, \] and notice that an element $\alpha \in B$ is a solution to the equation $\alpha + \F(R\alpha) = \beta$ if and only if $\alpha$ is a fixed point of the map $H$. Let us show that if $\delta$ is small enough then $H(B)\subset B$. Indeed, if $\alpha \in B$ then using \eqref{eqR2.4.1} we have \[ \|H(\alpha)-\beta\|_X = \|\F(R\alpha)\|_X \leq \|R \alpha\|_Y \leq C \|\alpha\|^2_X \leq C (1+\eps)^2 \|\beta\|^2_X. \] Hence if we choose $\delta$ such that $C(1+\eps)^2 \delta \leq \eps$ then we obtain \[ \|H(\alpha)-\beta\|_X \leq \eps \|\beta\|_X, \] and it follows that $H(B)\subset B$. It also follows from \lemref{lemR3.1} that if $\delta$ is small enough, then $H$ is a contractive mapping from the closed set $B$ into itself. Indeed, let $\alpha', \alpha'' \in B$, then we have \[ \|H(\alpha'') - H(\alpha') \|_X = \|\F(R\alpha'' - R\alpha') \|_X \le \|R\alpha'' - R\alpha' \|_Y \le \rho \|\alpha'' - \alpha'\|_X, \] where $0<\rho<1$. Then the Banach fixed point theorem implies that $H$ has a (unique) fixed point $\alpha \in B$, which yields the desired solution. \end{proof} \subsubsection{Proof of \thmref{thmR7.2}} Let $r>0$, and let $\beta = \{\beta_n\}$, $n \in \Z$, be a real sequence defined by $\beta_n := (-1)^n r$. Then $\|\beta\|_X = r$. If $r=r(\eps)>0$ is small enough, then by \lemref{lemR4.2} there is an element $\alpha \in X$, $\|\alpha - \beta\|_X \le \eps \|\beta\|_X$, which solves the equation $\alpha + \F(R \alpha) = \beta$. We observe that the estimate $\|\alpha - \beta\|_X \le \eps \|\beta\|_X$ implies \eqref{eqR7.2.5}. The relation $\alpha + \F(R \alpha) = \beta$ means that $\beta - \alpha$ is the sequence of Fourier coefficients of the function $R \alpha$. This implies that the series \[ \sum_{n \in \Z} (\beta_n - \alpha_n) e^{2\pi int} \] converges in $ L^2(I)$ to $R \alpha$. Since we have $\beta_n = (-1)^n r$, $n \in \Z$, the series \[ \sum_{n \in \Z} \beta_n e^{2\pi int} \] converges in the distributional sense to the measure $r \cdot \del_{\Z + \frac1{2}}$ on $\R$. In particular, this series converges to zero in the open interval $(-\frac1{2},\frac1{2})$. Let $F$ be the function given by \eqref{eqR21.4} associated to the sequence $\alpha=\{\alpha_n\}$. Then \[ \hat F (-t) = \lim_{N\to\infty}\sum_{|n|\le N}\hat F_n(-t) \] in the sense of distributions, and by \eqref{eqR21.5} we have \begin{equation} \label{eqR21.6} \ft{F}_n(-t) = e^{2 \pi i n t} \cdot \alpha_n \, \ft{\chi}_{n^2+1}(-\alpha_n t). \end{equation} Hence \[ \hat F(-t) = \lim_{N\to\infty} \Big[ \sum_{|n|\le N} \beta_n e^{2\pi int} - \sum_{|n|\le N} (\beta_n - \alpha_n )e^{2\pi int} + \sum_{|n|\le N} e^{2\pi int}\cdot \alpha_n \big( \ft{\chi}_{n^2+1}(-\alpha_n t) - 1\big) \, \Big]. \] The first sum converges in the distributional sense to zero in $(-\tfrac{1}{2}, \tfrac{1}{2})$. The second sum converges in $L^2(I)$ to $R \alpha$. The third sum converges to $R\alpha$ uniformly on $I$. It follows that the distribution $\hat F$ vanishes in the open interval $(-\tfrac{1}{2}, \tfrac{1}{2})$. \qed \subsection{Proof of \thmref{thmR7.43}} Let $0<\eps<1$ be given, and for $r=r(\eps)>0$ sufficiently small let $\{\alpha(n)\}$, $n \in \Z$, be the sequence given by \thmref{thmR7.2}. Define \begin{equation} \label{eqR7.40.1} \Lam_n = \Big\{ n + \frac{2 j\alpha_n }{(n^2+1)(n^2+2)} , \quad 1 \leq j \leq n^2+1 \Big\} \end{equation} and \begin{equation} \label{eqR7.40.2} \Lam = \bigcup_{n \in \Z} \Lam_n. \end{equation} We have $\alpha_n \neq 0$ for every $n \in \Z$, due to \eqref{eqR7.2.5}. Hence $\Lam_n$ is a set with exactly $n^2+1$ elements. The set $\Lam_n$ is contained in the interval $[n - |\alpha_n|, n + |\alpha_n|]$. This yields \eqref{eqR7.30.12} provided that $r>0$ is small enough, again due to \eqref{eqR7.2.5}. In particular we may assume that the sets $\Lam_n$ are pairwise disjoint. Observe that the distributional derivative of the function $F$ in \eqref{eqR21.4} is \[ F'=\sum_{n\in \Z} F'_n = \sum_{n\in \Z}((n^2+1) \delta_n - \delta_{\Lam_n}), \] and hence \[ \delta_{\Lam} = \sum_{n\in \Z} (n^2+1) \delta_n - F'. \] The Fourier transform of the measure $\sum_{n \in \Z} (n^2+1) \delta_n$ is $\sum_{k \in \Z} \big(\delta_k - (4\pi^2)^{-1} \delta''_k \big)$, which follows from Poisson's summation formula. This implies that \[ \hat \delta_\Lambda= \delta_0 - \frac{\delta''_0}{4\pi^2} -\ft{{F'}} \quad \text{in $(-\tfrac{1}{2}, \tfrac{1}{2})$.} \] But since $\ft{F}$ vanishes in $(-\tfrac{1}{2},\tfrac{1}{2})$, then the same is true for $\ft{{F'}}$, so \eqref{eqR7.30.9} is established. \qed \subsection{Proof of \thmref{thmA9}} Finally we show that if $f+\Lam$ is a tiling at some level $w$, where $\Lambda \sbt \R$ is any set of tempered growth but not of bounded density, and $f$ is any function in the Schwartz class, then $f$ must have zero integral. Suppose to the contrary that $\int f = \ft{f}(0)$ is nonzero. Then, due to the continuity and smoothness of $\ft{f}$, there is a Schwartz function $g$ such that $\ft{f} \cdot \ft{g}=1$ in some neighborhood $(-a,a)$ of the origin. Let $h>0$ be a Schwartz function with $\operatorname{supp}(\ft{h})\subset (-a,a)$, then \[ \ft{h} \cdot \ft{g} \cdot \ft{f} =\ft{h} \] and hence \[ h \ast g \ast f = h. \] It follows that \[ h \ast \del_\Lam = (h \ast g \ast f) \ast \del_\Lam = (h \ast g) \ast (f \ast \del_\Lam) = w \cdot \textstyle\int (h \ast g), \] where in the last equality we used the tiling assumption $f \ast \del_\Lam = w$ (the associativity of the convolution is justified since $\del_\Lam$ is a tempered distribution and $f,g,h$ are Schwartz functions, see \cite[Theorem 7.19]{Rud91}). We conclude that $h + \Lam$ is a tiling (at a certain level). But it is known, see \cite[Lemma 2.1]{KL96}, that if $h$ is a \emph{nonnegative}, nonzero function and if $h + \Lam$ is a tiling, then the set $\Lam$ must have bounded density. We thus arrive at a contradiction. \qed \section{Open problems} \label{secOP1} We conclude the paper by posing some open problems. \subsection{Tiling at level zero} The following problem was already mentioned in \secref{secTLZ} above: Let $f+\Lam$ be a tiling at level zero, where the function $f \in L^1(\R)$ is nonzero and the set $\Lam \sbt \R$ is nonempty and has bounded density. Does it follow that $\Lam$ has a \emph{uniform density} $D(\Lam)$\,? In Fourier analytic terms, the problem can be equivalently stated as follows: Let $\Lam \sbt \R$ be nonempty and have bounded density, and suppose that $\ft{\del}_\Lam$ vanishes on some open interval $(a,b)$. Does $\Lam$ necessarily have a uniform density? What makes the problem nontrivial is the existence of tilings $f+\Lam$ at level zero such that $\ft{\del}_\Lam$ is not a scalar multiple of $\del_0$ in any neighborhood of the origin, see \cite[Section 5]{Lev20}. In particular, \lemref{lem5.2} does not apply. We note that by \thmref{thmA6} the set $\Lam$ must be relatively dense, so if the density $D(\Lam)$ exists then it is a strictly positive number. \subsection{Non-periodic tilings} Let $f$ be a nonzero function in $L^1(\R)$, and suppose that the set $\{x : f(x) \neq 0\}$ has \emph{finite measure}. If $f$ tiles at some level $w$ by a translation set $\Lam \sbt \R$ of bounded density, does it follows that $\Lam$ has a periodic structure? \thmref{thmLM91} does not apply here, since $f$ is \emph{not} assumed to have compact support. Does there exist a measurable set $\Om \sbt \R$, $0<\operatorname{mes}(\Om)<+\infty$, whose indicator function $\1_\Om$ can tile at level one, or, a weaker requirement, at some other integer level $w$, with a translation set $\Lam \sbt \R$ that \emph{does not} have a periodic structure? Notice that such a set $\Om$ (if it exists) must be unbounded, again due to \thmref{thmLM91}. \subsection{Tilings of unbounded density} Let $f \in L^1(\R)$ be nonzero and have \emph{compact support}, and suppose that $f+\Lam$ is a tiling at some level $w$, where $\Lam \sbt \R$ is a discrete set (not a multi-set) of tempered growth. Does it follow that $\Lam$ is of the form \eqref{eqI2.1}, i.e. $\Lambda$ is a set of bounded density having a periodic structure? In other words, the question is whether \thmref{thmLM91} remains valid if the set $\Lambda$ is not assumed to have bounded density, but only tempered growth. We note that \thmref{thmA3} does \emph{not} provide a negative answer to this question, since the function $f$ constructed in the proof of this theorem has \emph{unbounded} support. \subsection{Lattice tilings} Let $f \in L^1(\R^d)$, $d \geq 1$, and suppose that $f$ tiles at some level $w$ with a translation set $\Lambda \sbt \R^d$ of bounded density.\footnote{A set $\Lambda \sbt \R^d$ is said to have \emph{bounded density} if there exists $M>0$ such that $\#(\Lam \cap (x+B)) \leq M$ for all $x \in \R^d$, where $B$ is the open unit ball in $\R^d$.} Does there necessarily exist a \emph{lattice} $L \sbt \R^d$ such that $f+L$ is also a tiling, possibly at a different level $w'$\,? The answer is known to be affirmative in the special case where $\Lambda$ is assumed to be a disjoint union of finitely many translated lattices, namely, $\Lam = \biguplus_{j=1}^{N} (L_j + \tau_j)$ where each $L_j$ is a lattice in $\R^d$ and the $\tau_j$ are translation vectors. This was proved in dimension one in \cite[p.\ 673]{KL96}, while in several dimensions the result was proved more recently in \cite[Theorem 1.6]{Liu18}. In both proofs, number theory plays an essential role: the proof in $\R$ uses the classical Skolem--Mahler--Lech theorem, while in $\R^d$ the proof relies on a result due to Evertse, Schlickewei and Schmidt \cite{ESS02}. \bibliographystyle{amsplain}

https://arxiv.org/abs/2208.03959

Partial reconstruction of measures from halfspace depth

The halfspace depth of a $d$-dimensional point $x$ with respect to a finite (or probability) Borel measure $\mu$ in $\mathbb{R}^d$ is defined as the infimum of the $\mu$-masses of all closed halfspaces containing $x$. A natural question is whether the halfspace depth, as a function of $x \in \mathbb{R}^d$, determines the measure $\mu$ completely. In general, it turns out that this is not the case, and it is possible for two different measures to have the same halfspace depth function everywhere in $\mathbb{R}^d$. In this paper we show that despite this negative result, one can still obtain a substantial amount of information on the support and the location of the mass of $\mu$ from its halfspace depth. We illustrate our partial reconstruction procedure in an example of a non-trivial bivariate probability distribution whose atomic part is determined successfully from its halfspace depth.

\section{The Depth Characterization/Reconstruction Problem} Let $x$ be a point in the $d$-dimensional Euclidean space $\R^d$ and let $\mu$ be a finite Borel measure in $\R^d$. We write $\half$ for the collection of all closed halfspaces\footnote{A halfspace is one of the two regions determined by a hyperplane in $\R^d$; any halfspace can be written as a set $\left\{ y \in \R^d \colon \left\langle y, u \right\rangle \leq c \right\}$ for some $c \in \R$ and $u \in \R^d \setminus \left\{ 0 \right\}$.} in $\R^d$ and $\half(x)$ for the subset of those halfspaces from $\half$ that contain $x$ in their boundary hyperplane. The \emph{halfspace depth} (or \emph{Tukey depth}) of the point $x$ with respect to $\mu$ is defined as \begin{equation} \label{halfspace depth} \D\left(x;\mu\right) = \inf_{H\in\half(x)} \mu(H). \end{equation} The history of the halfspace depth in statistics goes back to the 1970s \cite{Tukey1975}. The halfspace depth plays an important role in the theory and practice of nonparametric inference of multivariate data; for many references see \cite{Liu_etal1999,Nagy_etal2019,Zuo_Serfling2000}. The depth~\eqref{halfspace depth} was originally designed to serve as a multivariate generalization of the quantile function. As such, it is desirable that just as the quantile function in $\R$, the depth function $x \mapsto \D(x;\mu)$ in $\R^d$ characterizes the underlying measure $\mu$ uniquely, and $\mu$ is straightforward to be retrieved from its depth. The question whether the last two properties are valid for $\D$ are known as the \emph{halfspace depth characterization and reconstruction problems}. They both turned out not to have an easy answer. In fact, only the recent progress in the theory of the halfspace depth gave the first definite solutions to some of these problems. In \cite{Nagy2019b}, the general characterization question for the halfspace depth was answered in the negative, by giving examples of different probability distributions in $\R^d$ with $d \geq 2$ with identical halfspace depth functions. On the other hand, several authors have obtained also partial positive answers to the characterization problem; for a recent overview of that work see \cite{Nagy2020c}. Only three types of distributions are known to be completely characterized by their halfspace depth functions: \begin{enumerate*}[label=(\roman*)] \item univariate measures, in which case the depth~\eqref{halfspace depth} is just a simple transform of the distribution function of $\mu$; \item atomic measures with finitely many atoms (which we subsequently call \emph{finitely atomic measures} for brevity) in $\R^d$ \cite{Struyf_Rousseeuw1999,Laketa_Nagy2021}; and \item measures that possess all Dupin floating bodies\footnote{A Borel measure $\mu$ on $\R^d$ is said to possess all Dupin floating bodies if each tangent halfspace to the halfspace depth upper level set $\left\{ x \in \R^d \colon \D(x;\mu) \geq \alpha \right\}$ is of $\mu$-mass exactly $\alpha$, for all $\alpha \geq 0$.} \cite{Nagy_etal2019}.\end{enumerate*} In this contribution we revisit the halfspace depth reconstruction problem. We pursue a general approach, and do not restrict only to atomic measures, or to measures with densities. Our results are valid for any finite (or probability) Borel measure $\mu$ in $\R^d$. As the first step in addressing the reconstruction problem, our intention is to identify the support and the location of the atoms of $\mu$, based on its depth. We will see at the end of this note that without additional assumptions, neither of these problems is possible to be resolved. We, however, prove several positive results. We begin by introducing the necessary mathematical background in Section~\ref{section:preliminaries}. In Section~\ref{section:main} we state our main theorem; a detailed proof of that theorem is given in the Appendix. We show that \begin{enumerate*}[label=(\roman*)] \item the support of the measure $\mu$ must be concentrated only in the boundaries of the level sets of its halfspace depth; \item each atom of $\mu$ is an extreme point of the corresponding (closed and convex) upper level sets of the halfspace depth; and \item each atom of $\mu$ induces a jump in the halfspace depth function on the line passing through that atom. \end{enumerate*} These advances enable us to identify the location of the atoms of $\mu$, at least in simpler scenarios. We illustrate this in Section~\ref{section:examples}, where we give an example of a non-trivial bivariate probability measure $\mu$ whose atomic part we are able to determine from its depth. We conclude by giving an example of two measures whose depth functions are the same, yet both their supports and the location of their atoms differ. \section{Preliminaries: Flag Halfspaces and Central Regions} \label{section:preliminaries} \subsubsection*{Notations.} When writing simply a subspace of $\R^d$ we always mean an affine subspace, that is the set $a + L = \left\{ a + x \in \R^d \colon x \in L \right\}$ for $a \in \R^d$ and $L$ a linear subspace of $\R^d$. The intersection of all subspaces in $\R^d$ that contain a set $A \subseteq \R^d$ is called the affine hull of $A$, and denoted by $\aff{A}$. It is the smallest subspace that contains $A$. The affine hull $\aff{\{x,y\}}$ of two different points $x, y \in \R^d$ is the infinite line passing through both $x$ and $y$; another example of a subspace is any hyperplane in $\R^d$. For a set $A\subseteq \R^d$ we write $\intr(A)$, $\cl(A)$ and $\bd(A)$ to denote the interior, closure, and boundary of $A$, respectively. The interior, closure, and boundary of a set $B\subseteq A$ when considered only as a subset of a subspace $A \subseteq \R^d$ are denoted by $\intr_A(B)$, $\cl_A(B)$ and $\bd_A(B)$, respectively. For two different points $x,y\in\R^d$, $x \ne y$, we denote by $L(x,y)$ the interior of the line segment between $x$ and $y$ when considered inside the infinite line $\aff{\{x,y\}}$. In other words, $L(x,y)$ is the open line segment between $x$ and $y$. In the special case of $A=\aff{B}$ we write $\relint(B)=\intr_A(B)$, $\relbd(B)=\bd_A(B)$ and $\relcl(B)=\cl_A(B)$ to denote the relative interior, relative boundary, and relative closure of $B$, respectively. For instance, $\relbd(L(x,y)) = \{x, y \}$ and $L(x,y) = \relint(L(x,y))$, but $\intr(L(x,y)) = \emptyset$ if $d>1$. We write $\Meas$ for the collection of all finite Borel measures in $\R^d$. For a subspace $A\subseteq \R^d$ and $\mu\in\Meas$ we write $\mu|_A$ to denote the measure obtained by restricting $\mu$ to the subspace $A$, that is the finite Borel measure given by $\mu|_A(B) = \mu(B \cap A)$ for any Borel set $B \subseteq \R^d$. By $\Support{\mu}$ we mean the support of $\mu\in\Meas$, which is the smallest closed subset of $\R^d$ of full $\mu$-mass. \subsection{Minimizing halfspaces and flag halfspaces} For $\mu\in\Meas$ and $x \in \R^d$ we call $H \in \half(x)$ a \emph{minimizing halfspace} of $\mu$ at $x$ if $\mu(H) = \D\left(x;\mu\right)$. For $d = 1$ a minimizing halfspace always trivially exists. It also exists if $\mu$ is smooth in the sense that $\mu(\bd(H)) = 0$ for all $H \in \half(x)$, or if $\mu \in \Meas$ is finitely atomic. In general, however, the infimum in~\eqref{halfspace depth} does not have to be attained. We give a simple example. \begin{example} \label{example:flag} Take $\mu \in \Meas[\R^2]$ the sum of a uniform distribution on the disk $B = \left\{ x \in \R^2 \colon \left\Vert x \right\Vert \leq 2 \right\}$ and a Dirac measure at $a = (1,1) \in \R^2$. For $x = (1,0) \in \R^2$ no minimizing halfspace of $\mu$ at $x$ exists. As can be seen in Fig.~\ref{figure:flag halfspace}, the depth $\D(x;\mu)$ is approached by $\mu(H_{n})$ for a sequence of halfspaces $H_n\in\half(x)$ with inner normals $v_n = \left( \cos(-1/n), \sin(-1/n) \right)$ that converge $H_{v} \in \half(x)$ with inner normal $v = (1,0)$, yet $\D(x;\mu) = \lim_{n \to \infty} \mu(H_n) < \mu(H_{v})$. \end{example} \begin{figure}[htpb] \begin{center} \includegraphics[width=\twofigb\textwidth]{Flag1.eps} \qquad \includegraphics[width=\twofigb\textwidth]{Flag2.eps \end{center} \caption{The support of $\mu \in \Meas[\R^2]$ from Example~\ref{example:flag} (colored disk) and its unique atom $a$ (diamond). No minimizing halfspace of $\mu$ at $x = (1,0)\in\R^2$ exists. On the left hand panel we see a halfspace $H_n \in \half(x)$ whose $\mu$-mass is almost $\D(x;\mu)$. The halfspace $H_n$ does not contain $a$. On the right hand panel the unique minimizing flag halfspace $F \in \flag(x)$ of $\mu$ at $x$ is displayed.} \label{figure:flag halfspace} \end{figure} For certain theoretical properties of the halfspace depth of $\mu$ to be valid, the existence of minimizing halfspaces appears to be crucial. As a way to alleviate the issue of their possible non-existence, in \cite{Pokorny_etal2022} a novel concept of the so-called flag halfspaces was introduced. A \emph{flag halfspace} $F$ centered at a point $x\in\R^d$ is defined as any set of the form \begin{equation}\label{flag halfspace} F=\{x\} \cup \left( \bigcup_{i=1}^d \relint(H_i) \right). \end{equation} In this formula, $H_d\in\half(x)$ and for each $i\in \{1,\dots,d-1\}$, $H_i$ stands for an $i$-dimensional halfspace inside the subspace $\relbd(H_{i+1})$ such that $x\in\relbd(H_i)$. The collection of all flag halfspaces in $\R^d$ centered at $x\in\R^d$ is denoted by $\flag(x)$. An example of a flag halfspace in $\R^2$ is displayed in the right hand panel of Fig.~\ref{figure:flag halfspace}. That flag halfspace is a union of an open halfplane $H_2$ (light-colored halfplane) whose boundary passes through $x$, a halfline (thick halfline) in the boundary line $\bd(H_2)$ starting at $x$, and the point $x$ itself. The results derived the present paper lean on the following crucial observation, whose complete proof can be found in \cite[Theorem~2]{Pokorny_etal2022}. \begin{lemma}\label{theorem:Pokorny} For any $x\in\R^d$ and $\mu\in\Meas$ it holds true that \[ \D\left(x;\mu\right)=\min_{F\in\flag(x)}\mu(F). \] In particular, there always exists $F\in\flag(x)$ such that $\mu(F)=\D\left(x;\mu\right)$. \end{lemma} Any flag halfspace $F \in \flag(x)$ from Lemma~\ref{theorem:Pokorny} that satisfies $\mu(F)=\D\left(x;\mu\right)$ is called a \emph{minimizing flag halfspace} of $\mu$ at $x$. This is because it minimizes the $\mu$-mass among all the flag halfspaces from $\flag(x)$. Lemma~\ref{theorem:Pokorny} tells us two important messages. First, the halfspace depth $\D(x;\mu)$ can be introduced also in terms of flag halfspaces instead of the usual closed halfspaces in~\eqref{halfspace depth}, and the two formulations are equivalent to each other. Second, in contrast to the usual minimizing halfspaces that do not exist at certain points $x \in \R^d$, according to Lemma~\ref{theorem:Pokorny} there always exists a minimizing flag halfspace of any $\mu$ at any $x$. \subsection{Halfspace depth central regions} The upper level sets of the halfspace depth function $\D(\cdot;\mu)$, given by \begin{equation} \label{central region} \Damu = \left\{ x \in \R^d \colon \depth{x}{\mu} \geq \alpha \right\} \mbox{ for }\alpha\geq 0, \end{equation} play the important role of multivariate quantiles in depth statistics. The set $\Damu$ is called the \emph{central region} of $\mu$ at level $\alpha$. All central regions are known to be convex and closed. The sets~\eqref{central region} are clearly also nested, in the sense that $\Damu \subseteq \D_\beta(\mu)$ for $\beta\leq \alpha$. Besides~\eqref{central region}, another collection of depth-generated sets of interest considered in \cite{Laketa_Nagy2021b,Pokorny_etal2022} is \begin{equation*} \Uamu = \left\{ x \in \R^d \colon \depth{x}{\mu} > \alpha \right\} \mbox{ for }\alpha\geq 0. \end{equation*} We conclude this collection of preliminaries with another result from \cite{Pokorny_etal2022}, which tells us that no set difference of the level sets $\Damu\setminus\Uamu$ can contain a relatively open subset of positive $\mu$-mass. That result lends an insight into the properties of the support of $\mu$, based on its depth function $\D\left(\cdot;\mu\right)$. It will be of great importance in the proof of our main result in Section~\ref{section:main}. The complete proof of the next technical lemma can be found in~\cite[Lemma~6]{Pokorny_etal2022}. \begin{lemma}\label{main lemma for support} Let $\mu \in \Meas$ and let $K \subset \R^d$ be a relatively open set of points of equal depth of $\mu$ that contains at least two points. Then $\mu(K) = 0$. \end{lemma} \section{Main Result} \label{section:main} The preliminary Lemma~\ref{main lemma for support} hints that the mass of $\mu$ cannot be located in the interior of regions of constant depth. We refine and formalize that claim in the following Theorem~\ref{theorem:support}, which is the main result of the present work. In part~\ref{support2} of Theorem~\ref{theorem:support} we bound the support of $\mu\in\Meas$, based on the information available in its depth function $\D\left(\cdot;\mu\right)$. We do so by showing that $\mu$ may be supported only in the closure of the boundaries of the central regions $\Damu$. That is a generalization of a similar result, known to be valid in the special case of finitely atomic measures $\mu\in\Meas$ \cite{Laketa_Nagy2021,Liu_etal2020,Struyf_Rousseeuw1999}. In the latter situation, all central regions $\Damu$ are convex polytopes, there is only a finite number of different polytopes in the collection $\left\{ \Damu \colon \alpha \geq 0 \right\}$, and the atoms of $\mu$ must be located in the vertices of the polytopes from that collection. Nevertheless, not all vertices of $\Damu$ are atoms of $\mu$; an algorithmic procedure for the reconstruction of the atoms, and the determination of their $\mu$-masses, is given in \cite{Laketa_Nagy2021}. Extending the last observation about the possible location of atoms from finitely atomic measures to the general scenario, in part~\ref{support1} of Theorem~\ref{theorem:support} we show that all atoms of $\mu$ are contained in the extreme points\footnote{For a convex set $C \subset \R^d$, a face of $C$ is a convex subset $F \subseteq C$ such that $x, y \in C$ and $(x+y)/2 \in F$ implies $x, y \in F$. If $\{z\}$ is a face of $C$, then $z$ is called an extreme point of $C$.} of the central regions $\Damu$. Note that this indeed corresponds to the known theory for finitely atomic measures --- the extreme points of polytopes are exactly their vertices. Our last observation in part~\ref{jump} of Theorem~\ref{theorem:support} is that each atom $x\in\R^d$ of $\mu$ induces a jump discontinuity in the halfspace depth, when considered on the straight line connecting any point of higher depth with $x$. This will be useful in detecting possible locations of atoms for general measures. \begin{theorem} \label{theorem:support} Let $\mu\in\Meas$. \begin{enumerate}[label=(\roman*)] \item \label{support2} Let $A$ be a subspace of $\R^d$ that contains at least two points. Then \[ \Support{\mu|_A} \subseteq \cl_A\left(\bigcup_{\alpha \geq 0}\bd_A\left(\Damu\cap A\right)\right), \] In particular, for $A = \R^d$ we have \[ \Support{\mu} \subseteq \cl\left(\bigcup_{\alpha \geq 0}\bd\left(\Damu\right)\right). \] \item \label{support1} Each atom $x$ of $\mu$ with $\depth{x}{\mu} = \alpha$ is an extreme point of $\D_\beta(\mu)$ for all $\beta \in (\alpha - \mu(\{x\}), \alpha]$. \item \label{jump} For any $x \in \R^d$ with $\depth{x}{\mu} = \alpha$, any $z\in\Uamu$, and any $y \in \R^d$ such that $x$ belongs to the open line segment $L(y,z)$ between $y$ and $z$, it holds true that \[ \D\left(y;\mu\right)\leq\D\left(x;\mu\right)-\mu(\{x\}).\] \end{enumerate} \end{theorem} The proof of Theorem~\ref{theorem:support} is given in the Appendix. Theorem~\ref{theorem:support} sheds light on the support and the location of the atoms of a measure. Its part~\ref{support2} tells us that if a depth function $\D\left(\cdot;\mu\right)$ attains only at most countably many different values, and each level set $\Damu$ is a polytope, the mass of $\mu$ must be concentrated in the closure of the set of vertices of the level sets $\Damu$. A special case is, of course, the setup of finitely atomic measures treated in \cite{Struyf_Rousseeuw1999,Laketa_Nagy2021}. \section{Examples} \label{section:examples} We conclude this note by giving two examples. Parts~\ref{support1} and~\ref{jump} of Theorem~\ref{theorem:support} show a way, at least in special situations, to locate the atomic parts of measures. We start by reconsidering our motivating Example~\ref{example:flag}. The distribution $\mu\in\Meas[\R^2]$ is not purely atomic, and can be shown not to possess Dupin floating bodies. Thus, it is currently unknown whether its depth function $\D\left(\cdot;\mu\right)$ determines $\mu$ uniquely. In our first example of this section we show how Theorem~\ref{theorem:support} recovers the position the atomic part of $\mu$. Then, in Example~\ref{example:Nagy} we argue that the general problem of determining the support, or the location of the atoms of $\mu \in \Meas$ from its halfspace depth is impossible to be solved without further restrictions. \addtocounter{example}{-1} \begin{example}[continued] Suppose that in Example~\ref{example:flag} we have $\mu(\{a\}) = \delta$ for $\delta \in (0,1/2)$ small enough, and that the non-atomic part of $\mu$ is $\nu\in\Meas[\R^2]$ uniform on the disk $B$, with $\nu(B) = 1$. Hence, $\mu(\R^2) = \nu(B) + \mu(\{a\}) = 1 + \delta$. We first compute the halfspace depth function $\D\left(\cdot;\mu\right)$ of $\mu$, and then show how to use Theorem~\ref{theorem:support} to find the atom $a$ of $\mu$ from its depth. The computation of the depth function is done by means of determining all the central regions~\eqref{central region} at levels $\beta \geq 0$ of $\mu$. We denote $\alpha = \D(x;\mu)$, and split our argumentation into three situations according to the behavior of the regions $\D_{\beta}(\mu)$: \begin{enumerate*}[label=(\roman*)] \item $\beta < \alpha$ where $x$ is contained in the interior of $\D_{\beta}(\mu)$; \item $\beta \in (\alpha, \alpha + \delta]$ where $x$ lies in the boundary of $\D_{\beta}(\mu)$; and \item $\beta > \alpha + \delta$ where $\D_{\beta}(\mu)$ does not contain $x$.\end{enumerate*} First note that because $\nu$ is uniform on a unit disk, all non-empty depth regions $D_{\beta}(\nu)$ of $\nu$ are circular disks centered at the origin, and all the touching halfspaces\footnote{We say that $H\in\half$ is \emph{touching} $A\subset\R^d$ if $H\cap A\neq \emptyset$ and $\intr(H)\cap A=\emptyset$.} of $\D_{\beta}(\nu)$ carry $\nu$-mass exactly $\beta$. \begin{figure}[htpb] \begin{center} \includegraphics[width=\twofigb\linewidth]{Example1continued.eps} \qquad \includegraphics[width=\twofigb\linewidth]{Cauchy.eps} \end{center} \caption{Left panel: The measure $\mu$ from Example~\ref{example:flag} being the sum of a uniform distribution on a disk and a single atom at $a \in \R^2$ (black point) with $\delta = 1/10$, with several contours of its depth $\D\left(\cdot;\mu\right)$ (thick colored lines). The halfspace median is the red line segment in the middle of the plot. From the depth only, Theorem~\ref{theorem:support} allows us to determine the mass and the location of the atom. Two depth contours that share $a \in \R^2$ as an extreme point are plotted with boundaries in green. Right panel: Example~\ref{example:Nagy} with $d=2$. Several density contours of the measure $\mu \in \Meas[\R^2]$ (solid lines) and its atom (point at the origin), together with multiple contours of the corresponding depth $\D(\cdot;\mu) \equiv \D(\cdot; \nu)$ (dashed lines).} \label{figure:atom} \end{figure} \smallskip\noindent \textbf{Case I:} $\beta\leq \alpha$. For $\alpha = \depth{a}{\mu} = \depth{a}{\nu}$ we have that $\D_{\alpha}(\nu)$ is a disk centered at the origin containing $a$ on its boundary. Note that the halfspace depths of $\mu$ and $\nu$ remain the same outside $\D_{\alpha}(\nu)$, since the added atom $a$ does not lie in any minimizing halfspace of $x \notin \D_\alpha(\nu)$, so we have $\D_{\beta}(\mu)=\D_{\beta}(\nu)$ for all $\beta \leq \alpha$. \smallskip\noindent \textbf{Case II:} $\beta\in (\alpha,\alpha+\delta]$. We have $\depth{a}{\mu}=\alpha+\delta\geq \beta$, meaning that $a\in D_{\beta}(\mu)$. Because $\mu$ is obtained by adding mass to $\nu$, it must be $D_{\beta}(\nu)\subseteq D_{\beta}(\mu)$ and due to the convexity of the central regions \eqref{central region}, the convex hull $C$ of $D_{\beta}(\nu) \cup \{a\}$ must be contained in $D_{\beta}(\mu)$. Denote by $H \in \half(a)$ a touching halfspace of $D_{\beta}(\nu)$ that contains $a$ on its boundary. Then $\nu(H)=\beta$, and hence $\intr(H)\cap D_{\beta}(\mu)=\emptyset$. We obtain that $D_{\beta}(\mu)$ is equal to the convex hull of $D_{\beta}(\nu)\cup\{a\}$. \smallskip\noindent \textbf{Case III:} $\beta>\alpha+\delta$. In a manner similar to Case II one concludes that $D_{\beta}(\mu)$ is the convex hull of a circular disk $D_{\beta}(\nu)$ and $a$, intersected with the disk $D_{\beta-\delta}(\nu)$. \smallskip In order to complete the reconstruction of the atomic part of measure $\mu$ from Example~\ref{example:flag} based on its depth function, we present Lemma~\ref{lemma:Laketa}, which is a special case of a more general result (called the \emph{generalized inverse ray basis theorem}) whose complete proof can be found in \cite[Lemma~4]{Laketa_Nagy2021b}. \begin{lemma} \label{lemma:Laketa} Suppose that $\mu\in\Meas$, $\alpha > 0$, a point $x\notin \Damu$ and a face $F$ of $\Damu$ are given so that the relatively open line segment $L(x,y)$ does not intersect $\Damu$ for any $y \in F$. Then there exists a touching halfspace $H \in \half$ of $\Damu$ such that $\mu(\intr(H))\leq\alpha$, $x\in H$, and $F\subset\bd(H)$ \end{lemma} \smallskip\noindent \emph{Reconstruction.} We now know the complete depth function $\D\left(\cdot;\mu\right)$ of $\mu$, see also Fig.~\ref{figure:atom}. From this depth only, we will locate the atoms of $\mu$ and their mass. The only point in $\R^2$ that is an extreme point of more than one depth region is certainly $a$, so that $a$ is the only possible candidate for an atom of $\mu$ by part~\ref{support1} of Theorem~\ref{theorem:support}. Take any $\beta\in(\alpha,\alpha+\delta)$. Then $D_{\beta}(\mu)$ is the convex hull of a circular disk and the point $a$ outside that disk, so its boundary contains a line segment $L(a,y_\beta)$ for $y_\beta\in \bd(D_{\beta}(\nu))$. Due to Lemma~\ref{lemma:Laketa}, there is a halfspace $H_{\beta} \in \half$ such that $L(a,y_\beta)\subset\bd(H_{\beta})$ and $\mu(\intr(H_{\beta}))\leq \beta < \alpha+\delta=\depth{a}{\mu} \leq \mu(H_\beta)$, the last inequality because $a \in H_\beta$. We obtain $\mu(\bd(H_{\beta}))\geq \alpha + \delta - \beta$. This is true for any $\beta \in (\alpha, \alpha + \delta)$, and for different $\beta_1, \beta_2 \in (\alpha,\alpha+\delta)$ we have $H_{\beta_1}\neq H_{\beta_2}$ with $x \in \bd(H_{\beta_i})$ and $\mu\left(\bd(H_{\beta_i})\right) \geq \alpha + \delta - \beta_i$, $i=1,2$. In conclusion, we obtain uncountably many different lines $\bd(H_\beta)$ of positive $\mu$-mass, all passing through $a$. That is possible only if $a$ is an atom of $\mu$, and $\mu(\{a\}) \geq \delta$. Theorem~\ref{theorem:support} again guarantees that $\mu(\{a\}) = \delta$ and that there is no other atom of $\mu$. \end{example} The complete Example~\ref{example:flag} gives a partial positive result toward the halfspace depth characterization problem, and promises methods allowing one to determine features of $\mu$ from its depth $\D\left(\cdot;\mu\right)$, at least for special sets of measures. The complete determination of the support or the atoms of $\mu$ from its depth is, however, a problem considerably more difficult, impossible to be solved in full generality. Follows an example of mutually singular measures\footnote{Recall that $\mu, \nu \in \Meas$ are called \emph{singular} if there is a Borel set $A \subset \R^d$ such that $\mu(A) = \nu(\R^d\setminus A) = 0$.} $\mu, \nu \in \Meas$ sharing the same depth function from \cite[Section~2.2]{Nagy2020c}. \begin{example} \label{example:Nagy} For $\mu_1 \in \Meas[\R^d]$ with independent Cauchy marginals and $\mu_2 \in \Meas[\R^d]$ the Dirac measure at the origin, define $\mu \in \Meas[\R^d]$ by the sum of $\mu_1$ and $\mu_2$ with weights $1/d$ and $1/2 - 1/(2d)$, respectively. The total mass of $\mu$ is hence $\mu\left(\R^d\right) = 1/2 + 1/(2d)$, and its support is $\R^d$. For the other distribution take $\nu \in \Meas[\R^d]$ the probability measure supported in the coordinate axes $A_i = \left\{ x = \left(x_1, \dots, x_d \right) \in \R^d \colon x_j = 0 \mbox{ for all }j \ne i \right\}$, $i=1,\dots,d$. The density $g$ of $\nu$ with respect to the one-dimensional Hausdorff measure on its support $\Support{\nu} = \bigcup_{i=1}^d A_i$ is given as a weighted sum of densities of univariate Cauchy distributions in $A_i$ \[ g(x) = \frac{1}{d} \sum_{i=1}^d \frac{\I{x \in A_i}}{\pi(1+x_i^2)} \quad\mbox{for }x = \left(x_1, \dots, x_d\right)\in\R^d. \] It can be shown \cite[Section~2.2]{Nagy2020c} that the depths of $\mu$ and $\nu$ coincide at all points $x = \left(x_1, \dots, x_d\right)$ in $\R^d$ \[ \depth{x}{\mu} = \depth{x}{\nu} = \begin{cases} \frac{1}{d} \left( \frac{1}{2} - \frac{{\arctan(\max_{i=1,\dots,d} \left\vert x_i \right\vert)}}{\pi} \right) & \mbox{if }x \in \R^d \setminus \{ 0 \}, \\ 1/2 & \mbox{for }x = 0 \in \R^d. \end{cases} \] The two measures $\mu$ and $\nu$ are, however, singular as for $A = \Support{\nu}\setminus\{0\}$ we have $\mu(A) = \nu(\R^d \setminus A) = 0$. For an arbitrary finite Borel measure, it is therefore impossible to retrieve the full information about its support only from its depth function. For a visualization of measure $\mu$ and its halfspace depth see Fig.~\ref{figure:atom}. The same example demonstrates that in general, also the location of the atoms of $\mu\in\Meas$, or even the number of them, cannot be recovered from the depth function $\D\left(\cdot;\mu\right)$ only --- the measure $\nu$ in Example~\ref{example:Nagy} does not contain any atoms, but $\mu$ has a single atom at its unique halfspace median (the smallest non-empty central region~\eqref{central region}). Because of the very special position of the atom of $\mu$, it is impossible to use our results from parts~\ref{support1} and~\ref{jump} of Theorem~\ref{theorem:support} to decide whether the origin is an atom of $\mu$, or not. \end{example} \section{Conclusion} The halfspace depth has many applications, for example in classification or in nonparametric goodness-of-fit testing. However, in order to apply it properly, one needs to make sure that the measure $\mu$ is characterized by its halfspace depth function, so that we can use the halfspace depth to distinguish $\mu$ from other measures. For that reason, it is important to know which collections of measures satisfy this property. The partial reconstruction procedure provided in this paper may be used to narrow down the set of all possible measures that correspond to a given halfspace depth function. That can be used to guide the selection of an appropriate tool of depth-based analysis. The problem of determining those distributions that are uniquely characterized by their halfspace depth, however, remains open.

https://arxiv.org/abs/2009.08573

Detection of Change Points in Piecewise Polynomial Signals Using Trend Filtering

While many approaches have been proposed for discovering abrupt changes in piecewise constant signals, few methods are available to capture these changes in piecewise polynomial signals. In this paper, we propose a change point detection method, PRUTF, based on trend filtering. By providing a comprehensive dual solution path for trend filtering, PRUTF allows us to discover change points of the underlying signal for either a given value of the regularization parameter or a specific number of steps of the algorithm. We demonstrate that the dual solution path constitutes a Gaussian bridge process that enables us to derive an exact and efficient stopping rule for terminating the search algorithm. We also prove that the estimates produced by this algorithm are asymptotically consistent in pattern recovery. This result holds even in the case of staircases (consecutive change points of the same sign) in the signal. Finally, we investigate the performance of our proposed method for various signals and then compare its performance against some state-of-the-art methods in the context of change point detection. We apply our method to three real-world datasets including the UK House Price Index (HPI), the GISS surface Temperature Analysis (GISTEMP) and the Coronavirus disease (COVID-19) pandemic.

\section{Discussion} \label{sec:discussion.proj1} This paper proposed an algorithm, PRUTF, to detect change points in piecewise polynomial signals using trend filtering. We demonstrated that the dual solution path produced by the PRUTF algorithm forms a Gaussian bridge process for any given value of the regularization parameter $\lambda$. This conclusion allowed us to derive an efficient stopping rule for terminating the search algorithm, which is vital in change point analysis. We then proved that when there is no staircase block in the signal, the method guarantees consistent pattern recovery. However, it fails in doing so when there is a staircase in the underlying signal. To address this shortcoming in such a case, we suggested a modification in the procedure of constructing the solution path, which effectively prevents false discovery of change points. Evidence from both simulation studies and real data analysis reveals the accuracy and the high detection power of the proposed method. \chapter*{APPENDICES} \section{Introduction} \label{sec:introduction.proj1} The problem of change point detection is more than sixty years old. It was first studied by \cite{page1954continuous,page1955test}, and since then, has been of interest to many scientists including statisticians. Many of the earlier developments concerned the existence of at most one change point; however, considerable attention in recent years has been given to multiple change point analysis, which has found applications in many fields such as finance and econometrics, \cite{bai2003computation, hansen2001new}, bioinformatics and genomics, \cite{futschik2014multiscale, lavielle2005using}, climatology, \cite{liu2010impacts, pezzatti2013fire}, and technology, \cite{siris2004application, oudre2011segmentation, lung2012distributed, ranganathan2012pliss, galceran2017multipolicy}. Consequently, there is a vast and rich literature on the subject. In the following, we only review a body of literature on a retrospective change point framework closely related to our work and refer the interested readers to \cite{eckley2011analysis, lee2010change, horvath2014extensions, truong2018review} for more comprehensive reviews. We consider the univariate signal plus noise model \begin{align}\label{fmodel.proj1} y_{_i}=f_{_i}+\varepsilon_{_i}, \qquad\qquad i=1,\,\ldots,\,n, \end{align} where $f_{_i}=f(i/n)$ is a deterministic and unknown signal with equally spaced input points over the interval $[0,\,1]$. The error terms $\varepsilon_{_1},\,\ldots,\,\varepsilon_{_n}$ are assumed to be independently and identically distributed Gaussian random variables with mean zero and finite variance $\sigma^2$. We assume that $f(\cdot)$ undergoes $J_{_0}$ unknown and distinct changes at point fractions $0=\omega_{_0}<\omega_{_1}< \ldots< \omega_{_{J_0}}< \omega_{_{J_0+1}}=1$, where the number of change point fractions, $J_{_0}$ can grow with the sample size $n$. Additionally, we assume that $f(\cdot)$ is a piecewise polynomial function of order $r \in \mbb N$. These assumptions imply that, associated with $\omega_{_0}, \ldots, \omega_{_{J_0+1}}$, there are change points locations $0=\tau_{_0}<\tau_{_1}< \ldots< \tau_{_{J_0}}< \tau_{_{J_0+1}}=n$, which partition the entire signal $\mbf f=(f_{_1}, \ldots, f_{_n})$ into $J_0+1$ segments. More specifically, any subsignal of $\mbf f$ within segments created by the change points follows an $r$-degree polynomial structure with or without a continuity constraint at the change points. For more detail, see Figure \ref{fig:coor-removal}. Change in the level of a piecewise constant signal, known as the canonical multiple change point problem, and change in the slope of a piecewise linear signal are examples of the problem under consideration in this paper. In change point analysis, the objective is to estimate the number of change points, $J_{_0}$, as well as their locations $\bsy \tau=\{\tau_{_1},\,\ldots,\,\tau_{_{_{J_0}}}\}$ based on the observations $\mbf y=(y_{_1},\,\ldots,\,y_{_n})$. The canonical multiple change point problem, where the signal $\mbf f$ is modelled as a piecewise constant function, has been extensively studied in the literature. In this framework, there are many approaches and we only attempt to list a selection of them here. The majority of these techniques seek to identify all change points at once by solving an optimization problem consisting of a loss function, often the negative log-likelihood, and a penalty criterion. \cite{yao1988estimating, yao1989least} used the square error loss along with the Schwarz Information Criterion (SIC) as a penalty function to consistently estimate the bounded number of change points and their locations for the data drawn from a Gaussian distribution. Within the same setting, incorporation of various penalty functions including Modified Information Criterion (MIC) \cite{pan2006application}, modified Bayesian Information Criterion (mBIC) \cite{zhang2007modified}, Simultaneous Information Theoretic Criterion (SITC) \cite{wu2008simultaneous} and modified SIC \cite{ciuperca2011general, ciuperca2014model}, have been studied. Specific algorithms such as Optimal Partitioning \cite{auger1989algorithms}, Segment Neighbourhood \cite{jackson2005algorithm}, and pruning approaches such as PELT \cite{killick2012optimal} and PDPa \cite{rigaill2015pruned} are developed to solve such optimization problems. Apart from penalty-based techniques, another frequently used class of change point detection approaches encompasses the greedy search procedures in which they search sequentially for one single change point at a time. The most popular methods in this class are Binary Segmentation \cite{vostrikova1981detecting} and its variants such as Circular Binary Segmentation (CBS) \cite{olshen2004circular}, and Wild Binary Segmentation (WBS) \cite{fryzlewicz2014wild}. In recent years, researchers have attempted to improve Binary Segmentation's performance from statistical and computational viewpoints. \cite{fryzlewicz2018tail} suggested a backward (bottom-up) mechanism, called Tail Greedy Unbalanced Haar (TGUH), which is computationally fast and statistically consistent in estimating both the number and the locations of change points. Also, \cite{fryzlewicz2018detecting} introduced Wild Binary Segmentation 2 (WBS2) to deal with the shortcomings of WBS in datasets with frequent changes. It has been shown that the method is fast in run time and accurate in detection. Beyond the canonical change point problem, signals in which $f$ is modelled as a piecewise polynomial of order $r\geq 1$ have attracted less attention in the literature despite many applications. For instance, piecewise linear signals are applied in monitoring patient health (\cite{aminikhanghahi2017survey}, \cite{stasinopoulos1992detecting}), climate change (\cite{robbins2011changepoints}), and finance (\cite{bianchi1999comparison}). In such a framework, \cite{bai1997estimating} introduced a method based on Wald-type sequential tests, and \cite{maidstone2017optimal} devised a dynamic programming applied to an $\ell_{_0}$-penalized least square procedure. In continuous piecewise linear models, \cite{kim2009ell_1} developed a methodology called $\ell_{_1}$-trend filtering. Furthermore, \cite{baranowski2019narrowest} put forward the method of Narrowest Over Threshold (NOT), and \cite{anastasiou2019detecting} developed an approach called Isolate-Detect (ID) which both provide asymptotically consistent estimators of the number and locations of change points. Our goal in this paper is to introduce a unifying method covering the canonical change point problem and beyond. More precisely, the method is cable of detecting change points in piecewise polynomial signals of order $r$ ($r=0,\,1,\,2,\,\ldots$) with and without continuity constraint at the locations of change points. The detection of change points in a sequence of data can be formulated as a penalized regression fitting problem. According to our notation, the quantity $f_{\tau}-f_{\tau+1}$ is nonzero if the signal $ f$ undergoes a change at point $\tau$, and is zero otherwise. Moreover, if we assume that change points are sparse, that is, the number of locations where $f$ changes, $J_{_0}$, is much smaller than the number of observations $n$, change points can be estimated using the one-dimensional fused lasso problem \begin{align* \min_{\mbf f\in \mathbb{R}^n} \frac{1}{2}\, \big\|\,\mbf y- \mbf f \, \big\|_2^2 \,+\, \lambda\, \ssum{1}{n-1} \big| f_{_{i+1}}-f_{_i} \big|\,, \end{align*} where $\mbf f=(f_{_1},\,\ldots,\,f_{_n})$. This formulation of the canonical change point problem was first considered in \cite{huang2005detection} and was applied to analyze a DNA copy number dataset. \cite{harchaoui2010multiple} considered the same formulation and proved the consistency of the respective change point estimates when the number of change points is bounded. Employing sparse fused lasso which is composed of both the $\ell_{_1}$-norm and the total variation seminorm penalties, \cite{rinaldo2009properties} proposed a sparse piecewise constant fit and established the consistency of the corresponding estimates when the variance of the noise terms vanishes, and the minimum magnitude of jumps is bounded from below. However, \cite{rojas2014change} argued that the consistency results achieved by \cite{rinaldo2009properties} are incorrect when a frequently viewed pattern, called {\it staircase}, exists in the signal. The staircase phenomenon occurs in a piecewise constant model when there are either two consecutive downward jumps or upward jumps in its mean structure. The staircase pattern will be discussed in more detail in Section \ref{sec:pattern.recovery.proj1}. Additionally, \cite{qian2016stepwise} showed that the lasso problem of \cite{tibshirani1996regression} when derived by transforming fused lasso does not satisfy the Irrepresentable Condition (\cite{zhao2006model}) that is necessary and sufficient for exact pattern recovery. In particular, \cite{qian2016stepwise} proposed an approach called preconditioned fused lasso based on the puffer transformation of \cite{jia2015preconditioning} and established that it can recover the exact pattern with probability approaching one. A similar approach to that of the piecewise constant signals can be considered for estimating change points in piecewise polynomial signals. In particular, a positive integer $\tau$ is a change location in an $r$-th degree piecewise polynomial signal $f$ if $\tau$-th element of the vector $\mbf D^{(r+1)}\, \mbf f$ is non-zero, denoted by $[\,\mbf D^{(r+1)}\, \mbf f\,]_{\tau}\neq 0$. Here $\mbf D^{(r+1)}$ is a penalty matrix that will be defined in Section \ref{sec:dual.tf.proj1}. Hence, change points can be estimated from nonzero elements of $\mbf D^{(r+1)}\, \widehat{\mbf f}$, where $\widehat{\mbf f}$ is the solution of \begin{align}\label{tf.obj.proj1} \min_{\mbf f\in \mathbb{R}^{^n}} \frac{1}{2}\|\,\mathbf{y}-\mbf f \,\|_{_2}^2+\lambda \,\|\mathbf{D}^{(r+1)}\mbf f \|_{_1}\,. \end{align} The aforementioned problem was first studied by \cite{steidl2006splines} in the context of image processing and was called {\it higher order total variation regularization}. It was later rediscovered by \cite{kim2009ell_1} and termed {\it trend filtering} in the nonparametric regression setting. \cite{kim2009ell_1} specifically explored linear trend filtering ($r=1$) which fits piecewise linear models. \cite{tibshirani2014adaptive} extensively studied trend filtering and compared its performance with smoothing splines \cite{green1993nonparametric} and locally adaptive regression splines \cite{mammen1997locally} in the context of nonparametric regression. \cite{tibshirani2014adaptive} also established that trend filtering enjoys desirable and strong theoretical properties of locally adaptive regression splines while being computationally less intensive due to its banded penalty matrix. Moreover, trend filtering has an adaptive knot selection property, which makes it well suited for change point analysis. From a computational and algorithmic standpoint, \cite{kim2009ell_1} described Primal-Dual Interior Point (PDIP) method for deriving the estimates of the linear trend filtering problem at a fixed value of $\lambda$. This can be easily carried over to the trend filtering problem of any order. \cite{wang2014falling} suggested an algorithm based on a falling factorial basis while \cite{ramdas2016fast} derived an algorithm based on the Alternating Direction Method of Multipliers (ADMM) discussed in \cite{boyd2011distributed}. The computational complexity of all these algorithms is of order $O(n)$. In this paper, we develop a new methodology called {\it Pattern Recovery Using Trend Filtering} (PRUTF) for identifying unknown change points in piecewise polynomial signals with no continuity restriction at change point locations. Therefore, a change point is defined as a sudden jump in the signal and its all derivatives up to order $r$. Figure \ref{fig:coor-removal} displays such change points for various $r$. In this paper, we make the following contributions. \begin{itemize} \item We propose a generic dual solution path algorithm along with the regularization parameter for trend filtering. This solution path, whose basic idea is borrowed from \cite{tibshirani2011solution} enables us to determine change points at each level of the regularization parameter. Our algorithm, PRUTF, is different from that of \cite{tibshirani2011solution} as we remove $(r+1)$ coordinates of dual variables after identifying each change point. This adjustment to the algorithm allows us to have independent dual variables between each pair of neighbouring change points. Besides, the elimination of $(r+1)$ coordinates at each step leads to faster implementation of the algorithm. \item We establish a stopping criterion that plays an essential rule in the PRUTF algorithm used to find change points. Notably, we show that the dual variables of trend filtering between consecutive change points constitute a Gaussian bridge process. This finding allows us to introduce a threshold for terminating our proposed algorithm. \item If the signal contains a staircase pattern, we prove that the method is statistically inconsistent, which makes it unfavourable. Explaining the reason for this end, we modify the PRUTF algorithm to produce estimates consistent in terms of both the number and location of change points. \end{itemize} This paper is organized as follows: we first describe how to characterize the dual optimization problem of trend filtering. In Section \ref{sec:solution-path-algorithm}, we develop our main algorithm, PRUTF, to use in constructing the dual solution path of trend filtering and, in turn, identifying the locations of change points. Section \ref{sec:property.solution.path.proj1} discusses the properties of this dual solution path. We establish that the dual variables derived from the solution path form a Gaussian bridge process that makes them favourable for statistical inference. Applying these properties, we develop a stopping rule for the change point search algorithm in Section \ref{sec:stop.rule.proj1}. The quality of the PRUTF algorithm is validated in terms of pattern recovery of the true signal in Section \ref{sec:pattern.recovery.proj1}. It is established that the proposed technique in its naive form fails to consistently identify the true signal when a special pattern, called staircase, is present in the signal. Section \ref{sec:modified.trend.filtering.proj1} elaborates on how to modify the algorithm in order to estimate the true pattern consistently. Simulation results, and real-world applications are presented in Section \ref{sec:simulation.proj1}. We explore the performance of our proposed method for signals with frequent change points as well as models with dependent error terms in Section \ref{sec:model_misspecification.proj1}. We conclude the paper with a discussion in Section \ref{sec:discussion.proj1}. \section{Notations and Fundamental Concepts}\label{sec:notations.concepts.proj1} \subsection{Notations}\label{sec:notations.proj1} We begin this section with setting up notations that will be used throughout this article. For an $m\times n$ matrix $\mbf A$, we denote its rows by $\mbf A_{_1},\ldots,\mbf A_{_m}$ and express the matrix as $\mbf A=(\mbf A_{_1}^{^T},\ldots,\mbf A_{_m}^{^T})^T$. Now for the set of indices $\mca I=\{i_{_1},\,\ldots,\,i_{_k}\}\subseteq\{1,\,\ldots,\,m\}$, the notation $\mbf A_{_{\mca I}}=(\mbf A_{i_{_1}}^{^T},\,\ldots,\,\mbf A_{i_{_k}}^{^T})^T$ represents the submatrix of $\mbf A$ whose row labels are in the set $\mca I$. In a similar manner, for a vector $\bsy a$ of length $m$, we let $\bsy a_{_{\mca I}}=( a_{_{i_{_1}}},\ldots, a_{_{i_{_k}}})^{^T}$ denote a subvector of $\bsy a$ whose coordinate labels are in $\mca I$. We write $\mbf A_{_{-\mca I}}$ and $\bsy a_{_{-\mca I}}$ to denote $\mbf A_{_{\{1,\,\ldots,\,m\} \backslash \mca I}}$ and $\bsy a_{_{\{1,\,\ldots,\,m\} \backslash \mca I}}$\,, respectively, where $\mca J\backslash\mca I$ is the set of indices in $\mca J$ but not in $\mca I$. Furthermore, for selecting $i$-th row of $\mbf A$, the notation $[\mbf A]_i$ and for its $(i,j)$-th element the notation $[\mbf A]_{ij}$ are used. Also, $[\bsy a]_i$ extracts the $i$-th elements of the vector $\bsy a$. We write $\diag{\mbf A}$ to denote the vector of the main diagonal entries of the matrix $\mbf A$. Moreover, for a real number $x$, $\lfloor x\rfloor$ denotes the greatest integer less than or equal $x$, and $\lceil x \rceil$ denotes the least integer greater or equal $x$. For a set $A$, the indicator function is denoted by $\mathbbm{1}(A).$ \subsection{The Dual Problem of Trend Filtering} \label{sec:dual.tf.proj1} Recall the trend filtering problem \begin{align}\label{tf2.obj.proj1} \min_{\mbf f\in \mathbb{R}^{^n}} \frac{1}{2}\|\mathbf{y}-\mbf f \|_{_2}^2+\lambda \|\mathbf{D}^{(r+1)}\mbf f \|_{_1}, \end{align} where $\lambda\geq 0$ is the regularization parameter for controlling the effect of smoothing, and the $(n-r-1)\times n$ penalty matrix $\mathbf{D}^{(r+1)}$ is the difference operator of order $(r+1)$. For $r=0$, the first order difference matrix $\mathbf{D}^{(1)}$ is defined as \begin{align*} \mathbf{D}^{(1)}=\begin{pmatrix} -1 & 1 & 0 & \ldots & 0 & 0 \\ 0 & -1 & 1 & \ldots & 0 & 0 \\ \vdots & & & & & \vdots \\ 0 & 0 & 0 & \ldots & -1 & 1 \\ \end{pmatrix}, \end{align*} and for $r\geq 1$, the difference operator of order $r+1$ can be recursively computed by $\mathbf{D}^{(r+1)}=\mathbf{D}^{(1)}\times \mathbf{D}^{(r)}$. Notice that, in this matrix multiplication, we consider only the submatrix consisting of the first $n-r-2$ rows and $n-r-1$ columns of the matrix $\mathbf{D}^{(1)}$. Figure \ref{fig:tf-splines} displays the trend filtering fits for $r=0,1,2$ for simulated data. \begin{figure}[!ht] \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=1\linewidth]{img/sigspln1.jpg} \caption{Linear} \end{subfigure} \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=1\linewidth]{img/sigspln2.jpg} \caption{Quadratic} \end{subfigure} \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=1\linewidth]{img/sigspln3.jpg} \caption{Cubic} \end{subfigure} \caption[]{Trend filtering solutions for $r=1,\,2,\,3$ producing (a) piecewise linear, (b) piecewise quadratic and (c) piecewise cubic fits, respectively.} \label{fig:tf-splines} \end{figure} Although the objective function in the $r$-th order trend filtering \eqref{tf2.obj.proj1} is strictly convex and thus the minimization has a guaranteed unique solution, the penalty term is not differentiable in $\mbf f $, so solving the optimization in its current form is difficult. To overcome this difficulty, we follow the argument in \cite{tibshirani2011solution} and convert this optimization problem into its dual form. Since the objective function in the primal problem is strictly convex with no constraint, the strong duality holds, meaning that the primal and the dual solutions coincide. The trend filtering problem \eqref{tf2.obj.proj1} can be rewritten as \begin{equation*} \min_{\mbf f \in \mathbb{R}^{^n}} \frac{1}{2}\|\mathbf{y}-\mbf f \|_{_2}^2+\lambda \|\mathbf{z}\|_{_1}, \quad \text{subject to } \mathbf{z}=\mathbf{D}\mbf f \,, \end{equation*} where, for ease in the notation, we use $\mathbf{D}=\mathbf{D}^{(r+1)}$. For any given $\lambda>0$, the Lagrangian is \begin{align*} \mathcal{L}(\mbf f ,\,\mathbf{z},\,\mathbf{u})&=\frac{1}{2}\|\mathbf{y}-\mbf f \,\|_{_2}^2+\lambda \|\mathbf{z}\|_{_1}+\mathbf{u}^T(\mathbf{D}\mbf f -\mathbf{z}) \end{align*} and, thus the dual function is given by \begin{equation*} g(\mathbf{u})= \inf_{\mbf f \in \mathbb{R}^{^n},\,\mathbf{z}\in\mathbb{R}^{^m}}\mathcal{L}(\mbf f ,\,\mathbf{z},\,\mathbf{u}), \end{equation*} which is a concave function defined on $\mathbb{R}^{^m}$, where $m=n-r-1$ and takes values in the extended real line $\mathbb{R}\cup\lbrace -\infty,\,\infty\rbrace$. The vectors $\mbf f $ and $\mathbf{u}$ are called the primal and dual variables, respectively. Taking the derivative of the Lagrangian $\mathcal{L}(\mbf f ,\,\mathbf{z},\,\mathbf{u})$ with respect to $\mbf f $ and setting it to be equal to zero, we obtain \begin{align}\label{primal.dual.exact.proj1} \mbf f =\mathbf{y-D}^T\mathbf{u}. \end{align} Now substituting this back into the Lagrangian $\mathcal{L}(\mbf f ,\,\mathbf{z},\,\mathbf{u})$, and performing certain algebraic manipulations, we obtain \begin{align*} \mathcal{L}^\ast(\mathbf{z},\,\mathbf{u})&=\inf_{\mbf f \in \mathbb{R}^{^n}}\mathcal{L}(\mbf f ,\,\mathbf{z},\,\mathbf{u})\\ &=-\frac{1}{2}\|\mathbf{y}-\mathbf{D}^T\mathbf{u}\|_{_2}^2+\frac{1}{2}\|\mathbf{y}\|^2+\lambda \|\mathbf{z}\|_{_1}-\mathbf{u}^T\mathbf{z}\,. \end{align*} Minimizing $\mathcal{L}^\ast(\mathbf{z},\,\mathbf{u})$, or equivalently maximizing $\mathbf{u}^T\mathbf{z}-\lambda \|\mathbf{z}\|_{_1}$, with respect to $\mathbf{z}\in\mathbb{R}^{^m}$ leads us to the dual function $g(\mathbf{u})$. Notice that $\sup\limits_{\mathbf{z}}\lbrace\mathbf{u}^T\mathbf{z}-\lambda \|\mathbf{z}\|_{_1}\rbrace$ is the conjugate of the function $\lambda \|\mathbf{z}\|_{_1}$ in the context of conjugate convex functions. See \cite{brezis2010functional} and \cite{boyd2004convex}. This conjugate function is given by \begin{align*} \sup\limits_{\mathbf{z}}\lbrace\mathbf{u}^T\mathbf{z}-\lambda \|\mathbf{z}\|_{_1}\rbrace=\begin{cases} 0 & \text{ if }\|\mathbf{u}\|_{_\infty}\le \lambda\\ \infty & \text{ otherwise\,.} \end{cases} \end{align*} From all these, the dual function is given as \begin{align*} g(\mathbf{u})= -\frac{1}{2}\|\mathbf{y}-\mathbf{D}^T\mathbf{u}\|_{_2}^2+\frac{1}{2}\|\mathbf{y}\|^2 \quad \textrm{ for }\quad \|\mathbf{u}\|_{_\infty}\leq\lambda\,, \end{align*} and, thus the dual problem is to find the maximum of the dual function $g(\mathbf{u})$. This is equivalent to \begin{equation}\label{tf.dual.obj.proj1} \min\limits_{\mathbf{u}\in \mathbb{R}^{^m}} \frac{1}{2}\|\mathbf{y}-\mathbf{D}^T\mathbf{u}\|_{_2}^2 \quad \textrm{subject to }\quad \|\mathbf{u}\|_{_\infty}\leq\lambda\,. \end{equation} The constraint in \eqref{tf.dual.obj.proj1} is an $\ell_{_\infty}$-ball or a hypercube centered at the origin with the boundaries given by the set $\lbrace -\lambda,\,+\lambda\rbrace^{m}$. Since the matrix $\mbf D$ is full row rank, the problem \eqref{tf.dual.obj.proj1} is strictly convex and has a unique solution, see \cite{ali2019generalized}. In addition, notice that the dimension of the dual vector $\mbf u$ is $m$, which is smaller than that of the primal vector $\mbf f $ and may lead to relatively faster computations. The connection between the primal and the dual solutions is given by the equations \begin{align}\label{primal.to.dual} \hspace{-1.45cm}\widehat{\mbf u}_{_\lambda}=\lambda \,\widehat{\bsy\gamma}, \end{align} \vspace{-1.2cm} \begin{align}\label{dual.to.primal} \widehat{\mbf f }_{_\lambda}=\mathbf{y}-\mathbf{D}^T\widehat{\mathbf{u}}_{_\lambda}\,, \end{align} where $\widehat{\bsy\gamma} \in \mathbb{R}^{^m}$ is a subgradient of $\| \mbf x\|_{_1}$ computed at $\mbf x=\mbf D\widehat{\mbf f }_{_\lambda}$. This subgradient is given by \begin{align}\label{gamma.subgrad} \widehat{\gamma}_{_i}\in \left\{ \begin{array}{lcl} \{+1\} & \textrm{if} & [\mathbf{D\widehat{\mbf f }_{_\lambda}}]_i>0 \\ \{-1\} & \textrm{if} & [\mathbf{D\widehat{\mbf f }_{_\lambda}}]_i<0 \\ \,[-1,+1] & \textrm{if} & [\mathbf{D\widehat{\mbf f }_{_\lambda}}]_i=0\,. \end{array} \right. \end{align} The statements in Equations \eqref{primal.to.dual}-\eqref{gamma.subgrad} are equivalent to the KKT optimality conditions of the primal problem \eqref{tf2.obj.proj1}. The dual problem \eqref{tf.dual.obj.proj1} demonstrates that $\mbf D^T\widehat{\mbf u}_{_\lambda}$ is the projection, $P_{_{\mbb C}}(\mbf y)$, of $\mathbf{y}$ onto the convex polyhedron (or hypercube here) $\mathbb{C}=\{\mbf x\in \mathbb{R}^{^m}:\, \|\mbf x\|_{_\infty}\leq \lambda\}\,$. From this, the primal solution \eqref{dual.to.primal} can be rewritten in the form of $(\mbf I- P_{_{\mathbb{C}}})\,(\mbf y)$, representing the residual projection map of $\mbf y$ onto the polyhedron $\mathbb{C}$. Our idea of applying trend filtering to discover change points in piecewise polynomial signals is inspired by \cite{rinaldo2009properties} and its correction \cite{rinaldo2014corrections}, in which change point detection is studied using fused lasso. Besides extending to piecewise polynomial signals, the novelty of our work is in providing an exact stopping criterion, which is based on the Gaussian bridge property of the trend filtering dual variables. In addition, we propose an algorithm which, unlike that proposed in \cite{rinaldo2009properties}, always produces consistent change points even in the presence of staircase patterns. \section{Solution Path of Trend Filtering and PRUTF Algorithm}\label{sec:solution-path-algorithm} In this section, we construct and study the solution path of dual variables $\widehat{\mbf u}_{\lambda}$ as the regularization parameter decreases from $\lambda=\infty$ to $\lambda=0$. In the following, the PRUTF algorithm is given to compute the entire dual solution path. This dual solution path identifies the corresponding primal solution using \eqref{dual.to.primal}. For any given $\lambda$, we call any coordinate of $\widehat{\mbf u}_{\lambda}$ a boundary coordinate if it is a vertex of the polyhedron $\mathbb{C}= \big\{ \mbf x\in \mathbb{R}^{^m}:\, \|\mbf x\|_{_\infty}\leq \lambda \big\}\,$, meaning that its absolute value becomes $\lambda$. In the process of constructing the solution path, for any $\lambda$, we trace several sets, introduced below. \begin{itemize} \item The set $\mca B=\mca B(\lambda)$, called the boundary set, contains the boundary coordinates identified by $\widehat{\mbf u}_{\lambda}$. \item The vector $\mbf s_{\mca B}=\mbf s_{\mca B}(\lambda)$, called the sign vector, represents collectively the signs of the boundary points in $\mca B(\lambda)$. \item The set $\mca A=\mca A(\lambda)$, called the augmented boundary set, contains the boundary coordinates in $\mca B(\lambda)$ as well as the first $r_{_a}=\lfloor (r+1)/2\rfloor$ coordinates immediately after. \item The vector $\mbf s_{\mca A}=\mbf s_{\mca A}(\lambda)$ represents collectively the signs of the augmented boundary points in $\mca A(\lambda)$. \end{itemize} In the following, we discuss the need for the augmented boundary set $\mca A$. We begin by studying the structure of the dual vector $\mbf u=\mbf D\mbf f$ in a piecewise polynomial signal of order $r$, where the signal is partitioned into a number of blocks defined by the position of the change points. Because the signal $f$ is a piecewise polynomial of order $r$, to compute the $i$-th coordinate of the vector $\mbf{u}$, we need $r_{_b}=\lceil (r+1)/2\rceil-1$ points directly before the $i$-th element of $\mbf{f}$ as well as $r_{_a}=\lfloor (r+1)/2\rfloor$ points immediately after that. Consequently, the first $r_{_a}$ elements of $\mbf D\mbf f$ within each block cannot be computed. Moreover, within each block, the last $r_{_b}+1$ elements of $\mbf D\mbf f$ are all nonzero due to the existence of a change point. This observation is depicted in Figure \ref{fig:coor-removal} for $r=0,\, 1,\, 2,\, 3$. To explain this point clearly, consider the case of $r=2$ in Figure \ref{fig:coor-removal} in which the structure of $\mbf {Df}$ is shown, where the true change points are at $6$ and $13$. As can be seen, the points on the boundary -- the nonzero coordinates of $\mbf{Df}$ -- are $\mca B(\lambda)= \{ 5,\,6,\,12,\,13\}$ with their respective signs $\mbf s_{\mca B}(\lambda)=\{ 1,\,1,\,-1,\,-1 \}$. Notice that $\mbf{Df}$ does not exist at points 7 and 14. The augmented boundary set contains these points as well as the boundary points; that is $\mca A(\lambda)= \{5,\,6,\,7,\,12,\,13,\,14\}$. The respective signs of the coordinates in the augmented boundary set $\mca A(\lambda)$ are given by $\mbf s_{\mca A}(\lambda)=\{1,\,1,\,1,\,-1,\,-1,\,-1\}$. At each value of $\lambda$, we call the coordinates that belong to the augmented boundary set $\mca A(\lambda)$ the augmented boundary coordinates, and the rest, the interior coordinates. \begin{figure}[!t] \begin{center} \begin{subfigure}{.43\textwidth} \centering \includegraphics[width=1\linewidth]{img/removal0.jpg} \caption{Piecewise constant, $r=0$.} \end{subfigure} \begin{subfigure}{.43\textwidth} \centering \includegraphics[width=1\linewidth]{img/removal1.jpg} \caption{Piecewise linear, $r=1$.} \end{subfigure} \\ \begin{subfigure}{.43\textwidth} \centering \includegraphics[width=1\linewidth]{img/removal2.jpg} \caption{Piecewise quadratic, $r=2$.} \end{subfigure} \begin{subfigure}{.43\textwidth} \centering \includegraphics[width=1\linewidth]{img/removal3.jpg} \caption{Piecewise cubic, $r=3$.} \end{subfigure} \caption[Structure of $\mbf D\mbf f$ For Piecewise Polynomial Signals]{Structure of $\mbf D\mbf f$ for piecewise polynomial signals with various orders $r=0,\,1,\,2,\,3$. The olive lines display the true signals with two change points at the locations $6$ and $13$. Empty circles represent the indices that $\mbf D\mbf f$ does not exist.} \label{fig:coor-removal} \end{center} \end{figure} At the $j$-th iteration with $\lambda=\lambda_{_j}$, we assume that the boundary set and its corresponding sign vector are $\mca B=\mca B(\lambda)$ and $\mbf s_{\mca B}=\mbf s_{\mca B}(\lambda)$, respectively. Furthermore, we assume the augmented boundary set and its sign vector are $\mca A=\mca A(\lambda)$ and $\mbf s_{\mca A}=\mbf s_{\mca A}(\lambda)$, respectively. Dual coordinates can be split into augmented boundary coordinates $\widehat{\mathbf{u}}_{_{\lambda_j,\,\mathcal{A}}}$ and interior coordinates $\widehat{\mathbf{u}}_{_{\lambda_j,\,-\mathcal{A}}}$. Recall from Section \ref{sec:notations.proj1} that $\widehat{\mathbf{u}}_{_{\lambda_j,\,\mathcal{A}}}$ represents the subvector of $\widehat{\mbf u}_{_\lambda}$ with the coordinate labels in the set $\mca A$ and $\widehat{\mathbf{u}}_{_{\lambda_j,\,-\mathcal{A}}}$ represents the subvector of $\widehat{\mbf u}_{_\lambda}$ with the coordinate labels in the set $\lbrace 1,\,2,\, \cdots, \,m\rbrace \backslash \mca A$. It is apparent from the definition of the boundary coordinates that \begin{align}\label{u.boundary} \widehat{\mathbf{u}}_{_{\lambda_j,\,\mathcal{A}}}=\lambda_{_j}\, \mathbf{s}_{\mathcal{A}}\,. \end{align} Replacing the boundary coordinate with $\lambda_{_j}\,\mbf s_{\mca A}$ in \eqref{tf.dual.obj.proj1} and solving the resulting quadratic problem with respect to the interior coordinates, lead to their least square estimates, given by \begin{align}\label{u.interior} \widehat{\mathbf{u}}_{_{\lambda_j,\mathcal{-A}}}&=\left(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\right)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\left(\mathbf{y}-\lambda_{_j} \mathbf{D}_{_{\mathcal{A}}}^T~\mathbf{s}_{\mathcal{A}}\right). \end{align} It should be noted that for the purpose of simplicity, we denote $(\mbf D_{\mca A})^T$ and $(\mbf D_{_{-\mca A}})^T$ with $\mbf D_{\mca A}^T$ and $\mbf D_{_{-\mca A}}^T$, respectively. Notice that in \eqref{u.interior}, the first term $\big( \mathbf{D}_{_{-\mathcal{A}}} \mathbf{D}_{_{-\mathcal{A}}}^T \big)^{-1} \mathbf{D}_{_{-\mathcal{A}}}\, \mathbf{y}$ simply yields the least square estimate of regressing the response vector $\mbf y$ on the design matrix $\mathbf{D}_{_{-\mathcal{A}}}$. The second term $-\lambda_{_j}\, \big( \mathbf{D}_{_{-\mathcal{A}}} \mathbf{D}_{_{-\mathcal{A}}}^T \big)^{-1} \mathbf{D}_{_{-\mathcal{A}}}\, \mathbf{D}_{_{\mathcal{A}}}^T~ \mathbf{s}_{\mathcal{A}}$ can be interpreted as a shrinkage term due to the condition $\|\mbf u\|_{_\infty}\leq \lambda$. The expression \eqref{u.interior} is true for $\lambda\leq \lambda_{_j}$ until either an interior coordinate joins to the boundary or a coordinate in the boundary set leaves the boundary. The following argument explains how to specify values of $\lambda$ while the interior coordinates change. We define the joining time associated with the interior coordinate $i\in \lbrace 1,\,2,\, \cdots, \,m\rbrace \backslash \mca A$ as the time at which this interior coordinate joins the boundary. To determine the next joining time, we reduce the value of $\lambda$ in a linear direction starting from $\lambda_{_j}$ and solve $\widehat{\mathbf{u}}_{_{\lambda,\mathcal{-A}}}=(\pm\lambda,\,\cdots,\,\pm\lambda)^T$. Note that the right-hand side of \eqref{u.interior} can be expressed as $\bsy a-\lambda_{_j}\,\mbf b$, where \begin{align}\label{ab} \bsy a&=\big(\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T\big)^{-1}\mathbf{D}_{_{-\mca A}}\,\mathbf{y}\,,\\[8pt] \mathbf{b}&=\big(\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T\big)^{-1}\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{\mca A}}^T\,\mathbf{s}_{_{\mca A}}\,. \end{align} The joining time for every $i\in \lbrace 1,\,2,\, \cdots, \,m\rbrace \backslash \mca A\,$ is hence the solution of the equation $a_{_i}-\lambda \,b_{_i}=\pm\lambda$ with respect to $\lambda$, which is given by \begin{align*} \lambda_{_i}^{^\textrm{join}}=\frac{a_{_i}}{b_{_i}\pm 1}\,, \qquad\qquad i\in\lbrace 1,\,2,\, \cdots, \,m\rbrace \backslash \mca A\,. \end{align*} Note that $\lambda_{_i}^{^\textrm{join}}$ is uniquely defined because only one of the signs $-1$ or $+1$ yields $\lambda_{_i}\in [0,\, \lambda_{_j}]$. Now we turn the attention to the characterization of a coordinate which leaves the boundary set $\mca B$. For $i\in \mca B$, the leaving time is defined as the time that the coordinate $i$ leaves the boundary set $\mca B$. Since $\mbf s_{\mca B}$ is the sign vector of changes captured by $\big[ \mbf D\, \widehat{\mbf f} \big]_{\mca B}$, then $\diag{\mbf s_{\mca B}}\, \big[ \mbf D\, \widehat{\mbf f} \big]_{\mca B}> \mbf 0$, which in turn, along with Equation \eqref{dual.to.primal}, implies $\diag{\mbf s_{\mca B}} \big[ \mbf D\, \big(\mbf y-\mbf D^T \widehat{\mbf u}_{_\lambda}\big)\, \big]_{\mca B}> \mbf 0$. Here, for any vector $\bsy{\eta}$, $\diag{\bsy{\eta}}$ denotes the diagonal matrix with the diagonal elements given by $\bsy{\eta}$, and $\bsy{\eta} > \mbf{0}$ holds element-wise. Therefore, a coordinate $i\in\mca B$ leaves the boundary set $\mca B$ if $\diag{\mbf s_{\mca B}} \big[ \mbf D\, \big(\mbf y-\mbf D^T \widehat{\mbf u}_{_\lambda}\big)\, \big]_{\mca B}> \mbf 0$ is violated. Using the relation \begin{align*} \Big[ \mbf D\, \big(\mbf y-\mbf D^T \widehat{\mbf u}_{_\lambda}\big)\, \big]_{\mca B}=\mbf D_{\mca B}\, \big( \mbf y-\mbf D^T \widehat{\mbf u}_{_\lambda} \big)\,, \end{align*} and the decomposition $\mbf D^T \widehat{\mbf u}_{_\lambda}=\mbf D_{\mca A}^T\, \widehat{\mbf u}_{_{\lambda,\,\mca A}}+\mbf D_{_{-\mca A}}^T\, \widehat{\mbf u}_{_{\lambda,-\mca A}}$, we obtain \begin{align} \diag{\mbf s_{\mca B}} \Big[ \mbf D\, \big(\mbf y-\mbf D^T \widehat{\mbf u}_{_\lambda}\big) \Big]_{\mca B}=\mbf c-\lambda\,\mbf d\,, \end{align} where \begin{align}\label{cd} \mathbf{c}&=\mathrm{diag}(\mathbf{s}_{\mathcal{B}})\,\mathbf{D}_{\mathcal{B}}\big(\mathbf{y}-\mathbf{D}_{_{-\mathcal{A}}}^T\,\bsy{a}\big)\,,\\[8pt] \mathbf{d}&=\mathrm{diag}(\mathbf{s}_{\mathcal{B}})\,\mathbf{D}_{\mathcal{B}}\big(\mathbf{D}_{\mathcal{A}}^T\,\mathbf{s}_{\mathcal{A}}-\mathbf{D}_{_{-\mathcal{A}}}^T\,\mathbf{b}\big). \end{align} Hence, a leaving time is obtained from the equation $c_{_i}-\lambda\, d_{_i} > 0$ as \begin{align*} \lambda_{_i}^{^\textrm{leave}}=\left\{\begin{array}{lll} \dfrac{c_{_i}}{d_{_i}}, && \textrm{if}~ c_{_i}<0~ \textrm{~and~} ~ d_{_i}<0\,, \\[8pt] 0, && \textrm{otherwise}\,. \end{array}\right. \end{align*} The conditions in the aforementioned equation is due to the fact that at the $j$-th iteration with $\lambda\leq \lambda_{_j}$, the expression $c_{_i}-\lambda\, d_{_i} > 0$ fails for $i\in \mca B$, if both $c_{_i}$ and $d_{_i}$ are negative. An alternative way to determine the next leaving time is to use the KKT optimality conditions of \eqref{tf.dual.obj.proj1}. We refer the reader to the supplementary materials of \cite{tibshirani2011solution}. The following algorithm, PRUTF, describes the process of constructing the entire dual solution path of trend filtering. \begin{algorithm}[PRUTF Algorithm] \label{tf.path.alg} \begin{enumerate} \item[] \item Initialize the set of change points locations as $\bsy\tau_{_0}=\emptyset$, the empty set. \item At step $j=1$, initialize the boundary set $\mathcal{B}_{_1}=\big\{\tau_{_1}-r_{_b},\,\tau_{_1}-r_{_b}+1,\, \ldots,\tau_{_1}\big\}$ and its associated sign vector $s_{_{\mathcal{B}_1}}=\{s_{_1},\ldots,s_{_1}\}$, both with cardinality of $r_{_b}+1$, where $\tau_{_1}$ is obtained by \begin{align}\label{firstjoin} \tau_{_1}=\underset{i=1,\,\ldots,\,m}{\rm argmax} \mid \widehat{u}_{_i}\mid \,, \end{align} and $s_{_1}=\sign{ \widehat{u}_{_{ \tau_{_1}}}}$, where $\widehat{u}_{_i}$ is the $i$-th element of the vector $\widehat{\mathbf{u}}=\left(\mathbf{DD}^T\right)^{-1}\mathbf{D}\,\mathbf{y}$. The updated set of change points locations is now $\bsy\tau_{_1}=\{\tau_{_1}\}$. We also record the first joining time $\lambda_{_1}= \mid\widehat{u}_{\tau_{_1}}\mid$ and keep track of the augmented boundary set $\mathcal{A}_{_1}=\{\tau_{_1}-r_{_b},\ldots,\tau_{_1}+r_{_a}\}$ and its corresponding sign vector $\mathbf{s}_{_{\mathcal{A}_{_1}}}=\{s_{_1},\,\ldots,\,s_{_1}\}$ of length $r+1$. The dual solution is regarded as $\widehat{\mbf u}(\lambda)=\left(\mathbf{DD}^T\right)^{-1}\mathbf{D}\,\mathbf{y}$, for $\lambda\geq \lambda_{_1}$. \item For step $j=2,\,3,\,\ldots\,,$ \begin{enumerate} \item Obtain the pair $\big( \tau_{_j}^{^\mathrm{join}},s_{_j}^{^\mathrm{join}} \big)$ from \begin{align}\label{joinpair} \big(\tau_{_j}^{^\mathrm{join}},s_{_j}^{^\mathrm{join}}\big)=~ \underset{i\notin \mathcal{A}_{_{j-1}},\,s\in\{-1,\,1\}}{\rm argmax}~~ \frac{a_{_i}}{s+b_{_i}}\cdot \mathbbm{1} \left\{0\leq \dfrac{a_{_i}}{s+b_{_i}} \leq \lambda_{_{j-1}}\right\}, \end{align} and set the next joining time $\lambda_{_j}^{^\mathrm{join}}$ as the value of $\frac{a_{_i}}{s+b_{_i}}$, for $i=\tau_{_j}^{^\mathrm{join}}$ and $s= s_{_j}^{^\mathrm{join}}$. \item Obtain the pair $\big( \tau_{_j}^{^\mathrm{leave}},s_{_j}^{^\mathrm{leave}}\big)$ from \begin{align}\label{leavepair} \big(\tau_{_j}^{^\mathrm{leave}},\,s_{_j}^{^\mathrm{leave}}\big)=~ \underset{i\in \mathcal{B}_{_{j-1}},\,s\in\{-1,\,1\}}{\rm argmax}~~ \dfrac{c_{_i}}{d_{_i}}\cdot\, \mathbbm{1} \Big\{c_{_i} < 0~,~ d_{_i}< 0\Big\}, \end{align} and assign the next leaving time $\lambda_{_j}^{^\mathrm{leave}}$ as the value of $\dfrac{c_{_i}}{d_{_i}}$, for $i=\tau_{_j}^{^\mathrm{leave}}$ and $s=s_{_j}^{^\mathrm{leave}}$. \item Let $\lambda_{_j}=\max \big\{\lambda_{_j}^{^\mathrm{join}}\, ,\, \lambda_{_j}^{^\mathrm{leave}}\big\}$, then the boundary set $\mathcal{B}_{_j}$ and its sign vector $\mathbf{s}_{_{\mathcal{B}_{j}}}$ are updated in the following fashion: \begin{itemize} \item[-- ] Either append $\big\{\tau_{_j}^{^\mathrm{join}}-r_{_b},\,\tau_{_j}^{^\mathrm{join}}-r_{_b}+1,\,\ldots,\tau_{_j}^{^\mathrm{join}}\big\}$ and the corresponding signs $\big\{s_{_j}^{^\mathrm{join}},\,\ldots,\,s_{_j}^{^\mathrm{join}}\big\}$ to $\mathcal{B}_{_{j-1}}$ and $\mathbf{s}_{_{\mathcal{B}_{j-1}}}$, respectively, provided that $\lambda_{_j}=\lambda_{_j}^{^\mathrm{join}}$. Also, add $\tau_{_j}^{^\mathrm{join}}$ to $\bsy\tau_{_{j-1}}$. \item[-- ] Or remove $\big\{\tau_{_j}^{^\mathrm{leave}},\, \tau_{_j}^{^\mathrm{leave}}+1,\, \ldots,\,\tau_{_j}^{^\mathrm{leave}}+r_{_b}\big\}$ and the corresponding signs $\big\{s_{_j}^{^\mathrm{leave}}, \,\ldots$, $ \,s_{_j}^{^\mathrm{leave}}\big\}$ from $\mathcal{B}_{_{j-1}}$ and $\mathbf{s}_{_{\mathcal{B}_{j-1}}}$, respectively, provided that $\lambda_{_j}=\lambda_{_j}^{^\mathrm{leave}}$. Also, remove $\tau_{_j}^{^\mathrm{leave}}$ from $\bsy\tau_{_{j-1}}$. \end{itemize} In the same manner, the augmented boundary set, $\mathcal{A}_{_j}$ and its sign, $\mathbf{s}_{_{\mathcal{A}_{_j}}}$ are formed by adding $\big\{\tau_{_j}^{^\mathrm{join}}-r_{_b},\,\ldots,\, \tau_{_j}^{^\mathrm{join}}+r_{_a}\big\}$ and $\big\{s_{_j}^{^\mathrm{join}},\,\ldots,\,s_{_j}^{^\mathrm{join}}\big\}$ to $\mathcal{A}_{_{j-1}}$ and $\mathbf{s}_{_{\mathcal{A}_{j-1}}}$, respectively, if $\lambda_{_j}=\lambda_{_j}^{^\mathrm{leave}}$ or, otherwise, by removing the associated set $\big\{\tau_{_j}^{^\mathrm{leave}},\,\ldots,\, \tau_{_j}^{^\mathrm{leave}}+r\big\}$ and $\big\{s_{_j}^{^\mathrm{leave}},\,\ldots,\,s_{_j}^{^\mathrm{leave}}\big\}$ from $\mathcal{A}_{_{j-1}}$ and $\mathbf{s}_{_{\mathcal{A}_{_{j-1}}}}$. Thus, the dual solution is computed as $\widehat{\mbf u}_{_{\mca A_{_j}}}(\lambda)=\bsy a-\lambda\, \mbf b$ for interior coordinates and $\widehat{\mbf u}_{_{-\mca A_{_j}}}(\lambda)=\lambda\,\mbf s_{_{\mca A_{_j}}}$ for boundary coordinates over $\lambda_{_j}\leq\lambda\leq\lambda_{_{j-1}}$. \end{enumerate} \item Repeat step 3 until $\lambda_{_j}> 0$. \end{enumerate} \end{algorithm} The critical points $\lambda_{_1}\, \geq\, \lambda_{_2}\, \geq\, \ldots\, \geq\, 0$ indicate the values of the regularization parameter at which the boundary set changes. \begin{remark} Notice that the vector $\bsy\tau$ derived by the PRUTF algorithm represents the locations of change points for the dual variables. In order to obtain the locations of change points in primal variables, we must add $r_{_a}$ to any element of $\bsy\tau$, that is, $\big\{ \tau_{_1}+r_{_a}, \, \tau_{_2}+r_{_a},\, \ldots\, \big\}$. This relationship between the primal and dual change point sets is visible from Figure \ref{fig:coor-removal}. \end{remark} \begin{remark} For fused lasso, $r=0$, Lemma 1 of \cite{tibshirani2011solution}, known as the boundary lemma, is satisfied since the matrix $\mbf D\mbf D^T$ is diagonally dominant, meaning that $\big[\,\mbf D \mbf D^T\,\big]_{i,i} \geq \sum_{j\neq i}\big| \big[\,\mbf D \mbf D^T\, \big]_{i,j} \big|$, for $i=1,\ldots,m$. This lemma states that when a coordinate joins the boundary, it will stay on the boundary for the rest of the path. Consequently, part (b) of step 3 in Algorithm \ref{tf.path.alg} is unnecessary, and hence the next leaving time in part (c) is set to zero, i.e., $\lambda_{_j}^{^\mathrm{leave}}=0$, for every step $j$. However, the boundary lemma is not satisfied for $r=1,\,2,\,3,\,\ldots$. \end{remark} \begin{remark}\label{rem:tibshirani_difference:proj1} There is a subtle and important distinction between our proposed algorithm, PRUTF, and the one presented in \cite{tibshirani2011solution}. The latter work studies the generalized lasso problem for any arbitrary penalty matrix $\mbf D$ (unlike $\mbf D$ used in trend filtering, which must have a certain structure). The proposed algorithm in \cite{tibshirani2011solution} relies on adding or removing only one coordinate to or from the boundary set at every step. The key attribute of our algorithm is to add or remove $r+1$ coordinates to or from the augmented boundary set, an approach inspired by the argument presented at the beginning of this section. Essentially, this attribute makes PRUTF, presented in Algorithm \ref{tf.path.alg}, well-suited for change point analysis. It is important to mention that PRUTF requires at least $r+1$ data points between neighbouring change points. \end{remark} \begin{remark} For a given $\lambda$, equations \eqref{u.boundary} and \eqref{u.interior} give the values of the dual variables in $\widehat{\mbf u}_{_\lambda}$. The equations demonstrate that the dual solution path is a linear function of $\lambda$ with change in the slopes at joining or leaving times $\lambda_1\geq\lambda_2\geq\ldots\geq 0$. \end{remark} \begin{remark} The number of iterations required for PRUTF, presented in Algorithm \ref{tf.path.alg}, is at most $(3^{\,p}+1)/2$, where $p=\lceil \frac{m}{r+1}\rceil$, see \cite{tibshirani2013lasso}, Lemma 6. However, this upper bound for the number of iterations is usually very loose. The upper bound comes from the following realization discovered by \cite{osborne2000lasso} and later improved by \cite{mairal2012complexity}. Any pair $\big(\mca A\,,\,\mbf s_{\mca A}\big)$ appears at most once throughout the solution path. In other words, if $\big(\mca A\,,\,\mbf s_{\mca A}\big)$ is visited in one iteration of the algorithm, the pair $\big(\mca A\,,\,-\mbf s_{\mca A}\big)$ as well as $\big(\mca A\,,\,\mbf s_{\mca A}\big)$ cannot reappear again for the rest of the algorithm. Interestingly, this fact says that once a coordinate enters the boundary set, it cannot immediately leave the boundary set at the next step. \noindent Moreover, note that at one iteration of the PRUTF algorithm with the augmented boundary set $\mca A$, the dominant computation is in solving the least square problem \begin{align}\label{lsproblem} \min\limits_{\mathbf{u}\, \in\, \mbb R^{m}}~ \frac{1}{2}\, \big\| \,\mathbf{y}-\mathbf{D}_{_{\mca A}}^T\, \mathbf{u}\, \big\|_{2}^2 \,. \end{align} One can apply QR decomposition of $\mathbf{D}_{_{\mca A}}^T$ to solve the least square problem, and then update the decomposition as set $\mca A$ changes. However, since $\mathbf{D}_{_{-\mathcal{A}}} \mathbf{D}_{_{-\mathcal{A}}}^T$ is a banded Toeplitz matrix (see Section \ref{sec:property.solution.path.proj1}), a solution of \eqref{lsproblem} always exists and can be computed using a banded Cholesky decomposition. Thus, the computational complexity for the iteration is of order $O \big((m-|\mca A|)\, r^2 \big)$, which is linear in the number of interior coordinates as $r$ is fixed and usually small. Overall, if $K$ is the total number of steps run by the PRUTF algorithm, then the total computational complexity is $O \big(K(m-|\mca A|)\, r^2 \big)$. See \cite{tibshirani2011solution} and \cite{arnold2016efficient}. \end{remark} \section{Statistical Properties of the Solution Path} \label{sec:property.solution.path.proj1} An important component of the methodology that we develop in this work involves computing algebraic expressions based on the matrix $\mbf D=\mbf D^{(r+1)}$. In this section, we describe the properties of such expressions. To begin with, let $\mathcal{A}=\{A_{_1},\,\ldots,\,A_{_J}\}$ and $\mathbf{s}_{\mathcal{A}}=\{\mbf s_{_1},\,\ldots,\,\mbf s_{_J}\}$ be the augmented boundary set and its corresponding sign vector, respectively, after a number of iterations of Algorithm \ref{tf.path.alg}, where $A_{_j}=\big\{\tau_{_j}-r_{_b},\,\tau_{_j}-r_{_b}+1,\,\ldots,\,\tau_{_j}+r_{_a}\big\}$ and $\mathbf{s}_{_j}=\{s_{_j},\,\ldots,\,s_{_j}\}$ for $j=1,\,\ldots,\,J$. This augmented boundary set corresponds to $J$ change points $\{\tau_{_1},\,\ldots,\,\tau_{_J}\}$ that partition all the dual variables into $J+1$ blocks $B_{_j}=\big\{\tau_{_j}+1,\,\ldots,\tau_{_{j+1}}\big\}$ for $j=0,\,1,\,\ldots,\,J$, with the conventions that $\tau_{_0}=0$ and $\tau_{_{J+1}}=m$. In the following, we list some properties of matrix multiplications involving $\mbf D$. \begin{itemize} \item It follows from the definition of the matrix $\mbf D$ that it is a banded Toeplitz matrix with bandwidth $r+1$. It tuns out that the matrix $\mbf D\mbf D^T$ reveals the same property, meaning that it is a square banded Toeplitz matrix. Moreover, its $r+1$ nonzero row elements are consecutive binomial coefficients of order $2\,r+2$ with alternating signs. In other words, $(i\, ,\, j)$-th element of $\mbf D\mbf D^T$ for $i\geq j$ is $(-1)^{\, i-j}{2\,r+2 \choose r+1+i-j}$. An example, for $r=1$, is given in panel (a) of Figure \ref{fig.Dstruct}. Note that $\mbf D\mbf D^T$ is a symmetric, nonsingular and positive definite matrix \citep{demetriou2001certain}. \item The matrix $\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T$ is a block diagonal matrix whose diagonal submatrices correspond to $J+1$ blocks. More precisely, the $j$-th submatrix on the diagonal of $\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T$ is a matrix with the first $(\tau_{_{j+1}}-\tau_{_j}-r)$ rows and columns of $\mbf D\mbf D^T$, see panel (b) of Figure \ref{fig.Dstruct}. Notice that, due to its non-singularity, $\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T$ is always invertible. In fact, both $\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}$ and $\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}$ are block diagonal matrices. Another interesting result is that every row of the matrix $\big(\mathbf{D}_{_{-\mathcal{A}}}\, \mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1} \mathbf{D}_{_{-\mathcal{A}}}$ is a contrast vector, meaning that for any $t=1,\,\ldots,\,m$, \begin{align*} \ssum{1}{n}\left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_{t,\,i}=0\,. \end{align*} \begin{figure}[!t] \centering \begin{subfigure}[b]{0.3\textwidth} \centering {\small \begin{align*} \left(\begin{array}{rrrrrrrrrrr} 6 & -4 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\ -4 & 6 & -4 & 1 & 0 & 0 & 0 & 0 & 0 \\ 1 & -4 & 6 & -4 & 1 & 0 & 0 & 0 & 0 \\ 0 & 1 & -4 & 6 & -4 & 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & -4 & 6 & -4 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 & -4 & 6 & -4 & 1 & 0 \\ 0 & 0 & 0 & 0 & 1 & -4 & 6 & -4 & 1 \\ 0 & 0 & 0 & 0 & 0 & 1 & -4 & 6 & -4 \\ 0 & 0 & 0 & 0 & 0 & 0 & 1 & -4 & 6 \\ \end{array}\right) \end{align*}} \caption{Structure of $\mbf D \mbf D^T$.} \label{fig.Dstruct.a} \end{subfigure} \quad \begin{subfigure}[b]{0.3\textwidth} \centering {\small \begin{align*} \left(\begin{array}{ccc|cccc} 6 & -4 & 1 & 0 & 0 & 0 & 0 \\ -4 & 6 & -4 & 0 & 0 & 0 & 0 \\ 1 & -4 & 6 & 0 & 0 & 0 & 0 \\\hline 0 & 0 & 0 & 6 & -4 & 1 & 0 \\ 0 & 0 & 0 & -4 & 6 & -4 & 1 \\ 0 & 0 & 0 & 1 & -4 & 6 & -4 \\ 0 & 0 & 0 & 0 & 1 & -4 & 6 \\ \end{array}\right) \end{align*}} \caption{The structure of $\mbf D_{_{-\mca A}}\, \mbf D^T_{_{-\mca A}}$.} \label{fig.Dstruct.b} \end{subfigure} \caption[Structure of Quadratic Forms of Matrix $\mbf D$]{Structure of quadratic forms of matrix $\mbf D$.} \label{fig.Dstruct} \end{figure} \item Another interesting term in analyzing the behaviour of the dual variables is $\mbf D_{\mca A}^T\,\mbf s_{\mca A}$. It can be shown that the vector $\mbf D_{\mca A}^T\,\mbf s_{\mca A}$ can be partitioned into $J+1$ subvectors associated with the change points $\tau_{_j},~j=1,\,\ldots,\,J$. The subvector associated with $\tau_{_j},\, \,j=2,\, \ldots, \,J-1$, is $\mbf D_{_{A_j}}^T\,\mbf s_{_{A_j}}$, whose elements are zero, except the first consecutive $r+1$ as well as the last consecutive $r+1$ elements. The first $r+1$ nonzero elements of $\mbf D_{_{A_j}}^T\,\mbf s_{_{A_j}}$ are the binomial coefficients in the expansion of $s_{_j}\,(x-1)^r$, and its last $r+1$ elements are the binomial coefficients in the expansion of $-s_{_{j+1}}\,(x-1)^r$. Furthermore, the first $r+1$ elements of the first subvector and the last $r+1$ elements of the last subvector are also equal to zero. For example, for a piecewise cubic signal, $r=3$, with two change points $\big(\tau_{_1}\, ,\, \tau_{_2}\big)$ and signs $\big(-1\, ,\, 1\big)$, the vector $\mathbf{D}_{_{\mathcal{A}}}^T\, \mathbf{s}_{_{\mathcal{A}}}$ becomes {\small \begin{align*} \left(\underbrace{0,\ldots,\,0,\,1,\, -3,\, 3,\,-1}_{1\, :\, (\tau_{_1}+r_{_a})}\, ,\, \underbrace{-1,\, 3,\, -3,\, 1,\, 0,\, \ldots, \,0,\, -1,\, 3,\, -3,\, 1}_{(\tau_{_1}+r_{_a}+1)\, :\, (\tau_{_2}+r_{_a})}\, ,\, \underbrace{1,\, -3,\, 3,\, -1,\, 0,\,\ldots,\,0 }_{(\tau_{_2}+r_{_a}+1)\, :\, m}\right). \end{align*}} Consequently, the structure of $\mbf D_{_{A_j}}^T\,\mbf s_{_{A_j}}$ allows us to write $\mathbf{D}_{_{\mathcal{A}}}^T\,\mathbf{s}_{_{\mathcal{A}}}=\sum_{j=0}^J \mathbf{D}_{_{{A}_j}}^T\,\mathbf{s}_{_j}$. Additionally, if the signs of two consecutive change points $\tau_{_{j}}$ and $\tau_{_{j+1}}$ are the same, then \begin{align}\label{drift.term.staircase.proj1} \left[ \big(\mathbf{D}_{_{-\mca A}} \mathbf{D}_{_{-\mca A}}^T \big)^{-1}\mathbf{D}_{_{-\mca A}}\right]_t\,\left(\mathbf{D}_{_{A_{j+1}}}^T \mathbf{s}_{_{j+1}}+\mathbf{D}_{_{A_j}}^T \mathbf{s}_{_j}\right)= -s_j, \end{align} for $t=\tau_{_j}+r_{_a}\, ,\, \ldots\, ,\, \tau_{_{j+1}}+r_{_b}$. \item Let $\mbf P_D=\mbf D_{_{-\mca A}}^T \big( \mbf D_{_{-\mca A}} \mbf D_{_{-\mca A}}^T \big)^{-1}\mbf D_{_{-\mca A}}$ be the projection matrix onto the row space of the matrix $\mbf D_{-\mca A}$. Moreover, let $\mbf X_j$ be the design matrix of the $r$-th polynomial regression on the indices of $j$-th segment $\big\{ \tau_{j}+1\, ,\, \ldots\, ,\, \tau_{j+1} \big\}$, that is, \begin{align*} \mbf X_{j}=\begin{pmatrix} 1 & \frac{\tau_{_j}+1}{n} & \left(\frac{\tau_{_j}+1}{n}\right)^2 & \cdots & \left(\frac{\tau_{_j}+1}{n}\right)^r \\[10pt] 1 & \frac{\tau_{_j}+2}{n} & \left(\frac{\tau_{_j}+2}{n}\right)^2 & \cdots & \left(\frac{\tau_{_j}+2}{n}\right)^r \\ \vdots & \vdots & \vdots & & \vdots \\ 1 & \frac{\tau_{_{j+1}}}{n} & \left(\frac{\tau_{_{j+1}}}{n}\right)^2 & \cdots & \left(\frac{\tau_{_{j+1}}}{n}\right)^r \\ \end{pmatrix}. \end{align*} Therefore, the orthogonal projection matrix $\mbf I - \mbf P_{D}$ is a block diagonal matrix whose $j$-th block associated with the segment $\big\{ \tau_{j}+1\, ,\, \ldots\, ,\, \tau_{j+1} \big\}$ is equal to the projection map onto the column space of $\mbf X_j$, i.e., \begin{align}\label{Dprojection.map.proj1} \mbf I - \mbf P_{D} = \mbf X_j \big(\mbf X_j^T\, \mbf X_j \big)^{-1}\, \mbf X_j^T. \end{align} \end{itemize} Equation \eqref{u.boundary} says that the absolute values of the boundary coordinates are $\lambda$, that is, \begin{align} \widehat{u}(t;\lambda)=\lambda\,s_{_j}\qquad\qquad\textrm{for }t\in A_j. \end{align} On the other hand, the values of the interior coordinates are given by {\small \begin{align}\label{usegment} \widehat{u}(t;\lambda)=\left\{\begin{array}{ll} \left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_t\,\left(\mathbf{y}-\lambda \,\mathbf{D}_{_{{A}_1}}^T \mathbf{s}_{_1}\right), & 1\leq t<\tau_{_1}-r_{_b}\\\\ \left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_t\,\left(\mathbf{y}-\lambda\, \left(\mathbf{D}_{_{A_{j+1}}}^T \mathbf{s}_{_{j+1}}+\mathbf{D}_{_{A_j}}^T \mathbf{s}_{_j}\right)\right), & \tau_{_j}+r_{_a}< t<\tau_{_{j+1}}-r_{_b}\\\\ \left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_t\,\left(\mathbf{y}-\lambda \,\mathbf{D}_{_{A_{_J}}}^T \mathbf{s}_{_J}\right), & \tau_{_J}+r_{_a} < t \leq m\,. \end{array}\right. \end{align}} For a given $\lambda$, the dual variables $\widehat{u}(t;\,\lambda)$ for $t=0,\,\ldots, \,m$ can be collectively viewed as a random bridge, that is, a conditioned random walk with drift whose end points are set to zero. Moreover, $\widehat{u}(t;\,\lambda)$ is bounded between $-\lambda$ and $\lambda$. The quantity $\widehat{u}(t;\,\lambda)$ can also be decomposed into a sum of several smaller random bridges which are formed by blocks created from the change points. Recall that the last consecutive $r_{_b}+1$ elements of the block $B_{_j}$ are $\lambda\, s_{_j}$, for any $j=0,\,1,\,\cdots, \,J$. Hence, for $t=\tau_{_j}+r_{_a},\,\ldots,\,\tau_{_{j+1}}-r_{_b}$, the random bridge associated with the $j$-th block is given by \begin{align}\label{z.dual.seg} \widehat{u}_{_j}(t;\, \lambda)=\left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_t\, \left(\mathbf{y}-\lambda \,\big(\mathbf{D}_{_{A_{j+1}}}^T \mathbf{s}_{_{j+1}}+\mathbf{D}_{_{A_j}}^T \mathbf{s}_{_j}\big)\right),\quad j=0,\,\ldots,\,J\,, \end{align} with the conventions $\mbf s_{_0}=\mbf s_{_{J+1}}=\mbf 0\in \mathbb{R}^{^{r+1}}$. It is important to note that similar to $\widehat{u}(t;\,\lambda)$, the process $\widehat{u}_{_j}(t;\,\lambda)$ satisfies the conditions $\widehat{u}_{_j}(\tau_{_j}+r_{_a};\,\lambda)=\lambda\,s_{_j}$ and $\widehat{u}_{_j}(\tau_{_{j+1}}-r_{_b};\lambda)=\lambda\,s_{_{j+1}}$. From \eqref{z.dual.seg}, the process $\widehat{u}_{_j}(t;\,\lambda)$ is composed of the stochastic term \begin{align}\label{u.stoch} \widehat{u}_{_j}^{\,\textrm{st}}(t)=\left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_t\, \mathbf{y}, \end{align} and the drift term \begin{align}\label{u.drift} \widehat{u}_{_j}^{\,\textrm{dr}}(t;\,\lambda)=-\lambda\left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_t\,\left(\mathbf{D}_{_{A_{j+1}}}^T \mathbf{s}_{_{j+1}}+\mathbf{D}_{_{A_j}}^T \mathbf{s}_{_j}\right). \end{align} According to model \eqref{fmodel.proj1} with Gaussian noises, it turns out that the discrete time stochastic process term $\widehat{u}_{_j}^{\,\textrm{st}}(t)$ can be embedded in a continuous time Gaussian bridge process. The following theorem describes the characteristics of this process. \begin{theorem}\label{thm:gaussian.bridge.proj1} Suppose the observation vector $\mbf y$ is drawn from the model \eqref{fmodel.proj1}, where the error vector $\bsy\varepsilon$ has a Gaussian distribution with mean zero and covariance matrix $\sigma^2\mbf\, \mbf I$. For given $\mbf D$ and $\mca A$, \begin{enumerate}[label=(\alph*)] \item Define \begin{align}\label{wj.process.proj1} W_j(t)= \big( \tau_{j+1}-\tau_j-r \big)^{-(2r+1)/2} \left[ \big(\mathbf{D}_{_{-\mathcal{A}}} \mathbf{D}_{_{-\mathcal{A}}}^T \big)^{-1} \mathbf{D}_{_{-\mathcal{A}}} \right]_{\lfloor mt\rfloor}\mathbf{y}, \end{align} for $(\tau_{_j}+r_{_a})/m ~\leq~ t ~\leq~ (\tau_{_{j+1}}-r_{_b})/m$, where \begin{align}\label{wbridge.tails.proj1} W_{_{j}} \Big(\, \frac{\tau_{_j}+r_{_a}}{m}\, \Big)= W_{_{j}} \Big(\, \frac{\tau_{_{j+1}}-r_{_b}}{m}\, \Big)=0, \end{align} and, for $j=0\, ,\, \ldots\, ,\, J$. Then the stochastic process $\mbf W_j=\big\{ W_j(t):~(\tau_{_j}+r_{_a})/m\leq t\leq (\tau_{_{j+1}}-r_{_b})/m\big\}$ is a Gaussian bridge process with mean vector zero and covariance function \begin{align}\label{w.cov} {\rm Cov} \Big( W_j(t)\, ,\, W_j(t')\Big)= \sigma^2 \left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\right]_{\lfloor mt\rfloor,\lfloor mt'\rfloor}, \end{align} for any $(\tau_{_j}+r_{_a})/m ~\leq~ t\, ,\, t' ~\leq~ (\tau_{_{j+1}}-r_{_b})/m$. \item The processes $\mbf W_j$ and $\mbf W_{j'}$ are independent, for $j'\neq j$. \end{enumerate} \end{theorem} A proof is given in Appendix \ref{prf:gaussian.bridge.proj1}. This theorem could be extended to the case of non-Gaussian random variables and therefore establishes a Donsker type Central Limit Theorem for $\mbf W_j$. Theorem \ref{thm:gaussian.bridge.proj1} guarantees that the dual variable process associated with the $j$-th block, i.e. \begin{align*} \mbf u_j= \Big\{\, \widehat{u} \big( \lfloor mt\rfloor;\,\lambda \big):~ (\tau_{_j}+r_{_a})/m\leq t\leq (\tau_{_{j+1}}-r_{_b})/m \Big\} \end{align*} is a Gaussian bridge process with the drift term \begin{align}\label{W.drift.simple} -\lambda\left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_{\lfloor mt\rfloor}\,\left(\mathbf{D}_{_{A_{j+1}}}^T \mathbf{s}_{_{j+1}}+\mathbf{D}_{_{A_j}}^T \mathbf{s}_{_j}\right), \end{align} and the covariance matrix stated in \eqref{w.cov}. Recall that a standard Brownian bridge process defined on the interval $[a,\,b]$ is a standard Brownian motion $B(t)$ conditioned on the event $B(a)=B(b)=0$. It is often characterized from a Brownian motion $B(t)$ with $B(a)=0$, by setting $$B_{_0}(t)=B(t)-\frac{t-a}{b-a}\,B(b)\,.$$ The mean and covariance functions of the Brownian bridge $B_{_0}(t)$ are given by ${\rm E} \big(B_{_0}(t)\big)=0$ and ${\rm Cov} \big( B_{_0}(s),\,B_{_0}(t) \big)= \min\lbrace s-a,\,t-a\rbrace -(b-a)^{-1}(s-a)(t-a)$ for any $s,\, t\in[a,\,b]$, respectively. A Gaussian bridge process is an extension of the Brownian bridge process when the Brownian motion $B(t)$, in the definition of the Brownian bridge $B_{_0}(t)$, is replaced by a more general Gaussian process $G(t)$. See, for example, \cite{gasbarra2007gaussian}. \begin{remark} The celebrated Donsker theorem \cite{donsker1951invariance} states that the partial sum process of a sequence of i.i.d. random variables, with mean zero and variance 1, converges weakly to a Brownian bridge process. See \cite{van1996weak} or \cite{billingsley2013convergence}. A version of Theorem \ref{thm:gaussian.bridge.proj1} involving non-Gaussian random variables would extend this result to weighted partial sum processes and show that the limiting process is a Gaussian bridge with a certain covariance structure. So the Gaussian assumption in Theorem \ref{thm:gaussian.bridge.proj1} is not restrictive. It is also interesting to show that for $r=0, \,1$, the process $\widehat{u}_{_j}^{\textrm{\,st}} \big( \lfloor mt\rfloor \big)$ boils down to its respective CUSUM processes. To show this, consider the interval $\big[(\tau_{_j}+r_{_a})/m\, , (\tau_{_{j+1}}-r_{_b})/m\big]$, \begin{itemize} \item For the piecewise constant signals, $r=0$, the quantity $\left[\big(\mathbf{D}_{_{-\mathcal{A}}} \mbf D_{_{-\mathcal{A}}}^T\big)^{-1}\mbf D_{_{-\mathcal{A}}}\right]_{\lfloor mt\rfloor}\mbf y$ can be written as {\small \begin{align*} \bigg(0,\ldots,\underbrace{0}_{ \tau_{_j}},1-\frac{\lfloor mt\rfloor}{\tau_{_{j+1}}-\tau_{_j}}, \ldots, \underbrace{1-\frac{\lfloor mt\rfloor}{\tau_{_{j+1}}-\tau_{_j}}}_{\lfloor mt\rfloor}, -\frac{\lfloor mt\rfloor}{\tau_{_{j+1}}-\tau_{_j}}, \ldots,-\frac{\lfloor mt\rfloor}{\tau_{_{j+1}}-\tau_{_j}},\, \underbrace{0}_{\tau_{_{j+1}}},\ldots,0 \bigg)\, \mbf y. \end{align*}} Notice that the above statement is the CUSUM statistic for the $j$-th segment, that is \begin{align}\label{CUSUMr0} \sum\limits_{k=\tau_{_j}+1}^{\lfloor mt \rfloor}\, \Big( y_{_k}-\overline{y}_{_{(\tau_{_j}+1):\tau_{_{j+1}}}}\Big)\,, \end{align} where $\overline{y}_{_{(\tau_{_j}+1):\tau_{_{j+1}}}}$ is the sample average of $\big( y_{_{\tau_{_j}+1}}, \,\ldots,\, y_{_{\tau_{_{j+1}}}} \big)$. It is well known that the CUSUM statistic \eqref{CUSUMr0} converges weakly to the Brownian bridge. In addition, for any $(\tau_{_j}+r_{_a})/m\leq t' \leq t\leq (\tau_{_{j+1}}-r_{_b})/m$, the covariance function becomes \begin{align*} \left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\right]_{(\lfloor mt'\rfloor,\lfloor mt\rfloor)}= (\lfloor mt'\rfloor-\tau_{_j})- \frac{(\lfloor mt'\rfloor-\tau_{_j}) (\lfloor mt\rfloor-\tau_{_j})}{\tau_{_{j+1}}-\tau_{_j}}, \end{align*} which is identical to the covariance function of the Brownian bridge. \item For the piecewise linear signals $r=1$, the quantity $\left[\big(\mathbf{D}_{_{-\mathcal{A}}} \mbf D_{_{-\mathcal{A}}}^T\big)^{-1}\mbf D_{-\mathcal{A}}\right]_{\lfloor mt\rfloor}\mbf y$ reduces to \begin{align}\label{CUSUMr1} \sum\limits_{k=\tau_{_j}+1}^{\lfloor mt \rfloor} \, k\, \big( y_{_k}-\widehat{ f}_{_k} \big)\,, \end{align} where $\widehat{f}$ is the least square fit of the simple linear regression of $\big(y_{_{\tau_{_j}+1}}, \,\ldots,\, y_{_{\tau_{_{j+1}}}}\big)$ onto $\big(\tau_{_j}+1, \,\ldots,\, \tau_{_{j+1}}\big)$. As proved in Theorem \ref{thm:gaussian.bridge.proj1}, the preceding statistic \eqref{CUSUMr1} is also a Gaussian bridge process. Furthermore, using the results in \cite{hoskins1972some}, for any $(\tau_{_j}+r_{_a})/m\leq t' \leq t\leq (\tau_{_{j+1}}-r_{_b})/m$, the covariance function of this Gaussian bridge process is given by {\small \begin{align*} \left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T \big)^{-1} \right]_{(\lfloor mt'\rfloor,\lfloor mt\rfloor)}&= \frac{\big(\Delta_{_j}- \lfloor mt\rfloor+\tau_{_j}\big)\big(\Delta_{_j}- \lfloor mt\rfloor+\tau_{_j}+1\big)}{3\, \Delta_{_j} \big(\Delta_{_j}+1\big) \big(\Delta_{_j}+2\big)}\\\\ &\hspace{-3cm}\times \big(\lfloor mt'\rfloor-\tau_{_j}\big) \big(\lfloor mt'\rfloor-\tau_{_j}+1\big)\\\\ &\hspace{-3cm}\times \Big [ \big(\lfloor mt\rfloor-\tau_{_j}+1\big) \big(\lfloor mt'\rfloor-\tau_{_j}-1\big) \big(\Delta_{_j}+2\big)-\big(\lfloor mt\rfloor-\tau_{_j}\big)\big(\lfloor mt'\rfloor-\tau_{_j}+2\big)\Delta_{_j} \Big ], \end{align*}} where $\Delta_{_j}=\tau_{_{j+1}}-\tau_{_j}$. \end{itemize} \end{remark} \section{Stopping Criterion} \label{sec:stop.rule.proj1} This section concerns developing a stopping criterion for the PRUTF algorithm. We provide tools for deriving a threshold value at which the PRUTF algorithm terminates the search if no values of dual variables exceed this threshold. Consider the dual variables at the first step of the algorithm, i.e. $\widehat{u}^{\,\trm{st}}(t)= \big[\left(\mathbf{D}\mathbf{D}^T\right)^{-1}\mathbf{D}\big]_t\,\mathbf{y}$, for $t=0,\ldots,m$, which correspond to $\widehat{u}^{\,\trm{st}}(t)$ in \eqref{u.stoch} with $\mca A=\emptyset$. It turns out that $\widehat{u}^{\,\trm{st}}(t)$ is a stochastic process with local minima and maxima attained at the change points. This structure is displayed with cyan-colored lines (\tikz\draw [color=cyan, thick, solid] (0,0) -- (.5,0);) in Figure \ref{fig:stopping-rule} for both piecewise constant $r=0$ and piecewise linear $r=1$ signals. As the PRUTF algorithm detects more change points and forms the augmented boundary set $\mca A$, the local minima or maxima corresponding to these change points are removed from the stochastic process \begin{align}\label{Uhatst} \widehat{u}_{_{-\mca A}}^{\,\trm{st}}(t)= \left[\big(\mathbf{D}_{_{-\mathcal{A}}}\mathbf{D}_{_{-\mathcal{A}}}^T\big)^{-1}\mathbf{D}_{_{-\mathcal{A}}}\right]_t\mathbf{y}= \ssum[j]{0}{J} ~\widehat{u}_{_j}^{\,\trm{st}}(t)\, \mathbbm 1 \big\{ t \in B_j \big\}, \end{align} \begin{figure}[!t] \begin{subfigure}{.45\textwidth} \centering \includegraphics[width=1\linewidth]{img/stopRule0.jpg} \caption{Piecewise constant with $r=0$} \end{subfigure} \qquad \begin{subfigure}{.45\textwidth} \centering \includegraphics[width=1\linewidth]{img/stopRule1.jpg} \caption{Piecewise linear with $r=1$} \end{subfigure} \caption[Structure of $\widehat{u}^{\,\trm{st}}(t)$ With Removed Change Points]{The cyan-colored lines show the dual variables for the full matrix $\mbf D$. Dual variables computed after removing rows of the matrix $\mbf D$ associated with $\tau_{_1}$, that is $\mbf D_{_{-\mca A_{_1}}}$, are displayed by the olive-colored lines. The augmented boundary set $\mca A_{_2}$ corresponding to $\tau_{_1}$ and $\tau_{_2}$ results to the dual variables shown by orange-colored lines. } \label{fig:stopping-rule} \end{figure} \noindent for $t=1,\,\ldots,\, m-|\mca A|$. This fact is shown by olive-colored lines (\tikz\draw [color=olive,thick,solid] (0,0) -- (.5,0);) in Figure \ref{fig:stopping-rule}. The last equality in \eqref{Uhatst} expresses that the $\widehat{u}_{_{-\mca A}}^{\,\trm{st}}(t)$ is the stochastic term of the dual variables for all the interior coordinates and is derived by stacking the stochastic terms of the dual variables associated with $j$-th block, $\widehat{u}_{_j}^{\,\trm{st}}(t)$, as defined in \eqref{u.stoch}, for $j=0,\,\ldots,\,J$. This behaviour suggests a way to introduce a stopping rule for the PRUTF algorithm. As can be viewed from the orange-colored lines (\tikz\draw [color=orange,thick,solid] (0,0) -- (.5,0);) of Figure \ref{fig:stopping-rule}, if all true change points are captured by the algorithm and stored in the augmented set $\mca A_0$, the resulting process \begin{align*} \widehat{u}_{_{-\mca A_0}}^{\,\trm{st}}(t)=\left[\big(\mathbf{D}_{_{-\mca A_0}}\mathbf{D}_{_{-\mca A_0}}^T\big)^{-1}\mathbf{D}_{_{-\mca A_0}}\right]_t\,\mathbf{y} \qquad \text{ for }\quad t=0,\,\ldots,\,m-|\mca A_0|\,, \end{align*} contains no noticeable optimum points and tends to fluctuate close to the zero line (x-axis). We terminate the search in Algorithm \ref{tf.path.alg} at step $j$ by checking whether the maximum of $\big|\,\widehat{ u}_{_{-\mca A_j}}^{\,\trm{st}} (t)\,\big|$, for $t=0,\,\ldots,\,m-|\mathcal{A}_{_j}|$, is smaller than a certain threshold. To exactly specify this threshold, as suggested by Theorem \ref{thm:gaussian.bridge.proj1}, we need to calculate the {\it excursion probabilities} of a Gaussian bridge process. As stated in \cite{adler2009random}, analytic formulas for the excursion probabilities are known to be available only for a small number of Gaussian processes. One of such Gaussian processes is the Brownian bridge process. It is well known that for the Brownian bridge process $B_{_0}(t)$ defined on the interval $[a,\,b]$ \begin{align}\label{BBmax} \Pr\Big(\sup_{a\leq t\leq b} \big| \,B_{_0}(t)\, \big| \geq x \Big)=2\,\sum\limits_{i=1}^{\infty}\, (-1)^{i+1}\exp\left(\frac{-2\,i^2\,x^2}{b-a}\right)\,. \end{align} See, for example, \cite{adler2009random}, and \cite{shorack2009empirical}. Hence for the piecewise constant signals, the required threshold for stopping Algorithm \ref{tf.path.alg} can be obtained from \eqref{BBmax}, for a suitably chosen interval $[a,\,b]$. That is, for a given value $\alpha$, we choose $x_{_\alpha}$ such that $\Pr\big( \sup_{a\, \leq\, t\, \leq\, b} |\,B_{_0}(t)\,| \geq x_{_\alpha} \big)=1-\alpha$. Therefore, for $r=0$ and $a=0$, $b=1$, we stop Algorithm \ref{tf.path.alg} at the iteration $j_{_0}$ if \begin{align*} \max_{0\, \leq\, t\, \leq\, 1}~ \Big|\widehat{\mbf u}_{_{-\mca A_{j_{_0}}}}^{\,\trm{st}}\big( \lfloor\, kt\,\rfloor \big) \Big| ~\leq~ \sigma\, x_{_\alpha}\,\sqrt{k}\,, \qquad\qquad \text{ for } \quad t=0,\,\ldots,\,m-|\mca A_{_{j_0}}|\,, \end{align*} and $k=m-|\mca A_{j_0}|$. For $r\geq 1$, the threshold is obtained in a similar fashion. Although the excursion probabilities for the Gaussian bridge processes are not known, we notice that by adopting the steps for the proof of \eqref{BBmax} in \cite{beghin1999maximum}, we can establish a similar formula for the Gaussian bridge process $G_{_0}(t)$ in Theorem \ref{thm:gaussian.bridge.proj1} as \begin{align}\label{GBmax} \Pr\Big(\sup_{a\, \leq\, t\, \leq\, b}\, \big|\,G_{_0}(t)\, \big| \geq x \Big)=2\, \sum\limits_{i=1}^{\infty}~(-1)^{i+1}\, \exp\left(\frac{-2\,i^2\,x^2}{S_r^2(k)}\right)\,, \end{align} where $k=m-|\mca A_{_{j_{_0}}}|$, and the quantity $S_r^2(k)$ is the $k$-th diagonal element of the matrix $$\Big(\mathbf{D}_{_{-\mathcal{A}_{j_{_0}}}}\mathbf{D}_{_{-\mathcal{A}_{j_{_0}}}}^T\Big)^{-1}\,.$$ Hence, we stop Algorithm \ref{tf.path.alg} at the iteration $j_{_0}$ if \begin{align}\label{stop.rule.proj1} \max_{0\, \leq\, t\, \leq\, 1}~ \Big|\widehat{\mbf u}_{_{-\mca A_{j_{_0}}}}^{\,\trm{st}}\big( \lfloor\, kt\,\rfloor\big) \Big| \leq \sigma x_{_\alpha}\left(k-r\right)^{(2\,r+1)/2}\,, \qquad \text{ for }\quad t=0,\,\ldots,\,m-|\mca A_{_{j_0}}, \end{align} where $x_{_\alpha}$ is derived from the equation \begin{align}\label{thresh.stop.rule} \sum\limits_{i=1}^{\infty}(-1)^{i+1}\exp\left(\frac{-2\,i^2\,x_{_\alpha}^2}{S_r^2(k)}\right)=\frac{\alpha}{2}\,. \end{align} \section{Pattern Recovery and Theories} \label{sec:pattern.recovery.proj1} The main purpose of this section is to investigate whether the PRUTF algorithm can recover features of the true signal $\mbf f$. We also demonstrate conditions under which the structure of the estimated signal $\widehat{\mbf f}$ matches the true signal $\mbf f$. To verify the performance of PRUTF in the discovery the true signal, we first define what we mean by pattern recovery. \begin{definition}{(Pattern Recovery):} A trend filtering estimate $\widehat{\mbf f}$ recovers the pattern of the true signal $\mbf f$ if \begin{align}\label{pat.rec} \trm{sign} \big( \big[\,\mbf D\widehat{\mbf f}\, \big]_i \big)=\trm{sign}\big(\big[\,\mbf D\mbf f\,\big]_i\big), \qquad\qquad \textrm{for} \quad i=1,\ldots,m, \end{align} where $m=n-r-1$ is the number of rows of matrix $\mbf D$. We use the notation $\widehat{\mbf f} \stackrel{pr}{=}\mbf f$ to briefly denote the pattern recovery feature of $\widehat{\mbf f}$. \end{definition} In the asymptotic framework, a trend filtering estimate is called pattern consistent if \begin{align} \Pr \big(\, \widehat{\mbf f}\, \stackrel{pr}{=} \, \mbf f \, \big) ~\longrightarrow 1\qquad\qquad \textrm{as} \quad n\longrightarrow \infty, \end{align} where $\widehat{\mbf f}=\widehat{\mbf f}_n$, to denote its dependency to the sample size $n$. Pattern recovery is very similar to the concept of sign recovery in lasso \citep{zhao2006model,wainwright2009sharp} as it deals with the specification of both locations of non-zero coefficients and their signs. The problem of pattern recovery is studied for the special case of the fused lasso in several papers. \cite{rinaldo2009properties} derived conditions under which fused lasso consistently identifies the true pattern. This was contradicted by \cite{rojas2014change}, who argued that fused lasso does not always succeed in discovering the exact change points. \cite{rojas2014change} showed that fused lasso can be reformulated as the usual lasso, for which the necessary conditions for exact sign recovery have been established in the literature. Then, they proved that one such necessary condition, known as the irrepresentable condition, is not satisfied for the transformed lasso when there is a specific pattern called a staircase (Definition \ref{staircase.def}). Corrections to \cite{rinaldo2009properties} were appeared in \cite{rinaldo2014corrections}. Later on, \cite{qian2016stepwise} proposed a method called puffer transformation, which is shown to be consistent in specifying the exact change points, including in the presence of staircases. In the remaining part of this section, we use the dual variables to demonstrate the situations in which PRUTF can correctly recover the pattern of the true signal. Exact pattern recovery implies that the dual variables are comprised of $J_{_0}+1$ consecutive bounded processes whose endpoints correspond to the true change points. The following lemma describes the situations in which exact pattern recovery can be attained. A particular case of this result in the context of piecewise constant signals was established in \cite{rinaldo2014corrections}. \begin{theorem}\label{thm:consistency.constraints.proj1} Exact pattern recovery in PRUTF occurs when the discrete time processes $\big\{ \widehat{u}_{_j}^{\,\mathrm{st}}(t)\, ,$ $t=\tau_{_j}+r_{_a}\, ,\, \ldots\, ,\, \tau_{_{j+1}}-r_{_b} \big\}$, for $j=0,\,\ldots,\,J_{_0}$, satisfy the following conditions simultaneously with probability one: \begin{enumerate}[label=(\alph*)] \item {\bf First block constraint:} for $t=1\, ,\,\ldots,\,\tau_{_1}-r_{_b}$, {\footnotesize \begin{align}\label{block.const.first} -\lambda\left(1-\left[\big( \mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T \big)^{-1}\mathbf{D}_{_{-\mca A}}\right]_t\,\mbf D_{_{A_1}}^T\mbf 1\right)& ~\leq~ \widehat{u}_{_0}^{\,\mathrm{st}}(t) ~\leq~ \lambda\left(1+\left[\big(\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T \big)^{-1}\mathbf{D}_{_{-\mca A}}\right]_t\,\mbf D_{_{A_1}}^T\mbf 1\right)\,. \end{align}} \item {\bf Last Block constraint:} for $t=\tau_{_{J_0}}+r_{_a}\, ,\, \ldots\, ,\, m$, {\footnotesize \begin{align}\label{block.const.last} \hspace{-.4cm}-\lambda\left(1+\left[ \big(\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T\big)^{-1}\mathbf{D}_{_{-\mca A}}\right]_t\,\mbf D_{_{{A}_{_{J_0}}}}^T\mbf 1\right)& ~\leq~ \widehat{u}_{_{J_0}}^{\,\mathrm{st}}(t) ~\leq~ \lambda \left(1-\left[\big(\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T \big)^{-1}\mathbf{D}_{_{-\mca A}}\right]_t\,\mbf D_{_{{A}_{_{J_0}}}}^T\mbf 1\right)\,. \end{align}} \item {\bf Interior Block constraints:} for $t=\tau_{_j}+r_{_a}\, ,\, \ldots\, ,\, \tau_{_{j+1}}-r_{_b}$, if $s_{_{j}}\neq s_{_{j+1}}$ {\footnotesize \begin{align}\label{block.const.inter} &-\lambda\left(1-\left[ \big(\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T \big)^{-1}\mathbf{D}_{_{-\mca A}}\right]_t\, \big(\mbf D_{_{A_{j+1}}}^T\mbf 1-\mbf D_{_{A_j}}^T\mbf 1 \big)\right)\, \leq \, \widehat{u}_{_j}^{\,\mathrm{st}}(t) \\[10pt] &\hspace{5cm}\leq\lambda\left(1+\left[\big(\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T \big)^{-1}\mathbf{D}_{_{-\mca A}}\right]_t\, \big(\mbf D_{_{A_{j+1}}}^T\mbf 1-\mbf D_{_{A_j}}^T\mbf 1 \big)\right)\,, \end{align}} and if $s_{_{j}}\neq s_{_{j+1}}$, which corresponds to a staircase block, $\widehat{u}_{_j}^{\,\mathrm{st}}(t) \,\leq\, 0$ or $\widehat{u}_{_j}^{\,\mathrm{st}}(t) \,\geq\, 0$. \end{enumerate} \end{theorem} \noindent In the foregoing equations, $\mbf 1 \in \mbb R^{r+1}$ is a vector of size $r+1$ whose elements are all 1. A proof of the theorem is given in Appendix \ref{prf:consistency.constraints.proj1}. We analyze the performance of the PRUTF algorithm in pattern recovery in two different scenarios; signals with staircase patterns and signals without staircase patterns. To our knowledge, \cite{rojas2014change} was the first paper to carefully investigate the staircase pattern for the piecewise constant signals in the change points analysis setting. In \cite{rojas2014change}, a staircase pattern for a piecewise constant signal refers to the phenomenon of equal signs in two consecutive changes. We extend this concept to the general case, which covers any piecewise polynomial signals of order $r$, by applying the penalty matrix $\mbf D=\mathbf{D}^{(r+1)}$. \begin{definition}\label{staircase.def} Suppose that the true signal $\mathbf{f}$ is a piecewise polynomial of order $r$ with change points at the locations $\boldsymbol\tau= \big\{\tau_{_1},\,\ldots,\,\tau_{_{J_{_0}}} \big\}$. Moreover, let $\mathbf{B}= \big\{B_{_0},\,\ldots,\,B_{_{J_{_0}}}\big\}$ be blocks created by the change points in $\bsy \tau$. A staircase occurs in block $B_{_j},~ j=1,\,\ldots,\,J_{_0}-1$ if \begin{align}\label{stair} \mathrm{sign}\big( \big[\,\mathbf{D}\mbf f\,\big]_{\tau_{_j}}\big)=\mathrm{sign}\big(\big[\,\mathbf{D}\mbf f\,\big]_{\tau_{_{j+1}}}\big). \end{align} \end{definition} The following theorem investigates the consistency of PRUTF in pattern recovery, in both with and without staircases. Specifically, it shows that for a signal without a staircase, the exact pattern recovery conditions are satisfied with probability one. On the other hand, in the presence of staircases in the signal, the probability of these conditions holding, which is equivalent to the probability of a Gaussian bridge process never crossing the zero line, converges to zero. In the literature, the consistency of a change point method is usually characterized by the signal size $n$, the number of change points $J_0$, the noise variance $\sigma_n^2$, the minimal spacing between change points, \begin{align*} \underline{L}_n= \min\limits_{j=0,\,\ldots,\,J_{_0}}~ \big|L_{n,\, j} \big|= \min\limits_{j=0,\,\ldots,\,J_{_0}}~ \big|\tau_{_{j+1}}-\tau_{_j}\big|, \end{align*} and the minimum magnitude of jumps between change points, \begin{align*} \delta_n= \min\limits_{j=1,\,\ldots,\,J_{_0}}~ \big| \mbf D_{\tau_j} \mbf f \big|. \end{align*} All the above quantities are allowed to change as $n$ grows. In the following, we present our main theorem providing conditions under which the output of the PRUTF algorithm consistently recovers the pattern of the true signal $\mbf f$. \begin{theorem}\label{thm:consistency.proj1} Suppose that $\mbf y$ follows the model in \eqref{fmodel.proj1}. Let $\bsy\tau$ be the set of $J_0$ change points for the true signal $\mbf f$. Additionally, assume that $\widehat{\bsy\tau}_n$ and $\widehat{\mbf f}_n$ are the set of change points estimates and the corresponding signal estimate obtained by the PRUTF algorithm, respectively. The followings hold for the PRUTF algorithm. \begin{enumerate}[label=(\alph*)] \item {\bf Non-staircase Blocks:} Suppose there is no staircase block in the true signal $\mbf f$. For some $\xi>0$ and with \begin{align*} \lambda_n < \dfrac{\delta_n\, \underline{L}_n^{2r+1}}{n^{2r}\, 2^{\,r+2}}, \end{align*} if the conditions \begin{itemize} \item \begin{minipage}{0.875\textwidth} \begin{align}\label{consistent.conditions1.proj1} \frac{\delta_{n}\, \underline{L}_{n}^{r+1/2}}{n^r\, \sigma_n} \longrightarrow \infty \qquad\quad \trm{and} \qquad\quad \frac{ \delta_{n}\, \underline{L}_{n}^{r+1/2}}{ 2^{\,r/2+2}\,n^r\, \sigma_n\, \sqrt{\log (J_0)}} ~>~ (1+\xi), \end{align} \end{minipage} \item \begin{minipage}{0.875\textwidth} \centering \begin{align}\label{consistent.conditions2.proj1} \frac{\lambda_n\,\, \underline{L}_{n}^{r+1/2}}{n^r\, \sigma_n} \longrightarrow \infty \qquad\quad \trm{and} \qquad\quad \frac{ 2^{\,r/2+1}\, \lambda_n\,\, \underline{L}_{n}^{r+1/2}}{n^r\, \sigma_n\, \sqrt{\log (n-J_0)}} ~>~ (1+ \xi), \end{align} \end{minipage} \end{itemize} hold, then the PRUTF algorithm guarantees exact pattern recovery with probability approaching one. That is, \begin{align*} \Pr \big(\, \widehat{\mbf f}_n\, \stackrel{pr}{=}\, \mbf f \big)\longrightarrow 1 \qquad\qquad\qquad \trm{as} \qquad n \longrightarrow \infty. \end{align*} \item {\bf Staircase Blocks:} On the other hand, if the true signal $\mbf f$ contains at least one staircase block, then the probability of exact pattern recovery by the PRUTF algorithm converges to zero. That is, \begin{align*} \Pr \big(\, \widehat{\mbf f}_n\, \stackrel{pr}{=}\, \mbf f \big)\longrightarrow 0 \qquad\qquad\qquad \trm{as} \qquad n \longrightarrow \infty. \end{align*} \end{enumerate} \end{theorem} A proof is given in Appendix \ref{prf:consistency.proj1}. \begin{remark} The performance of PRUTF in terms of consistent pattern recovery relies on the quantity $\delta_n\, \underline{L}_{n}^{r+1/2}/\sigma_n$ and the choice of $\lambda_n$. In the piecewise constant case, the former quantity reduces to the well-known signal-to-noise-ratio quantity, which is crucial for a consistent change point estimation \citep{fryzlewicz2014wild,wang2020univariate}. The statements in \ref{consistent.conditions1.proj1} illustrate that the consistency of PRUTF in non-staircase blocks is achievable if the quantity $\delta_n\,\underline{L}_{n}^{\,r+1/2}/\sigma_n$ is of order $O(n^{r+c})$, for some $c>0$. In addition, the number of the change points $J_0$ is allowed to diverge, provided \begin{align*} \log \big( J_0 \big) ~\lesssim~ \frac{ \delta_{n}^2\,\, \underline{L}_{n}^{2r+1}}{ n^{2r}\, \sigma_n^2}\,. \end{align*} \end{remark} The drift term \eqref{u.drift} plays a key role in assessing the performance of PRUTF in pattern recovery. From \eqref{drift.term.staircase.proj1}, this drift for a staircase block $B_{_j}$ becomes $\lambda \,s_{_j}$, which is constant in $t$ for the entire block. Consequently, the interior dual variables $\widehat{u}_{_j}(t;\lambda)$ for the staircase block $B_{_j}$ contain only the stochastic term $\widehat{u}_{_j}^{\,\textrm{st}}(t)=\left[\big(\mathbf{D}_{_{-\mca A}}\mathbf{D}_{_{-\mca A}}^T\big)^{-1}\mathbf{D}_{_{-\mca A}}\right]_t\, \mathbf{y}$, which fluctuates around the line $\lambda\, s_{_j}$. Recall that the KKT conditions for the dual problem of trend filtering require $\widehat{u}_{_j}(t;\,\lambda)$ to stay within the lines $-\lambda$ and $\lambda$. Thus, for a signal with staircase patterns, the PRUTF algorithm is sensitive to the variability of random noises and identifies change points once $\widehat{u}_{_{j}}^{\,\textrm{st}}(t)$ touches the $\pm\lambda$ boundaries. Examples of piecewise constant and piecewise linear signals, along with their corresponding dual variables, are depicted in Figure \ref{fig:stair-dual}, in which the above argument can be clearly seen. \begin{figure}[!ht] \begin{center} \begin{subfigure}{.38\textwidth} \centering \includegraphics[width=1\linewidth]{img/PrimalDual-Stair0.jpg} \caption{Piecewise constant signal with staircase block (50, 80].} \end{subfigure} \qquad\qquad \begin{subfigure}{.38\textwidth} \centering \includegraphics[width=1\linewidth]{img/PrimalDual-Stair1.jpg} \caption{Piecewise linear signal with staircase block (20, 55].} \end{subfigure} \caption[Piecewise Constant and Piecewise Linear Signals With Staircase Patterns]{Piecewise constant and piecewise linear signals with staircase pattern at blocks (50, 80] and (20, 55] and their corresponding dual variables. } \label{fig:stair-dual} \end{center} \end{figure} According to Theorem \ref{thm:consistency.proj1}, if there is no staircase pattern in the underlying signal, the PRUTF algorithm consistently estimates the true signal, and fails to do so, otherwise. Given the results in Theorem \ref{thm:consistency.proj1}, the natural question is whether Algorithm \ref{tf.path.alg} could be modified to enjoy the consistent pattern recovery in any case. In the next section, we will present an effective remedy based on altering the sign of a change associated with a staircase block. \section{Modified PRUTF Algorithm} \label{sec:modified.trend.filtering.proj1} In this section, we attempt to modify the PRUTF algorithm in such a way that it produces consistent estimates of the number and locations of change points even in the presence of staircase patterns. As previously mentioned, for a staircase block, the drift term \eqref{u.drift} is constant and leads to false discoveries in change points. This is shown in Figure \ref{fig:stair.false.discovery} with a piecewise constant signal of size $n=100$ and the true change points at $\bsy\tau= \big\{15,\, 40,\, 50,\, 80 \big\}$. The figure reveals that the staircase block $(50,\, 80]$ leads to three false discoveries at the locations 52, 54 and 76. \begin{figure}[!t] \begin{center} \begin{subfigure}[b]{.36\textwidth} \centering \includegraphics[width=1\linewidth]{img/dual_orig_alg1.jpg} \caption{First change point at $\tau=50$.} \end{subfigure} \qquad\quad \begin{subfigure}[b]{.36\textwidth} \centering \includegraphics[width=1\linewidth]{img/dual_orig_alg2.jpg} \caption{Second change point at $\tau=15$.} \end{subfigure} \\ \begin{subfigure}[b]{.36\textwidth} \centering \includegraphics[width=1\linewidth]{img/dual_orig_alg3.jpg} \caption{Third change point at $\tau=80$.} \end{subfigure} \qquad\quad \begin{subfigure}[b]{.36\textwidth} \centering \includegraphics[width=1\linewidth]{img/dual_orig_alg4.jpg} \caption{Fourth change point at $\tau=40$.} \end{subfigure} \caption[Impact of Staircase Patterns in Change Point False Discovery] {The process of detecting change points using PRUTF for a signal with a staircase pattern. In panel (b), there are three falsely detected change points $\{52, \,54,\, 76\}$ which is due to the staircase block $(50, \,80]$.} \label{fig:stair.false.discovery} \end{center} \end{figure} \begin{figure}[!hb] \begin{center} \begin{subfigure}{.36\textwidth} \centering \includegraphics[width=1\linewidth]{img/dual_mod_alg1.jpg} \caption{First change point at $\tau=50$.} \end{subfigure} \qquad\quad \begin{subfigure}{.36\textwidth} \centering \includegraphics[width=1\linewidth]{img/dual_mod_alg2.jpg} \caption{Second change point at $\tau=15$.} \end{subfigure} \\ \begin{subfigure}{.36\textwidth} \centering \includegraphics[width=1\linewidth]{img/dual_mod_alg3.jpg} \caption{Third change point at $\tau=80$.} \end{subfigure} \qquad\quad \begin{subfigure}{.36\textwidth} \centering \includegraphics[width=1\linewidth]{img/dual_mod_alg4.jpg} \caption{Fourth change point at $\tau=40$.} \end{subfigure} \caption[Steps of the mPRUTF Algorithm]{Steps of the mPRUTF algorithm until all four true change points are identified.} \label{fig:modified-tf} \end{center} \end{figure} The inconsistency of PRUTF in the presence of a staircase as established in Theorem \ref{thm:consistency.proj1}, stems from the fact that the change signs of the two consecutive change points at both ends of the staircase block are identical. That is, for the staircase block $B_{_j}$, $\mathrm{sign} \big( [\,\mathbf{D} \mbf f\,]_{\tau_{_j}} \big)= \mathrm{sign} \big( [\,\mathbf{D} \mbf f\,]_{\tau_{_{j+1}}} \big)$. Therefore, a question arises: can we modify Algorithm \ref{tf.path.alg} in such a way that the change signs of two neighbouring change points never become equal but still yield the solution path of trend filtering? We suggest a simple but very efficient solution to the above question. Once a new change point is identified, we check whether its $r$-th order difference sign is the same as that of the change points right before and after. If these change signs are not identical, then the procedure continues to search for the next change point. Otherwise, we replace the sign of the neighbouring change point with zero. This replacement of the sign prevents the drift term \eqref{u.drift} from becoming zero. This idea is implemented for the above signal, and the result is displayed in Figure \ref{fig:modified-tf}. As shown in panel (b), the sign of the first change point at location 50 is set to zero since its sign is identical to the sign of the second change point at 15. This sign replacement vanishes false discoveries appeared in panel (b) of Figure \ref{fig:stair.false.discovery}. Based on the above argument, PRUTF presented in Algorithm \ref{tf.path.alg} can be modified as follows to avoid false discovery and to produce consistent pattern recovery. \begin{algorithm}[mPRUTF] \label{tf.modified.path.alg} \begin{enumerate} \item[] \item Execute steps 1 and 2 of Algorithm \ref{tf.path.alg}. \item \begin{enumerate} \item Execute part (a) of step 3 in Algorithm \ref{tf.path.alg} to obtain $\tau_{_j}^{^\mathrm{join}}$ and its sign $s_{_j}^{^\mathrm{join}}$. At this point, the algorithm checks whether $s_{_j}^{^\mathrm{join}}$ is identical to the signs of the change points just before and after $\tau_{_j}^{^\mathrm{join}}$. If so, set the sign of change point which is identical to $s_{_j}^{^\mathrm{join}}$ to zero. Then, repeat part (a) of step 3 again to obtain new $\tau_{_j}^{^\mathrm{join}}$ and $s_{_j}^{^\mathrm{join}}$ and update the sets $\mca A_{_j}$ and $\mca B_{_j}$. \item Execute parts $(b)$ and $(c)$ of step 3 in Algorithm \ref{tf.path.alg}. \end{enumerate} \item Repeat step 3 until either $\lambda_{_j}>0$ or a stopping rule is met. \end{enumerate} \end{algorithm} The modified PRUTF (mPRUTF) algorithm produces consistent change point estimations, even in the presence of staircase patterns. This consistency has been achieved by converting the staircase patterns to non-staircase patterns that avoid false change point detection. In other words, running mPRUTF on an arbitrary signal (with or without staircases) is equivalent to running PRUTF on a signal without any staircase; see Figures \ref{fig:stair.false.discovery} and \ref{fig:modified-tf}. Thus, from part (a) of Theorem \ref{thm:consistency.proj1}, the mPRUTF algorithm is consistent in pattern recovery. \begin{remark} In step 2, part (a) of the mPRUTF algorithm, presented in Algorithm \ref{tf.modified.path.alg}, it is impossible for the sign $s_{_j}^{^\mathrm{join}}$ of the new change point to be identical to the sign of both of its immediate neighbouring change points, because the algorithm has already checked the equality of signs at previous steps. If they are equal, the sign of the immediate neighbouring change point will be set to zero. \end{remark} Recall that the KKT optimality conditions for solutions of the trend filtering problem in \eqref{tf.dual.obj.proj1} requires the dual variables $\widehat{\mbf u}_{_\lambda}$ to be less than or equal to $\lambda$ in absolute values, i.e., $|\widehat{\mbf u}_{_\lambda}|\leq \lambda$. This condition still holds when we replace the sign values ($+1$ or $-1$) with 0. Consequently, we have the following theorem. \begin{theorem}\label{lem:modif.alg} The mPRUTF algorithm presented in Algorithm \ref{tf.modified.path.alg} is a solution path of trend filtering. \end{theorem} For brevity, we do not provide the proof of Theorem \ref{lem:modif.alg} here. We refer the reader to the similar arguments for the LARS algorithm of lasso in \cite{tibshirani2013lasso}. It is worth pointing out that the mPRUTF algorithm requires slightly more computation than the original PRUTF algorithm. The increase in computation time directly depends on the number of staircase blocks in the underlying signal. To show how mPRUTF resolves the problem of false discovery in signals with staircases, we ran both algorithms for 1000 generated datasets from a piecewise constant and piecewise linear signals. The frequency plot of the estimated change points for both algorithms are represented in Figure \ref{fig:simul-mod-vs-orig}. The figure reveals that the original algorithm produces false discoveries within staircase blocks for both signals, whereas mPRUTF resolves this issue. \begin{figure}[!h] \begin{center} \begin{subfigure}{.4\textwidth} \centering \includegraphics[width=\linewidth]{img/cpfreq_pc.jpg} \caption{A piecewise constant signal with blocks 4 and 8 as staircase blocks.} \end{subfigure} \quad \begin{subfigure}{.4\textwidth} \centering \includegraphics[width=\linewidth]{img/cpfreq_pl.jpg} \caption{A piecewise linear signal with blocks 3 and 5 as staircase blocks.} \end{subfigure} \caption[Change Point Frequency Plots For the PRUTF and mPRUTF Algorithms]{The frequency plots of estimated change points using the PRUTF and mPRUTF algorithms.} \label{fig:simul-mod-vs-orig} \end{center} \end{figure} \section{Numerical Studies} \label{sec:simulation.proj1} In this section, we provide numerical studies to demonstrate the effectiveness and performance of our proposed algorithm, mPRUTF . We begin with a simulation study and then provide real data analyses. \subsection{Simulation Study} In this section, we investigate the performance of our proposed method, mPRUTF, by a simulation study. We consider two scenarios, namely piecewise constant and piecewise linear signals with staircase patterns. We compare our method to some powerful state-of-the-art approaches in change point analysis. These methods, a list of their available packages on CRAN, and their applicability for different scenarios are listed in Table \ref{tab:CPmethods}. \begin{table}[!t] \centering {\small \begin{tabular}{ll ll ll llll} \hline Method && Reference && R Package && \multicolumn{3}{c}{Signal} \\ \cline{7-9} && && && PWC && PWL \\ \cline{1-1}\cline{3-3}\cline{5-5}\cline{7-9} PELT && \cite{killick2012optimal} && {\bf changepoint} && \checkmark && \ding{53} \\ WBS && \cite{fryzlewicz2014wild} && {\bf wbs} && \checkmark && \ding{53} \\ SMUCE && \cite{frick2014multiscale} && {\bf stepR} && \checkmark && \ding{53} \\ NOT && \cite{baranowski2019narrowest} && {\bf not} && \checkmark && \checkmark \\ ID && \cite{anastasiou2019detecting} && {\bf IDetect} && \checkmark && \checkmark \\ \hline \end{tabular}} \caption[A List of Change Point Detection Methods With Their Packages in CRAN]{A list of change point detection and estimation methods with their packages in CRAN. The last two columns indicate which methods can be applied to piecewise constant or/and piecewise linear signals.} \label{tab:CPmethods} \end{table} We have adopted the simulation setting of \cite{baranowski2019narrowest}, and consider piecewise constant and piecewise linear signals as follows. \begin{enumerate}[label=(\roman*)] \item A piecewise constant signal (PWC) of size $n=2024$ with the number of change points $J_{_0}=8$. The locations of the true change points are $\bsy\tau= \big\{$205, 308, 512, 820, 902, 1332, 1557, 1659$\big\}$ with jump sizes 1.464, -0.656, 0.098, 1.830, 0.537, 0.768, -0.574, -3.335. We set the starting intercept to 0. \item A piecewise linear signal (PWL) of size $n=1408$ and the number of change points $J_{_0}=7$. The true change points are located at $\bsy\tau=\big\{$256, 512, 768, 1024, 1152, 1280, 1344$\big\}$. The corresponding intercepts and slopes for 8 created blocks by $\bsy\tau$ are 0.111, 0.553, -0.481, 3.002, -7.169, -0.030, 7.217, -0.958 and -8, 6, -3, -11, 12, 4, -7, 8, respectively. \end{enumerate} \begin{figure}[!b] \begin{center} \begin{subfigure}{.475\textwidth} \centering \includegraphics[width=1\linewidth]{img/signal_pwc.jpg} \caption{PWC signal with staircases at blocks $(512\,,\, 820]$ and $(1557\,,\, 1659]$.} \end{subfigure} \quad \begin{subfigure}{.475\textwidth} \centering \includegraphics[width=1\linewidth]{img/signal_pwl.jpg} \caption{PWL signal with staircases at blocks $(512\,,\,768]$ and $(1024\,,\,1152]$.} \end{subfigure} \caption[Simulated Piecewise Constant and Piecewise Linear Signals]{The piecewise constant (PWC) and piecewise linear (PWL) signals with the generated samples used in the simulation study.} \label{fig:signal-scenarios} \end{center} \end{figure} \noindent Figure \ref{fig:signal-scenarios} displays the true PWC and PWL signals, with their representative datasets generated using model \eqref{fmodel.proj1}. We note that both PWC and PWL signals contain two staircase blocks. These blocks for the PWC signal are $(512\,,\, 820]$, $(1557\,,\, 1659]$ and for PWL signal are $(512\,,\,768]$ and $(1024\,,\,1152]$. We apply mPRUTF presented in Algorithm \ref{tf.modified.path.alg} to estimate the number and the locations of the change points for the PWC and PWL signals. In each iteration of the simulation study, we simulate a dataset according to model \eqref{fmodel.proj1} under the assumption that the error terms are independently and identically distributed as $N(0\,,\,\sigma^2)$. Moreover, we set the significance level to $\alpha=0.05$ for the stopping rule in \eqref{stop.rule.proj1}. \begin{figure}[!b] \begin{center} \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=\linewidth]{img/ncpts_pc.jpg} \caption{Average number of change points.} \label{fig:simul-pc-a} \end{subfigure} \quad \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=\linewidth]{img/mse_pc.jpg} \caption{MSE estimations.} \label{fig:simul-pc-b} \end{subfigure} \\ \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=\linewidth]{img/dh_pc.jpg} \label{fig:simul-pc-c} \end{subfigure} \quad \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=\linewidth]{img/time_pc.jpg} \caption{Computation time.} \label{fig:simul-pc-d} \end{subfigure} \caption[Comparison of mPRUTF With the State-of-the-Art Change Point Methods: Piecewise Constant Signal]{The estimated average number of change points, MSE and Hausdorff distance, as well as the computation time of various methods for PWC signal. The results are provided for different values of the noise variability $\sigma$.} \label{fig:simul-pc} \end{center} \end{figure} \begin{figure}[!b] \begin{center} \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=\linewidth]{img/ncpts_pl.jpg} \caption{Average number of change points.} \label{fig:simul-pl-a} \end{subfigure} \quad \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=\linewidth]{img/mse_pl.jpg} \caption{MSE estimations} \label{fig:simul-pl-b} \end{subfigure} \newline \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=\linewidth]{img/dh_pl.jpg} \caption{Hausdorff distance.} \label{fig:simul-pl-c} \end{subfigure} \quad \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=\linewidth]{img/time_pl.jpg} \caption{Computation time.} \label{fig:simul-pl-d} \end{subfigure} \caption [Comparison of mPRUTF With the State-of-the-Art Change Point Methods: Piecewise Linear Signal]{The estimated average number of change points, MSE and Hausdorff distance, as well as the computation time of various methods for PWL signal. The results are provided for different values of the noise variability $\sigma$.} \label{fig:simul-pl} \end{center} \end{figure} In order to explore the impact of different noise levels on the change point methods, we run each simulation for various values of $\sigma$ in $\big\{ 0.5 , \,1 , \,1.5 ,\, \ldots\, ,\, 4.5 , \,5 \big\}$. We run the simulation $N=5000$ times and report the results for each change point technique in terms of estimates of the number of change points, estimates of the mean square error given by $\textrm{MSE}=N^{-1} \sum_{\, i=1}^{N} \big(\widehat{f}_{_i}-f_{_i} \big)^2$, estimates of the scaled Hausdorff distance given by \begin{align*} d_H=\frac{1}{N}\,\max\left\{\max_{j=0,\,\ldots,\,J_{_0}}\,\,\min_{i=0,\,\ldots,\,\widehat{J}_{_0}}\,\big|\widehat{\tau}_{_i}-\tau_{_j} \big|~~,~~\max_{i=0,\,\ldots,\,\widehat{J}_{_0}}\,\min_{j=0,\,\ldots,\,J_{_0}}\, \big|\widehat{\tau}_{_i}-\tau_{_j} \big|\right\}, \end{align*} and the computation time in seconds. These quantities are frequently used to assess the performance of a change point detection technique in the literature, for example, see \cite{baranowski2019narrowest}, \cite{anastasiou2019detecting}. The signal estimate, $\widehat{f}$, is computed by the least square fit of a polynomial of order $r$ to the observations within segments created by each change point method. We also remark that the tuning parameters and stopping criteria for the methods listed in Table \ref{tab:CPmethods} are set to the default values by the packages. The results for the PWC and PWL signals are presented in Figures \ref{fig:simul-pc} and \ref{fig:simul-pl}, respectively. In the case of piecewise constant signal, as in Figure \ref{fig:simul-pc}, mPRUTF performs comparable to PELT and SMUCE in terms of the average number of change points, MSE and the scaled Hausdorff distance up to $\sigma=3$, and outperforms them as $\sigma$ increases. For $\sigma \geq 4$, similar performance to WBS, NOT and ID is viewed from these measurements. As indicated by the average number of change points, MSE and the scaled Hausdorff distance, WBS, NOT and ID outperform the other methods in almost all noise levels. From a computational point of view, mPRUTF takes a slightly longer time, mainly due to the matrix $\mbf D$ multiplications, however, this computation time decreases as noise level $\sigma$ increases. As in panel (d) of Figure \ref{fig:simul-pc}, the methods PELT, SMUCE and ID are the fastest ones. In the case of piecewise linear signal, mPRUTF is only compared to NOT and ID methods, which are applicable to the piecewise polynomials of order $r \geq 1$. As in Figure \ref{fig:simul-pl} mPRUTF outperforms both NOT and ID in terms of the average number of change points and the scaled Hausdorff distance for all noise levels. In terms of MES, mPRUTF outperforms the other two for $\sigma \geq 2$. As shown in Panel (d) of Figure \ref{fig:simul-pl}, the computation time of mPRUTF ranks second after ID. The mPRUTF method performs well in terms of the estimation of the number of change points, their locations, as well as the true signals. In fact, simulation results for most of the scenarios indicate that mPRUTF is among the most competitive change point detection approaches in the literature. \subsection{Real Data Analysis} In this section, we have analyzed UK HPI and GISTEMP and COVID-19 datasets, using our proposed algorithm. Because $\sigma^2$ is unknown for these real datasets, we applied median absolute deviation (MAD) proposed by \cite{hampel1974influence}, to robustly estimate $\sigma^2$. More specifically, a MAD estimate of $\sigma$ for piecewise constant signals is given by $\widehat{\sigma}=\textrm{Median}\big(\mbf D^{(1)}\,\mbf y \big) \big/ \big[\sqrt{2}\, \Phi^{-1}(0.75) \big]$ and for piecewise linear signals by $\widehat{\sigma}=\textrm{Median} \big( \mbf D^{(2)}\,\mbf y \big) \big/ \big[\sqrt{6}\,\Phi^{-1}(0.75) \big]$, where $\Phi^{-1}(\cdot)$ represents the inverse cumulative density function of the standard normal distribution. \begin{example}[UK HPI Data] \label{HPIdata.example} The UK House Price Index (HPI) is a National Statistic that shows changes in the value of residential properties in England, Scotland, Wales and Northern Ireland. The HPI measures the price changes of residential housing by calculating the price of completed houses sale transactions as a percentage change from some specific start date. The UK HPI uses the hedonic regression model as a statistical approach to produce estimates of the change in house prices for each period. For more details, see \url{https://landregistry.data.gov.uk/app/ukhpi}.Many researchers, including \cite{fryzlewicz2018detecting}, \cite{fryzlewicz2018tail} and \cite{anastasiou2019detecting}, have studied the UK HPI dataset in carrying out change point analysis. In the current study, we consider monthly percentage changes in the UK HPI at Tower Hamlets (an east borough of London) from January 1996 to November 2018. We have applied the mPRUTF algorithm to the dataset. The algorithm have found five change points located at the dates December 2002, April 2008 and August 2009 (may be attributed to the Credit Crunch and Financial Crises), May 2012 (may be attributed to The London 2012 Summer Olympics) and August 2015 (may be attributed to regulatory and tax changes, and also by lower net migration from the EU). The dataset, the change points derived by mPRUTF and its piecewise constant fit are presented in panel (a) of Figure \ref{fig:HPI-GISTEMP}. \end{example} \begin{figure}[!t] \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=1\linewidth]{img/hpi_cpts.jpg} \caption{UK HPI dataset and its piecewise constant fit.} \end{subfigure} \quad \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=1\linewidth]{img/gistemp_cpts.jpg} \caption{GISTEMP dataset and its piecewise linear fit.} \end{subfigure} \caption[Detected Change Points Using mPRUTF For UK HPI and GISTEMP Datasets]{The time series and fitted signals for both Tower Hamlet HPI and GISTEMP datasets presented in examples } \label{fig:HPI-GISTEMP} \end{figure} \begin{example} [GISTEMP Data] \label{GISTEMPdata.example} The Goddard Institute for Space Studies (GISS) monitors broad global changes around the world. The GISS Surface Temperature Analysis (GISTEMP) is an estimate of the global surface temperature changes (see \url{https://www.giss.nasa.gov}). In the analysis of GISTEMP data, the temperature anomalies are used rather than the actual temperatures. A temperature anomaly is a difference from an average or baseline temperature. The baseline temperature is typically computed by averaging thirty or more years of temperature data (1951 to 1980 in the current dataset). A positive anomaly indicates the observed temperature was warmer than the baseline, while a negative anomaly indicates the observed temperature was cooler than the baseline. For more details see \cite{hansen2010global} and \cite{lenssen2019improvements}. The GISTEMP dataset has been frequently explored in change point literature, for example see \cite{ruggieri2013bayesian}, \cite{james2015change} and \cite{baranowski2019narrowest}. Panel (b) of Figure \ref{fig:HPI-GISTEMP} displays the monthly land-ocean temperature anomalies recorded from January 1880 to August 2019 (see \url{https://data.giss.nasa.gov/gistemp}). The plot reveals the presence of a linear trend with several change points in the dataset. For this dataset, we have identified six change points using mPRUTF located in September 1899, February 1911, May 1929, April 1941, March 1960, October 1984. The locations of change points and an estimate of the piecewise linear signal are presented in panel (b) of Figure \ref{fig:HPI-GISTEMP}. \end{example} \begin{example}[COVID-19 Data] \label{example:covid19.proj1} Since the initial outbreak of Novel Coronavirus Disease 2019 (COVID-19) in Wuhan, China, in mid-November 2019, the virus has rapidly spread throughout the world. The pandemic has infected 21.26 million people and killed more than 761,000 \url{https://covid19.who.int/}, greatly stressing public health systems and adversely influencing global society and economies. Therefore, every country has attempted to slow down the transmission rate by various regional and national policies such as the declaration of national emergencies, quarantines and mass testing. Of vital interest to governments is understanding the pattern of the epidemic growth and assessing the effectiveness of policies undertaken. A scientist can investigate these matters by analyzing the sequence of infection data for COVID-19. Changepoint detection is one possible framework for studying the behaviour of COVID-19 infection curves. By detecting the locations of alterations in the curves, change point analysis gives us insights into changes in the transmission rate or efficiency of interventions. It also enables us to raise warning signals if the disease pattern changes. For this example, we consider the log-scale of the cumulative number of confirmed cases for Australia, Canada, the United Kingdom and the United States, during the period March 10, 2020 through April 30, 2021. We have applied mPRUTF to detect change points that have occurred in the data for each country. We then fitted a piecewise linear model to the data using the selected change points, which provides a more direct perception of how the growth rate changes over time. \begin{figure}[!t] \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=1\linewidth]{img/aus.jpg} \end{subfigure} \quad \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=1\linewidth]{img/can.jpg} \end{subfigure} \\ \vspace{.6cm} \\ \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=1\linewidth]{img/uk.jpg} \end{subfigure} \quad \begin{subfigure}{.47\textwidth} \centering \includegraphics[width=1\linewidth]{img/us.jpg} \end{subfigure} \caption[Detected Change Points Using mPRUTF For COVID-19 Datasets]{The change point locations and estimated linear trend for the transformed COVID-19 datasets in Example \ref{example:covid19.proj1}. The dates indicated on each plot show the detected change points.} \label{fig:covid19} \end{figure} Figure \ref{fig:covid19} displays the locations of change points detected by the mPRUTF algorithm as well as the estimated linear trends for the four countries. For example, our algorithm has identified eight change points for Canada, on March 26, 2020; April 9, 2020; May 11, 2020; July 14, 2020; August 31, 2020; October 10, 2020; January 12, 2021 and March 18, 2021. The figure shows segments created by the estimated change points as well as their growth rate. The growth rate for the first segment (from March 10, 2020 to March 26, 2020) is remarkably high, but starts to slightly decline after the first change point on March 26, 2020. This mild decline may be linked to the declaration of the the state of emergency, quarantine and international travel ban declared by the Government of Canada. The third segment (from April 9, 2020 to May 11, 2020), the fourth segment (from May 11, 2020 to July 14, 2020) and the fifth segment (from July 14, 2020 to August 31, 2020) have witnessed noticeable decreases in the growth rate. The decrease can perhaps/probably be explained by the mandatory use of face-coverings and the border closure with the United States for the third segment, and the use of COVID-19 serological tests and the national contact tracing for the fourth and fifth segments. An upward trend in the growth rate observed from August 31, 2020 to October 10, 2020 could have resulted from the opening of businesses and public spaces. It seems that the second wave started on October 10, 2020, with a remarkable increase in the rate that continued until January 12, 2021. After this date, the rate again declined until March 18, 2021, which could be the result of provincial states of emergency and lockdowns. The last segment witnessed another surge in the rate, perhaps due to new variants of Coronavirus. The mPRUTF algorithm has also detected seven change points for the United Kingdom on the following dates: April 4, 2020; April 28, 2020; May 25, 2020; June 22, 2020; September 9, 2020; November 26, 2020 and February 5, 2021. As can be viewed from the figure, there are remarkable declines in the growth rates for the second segment (perhaps due to the nationwide lockdown), the third segment (perhaps due to the international travel ban) and the segments from May 25, 2020 to September 9, 2020 (perhaps due to mandatory use of face masks and comprehensive contact tracing). The country witnessed a significant increase in the growth rate starting from September 9, 2020, which aligns with the reopening of businesses, schools and universities. The second national lockdown could be linked to the very small decrease in the slope of the segment from November 26, 2020 to February 5, 2021. Finally, the growth rate in the last segment seemed to be under control, which could be the result of COVID vaccinations. \end{example} \section{More on Models With Frequent Change Points or With Dependent Errors} \label{sec:model_misspecification.proj1} This section empirically investigates the performance of mPRUTF in models with frequent change points as well as models with dependent random errors. \subsection{mPRUTF in Signals With Frequent Change Points} \label{discussion:frequent.changepoint.proj1} In order to evaluate the detection power of mPRUTF in signals with frequent change points, we employ a teeth signal for the piecewise constant case and a wave signal for the piecewise linear case. For the teeth signal, we consider a signal with 29 change points and varying segment lengths defined as follows: \begin{itemize} \item for $1 \leq t \leq 50$, $f_t=0$ if $(t~ \trm{mod}~ 10) \in \{1, \ldots, 5\}$; $f_t=1$, otherwise, \item for $51 \leq t \leq 150$, $f_t=0$ if $(t~ \trm{mod}~ 20) \in \{1, \ldots, 10\}$; $f_t=1$, otherwise, \item for $151 \leq t \leq 250$, $f_t=0$ if $(t~ \trm{mod}~ 40) \in \{1, \ldots, 20\}$; $f_t=1$, otherwise, \item for $251 \leq t \leq 500$, $f_t=0$ if $(t~ \trm{mod}~ 100) \in \{1, \ldots, 50\}$; $f_t=1$, otherwise. \end{itemize} The signal is displayed in the top-left panel of Figure \ref{fig:cpts_frequent_changepoint.proj1}. \begin{figure}[!b] \begin{center} \begin{subfigure}{.45\textwidth} \centering \includegraphics[width=\linewidth]{img/pwc_frq_cpts.jpg} \caption{Teeth signal} \end{subfigure} \quad \begin{subfigure}{.45\textwidth} \centering \includegraphics[width=\linewidth]{img/pwl_frq_cpts.jpg} \caption{Wave signal} \end{subfigure} \caption[Histograms of the Locations of Change Points For the Teeth and Wave Signals]{Histograms of the locations of change points for the teeth and wave signals. The histograms show the frequencies of the change points detected using mPRUTF in both signals. The result are displayed for two different $\sigma$ values.} \label{fig:cpts_frequent_changepoint.proj1} \end{center} \end{figure} \noindent The wave signal also has 29 change points with varying slopes which is defined as follows: \begin{itemize} \item for $1 \leq t \leq 50$, $f_t=-1+0.4\, t$ if $(t~ \trm{mod}~ 10) \in \{1, \ldots, 5\}$; $f_t=1-0.4\,t$, otherwise, \item for $51 \leq t \leq 150$, $f_t=-1+0.2\,t$ if $(t~ \trm{mod}~ 20) \in \{1, \ldots, 10\}$; $f_t=1-0.2\,t$, otherwise, \item for $151 \leq t \leq 250$, $f_t=-1+0.1\,t$ if $(t~ \trm{mod}~ 40) \in \{1, \ldots, 20\}$; $f_t=1-0.1\,t$, otherwise, \item for $251 \leq t \leq 500$, $f_t=-1+0.04\,t$ if $(t~ \trm{mod}~ 100) \in \{1, \ldots, 50\}$; $f_t=1-0.04\,t$, otherwise. \end{itemize} The top-right panel of Figure \ref{fig:cpts_frequent_changepoint.proj1} shows this signal. We generated $1000$ independent samples of $y_{_t}$ in model \eqref{fmodel.proj1} with $\varepsilon_t \stackrel{\trm{i.i.d}} {\sim} N(0\, ,\, \sigma^2)$ for both signals. The mPRUTF algorithm was then applied to these samples to estimate their change point locations. Figure \ref{fig:cpts_frequent_changepoint.proj1} shows the histograms of the locations of these change points for the signals. The figure provides evidence that mPRUTF is unable to effectively detect change points in signals with frequent change points and short segments. It also shows that the results deteriorate when the noise variance $\sigma^2$ or the polynomial order $r$ increase. It turns out that the success of the mPRUTF algorithm critically relies on its stopping rule. Equation \eqref{stop.rule.proj1} verifies that estimating the noise variance $\sigma^2$ and specifying the threshold $x_{\alpha}$ from a Gaussian bridge process play crucial roles in the stopping rule. As discussed in \cite{fryzlewicz2018detecting}, the two widely used robust estimators of $\sigma$, Mean Absolute Deviation (MAD) (used here) and Inter-Quartile Range (IQR), overestimate $\sigma$ in frequent change point scenarios. In addition, determining the accurate value of the threshold $x_{\alpha}$ using \eqref{thresh.stop.rule} is affected in such scenarios. These two factors prevent the stopping rule from being effective in the mPRUTF algorithm and lead to the underestimation of change points for these scenarios. We must note that such poor performance in frequent change point scenarios is not specific to mPRUTF. As investigated in \cite{fryzlewicz2018detecting}, state-of-the-art methods such as PELT, WBS, MOSUM, SMUCE and FDRSeg are among the approaches that fail in such scenarios. \subsection{mPRUTF in Models With Dependent Error Terms} \label{discussion:numeric.dependent.proj1} How can mPRUTF's performance be affected by various types of random errors, such as non-Gaussian or dependent errors? This is of course an important question and will be the topic of future works. Notice that the dual solution path of trend filtering is not impacted by the type of random errors. However, the type of random errors plays a key role in the stopping rule of mPRUTF because the stopping rule is built based on Gaussian bridge processes established by Donsker's Theorem. \begin{figure}[!ht] \begin{center} \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=\linewidth]{img/ncpts_pc_dep.jpg} \label{fig:pwc.ncpts.supp.proj1} \end{subfigure} \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=\linewidth]{img/mse_pc_dep.jpg} \caption{PWC signal } \label{fig:pwc.mse.supp.proj1} \end{subfigure} \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=\linewidth]{img/dh_pc_dep.jpg} \label{fig:pwc.dh.supp.proj1} \end{subfigure} \newline \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=\linewidth]{img/ncpts_pc_dep.jpg} \label{fig:pwl.ncpts.supp.proj1} \end{subfigure} \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=\linewidth]{img/mse_pc_dep.jpg} \caption{PWL signal } \label{fig:pwl.mse.supp.proj1} \end{subfigure} \begin{subfigure}{.32\textwidth} \centering \includegraphics[width=\linewidth]{img/dh_pc_dep.jpg} \label{fig:pwl.dh.supp.proj1} \end{subfigure} \caption[mPRUTF Results for PWC and PWL in Models With Dependent Random Errors]{The estimated average number of change points, MSEs and Hausdorff distances of various methods for both PWC and PWL signals. The results are based on weakly dependent observations and provided for various values of the error variability $\sigma$.} \label{fig:simul-pc-pl_dep.proj1} \end{center} \end{figure} To empirically investigate the performance of mPRUTF for weakly dependent random errors, a simulation study is carried out here. To this end, we generate $N=5000$ samples from model \eqref{fmodel.proj1} with the PWC and PWL signals. We consider errors $\varepsilon_i$ from an $AR(1)$ model with $\varepsilon_i = \rho \, \varepsilon_{i-1} + e_i$, for $i=1,\ldots,n$. Here, $e_i$'s are independent and identical random errors drawn from $N\big(0\, ,\, (1-\rho^2)\, \sigma^2\big)$ with $\rho \in \{-0.75,\, -0.5,\, -0.25,\, 0,\, 0.25,\, 0.5,\, 0.75 \}$ and $\sigma \in \big\{ 0.5,\, 1,\, 1.5,\, \ldots,\, 4.5,\, 5 \big\}$. The results of mPRUTF for both PWC and PWL signals are provided in Figure \ref{fig:simul-pc-pl_dep.proj1}. As can be seen, the results are very similar, in terms of the average number of change points, MSEs and the scaled Hausdorff distances, for various values of $\rho$. Therefore, it appears that the mPRUTF algorithm is quite robust against dependent error terms. Extensive studies of mPRUTF for non-Gaussian and dependent random errors will be carried out in future research.

https://arxiv.org/abs/2007.11828

Exponential node clustering at singularities for rational approximation, quadrature, and PDEs

Rational approximations of functions with singularities can converge at a root-exponential rate if the poles are exponentially clustered. We begin by reviewing this effect in minimax, least-squares, and AAA approximations on intervals and complex domains, conformal mapping, and the numerical solution of Laplace, Helmholtz, and biharmonic equations by the "lightning" method. Extensive and wide-ranging numerical experiments are involved. We then present further experiments showing that in all of these applications, it is advantageous to use exponential clustering whose density on a logarithmic scale is not uniform but tapers off linearly to zero near the singularity. We give a theoretical explanation of the tapering effect based on the Hermite contour integral and potential theory, showing that tapering doubles the rate of convergence. Finally we show that related mathematics applies to the relationship between exponential (not tapered) and doubly exponential (tapered) quadrature formulas. Here it is the Gauss--Takahasi--Mori contour integral that comes into play.

\section{\label{sec-intro}Introduction} Analytic functions can be approximated by polynomials with exponential convergence, i.e., $\|f-p_n\| = O( \exp(-Cn))$ for some $C>0$ as $n\to\infty$. Here $n$ is the polynomial degree and $\|\cdot\|$ is the $\infty$-norm on an approximation domain $E$, which may be a closed interval of the real axis or more generally a simply connected compact set in the complex plane. This result is due to Runge~\cite{runge,walsh} and explains the exponential convergence of many numerical methods when applied to analytic functions, including Gauss and Clenshaw--Curtis quadrature~\cite{gaussCC,atap} and spectral methods for ordinary and partial differential equations~\cite{smim,series}. It is also the mathematical basis of Chebfun~\cite{chebfun}. If $f$ is not analytic in a neighborhood of $E$, then Bernstein showed in 1912 that exponential convergence of polynomial approximations is impossible~\cite{bernstein,atap}. Bernstein also showed that in approximation of functions with derivative discontinuities such as $f(x) = |x|$ on $[-1,1]$, polynomials can converge no faster than $O(n^{-1})$~\cite{b14}. Now from the beginning, going back to Chebyshev in the mid-19th century, approximation theorists had investigated approximation by rational functions as well as polynomials. Yet it was not until fifty years after these works by Bernstein that it was realized that for this problem of approximating $|x|$ on $[-1,1]$, rational functions can achieve the much faster rate of {\em root-exponential convergence,} that is, $\|f-r_n\| = O( \exp(-C\sqrt n\kern 1pt))$ for some $C>0$. This result was published by Newman in 1964~\cite{newman}, who also showed that faster convergence is not possible. With hindsight, it can be seen that the root-exponential effect was implicit in the results of Chebyshev's student Zolotarev nearly a century earlier~\cite{gonZ,nf,stahl93,zol}, but this was not noticed. Newman's theorem has been a great stimulus to further research in rational approximation theory~\cite{gonZ,gon67,gonchar,levsaff,safftotik,stahlgeneral,stahl93,vyach}. It has not, however, had much impact on scientific computing until very recently with the discovery that it can be the basis of root-exponentially converging numerical methods for the solution of partial differential equations (PDE\kern .3pt s) in domains with corner singularities~\cite{stokes,conf,lightning,PNAS,laplace}. The aim of this paper is to contribute to building the bridge between approximation theory and numerical computation. In particular, we shall focus on the key feature that gives rational approximations their power: the exponential clustering of poles near singularities. (The zeros are also exponentially clustered, typically interlacing the poles, with the alternating pole-zero configuration serving as proxy for a branch cut.) This has been a feature of the theory since Newman's explicit construction. Our aim is, first, to show how widespread this effect is, not only with minimax approximations (i.e., optimal in the $\infty$-norm), the focus of most theoretical studies, but also for other kinds of approximations that may be more useful in computation. Section~\ref{sec2} explores this effect in a wide range of applications. In section~\ref{sec3} we turn to a new contribution of this paper, the observation that good approximations tend to make use of poles which, although exponentially clustered, have a density on a logarithmic scale that tapers to zero at the endpoint. Specifically, the distances of the clustered poles to the singularity appear equally spaced when the log of the distance is plotted against the square root of the index. We show experimentally that this scaling appears not just with minimax approximations but more generally. To explain this effect, we begin with a review in section~\ref{sec4} of the Hermite contour integral, which is the basis of the application of potential theory in approximation. We show how this leads to the idea of condenser capacity for the analysis of rational approximation of analytic functions. Section~\ref{sec5} then turns to functions with singularities and explains the tapering effect. In this case the condenser is short-circuited, and it is not possible to estimate the Hermite integral by considering the $\infty$-norm of the factors of its integrand, but the $1$-norm gives the required results. Analysis of a model problem shows how the tapered exponential clustering of poles enables better overall resolution, potentially doubling the rate of convergence. These arguments are related to those developed in the theoretical approximation theory literature by Stahl and others~\cite{stahlgeneral,stahl93,stahl}, but we believe that section~\ref{sec5} of this paper is the first to connect this theory with numerical analysis. Finally in section~\ref{sec6} we turn to a different problem, the quadrature of functions with endpoint singularities on $[-1,1]$. Here the famous methods are the exponential (tanh) and double exponential (tanh-sinh) formulas~\cite{haber,IMT,mori,ms01,oms,sugihara,tm73,tm74,tanaka}. Making use of the link to another contour integral formula, the Gauss--Takahasi--Mori integral~\cite{gauss,gaussCC,tm71}, we show that the distinction between straight and tapered exponential clustering arises here too. Throughout the paper, $R_n$ denotes the set of rational functions of {\em degree~$n$}, that is, functions that can be written as $r(x) = p(x)/q(x)$ where $p$ and $q$ are polynomials of degree~$n$. The norm $\|\cdot\|$ is the $\infty$-norm on $E\kern 1pt$, but, as mentioned above, other measures will come into play in sections~\ref{sec5} and~\ref{sec6}, and indeed, a theme of our discussion is that certain aspects of rational approximation are often concealed by too much focus on the $\infty$-norm. The numerical experiments in this paper are a major part of the contribution; we are not aware of comparably detailed studies elsewhere in the literature. Our emphasis is on the results, not the algorithms, but our numerical methods are briefly summarized in the discussion section at the end. \section{\label{sec2}Root-exponential convergence and exponential clustering of poles} In this section we explore the convergence of a variety of rational approximations to analytic functions with boundary branch point singularities. Our starting point is Fig.~\ref{fig1}, which presents results for six kinds of approximations of $f(x) =\sqrt x$ on $[\kern .5pt 0,1]$ by rational functions of degrees $1\le n \le 20$. (By the substitution $x = t^2$, this is equivalent to Newman's problem of approximation of $|t|$ on $[-1,1]$.) The choice of $f$ is not special; as we shall illustrate in Figs.~\ref{fig2} and~\ref{fig4}, other functions with endpoint singularities give similar results. First, the big picture. The upper-left image of the figure shows $\infty$-norm errors $\|f-r_n\|$ plotted on a log scale as functions not of $n$ but of $\sqrt n$. With the exception of the erratic case labeled AAA, all the curves plainly approach straight lines as $n\to\infty$: root-exponential convergence. (The shapes would be parabolas if we plotted against $n$.) The upper-right image shows the absolute values of the 20 poles for the approximations with $n=20$, that is, their distances from the singularity at $x=0$. On this logarithmic scale the poles are smoothly distributed: exponential clustering. This clustering is further shown in the lower images, for the approximation labeled minimax, by a phase portrait~\cite{wegert} of the square root function (the standard branch) and its degree 20 rational approximation after an exponential change of variables. The top four approximations have preassigned poles, making the approximation problems linear; indeed the Stenger, trapezoidal, and Newman approximations are given by explicit formulas. The AAA and minimax approximations are nonlinear, with poles determined during the computation. Although it is tempting to rank these candidates from worst at the top to best at the bottom (the minimax approximation is best by definition), this is not the point. All these approximations converge root-exponentially, and the differences in efficiency among them amount to constant factors of order 10, which can in fact be improved in most cases by introducing a scaling parameter or two. In particular, minimax and other nonlinear approximations can approximately double the rate of convergence of the linear approximations~\cite{rakh}. All these approximations can achieve accuracy $10^{-6}$ with degrees $n\approx 100$, whereas with polynomials one needs $n=\hbox{140,085}$. \begin{figure} \vskip 1.4em \begin{center} \includegraphics[scale=.99]{fig1.eps} \end{center} \caption{\label{fig1}Root-exponential convergence of six kinds of degree $n$ rational approximations of $f(x) = \sqrt x$ on $[\kern .5pt 0,1]$ as $n\to\infty$. On the upper-left, the asymptotically straight lines on this log scale with $\sqrt n$ on the horizontal axis (except for AAA) show the root-exponential effect. On the upper-right, the distances of the poles in $(-\infty,0)$ from the singularity at $x=0$ show the exponential clustering. Below, phase portraits in the complex plane of the square root function (the standard branch) and its degree 20 minimax approximation on $[\kern .5pt 0,1]$, after an exponential change of variables, show how a branch cut is approximated by interlacing exponentially clustered poles and zeros. Red before yellow going counterclockwise indicates a zero, and yellow before red indicates a pole. We use $10^z$ instead of $e^z$ to enable comparison with the axis labels in the images above.} \end{figure} We comment now on the individual approximations of Fig.~\ref{fig1}. The Newman approximation comes from the explicit formula presented in his four-page paper~\cite{newman}. The approximation is\/ $r(x) = {\sqrt x\kern 1pt} (\kern .7pt p({\sqrt x\kern 1pt})-p(-{\sqrt x\kern 1pt}))/(\kern .7pt p({\sqrt x\kern 1pt})+p(-{\sqrt x\kern 1pt}))$, where $p(t) = \prod_{k=0}^{2n-1} (t+\xi^k)$ and $\xi = \exp(-1/\sqrt{\kern .5pt 2n}\kern 1pt )\kern .7pt$; this can be shown to be a rational function in $x$ of degree $n$. The asymptotic convergence rate is $\exp(-\sqrt{\kern .5pt 2\kern .3pt n}\kern 1pt )$~\cite{xz}. This can be improved to approximately $\exp(-(\pi/2)\sqrt{\kern .5pt 2\kern .3pt n}\kern 1pt )$ by defining $\xi = \exp(-(\pi/2)/\sqrt{\kern .5pt 2\kern .3pt n}\kern 1pt )$, an example of the scaling parameters mentioned in the last paragraph (these values are conjectured to be optimal based on numerical experiments). The trapezoidal approximation originates with Stenger's investigations of sinc functions and associated approximations~\cite{stengersurvey,stenger,stengerbook}. Following p.~211 of~\cite{atap}, we approximate $\sqrt x$ by starting from the identity $\sqrt x = {2\kern .3pt x/\pi} \int_0^\infty (t^2 + x)^{-1}dt$, which with the change of variables $t = e^s$ becomes \begin{equation} \sqrt x = {2\kern .3pt x\over \pi} \int_{-\infty}^\infty {e^s ds\over e^{2s} + x}. \end{equation} For $n\ge 1$, we approximate this integral by an equispaced $n$-point trapezoidal rule with step size $h>0$, \begin{equation} r(x) = {2h x\over \pi} \sum_{k =-(n-1)/2}^{(n-1)/2} {e^{kh}\over e^{2kh}+x}. \label{traprule} \end{equation} (If $n$ is even, the values of $k$ are half-integers.) There are $n$ terms in the sum, so $r$ is a rational function of degree $n$ with simple poles at the points $p_k = -\exp(2kh)$. Two sources of error make $r(x)$ differ from $\sqrt x$. The termination of the sum at $n<\infty$ introduces an error of the order of $\exp(-nh/2)$, and the finite step size introduces an error on the order of $\exp(-\pi^2/h)$, since the integrand is analytic in the strip around the real $s$-axis of half-width $\pi/2$~\cite[Thm.~5.1]{trap}. Balancing these errors gives the optimal step size $h \approx \pi\sqrt{\kern .5pt 2/n}$ and approximation error $\|r - {\sqrt x\kern 1pt} \| \approx \exp(-\pi\sqrt{n/2}\kern .5pt)$. Note that the poles for this approximation cluster at $\infty$ as well as at $0$, and indeed, it converges root-exponentially not just on $[\kern .5pt 0,1]$ but on any interval $[\kern .5pt 0,L]$ with $L>0$. The derivation by the trapezoidal rule just given explains in a general way why root-exponential convergence is achievable for a wide range of problems with endpoint singularities. With any exponentially graded discretization, there will be errors associated with finite grid sizes and errors associated with truncation of an infinite series. If both sources of error follow an exponential dependence, then an optimal balance with step sizes scaling with $1/\sqrt n$ can be expected to lead to a root-exponential result. Such effects are familiar in the analysis of $hp$ discretizations of partial differential equations when the step sizes $h$ and orders $p$ of multiscale discretizations are balanced to achieve optimal rates of convergence near corners~\cite{schwab}. A drawback of the trapezoidal approximation is that its derivation depends on the precise spacing of the poles, since it relies on the property that the trapezoidal rule is exponentially accurate in this special case~\cite{trap}. The curves labeled Stenger in Fig.~\ref{fig1} come from a more flexible alternative approach, also proposed by Stenger~\cite{stenger}, where we fix $n$ distinct poles $p_k\in (-\infty,0)$, $1\le k \le n$ and $n+1$ interpolation points $x_k\in [\kern .5pt 0,1]$, $0\le k \le n$, and then take~$r$ to be the unique rational function of degree $n$ with these poles that interpolates $f(x)$ in these points. The theory of rational interpolation with preassigned poles was developed by Walsh~\cite{walsh} and will be discussed in section~\ref{sec4}. For our problem of approximation on $[\kern .5pt 0,1]$ with a singularity at $x=0$, a good choice is to take $x_0 = 0$ and $x_k = -p_k$ for $k\ge 1$. In particular, our {\em Stenger approximant\/}\footnote{Stenger considered rational approximations of this kind, though not in this precise setting of a finite interval with just one endpoint singularity.} is the rational function $r$ resulting from the choices \begin{equation} -p_k = x_k = \exp(-(k-1) h) , \quad 1\le k \le n, \label{stenginterp} \end{equation} with $h = O(1/\sqrt n\kern 1pt)$. Figure~\ref{fig1} takes $h = \pi /\sqrt n$. Interpolation is important for theoretical analysis, but for practical computation, least-squares fitting is more robust and more accurate, since it does not require knowledge of good interpolation points. The least-squares data of Fig.~\ref{fig1} come from fixing the same exponentially clustered poles as in (\ref{stenginterp}), but now choosing approximation coefficients by minimizing the least-squares error $f-r$ on a discretization of $[\kern .5pt 0,1]$ by standard methods (\hbox{MATLAB}\ backslash). As always when discretizing near singularities, we use an exponentially graded mesh ({\tt logspace(-12,0,2000})), and a weight function $w(x) = \sqrt x$ is introduced in the discrete least-squares problem so that it approximates a uniformly weighted problem on the continuum. The error curve $r(x) - \sqrt x$ for $x\in [\kern .5pt 0,1]$ for this approximation (not shown) approximately equioscillates between $n+2$ extrema, indicating that it is a reasonable approximation to the best $L^\infty$ approximation with these fixed poles. The minimax data in Fig.~\ref{fig1} correspond to the true optimal (real) approximations, rational approximations with free poles. Here the error curve equioscillates between $2\kern .3pt n+2$ extrema~\cite{atap}, and the error is approximately squared; the asymptotic convergence rate is $\exp(-\pi\sqrt{\kern .5pt 2\kern .3pt n}\kern 1pt )$~\cite{stahl,vyach}. Computing minimax approximations, however, can be challenging~\cite{minimax}, and on a complex domain they need not even be unique~\cite{gt}. This brings us to the data in the figure for AAA (adaptive Antoulas--Anderson) approximation, a fast method of near-best rational approximation introduced in~\cite{aaa}. AAA approximation is at its least robust on real intervals, as reflected in the erratic data of the figure, but for more complicated problems and in the complex plane, it is often the most practical method for rational approximation. \begin{figure} \begin{center} \vskip 1em \includegraphics[scale=.99]{fig2.eps} \end{center} \caption{\label{fig2}Four more minimax approximations, showing the same root-exponential convergence and exponential clustering of poles as in Fig.~{\rm\ref{fig1}}. Two involve the functions $x^{1/\pi}$ and $x\log x$ on $[\kern .5pt 0,1]$, one involves $x$ on $[\kern .5pt 0,1]$ but with the $\infty$-norm weighted by $x$, and one involves $\sqrt z$ on the disk about ${\textstyle{1\over 2}}$ of radius ${\textstyle{1\over 2}}$. In the right image, $n$ takes its final value from the left image for each problem, $14$ for the weighted approximation and $20$ for the other cases.} \end{figure} This concludes our discussion of Fig.~\ref{fig1}. The next figure, Fig.~\ref{fig2}, illustrates that these effects are not confined to approximation on a real interval or to the function $\sqrt x$. The figure presents data for four further examples of minimax approximations. One set of curves shows approximation of $x^{1/\pi}$ on $[\kern .5pt 0,1]$, with the value $1/\pi$ chosen to dispel any thought that rational exponents might be special. This problem requires poles particularly close to the singularity since the exponent is so small. Another shows approximation of $x\log x$ on $[\kern .5pt 0,1]$. With a much weaker singularity, this problem shows higher approximation accuracy. A third shows approximation of $\sqrt x$ again, but now it is weighted minimax approximation, with a weight function $x$ (and the error measured is now the weighted error, notably smaller than before). Finally the fourth set of data shows minimax approximation of $\sqrt z$ on the complex disk $\{z {\kern 1pt:\kern 3pt} |z-{\textstyle{1\over 2}}|< {\textstyle{1\over 2}}\}$. \begin{figure} \begin{center} \vskip 1em \includegraphics[scale=.95]{fig3.eps} \end{center} \caption{\label{fig3}The conformal map of a circular pentagon onto the unit disk has been computed and then approximated numerically by a rational function of degree $70$~{\rm\cite{conf,conformal}} by the AAA algorithm. The poles cluster exponentially at the corners, where the map is singular.} \end{figure} Figure~\ref{fig3} turns to our first problem of scientific computing. Following methods presented in~\cite{conf} and~\cite{conformal}, a region $E$ in the complex plane bounded by three line segments and two circular arcs has been conformally mapped onto the unit disk, and the map has then been approximated to about eight digits of accuracy by AAA approximation, which finds a rational function with $n=70$. This process is entirely adaptive, based on no a priori information about corners or singularities, yet it clusters the poles near the corners just as in Figs.~\ref{fig1} and~\ref{fig2}.\ \ Many poles cluster at the strong singularity {\sf A} and only a few at the weak singularity {\sf B}.\ \ Note that the poles lie asymptotically on the bisectors of the external angles. This effect is well known especially from the theory of Pad\'e approximation as worked out initially by Stahl~\cite{stahl12,suetin}. Optimal approximations line up their poles along curves which balance the normal derivatives of a potential gradient on either side, and evidently the AAA method comes close enough to optimal for the same effect to appear. We finish this section with a look at lightning solvers for PDE\kern .3pt s in two-dimensional domains, introduced in 2019 and applied to date to Laplace~\cite{lightning,PNAS,laplace}, Helmholtz~\cite{PNAS}, and biharmonic equations (Stokes flow)~\cite{stokes}. In the basic case of a Laplace problem $\Delta u = 0$, the idea is to represent the solution on a domain $E$ as $u(z) \approx \hbox{\rm Re\kern .7pt} r(z)$, the real part of a rational function with no poles in $E$ that approximates the boundary data to an accuracy typically of 6--10 digits. The rational functions have preassigned poles that cluster exponentially at the corners, where the solution will normally have singularities~\cite{lehman,wasow}, and the name ``lightning'' alludes to this exploitation of the same mathematics that makes lightning strike objects at sharp corners. Coefficients for the solution are found by least-squares fitting, making this an approximation process of the same structure as in the least-squares example of Fig.~\ref{fig1}.\ \ The difference is that the approximations are now applied to give values of $u(z)$ in the interior of the domain $E$, where it is not known a priori. See Fig.~\ref{fig4} for an example on a ``snowflake'' with boundary data $\log|z|$. \begin{figure} \begin{center} \vskip 2em \includegraphics[scale=.85]{fig4.eps} \end{center} \caption{\label{fig4}Example of the lightning Laplace solver~{\rm\cite{lightning,PNAS}} as implemented in the code {\tt laplace.m}~{\rm\cite{laplace}}. For each number of degrees of freedom (DoF), poles are clustered exponentially near the $12$ corners of the domain $E$, and the numbers are increased until a solution to $10$-digit accuracy is obtained in the form of a rational function with $480$ poles. This takes $2.3$ s on a laptop, and subsequent evaluations take $22$ $\mu$s per point, with the accuracy of each evaluation guaranteed by the maximum principle.} \end{figure} Lightning solvers have been generalized to the Helmholtz equation $\Delta u + k^2u = 0$~\cite{PNAS} and the biharmonic equation $\Delta^2 u = 0$~\cite{stokes}, as illustrated in Fig.~\ref{fig5}. In the Helmholtz case, poles $(z-z_k)^{-1}$ of rational functions become singularities of complex Hankel functions $H_1(k|z-z_k|)\exp(\pm i \kern .7pt \hbox{\rm arg\kern .2pt}(z-z_k))$, and the biharmonic case is handled by the Goursat reduction $u(z) = \hbox{\rm Im\kern .7pt} (\kern .5pt \overline z f(z) + g(z))$ to a coupled pair of analytic functions $f$ and $g$, each of which is approximated by its own rational function. The mathematics of lightning methods for Helmholtz and biharmonic problems has not yet been worked out fully, and the analysis given in section~\ref{sec5} applies just to the Laplace case. \begin{figure} \begin{center} \vskip 10pt \raisebox{12pt}{\includegraphics[scale=.66]{helmfig.eps}~~~~~~~~~~~~~~~~~}% \includegraphics[scale=.80]{stokesfig.eps} \vspace{-5pt} \end{center} \caption{\label{fig5}Lightning solvers have been generalized to the two-dimensional Helmholtz (left)~{\rm\cite{PNAS}} and biharmonic equations (right)~{\rm\cite{stokes}}. The Helmholtz image shows a plane wave incident from the left scattered from a sound-soft equilateral triangle. The biharmonic image shows contours of the stream function for Stokes flow in a cavity driven by a quarter-circular boundary segment rotating at speed 1 and with zero velocity on the remainder of the boundary. The black contours in the corners, representing the stream function value $\psi = 0$, delimit counter-rotating Moffatt vortices. Tapered exponentially clustered singularities are used in both computations.} \end{figure} Although it is not the purpose of this article to give details about lightning PDE solvers, they are at the heart of our motivation. Usually in approximation theory, minimax approximations are investigated as an end in themselves, and the locations of their poles may be examined as an outgrowth of this process; a magnificent example is~\cite{stahl94}. Here, the order is reversed. Our aim is to exploit an understanding of how poles cluster to construct approximations on the fly to solve problems of scientific computing. \section{\label{sec3}Tapered exponential clustering} In the last section, 13 plots were presented of the distances of poles to singularities on a log scale, the right-hand images of Figs.~\ref{fig1}, \ref{fig2}, and~\ref{fig3}. All showed exponential clustering, and all but three showed a further effect which we call {\em tapered exponential clustering\/\kern .4pt}, the main subject of the rest of this paper: on the log scale, the spacing of the poles grows sparser near the singularity. This was also colorfully evident in the phase portrait at the bottom of Fig.~\ref{fig1}. The three exceptions were the Stenger, least-squares, and trapezoidal approximations of Fig.~\ref{fig1}, all of which are based on poles preassigned with strictly uniform exponential clustering. These examples illustrate that tapering of the pole distribution is not necessary for root-exponential convergence. A fourth set of data in Fig.~\ref{fig1} also involves preassigned poles, the Newman data, and some tapering is apparent in this case. \begin{figure} \begin{center} \vskip 15pt \includegraphics[scale=1]{fig6} \vspace{-8pt} \end{center} \caption{\label{fig6}Tapered exponential clustering of poles near singularities for the nine examples with free poles from Figs.~{\rm\ref{fig1}--\ref{fig3}} of the last section. The crucial feature is that the curves appear straight with this horizontal axis marking $\sqrt k$ rather than $k$, where $\{d_k\}$ are the sorted distances of the poles from the singularities. The data for the poles at vertex {\sf A} of Fig.~{\rm \ref{fig3}} have been deemphasized to diminish clutter (black dots), since they lie at such a different slope from the others.} \end{figure} Figure~\ref{fig6} shows the nine remaining examples of exponential clustering of poles from Figs.~\ref{fig1}--\ref{fig3}, the ones with free poles, presenting the distances $\{d_k\}$ of the poles from their nearest singularities on a log scale. What is immediately apparent is that all the curves look straight for smaller values of $k$. Note that five of them stop at $n=20$, one at $n=14$, and the remaining three, from the approximation of a conformal map of Fig.~\ref{fig3}, at different values determined adaptively by the AAA algorithm. Yet the horizontal axis in Fig.~\ref{fig6} is not $k$ but $\sqrt k$. Plotted against $k$ (not shown), the data would look completely different. Evidently in a wide range of rational approximations, both best and near-best, the distances $\{d_k\}$ of poles to singularities is well approximated by the formula \begin{equation} \log d_k \approx \alpha + \sigma \sqrt k \label{model1} \end{equation} for some constants $\alpha$ and $\sigma$, that is, \begin{equation} d_k \approx \beta\exp(\sigma \sqrt k \kern 1pt) \label{model2} \end{equation} for some $\beta$ and $\sigma$. To make sense of the $\sqrt k$ scaling, let us remove the exponential from the problem by defining a distance variable $s= \log d$, thereby transplanting an interval such as $d\in [\kern .5pt 0,1]$ to $s\in (-\infty, 0\kern .5pt]$. We ask, what can be said of the density $\rho(s)$ of poles with respect to $s\kern .7pt$? If $\rho(s)$ were constant, this would correspond to a uniform exponential distribution of poles, requiring an infinite number of poles since $s$ goes to $-\infty$. So some kind of cutoff of $\rho(s)$ to $0$ must occur as $s\to-\infty$. An abrupt cutoff, as with the Stenger, trapezoidal, and least-squares distributions of Fig.~\ref{fig1}, leads to a linear cumulative distribution, as shown in the left column of Fig.~\ref{relu}. By contrast, a linear cutoff gives a quadratic cumulative distribution, as shown in the right column, and when this is inverted, the result is the $\sqrt k$ distribution we have observed. Thus the straight lines of Fig.~\ref{fig6} can be explained if pole density functions $\rho(s)$ for good rational approximations tend to take the form sketched in the upper-right of Fig.~\ref{relu}. (Aficionados of deep learning may call this the ``ReLU'' shape.) In section~\ref{sec5} we will explain why this is the case and continue the story of Fig.~\ref{relu} in Fig.~\ref{relupot}. We have not presented data in this section for lightning PDE solutions, but it was in this context that we first became aware of the importance of tapered exponential clustering. In the course of the work leading to~\cite{lightning}, the first author noticed that although straight exponential spacing of preassigned poles gave root-exponential convergence, better efficiency could be achieved if the resulting approximations were re-approximated a second time by the AAA algorithm. On examination it was found that the AAA approximations had poles in a tapered distribution, just like cases {\sf A}--{\sf C} of Fig.~\ref{fig6}. The model (\ref{model1})--(\ref{model2}) was developed empirically in this context, with $\sigma\approx 4$ found to be an effective choice. This became the formula for preassignment of poles in the lightning Laplace software~\cite{laplace}, where it improved the overall speed by a good factor, and it appears as equation (3.6) in~\cite{lightning}. \begin{figure} \vskip 15pt \begin{center} \includegraphics[scale=1.2]{relu} \vspace{-3pt} \end{center} \caption{\label{relu}The algebra of exponential clustering. With respect to the variable $s = \log d$, where $d$ is the distance to the singularity, the simplest exponential clustering of poles would have uniform density $\rho(s)$ down to a certain value and then cut off abruptly (left column). A tapered distribution cuts off linearly instead (right column), resulting in poles exponentially clustered in the $\sqrt k$ fashion seen in Fig.~{\rm \ref{fig6}}.} \end{figure} \section{\label{sec4}Hermite integral formula and potential theory} The basic tool for estimating accuracy of rational approximations is the Hermite integral formula~\cite{levsaff,walsh}. In this section we review how this formula leads to the use of potential theory~\cite{ransford}, and in particular the quantity known as the condenser capacity, for approximations of analytic functions. Building on the work of Walsh~\cite{walsh}, these ideas began to be developed by Gonchar and Rakhmanov in the Soviet Union not long after the appearance of Newman's paper~\cite{gon67,gonchar}. The following statement is adapted from Thm.~8.2 of~\cite{walsh}. \begin{theorem} Let\/ $\Omega$ be a simply connected domain in ${\bf C}$ bounded by a closed curve $\Gamma$, and let $f$ be analytic in $\Omega$ and extend continuously to the boundary. Let distinct interpolation points $x_0,\dots, x_n\in \Omega$ and poles $p_1,\dots,p_n$ anywhere in the complex plane be given. Let $r$ be the unique degree $n$ rational function with simple poles at $\{p_k\}$ that interpolates $f$ at $\{x_k\}$. Then for any $x\in\Omega$, \begin{equation} f(x) - r(x) = {1\over 2\pi i} \int_\Gamma {\phi(x)\over \phi(t)} {f(t)\over t-x} \kern 1pt dt, \label{herm} \end{equation} where \begin{equation} \phi(z) = \prod_{k=0}^n(z-x_k)\biggl/\kern 1.5pt \prod_{k=1}^n(z-p_k). \label{phidef} \end{equation} \label{hermthm} \end{theorem} \medskip To see how this theorem is applied, let\/ $\Omega$ be a simply connected domain bounded by a closed curve~$\Gamma$, as indicated in Fig.~\ref{figpotent} (see also Fig.~9 in the next section), and let $f$ be analytic in $\Omega$ and extend continuously to $\Gamma$. Suppose $f$ is to be approximated on a compact set $E\subset \Omega$, which in this section we take to be disjoint from $\Gamma$. Theorem~\ref{hermthm} implies that for any $x\in E$, \begin{equation} |f(x) - r(x)| \le C \tau, \label{tauest} \end{equation} where $C$ is a constant independent of $n$ and $\tau$ is the ratio \begin{equation} \tau = {\max_{z\in E} |\phi(z)|\over \min_{z\in \Gamma} |\phi(z)|}. \label{rhodef} \end{equation} If $\phi$ is much smaller on $E$ than on $\Gamma$, then $\tau$ and hence $f-r$ must be small. \begin{figure} \begin{center} \vskip 10pt \includegraphics[scale=.8]{figpotent.eps}~~~~ \end{center} \caption{\label{figpotent}Potential theory and rational approximation. In each image, the shaded region is an approximation domain $E$ for a function $f$ analytic in the region $\Omega$ bounded by $\Gamma$.\ \ If we think of the poles of an approximation $r\approx f$ as positive point charges and the interpolation points as negative point charges, then a minimal-energy equilibrium distribution of the charges gives a favorable configuration for approximation. This is a discrete problem of potential theory that becomes continuous in the limit $n\to \infty$, enabling one to take advantage of invariance under conformal maps. In these images $E$ and\/ $\Gamma$ are disjoint and the convergence is exponential, but the third domain and its close-up illustrate the clustering effect, which will become more pronounced as the gap shrinks to zero. The pairs of interpolation points and poles marked by hollow dots delimit one half of the total, highlighting how both sets of points accumulate close to the singularity.} \end{figure} Figure~\ref{figpotent} gives an idea of how this can happen. In each image, the red dots on~$\Gamma$ represent a good choice of poles $\{p_k\}$ and the blue dots on the boundary of $E$ a corresponding good choice of interpolation points $\{x_k\}$. Consider first the upper-left image, where $E$ and $\Gamma$ define a circular annulus. The equispaced configurations of $\{p_k\}$ and $\{x_k\}$ ensure that $\tau$ will decrease exponentially as $n\to\infty$. To see this, in view of (\ref{phidef}), we define \begin{equation} u(z) = n^{-1}\sum_{k=0}^n \log|z-x_k| - n^{-1}\sum_{k=1}^n\log |z-p_k|. \end{equation} This is the potential function generated by $n+1$ negative point charges of strength $n^{-1}$ at the interpolation points and $n$ positive point charges of strength $-n^{-1}$ at the poles. Then $\exp(n u(z)) = |\phi(z)|$, and therefore \begin{equation} \tau = \exp(-n \kern 1pt [\kern 1pt \min_{z\in \Gamma} u(z) - \max_{z\in E} u(z)\kern 1pt]\kern 1pt ). \end{equation} For $\tau$ to be small, we want $u$ to be uniformly bigger on $\Gamma$ than on $E$. Finding the best such configuration is an extremal problem that will be approximately solved if the points are placed in an energy-minimizing equilibrium position. In each of the images of Fig.~\ref{figpotent}, the points are close to such an equilibrium. Each charge is attracted to the charges of the other color, but repelled by charges of its own color. Finding an optimal configuration (for the given choice of $\Gamma$) is complicated for finite~$n$, but the problem becomes cleaner in the limit $n\to \infty$, and this is where the power of potential theory is fully revealed. We now imagine continua of interpolation points and poles defined by a signed measure $\mu$ supported on $E$, where it is nonpositive with total mass $-1$, and on $\Gamma$, where it is nonnegative with total mass $1$. It can be shown that there is a unique measure of this kind that minimizes the energy \begin{equation} I(\kern .7pt\mu) = - \int\kern -4pt \int \log|z-t| \kern .7pt d\mu(z) \kern .7pt d\mu(t), \label{energy} \end{equation} with associated potential function \begin{equation} u(z) = -\int \log|z-t|\kern .7pt d\mu(t), \label{potential} \end{equation} and $u$ takes constant values $u_E^{} < 0$ on $E$ and $u_\Gamma^{} = 0$ on $\Gamma$. The minimum $I_{\min{}} = \inf_\mu I(\kern .5pt \mu)$ is known to be positive, and for minimax degree $n$ rational approximations $r_n^*$ one has exponential convergence as $n\to \infty$ at a corresponding rate: \begin{equation} \limsup \| f- r_n^*\|^{1/n} \le \exp(-I_{\min{}}). \label{optrate} \end{equation} (The actual rate is in fact twice as fast as this, $\exp(-2\kern .3pt I_{\min{}})$, for functions whose singularities in the complex plane are just isolated algebraic branch points~\cite[p.~93]{levsaff}, \cite{stahlgeneral}.) The reciprocal of $I_{\min{}}$ is known as the {\em condenser capacity} for the $(E,\Gamma)$ pair, a term that reflects an electrostatic interpretation of the approximation problem. In electronics, capacitance is the ratio of charge to voltage difference. A capacitor has high capacitance if its positive and negative plates are close to one another, so that the attraction of charges of opposite sign enables a great deal of charge to be accumulated on them without the need for much of a voltage difference. For fast-converging rational approximation, on the other hand, we want $\Gamma$ and $E$ to be far apart, corresponding to a {\em small\/} amount of charge relative to the voltage difference, hence small numbers of poles and interpolation points needed to achieve a given ratio $\tau$. We can now see how the second and third images of Fig.~\ref{figpotent} were drawn. They were obtained by conformal transplantation, exploiting the invariance of problems of potential theory under conformal maps. The eccentric domain of the second image comes from a M\"obius transformation, and the pinched domain of the third image comes from a further squaring. The blue and red points obtained as conformal images of equispaced points in the symmetric annulus are known as Fej\'er--Walsh points~\cite{starke93}. One might wonder, for arguments of this kind, is it necessary to place the poles of\/ $r$ on the boundary of the region of analyticity of $f\kern .7pt $? In fact, $\Gamma$ does not have to lie as far out as that boundary, nor do the poles have to be on $\Gamma$, for as stated in Theorem~\ref{hermthm}, the integral representation (\ref{herm}) is valid for any placement of the poles. Asymptotically as $n\to\infty$, however, it is known that the convergence rate cannot be improved by placing poles beyond the region of analyticity of $f$~\cite{levsaff}. A special choice is to put all the poles at $x=\infty$, in which case rational approximation reduces to polynomial approximation, still with exponential convergence though at a lower rate than in~(\ref{optrate}). \section{\label{sec5}Explanation of tapered exponential clustering} Now we examine how the analysis of the last section must change for approximations with singularities. There is a considerable specialist literature here by authors including Aptekarev, Saff, Stahl, Suetin, and Totik~\cite{bt,levsaff,rtw,safftotik,stahl,stahl12,suetin}, which investigates certain best approximations in detail. Our emphasis is on the broad ideas applicable to near-best approximations too. \begin{figure} \begin{center} \vskip .2in \includegraphics{sketchfig.eps} \end{center} \caption{\label{sketchfig}Two kinds of problems of rational approximation of a function $f$ on a domain $E$. On the left (section~\ref{sec4}), $f$ is analytic on $E$ and poles can be placed on a contour\/ $\Gamma$ enclosing $E$ in the region of analyticity:\ convergence is exponential with accuracy on the order of $\exp(-n\kern .5pt \delta\kern .5pt)$ for a constant $\delta>0$. On the right (section~\ref{sec5}), $f$ has a singularity at a point $z_c^{}$ on the boundary of $E$, and $\Gamma$ must touch $E$ at $z_c^{}$:\ convergence is root-exponential with accuracy of order $\exp(-n\kern .5pt \delta\kern .5pt)$ again, but now with~$\delta$ diminishing at the rate $1/\sqrt n$ as $n\to\infty$. In the circled region, the potential makes the transition from $u_\Gamma^{}= 0$ to $u_E^{} = -\delta$.} \end{figure} From Fig.~\ref{figpotent} it is clear that potential theory should give some insight when $f$ has a singularity on the boundary of $E$. The lower pair of images shows clustering of poles where $\Gamma$ has a cusp close to the boundary of $E$, and as the cusp is brought closer to $E$, the clustering will grow more pronounced. However, the argument we have presented breaks down when $\Gamma$ actually touches $E$. The situation is sketched in Fig.~\ref{sketchfig}. Physically, this would be a capacitor of infinite capacitance, implying that an equipotential distribution $u$ with a nonzero voltage difference would require an infinite quantity of charge. Mathematically, the estimate (\ref{tauest}) fails because $\tau$ cannot be smaller than $1$. To see what happens in such cases, we can examine the function $\phi$ computed numerically for an example problem. The left column of Fig.~\ref{figmini} shows error curves in type $(9,10)$ minimax approximation of $\sqrt x$ on $[\kern .5pt 0.01,1]$ (above) and $[\kern .5pt 0,1]$ (below). (Type $(m,n)$ means numerator degree at most $m$ and denominator degree at most $n\kern .3pt$; we choose these parameters rather than $(n,n)$ to make the plots slightly cleaner.) The curves each equioscillate between $m+n+2 = 21$ extrema, and in the lower curve, on the semilogx scale, we see the wavelength increasing as $x\to 0$. As a minimax approximation with free poles, this rational function has $m+n+1 = 20$ points of interpolation rather than the standard number $n = 10$ for an approximation with preassigned poles, so for the cleanest display of the potential function $\phi$ in the right column we have picked out just half of these, marked by the red dots. \begin{figure} \begin{center} \vskip 10pt \includegraphics[scale=.9]{figmini.eps}~~~~ \end{center} \caption{\label{figmini}On the left, error curves in type $(9,10)$ minimax approximation of $\sqrt x$ on $[\kern .5pt 0.01,1]$ and $[\kern .5pt 0,1]$. On the right, plots of $\phi(z)$ as defined by $(\ref{phidef})$ on these approximation intervals and on $[-1,0\kern .5pt]$. The curves in the upper-right image show a reasonable approximation to constant values on $[-1,0\kern .5pt]$ (upper curve) and on $[\kern .5pt 0.01,1]$ (lower curve), but in the lower-right image, nothing like constant behavior of $|\phi(z)|$ on $[-1,0\kern .5pt]$ is evident. We explain this by noting that what matters to the accuracy of an approximation is the integral (\ref{hermagain}) of $|\phi(x)/\phi(t)|$ with respect to $t\in \Gamma$, not its maximum. Taking advantage of this property, poles and interpolation points distribute themselves more sparsely near the singularity, freeing more of them to contribute to the approximation further away---the phenomenon of tapered exponential clustering.} \end{figure} The right column of Fig.~\ref{figmini} shows the function $|\phi|$ plotted on the approximation interval (the lower blue curve) and on the important portion $[-1,0\kern .5pt]$ of the integration contour $\Gamma$ (the upper red curve). (To be precise, for these plots the numerator of (\ref{phidef}) ranges over just the interpolation points $x_1,\dots, x_n$ marked by red dots.) In the upper image, for $[\kern .5pt 0.01,1]$, the curves reveal a reasonable approximation to what the last section has led us to expect from potential theory. The blue curve has approximately even magnitude, and this is about five orders of magnitude below the red curve, also of approximately even magnitude. Thus the ratio $\tau$ of (\ref{rhodef}) is far below $1$, and the estimate (\ref{tauest}) serves to bound the approximation error. (The actual error is about the square of this bound since we have omitted half the interpolation points.) The lower image, which is a centerpiece of this paper, tells a strikingly different story. Here again the blue curve is flat, showing the even dependence on $x$ we expect in a minimax approximation. The red curve for $|\phi(z)|$ on $[-1,0\kern .5pt]$, however, is now tilted at an angle on these log-log axes, showing a steady closing of the gap between the curves as\/ $t$ moves from $-1$ to $0$. Clearly in this case $[-1,0\kern .5pt]$ is not at all a curve of constant $|\phi|$. To understand the linearly closing gap in Fig.~\ref{figmini}, we note that what fails in the analysis of the last section for an approximation problem with a singularity is not the Hermite integral, \begin{equation} f(x) - r(x) = {1\over 2\pi i} \int_\Gamma {\phi(x)\over \phi(t)} {f(t)\over t-x} \kern 1pt dt, \label{hermagain} \end{equation} but the estimate (\ref{tauest}) we derived from it. Implicitly (\ref{tauest}) came from bounding (\ref{hermagain}) by H\"older's inequality, \begin{equation} |f(x) - r(x) | \le {1\over 2\pi} \left\|{\phi(x)\over \phi(t)}\right\|_\infty \left\|{f(t)\over t-x}\right\|_\infty \|1\|_1^{}, \label{option1} \end{equation} where the $\infty$- and $1$- norms are defined over $t\in \Gamma$. (The norm $\|1\|_1^{}$ is equal to the length of $\Gamma$.) When $\Gamma$ and $E$ are disjoint, the first $\infty$-norm in (\ref{option1}) is exponentially small as $n\to\infty$ and the second is bounded. However, these properties fail as $\Gamma$ and $E$ touch. We can rescue the argument by noting that $|\phi(x)/\phi(t)|$ does not have to be small for all $t$ so long as its integral is small. More precisely, the quantity $f(t)/(t-x)$ of (\ref{option1}) may not be bounded as $t,x\to z_c^{}$ but $f(t)|\kern .5pt t-z_c^{}|^{1-\alpha} /(t-x)$ will be bounded if we assume $f(t-z_c^{}) = O(|\kern .5pt t-z_c^{}|^\alpha)$ for some constant $\alpha$. So what actually matters is that the integral of $|\kern .5pt t-z_c^{}|^{\alpha - 1} |\phi(x)/\phi(t)|$ should be small, and we accordingly replace (\ref{option1}) by the alternative H\"older estimate \begin{equation} |f(x) - r(x) | \le {1\over 2\pi} \left\|{\phi(x)\over \phi(t) }|\kern .5pt t-z_c^{}|^{\alpha-1}\right\|_1 \left\|{f(t)\over t-x}|\kern .5pt t-z_c^{}|^{1-\alpha}\right\|_\infty. \label{option3} \end{equation} \begin{figure} \begin{center} \vskip 18pt \includegraphics[scale=1.2]{relupot.eps} \end{center} \caption{\label{relupot}The potential theory of exponential clustering (continuation of Fig.~{\rm \ref{relu}}). The first two rows (right column) show the function $|\phi(t)|$ of $(\ref{phidef})$ and the associated potential $u(t) = n^{-1} \log |\phi(t)|$ for the model problem $(\ref{modelprob0})$--$(\ref{modelprob})$. The third row shows the behavior along the real $s$-axis after the change of variables to $s = \log t$; the domain is now the infinite strip $0 < \hbox{\rm Im\kern .7pt} s < \pi$, with $u=0$ for $\hbox{\rm Im\kern .7pt} s = \pi$. The final row shows the charge density $\rho(s) = n u_n(s)/\pi$, where $u_n$ is the normal derivative of $u$ on the boundary of the strip. The intervals that matter (emphasized by solid rather than dashed lines) are $\varepsilon < |t| < 1$ in the $t$ variable and $\log \varepsilon < \hbox{\rm Re\kern .7pt} s < 0$ in the $s$ variable. Smaller values of $|t|$ and $s$ contribute negligibly to the integral {\rm (\ref{hermagain})}, and larger values are far from the singularity.} \end{figure} For simplicity let us assume that the singularity lies at $z_c^{} = 0$ and the part of the contour $\Gamma$ that matters is $[\kern .5pt 0,1]$, and let the domain be scaled so that $|\phi(x)|\approx 1$ for $x\in E$. We want $|\phi(t)|$ to be at most $1$ for $t<0$ and at least $(t/\varepsilon)^\alpha$ for $t> \varepsilon$, where $\varepsilon$ is the distance of the closest pole to the singularity. See the upper-right image of Fig.~\ref{relupot}. Defining $u(t) = n^{-1} \log \phi(t)$ leads us to the model problem sketched in the image below this in the figure: find a harmonic function $u(t)$ in the upper half $t$-plane such that \begin{equation} u(t) = \cases{ 0, & $t \le \varepsilon$, \cr\noalign{\vskip 3pt} \alpha \kern .5pt n^{-1}\log(t/\varepsilon) , & $t> \varepsilon$.\cr } \label{modelprob0} \end{equation} We now make the change of variables $s = \log t$, which transplants the Laplace problem to the infinite strip $S = \{s\in{\bf C}:\, 0 <\hbox{\rm Im\kern .7pt} s < \pi\}$, as sketched in the $(3,2)$ position of the figure: find a harmonic function $u(s)$ in $S$ satisfying \begin{equation} u(s) = \cases{ 0, & $\hbox{\rm Im\kern .7pt} s = \pi$, \cr\noalign{\vskip 3pt} 0, & $\hbox{\rm Im\kern .7pt} s = 0$ and $\hbox{\rm Re\kern .7pt} s \le \log \varepsilon$, \cr\noalign{\vskip 3pt} \alpha \kern .5pt n^{-1}(s-\log\varepsilon), & $\hbox{\rm Im\kern .7pt} s = 0$ and $\hbox{\rm Re\kern .7pt} s > \log \varepsilon$.\cr } \label{modelprob} \end{equation} This change of variables is convenient mathematically, and it is also important conceptually, since it is well known that influences on harmonic functions decay exponentially with distance along a strip. Consequently, if $\varepsilon$ is small, the solution to a Laplace problem for $\log \varepsilon \ll \hbox{\rm Re\kern .7pt} s \ll 0$ will be essentially (though not exactly) determined by the boundary conditions in that region. This just matches what we need for the model problem as posed in the original~$t$ variable, where behavior for $|t|$ of order $\varepsilon$ or less is unimportant because it contributes negligibly to the integral (\ref{hermagain}) and behavior for $|t|$ of order $1$ or more is unimportant because it is far from the singularity under investigation. So we address our attention to (\ref{modelprob}). An exact solution can be obtained via the Poisson integral formula for an infinite strip~\cite{widder}, \begin{equation} u(x+iy) = {\alpha \kern .5pt n^{-1}\over 2\pi} \int_0^\infty {\xi \sin(y)\over \,\cosh(\xi - (x-\log \varepsilon)) - \cos(y)\,} \kern 1pt d\xi, \label{solnexact} \end{equation} where we have set $s = x+iy$. However, we do not need exactly this since the region where our model applies is $\log \varepsilon \ll \hbox{\rm Re\kern .7pt} s \ll 0$. In this region, the bilinear harmonic function \begin{equation} u(x+iy) = \alpha\kern .5pt n^{-1} (1-{y\over \pi}) (x - \log\varepsilon) \label{solnapprox} \end{equation} satisfies the boundary conditions and is accordingly a good approximation to the solution to~(\ref{modelprob}). The corresponding pole density distribution on the real $x$ axis is $(n/\pi)$ times the normal derivative, \begin{equation} \rho(s) = - {n\over \pi} {\partial \over \partial y} u(x+iy) = {\alpha\over \pi^2} (x - \log\varepsilon). \end{equation} This linear growth, sketched in the bottom-right image of Fig.~\ref{relupot}, is just what we set out to explain in Fig.~\ref{relu}. Let us now look at the quantitative implications of this argument, comparing uniform exponential clustering (left column of Fig.~\ref{relupot}) with tapered exponential clustering (right column). According to our model, the integral of the solid portions of the $\rho(s)$ curves in the bottom should be equal to $n$, the total number of poles. For uniform clustering the integral is $(-\alpha\kern .5pt\log \varepsilon)^2/\pi^2$, leading to the estimates \begin{equation} \hbox{Closest pole:~~} \varepsilon \approx \exp(-\pi \sqrt{n/\alpha}\kern 1pt), \quad \hbox{Accuracy:~~} \varepsilon^\alpha \approx \exp(-\pi \sqrt{\alpha\kern .5pt n}\kern 1pt). \label{est1} \end{equation} For tapered clustering the integral is ${1\over 2}(-\alpha\kern .5pt\log \varepsilon)^2/\pi^2$, leading to the estimates \begin{equation} \hbox{Closest pole:~~} \varepsilon \approx \exp(-\pi \sqrt{2\kern .3pt n/\alpha}\kern 1pt), \quad \hbox{Accuracy:~~} \varepsilon^\alpha \approx \exp(-\pi \sqrt{2\kern .3pt \alpha\kern .5pt n}\kern 1pt). \label{est2} \end{equation} Thus, as mentioned in the abstract, our model leads to the prediction of a factor of 2 speedup with tapered clustering. It would be interesting to investigate whether, for certain problems, exactly this ratio could be established theoretically in the limit $n\to\infty$. As an example of a problem in which we may make such a comparison numerically, consider Fig.~\ref{fig12}. These data show $\infty$-norm errors for rational linear-minimax approximations of even degrees $n$ from 2 to 50 with preassigned exponentially clustered poles. That is, the approximations are optimal in the $\infty$-norm among rational functions in $R_n$ with simple poles at the prescribed points; they are characterized by error curves equioscillating between $n+2$ extrema. The upper curves correspond to uniformly clustered poles $p_k = -\exp(-\pi k/\sqrt n)$, $0\le k \le n-1$, and the lower curves to tapered poles $p_k = -\exp(\sqrt 2 \pi(\sqrt k-\sqrt n))$, $1\le k \le n$. The asymptotic errors appear to be about $\exp(-\sqrt{2.3 n})$ for uniform clustering and $\exp(-\sqrt{4.7 n})$ for tapered clustering. With $\alpha = 1/2$ for $f(x) = \sqrt x$, the corresponding estimates (\ref{est1}) and (\ref{est2}) are $\exp(-\sqrt{2.2 n})$ and $\exp(-\sqrt{4.4 n})$. \begin{figure} \vskip 1.4em \begin{center} \includegraphics[scale=.99]{fig12.eps} \end{center} \caption{\label{fig12}Linear-minimax approximation of $f(x) =\sqrt x$ on $[\kern .5pt 0,1]$ with preassigned exponentially clustered poles in $[-1,0\kern .5pt]$, $n = 2,4,\dots, 50$. Tapering the distribution makes the convergence rate approximately double, as predicted by the model of Section~$\ref{sec5}$.} \end{figure} Analyses related to the argument we have presented were published by Stahl for rational minimax approximation of $|x|$ on $[-1,1]$ and $x^\alpha$ on $[\kern .5pt 0,1]$~\cite{stahl92,stahl93,stahl94,stahl}. For $x^\alpha$ Stahl gives the result \begin{equation} \hbox{Accuracy:~~} \varepsilon^\alpha \approx \exp(-\pi \sqrt{4\kern .3pt \alpha\kern .5pt n}\kern 1pt), \label{stahlest2} \end{equation} which is not just an estimate but a theorem concerning the limit $n\to\infty$ (assuming $\alpha$ is not an integer), with precise constants. This is exactly what one would expect based on (\ref{est2}), since, as mentioned earlier, the effective value of $n$ is doubled in the case of true minimax approximants~\cite{rakh}. Stahl worked essentially in the variable $t$ rather than $s$, so his boundary conditions involved logarithms, as in the second image of the right column of Fig.~\ref{relupot}. Whenever one has a Laplace problem with Dirichlet boundary data, one can interpret it as the problem of finding an equipotential distribution in the presence of an external field defined by that boundary data, and this interpretation has been carried far in approximation theory~\cite{safftotik}. From this point of view one can say that tapered exponential clustering results from poles and zeros being slightly pushed away from a singular point by a logarithmic potential field. \section{\label{sec6}Exponential and double exponential quadrature} In this final section we turn to another problem where exponential clustering appears. Let $f$ be a continuous function on $[-1,1]$. We wish to approximate the integral of $f$ by a linear combination \begin{equation} I_n = \sum_{k=1}^n w_k f(x_k), \label{quadsum} \end{equation} where $\{x_k\}$ are distinct nodes in $[-1,1]$ and $\{w_k\}$ are corresponding weights, in such a way that the accuracy is good even if $f$ has branch point singularities at the endpoints. To this end, we introduce a change of variables $g(s)$ from the real line to $[-1,1]$, so that the integral becomes \begin{equation} I = \int_{-1}^1 f(x) \kern .7pt dx = \int_{-\infty}^\infty f(\kern .5pt g(s)) \kern .7pt g'(s) \kern .7pt ds, \label{cov} \end{equation} and we apply the equispaced trapezoidal rule. This involves an infinity of sample points in principle, but if $g'(s)$ decays rapidly, we may truncate these to an $n$-point rule like (\ref{traprule}): \begin{equation} I_n = \kern 2pt h \kern -6pt \sum_{k =-(n-1)/2}^{(n-1)/2} f(\kern .3pt g(kh)) \kern .7pt g'(kh). \label{DESE} \end{equation} Quadrature formulas of this kind were introduced around 1970 by Mori, Takahasi, and other Japanese researchers and also in the analysis of sinc methods by Stenger. See~\cite{haber,IMT,stengersurvey,stengerbook,tm73,trap}, as well as~\cite{mori} for the history as told by Mori himself. The standard ``exponential'' choice of $g$ is \begin{equation} g(s) = \tanh(s), \quad g'(s) = \hbox{sech}^2(s), \label{tanh} \end{equation} with which (\ref{DESE}) becomes the {\em tanh formula}. As in section 2, we estimate the truncation error as of order $\exp(-nh)$ and the discretization error of order $\exp(-\pi^2/h)$. (The latter could be worse if $f$ has additional singularities near $(-1,1)$.) This gives a balance $h \approx \pi/\sqrt n$, with convergence rate of order $\exp(-\pi\sqrt n\kern .7pt )$. An estimate of this form is valid for any H\"older continuous branch point singularity; see~\cite[Thm.~3.4]{stengersurvey}, \cite[Thm.~2.1]{tanaka}, and~\cite[Thm.~14.1]{trap}. Root-exponential convergence! This is much better than any algebraic order, but for practical applications on one-dimensional domains, methods of this kind often seem very wasteful, with almost all the points being used up in resolving the singularity (100$\%$ of them, in the limit $n\to \infty$)~\cite{sincfun}. A year or two after the first exponential formulas appeared, it was realized that one can do better with ``double exponential'' formulas. We focus on the {\em tanh-sinh formula\/} proposed by Takahasi and Mori in~\cite{tm74} and subsequently used and analyzed by many others including Okayama, Sugihara, and Tanaka as well as Bailey and Borwein~\cite{bb,ms01,oms,sugihara,tanaka}. Here (\ref{tanh}) is replaced by \begin{equation} g(s) = \tanh({\textstyle{\pi\over 2}} \sinh(s)), \quad g'(s) = {\textstyle{\pi\over 2}}\cosh(s)\kern 1.3pt \hbox{sech}^2({\textstyle{\pi\over 2}}\sinh(s)). \label{tanhsinh} \end{equation} Under suitable assumptions we can now estimate the truncation and discretization errors as of orders $\exp(-(\pi/2) \exp(nh/2))$ and $\exp(-\pi^2/h)$. The first of these estimates is the big improvement, for this quantity can be almost-exponentially small with a much smaller value of $h$ than before, of order $\log(n)/ n$ rather than $1/\sqrt n$. By {\em almost-exponential,} we mean of order\break $\exp(-C n/\log n)$ for some \hbox{$C>0$}. With this reduced value of $h$, the second estimate becomes almost-exponentially small too. \begin{figure} \begin{center} \vspace{12pt} \includegraphics[scale=.96]{fig13.eps} \vspace{-7pt} \end{center} \caption{\label{fig13} On the left, root-exponential convergence of the tanh quadrature formula applied to integration of $\sqrt{1+x}$ (note the $\sqrt n$ axis as usual); the tanh-sinh formula converges much faster down to machine precision. On the right, the distances of nodes from poles (with a $\sqrt k$ axis) show uniform exponential clustering for the tanh formula with $n=40$ and tapered exponential clustering for tanh-sinh.} \end{figure} Figure~\ref{fig13} shows data for the tanh and tanh-sinh formulas. (We used the empirical choices $h = \pi/\sqrt n$ and $h = 1.2\log(2\pi n)/n$, respectively.) The left image plots $|I_n- I|$ against $\sqrt n$ for $n$ from $1$ to $40$ for the integration of $f(x) = \sqrt{1+x}$. The tanh curve appears straight, confirming the root-exponential convergence, and the tanh-sinh curve bends downward, confirming that its rate is faster. The unexpected image is on the right, a plot of distances of the nodes from the endpoint $x=-1$. For tanh quadrature, these distances are uniformly exponentially spaced, appearing as a parabola on these axes. The curve for tanh-sinh quadrature, however, is almost perfectly straight. It would seem that tanh-sinh quadrature exploits tapered exponential clustering! It surprised us when we first saw curves like this. Why is there a resemblance between the tanh and tanh-sinh quadrature formulas and the phenomena of rational approximation discussed in the earlier sections of this article? Some steps toward an answer come from a beautiful connection introduced by Gauss and exploited by Takahasi and Mori~\cite{gauss,tm71,gaussCC}: every quadrature formula can be associated with a rational approximation. Suppose first that $f$ can be analytically continued to a neighborhood $\Omega$ of $[-1,1]$ bounded by a contour $\Gamma$. Then the integral can be written \begin{equation} I = \int_{-1}^1 f(x) \kern .7pt dx = {1\over 2\pi i} \int_\Gamma f(t) \kern .7pt \varphi(t)\kern .7pt dt, \label{Iint} \end{equation} where the {\em characteristic function} $\varphi$ is defined by \begin{equation} \varphi(t) = \int_{-1}^1 {dx\over t-x} = \log{t+1\over t-1}. \end{equation} On the other hand the quadrature sum (\ref{quadsum}) can be written \begin{equation} I_n = {1\over 2\pi i} \int_\Gamma f(t) \kern .7pt r(t)\kern .7pt dt, \label{Inint} \end{equation} where $r$ is the degree $n$ rational function defined by \begin{equation} r(t) = \sum_{k=1}^n {w_k\over t-x_k}. \label{ratdef} \end{equation} Subtracting (\ref{Inint}) from (\ref{Iint}) gives what we call the {\em Gauss--Takahasi--Mori (GTM) contour integral,} \begin{equation} I - I_n = {1\over 2\pi i} \int_\Gamma^{} f(t) \kern .7pt (\varphi(t)-r(t)) \kern .7pt dt \label{rel1} \end{equation} and the corresponding error bound \begin{equation} | I - I_n | \le {1\over 2\pi} \|f\|_\infty^{}\kern .7pt \|\varphi - r\|_\infty^{}\kern .7pt \|1\|_1^{}, \label{rel2} \end{equation} which we have written in the style of (\ref{option1}), with the norms defined over $\Gamma$. Equations (\ref{rel1}) and (\ref{rel2}) relate accuracy of a quadrature formula to an approximation problem: if the nodes and weights are such that $\varphi - r$ is small on the boundary $\Gamma$ of a region where $f$ is analytic, then $|I-I_n|$ must be small. This reasoning was applied by Takahasi and Mori to a range of quadrature formulas~\cite{tm71}. Now $\varphi$ is an analytic function in the extended complex plane minus the segment $[-1,1]$. It follows that so long as $\Gamma$ is disjoint from $[-1,1]$, there exist rational approximations to $\varphi$ that converge exponentially on $\Gamma$ as $n\to\infty$. In particular, this holds for the rational functions associated with Gauss and Clenshaw--Curtis quadrature~\cite{gaussCC}, where it is convenient to take $\Gamma$ in the form of an ellipse about $[-1,1]$ with foci $\pm 1$. It follows that both these quadrature formulas converge exponentially as $n\to\infty$ for analytic integrands (cf.\ \cite[Thm.~19.3]{atap}). But what if $f$ has endpoint singularities? Now $\Gamma$ must touch $[-1,1]$ at the endpoints, and (\ref{rel2}) fails just as (\ref{option1}) did in such a case. In fact, this failure is more severe, since $\|\varphi - r\|_\infty = \infty$ for any $r$ because of the logarithmic singularities of $\varphi$. The last section, however, suggests a solution. Instead of (\ref{rel2}), we can derive from (\ref{rel1}) the bound \begin{equation} | I - I_n | \le {1\over 2\pi} \kern .7pt\|f\|_\infty{} \|\varphi - r\|_1^{} \kern 1pt . \label{rel3} \end{equation} The switch from the $\infty$- to the $1$-norm changes the problem of rational approximation of $\varphi$ profoundly. Since the dominant effects just concern approximation of a logarithmic singularity near the singular point, the essential question becomes, how fast can $\log t$ be approximated by rational functions over both sides of the interval $[-1,0\kern .5pt]$ in the $1$-norm? \begin{figure} \begin{center} \vspace{12pt} \includegraphics[scale=1]{fig14.eps} \vspace{-7pt} \end{center} \caption{\label{fig14} Error $|\varphi(t)-r(t)|$ as a function of distance to the left from $t=-1$ for the tanh and tanh-sinh approximations with $n=40$ with $h = \pi/\sqrt n$ and $h = 1.2\log(2\pi n)/n$, respectively. By symmetry, the same behavior would appear to the right from $t=1$. Compare Fig.~\ref{figmini}, where the ratio of the blue and red values in the lower-right image is closely analogous to the blue curve here. The $1$-norms of the approximation errors over $[-2,-1]$ are indicated. The slight irregularities at the left are the result of rounding error.} \end{figure} As we did with Fig.~\ref{figmini}, let us get some insight by looking at the details of the approximation problem. The rational function (\ref{ratdef}) for the tanh rule is \begin{equation} r(t) = \kern 2pt h \kern -6pt \sum_{k =-(n-1)/2}^{(n-1)/2} {\hbox{sech}^2(kh) \over t - \tanh(kh)}, \label{tanhrule} \end{equation} and for the tanh-sinh rule, it is \begin{equation} r(t) = \kern 2pt h \kern -6pt \sum_{k =-(n-1)/2}^{(n-1)/2} {{\textstyle{\pi\over 2}}\kern -.5pt \cosh(kh) \kern 1.3pt\hbox{sech}^2({\textstyle{\pi\over 2}}\sinh(kh)) \over t - \tanh({\textstyle{\pi\over 2}}\sinh(kh))}. \label{tanhsinhrule} \end{equation} Figure~\ref{fig14} plots $|\kern .5pt \varphi(t)-r(t)|$ for these two approximations. For tanh quadrature, we know that $|\kern .5pt \varphi(t)-r(t)|$ must diverge to $\infty$ as $t\to -1$ because of the log singularity of $\varphi$ at $t=-1$. Yet the singularity is so weak that the divergence only shows up as a gentle upward drift in the blue curve at the left. Over the main part of the plot, $|\kern .5pt\varphi(t)-r(t)|$ decreases steadily down to around $10^{-7}$. The $1$-norm, measured here over $t\in [-2,-1]$, is consequently very small, confirming via (\ref{rel3}) the high accuracy of this quadrature rule. As $n\to\infty$, this $1$-norm decays root-exponentially. For tanh-sinh quadrature, again no approximation of $\varphi$ is possible in the $\infty$-norm. In the $1$-norm, however, one might expect that the convergence will now be almost-exponential. Indeed, $|\kern .5pt \varphi - r|$ decays almost-exponentially as $n\to\infty$ over any domain bounded away from the singularity. But the 1-norm decay over the whole interval is in fact just root-exponential, as is suggested by the number listed being barely smaller than before. The following reasoning suggests why this must be. Consider approximation of $f(x) = \log x$ on $[\kern .5pt 0, 1]$. Suppose rational approximations existed with faster than root-exponential convergence in the $1$-norm. Then by integrating, we would get rational approximations to $g(x) = x\log x - x$ with faster than root-exponential convergence in the $\infty$-norm, which would contradict the evidence of Fig.~\ref{fig2}. If $\|\kern .7pt \varphi - r\kern .7pt \|_1$ decreases only root-exponentially as $n\to\infty$, how does the quadrature formula converge almost-exponentially? It appears that this depends on additional properties that go beyond rational approximation, involving analytic continuation of the integrand onto an infinitely-sheeted Riemann surface in exponentially small neighborhoods of the endpoints~\cite{sugihara,tanaka}. There remains the phenomenon of tapered exponential clustering, so vividly evident in Fig.~\ref{fig13}. We do not yet have an explanation for this, nor a view of whether an approximate $\sqrt k$ dependence is genuine or just an artifact. This is a topic for ongoing research, where it would be good to investigate also the distributions of exponentially clustered nodes, also apparently tapered, that arise with the ``universal quadrature'' formulas of Bremer, et al.~\cite{brs,serkh}. \section{Discussion} \label{sec-discussion} Exponential clustering of poles at singularities has been part of the landscape of rational approximation for half a century, but we believe this is the first study to focus on this effect. Our motivation is that this clustering is what makes rational approximations so powerful, and understanding it enables one to improve existing numerical algorithms and develop new ones. We find these phenomena fascinating, especially the tapered clustering effect, and discovering that tapering also appears in double exponential quadrature was a bonus. The elucidation of these matters with the help of a sometimes seemingly endless program of numerical experiments will forever be associated in our minds with the Covid-19 shutdowns of 2020. Here are some details of our computations. Figures~\ref{fig1}, \ref{fig2}, \ref{fig6} and~\ref{figmini} made use of the Chebfun {\tt minimax} command~\cite{minimax}, principally due to Silviu Filip, and Filip also kindly provided us with a modified code for the weighted minimax approximations of Figs.~\ref{fig2} and~\ref{fig6}.\ \ For successful results in some of these problems, we applied a M\"obius transformation of $[\kern .5pt 0,1]$ to itself to weaken the singularity while preserving the space $R_n$. For the approximations of Figs.~\ref{fig2} and~\ref{fig6} on a complex disk, the AAA-Lawson algorithm was used as implemented in Chebfun~\cite{chebfun,lawson}, again with a M\"obius transformation. Figure~\ref{fig3} was produced with the {\tt confmap} code available at~\cite{laplace}, which in turn calls {\tt aaa} from Chebfun~\cite{aaa} and {\tt laplace} from~\cite{laplace}. The {\tt aaa} code was also used directly in Figs.~\ref{fig1} and~\ref{fig6}, and {\tt laplace} in Fig.~\ref{fig5}. The Stokes and Helmholtz results of Fig.~\ref{fig5} were produced by experimental codes that are not yet publicly available developed with Abi Gopal and Pablo Brubeck, respectively. In Fig.~\ref{fig12}, a least-squares problem was extended by a Lawson iteration (iteratively reweighted least-squares) to compute minimax approximations with preassigned poles. All the remaining results are based on straightforward computations in \hbox{MATLAB}\ and Chebfun. \begin{acknowledgements} We have benefited from helpful advice from Bernd Beckermann, Pablo Brubeck, Silviu Filip, Abi Gopal, Stefan G\"uttel, Arno Kuijlaars, Andrei Mart\'inez-Finkelshtein, Ed Saff, Kirill Serkh, Alex Townsend, and Heather Wilber. \end{acknowledgements}

https://arxiv.org/abs/2002.04664

Universal Average-Case Optimality of Polyak Momentum

Polyak momentum (PM), also known as the heavy-ball method, is a widely used optimization method that enjoys an asymptotic optimal worst-case complexity on quadratic objectives. However, its remarkable empirical success is not fully explained by this optimality, as the worst-case analysis -- contrary to the average-case -- is not representative of the expected complexity of an algorithm. In this work we establish a novel link between PM and the average-case analysis. Our main contribution is to prove that any optimal average-case method converges in the number of iterations to PM, under mild assumptions. This brings a new perspective on this classical method, showing that PM is asymptotically both worst-case and average-case optimal.

\section{Introduction} Polyak momentum (PM), also known as the heavy-ball method, is a widely used optimization method. Originally developed to solve linear equations~\citep{frankel1950convergence, rutishauser1959theory}, it was generalized to smooth functions and popularized in the optimization community by Boris Polyak~\citep{polyak1964some,polyak1987introduction}. This method has seen a renewed interest in recent years, as its stochastic variant which replaces the gradient with a stochastic estimate is effective on deep learning problems~\citep{sutskever2013importance, zhang20202dive}. PM also enjoys a locally optimal rate of convergence for strongly convex and twice differentiable objectives. As is common within the optimization literature, this optimality is relative to the \emph{worst-case} analysis, that provides complexity bounds for \emph{any} input from a function class, no matter how unlikely. Despite its widespread use, the worst-case is not representative of the typical behavior of optimization methods. The simplex method, for example, has a worst-case exponential complexity, that becomes polynomial when considering the average-case \citep{spielman2004smoothed}. A more representative analysis of the typical behavior is given by the \emph{average-case} complexity, which averages the algorithm's complexity over all possible inputs. The average-case analysis is standard for analyzing sorting~\citep{knuth1997art} and cryptography~\citep{katz2014introduction} algorithms, to name a few. However, little is known of the average-complexity of optimization algorithms, whose analysis depends on the often unknown probability distribution over the inputs. The recent work of \citet{pedregosa2020acceleration, lacotte2020optimal} overcame this dependency on the input probability distribution through the use of random matrix theory techniques. In the same papers, the authors noticed the convergence of some optimal average-case methods to PM, as the number of iterations grows (see Figure~\ref{fig:convergence_pm}). This is rather surprising given their crucial differences. For instance, average-case optimal methods use knowledge of the full spectral distribution, while PM only requires knowledge of its edges (i.e., smallest and largest eigenvalue). Since this convergence was only shown on specific methods, it raises the question on whether this is a spurious phenomenon or if this holds more generally: \begin{myenv}{Conjecture} \centering As the number of iterations grows, all average-case optimal methods converge to Polyak momentum. \vspace{-0.8em}\vphantom{1} \end{myenv} \begin{figure*} \centering \includegraphics[width=1\linewidth]{figures/polyak_convergence.pdf} \caption{{\bfseries Convergence of optimal average-case methods to Polyak Momentum}. For the Marchenko-Pastur and uniform distribution of eigenvalues (left), we construct the method that has optimal average-case complexity and plot the momentum (middle) and steps-size (right) parameters. For the two methods considered, the momentum and step-size parameters converge as the number of iterations grows to those of Polyak momentum, displayed here as a straight line. } \label{fig:convergence_pm} \end{figure*} The {\bfseries main contribution} of this paper is to give a positive answer to this conjecture. The main, but not so restrictive assumption, is that the probability density function of the eigenvalues is non-zero on the interval containing its support. With this we can show the previously unknown property that PM is \emph{asymptotically} optimal under the average-case analysis, bringing a new perspectiveon the remarkable empirical performance of this classical method. Furthermore, this statement is \emph{universal}, i.e., independent of the probability distribution over the inputs. \subsection{Related work} This work draws from the fields of optimization, complexity analysis and orthogonal polynomials, of which we comment on the most closely related ideas. \paragraph{Average-case analysis.} The average-case analysis has a long history in computer science and numerical analysis. Often it is used to justify the superior performances of algorithms such as Quicksort \citep{Hoare1962Quicksort} and the simplex method in linear programming~\citep{spielman2004smoothed}. Despite this rich history, it's challenging to transfer these ideas into continuous optimization due to the ill-defined notion of a typical continuous optimization problem. In the context of optimization, \citet{pedregosa2020acceleration} derived a framework for analyzing the average-case gradient-based methods and developed methods that are non-asymptotic optimal algorithms with respect to the average-case. Such average-case analysis finds applications in various domains. For instance, \citet{lacotte2020optimal} use this framework to derive optimal average-case algorithms to minimize least-squares with random matrix sketching. Prior to this stream of papers, \citet{berthier2018accelerated} use methods based on Jacobi polynomials to design average-case optimal gossip methods, but without generalizing the framework. In the numerical analysis literature,~\citet{deift2019conjugate} have recently developed an average-case complexity of conjugate gradient. \paragraph{Asymptotics or orthonormal polynomials.} A key ingredient of the proof are asymptotics or orthonormal polynomials. This is a vast subject with applications in stochastic processes~\citep{grenander1958toeplitz}, random matrix theory~\citep{deift1999orthogonal} and numerical integration~\citep{mhaskar1997introduction} to name a few. The monograph of \citep{lubinsky2000asymptotics} discusses all results used in this paper. \paragraph{Notation.} Throughout the paper we denote vectors in lowercase boldface (${\boldsymbol x}$), matrices in uppercase boldface letters ($\boldsymbol H$), and polynomials in uppercase latin letter ($P, Q$). We will sometimes omit integration variable, with the understanding that $\int \varphi \mathop{}\!\mathrm{d} \mu$ is a shorthand for $\int \varphi(\lambda) \mathop{}\!\mathrm{d}\mu(\lambda)$. \section{Average-Case Analysis of Gradient-Based Methods}\label{scs:average_analysis} The goal of the average-case analysis is to quantify the expected error ${\mathbb{E}\,} \|{\boldsymbol x}_t - {\boldsymbol x}^\star\|^2$, where ${\boldsymbol x}_t$ is the $t$-th update of some optimization method and the expectation is taken over all possible problem instances. To make this analysis tractable, and following \citep{pedregosa2020acceleration}, we consider quadratic optimization problems of the form \begin{empheq}[box=\mybluebox]{equation*}\tag{OPT}\label{eq:quad_optim} \vphantom{\sum_0^i}\min_{{\boldsymbol x} \in {\mathbb R}^d} \Big\{ f({\boldsymbol x}) \stackrel{\text{def}}{=}\!\mfrac{1}{2}({\boldsymbol x}\!-\!{\boldsymbol x}^\star)^\top\!{\boldsymbol H}({\boldsymbol x}\!-\!{\boldsymbol x}^\star) \Big\}, \end{empheq} where ${\boldsymbol H} \in {\mathbb R}^{d \times d}$ is a \textit{random} symmetric positive-definite matrix and ${\boldsymbol x}^\star$ is a \textit{random} $d$-dimensional vector which is a solution of \eqref{eq:quad_optim}. \begin{remark} Problem \eqref{eq:quad_optim} subsumes the quadratic minimization problem $\min_{{\boldsymbol x}} {\boldsymbol x}^\top{\boldsymbol H}{\boldsymbol x} + {\boldsymbol b}^\top {\boldsymbol x} + c$ but the notation above will be more convenient for our purposes. \end{remark} \begin{remark} The expectation in ${\mathbb{E}\,} \|{\boldsymbol x}_t - {\boldsymbol x}^\star\|^2$ is over the inputs and not over any randomness of the algorithm, as is common in the stochastic literature. In this paper we only consider deterministic algorithms. \end{remark} We consider in this paper the class of \textit{first order methods}, which build ${\boldsymbol x}_t$ using a pre-defined linear combination of an initial guess and previous gradients: \begin{equation} {\boldsymbol x}_{t} \in {\boldsymbol x}_0 + \textbf{span}\{\nabla f({\boldsymbol x}_0),\;\ldots,\;\nabla f({\boldsymbol x}_t)\}. \label{eq:first_order_definition} \end{equation} This wide class includes most gradient-based optimization methods, such as gradient descent and momentum. However, it excludes quasi-Newton methods, preconditioned gradient descent or Adam (to cite a few), as the preconditioning allows the iterates to go outside span. \subsection{Tools of the trade: orthogonal polynomials and spectral densities} Average-case optimal methods rely on two key concepts that we now introduce: \textit{residual orthogonal polynomials} and the \textit{expected spectral distribution}. \subsubsection{Orthogonal (residual) polynomials} This section defines orthogonal polynomials and residual polynomials. \begin{definition} \label{def:orthogonal_polynomial} Let $\alpha$ be a non-decreasing function such that $\int Q\mathop{}\!\mathrm{d}\alpha$ is finite for all polynomials $Q$. We will say that the sequence of polynomials $P_0, P_1, \ldots$ is orthogonal with respect to $\mathop{}\!\mathrm{d}\alpha$ if $P_i$ has degree $i$ and \begin{equation}\label{eq:optimal_orthogonal_polynomials} \int_{\mathbb{R}} P_i\,P_j\mathop{}\!\mathrm{d}\alpha \begin{cases} = 0 & \text{if } i\neq j \\ > 0 & \text{if } i = j \end{cases}. \end{equation} Furthermore, if they verify $P_i(0) = 1$ for all $i$, we call these {\bfseries residual orthogonal} polynomials. \end{definition} Residual orthogonal polynomials verify a three-term recurrence \citep[\S 2.4]{fischer1996polynomial}, that is, there exists a sequence of real values $a_t, b_t$ such that \begin{equation}\label{eq:recurence_orthogonal_polynomials} P_{t}(\lambda) = (a_t + b_t \lambda) P_{t-1}(\lambda) + (1-a_{t})P_{t-2}(\lambda)\,, \end{equation} where $P_0(\lambda) = 1$ and $P_1(\lambda) = 1+b_1\lambda$\,. \subsubsection{Expected spectral distribution} The expected spectral distribution and the extreme eigenvalues of the matrix ${\boldsymbol H}$ play similar roles in the case of, respectively, average-case and worst-case optimal methods. They measure the problem's difficulty and define the optimal method's parameters. \begin{definition}[Empirical/Expected Spectral Measure] Let ${\boldsymbol H}$ be a random matrix with eigenvalues $\{\lambda_1, \ldots, \lambda_d\}$. The {\bfseries empirical spectral measure} of ${\boldsymbol H}$, called ${\mu}_{{\boldsymbol H}}$, is the probability measure \begin{equation}\label{eq:empirical_spectral_density} \mu_{{\boldsymbol H}}(\lambda) \stackrel{\text{def}}{=} {\textstyle{\frac{1}{d}\sum_{i=1}^d}} \delta_{\lambda_i}(\lambda) ~, \end{equation} where $\delta_{\lambda_i}$ is the Dirac delta, a distribution equal to zero everywhere except at $\lambda_i$ and whose integral over the entire real line is equal to one. Since ${\boldsymbol H}$ is random, the empirical spectral measure $\mu_{{\boldsymbol H}}$ is a random measure. Its expectation over ${\boldsymbol H}$, \begin{equation} \mu \stackrel{\text{def}}{=} {\mathbb{E}\,}_{{\boldsymbol H}}[\mu_{{\boldsymbol H}}]\,, \end{equation} is called the {\bfseries expected spectral distribution}. \end{definition} \begin{example}[Marchenko-Pastur density and large least squares problems] Consider a matrix ${\boldsymbol A} \in {\mathbb R}^{n \times d}$, where each entry is an iid random variable with mean zero and variance $\sigma^2$. Then it is known that the expected spectral distribution of ${\boldsymbol H} = \frac{1}{n}{\boldsymbol A}^\top\!{\boldsymbol A}$ converges to to the Marchenko-Pastur distribution \citep{marchenko1967distribution} as $n$ and $d \to \infty$ at a rate in which the asymptotic ratio $d / n \rightarrow r$ is finite. The Marchenko-Pastur distribution $\dif\mu_{\mathrm{MP}}$ is defined as \begin{equation} \label{eq:mp_distribution} \max\{1 - \tfrac{1}{r}, 0\}\delta_0(\lambda) + \frac{\sqrt{(L - \lambda)(\lambda - \ell)}}{2 \pi \sigma^2 r \lambda}1_{\lambda \in [\ell, L]}\,. \end{equation} Here $\ell\stackrel{\text{def}}{=} \sigma^2(1 - \sqrt{r})^2$, $L \stackrel{\text{def}}{=} \sigma^2(1 + \sqrt{r})^2$ are the extreme nonzero eigenvalues, $\delta_0$ is a Dirac delta at zero (which disappears if $r \geq 1$) and $1_{\lambda \in [\ell, L]}$ is a rectangular window function, equal to 1 for $\lambda \in [\ell, L]$ and 0 elsewhere. \end{example} \subsection{Average-case optimal methods} With these two ingredients, we can construct the method with optimal average-case complexity. We rewrite the expected error as an integral with weight function the expected spectral density $\mu$. \begin{theorem}\label{thm:pedregosa_rate}\citep{pedregosa2020acceleration} Assume ${\boldsymbol x}_0$, ${\boldsymbol x}^{\star}$ are random variables independent of ${\boldsymbol H}$, satisfying $\mathbb{E}[({\boldsymbol x}_0-{\boldsymbol x}^\star)({\boldsymbol x}_0-{\boldsymbol x}^\star)^\top] = R^2\boldsymbol{I}$. Let ${\boldsymbol x}_t$ be generated by a first-order method, associated to the polynomial $P_t$. Then the expected error at iteration $t$ reads \begin{empheq}{equation}\label{eq:error_norm_x} \vphantom{\sum_0^i}\mathbb{E}\|{\boldsymbol x}_t-{\boldsymbol x}^\star\|^2 = {\overbrace{R^2\vphantom{R_t}}^{\text{initialization}}}\int_{\mathbb R} {\underbrace{P_t^2}_{\text{algorithm}}} {\overbrace{\mathop{}\!\mathrm{d}\mu}^{\text{problem}}} \,. \end{empheq} \end{theorem} The optimal first order method is obtained by minimizing the above identity over the space of residual polynomials of degree~$t$. This turns out to be equivalent to finding a sequence of residual polynomials $\{P_i\}$ orthogonal w.r.t. the weight function $\lambda \mu(\lambda)$, as shown in the following theorem. \begin{theorem}\label{thm:optimal_polynomial} \citep{pedregosa2020acceleration} Let $a_t$ and $b_t$ be the coefficients of the three-term recurrence \eqref{eq:recurence_orthogonal_polynomials} for the sequence of residual polynomials orthogonal w.r.t. $\lambda \mathop{}\!\mathrm{d}\mu(\lambda)$. Then the following method has optimal average-case complexity over the class of problems \eqref{eq:quad_optim}:\footnote{Throughout the paper, we will color-code {\color{colormomentum} momentum} and {\color{colorstepsize}step-size} parameters.} \begin{align} \label{eq:optimal_algorithm} &{\boldsymbol x}_1 = {\boldsymbol x}_0 + {\color{colorstepsize}b_1} \nabla f({\boldsymbol x}_0),\\ &{\boldsymbol x}_{t} = {\boldsymbol x}_{t-1} + {\color{colormomentum}(a_t-1)}({\boldsymbol x}_{t-1}-{\boldsymbol x}_{t-2}) + {\color{colorstepsize}b_t} \nabla f({\boldsymbol x}_{t-1})\nonumber\,. \end{align} \end{theorem} Due to the dependency of the coefficients $a_t, b_t$ on the expected spectral distribution, equation \eqref{eq:optimal_algorithm} does not represents a single scheme, but rather a family of algorithms: each different expected spectral distribution generates a different optimal method. Below is an example of such optimal algorithm w.r.t the Marchenko-Pastur expected spectral distribution. \begin{example}[Marchenko-Pastur acceleration]\label{ex:mp} Let $\mathop{}\!\mathrm{d}\mu$ be the density associated with the Marchenko-Pastur distribution. Then, the recurrence of the optimal average-case method associated with this distribution is \begin{equation*}\label{algo:mp_algo} \begin{split} &\rho = \mfrac{1+r}{\sqrt{r}}\,,~\delta_{0} = 0;\\ & {\boldsymbol x}_1 = {\boldsymbol x}_0-{\color{colorstepsize}\mfrac{1}{(1+r)\sigma^2}}\nabla f({\boldsymbol x}_0)\,;~\\ &\delta_{t} = -({\rho+\delta_{t-1}})^{-1}\,; \\ &{\boldsymbol x}_{t} = {\boldsymbol x}_{t-1} + {\color{colormomentum}\left(1 + \rho\delta_t\right)}({\boldsymbol x}_{t-2} - {\boldsymbol x}_{t-1}) + {\color{colorstepsize}\mfrac{ \delta_t}{\sigma^2\sqrt{r}}}\nabla f({\boldsymbol x}_{t-1})\,. \end{split} \end{equation*} The coefficients come from the orthogonal polynomials w.r.t. $\lambda \mathop{}\!\mathrm{d}\mu(\lambda)$, which is a shifted Chebyshev polynomials of the second kind. \end{example} \subsection{Polyak Momentum and worst-case optimality} The Polyak momentum algorithm~\citep{polyak1964some} has an optimal worst-case convergence rate over the class of first order methods with constant coefficients~\citep{polyak1987introduction, scieur2017integration}. The method requires knowledge of the smallest and largest eigenvalue of the Hessian ${\boldsymbol H}$ (denoted $\ell$ and $L$ respectively) and iterates as follows: \begin{align}\label{algo:pm_algo}\tag{PM} &{\boldsymbol x}_1 = {\boldsymbol x}_0 - {\color{brown}\mfrac{2}{L + \ell}}\nabla f({\boldsymbol x}_0)\\ &{\boldsymbol x}_{t+1} = {\boldsymbol x}_t + {\color{colormomentum}\textstyle{\Big(\frac{\sqrt{L}-\sqrt{\ell}}{\sqrt{L}+\sqrt{\ell}}\Big)^2}}({\boldsymbol x}_{t} - {\boldsymbol x}_{t-1}) - {\color{colorstepsize} \textstyle\Big(\frac{ 2}{\sqrt{L}+\sqrt{\ell}}\Big)^2}\nabla \nonumber f({\boldsymbol x}_t) \end{align} \vspace{0.5em}\begin{remark} Unlike the Marchenko-Pastur accelerated method of Example~\ref{ex:mp}, coefficients of this method are constant in the iterations. Furthermore, these coefficients only depend on the edges of the spectral distribution and not on the full density. \end{remark} \break \section{All Roads Lead to Polyak Momentum}\label{scs:asymptotic_heavyball} \begin{myenv}{Main result} \begin{theorem}\label{thm:conv_hb} Assume the density function $\mathop{}\!\mathrm{d}\mu$ is strictly positive in the interval $[\ell, L]$ with $\ell > 0$ and let ${a_t}, b_t$ be the parameters of the optimal average-case method (Theorem~\ref{thm:optimal_polynomial}). Then these parameters converge to those of \eqref{algo:pm_algo}. More precisely, we have the limits: \begin{align} \label{eq:limit_method} &\lim_{t \to \infty} {\color{colormomentum} a_t-1} = \underbrace{\left(\mfrac{\sqrt{L}-\sqrt{\ell}}{\sqrt{L}+\sqrt{\ell}}\right)^2}_{=\,\text{\eqref{algo:pm_algo} momentum}}\,, \;\;\quad\text{ and }\\ &\lim_{t \to \infty}{\color{colorstepsize} b_t} = \underbrace{- \left(\mfrac{2}{\sqrt{L} + \sqrt{\ell}}\right)^2}_{=\,\text{\eqref{algo:pm_algo} step-size}}\,. \end{align} ~ \end{theorem} \end{myenv} The key insight of the proof is to cast the three-term recurrence of residual \textit{orthogonal} polynomials into orthonormal polynomials\footnote{A sequence $Q_1, Q_2, ...$ of orthogonal polynomials with respect to $\mathop{}\!\mathrm{d}\omega$ is orthonormal if $\int Q_i^2 \mathop{}\!\mathrm{d} \omega = 1$.} in the interval $[-1,\,1]$. Once this is done, we will use asymptotic properties of these polynomials. The proof is split into three steps. \begin{itemize}[leftmargin=*] \item Step 0 introduces notation and some known results. \item Step 1 writes the coefficients of optimal average-case methods in terms of properties of a class of orthonormal polynomials in the $[-1, 1]$ interval. \item Step 2 computes the limits of the expressions derived in the previous step by using known asymptotic properties of orthonormal polynomials. \end{itemize} {\bfseries \underline{Step 0}: Definitions.} In the classical theory of orthogonal polynomials, the weight function associated with orthogonal polynomials is defined in the interval $[-1, 1]$. However, in our case the spectral densities are instead defined in $[\ell, L]$. To translate results from one setting to the other we define the following linear mapping from $[\ell, L]$ to $[-1, 1]$: \begin{equation} m(\lambda) = \mfrac{L+\ell}{L - \ell} - \mfrac{2}{L-\ell}\lambda\,. \end{equation} For notational convenience, we will also use the shorthand $m_0 \stackrel{\text{def}}{=} m(0)$. We now define $Q_i(m(\cdot))$ as the $i$-th degree orthonormal polynomial with respect to the weight function $\lambda\mu(\lambda)$. That is, the sequence $Q_1, Q_2, \ldots$ verifies \begin{equation} \int_{\ell}^L Q_i(m(\lambda))Q_j(m(\lambda)) \lambda \mathop{}\!\mathrm{d}\mu(\lambda) = \begin{cases} 1 \text{ if $i=j$} \\ 0 \text{ otherwise\,.}\end{cases}\,, \end{equation} where $\delta_{ij}$ represents Kronecker's delta. Like residual orthogonal polynomials, orthonormal polynomials also verify a three-term recurrence relation. This time, the relation is of the form \begin{equation}\label{eq:recurrence_orthonormal_poly} \alpha_t Q_t(\xi) = (\xi - \beta_t)Q_{t-1}(\xi) - \alpha_{t-1}Q_{t-2}(\xi)\,, \end{equation} and depends on coefficients $\alpha_t, \beta_t$: {\bfseries \underline{Step 1}: Parameters of optimal method and orthonormal polynomials}. In this step we derive the recurrence relation for an \emph{orthonormal} family with respect to the density $\mathop{}\!\mathrm{d}\nu$. This will allow us to use existing results on the asymptotics of orthonormal polynomials. \begin{lemma}\label{lemma:orthonormal_a_b} Let $a_t, b_t$ be the parameters associated with optimal average-case method (Theorem~\ref{thm:optimal_polynomial}). These coefficients verify the following identity, \begin{align}\label{eq:reformation_a_b} &{\color{colormomentum}1 - a_t} = -\frac{\alpha_{t-1}}{ \alpha_t}\frac{Q_{t-2}(m_0)}{ Q_t(m_0)} ~, \quad \text{ and }\\ &{\color{colorstepsize} b_t} = - \frac{2}{\alpha_t (L - \ell)} \frac{Q_{t-1}(m_0)}{Q_{t}(m_0)}~. \end{align} \end{lemma} \begin{proof} Since orthogonality is preserved after multiplication by a scalar, the polynomial $P_t(\lambda) \stackrel{\text{def}}{=} Q_t(m(\lambda)) / Q_t(m_0)$ is also orthogonal with respect to the weight function $\lambda \mathop{}\!\mathrm{d}\mu(\lambda)$. The normalization $1/Q_t(m_0)$ ensures $P_t$ is a residual polynomial. Note that $Q_t(m_0)$ cannot be zero because $m_0$ lies outside of the weight function's support $[-1,\,1]$. Using Theorem~\ref{thm:optimal_polynomial}, the coefficients of the optimal average-case method can be derived from the three-term recurrence of this polynomial. Indeed, starting from the three-term recurrence of $Q_i$ \eqref{eq:recurrence_orthonormal_poly}, we obtain for $P_t$ \begin{align*} P_t(\lambda) &= (m(\lambda) - \beta_{t-1})\mfrac{Q_{t-1}(m(\lambda))}{ \alpha_t \, Q_t(m_0)} - \alpha_{t-1}\mfrac{Q_{t-2}(m(\lambda))}{ \alpha_t \, Q_t(m_0)}\\ & = \underbrace{\mfrac{1}{\alpha_t}\left(\mfrac{L+\ell}{L - \ell} - \beta_{t-1} - \mfrac{2}{L-\ell}\lambda\right)\mfrac{Q_{t-1}(m_0)}{ \, Q_t(m_0) }}_{= (a_t + b_t\lambda)} P_{t-1}\\ &\qquad - \underbrace{\mfrac{\alpha_{t-1}}{ \alpha_t}\mfrac{Q_{t-2}(m_0)}{ Q_t(m_0)}}_{=-(1 - a_t)} P_{t-2}(\lambda)\,, \end{align*} where in the last line we used the definition of $m$ and the identity $P_{i}(\lambda) = Q_i(m(\lambda))/Q_i(m_0)$ for $i=t-1$ and $i=t-2$. Finally, matching the coefficients of this recurrence with \eqref{eq:recurence_orthogonal_polynomials} yields the identity in the Lemma. \end{proof} {\bfseries \underline{Step 2}: Asymptotics of orthonormal polynomials.} This step uses known result on asymptotics of orthonormal polynomials to compute the limit $t \to \infty$ of expressions derived in the previous step. We use the following theorem on the asymptotic ratio between two successive orthonormal polynomials. \begin{theorem}[\citep{rakhmanov1983asymptotics};\footnote{The original version of this theorem was stated for monic orthogonal polynomials but is valid for polynomials with other normalizations like orthonormal, see for instance \citep{lubinsky2000asymptotics, denisov2004rakhmanov}. } Ratio Asymptotics] \label{thm:rakhmanov1983asymptotics} Let $\{Q_i\}$ be a sequence of orthonormal polynomials with respect to a weight function strictly positive in $]-1, 1[$, and zero elsewhere. Then we have the following limit for the ratio of polynomials evaluated outside of the support, \begin{equation}\label{eq:rakhmanov} \lim_{t\rightarrow\infty} \frac{Q_t(\xi)}{Q_{t-1}(\xi)} = \xi + \sqrt{\xi^2-1} \quad \text{for} \;\; \xi > 1\,. \end{equation} \end{theorem} We can use this result to compute the limit of the ratio $Q_{t-1}(m_0)/Q_t(m_0)$, that appears in \eqref{eq:reformation_a_b}, as $m(0)> 1$ (and thus is not in the interval $[-1,\,1]$): \begin{align} \lim_{t \to \infty} \mfrac{Q_{t-1}(m_0)}{Q_t(m_0)} &\stackrel{\eqref{eq:rakhmanov}}{=} \Big(\mfrac{L + \ell}{L - \ell} + \sqrt{\big(\mfrac{L + \ell}{L - \ell}\big)^2-1}\big)\Big)^{-1} \nonumber\\ &= \mfrac{\sqrt{L}-\sqrt{\ell}}{\sqrt{L}+\sqrt{\ell}}\,. \label{eq:limit_ratio_q} \end{align} The other dependency of Eq. \eqref{eq:reformation_a_b} on the iteration $t$ is through the coefficients $\alpha_t, \beta_t$. To compute the limits of these we use the following known asymptotics:\footnote{It can be shown that the last two theorems are equivalent \citep[Theorem 13]{nevai1979orthogonal}. However, it will be more convenient for our purposes to present them as independent results.} \begin{theorem}[\citet{mate1985asymptotics}; Limits of recurrence coefficients] Under the same assumptions as Theorem \ref{thm:rakhmanov1983asymptotics}, the limits of the coefficients $\alpha_t, \beta_t$ in the orthonormal three-terms recurrence (Eq. \ref{eq:recurrence_orthonormal_poly}) is \begin{equation}\label{eq:mate} \lim_{t\rightarrow\infty}\alpha_t = \mfrac{1}{2}~,\qquad \lim_{t\rightarrow\infty}\beta_t = 0 ~. \end{equation} \end{theorem} Using this last theorem together with \eqref{eq:limit_ratio_q}, we have \begin{align*} \lim_{t \to \infty} {\color{colormomentum}(1 - a_t)} &~~\stackrel{\eqref{eq:reformation_a_b}}{=} -\left(\lim_{t \to \infty} \mfrac{\alpha_{t-1}}{ \alpha_t}\right) \left( \lim_{t \to \infty}\mfrac{Q_{t-2}(m_0)}{ Q_t(m_0)}\right) \nonumber\\ &\!\stackrel{(\ref{eq:limit_ratio_q}, \ref{eq:mate})}{=} -\Big(\mfrac{\sqrt{L}-\sqrt{\ell}}{\sqrt{L}+\sqrt{\ell}}\Big)^2, \end{align*} which is the claimed limit. To conclude the proof, we compute the same limit for the step-size ${\color{colorstepsize}b_t}$: \begin{align} \lim_{t \to \infty} {\color{colorstepsize}b_t} &\stackrel{\eqref{eq:reformation_a_b}}{=} - \mfrac{2}{ L - \ell} \left(\lim_{t \to \infty} \alpha_t^{-1}\right) \left(\lim_{t \to \infty} \mfrac{Q_{t-1}(m_0)}{Q_{t}(m_0)}\right) \\ &\!\!\!\!\stackrel{(\ref{eq:limit_ratio_q}, \ref{eq:mate})}{=} - \left(\mfrac{2}{\sqrt{L} + \sqrt{\ell}}\right)^2\,\,,~\widetilde{r}_1 = r_1, \widetilde{r}_0 = r_0. \end{align} \section{Asymptotic Expected Convergence Rates} \label{sec:proof} The previous section showed convergence of the method's parameters to PM, but said nothing about its rate of convergence. This section fills this gap by providing the asymptotic convergence of the expected convergence rate $\mathbb{E}\|{\boldsymbol x}_t-{\boldsymbol x}^\star\|^2$. More precisely, we show that the expected convergence rate converges to the rate of convergence of Polyak, and that this convergence rate is \textit{independent} of the probability distribution. \begin{restatable}{thm}{conv_hb_rate} \label{thm:conv_hb_rate} Under the same assumptions of Theorem \ref{thm:conv_hb}, the asymptotic expected rate of convergence of the optimal method converges to the worst-case rate of convergence, \begin{equation}\label{eq:asympt_rate} \limsup_{t\rightarrow \infty}\sqrt[t]{\mathbb{E}\left[\mfrac{\|{\boldsymbol x}_t-{\boldsymbol x}^\star\|^2}{\|{\boldsymbol x}_0-{\boldsymbol x}^\star\|^2}\right]} = \left(\mfrac{\sqrt{L}-\sqrt{\ell}}{\sqrt{L}+\sqrt{\ell}}\right)^2. \end{equation} \end{restatable} \begin{proof} Let $P_t$ be the residual orthogonal polynomial w.r.t. $\lambda \mathop{}\!\mathrm{d}\mu(\lambda)$. \citet[Theorem 3.1]{pedregosa2020acceleration} showed that the expected rate of convergence for average-case optimal methods admits the following simple form \begin{equation} \mathbb{E}\|{\boldsymbol x}_t-{\boldsymbol x}^\star\|^2 = R^{2}\int_{\mathbb{R}} P_t \mathop{}\!\mathrm{d}\mu~. \end{equation} This form is particularly convenient for us, as we can then use the the three-term recurrence to obtain a recurrence of this expression. Let $r_t = \int_{\mathbb{R}} P_t \mathop{}\!\mathrm{d}\mu$. After using the recurrence over $P_t$, \begin{align} r_t &= \int_{\mathbb{R}} (a_t+\lambda b_t)P_{t-1}(\lambda) + (1-a_t) P_{t-2}(\lambda) \mathop{}\!\mathrm{d}\mu(\lambda)\nonumber\\ &= a_t \underbrace{\int_{\mathbb{R}} P_{t-1}\mathop{}\!\mathrm{d}\mu}_{=r_{t-1}} + (1-a_t) \underbrace{\int_{\mathbb{R}}P_{t-2} \mathop{}\!\mathrm{d}\mu}_{=r_{t-2}}\,, \end{align} where in the last identity we have used the orthogonality between $P_t$ and $P_0(\lambda)=1$ w.r.t. $\lambda \mathop{}\!\mathrm{d}\mu(\lambda)$. In all, we have that the convergence rate $r_t$ is described by the recurrence \begin{equation} \begin{split} r_t &= a_t r_{t-1} + (1-a_t) r_{t-2}\,,\\ r_1 &= 1 + b_1 \int_{\mathbb{R}} \lambda \mathop{}\!\mathrm{d}\mu\,,~ r_0 = 1 \,. \end{split} \end{equation} A classical result, often referred to as the Poincar{\'e}-Perron theorem (see for example \citet[Theorem C]{pituk2002more} or \citep[Thm. 8.11]{elaydi2005introduction} ), states that if $a_t$ has a finite limit --guaranteed by the previous theorem and which we denote $a_{\infty}$-- then the recurrence has a fundamental set of solutions $\{r_t^1, r_t^2\}$ such that \begin{equation} \limsup_{t \to \infty} \sqrt[t]{r^i_t} = |\lambda_i|\quad \text{ $i =1, 2$}\,, \end{equation} where $\lambda_i$ are the roots of the characteristic equation ${\lambda^2 - a_{\infty} \lambda - (1 - a_{\infty})}$. In our case, these roots are $1$ and $1 - a_{\infty}$. Now, since the method we're considering is average-case optimal, this limit cannot be larger than that of Polyak momentum, known to be $(\tfrac{\sqrt{L}-\sqrt{\ell}}{\sqrt{L}+\sqrt{\ell}})^2 < 1$. Hence, we can eliminate the solution $r^1_t = 1$ and conclude \begin{equation} \limsup_{t \to \infty} \sqrt[t]{r_t} = (1 - a_{\infty}) = \left(\mfrac{\sqrt{L}-\sqrt{\ell}}{\sqrt{L}+\sqrt{\ell}}\right)^2\,. \end{equation} \end{proof} \section{Discussion and Simulations: Speed of Convergence to PM} The main result (Theorem~\ref{thm:conv_hb}) shows that, asymptotically, any average-case optimal method converge towards Polyak momentum. This could be interpreted as evidence against average-case optimal methods, as average-case optimal methods are not ``essentially different'' from PM. However, simulations show other dynamics at play. \begin{figure*} \centering \includegraphics[width=\linewidth]{figures/speed_convergence.pdf} \includegraphics[width=\linewidth]{figures/speed_convergence_uniform.pdf} \caption{{\bfseries Speed of convergence to Polyak Momentum}. For different parametrizations of the Marchenko-Pastur (\textit{top line}) and uniform (\textit{bottom line}) distributions, we plot the absolute difference between the average-case optimal momentum parameter (\textit{middle}) and average-case optimal step-size (\textit{right}) and the momentum and step-size of the Polyak method. The plots show a high anti-correlation between the speed of convergence of optimal average-case methods to PM and problem conditioning: for well-conditioned problems (small condition number) the parameters converge faster to PM than for ill-conditioned (large condition number) problems. Thus, in a regime were we perform only a few iterations, Polyak momentum may not be the best choice.} \label{fig:speed_convergence} \end{figure*} In Figure \ref{fig:speed_convergence} we plot the speed of convergence of the parameters of the optimal average-method for the Marchenko-Pastur distribution with different ratios $r = \frac{d}{n}$ (and hence condition number) and for the uniform distribution with different intervals. We see a clear effect of the condition number on the speed of convergence. The more ill-conditioned the problem, the slower the convergence of the optimal method to PM, implying that PM behaves sub-optimally for a larger number of iterations. This observation is consistent with the results of \citep{pedregosa2020acceleration}, who showed important speedups in the ill-conditioned regime. \section{Conclusion and Perspectives} In this work, we've shown that optimal average-case methods for minimizing quadratics converge to PM under mild assumptions on the expected spectral distribution. This universality over the probability measure is somewhat surprising, as Polyak momentum method only depends on the edges of the spectrum, while on the other hand optimal average-case methods depend on the whole spectrum. A potential area for future work is the analysis of the rate of convergence of optimal method to Polyak momentum algorithm. It seems the convergence of the step-size and momentum parameters are bounded polynomially in the number of iterations. This observation indicates the potential benefit of optimal methods over PM in the case where we perform a small number of iteration, typical in machine-learning problems. A second research direction is the study of optimal polynomials on the \textit{complex} plane. In this case, we are no longer solving the optimization problem \eqref{eq:quad_optim}. Instead, we aim to solve the linear system ${\boldsymbol A} {\boldsymbol x}={\boldsymbol b}$, where the matrix ${\boldsymbol A}$ is non-symmetric, with potentially complex eigenvalues. This has implication in the study of optimal algorithm in game theory \citep{azizian2020accelerating} or in the acceleration of primal-dual algorithms \citep{bollapragada2018nonlinear}. Finally, our results are only valid in the strongly convex regime ($\ell > 0$), ruling out the important case $r=1$ in the Marchenko-Pastur distribution, which corresponds to large least squares problems with a square matrix. After the first version of this paper appeared, \citet{paquette2020} derived an average-case analysis for gradient descent and showed a gap between the asymptotic average-case and worst-case convergence rate. The development of average-case optimal methods and the study of their asymptotic limits in this regime remains an open problem. \clearpage \section*{Acknowledgements} We would like to thank our colleague Gauthier Gidel for identifying and reporting some gaps in the proof of Theorem \ref{thm:conv_hb_rate}. A note of gratitude also goes to Reza Babanezad, Simon Lacoste-Julien, Remi Lepriol, Nicolas Loizou, Adam Ibrahim, Nicolas Leroux and Courtney Paquette for their insightful discussions and relevant remarks. We also thank Francis Bach and Raphaël Berthier for their useful remarks and pointers.

https://arxiv.org/abs/2005.08304

Understanding Nesterov's Acceleration via Proximal Point Method

The proximal point method (PPM) is a fundamental method in optimization that is often used as a building block for designing optimization algorithms. In this work, we use the PPM method to provide conceptually simple derivations along with convergence analyses of different versions of Nesterov's accelerated gradient method (AGM). The key observation is that AGM is a simple approximation of PPM, which results in an elementary derivation of the update equations and stepsizes of AGM. This view also leads to a transparent and conceptually simple analysis of AGM's convergence by using the analysis of PPM. The derivations also naturally extend to the strongly convex case. Ultimately, the results presented in this paper are of both didactic and conceptual value; they unify and explain existing variants of AGM while motivating other accelerated methods for practically relevant settings.

\section{Introduction} The \emph{proximal point method (PPM)}~\cite{moreau1965proximite,martinet1970regularisation,rockafellar1976monotone} is a fundamental method in optimization which solves the minimization of the cost function $f:\mathbb{R}^d\to \mathbb{R}$ by iteratively solving the subproblem \begin{align} \label{prox} x_{t+1} \leftarrow \argmin_{x\in \mathbb{R}^d} \left\{ f(x) + \frac{1}{2\eta_{t+1}} \norm{x-x_t}^2 \right\} \end{align} for a step size $\eta_{t+1}>0$, where the norm is chosen as the $\ell_2$ norm. The motivation of the method is clear: we add a quadratic regularization to make the cost function well conditioned for faster optimization\footnote{For instance, when $f$ is nonconvex, the regularization term can make each subproblem \eqref{prox} convex, and even when $f$ is convex, the regularization term will serve to increase the strong convexity parameter, which results in faster optimization. }. Nevertheless, solving \eqref{prox} is in general as difficult as solving the original optimization problem, and PPM is largely regarded as a ``conceptual'' guiding principle for accelerating optimization algorithms~\cite{drusvyatskiy2017proximal}. On the other hand, there is another prevalent accelerated method called \emph{Nesterov's accelerated gradient method} (AGM)~\cite{nesterov1983method}. In constrast to PPM, AGM is \emph{implementable} and has been applied to a myriad of applications, including sparse linear regression~\cite{beck2009fast}, compressed sensing~\cite{becker2011nesta}, the maximum flow problem~\cite{lee2013new}, and deep neural networks~\cite{sutskever2013importance}. Nonetheless, in contrast with the clear motivation of PPM, AGM has an obscure driving principle. In particular, original construction of AGM relies on an ingenious yet abstruse technique called \emph{estimate sequence}~\cite[Section 2.2.1]{nesterov2018lectures}, which has motivated researchers to investigate numerous alternative explanations (see Section~\ref{sec:related} for details). Recently, Defazio~\cite{defazio2019curved} established an inspiring connection between PPM and AGM. The main observation is that for strongly convex costs, one can derive a version of AGM from the primal-dual form of PPM with a tweak of geometry. This observation constitutes an important step toward understanding AGM because PPM is not as difficult to understand. This inspiring result, nevertheless, leaves open many other important questions. Most importantly, \cite{defazio2019curved} lacks \emph{quantitative} explanations as to why AGM achieves the accelerated convergence rates of $O(\nicefrac{1}{T^2})$ for smooth (Definition~\ref{def:ell2}) costs and $O(\exp(\nicefrac{-T}{\sqrt{\kappa}}))$ for smooth and strongly convex (Definition~\ref{def:str}) costs, where $T$ is the number of iterations and $\kappa$ is the condition number of the problem. Moreover, it is not clear whether such connection can be made without assuming strong convexity and can be extended to other more general and practical versions of AGM. In this work, we build a \emph{thorough} understanding of Nesterov's acceleration from the proximal point method by strengthening the connection made in \cite{defazio2019curved}. The main observation in this paper is that the mysterious updates of AGM can be fully understood by viewing it as a simple approximation of PPM. In particular, this observation leads to a straightforward derivation of AGM that does not rely on duality unlike \cite{defazio2019curved}. Moreover, the PPM view of AGM offers a simple analysis of AGM based on the standard analysis of PPM~\cite{guler1991convergence}. We also demonstrate the generality of our view. More specifically, our view naturally extends to the strongly convex case and obtains a general\footnote{It is general in the sense that it smoothly interpolates between the strongly convex case and the non-strongly convex case.} version of AGM~\cite[(2.2.19)]{nesterov2018lectures}, and our view also gives rise to the key idea of the \emph{method of similar triangles}, a version of AGM shown to have simple extensions to practically relevant settings~\cite{tseng2008accelerated,gasnikov2018universal,nesterov2018lectures}. \section{Baseline: analysis of the proximal point method}\label{sec:ppm} The baseline of our discussion is the following convergence rate of PPM for convex costs proved in a seminal paper by G{\"u}ler~\cite{guler1991convergence} (here $\xs$ denotes a global optimum point, i.e., $\xs \in \argmin_x f(x)$): \begin{align} \boxed{\textstyle f(x_{T})- f(\xs) \leq \OO{\big(\sum_{t=1}^{T} \eta_t\big)^{-1}}\quad \text{for any $T\geq 1$.}} \label{conv:prox} \end{align} In words, one can achieve an arbitrarily fast convergence rate by choosing step sizes $\eta_t$'s large. Below, we review a short Lyapuov function proof of \eqref{conv:prox}, which will serve as a backbone to other analyses. \begin{proof}[{\bf Proof of \eqref{conv:prox}}] It turns out that the following Lyapunov function is suitable: \begin{align} \boxed{\textstyle \Phi_t:= \big(\sum_{i=1}^{t} \eta_i \big)\cdot \big(f(x_t)-f(\xs)\big) + \frac{1}{2}\norm{\xs-x_t}^2,}\label{def:lya} \end{align} where $\Phi_0:= \frac{1}{2}\norm{\xs-x_0}^2$ and here and below, $\norm{\cdot}$ is the $\ell_2$ norm unless stated otherwise. Now, it suffices to show that $\Phi_t$ is decreasing, i.e., $\Phi_{t+1}\leq \Phi_t$ for all $t\geq 0$. Indeed, if $\Phi_t$ is decreasing, we have $\Phi_T\leq \Phi_0$ for any $T\geq 1$, which precisely recovers \eqref{conv:prox}. To that end, we use a standard result: \vspace{-5pt} \begin{propbox}[Proximal inequality~(see e.g. {\cite[Proposition 12.26]{bauschke2011convex}})] \label{prop:per} For a convex function $\phi:\mathbb{R}^d\to \mathbb{R}$, let $x_{t+1}$ be the unique minimizer of the following proximal step: \vspace{-5pt} \begin{align} \textstyle x_{t+1} \leftarrow \min_{x\in \mathbb{R}^d} \left\{\phi(x) +\frac{1}{2}\norm{x-x_t}^2\right\}\,. \label{prox:cond} \end{align} \vspace{-10pt} \noindent Then, for any $u\in \mathbb{R}^d$, $\phi(x_{t+1})-\phi(u) +\frac{1}{2}\norm{u-x_{t+1}}^2 +\frac{1}{2}\norm{ x_{t+1}-x_t}^2 -\frac{1}{2}\norm{u-x_t}^2\leq 0$. \end{propbox} Now Proposition~\ref{prop:per} completes the proof as follows: First, we apply Proposition~\ref{prop:per} with $\phi=\eta_{t+1} f$ and $u=\xs$ and drop the term $\frac{1}{2}\norm{x_{t+1}-x_t}^2$ to obtain: \begin{align} \tag{B1} \eta_{t+1}\left[f( x_{t+1}) -f( \xs)\right] + \frac{1}{2} \norm{\xs- x_{t+1}}^2 -\frac{1}{2} \norm{ \xs-x_t}^2\leq 0\,.\label{ineq:1} \end{align} Next, from the optimality of $x_{t+1}$ for \eqref{prox:cond}, it readily follows that \begin{align} \tag{B2} f(x_{t+1}) - f(x_t) \leq 0\,. \label{ineq:2} \end{align} Now, computing $\eqref{ineq:1}+( \sum_{i=1}^{t} \eta_i )\times $\eqref{ineq:2} yields $\Phi_{t+1}\leq \Phi_t$, which finishes the proof. \end{proof} \vspace{-10pt} \subsection{Our conceptual question} Although the convergence rate \eqref{conv:prox} seems powerful, it does not have any practical values as PPM is in general not implementable. Nevertheless, one can ask the following conceptual question: \begin{center} \emph{Can we develop an implementable approximation of PPM for large step sizes $\eta_t$'s?} \end{center} Perhaps, the most straightforward approximation would be to replace the cost function $f$ in \eqref{prox} with its lower-order approximations. We implement this idea in the next section. \section{Two simple approximations of the proximal point method} \label{sec:warmup} To analyze approximation errors, let us assume that the cost function $f$ is $L$-smooth. \begin{definition}[Smoothness] \label{def:ell2} For $L>0$, we say a differentiable function $f:\mathbb{R}^d\to \mathbb{R}$ is $L$-smooth if $f(x) \leq f(y) + \inp{\nabla f(y)}{x-y} +\frac{L}{2}\norm{x-y}^2 $ for any $x,y\in\mathbb{R}^d$. \end{definition} \noindent From the convexity and the $L$-smoothness of $f$, we have the following lower and upper bounds: \begin{align*} \textstyle f(y) +\inp{\nabla f(y)}{x-y} \leq f(x) \leq f(y) +\inp{\nabla f(y)}{x-y}+\frac{L}{2}\norm{x-y}^2\quad \text{for any $x,y\in \mathbb{R}^d$.} \end{align*} In this section, we use these bounds to approximate PPM. \subsection{First approach: using first-order approximation} \label{sec:app1} Let us first replace $f$ in the objective \eqref{prox} with its lower approximation: \begin{align} x_{t+1} \leftarrow \argmin_{x} \left\{ f(x_t)+\inp{\nabla f(x_t)}{x-x_t} + \frac{1}{2\eta_{t+1}} \norm{x-x_t}^2 \right\}\,. \label{gd} \end{align} Writing the optimality condition, one quickly notices that \eqref{gd} actually leads to gradient descent: \begin{align} x_{t+1} =x_t -\eta_{t+1} \nabla f(x_{t})\,.\label{gd:1} \end{align} Let us see how well \eqref{gd} approximates PPM: \vspace{-7pt} \begin{proof}[{\bf Analysis of the first approach}] We first establish counterparts of \eqref{ineq:1} and \eqref{ineq:2}. We begin with \eqref{ineq:1}. We first apply Proposition~\ref{prop:per} with $\phi(x)=\eta_{t+1}[f(x_t)+\inp{\nabla f(x_t)}{x-x_t}]$ and $u=\xs$: \begin{align*} &\phi(x_{t+1})-\phi(\xs)+ \frac{1}{2} \norm{\xs -x_{t+1}}^2 + \frac{1}{2} \norm{ x_{t+1}-x_t}^2 -\frac{1}{2 } \norm{\xs -x_{t}}^2\leq 0\,. \end{align*} Now using convexity and $L$-smoothness, we have $\phi(x)\leq \eta_{t+1} f(x)\leq \phi(x) +\frac{L\eta_{t+1}}{2}\norm{x-x_t}^2$, and hence the above inequality implies the following approximate version of \eqref{ineq:1}: \begin{align*} \eta_{t+1} \left[ f(x_{t+1})- f(\xs)\right] + \frac{1}{2} \norm{\xs -x_{t+1}}^2- \frac{1}{2 } \norm{\xs -x_{t}}^2 \leq \text{($\mathcal{E}_1$)}, \end{align*} where $\text{($\mathcal{E}_1$)}:= (\frac{L\eta_{t+1}}{2}-\frac{1 }{2} )\norm{x_{t+1}-x_t}^2$. Next, we use the $L$-smoothness of $f$ and the fact $\nabla f(x_t)= -\nicefrac{1}{\eta_{t+1}}(x_{t+1}-x_t)$ (due to \eqref{gd:1}), to obtain the following counterpart of \eqref{ineq:2}: \begin{align*} f(x_{t+1}) -f(x_t) \leq \inp{\nabla f(x_t)}{x_{t+1}-x_t} + \frac{L}{2} \norm{x_{t+1}-x_t}^2 = \text{($\mathcal{E}_2$)}, \end{align*} where $\text{($\mathcal{E}_2$)}:=( \frac{L}{2}- \frac{1}{\eta_{t+1}}) \norm{x_{t+1}-x_t}^2$. Now paralleling the proof of \eqref{conv:prox}, to show that $\Phi_t$~\eqref{def:lya} is a valid Lyapunov function, we need to find the step sizes $\eta_t$'s that satisfy the following relation: $\text{($\mathcal{E}_1$)} + ( \sum_{i=1}^t \eta_i) \times \text{($\mathcal{E}_2$)} \leq 0$. On the other hand, note that both \text{($\mathcal{E}_1$)}~and \text{($\mathcal{E}_2$)}~become positive numbers when $\eta_{t+1}>2/L$. Hence, the admissible choices for $\eta_t$ at each iteration are upper bounded by $2/L$, which together with the PPM convergence rate \eqref{conv:prox} implies that $O(\nicefrac{1}{\sum_{t=1}^T\eta_t}) =O(\nicefrac{1}{T})$ is the best convergence rate one can prove. Indeed, choosing $\eta_t\equiv 1/L$, then we have $\text{($\mathcal{E}_1$)}=0$ and $\text{($\mathcal{E}_2$)}<0$, obtaining the well-known bound of $f(x_T)-f(\xs) \leq \frac{L\norm{x_0-\xs}^2}{2T} =O(\nicefrac{1}{T})$. \end{proof}\vspace{-7pt} \noindent To summarize, the first approach only leads to a disappointing result: the approximation is valid only for the small step size regime of $\eta_t = \OO{1}$. We empirically verify this fact for a quadratic cost in Figure~\ref{fig:1}. As one can see from Figure~\ref{fig:1}, the lower approximation approach \eqref{gd} overshoots for large step sizes like $\eta_t = \Theta(t)$ and quickly steers away from PPM iterates.\vspace{-7pt} \begin{figure} \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=\linewidth]{p1.eps} \caption{$\eta_t\equiv 1/3$.} \end{subfigure} \begin{subfigure}{.5\textwidth} \centering \includegraphics[width=\linewidth]{p2.eps} \caption{$\eta_t=t/3$.} \end{subfigure} \caption{ Iterates comparison between PPM~\eqref{prox}, the first approach~\eqref{gd}, the second approach~\eqref{gd:2-1}, and the combined approach~\eqref{approx:ppm}. For the setting, we choose $f(x,y) = 0.1x^2+y^2$ and $x_0=(10,10)$. } \label{fig:1} \vspace{-7pt} \end{figure} \subsection{Second approach: using smoothness} \label{sec:app2} \vspace{-3pt} After seeing the disappointing outcome of the first approach, our second approach is to replace $f$ with its upper approximation due to the $L$-smoothness: \begin{align} x_{t+1} \leftarrow \argmin_{x} \left\{ f(x_t)+\inp{\nabla f(x_t)}{x-x_t} +\frac{L}{2} \norm{x-x_t}^2+ \frac{1}{2\eta_{t+1}} \norm{x-x_t}^2 \right\}\,. \label{gd:2-1} \end{align} Writing the optimality condition, \eqref{gd:2-1} actually leads to a conservative update of gradient descent: \begin{align} x_{t+1} = x_t -\frac{1}{L+\eta_{t+1}^{-1}} \nabla f(x_{t})\,. \label{gd:2-2} \end{align} Note that regardless of how large $\eta_{t+1}$ we choose, the actual update step size in \eqref{gd:2-2} is always upper bounded by $\nicefrac{1}{L}$. Although this conservative update prevents the overshooting phenomenon of the first approach, as we increase $\eta_{t}$, this conservative update becomes too tardy to be a good approximation of PPM; see Figure~\ref{fig:1}. \section{Nesterov's acceleration via alternating two approaches} \label{sec:deriv} In the previous section, we have seen that the two simple approximations of PPM both have limitations. Nonetheless, observe that their limitations are opposite to each other: while the first approach is too ``reckless,'' the second approach is too ``conservative.'' This observation motivates us to consider a \emph{combination} of the two approaches which could mitigate each other's limitation. Let us implement this idea by alternating between the two approximations \eqref{gd} and \eqref{gd:2-1} of PPM. The key modification is that for both approximations, we introduce an additional sequence of points $\{y_t\}$ for cost function approximation; i.e., we use the following approximations for the $t$-th iteration: \begin{align*} f(y_t) +\inp{\nabla f(y_t)}{x-y_t} \leq f(x) \leq f(y_t) +\inp{\nabla f(y_t)}{x-y_t}+\frac{L}{2}\norm{x-y_t}^2\,. \end{align*} Indeed, this modification is crucial: if we just use approximations at $x_t$, the resulting alternation merely concatenates \eqref{gd} and \eqref{gd:2-1} during each iteration, and the two limitations we discussed in Section~\ref{sec:warmup} will remain in the combined approach. Having introduced a separate sequence $\{y_t\}$ for cost approximations, we consider the following alternation where during each iteration, we update $x_{t}$ with \eqref{gd} and $y_{t}$ with \eqref{gd:2-1}: \begin{mdframed}[backgroundcolor=gray!20] {\bf Approximate PPM with alternating two approaches.} Given $x_0\in \mathbb{R}^d$, let $y_0=x_0$ and run: \begin{subequations} \label{approx:ppm} \begin{align} \label{ppm:a} &{\textstyle x_{t+1} \leftarrow \argmin_{x } \left\{ f(y_t)+\inp{\nabla f(y_t)}{x-y_t} + \frac{1}{2\eta_{t+1}} \norm{x-x_t}^2 \right\}}, \\ \label{ppm:b} &{\textstyle y_{t+1}\leftarrow \argmin_{x} \left\{ f(y_t)+\inp{\nabla f(y_t)}{x-y_t} +\frac{L}{2} \norm{x-y_t}^2+ \frac{1}{2\eta_{t+1}} \norm{x-x_{t+1}}^2 \right\} }. \end{align} \end{subequations} \end{mdframed} In Figure~\ref{fig:1}, we empirically verify that \eqref{approx:ppm} indeed gets the best of both worlds: this combined approach successfully approximates PPM even for the regime $\eta_t=\Theta(t)$. More remarkably, \eqref{approx:ppm} exactly recovers Nesterov's AGM. More specifically, turning \eqref{approx:ppm} into the equational form by writing the optimality conditions, and introducing an auxiliary iterate $z_{t+1}:= y_t -\nicefrac{1}{L}\nabla f(y_t)$, we obtain the following ($z_0:=x_0=y_0$): \begin{mdframed} \begin{minipage}{0.50\textwidth} {\bf An equivalent representation of \eqref{approx:ppm}:} \begin{subequations} \label{agm} \begin{align} & \textstyle y_{t} =\frac{\nicefrac{1}{L}}{ \nicefrac{1}{L}+\eta_t} x_t+\frac{ \eta_t }{ \nicefrac{1}{L}+\eta_t} z_t \,, \label{agm:a}\\ & \textstyle x_{t+1} = x_t -\eta_{t+1} \nabla f(y_{t})\,,\label{agm:b}\\ & \textstyle z_{t+1} = y_{t}-\frac{1}{L}\cdot \nabla f(y_{t})\,. \label{agm:c} \end{align} \end{subequations} \end{minipage} \begin{minipage}{0.50\textwidth} \vspace*{-12pt} \begin{figure}[H] \centering \begin{tikzpicture}[scale=0.35] \coordinate (a1) at (-2,0); \coordinate (a2) at (-2,-4.5*0.6); \coordinate (a3) at (-2,-4.5); \coordinate [label=left:{ $\boldsymbol{x_t}$}] (xt) at (0,0); \coordinate [label=right:{ $\boldsymbol{x_{t+1}}$}] (xt1) at (11,0); \coordinate [label=left:{ $\boldsymbol{z_{t}}$}] (zt) at (2,-4.5); \coordinate [label=left: {$\boldsymbol{y_{t}}$}] (yt1) at (2*0.6,-4.5*0.6); \coordinate [label=right: { $\boldsymbol{z_{t+1}}$}] (zt1) at (6.0,-4.5*0.6); \draw [->,>=stealth,dotted, line width=1pt,red] (xt) -- node[below] {\tiny $\textcolor{red}{\boldsymbol{-\eta_{t+1} \nabla f(y_{t})}}$} (xt1); \draw [->,>=stealth, dotted, line width=1pt,red] (yt1) -- node[below] {\tiny $\textcolor{red}{\boldsymbol{-\frac{1}{L}\nabla f(y_{t})}}$} (zt1); \draw [<->,dotted] (a1) -- node[left] {\small $ {\eta_t}$} (a2); \draw [<->,dotted] (a2) -- node[left] {\small $ {\nicefrac{1}{L}}$} (a3); \draw [dashed] (zt) -- (zt1); \draw [dashed] (zt1) -- (xt1); \draw [dashed] (xt) -- (zt); \end{tikzpicture} \vspace*{-4pt} \caption{The iterates of \eqref{agm}.} \label{fig:2} \end{figure} \end{minipage} \end{mdframed} Hence, we arrive at AGM without relying on any non-trivial derivations in the literature such as estimate sequence~\cite{nesterov2018lectures} or linear coupling~\cite{allen2014linear}. To summarize, we have demonstrated: \begin{center} \emph{Nesterov's accelerated method is a simple approximation of the proximal point method!} \end{center} \begin{remark} Our derivation is inspired by the one in the recent work by Defazio~\cite[Sections 5 and 6]{defazio2019curved}. However, unlike the approach in \cite{defazio2019curved}, our derivation does not rely on duality, which could be advantageous in the settings where duality fails. \end{remark} \begin{remark}[Understanding mysterious parameters of AGM] \label{rmk:par} It is often the case in the literature that the interpolation step \eqref{agm:a} is written as an abstract form $y_t = \tau_t x_t + (1-\tau_t) z_t$ with a weight parameter $\tau_t>0$ to be chosen~\cite{allen2014linear,lessard2016analysis,wilson2016lyapunov,bansal2019potential, ahn2020nesterov}. That said, in the previous works, $\tau_t$ is carefully chosen according to the analysis without conveying much intuition. One important aspect of our PPM view is that it reveals a close relation between the weight parameter $\tau_t$ and the step size $\eta_t$. More specifically, $\tau_t$ is chosen so that the ratio of the distances $\norm{y_t-x_t}:\norm{y_t-z_t}$ is equal to $\eta_t:\nicefrac{1}{L}$ (see Figure~\ref{fig:2}). \end{remark} \subsection{Understanding the accelerated convergence rate} \label{subsec:conv} In order to determine $\eta_t$'s in \eqref{agm}, we revisit the analysis of PPM from Section~\ref{sec:warmup}. In turns out that following Section~\ref{sec:app1}, one can derive from first principles the following inequalities using Proposition~\ref{prop:per} (we defer the derivations to Appendix~\ref{app:agm}): \begin{align*} \text{Counterpart of }\eqref{ineq:1}: &\quad\eta_{t+1}(f(z_{t+1})- f(\xs)) + \frac{1}{2} \norm{\xs -x_{t+1}}^2- \frac{1}{2} \norm{\xs -x_{t}}^2\leq \text{($\mathcal{F}_1$)}\quad \text{and}\\ \text{Counterpart of }\eqref{ineq:2}: &\quad f(z_{t+1}) -f(z_t) \leq \text{($\mathcal{F}_2$)}\,, \end{align*} where $\text{($\mathcal{F}_1$)}:=(\frac{\eta_{t+1}^2}{2}-\frac{\eta_{t+1}}{2L}) \norm{\nabla f(y_{t})}^2 + L\eta_t\eta_{t+1}\inp{\nabla f(y_{t})}{ z_{t}-y_t}$ and $\text{($\mathcal{F}_2$)}:= -\frac{1}{2L} \norm{\nabla f(y_{t})}^2$ $- \inp{\nabla f(y_{t})}{z_{t}-y_t}$. Hence, we modify the Lyapunov function \eqref{def:lya} by replacing the first $x_t$ with $z_t$: \begin{align} \textstyle \Phi_t:= (\sum_{i=1}^{t} \eta_i)\cdot (f(z_t)-f(\xs)) + \frac{1}{2}\norm{\xs-x_t}^2\,. \label{lya2} \end{align} Then as before, to prove the validity of the chosen Lyapunov function, it suffices to verify $\text{($\mathcal{F}_1$)} + ( \sum_{i=1}^t \eta_i)\cdot \text{($\mathcal{F}_2$)} \leq 0$, which is equivalent to \begin{align} \label{cond:1} \textstyle \frac{1}{2L}\left( L\eta_{t+1}^2 - \sum_{i=1}^{t+1} \eta_i\right) \norm{\nabla f(y_t)}^2 + \left( L\eta_t\eta_{t+1} - \sum_{i=1}^t \eta_i\right)\inp{\nabla f(y_{t})}{z_{t}-y_t} \leq 0 \end{align} From \eqref{cond:1}, it suffices to choose $\{\eta_t\}$ so that $L\eta_t\eta_{t+1}=\sum_{i=1}^t \eta_i$. Indeed, with such a choice, the coefficient of the inner product term in \eqref{cond:1} becomes zero and the coefficient of the squared norm term becomes $\nicefrac{1}{2L}(L\eta_{t+1}^2 - L\eta_{t+1}\eta_{t+2})\leq 0$ (if $\{\eta_t\}$ is increasing). Indeed, one can quickly notice that choosing $\eta_{t} =\nicefrac{t}{2L}$ satisfies the desired relation. Therefore, we obtain the well known accelerated convergence rate of $f(z_T)-f(\xs)\leq \frac{2L \norm{x_0-\xs}^2}{T(T+1)}= O(\nicefrac{1}{T^2})$~\cite{nesterov1983method}. \subsection{Separating step sizes for flexibility} \label{subsec:two} Since \eqref{approx:ppm} is an approximation of PPM, it is helpful to give \eqref{approx:ppm} more flexibility when we try to extend it to other settings. In particular, we relax \eqref{approx:ppm} by separating the two step sizes: \begin{mdframed} {\bf Approximate PPM with two separate step sizes $\{\eta_t\}$ and $\{\widetilde{\eta}_t\}$.} Given $x_0=y_0\in \mathbb{R}^d$, \begin{subequations}\label{approx:ppm2} \begin{align} \label{ppm2:a} &{\textstyle x_{t+1} \leftarrow \argmin_{x } \left\{ f(y_t)+\inp{\nabla f(y_t)}{x-y_t} + \frac{1}{2\eta_{t+1}} \norm{x-x_t}^2 \right\}}, \\ \label{ppm2:b} &{\textstyle y_{t+1}\leftarrow \argmin_{x} \left\{ f(y_t)+\inp{\nabla f(y_t)}{x-y_t} +\frac{L}{2} \norm{x-y_t}^2+ \frac{1}{2\widetilde{\eta}_{t+1}} \norm{x-x_{t+1}}^2 \right\} }. \end{align} \end{subequations} \end{mdframed} As we shall see in the next subsection, this simple relaxation allows us to recover a well known general version of AGM~\cite[Section 2.2]{nesterov2018lectures}. \subsection{Acceleration for strongly convex costs} \label{sec:str} Let us apply our PPM view to the strongly convex cost case. We begin with the definition: \begin{definition}[Strong convexity] \label{def:str} For $\mu>0$, we say a differentiable function $f:\mathbb{R}^d\to \mathbb{R}$ is $\mu$-strongly convex if $f(x) \geq f(y) + \inp{\nabla f(y)}{x-y} +\frac{\mu}{2}\norm{x-y}^2 $ for any $x,y\in\mathbb{R}^d$. \end{definition} \noindent Since $f$ is additionally assumed to be strongly convex, one can strengthen the step \eqref{ppm2:a} by \begin{align} \label{ppm:str} x_{t+1} \leftarrow \argmin_{x\in \mathbb{R}^d} \left\{ f(y_t)+\inp{\nabla f(y_t)}{x-y_t} + \frac{\mu}{2}\norm{x-y_t}^2 + \frac{1}{2\eta_{t+1}} \norm{x-x_t}^2 \right\}\,. \end{align} Writing the optimality condition of \eqref{ppm:str}, it is straightforward to check that \eqref{approx:ppm2} is equivalent to the following form (again, we introduce another auxiliary iterate $w_t$, and let $z_0:=x_0=y_0$): \begin{mdframed} \begin{minipage}{0.5\textwidth} {\bf Approximate PPM for strongly convex costs:} \vspace*{-10pt} \begin{subequations}\label{agm2} \begin{align} & \textstyle y_{t} =\frac{\nicefrac{1}{L}}{ \nicefrac{1}{L}+\widetilde{\eta}_t} x_t +\frac{ \widetilde{\eta}_t }{ \nicefrac{1}{L}+\widetilde{\eta}_t} z_t\,, \label{agm2:a}\\ & \textstyle w_t = \frac{\nicefrac{1}{\mu}}{ \nicefrac{1}{\mu}+\eta_{t+1}}x_t +\frac{\eta_{t+1}}{ \nicefrac{1}{\mu} +\eta_{t+1}}y_t\\ & \textstyle x_{t+1} = w_t- \frac{\nicefrac{1}{\mu}\cdot \eta_{t+1}}{ \nicefrac{1}{\mu}+\eta_{t+1}}\nabla f(y_{t})\,,\label{agm2:b}\\ & \textstyle z_{t+1} = y_{t}-\frac{1}{L}\nabla f(y_{t})\,. \label{agm2:c} \end{align} \end{subequations} \end{minipage} \begin{minipage}{0.5\textwidth} \vspace{-15pt} \begin{figure}[H] \centering \begin{tikzpicture}[scale=0.35] \coordinate (a1) at (-3,0); \coordinate (a2) at (-3,-7*0.7); \coordinate (a3) at (-3,-7); \coordinate (b1) at (-2.5,0); \coordinate (b2) at (-2.5,-7*0.3); \coordinate (b3) at (-2.5,-7*0.7); \coordinate [label=left:{ $\boldsymbol{x_t}$}] (xt) at (0,0); \coordinate [label=right:{$\boldsymbol{x_{t+1}}$}] (xt1) at (11,0); \coordinate [label=left:{ $\boldsymbol{w_t}$}] (wt) at (1*0.3,-7*0.3); \coordinate [label=left:{ $\boldsymbol{z_{t}}$}] (zt) at (1,-7); \coordinate [label=left: {$\boldsymbol{y_{t}}$}] (yt1) at (1*0.7,-7*0.7); \coordinate [label=below: {$\boldsymbol{z_{t+1}}$}] (zt1) at (1+10*0.3/0.7,-7+7*0.3/0.7); \draw [->,>=stealth,dotted, line width=1pt,red] (wt) -- node[above] {\tiny $\textcolor{red}{\boldsymbol{-\frac{\eta_{t+1}\cdot \nicefrac{1}{\mu}}{ \eta_{t+1}+\nicefrac{1}{\mu}} \nabla f(y_{t})}}$} (xt1); \draw [->,>=stealth, dotted, line width=1pt,red] (yt1) -- node[above] {\tiny $\textcolor{red}{\boldsymbol{-\frac{1}{L}\nabla f(y_{t})}}$} (zt1); \draw [<->,dotted] (a1) -- node[left] {\small $ {\widetilde{\eta}_t}$} (a2); \draw [<->,dotted] (a2) -- node[left] {\small $ {\nicefrac{1}{L}}$} (a3); \draw [dashed] (zt) -- (zt1); \draw [dashed] (zt1) -- (xt1); \draw [dashed] (xt) -- (zt); \draw [<->,dotted] (b1) -- node[right] {\scriptsize $ {\eta_{t+1}}$} (b2); \draw [<->,dotted] (b2) -- node[right] {\small $ {\nicefrac{1}{\mu}}$} (b3); \end{tikzpicture} \vspace*{-4pt} \caption{The iterates of \eqref{agm2}.} \label{fig:2-2} \end{figure} \end{minipage} \end{mdframed} Paralleling Remark~\ref{rmk:par}, our derivation provides new insights into the choices of the AGM step sizes by expressing them in terms of the PPM step sizes $\eta_t$'s and $\widetilde{\eta}_t$'s. Furthermore, our derivation actually demystifies the mysterious parameter choices made in the Nesterov's book~\cite[Section 2.2]{nesterov2018lectures}. To see this, let us recall the well known convergence rate of PPM for strongly convex costs due to Rockafellar~\cite[(1.14)]{rockafellar1976monotone}: \begin{align} \boxed{\textstyle f(x_{T})- f(\xs) \leq \OO{\prod_{t=1}^{T} (1+\mu\eta_t)^{-1}}\quad \text{for any $T\geq 1$.}} \label{conv:prox:str} \end{align} In light of \eqref{conv:prox:str}, it turns out that for the approximate PPM \eqref{agm2}, choosing the following step sizes \begin{center} $\eta_t \equiv \eta:= \mu^{-1}\cdot (\sqrt{\kappa}-1)^{-1}$ and $\widetilde{\eta}_t \equiv \widetilde{\eta}:= \mu^{-1}\cdot (\sqrt{\kappa})^{-1}$ (where $\kappa:=\nicefrac{L}{\mu}$) \end{center} actually recovers the well known parameters choice~\cite[(2.2.22)]{nesterov2018lectures} which leads to the convergence rate of $\OO{(1+\eta \mu)^{-T}}=\OO{(1+\nicefrac{1}{(\sqrt{\kappa}-1)})^{-T}}$. See \cite[Section 5.5]{bansal2019potential} for a simple Lyapunov function proof of this convergence rate. \begin{remark}[Nesterov's general method] Remarkably, \eqref{agm2} also exactly recovers a general version of AGM~\cite[(2.2.19)]{nesterov2018lectures} which smoothly interpolates between the strongly convex case and the non-strongly convex case. To see this, we follow a simple equivalent representation of Nesterov's general step sizes given in \cite{ahn2020nesterov}. More specifically, given a sequence $\{\xi_t\}$ of positive numbers defined as per the nonlinear recursion $\frac{\xi_{t+1}(\xi_{t+1}-\kappa^{-1})}{1-\xi_{t+1}}=\xi_t^2$ with an initial value $\xi_0>0$, choosing $\eta_t=\mu^{-1}\cdot (\kappa \xi_t-1)^{-1}$ and $\widetilde{\eta}_t = \mu^{-1}\cdot (\kappa \xi_{t+1}-1)^{-1}\cdot (1-\xi_{t+1})$ exactly recovers the choices in~\cite[Section 2]{ahn2020nesterov} which are shown to be equivalent to Nesterov's choices~\cite[(2.2.19)]{nesterov2018lectures}. \end{remark} \section{Simple generalizations with similar triangles} \label{sec:tri} In the previous section, we have demonstrated that Nesterov's method is nothing but an approximation of PPM. This view point has not only provided simple derivations of versions of AGM, but also offered clear explanations of the step sizes. In this section, we demonstrate that these interpretations offered by PPM actually lead to a great simplification of Nesterov's AGM in the form of the \emph{method of similar triangles}~\cite{nesterov2018lectures,gasnikov2018universal} which admits simple generalizations to practically relevant settings. Our starting point is the observations made in the previous section: (i) from Remark~\ref{rmk:par}, we have seen $\norm{y_t-x_t}:\norm{y_t-z_t}=\eta_t:\nicefrac{1}{L}$; (ii) from Section~\ref{subsec:conv}, we have seen that we need to choose $\eta_t=\Theta(t)$, and hence, $\eta_{t+1}\approx\eta_t \gg 1$. From these observations, one can readily see that the triangle $\triangle x_tx_{t+1}z_t$ is approximately similar to $\triangle y_tz_{t+1} z_t$. Therefore, one can simplify AGM by further exploiting this fact: we modify the updates so that the two triangles are indeed \emph{similar}: \begin{mdframed} \begin{minipage}{0.50\textwidth} {\bf Similar triangle approximation of PPM:} \begin{subequations}\label{sim} \begin{align} & \textstyle y_{t} = \frac{\nicefrac{1}{L}}{ \nicefrac{1}{L}+\eta_t} x_t+\frac{ \eta_t }{ \nicefrac{1}{L}+\eta_t} z_t\,, \label{sim:a}\\ & \textstyle x_{t+1} = x_t -\eta_{t+1} \nabla f(y_{t})\,,\label{sim:b}\\ & \textstyle z_{t+1} = \frac{\nicefrac{1}{L}}{ \nicefrac{1}{L}+\eta_t} x_{t+1}+\frac{ \eta_t }{ \nicefrac{1}{L}+\eta_t } z_t\,. \label{sim:c} \end{align} \end{subequations} \end{minipage} \begin{minipage}{0.50\textwidth} \vspace*{-12pt} \begin{figure}[H] \centering \begin{tikzpicture}[scale=0.35] \coordinate (a1) at (-2,0); \coordinate (a2) at (-2,-4.5*0.6); \coordinate (a3) at (-2,-4.5); \coordinate [label=left:{ $\boldsymbol{x_t}$}] (xt) at (0,0); \coordinate [label=right:{ $\boldsymbol{x_{t+1}}$}] (xt1) at (11,0); \coordinate [label=left:{ $\boldsymbol{z_{t}}$}] (zt) at (2,-4.5); \coordinate [label=left: {$\boldsymbol{y_{t}}$}] (yt) at (2*0.6,-4.5*0.6); \coordinate [label=right: { $\boldsymbol{z_{t+1}}$}] (zt1) at (2+9*0.4,-4.5*0.6); \draw [->,>=stealth, dotted, line width=0.8pt,red] (xt) -- node[below] {\tiny $\textcolor{red}{\boldsymbol{-\eta_{t+1} \nabla f(y_{t})}}$} (xt1); \draw [->,>=stealth,dotted, line width=0.8pt,red] (yt) -- (zt1); \draw [<->,dotted] (a1) -- node[left] {\small $ {\eta_t}$} (a2); \draw [<->,dotted] (a2) -- node[left] {\small $ {\nicefrac{1}{L}}$} (a3); \draw [dashed] (zt) -- (xt1); \draw [dashed] (xt) -- (zt); \draw pic[draw=black, fill=gray!10, -, angle eccentricity=1.2, angle radius=0.5cm, line width=1pt] {angle=xt--xt1--zt}; \draw pic[draw=black, fill=gray!10, -, angle eccentricity=1.2, angle radius=0.5cm, line width=1pt] {angle=yt--zt1--zt}; \draw pic[draw=black, fill=gray!10,-, densely dotted, angle eccentricity=1.2, angle radius=0.3cm, line width=1pt] {angle=zt--yt--zt1}; \draw pic[draw=black, fill=gray!10, -, densely dotted, angle eccentricity=1.2, angle radius=0.3cm, line width=1pt] {angle=zt--xt--xt1}; \end{tikzpicture} \vspace*{-5pt} \caption{The similar triangle updates \eqref{sim}.} \label{fig:sim} \end{figure} \end{minipage} \end{mdframed} We provide a PPM-based analysis of \eqref{sim} for a more general setting in Section~\ref{sec:gen}. \begin{remark} The updates akin to \eqref{sim} can be found in~\cite[Algorithm 1]{tseng2008accelerated} (note that the step sizes are slightly different). Our derivation based on the PPM view indeed clarifies why such similar triangles are natural updates to consider. We also remark that the updates based on similar triangles is useful in developing universal methods for stochastic composite optimizations~\cite{gasnikov2018universal}. \end{remark} \begin{remark}[Momentum methods] \label{rmk:momentum} An alternative way to approximate PPM with similar triangles is: \begin{mdframed} \begin{minipage}{0.50\textwidth} {\bf Alternative similar triangle approximation:} \begin{subequations}\label{sim2} \begin{align} & \textstyle y_{t} = \frac{\nicefrac{1}{L}}{ \nicefrac{1}{L}+\eta_t} x_t+\frac{ \eta_t }{ \nicefrac{1}{L}+\eta_t} z_t\,, \label{sim2:a}\\ & \textstyle z_{t+1} = y_t -\frac{1}{L} \nabla f(y_{t})\,,\label{sim2:b}\\ & \textstyle x_{t+1} =z_{t+1} + L\eta_t (z_{t+1}-z_t)\,. \label{sim2:c} \end{align} \end{subequations} \end{minipage} \begin{minipage}{0.50\textwidth} \vspace*{-12pt} \begin{figure}[H] \centering \begin{tikzpicture}[scale=0.35] \coordinate (a1) at (-2,0); \coordinate (a2) at (-2,-4.5*0.6); \coordinate (a3) at (-2,-4.5); \coordinate [label=left:{ $\boldsymbol{x_t}$}] (xt) at (0,0); \coordinate [label=right:{ $\boldsymbol{x_{t+1}}$}] (xt1) at (11,0); \filldraw (2,-4.5) circle (3pt) node[] {}; \coordinate [label=left:{ $\boldsymbol{z_{t}}$}] (zt) at (2,-4.5); \filldraw (2*0.6,-4.5*0.6) circle (3pt) node[] {}; \coordinate [label=left: {$\boldsymbol{y_{t}}$}] (yt) at (2*0.6,-4.5*0.6); \filldraw (2+9*0.4,-4.5*0.6) circle (3pt) node[] {}; \coordinate [label=below: { $\boldsymbol{z_{t+1}}$}] (zt1) at (2+9*0.4,-4.5*0.6); \filldraw (2+9*0.6,-4.5*0.4) circle (3pt) node[] {}; \coordinate [label=above: { $\boldsymbol{y_{t+1}}$}] (yt1) at (2+9*0.6,-4.5*0.4); \draw [->,>=stealth, dotted, line width=0.8pt,red] (xt) -- (xt1); \draw [->,>=stealth,dotted, line width=0.8pt,red] (yt) -- node[above] {\tiny $\textcolor{red}{\boldsymbol{-\frac{1}{L} \nabla f(y_{t})}}$} (zt1); \draw [<->,dotted] (a1) -- node[left] {\small $ {\eta_t}$} (a2); \draw [<->,dotted] (a2) -- node[left] {\small $ {\nicefrac{1}{L}}$} (a3); \draw [dashed] (zt) -- (xt1); \draw [dashed] (xt) -- (zt); \draw pic[draw=black, fill=gray!10, -, angle eccentricity=1.2, angle radius=0.5cm, line width=1pt] {angle=xt--xt1--zt}; \draw pic[draw=black, fill=gray!10, -, angle eccentricity=1.2, angle radius=0.5cm, line width=1pt] {angle=yt--zt1--zt}; \draw pic[draw=black, fill=gray!10,-, densely dotted, angle eccentricity=1.2, angle radius=0.3cm, line width=1pt] {angle=zt--yt--zt1}; \draw pic[draw=black, fill=gray!10, -, densely dotted, angle eccentricity=1.2, angle radius=0.3cm, line width=1pt] {angle=zt--xt--xt1}; \end{tikzpicture} \vspace*{-5pt} \caption{The updates of \eqref{sim2}.} \label{fig:sim2} \end{figure} \end{minipage} \end{mdframed} In fact, \eqref{sim2} can be equivalently expressed without $\{x_t\}$: at the $t$-iteration, we compute \eqref{sim2:b} and then update $y_{t+1}$ via $y_{t+1}= z_{t+1} + \frac{L\eta_t}{L\eta_{t+1}+1}(z_{t+1}-z_t)$ (these updates are illustrated with dots in Figure~\ref{fig:sim2}). Hence, \eqref{sim2} is equal to the well-known momentum based AGM~\cite{nesterov1983method,beck2009fast}. Notably, it turns out that our PPM-based analysis suggests the choice of $\{\eta_t\}$ as per the recursive relation $(L\eta_{t+1}+\frac{1}{2})^2= (L\eta_t+1)^2 +\frac{1}{4}$, which after substitution $L\eta_t+1\leftarrow a_t$ exactly recovers the popular recursive relation $a_{t+1} = \frac{1}{2} (1+\sqrt{1+4a_t^2})$ in \cite{nesterov1983method,beck2009fast}. See Appendix~\ref{app:rmk} for precise details. \end{remark} The main advantage of this similar triangles approximation \eqref{sim} becomes clearer in the constraint optimization case: when there is a constraint set, the steps \eqref{agm:b} and \eqref{agm:c} both become projections steps which could be costly when the constraint set does not admit simple projections. On the other hand, since \eqref{sim} only requires a single projection in each iteration, it minimizes such costly computations. \subsection{Extension to composite costs and general norms} \label{sec:gen} It turns out \eqref{sim} admits a simple extension to the practically relevant setting of composite costs and general norms (see e.g. \cite[Section 6.1.3]{nesterov2018lectures}). More specifically, for a closed convex set $Q\subseteq \mathbb{R}^d$ and a closed\footnote{This means that the epigraph of the function is closed. See \cite[Definition 3.1.2]{nesterov2018lectures}. } convex function $\Psi:Q\to \mathbb{R}$, consider \begin{align*} \textstyle \min_{x\in Q} f^{\Psi}(x):= f(x)+\Psi(x) \,, \end{align*} where $f:Q\to \mathbb{R}$ is a differentiable convex function which is $L$-smooth with respect to a norm $\norm{\cdot}$ that is not necessarily the $\ell_2$ norm (i.e., we regard the norm in Definition~\ref{def:ell2} to be our chosen norm). \noindent For the general norm case, we use the Bregman divergence for the regularizer\footnote{The reason why we need the Bregman divergence in place of the norm squared regularization is because for norms other than the $\ell_2$ norm, $\frac{1}{2}\norm{u-v}^2$ is not strongly convex with respect to the chosen norm. }: \begin{definition} Given a $1$-strongly convex (w.r.t the chosen norm $\norm{\cdot}$) function $h:Q\to \mathbb{R} \cup \{\infty\}$ that is differentiable on the interior of $Q$, $\breg{h}{u}{v}:=h(u)-h(v)-\inp{\nabla h(v)}{u-v}$ for all $u,v\in Q$. \end{definition} \noindent Under the above setting and assumption, \eqref{sim} admits a simple generalization: \begin{mdframed} {\bf Similar triangle approximations of PPM for composite costs and general norms:} \begin{subequations}\label{simg} \begin{align} &\textstyle y_{t} =\frac{\nicefrac{1}{L}}{\nicefrac{1}{L}+\eta_t } x_t+\frac{ \eta_t }{ \nicefrac{1}{L}+\eta_t } z_t\,, \label{simg:a}\\ &\textstyle x_{t+1} \leftarrow \argmin_{x\in Q} \left\{ f(y_t)+\inp{\nabla f(y_t)}{x-y_t} + \frac{1}{\eta_{t+1}} \breg{h}{x}{x_t} +\Psi(x) \right\}\,,\label{simg:b}\\ & \textstyle z_{t+1} =\frac{\nicefrac{1}{L}}{\nicefrac{1}{L}+\eta_t} x_{t+1}+\frac{ \eta_t }{\nicefrac{1}{L}+ \eta_t } z_t \,. \label{simg:c} \end{align} \end{subequations} \end{mdframed} Again, the similar triangle approximation \eqref{simg} is computationally advantageous in that it only requires a single projection in each iteration. Now we provide a simple PPM-based analysis of \eqref{simg}: \begin{proof}[{\bf PPM-based analysis of \eqref{simg}}] To obtain counterparts of \eqref{ineq:1} and \eqref{ineq:2}, we now use a generalization of Proposition \ref{prop:per} to the Bregman divergence (\cite[Lemma 3.1]{teboulle2018simplified}). With such a generalization, we obtain the following inequality for $\phi^{\Psi}(x):=\eta_{t+1}[f(y_t)+\inp{\nabla f(y_t)}{x-y_t} +\Psi(x)]$: \begin{align} \phi^{\Psi}(x_{t+1}) -\phi^{\Psi}(\xs) + \breg{h}{\xs}{x_{t+1}}+\breg{h}{x_{t+1}}{x_{t}}- \breg{h}{\xs}{x_{t}} \leq 0\,, \label{gen:1} \end{align} where $\xs\in \argmin_{x\in Q}f^{\Psi}(x)$. Now using \eqref{gen:1}, one can derive from first principles the following inequalities (we defer the derivations to Appendix~\ref{app:sim}): \begin{align*} \text{Counterpart of }\eqref{ineq:1}: &\quad \eta_{t+1}(f^{\Psi}(z_{t+1})- f^{\Psi}(\xs) ) + \breg{h}{\xs}{x_{t+1}}- \breg{h}{\xs}{x_{t}} \leq \text{($\mathcal{G}_1$)} \quad \text{and}\\ \text{Counterpart of }\eqref{ineq:2}: &\quad f^{\Psi}(z_{t+1}) -f^{\Psi}(z_t) \leq \text{($\mathcal{G}_2$)}\,, \end{align*} where $\text{($\mathcal{G}_1$)}:= \textstyle-\frac{1}{2}\norm{ x_{t+1}-x_t}^2 +\eta_{t+1}[\frac{L}{2}\norm{z_{t+1}-y_{t}}^2 +\inp{\nabla f(y_{t})}{z_{t+1}-x_{t+1}} +\Psi(z_{t+1}) -\Psi(x_{t+1})]$ and $\text{($\mathcal{G}_2$)}:= \frac{L}{2} \norm{z_{t+1}-y_t}^2 + \inp{\nabla f(y_{t})}{z_{t+1}-z_t} +\Psi(z_{t+1})-\Psi(z_{t})$. Similarly to Section~\ref{subsec:conv}, yet replacing the norm squared term with the Bregman divergence, we choose \begin{align*} \textstyle\Phi_t:= (\sum_{i=1}^{t} \eta_i)\cdot (f^{\Psi}(z_t)-f^{\Psi}(\xs)) +\breg{h}{\xs}{x_t}. \end{align*} Then, it suffices to show $\text{($\mathcal{G}_1$)} + ( \sum_{i=1}^t \eta_i)\cdot \text{($\mathcal{G}_2$)} \leq 0$. Using the facts (i) $z_{t+1}-x_{t+1} = L\eta_t (z_t-z_{t+1})$ and (ii) $\norm{x_{t+1}-x_t}= (L\eta_t+1)\norm{z_{t+1}-y_t}$ (both are immediate consequences of the similar triangles) and rearranging, one can easily check that $\text{($\mathcal{G}_1$)} + ( \sum_{i=1}^t \eta_i)\cdot \text{($\mathcal{G}_2$)}$ is equal to \begin{align} \textstyle \frac{1}{2}\left(-(L\eta_t+1)^2 +L \eta_{t+1}+ L\sum_{i=1}^{t}\eta_{i} \right)\norm{z_{t+1}-y_{t}}^2 \label{a1}\\ \textstyle +\left(L\eta_t \eta_{t+1} - \sum_{i=1}^t\eta_i\right)\inp{\nabla f(y_{t})}{z_{t}-z_{t+1}} \label{a2}\\ \textstyle +\eta_{t+1}[ \Psi(z_{t+1}) -\Psi(x_{t+1}) ] + \left( \sum_{i=1}^t \eta_i\right)\cdot [\Psi(z_{t+1})-\Psi(z_{t})]. \label{a3} \end{align} Now choosing $\eta_t= \nicefrac{t}{2L}$ analogously to Section~\ref{subsec:conv}, one can easily verify $\eqref{a1}+\eqref{a2}+\eqref{a3} \leq 0$. Indeed, for \eqref{a1}, since $L \eta_t \eta_{t+1}=\sum_{i=1}^t\eta_i$, the coefficient becomes $\nicefrac{1}{2}(L\eta_t+1)(L\eta_{t+1}-L\eta_t-1)$ which is a negative number since $L\eta_{t+1}-L\eta_t-1=-\nicefrac{1}{2}$; for \eqref{a2}, the coefficient becomes zero due to the relation $L \eta_t \eta_{t+1}=\sum_{i=1}^t\eta_i$; lastly, for \eqref{a3}, we have \begin{align} \eqref{a3}= \eta_{t+1}\left[ (1+L\eta_t) \Psi(z_{t+1}) -\Psi(x_{t+1}) -L\eta_{t}\Psi(z_{t})\right] \leq 0\,, \end{align} where the equality is due to the relation $L \eta_t \eta_{t+1} =\sum_{i=1}^t\eta_i$, and the inequality is due to the update \eqref{simg:c} (which can be equivalently written as $(1+L\eta_t) z_{t+1} = x_{t+1}+ L\eta_t z_t$) and the convexity of $\Psi$. Hence, we obtain the accelerated rate of $f^{\Psi}(z_T)-f^{\Psi}(\xs)\leq \frac{4L\breg{h}{\xs}{x_0}}{T(T+1)} = O(\nicefrac{1}{T^2})$. \end{proof} \section{Related work} \label{sec:related} Motivated by the obscure scope of Nesterov's estimate sequence technique, there have been a flurry of works on developing alternative approaches to Nesterov's acceleration. The most contributions are made based on understanding the continuous limit dynamics of Nesterov's AGM~\cite{su2014differential,krichene2015accelerated,wibisono2016variational}. These continuous dynamics approaches have brought about new intuitions about Nesterov's acceleration, and follow-up works have developed analytical techniques for such dynamics~\cite{wilson2016lyapunov,diakonikolas2019approximate}. However, these approaches share a limitation that when applying discretization techniques to obtain optimization algorithms, some auxiliary modifications are required to recover/obtain analyzable algorithms. In contrast, our PPM approach directly yields accelerated methods and does not require additional adjustments. Another notable contribution is made based on the linear coupling framework~\cite{allen2014linear}. The main observation is that the two most popular first-order methods, namely gradient descent and mirror descent, have complementary performances, and hence, one can come up with a faster method by linearly coupling the two methods. This view indeed offers a general framework of developing fast optimization algorithms; however, for understanding Nesterov's acceleration, this view has less expressive power compared to our PPM view. More specifically, it is \emph{a priori} not clear why one needs to couple two methods \emph{linearly}. Moreover, with the linear coupling view, one cannot interpret the interpolation weight as we did in Remark~\ref{rmk:par}. It is also important to note that PPM has been given new attention as a building block for designing and analyzing fast optimization methods~\cite{drusvyatskiy2017proximal}. To list few instances, PPM has given rise to methods for weakly convex problems~\cite{davis2019proximally}, the prox-linear methods for composite optimizations~\cite{burke1995gauss,nesterov2007modified,lewis2016proximal}, accelerated methods for stochastic optimizations~\cite{lin2015universal}, and methods for saddle-point problems~\cite{mokhtari2019unified}. Moreover, using Proposition~\ref{prop:per} (and its generalization to the Bregman divergence) as a unified tool for analyzing first-order methods has been discussed in~\cite{teboulle2018simplified}. \section{Conclusion} This work provides a complete understanding of Nesterov's acceleration by making analytical and quantitative connections to the proximal point method. The key observation is that an alternation of two simple approximations of the PPM exatly recovers Nesterov's AGM. Through this connection, we are able to explain all the step sizes of AGM in terms of the PPM step sizes, demystifying the mysterious choices made in the literature. This view naturally extends to the strongly convex case and recovers Nesterov's general accelerated method from his book. Moreover, our PPM view motivates a simplification of AGM using similar triangles, which admits a simple PPM-based analysis as well as a simple extension to the general norm and composite optimization case. For future directions, it would be interesting to connect our PPM view to accelerated stochastic methods ~\cite{lin2015universal,lan2018optimal} and other accelerated methods, including geometric descent~\cite{bubeck2015geometric}. \section*{Acknowledgement} The author thanks Suvrit Sra, Alp Yurtsever and Jingzhao Zhang for detailed comments and stimulating discussions, Aaron Defazio for clarifications that help the author develop Section~\ref{sec:gen}, and Heinz Bauschke for constructive suggestions on the presentation of Section~\ref{sec:deriv} as well as Remark~\ref{rmk:momentum}. \bibliographystyle{alpha}

https://arxiv.org/abs/2202.03821

A Lower Bound on the Failed Zero Forcing Number of a Graph

Given a graph $G=(V,E)$ and a set of vertices marked as filled, we consider a color-change rule known as zero forcing. A set $S$ is a zero forcing set if filling $S$ and applying all possible instances of the color change rule causes all vertices in $V$ to be filled. A failed zero forcing set is a set of vertices that is not a zero forcing set. Given a graph $G$, the failed zero forcing number $F(G)$ is the maximum size of a failed zero forcing set. An open question was whether given any $k$ there is a an $\ell$ such that all graphs with at least $\ell$ vertices must satisfy $F(G)\geq k$. We answer this question affirmatively by proving that for a graph $G$ with $n$ vertices, $F(G)\geq \lfloor\frac{n-1}{2}\rfloor$.

\section{Introduction} Given a graph $G=(V,E)$ and a subset $S\subseteq V$ of filled vertices, we consider the following the color change rule. If any filled vertex $v$ is adjacent to exactly one unfilled vertex $w$, then $v$ ``forces'' $w$, and we mark $w$ as filled. The set of filled vertices after all possible instances of the rule are applied is called the derived set of $S$. If the derived set of $S$ is all of $V$, we say that $S$ is a zero forcing set for $G$. Otherwise it is a failed zero forcing set. The zero forcing number of a graph $G$, denoted $Z(G)$ is the minimum size of any zero forcing set of $G$. The failed zero forcing number $G$, denoted $F(G)$ is the maximum size of any failed zero forcing set of $G$. Zero forcing numbers and their relation to minimum rank problems have been studied extensively, for example in \cite{barioli}. Failed zero forcing numbers were introduced and studied in \cite{fetcie}. Any path $P_n$ satisfies $Z(P_n)=1$. This can be seen by taking the initial set of filled vertices to consist of exactly one end vertex of the path. The end vertex forces its neighbor, which in turn forces its unfilled neighbor, and so on until the entire path is filled. Hence there are graphs with arbitrarily large vertex set whose zero forcing number is 1. In \cite{fetcie}, the authors asked whether a similar property holds for the failed zero forcing number. Intuition tells us that for any fixed integer $k$ there should not be arbitrarily large graphs satisfying $F(G)=k$. The larger the graph, the more room there is to distribute a fixed number of filled vertices sparsely in such a way that no filled vertex can force. We confirm this intuition by proving that for any graph $G$ on $n$ vertices $F(G)\geq\lfloor \frac{n-1}{2}\rfloor$. This is the best bound possible, as it was shown in \cite{fetcie} that any path $P_n$ satisfies $F(P_n)=\lfloor\frac{n-1}{2}\rfloor$. Strikingly, for any given number of vertices, the minimum possible zero forcing number and maximum possible failed zero forcing number are both obtained by paths. The bound makes it possible in principle to classify all graphs with a given failed zero forcing number $k$ by exhaustively checking all graphs such that $n\leq 2k + 2$. This classification was done in \cite{fetcie} for $k=0,1$, and in \cite{gomez} for $k=2$, the latter using a special case of the bound proven here. In what follows we prove our main result by showing a somewhat stronger result for graphs with minimum vertex degree of at least three. We then show the desired bound inductively for graphs with minimum degree one and two. \section{Graphs with minimum degree at least three} The main result of this section is the following lower bound on the failed zero forcing number of a graph of minimum degree three or more: \begin{theorem}\label{delta3} Let $G=(V,E)$ be a connected graph on $n$ vertices satisfying $\delta(G)\geq 3$. Then $F(G) \geq \lfloor \frac{n}{2}\rfloor$. If $G$ contains a circuit of even length (and $n$ is odd), we can improve the bound to $F(G) \geq \lceil\frac{n}{2}\rceil$. \end{theorem} \tikzset{ unfilled/.style = {circle, draw = black, minimum size = 0.5cm}, filled/.style = {unfilled, fill = blue!40, minimum size = 0.5cm}, collection/.style = {ellipse, draw = black, thick} } We develop machinery needed to prove this result, starting with some terminology. If a set of vertices $S\neq V$ is its own derived set, then we say that $S$ is stalled. Note that any maximal failed zero forcing set is stalled. We will distinguish between vertices that cannot force because all of their neighbors are filled and vertices that cannot force because they have multiple unfilled neighbors. \begin{defn} Let $G=(V,E)$, $S\subseteq V$ be a set of filled vertices, and $v\in S$. We say $v$ is \emph{spent} if all neighbors of $v$ are in $S$. Otherwise, $v$ is \emph{unspent}. \end{defn} \noindent\textbf{Observation:} If we have a maximal failed zero forcing set $S$ in $G$ in which all vertices are unspent, then all filled vertices are adjacent to at least two unfilled vertices. In this case, we can add any number of edges to $G$, and $S$ will still be a failed zero forcing set. This is a special case of lemma 3 in \cite{gomez}. We next have a result about bipartite graphs. \begin{lemma}\label{bipartite} If $G=(V,E)$ is a bipartite graph with bipartition $V=L\sqcup R$, and $\delta(G)\geq 2$, then $F(G)\geq\max\{|L|,|R|\}$. In particular, $F(G)\geq\lceil\frac{n}{2}\rceil$. \end{lemma} \begin{proof} If we fill all the vertices in, say, $L$, then every filled vertex is adjacent to at least two (unfilled) vertices in $R$, and therefore cannot force. \end{proof} Note that all vertices in the failed zero forcing set obtained in lemma (\ref{bipartite}) are unspent. Combining the lemma with the observation we have: \begin{corollary} If $G$ has a bipartite spanning subgraph $H$ such that $\delta(H)\geq 2$, then $F(G)\geq\lceil\frac{n}{2}\rceil$ \end{corollary} A vertex cut in a connected graph $G$ is a set of vertices whose removal causes $G$ to be disconnected or trivial. The connectivity of a connected graph $G$, denoted $\kappa(G)$, is the size of the smallest vertex cut. We need the following: \begin{lemma}\label{cut_vert} Let $G=(V,E)$ be a connected graph on $n$ vertices satisfying $\delta(G)\geq 3$ and $\kappa(G)=1$. Then $F(G)\geq \lfloor\frac{n+1}{2}\rfloor$ \end{lemma} \begin{figure}[H] \caption{The graph $G$ with the vertices in $S \cup \{v\}$ colored.} \label{fig:cut_vert} \begin{tikzpicture} \draw[fill = blue!40] (0, 0) ellipse(2.5cm and 1cm); \draw (0, -3) ellipse(0.75cm and 0.75cm); \node[filled, label=right:{$v$}] (v) at (0, -1.75){}; \node (d2) at (-1.5, 0){}; \node (d3) at (1.5, 0){}; \node (d1) at (0, -3){}; \node[unfilled, label=right:{$w$}] (w) at (0, -2.75){}; \node (w1) at (-.25, -3.5){}; \node (w2) at (.25, -3.5){}; \draw (v) -- (w); \draw (v) -- (d2); \draw (v) -- (d3); \draw (w) -- (w1); \draw (w) -- (w2); \end{tikzpicture} \centering \end{figure} \begin{proof} Since $\kappa(G) = 1$ there is a vertex cut consisting of a single vertex $v$. Let $G'$ be the graph induced by $V-\{v\}$. Since $G'$ is disconnected, we can label $k$ connected components of $G'$, $D_1, D_2,...,D_k$ with corresponding vertex sets $Y_1, Y_2,...,Y_k$, in order of increasing size. Let $S = \bigcup_{i = 2}^k Y_i$. Note that $|S| + 1 \ge \lfloor \frac{n+1}{2} \rfloor$. The only vertex in $V - D_1$ adjacent to a vertex in $D_1$ is $v$. No vertex in $S$ can force a vertex in $D_1$, so we need only consider the color changing rule applied to $v$. If $v$ is adjacent to at least two vertices in $D_1$, then $S \cup \{v\}$, is stalled. If $v$ is adjacent to exactly one vertex in $D_1$, call it $w$ (see Figure \ref{fig:cut_vert}). Since $\delta(G) \ge 3$, $w$ has at least two neighbors in $D_1$. The derived set of $S \cup \{v\}$ would then include $w$, but not $w$'s neighbors in $D_1$ since there would be at least two of them. In either case, the derived set of $S \cup v$ excludes vertices in $D_1$, making it a failed zero forcing set. Thus, $F(G) \ge |S \cup \{v\}| \ge |S| + 1 \ge \lfloor \frac{n+1}{2} \rfloor$. \end{proof} Now we need to handle the case where $G$ has no cut vertices. We give an algorithm for finding a failed zero forcing set with at least $\lfloor \frac{n-1}{2}\rfloor$ vertices in any graph $G$ satisfying $\delta(G)\geq 3$ and $\kappa(G)\geq 2$. The algorithm for produces a stalled zero forcing set of the required size, all of whose vertices are unspent. \begin{algo}\label{algo} Our strategy will be to partition $V$ into three subsets $L$, $R$ and $O$ such that any vertex in $L$ is adjacent to at least two vertices in $R\cup O$, and any vertex in $R$ is adjacent to at least two vertices in $L\cup O$. We may then obtain a stalled set by filling either all the vertices in $L$, or all of the vertices in $R$. We will have that $|O|= 0$ or $|O|=1$, so that we are guaranteed to be able to fill at least $\lfloor \frac{n-1}{2}\rfloor$ vertices by choosing the larger of $L$ and $R$. In what follows, we refer to $L$ and $R$ as the ``sides'' of the partition. At any stage of the algorithm we let $A=L\cup R\cup O$ be the set of vertices that have already been assigned to one set in the partition, and we use $U$ the set of all vertices that remain unassigned. Since $\delta(G)\geq 3$, $G$ has at least one cycle. If $G$ has a cycle $C $ of even length, then assign alternating vertices of $C$ to $L$ and $R$ and let $O=\emptyset$. This guarantees that all assigned vertices are adjacent to two vertices on the other side of the partition. If $G$ has no even cycles, then choose any odd cycle $C$ and let $O=\{v_0\}$ for an arbitrary chosen $v_0\in C$. Starting from either neighbor of $v_0$ in $C$ assign alternating vertices to $L$ and $R$. The neighbors of $v_0$ are each adjacent to a vertex in $O$ and a vertex on the other side of the partition, and all other vertices in $C$ are adjacent to two vertices on the other side of the partition. Now, while any vertices remained unassigned we find a new vertex or vertices to assign by checking the following conditions in order of priority: \begin{enumerate} \item If there is any vertex $v$ that is adjacent to at least two vertices in $L\cup O$, we assign $v$ to $R$. \item If there is any $v$ that is adjacent to at least two vertices in $R\cup O$, we assign $v$ to $L$. \item If there is any $v$ that is adjacent to exactly two vertices $w_1,w_2$ in $A$, where $w_1\in L$ and $w_2\in R$, we proceed as follows. Note that there must be a path $u_0,u_1,\ldots,v$ from some other vertex $u_0\in A$ to $v$, or else $v$ would be a cut vertex. We may assume the interior vertices of the path are in $U$. We assign $u_1$ to the opposite side of the partition as $u_0$ (or either side if $u_0\in O$), and proceed to alternate assignments until we reach (and assign) $v$. Interior assigned vertices are assigned two neighbors on the opposite side of the partition. The vertex $v $ itself has one neighbor that was just assigned to the opposite side of the partition. Together with one of its previously assigned neighbors, it now has two neighbors with the opposite assignment as itself. \begin{figure}[H] \caption{Case 3 of Algorithm. A red label indicates the label was newly assigned in this stage. }\label{case3} \centering \begin{tikzpicture} \tikzstyle{vertex} = [circle, draw=black] \tikzstyle{filled_vertex} = [circle, draw=black, fill = blue!40] \draw(0,0) ellipse(3.5cm and 1.75cm); \node (v0) at (0,1){$A$}; \node[vertex,label=above:{$w_1$}] (v1) at (-1.5,-1){\tiny L}; \node[vertex,label=above:{$w_2$}] (v2) at (-.5,-1){\tiny R}; \node[vertex,text=red,label=below:{$v$}] (v3) at (-1,-2.25){\tiny R}; \node[vertex,text=red] (v4) at (0,-2.25){\tiny L}; \node[vertex,text=red] (v5) at (1,-2.25){\tiny R}; \node[vertex,text=red,label=below:{$u_1$}] (v6) at (2,-2.25){\tiny L}; \node[vertex,label=above:{$u_0$}] (v7) at (2,-.75){\tiny R}; \draw (v1) -- (v3); \draw (v2) -- (v3); \draw (v3) -- (v4); \draw (v4) -- (v5); \draw (v6) -- (v5); \draw (v6) -- (v7); \end{tikzpicture} \end{figure} \item If none of the previous three cases apply at any stage, we have that all vertices in $U$ that have adjacencies in $A$ have exactly one neighbor in $A$. Let $H$ be the subgraph induced by $U$. Note that $\delta(H)\geq 2$, so $H$ contains a cycle. If $H$ contains a cycle of even length, then we may simply assign vertices in the cycle to alternating sides of the partition, as in the initial step of the algorithm. So we may assume that all cycles have odd length. Choose any cycle $C$ in $H$. Let $P$ be a path from any vertex $c_0\in C$ to a vertex $v_U\in U$, such that $v_u$ is adjacent to a vertex $v_A\in A$. We may choose $P$ so that $P\cap C=\{c_0\}$ and $P\cap A =\{v_A\}$. (It is possible that $c_0=v_U$.) Note that there must be a path $Q$ from some vertex $c_i\in C$, $c_i\neq c_0$, to a vertex $w_U\in U$ that is adjacent to a vertex $w_A\in A$, or else $c_0$ would be a cut vertex of $G$. We may also assume $Q\cap C=\{c_i\}$ and $Q\cap A=\{w_A\}$. It is possible that $v_A=w_A$. However, $P\cap Q=\emptyset$, for otherwise we would be able to produce two cycles of opposite parity containing $v_U$ using $P$, $Q$ and the two paths from $c_0$ to $c_i$ in $C$, and hence $U$ would contain an even cycle. We have two paths from $v_U$ to $w_U$ in $H$ via $c_0$ and $c_i$, taking opposite paths around $C$. Since $C$ is of odd length, the paths' lengths have opposite parity. If $v_A$ and $w_A$ are on opposite side of the partition, then give $v_U$ the opposite label of $v_A$, and proceed along the even-length path alternating labels. If $v_A$ and $w_A$ are on opposite side of the partition, then give $v_U$ the opposite label of $v_A$, and proceed along the odd-length path alternating labels. \end{enumerate} \begin{figure}[H] \caption{Case 4 of Algorithm. In the event that $v_A$ and $w_A$ have opposite labels, we alternate labels along an odd-length path $v_A$ and $w_A$ } \centering \begin{tikzpicture} \tikzstyle{vertex} = [circle, draw=black, minimum size = .5cm] \tikzstyle{filled_vertex} = [circle, draw=black, fill = blue!40] \node[vertex,label=below:{$v_A$}] (v1) at (0,0){\tiny L}; \node[vertex,text=red,label={[align=left]$c_0=$\\$v_U$}] (v2) at (1,0){\tiny R}; \node[vertex,text=red] (v3) at (2.5,1){\tiny L}; \node[vertex,text=red] (v4) at (2,-1){}; x\node[vertex,text=red] (v5) at (3,-1){}; \node[vertex,text=red,label=below:{$c_i$}] (v6) at (4,0){\tiny R}; \node[vertex,text=red,label=below:{$w_U$}] (v7) at (5,0){\tiny L}; \node[vertex,label=below:{$w_A$}] (v8) at (6,0){\tiny R}; \draw (v1)--(v2); \draw (v2)--(v3); \draw (v2)--(v4); \draw (v4)--(v5); \draw (v6)--(v5); \draw (v3)--(v6); \draw (v7)--(v6); \draw (v7)--(v8); \end{tikzpicture} \end{figure} \begin{figure}[H] \caption{Case 4 of Algorithm. In the event that $v_A$ and $w_A$ have the same label, we alternate labels along an even-length path $v_A$ and $w_A$ } \centering \begin{tikzpicture} \tikzstyle{vertex} = [circle, draw=black, minimum size = .5cm] \tikzstyle{filled_vertex} = [circle, draw=black, fill = blue!40] \node[vertex,label=below:{$v_A$}] (v1) at (0,0){\tiny L}; \node[vertex,text=red,label={[align=left]$c_0=$\\$v_U$}] (v2) at (1,0){\tiny R}; \node[vertex,text=red] (v3) at (2.5,1){}; \node[vertex,text=red] (v4) at (2,-1){\tiny L}; \node[vertex,text=red] (v5) at (3,-1){\tiny R}; \node[vertex,text=red,label=below:{$c_i$}] (v6) at (4,0){\tiny L}; \node[vertex,text=red,label=below:{$w_U$}] (v7) at (5,0){\tiny R}; \node[vertex,label=below:{$w_A$}] (v8) at (6,0){\tiny L}; \draw (v1)--(v2); \draw (v2)--(v3); \draw (v2)--(v4); \draw (v4)--(v5); \draw (v6)--(v5); \draw (v3)--(v6); \draw (v7)--(v6); \draw (v7)--(v8); \end{tikzpicture} \end{figure} Any time we make new assignments, we repeat the process, checking which case applies in order. The cases are exhaustive and in any of the cases we can assign labels to more vertices while maintaining the property that any labeled vertex has as least two neighbors each of which either has the opposite label or is labeled $O$. Hence the process will terminate with all vertices assigned a label and the final labeling satisfying the required property. \end{algo} In the case where $G$ contains an even cycle, note that $O=\emptyset$, and all vertices are labeled either $L$ or $R$ upon termination. Therefore $F(G)\geq \lceil\frac{n}{2}\rceil$ by Lemma \ref{bipartite}. Lemma \ref{cut_vert} and Algorithm \ref{algo} exhaust all possibilities for connected graphs with minimum degree at least three. Hence, we have proven Theorem \ref{delta3}. \section{The General Case} Next we handle the case where $G$ has vertices of degree less than three with an inductive argument. \begin{theorem} Let $G=(V,E)$ be a graph on $n$ vertices. $F(G)\geq \lfloor\frac{n-1}{2}\rfloor$ \end{theorem} \begin{proof} If $G$ is disconnected, we may find a failed zero set $S$ satisfying $|S|\geq \lceil\frac{n}{2}\rceil$ by filling all vertices in all but the smallest connected component of $G$. If $G$ is connected and $\delta(G)\geq 3$, we know from Theorem \ref{delta3} that the desired inequality holds. So we only need to consider the cases where $\delta(G)=1$ and $\delta(G)=2$. We proceed by induction on $n$. If $n=1$, $F(G)=0=\lfloor\frac{n-1}{2}\rfloor$ and the result holds. Now assume the inequality holds for all graphs with fewer than $n$ vertices. Suppose $\delta(G) = 1$ and let $v \in V$ with $\deg(v) = 1$. Let $v$ be adjacent to $w$ and let $G'$ be the subgraph induced by $V' = V-\{v,w\}$ . By the inductive hypothesis, $F(G') \ge \lfloor \frac{(n - 2) - 1}{2} \rfloor$. Let $S'\subset V'$ be a stalled set that achieves this bound. \begin{figure}[H] \caption{The graph $G$ with $\delta(G) = 1$. Case where $w$ is adjacent to $x \in V' - S'$.} \centering \begin{tikzpicture} \node[unfilled, label=below:{$v$}] at (0,0) (v) {}; \node[unfilled, label=below:{$w$}] at (1,0) (w) {}; \node[filled] at (2, -1) (w_1) {}; \node[unfilled, label=right:{$x$}] at (2, 0) (w_2) {}; \node[filled] at (2, 1) (w_3) {}; \draw (v) -- (w); \draw (w) -- (w_1); \draw (w) -- (w_2); \draw (w) -- (w_3); \node[collection, minimum height = 4cm, minimum width = 3.5cm, dotted] at (3.2, 0) (W) {$G'$}; \end{tikzpicture} \begin{tikzpicture} \node[unfilled, label=below:{$v$}] at (0,0) (v) {}; \node[filled, label=below:{$w$}] at (1,0) (w) {}; \node[filled] at (2, -1) (w_1) {}; \node[unfilled, label=right:{$x$}] at (2, 0) (w_2) {}; \node[filled] at (2, 1) (w_3) {}; \draw (v) -- (w); \draw (w) -- (w_1); \draw (w) -- (w_2); \draw (w) -- (w_3); \node[collection, minimum height = 4cm, minimum width = 3.5cm, dotted] at (3.2, 0) (W) {$G'$}; \end{tikzpicture} \end{figure} Suppose $w$ is adjacent to an unfilled vertex $x\in V'$. We claim $S = S' \cup \{w\}$ is stalled in $G$. Consider $y \in S'$. We see $y$ is either only adjacent to vertices in $S' \cup \{w \}$, or it is adjacent to at least 2 unfilled vertices in $G'$. Either way, $y$ cannot force any of its neighbors in $G$. We also see that $w$ cannot force its neighbors since $w$ is adjacent to $v$ and $x$ which are both unfilled. Thus, $S$ is stalled. Note $|S' \cup \{w\}| \ge \lfloor \frac{(n - 2) - 1}{2} \rfloor + 1 = \lfloor \frac{n - 1}{2} \rfloor$. \begin{figure}[H] \caption{The graph $G$ with $\delta(G) = 1$. Case where $w$ is adjacent to only vertices in $S' \cup v$.} \centering \begin{tikzpicture} \node[unfilled, label=below:{$v$}] at (0,0) (v) {}; \node[unfilled, label=below:{$w$}] at (1,0) (w) {}; \node[filled] at (2, -1) (w_1) {}; \node[filled] at (2, 0) (w_2) {}; \node[filled] at (2, 1) (w_3) {}; \draw (v) -- (w); \draw (w) -- (w_1); \draw (w) -- (w_2); \draw (w) -- (w_3); \node[collection, minimum height = 4cm, minimum width = 3.5cm, dotted] at (3.2, 0) (W) {$G'$}; \end{tikzpicture} \begin{tikzpicture} \node[filled, label=below:{$v$}] at (0,0) (v) {}; \node[filled, label=below:{$w$}] at (1,0) (w) {}; \node[filled] at (2, -1) (w_1) {}; \node[filled] at (2, 0) (w_2) {}; \node[filled] at (2, 1) (w_3) {}; \draw (v) -- (w); \draw (w) -- (w_1); \draw (w) -- (w_2); \draw (w) -- (w_3); \node[collection, minimum height = 4cm, minimum width = 3.5cm, dotted] at (3.2, 0) (W) {$G'$}; \end{tikzpicture} \end{figure} If $w$ is adjacent to only filled vertices in $G'$, we see then $S = S' \cup \{v, w\}$ is stalled. This is because $v$ and $w$ do not cause any forcing in $G'$. Since $S' \ne V'$, we have $S \ne V$. So $S$ is a failed forcing set with $|S| = |S'| + 2 \ge \lfloor \frac{n + 1}{2} \rfloor$. In either case, we find a stalled set of size at least $\lfloor \frac{n - 1}{2} \rfloor$. Now let $\delta(G) = 2$. Then there exists some vertex $v \in V$ such that $\deg(v) = 2$. Let $x,y$ be the neighbors of $v$. Let $X$ be the set of all the vertices in $V - \{v\}$ adjacent to $x$. Let $Y$ be the set of all the vertices in $V - \{v\}$ adjacent to $y$. We construct a graph $G'$ by condensing the $x-v-y$ subgraph to a single vertex $w$. More precisely, let $G'=(V',E')$ be obtained from the subgraph of $G$ induced by $V-\{v,x,y\}$, by adding a new vertex $w$, and adding the edge $(w,z)$ for each $z\in X\cup Y$. \begin{figure}[H] \caption{Graph $G$ (left) with $\delta(G) = 2$ with corresponding $G'$ (right).} \centering \begin{tikzpicture} \node[unfilled, label=above:{$v$}] at (0,0) (v) {}; \node[unfilled, label=above:{$x$}] at (-1,-1) (x) {}; \node[unfilled, label=above:{$y$}] at (1,-1) (y) {}; \node[] at (-1.5, -2) (x_1) {}; \node[] at (-0.5, -2) (x_2) {}; \node[] at (1.5, -2) (y_1) {}; \node[] at (0.5, -2) (y_2) {}; \draw (y) -- (v) -- (x); \draw (x) -- (x_1); \draw (x) -- (x_2); \draw (y) -- (y_1); \draw (y) -- (y_2); \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (-1, -2) (X) {$X$}; \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (1, -2) (Y) {$Y$}; \end{tikzpicture} \begin{tikzpicture} \node[unfilled, label=above:{$w$}] at (0,0) (w) {}; \node[] at (-1.5, -1) (x_1) {}; \node[] at (-0.5, -1) (x_2) {}; \node[] at (1.5, -1) (y_1) {}; \node[] at (0.5, -1) (y_2) {}; \draw (w) -- (x_1); \draw (w) -- (x_2); \draw (w) -- (y_1); \draw (w) -- (y_2); \node[collection, minimum width = 4cm, minimum height = 0.75cm] at (0, -1) (XY) {$X \cup Y$}; \end{tikzpicture} \end{figure} We inductively assume $F(G') \ge \lfloor \frac{(n - 2) - 1}{2} \rfloor$. Let $S'$ be a corresponding stalled set. We consider the following cases: \begin{enumerate} \item \begin{figure}[H] \caption{Case when $w \not \in S'$ and corresponding stalled set in $G$.} \centering \begin{tikzpicture} \node[unfilled, label=above:{$w$}] at (0,0) (w) {}; \node[] at (-1.5, -1) (x_1) {}; \node[] at (-0.5, -1) (x_2) {}; \node[] at (1.5, -1) (y_1) {}; \node[] at (0.5, -1) (y_2) {}; \draw (w) -- (x_1); \draw (w) -- (x_2); \draw (w) -- (y_1); \draw (w) -- (y_2); \node[collection, minimum width = 4cm, minimum height = 0.75cm] at (0, -1) (XY) {}; \end{tikzpicture} \begin{tikzpicture} \node[filled, label=above:{$v$}] at (0,0) (v) {}; \node[unfilled, label=above:{$x$}] at (-1,-1) (x) {}; \node[unfilled, label=above:{$y$}] at (1,-1) (y) {}; \node[] at (-1.5, -2) (x_1) {}; \node[] at (-0.5, -2) (x_2) {}; \node[] at (1.5, -2) (y_1) {}; \node[] at (0.5, -2) (y_2) {}; \draw (y) -- (v) -- (x); \draw (x) -- (x_1); \draw (x) -- (x_2); \draw (y) -- (y_1); \draw (y) -- (y_2); \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (-1, -2) (X) {}; \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (1, -2) (Y) {}; \end{tikzpicture} \end{figure} If $w \not \in S'$, then $S = S' \cup \{v\}$ is stalled in $G$. We see that any $z \in S'$ that could not force $G'$ also cannot force in $G$ since $x$ and $y$ are unfilled to match the unfilled $w$. We also see that $v$ cannot zero force in $G$ since $v$ is adjacent to $x,y \not \in S$. Thus, $S$ is a stalled set. \item If $w \in S'$, it is either adjacent to only filled vertices or adjacent to at least two unfilled vertices in $G'$. \begin{enumerate} \item \begin{figure}[H] \caption{Case when $w$ is adjacent to filled vertices in $G'$ and corresponding stalled set in $G$.} \centering \begin{tikzpicture} \node[filled, label=above:{$w$}] at (0,0) (w) {}; \node[filled] at (-1.5, -1) (x_1) {}; \node[filled] at (-0.5, -1) (x_2) {}; \node[filled] at (1.5, -1) (y_1) {}; \node[filled] at (0.5, -1) (y_2) {}; \draw (w) -- (x_1); \draw (w) -- (x_2); \draw (w) -- (y_1); \draw (w) -- (y_2); \node[collection, minimum width = 4cm, minimum height = 0.75cm] at (0, -1) (XY) {}; \end{tikzpicture} \begin{tikzpicture} \node[filled, label=above:{$v$}] at (0,0) (v) {}; \node[filled, label=above:{$x$}] at (-1,-1) (x) {}; \node[filled, label=above:{$y$}] at (1,-1) (y) {}; \node[filled] at (-1.5, -2) (x_1) {}; \node[filled] at (-0.5, -2) (x_2) {}; \node[filled] at (1.5, -2) (y_1) {}; \node[filled] at (0.5, -2) (y_2) {}; \draw (y) -- (v) -- (x); \draw (x) -- (x_1); \draw (x) -- (x_2); \draw (y) -- (y_1); \draw (y) -- (y_2); \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (-1, -2) (X) {}; \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (1, -2) (Y) {}; \end{tikzpicture} \end{figure} If $w$ is adjacent to only filled vertices in $G'$, then $S = (S' - \{w\}) \cup \{v,x,y\}$ is a stalled set in $G$. If a vertex was spent in $S'$, we see it must also be spent in $S$ since both $x$ and $y$ get filled. If a vertex $z$ was adjacent to at least two unfilled vertices in $S'$, it must also be adjacent to at least two unfilled vertices in $S$ as $z$ could not not have been adjacent to $x$ or $y$. Thus, $S$ is stalled. \item If $w$ is adjacent to at least two vertices not in $S'$, we will fill $x$ and $y$ in $G$. We first observe that every vertex in $(S' - \{w\})$ in $G$ is spent or adjacent to at least two unfilled vertices. If a vertex $z \in S'$ was adjacent to $w$, then $z$ is adjacent to $x$ or $y$ or both in $G$. Since we fill $x$ and $y$ to match $w$ being filled, we see $z$ cannot force in $G'$. Since $w$ is adjacent to at least two unfilled vertices in $G'$, at least two vertices in $X \cup Y$ are not in $S'$. \begin{figure}[H] \caption{Case when $w$ is adjacent to exactly one unfilled vertex in $X$ and one in $Y$. Corresponding stalled set in $G$.} \label{fig:oneXoneY} \centering \begin{tikzpicture} \node[filled, label=above:{$w$}] at (0,0) (w) {}; \node[filled] at (-1.5, -1) (x_1) {}; \node[unfilled] at (-0.5, -1) (x_2) {}; \node[filled] at (1.5, -1) (y_1) {}; \node[unfilled] at (0.5, -1) (y_2) {}; \draw (w) -- (x_1); \draw (w) -- (x_2); \draw (w) -- (y_1); \draw (w) -- (y_2); \node[collection, minimum width = 4cm, minimum height = 0.75cm] at (0, -1) (XY) {}; \end{tikzpicture} \begin{tikzpicture} \node[unfilled, label=above:{$v$}] at (0,0) (v) {}; \node[filled, label=above:{$x$}] at (-1,-1) (x) {}; \node[filled, label=above:{$y$}] at (1,-1) (y) {}; \node[filled] at (-1.5, -2) (x_1) {}; \node[unfilled] at (-0.5, -2) (x_2) {}; \node[filled] at (1.5, -2) (y_1) {}; \node[unfilled] at (0.5, -2) (y_2) {}; \draw (y) -- (v) -- (x); \draw (x) -- (x_1); \draw (x) -- (x_2); \draw (y) -- (y_1); \draw (y) -- (y_2); \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (-1, -2) (X) {}; \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (1, -2) (Y) {}; \end{tikzpicture} \end{figure} \begin{enumerate} \item If there is at least one vertex in $X - S'$ and at least one vertex in $Y - S'$, then both $x$ and $y$ are adjacent to at least two unfilled vertices in $G$ since both $x$ and $y$ are adjacent to some unfilled vertex in $(X \cup Y)$ as well as $v$. This means $S = (S' - \{w\}) \cup \{x, y\}$ is a stalled set in $G$. (As seen in Figure \ref{fig:oneXoneY}). \item \begin{figure}[H] \caption{Case when $w$ is adjacent to at least two unfilled vertices in $X$ or $Y$. Corresponding stalled set in $G$.} \centering \begin{tikzpicture} \node[filled, label=above:{$w$}] at (0,0) (w) {}; \node[filled] at (-1.5, -1) (x_1) {}; \node[filled] at (-0.5, -1) (x_2) {}; \node[unfilled] at (1.5, -1) (y_1) {}; \node[unfilled] at (0.5, -1) (y_2) {}; \draw (w) -- (x_1); \draw (w) -- (x_2); \draw (w) -- (y_1); \draw (w) -- (y_2); \node[collection, minimum width = 4cm, minimum height = 0.75cm] at (0, -1) (XY) {}; \end{tikzpicture} \begin{tikzpicture} \node[filled, label=above:{$v$}] at (0,0) (v) {}; \node[filled, label=above:{$x$}] at (-1,-1) (x) {}; \node[filled, label=above:{$y$}] at (1,-1) (y) {}; \node[filled] at (-1.5, -2) (x_1) {}; \node[filled] at (-0.5, -2) (x_2) {}; \node[unfilled] at (1.5, -2) (y_1) {}; \node[unfilled] at (0.5, -2) (y_2) {}; \draw (y) -- (v) -- (x); \draw (x) -- (x_1); \draw (x) -- (x_2); \draw (y) -- (y_1); \draw (y) -- (y_2); \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (-1, -2) (X) {}; \node[collection, minimum width = 2.5cm, minimum height = 0.75cm] at (1, -2) (Y) {}; \end{tikzpicture} \end{figure} If $X - S' = \emptyset$ or $Y - S' = \emptyset$, we claim $S = (S' - \{w\}) \cup \{v, x, y\}$ is stalled in $G$. Without loss of generality, assume $X - S' = \emptyset$ and $|Y - S'| \ge 2$. Note $x$ has neighbors $X \cup \{v\}$ in $G$. Since $v \in S$ and $X \subset S$, all of $x$'s neighbors are in $S$. Also, since $x, y \in S$, all of $v$'s neighbors are in $S$. Finally, $y$ has at least two vertices in $V - S$ since $|Y - S'| \ge 2$. Thus, none of $x, y$, or $v$ can force, making $S$ a stalled set. \end{enumerate} \end{enumerate} \end{enumerate} In any case, we see we can fill at least one more vertex in $G$ than we could $G'$. That is, $|S| \ge |S'| + 1$ and thus $|S| \ge \lfloor \frac{n-1}{2} \rfloor$. \end{proof} \section{Conclusions and Future Work} We now know there are finitely many graphs where $F(G) = k$ for any given $k$ and we can enumerate them by checking all graphs with fewer than $2k + 2$ vertices. For $k\leq 4$ we can do this fairly quickly by computing $F(G)$ for all graphs $G$ with no more than ten vertices. For $k > 4$ the number of graphs one must check gets very large. We let $F_k$ be the set of all connected graphs $G$ such that $F(G) = k$, and $F_k(n)$ be the number of graphs in $F_k$ on $n$ vertices. Using a exhaustive search on all connected graphs up to 10 vertices we get: \begin{figure}[H] \caption{} \centering \begin{tabular}{c|cccccccc} $n =$ & 3& 4& 5& 6& 7& 8& 9& 10 \\ \hline $F_1(n)$ & 2& 1& & & & & & \\ $F_2(n)$ & & 5& 5& 2& & & & \\ $F_3(n)$ & & &16&29&16& 1& & \\ $F_4(n)$ & & & &81&277&268&14& 1\\ \end{tabular} \end{figure} The rows of the table correspond to graphs in $F_k$ while the columns correspond to graphs with $n$ vertices. For example, $F_1$ contains three graphs: two with three vertices and one with four vertices. We can sum the rows to get $|F_2| = 12$, $|F_3| = 62$, and $|F_4| = 641$. This confirms the result by \cite{gomez} where it was shown that there are $15$ graphs with a failed zero forcing number of two, twelve of which were connected. We can apply our lower bound to another question problem posed in $\cite{fetcie}$: the classification of graphs that satisfy $Z(G) = F(G)$. For small $k$ we can enumerate all graphs where $F(G) = k$, and see which of these satisfy $Z(G) = k$. This gives us an exhaustive list of $G$ for which $F(G) = Z(G) = k$. We define $E_k$ to be the set of all connected graphs $G$ such that $F(G) = Z(G) = k$, and $E_k(n)$ to be the number of graphs on $n$ vertices in $E_k$. We have computed $E_k(n)$ for $1 \le k \le 4$ and have found the following:\\ \begin{figure}[H]\label{ekn} \caption{} \centering \begin{tabular}{c|cccccccc} $n = $ & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\ \hline $E_1(n)$ & 1& 1& & & & & & \\ $E_2(n)$ & & 4& 4& 1& & & & \\ $E_3(n)$ & & & 9&10& 4& & & \\ $E_4(n)$ & & & &19&29& 2& & \\ \end{tabular} \end{figure} The columns of the table correspond to graphs with $n$ vertices; the sum across a row is then $|E_k|$. For example, the pair of $1$s in the first row corresponds to two graphs in $E_1$, one with three vertices and the other with four. These graphs are $P_3$ and $P_4$, the only graphs such that $F(G) = Z(G) = 1$. Furthermore, we see that $|E_1| = 2$, $|E_2| = 9$, $|E_3| = 23$, and $|E_4| = 50$. In \cite{fetcie} it was claimed that graphs where $F(G) = Z(G)$ seem to be rare. The above gives us a quantitative method to evaluate this claim, by finding the number of graphs with a fixed failed zero forcing number $k$ that have the same zero forcing number. For example, we see that among the 641 graphs with failed zero forcing number of four, exactly 50 have zero forcing number four. Interesting questions to consider include finding bounds on $|E_k|$ and characterizing $n$ for which $E_k(n)\neq 0$. Toward the latter question, we note $k = F(G) \le n - 2$ for any connected graph $G$ and so any graph with $n$ vertices in $E_k$ must satisfy $n \ge k + 2$. This inequality gives us the lower diagonal in Figure 12 and Figure 13. However, we see an upper diagonal in Figure 13 that is stricter than the one in Figure 12. That is, given a connected graph $G$ with $n$ vertices in $E_k$, we ask whether $n \le k + 4$. We have checked this holds true for all connected graphs $G$ up to ten vertices.

https://arxiv.org/abs/1612.04018

Trigonometric Interpolation and Quadrature in Perturbed Points

The trigonometric interpolants to a periodic function $f$ in equispaced points converge if $f$ is Dini-continuous, and the associated quadrature formula, the trapezoidal rule, converges if $f$ is continuous. What if the points are perturbed? With equispaced grid spacing $h$, let each point be perturbed by an arbitrary amount $\le \alpha h$, where $\alpha\in [\kern .5pt 0,1/2)$ is a fixed constant. The Kadec 1/4 theorem of sampling theory suggests there may be be trouble for $\alpha\ge 1/4$. We show that convergence of both the interpolants and the quadrature estimates is guaranteed for all $\alpha<1/2$ if $f$ is twice continuously differentiable, with the convergence rate depending on the smoothness of $f$. More precisely it is enough for $f$ to have $4\alpha$ derivatives in a certain sense, and we conjecture that $2\alpha$ derivatives is enough. Connections with the Fejér--Kalmár theorem are discussed.

\section{Introduction and summary of results} The basic question of robustness of mathematical algorithms is, what happens if the data are perturbed? Yet little literature exists on the effect on interpolants, or on quadratures, of perturbing the interpolation points. The questions addressed in this paper arise in two almost equivalent settings: interpolation by algebraic polynomials (e.g., in Gauss or Chebyshev points) and periodic interpolation by trigonometric polynomials (e.g., in equispaced points). Although we believe essentially the same results hold in the two settings, this paper deals with just the trigonometric case. Let $f$ be a real or complex function on $[-\pi,\pi)$, which we take to be $2\pi$-periodic in the sense that any assumptions of continuity or smoothness made for $f$ apply periodically at $x=-\pi$ as well as at interior points. For each $N\ge 0$, set $K = 2N+1$, and consider the centered grid of $K$ equispaced points in $[-\pi,\pi)$, \begin{equation} x_k^{} = kh, \quad -N \le k \le N, \quad h = {2\pi\over K}. \end{equation} There is a unique degree-$N$ trigonometric interpolant through the data $\{f(x_k^{})\}$, by which we mean a function \begin{equation} t_N^{}(x) = \sum_{k=-N}^N c_k^{} e^{ikx} \end{equation} with $t_N^{}(x_k^{}) = f(x_k^{})$ for each $k$. If $I$ denotes the integral of $f$, \begin{equation} I = \int_{-\pi}^\pi f(x) \,dx, \end{equation} the associated quadrature approximation is the integral of $t_N^{}(x)$, which can be shown to be equal to the result of applying the trapezoidal rule to $f$: \begin{equation} I_N^{} = h\!\sum_{k=-N}^N f(x_k^{}) = \int_{-\pi}^\pi t_N^{}(x) \,dx = 2\pi c_0^{} . \label{traprule} \end{equation} It is known that if $f$ is continuous, then \begin{equation} \lim_{N\to\infty} | I - I_N^{} | = 0, \label{quadconv} \end{equation} and if $f$ is Dini-continuous, for which H\"older or Lipschitz continuity are sufficient conditions, then \begin{equation} \lim_{N\to\infty} \| f - t_N^{} \| = 0 . \label{approxconv} \end{equation} Moreover, the convergence rates are tied to the smoothness of $f$, with exponential convergence if $f$ is analytic. Here and throughout, $\|\cdot\|$ is the maximum norm on $[-\pi,\pi)$. The problem addressed in this paper is the generalization of these results to configurations in which the interpolation points are perturbed. For fixed $\alpha\in ( 0,1/2)$, consider a set of points \begin{equation} \tilde x_k^{} = x_k^{} + s_k^{} h, \quad -N\le k \le N, \quad |s_k|\le \alpha \, . \label{ppoints} \end{equation} Note that since $\alpha < 1/2$, the $\tilde x_k^{}$ are necessarily distinct. Let $\tilde t_N^{}(x)$ be the unique degree-$N$ trigonometric interpolant to $\{f(\tilde x_k^{})\}$, and let $\tilde I_N^{} = \int \tilde t_N^{}(x) \kern .7pt dx$ be the corresponding quadrature approximation. As in (\ref{traprule}), this will be a linear combination of the function values, although no longer with equal weights in general. Let $\sigma>0$ be any positive real number, and write $\sigma = \nu + \gamma$ with $\gamma\in (\kern .5pt 0,1]$. We say that $f$ {\em has $\sigma$ derivatives} if $f$ is $\nu$ times continuously differentiable and, moreover, $f^{(\nu)}$ is H\"older continuous with exponent $\gamma$. Note that if $\sigma$ is an integer, then for $f$ to ``have $\sigma$ derivatives'' means that $f$ is $\sigma-1$ times continuously differentiable and $f^{(\sigma-1)}$ is Lipschitz continuous. We will prove the following main theorem, whose central estimate is the bound on $\|f-\tilde t_N^{}\|$ in (\ref{thmrate2}). The estimates (\ref{thmconv})--(\ref{thmrate2}) are new, whereas (\ref{thmrate3}) follows from the work of Kis~\cite{kis}, as discussed in Section~\ref{confluent}. Numerical illustrations of these bounds can be found in~\cite{austin}. \medskip \begin{theorem} \label{thm1} For any $\alpha\in (\kern .5pt 0,1/2)$, if $f$ is twice continuously differentiable, then \begin{equation} \lim_{N\to\infty} | I - \tilde I_N^{} | = \lim_{N\to\infty} \| f - \tilde t_N^{} \| = 0. \label{thmconv} \end{equation} More precisely, if $f$ has $\sigma$ derivatives for some $\sigma > 4\alpha$, then \begin{equation} | I - \tilde I_N^{} | , \| f - \tilde t_N^{} \| = O(N^{4\alpha-\sigma}). \label{thmrate2} \end{equation} If $f$ can be analytically continued to a $2\pi$-periodic function for $-a < {\rm Im}\kern 2pt x < a$ for some $a>0$, then for any\/ $\hat a < a$, \begin{equation} | I - \tilde I_N^{} | , \| f - \tilde t_N^{} \| = O(e^{-\hat aN}). \label{thmrate3} \end{equation} \end{theorem} \medskip Our proofs are based on combining standard estimates of approximation theory, the Jackson theorems, with a new bound on the Lebesgue constants associated with perturbed grids, Theorem~\ref{thmbound}. Our bounds are close to sharp, but not quite. Based on extensive numerical experiments presented in Section 3.3.2 of~\cite{austin}, we conjecture that $4\alpha$ can be improved to $2\alpha$ in (\ref{thmrate2}) and (\ref{bigestimate}); for (\ref{bigestimate}) the result would probably then be sharp, but for (\ref{thmrate2}) a slight further improvement may still be possible. For the quadrature problem in particular, further experiments presented in Section 3.5.2 of~\cite{austin} lead us to conjecture that $\tilde I_N^{} \to I$ as $N\to\infty$ for all continuous functions $f$ for all $\alpha<1/2$. This conjecture is based on the theory of P\'olya in 1933~\cite{polya}, who showed that such convergence is ensured if and only if the sums of the absolute values of the quadrature weights are bounded as $N\to\infty$. Experiments indicate that for all $\alpha<1/2$, these sums are indeed bounded as required. On the other hand, $\tilde I_N^{}\to I$ cannot be guaranteed for any $\alpha\ge 1/2$, since in that case the interpolation points may come together, making the quadrature weights unbounded. Theorems~\ref{thm1} and~\ref{thmbound} suggest that from the point of view of approximation and quadrature, $\alpha = 1/4$ is not a special value. In Section~\ref{sampling} we comment on the significance of the appearance of this number in the Kadec 1/4 theorem and more generally on the relationship between approximation theory and sampling theory, two subjects that address closely related questions and yet have little overlap of literatures or experts. All the estimates reported here were worked out by the first author and presented in his D.\ Phil.\ thesis~\cite{austin}. This work was motivated by work of the second author with Weideman in the review article ``The exponentially convergent trapezoidal rule''~\cite{tw}. It is well known that on an equispaced periodic grid, the trapezoidal rule is exponentially convergent for periodic analytic integrands~\cite{davis,tw}. With perturbed points, it seemed to us that exponential convergence should still be expected, and we were surprised to find that there seemed to be no literature on this subject. A preliminary discussion was given in~\cite[Sec.~9]{tw}. Section~\ref{lebesguesec} reduces Theorem~\ref{thm1} to a bound on the Lebesgue constant, Theorem~\ref{thmbound}. Sections~\ref{confluent} and \ref{sampling} are devoted to comments on problems with $\alpha\ge 1/2$ and on the link with sampling theory and Kadec's theorem, respectively. Section~\ref{bound} outlines the proof of Theorem~\ref{thmbound}. \section{\label{lebesguesec}Reduction to a Lebesgue constant estimate} A fundamental tool of approximation theory is the Lebesgue constant: for any linear projection $L:f\mapsto t$, the Lebesgue constant is the operator norm $\Lambda = \|L\|$. For our problem the operator is the map $\tilde L_N^{}$ from a function $f$ to its trigonometric interpolant $\tilde t_N^{}$ through the values $\{f(\tilde{x}_k^{})\}$, and the norm on $L$ is the operator norm induced by $\|\cdot\|$, the $\infty$-norm on $[-\pi,\pi)$. We denote the Lebesgue constant by $\tilde \Lambda_N$. Lebesgue constants are linked to quality of approximations by the following well-known bound. If $\tilde \Lambda_N^{}$ is the Lebesgue constant associated with the projection $\tilde L_N^{}: f \mapsto \tilde t_N^{}$ and $t_N^*$ is the best approximation to $f$ of degree $N$, then \begin{equation} \|f-\tilde t_N^{}\| \le (1+\tilde \Lambda_N^{}) \| f-t_N^*\| . \label{basicbound} \end{equation} It follows that if $\tilde \Lambda_N^{}$ is small, then $\tilde t_N^{}$ is a near-optimal approximation to $f$. If $f$ has a certain smoothness property for which the optimal approximations $t_N^*$ are known to converge at a certain rate, this implies that the interpolants $\tilde t_N^{}$ converge at nearly the same rate. Applying (\ref{basicbound}), we prove Theorem~\ref{thm1} by combining a bound on the Lebesgue constants $\tilde \Lambda_N^{}$ with bounds on the best approximation errors $\|f- t_N^*\|$. Our estimates of best approximations are standard Jackson theorems, going back to Dunham Jackson in 1911 and 1912. The nonstandard part of the argument, which from a technical point of view is the main contribution of this paper, is the following estimate of the Lebesgue constant, the proof of which is outlined in Section~\ref{bound}. \medskip \begin{theorem} \label{thmbound} There is a universal constant $C$ such that \begin{equation} \tilde \Lambda_N^{} \le {C(N^{4\alpha}-1)\over \alpha(1-2\alpha)} \label{bigestimate} \end{equation} for all\/ $\alpha\in [\kern .5pt 0,1/2)$ and $N> 0$. For $\alpha=0$ this bound is to be interpreted by its limiting value given, for example, by L'Hopital's rule, $\tilde\Lambda_N^{} \le 4\kern1pt C\log N$. \end{theorem} \medskip The $\log N$ bound for an equispaced grid with $\alpha = 0$ is standard, so the substantive result here concerns $\alpha\in (0,1/2)$. This is what we can prove, but as mentioned in the previous section, based on numerical experiments, we conjecture that (\ref{bigestimate}) actually holds with $N^{4\alpha}$ replaced by $N^{2\alpha}$. Given Theorem~\ref{thmbound}, we prove Theorem~\ref{thm1} as follows. \medskip {\em Proof of Theorem\/~$\ref{thm1}$, given Theorem\/~$\ref{thmbound}$.} The Jackson theorems of approximation theory relate the smoothness of a function $f$ to the accuracy of its best approximations~\cite{jackson,meinardus}. According to one of these theorems, given for example as Theorem~41 of~\cite{meinardus}, if $f$ is a periodic function on $[-\pi,\pi)$ that has $\sigma$ derivatives for some $\sigma>0$ in the sense defined in Section 1, then \begin{equation} \| f-t_N^*\| = O(N^{-\sigma}). \end{equation} Combining this with Theorem~\ref{thmbound} gives (\ref{thmrate2}). The bound (\ref{thmrate3}) follows similarly from the estimate \begin{equation} \| f-t_N^*\| = O(e^{-\hat aN}) \end{equation} for any $2\pi$-periodic function $f$ analytic and bounded in the strip of half-width $\hat a>0$ about the real axis; see, for example, eq.~(7.17) of~\cite{tw}. \endproof \section{\label{confluent}\boldmath $\alpha \ge 1/2$, confluent points, and analytic functions} Our framework (\ref{ppoints}) for perturbed points can be generalized to values $\alpha \ge 1/2$. For $\alpha\in [1/2, 1)$, two grid points may coalesce, so one must assume that $f'$ exists in order to ensure that there are appropriate data to define an interpolation problem (in this case, trigonometric Hermite interpolation). Similarly for $\alpha\in [1,3/2)$, three points may coalesce, so one must assume $f''$ exists; and so on analogously for any finite value of $\alpha$. (We wrap grid points around as necessary if the perturbation moves them outside of $[-\pi, \pi)$; equivalently, one could extend $f$ periodically.) Looking at the statement of Theorem~\ref{thm1} but considering values $\alpha \ge 1/2$, one notes that the assumption of $\sigma>4\alpha$ derivatives is enough to ensure that the necessary derivatives exist for the interpolation problem to make sense; the conjectured sharper condition of $\sigma> 2\alpha$ derivatives is also (\kern .7pt just) enough. This coincidence seems suggestive, and we consider it possible that Theorem~\ref{thm1} and its conjectured improvement with $2\alpha$ may in fact be valid for arbitrary $\alpha>0$, not just $\alpha \in (\kern .5pt 0,1/2)$. We have not attempted to prove this, however. As a practical matter, trouble can be expected in floating-point arithmetic as sample points coalesce, so we regard the case $\alpha \ge 1/2$ as somewhat theoretical. Going further, what if we allow arbitrary perturbations of the interpolation points, so that each $\tilde x_k^{}$ may lie anywhere in $[-\pi,\pi)$? Doing so makes sense mathematically if $f$ is infinitely differentiable; so in particular, it makes sense if $f$ is analytic, which implies that it can be analytically continued to a $2\pi$-periodic function on the whole real line. We are now in an area of approximation theory (and potential theory) going back to the work of Runge~\cite{runge} and Fej\'er~\cite{fejer}, in which a major contributor was Joseph Walsh~\cite{gaier,walsh}. For arbitrary $x_k^{}$, convergence will occur if $f$ is analytic in a sufficiently wide strip around the real axis in the complex $x$-plane. Repeated points are permitted, with interpolation at such points interpreted in the Hermite sense involving values of both the function and its derivatives. If the points $x_k$ are {\em uniformly distributed\/} in the sense that the fraction of points falling in any interval $[a,b)\subseteq [-\pi,\pi)$ converges to $(b-a)/2\pi$ as $N\to\infty$, then it is enough for $f$ to be analytic in {\em any\/} strip around the real axis. Such results were first developed for polynomial approximation on the unit circle of functions analytic in the unit disk, the so-called Fej\'er--Kalm\'ar theorem~\cite{fejer,kalmar,walsh}. The extension to functions analytic in an annulus was considered by Hlawka~\cite{hlawka}, and the equivalent problem of trigonometric interpolation of $2\pi$-periodic functions on $[-\pi,\pi)$ was considered by Kis~\cite{kis}. All these results may fail in practice because of rounding errors on the computer, however. For example, Figure~3.7 of~\cite{austin} shows an example with uniformly distributed random interpolation points in $[-\pi,\pi)$, with rounding errors beginning to take over at $N\approx 20$. For the case of interpolation by algebraic polynomials, this kind of effect is familiar in the context of the Runge phenomenon, where polynomial interpolants in equispaced points in $[-1,1]$ will diverge on a computer as $N\to\infty$ even for a function like $f(x) = \exp(x)$ for which in principle they should converge. \section{\label{sampling}Sampling theory and the Kadec \boldmath$1/4$ theorem} The field of approximation theory goes back to Borel, de la Vall\'ee Poussin, Fej\'er, Jackson, Lebesgue, and others at the beginning of the 20th century, and its central question might be characterized like this: \medskip \begin{quotation} \noindent \em Given a function $f$ of a certain regularity, how fast do its approximations of a given kind converge? \end{quotation} \medskip \noindent For example, if $f$ is periodic and analytic on $[-\pi, \pi)$, then its equispaced trigonometric interpolants converge exponentially. The same holds if $f$ is analytic in a strip surrounding the whole real line and satisfies a decay condition at $\infty$, with trigonometric interpolants generalized to interpolatory series of sinc functions. The field of sampling theory goes back to Gabor, Kotelnikov, Nyquist, Paley, Shannon, J. M. and E. T. Whittaker, and Wiener a few years later. Its central question might be characterized like this: \medskip \begin{quotation} \noindent \em Given a function $f$ of a certain regularity, which of its approximations of a given kind are exactly equal to $f\kern 1pt ?$ \end{quotation} \medskip \noindent For example, if $f$ is periodic and analytic on $[-\pi,\pi)$, then its equispaced trigonometric interpolant is exact if $f$ is band-limited (has a Fourier series of compact support) and the grid includes at least two points per wavelength for each wave number present in the series. The same holds if $f$ is a band-limited analytic function on the whole real line, with the Fourier series generalized to the Fourier transform, and again with trigonometric interpolation generalized to sinc interpolation. Obviously we have worded these characterizations to highlight the similarities between the two fields, which in fact differ in significant ways. Still, it is remarkable how little interaction there has been between the two. What makes this relevant to the present paper is that our theorems and orientation are very much those of approximation theory, whereas most of the scientific interest in perturbed grids in the past has been from the side of sampling theory, and the Kadec 1/4 theorem is the known result in this general area. Kadec's theorem is an answer to a question of sampling theory that originates with Paley and Wiener~\cite{pw}. The exponentials $\{\exp(i\lambda_k^{} x)\}$, $-\infty < k < \infty$, form an orthonormal basis for $L^2[-\pi,\pi]$ if $\lambda_k^{} = k$ for each $k$. Thus, the sampling theorist would say that one can recover a function $f\in L^2[-\pi,\pi]$ from its inner products with the functions $\{\exp(i\lambda_k^{} x)\}$. Now suppose these wave numbers are perturbed so that $|\tilde \lambda_k^{} - k | \le \alpha$ for some fixed $\alpha$. Can one still recover the signal? Specifically, does the family $\{\exp(i\tilde \lambda_k^{} x)\}$ form a {\em Riesz basis} for $L^2[-\pi,\pi]$, that is, a basis that is related to the original one by a bounded transformation with a bounded inverse? Paley and Wiener showed that this is always the case for $\alpha < 1/\pi^2$, and Levinson showed it is not always the case for $\alpha \ge 1/4$. Kadec's theorem shows that Levinson's construction was sharp: for any $\alpha < 1/4$, the family $\{\exp(i\tilde \lambda_k^{} x)\}$ forms a Riesz basis~\cite{aldgro,christensen,kadec,young}. Note that the standard setting of Kadec's theorem involves perturbation of wave numbers from equispaced values, in contrast to the results of this paper, which involve perturbation of interpolation points from equispaced values. In view of the Fourier transform, however, these settings are related, so one might imagine, based on Kadec's theorem, that $\alpha = 1/4$ might be a critical value for trigonometric interpolation in perturbed points. Instead, we have found that the critical value is $\alpha = 1/2$. We explain this apparent discrepancy as follows. The Paley--Wiener theory and Kadec's theorem are results concerning the $L^2$ norm, which in many applications would represent energy. In our application of trigonometric interpolation, something related to the $L^2$ norm does indeed happen at $\alpha = 1/4$. Suppose we look at a {\em $2$-norm Lebesgue constant} $\tilde\Lambda_N^{(2)}$ for the perturbed grid interpolation problem, defined as the operator norm on $L: f\mapsto \tilde t_N^{}$ induced by the discrete $\ell^2$-norm on the data $\{f(\tilde x_k^{})\}$ and on the Fourier coefficients of the interpolant $\tilde t_k^{}$. Numerical experiments reported in Section 3.4.3 of~\cite{austin} indicate that whereas the usual $\infty$-norm Lebesgue constant is unbounded for all $\alpha$, $\tilde\Lambda_N^{(2)}$ is bounded as $N\to\infty$ for any $\alpha < 1/4$ but not always bounded for $\alpha \ge 1/4$. (Indeed Kadec's theorem may imply this result.) For $\alpha \in (1/4,1/2)$, we conjecture $\tilde \Lambda_N^{(2)} = O(N^{4\alpha - 1})$. Thus a sampling theorist might say that for $\alpha \in [1/4,1/2)$, trigonometric interpolation is {\em unstable} in the sense that it may amplify signals unboundedly in $\ell^2$ as $N\to\infty$. On the other hand the approximation theorist might note that the instability is very weak, involving not even one power of $N$. Assuming that the conjectured sharpening of the estimate (\ref{thmrate2}) of Theorem~\ref{thm1} is valid, one derivative of smoothness of $f$ is enough to suppress the instability, ensuring $\|f-\tilde t_N^{}\|\to 0$ as $N\to\infty$ for all $\alpha < 1/2$. The numerical analyst might add that on a computer, amplification of rounding errors by $o(N)$ is unlikely to cause trouble. For $\alpha \ge 1/2$, in strong contrast, the amplification is unbounded in any norm even for finite $N$, and trouble is definitely to be expected. \section{\label{bound}Proof of the Lebesgue constant estimate, Theorem~\ref{thmbound}} A full proof of Theorem~\ref{thmbound}, filling 20 pages, is the subject of Chapter~4 of the first author's D.\ Phil.\ thesis~\cite{austin}. Many detailed trigonometric estimates are involved, and we do not know how to shorten it significantly. For readers interested in full details, that chapter has been made available in the Supplementary Materials attached to this paper. Here, we outline the argument. To prove the bound (\ref{bigestimate}) on the Lebesgue constant, \begin{equation} \tilde \Lambda_N^{} \le {C(N^{4\alpha}-1)\over \alpha(1-2\alpha)}, \label{bigestimateagain} \end{equation} we begin by noting that $\tilde \Lambda_N^{}$ is given by \begin{equation} \tilde \Lambda_N^{}= \max_{x\in[-\pi,\pi]} \tilde L(x), \label{maxdef} \end{equation} where $\tilde L$ is the Lebesgue function \begin{equation} \tilde L(x) = \sum_{k=-N}^N |\tilde \ell_k^{}(x) |, \label{basicsum} \end{equation} where $\tilde{\ell}_k^{}$ is the $k$th Lagrange cardinal trigonometric polynomial for the perturbed grid, \begin{equation} \tilde \ell_k^{}(x) = \prod_{j\ne k}\left. \sin\Big({x-\tilde x_j^{}\over 2}\Big) \right/ \sin\Big({\tilde x_k^{}-\tilde x_j^{}\over 2}\Big). \end{equation} The function $\tilde\ell_k^{}(x)$ takes the values $1$ at $\tilde x_k^{}$ and $0$ at the other grid points $\tilde x_j^{}$, and the sum (\ref{basicsum}) adds up contributions at a point $x$ from all the $2N+1$ cardinal functions associated with grid points to its left and right. The argument begins by showing that on the interval $[x_{-(k + 1)}^*, x_{-k}^*]$, $\tilde{\ell}_0$ satisfies the bound \begin{equation} |\tilde{\ell}_0(x)| \leq M_k, \quad x\in [x_{-(k + 1)}^*, x_{-k}^*], \quad 0 \leq k \leq N, \label{Mbounds} \end{equation} for certain numbers $M_0, \ldots, M_N$, independently of the choice of perturbed points $\{\tilde{x}_k\}$. The points $x_k^*$ are defined by $x_0^* = 0$, $x_{-(N + 1)}^* = -\pi$, and \[ x_k^* = 2\arctan\left(\frac{\cos(kh) - \cos(\alpha h) + \tan(\tilde{x}_0/2)\sin(kh)}{\tan(\tilde{x}_0/2)\bigl(\cos(kh) + \cos(\alpha h)\bigr) - \sin(kh)}\right), \quad -N \leq k \leq N, k \neq 0; \] the most important fact about them is that they satisfy the inequalities \[ (k - \alpha)h \leq x_k^* \leq (k + \alpha)h, \quad -N \leq k \leq N. \] Thus, (\ref{Mbounds}) bounds $\tilde{\ell}_0$ on certain subintervals of $[-\pi, 0]$. By exploiting symmetry, these bounds yield similar bounds on $\tilde{\ell}_0$ on similar subintervals of $[0, \pi]$ as well as bounds on the other $2N$ contributions to $\tilde{L}$ in (\ref{basicsum}). We are eventually led to the estimate \[ \tilde{L}(x) \leq 9 \sum_{k = 0}^N M_k, \] which holds uniformly for $x \in [-\pi, \pi]$. For sufficiently large $N$, the $M_k$ satisfy \begin{equation} M_k^{} \le {10 \pi\over 1-2\alpha}, \quad k = 0, 1 \label{Mineq1} \end{equation} and \begin{equation} M_k^{} \le {3 \pi (k+1)^{2\alpha}\over (1-2\alpha) (k-1)^{1-2\alpha}}, \quad 2 \le k \le N. \label{Mineq2} \end{equation} The bound (\ref{bigestimateagain}) follows by an estimation of the sums of (\ref{Mineq1}) and (\ref{Mineq2}) over all $k$. The numbers $M_k^{}$ are defined by \begin{equation} M_k^{} = \max_{x \in [-\pi, 0] \cap R_k^{}} \frac{P_k^{}(x)}{Q_k^{}}, \quad 0 \le k \le N, \label{Mdef} \end{equation} with \begin{displaymath} P_k^{}(x) = \prod_{i=1}^N \left|\sin\left(\frac{x - (i - \alpha)h}{2}\right)\right| \times\prod_{i=1}^k \left|\sin\left(\frac{x + (i - \alpha)h}{2}\right)\right| \times\prod_{i = k + 1}^N \left|\sin\left(\frac{x + (i + \alpha)h}{2}\right)\right| \end{displaymath} and \begin{displaymath} Q_k^{} = \prod_{i = 1}^N \left|\sin\left(\frac{(2\alpha - i)h}{2}\right)\right| \times\prod_{i = 1}^k \left|\sin\left(\frac{ih}{2}\right)\right| \times\prod_{i = k + 1}^N \left|\sin\left(\frac{(2\alpha + i)h}{2}\right)\right|. \end{displaymath} The set $R_k^{}$ in the definition of the range of the maximum in (\ref{Mdef}) is the interval \begin{displaymath} R_k^{} = [(-k - 1 - \alpha)h, (-k + \alpha)h]. \end{displaymath} \section*{Acknowledgements} Many friends and colleagues have helped us along the way. In particular, we thank Stefan G\"uttel, Andrew Thompson, Alex Townsend, and Kuan Xu. Thompson showed us how we could improve $N^{4\alpha}$ to $N^{4\alpha}-1$ in our main estimate (\ref{bigestimate}). \vskip .1in ~~

https://arxiv.org/abs/1508.07633

The number of distinct eigenvalues of a matrix after perturbation

We prove a new theorem relating the number of distinct eigenvalues of a matrix after perturbation to the prior number of distinct eigenvalues, the rank of the update, and the degree of nondiagonalizability of the matrix. In particular, a rank one update applied to a diagonalizable matrix can at most double the number of distinct eigenvalues. The theorem applies to both symmetric and nonsymmetric matrices and perturbations, of arbitrary magnitudes. An an application, we prove that in exact arithmetic the number of Krylov iterations required to exactly solve a linear system involving a diagonalizable matrix can at most double after a rank one update.

\section{Distinct eigenvalues after perturbation} The spectrum of a matrix after perturbation is of interest in a wide variety of applications and has been studied extensively in various particular cases, with most work focussing on the case of symmetric rank one perturbations \cite{wilkinson1965,golub1973,bunch1978,ipsen2009}. More general results concern the Jordan form of the matrix after ``generic'' rank one perturbations, i.e.~the set of rank one perturbations for which the analysis does not hold has Lebesgue measure zero \cite{hormander1994,savchenko2003,moro2003,mehr2013}. In this work we prove a new theorem regarding the number of distinct eigenvalues of arbitrary matrices perturbed by updates of arbitrary rank. Let $\Lambda(M)$ be the set of distinct eigenvalues of a matrix $M$. Let $m_a(M, \lambda)$ and $m_g(M, \lambda)$ be the algebraic and geometric multiplicity of $\lambda$ as an eigenvalue of $M$, respectively. \begin{definition}[Defectivity of an eigenvalue] \label{def:defectivity} The defectivity of an eigenvalue $d(M, \lambda) \ge 0$ is the difference between its algebraic and geometric multiplicities, \begin{equation} d(M, \lambda) \equiv m_a(M, \lambda) - m_g(M, \lambda). \end{equation} \end{definition} \begin{definition}[Defectivity of a matrix] The defectivity of a matrix $d(M)$ is the sum of the defectivities of its eigenvalues: \begin{equation} d(M) \equiv \sum_{\lambda \in \Lambda(M)} \big( m_a(M, \lambda) - m_g(M, \lambda) \big). \end{equation} \end{definition} Recall that $m_a(M, \lambda) \ge m_g(M, \lambda)$ for all $M$ and $\lambda$. Thus, $d(M, \lambda) \ge 0$, and $d(M) \ge 0$. Defectivity is a quantitative measure of nondiagonalizability: a matrix is diagonalizable if and only if it has defectivity zero. \begin{remark} The defectivity of a matrix is clear from its Jordan form: it is the number of off-diagonal ones. \end{remark} We now give the central theorem of this paper. \begin{theorem} \label{thm:fullthm} Let $A, B \in \mathbb{C}^{n \times n}$. If $C = A + B$, then $\left|\Lambda(C)\right| \le (\rank{B}+1) \left|\Lambda(A)\right| + d(A)$. \end{theorem} \begin{proof} Clearly $\left|\Lambda(C)\right| = \left|\Lambda(C) \cap \Lambda(A)\right| + \left|\Lambda(C) \setminus \Lambda(A)\right|$, and the first term is bounded by $\left|\Lambda(A)\right|$. We seek an upper bound for the quantity \begin{equation} \sum_{\begin{subarray} \lambda \in \Lambda(C) \\ \lambda \notin \Lambda(A) \\ \end{subarray}} m_a(C, \lambda) \end{equation} as this bounds the number of new eigenvalues that the perturbation can introduce. (Every eigenvalue $\lambda$ of $C$ must have $m_a(C, \lambda) \ge 1$.) Since for $M \in \mathbb{R}^{n \times n}$, \begin{equation} \sum_{\lambda \in \Lambda(M)} m_a(M, \lambda) = n, \end{equation} it follows that \begin{equation} \sum_{\begin{subarray} \lambda \in \Lambda(C) \\ \lambda \notin \Lambda(A) \\ \end{subarray}} m_a(C, \lambda) + \sum_{\lambda \in \Lambda(A)} m_a(C, \lambda) = n, \end{equation} with the convention that $m_a(C, \lambda) = 0 \iff \lambda \notin \Lambda(C)$. Thus, the upper bound on the number of new eigenvalues introduced is maximized when \begin{equation} \sum_{\lambda \in \Lambda(A)} m_a(C, \lambda) \end{equation} is minimized. Let $\lambda \in \Lambda(A)$. We first investigate $m_g(C, \lambda)$, the geometric multiplicity of $\lambda$ as an eigenvalue of the perturbed matrix $C$. Using the fact that $\rank{X + Y} \le \rank{X} + \rank{Y}$, we derive a lower bound for $m_g(C, \lambda)$: \begin{subequations} \begin{alignat}{2} &\quad &\rank{A + B - \lambda I} &\le \rank{A - \lambda I} + \rank{B} \\ &\implies &n - \dk{A + B - \lambda I} & \le n - \dk{A - \lambda I} + \rank{B} \label{eqn:rankB} \\ &\implies &m_g(C, \lambda) &\ge m_g(A, \lambda) - \rank{B}. \label{eqn:rank_lowerbound} \end{alignat} \end{subequations} Hence, the geometric multiplicity of an eigenvalue can at most decrease by $r$ on perturbation by a rank-$r$ operator. It therefore follows that \begin{subequations} \label{eqn:allinequality} \begin{align} \sum_{\lambda \in \Lambda(A)} m_a(C, \lambda) &\ge \sum_{\lambda \in \Lambda(A)} m_g(C, \lambda) \\ &\ge \sum_{\lambda \in \Lambda(A)} \big(m_g(A, \lambda) - \rank{B}\big) \qquad \qquad \qquad \ \text{ (by \eqref{eqn:rank_lowerbound})} \label{eqn:beforediagonalisability} \\ &= \sum_{\lambda \in \Lambda(A)} \big(m_a(A, \lambda) - \rank{B} - d(A, \lambda)\big) \qquad \text{ (by Definition \ref{def:defectivity})} \label{eqn:diagonalisability} \\ &= n - \rank{B}\left|\Lambda(A)\right| - d(A). \label{eqn:lowerbound} \end{align} \end{subequations} The maximal number of new eigenvalues is achieved when \eqref{eqn:allinequality} is an equality, and \begin{equation} \sum_{\begin{subarray} \lambda \in \Lambda(C) \\ \lambda \notin \Lambda(A) \\ \end{subarray}} m_a(C, \lambda) = n - \left(n - \rank{B}\left|\Lambda(A)\right| - d(A)\right) = \rank{B}\left|\Lambda(A)\right| + d(A). \end{equation} Hence \begin{align} \left|\Lambda(C)\right| &= \left|\Lambda(C) \cap \Lambda(A)\right| + \left|\Lambda(C) \setminus \Lambda(A)\right| \nonumber\\ &\le \left|\Lambda(A)\right| + \rank{B}\left|\Lambda(A)\right| + d(A) = (\rank{B}+1)\left|\Lambda(A)\right| + d(A). \end{align} \end{proof} \begin{corollary} Let $A$ be diagonalizable $($i.e., $d(A) = 0)$ and let $B$ have rank one. If $C = A + B$, then $\left|\Lambda(C)\right| \le 2 \left|\Lambda(A)\right|$. \end{corollary} \section{Krylov iterations after a rank one update} Consider the linear systems $Ax = b$ and $Cy = d$. If $A$ is diagonalizable, then its minimal polynomial degree $\mpd{A} = \left|\Lambda(A)\right|$, and an optimal Krylov method (GMRES \cite{saad1986}, MINRES \cite{paige1975}, or CG \cite{hestenes1952}, if applicable) will compute $x$ exactly in the same number of iterations. (Here, and henceforth, exact arithmetic is assumed.) \begin{theorem} \label{thm:krylov} Consider the linear systems $Ax = b$ and $Cy = d$. Let $A$ be diagonalizable, and let $B$ have rank one. If $C = A + B$, then $y$ can be computed exactly with an optimal Krylov method in at most double the number of iterations required for $x$. \end{theorem} \begin{proof} If $C$ is diagonalizable, then $\mpd{C} = \left|\Lambda(C)\right| \le 2\left|\Lambda(A)\right| = 2\ \mpd{A}$, i.e.~the number of distinct eigenvalues bounds the number of Krylov iterations required to solve the perturbed matrix. If $C$ is not diagonalizable, we know from \eqref{eqn:rank_lowerbound} that the number of Jordan blocks associated with an eigenvalue $\lambda \in \Lambda(A)$ can decrease by at most $1 = \rank{B}$ in $C$. Since by diagonalizability of $A$ all its Jordan blocks are of size $1 \times 1$, the largest Jordan block of $C$ can be at most of size $\left|\Lambda(A)\right| \times \left|\Lambda(A)\right|$, which can occur when all eigenvalues of $A$ lose exactly one Jordan block. It is straightforward to calculate that with any arrangement of new Jordan blocks of $C$ with sizes adding to $\left|\Lambda(A)\right|$, the number of Krylov iterations required to compute $y$ is bounded by twice that of $x$. \end{proof} \section{Application: Schur complement preconditioners and deflation} Theorem \ref{thm:fullthm} is mainly of interest in situations where $|\Lambda(A)|$ is expected to be small. Such a situation arises in the application of preconditioners based on Schur complements. Let $F: \mathbb{R}^n \to \mathbb{R}^n$ be the (discretized) residual of a nonlinear problem \begin{equation} \label{eqn:problem} F(u) = 0 \end{equation} with block-structured Jacobian \begin{equation} J = \begin{bmatrix} X & Y \\ Z & 0 \end{bmatrix}, \end{equation} with $X$ invertible. This structure arises in many problems, including the Stokes and Navier--Stokes equations, and in equality-constrained optimization \cite{benzi2005}. Linear systems involving $J$ are typically solved with Schur complement preconditioners. Define \begin{equation} P = \begin{bmatrix} X & 0 \\ 0 & -S \end{bmatrix}, \end{equation} where the Schur complement $S = -ZX^{-1}Y$. If $P$ is used as a preconditioner, then the preconditioned operator $P^{-1}J$ is diagonalizable and has exactly three distinct eigenvalues (with exact inner solves for the application of $P^{-1}$) \cite{murphy2000}. Similar results hold for more general block-structured Jacobians and preconditioners based on the Schur complement: the preconditioned operator has a small number of distinct eigenvalues \cite{ipsen2001}. Suppose \eqref{eqn:problem} supports multiple solutions. One approach to compute them is to initialize Newton's method from many different initial guesses, hoping to start in different basins of convergence. A highly effective alternative is to \emph{deflate} known solutions \cite{farrell2014}. Suppose one solution $u_1^*$ of \eqref{eqn:problem} has been computed from an initial guess $u_0$ and additional solutions are sought. We construct a modified residual \begin{equation} \label{eqn:deflatedproblem} G(u) = M(u; u_1^*) F(u), \end{equation} via the application of a \emph{deflation operator} $M: \mathbb{R}^n \times \mathbb{R}^n \to \mathbb{R}$ to the residual $F$. This deflation operator guarantees two properties. The first is the preservation of solutions of $F$, i.e. for $u \neq u_1^*$, $G(u) = 0 \iff F(u) = 0$. The second is that Newton's method (or other rootfinding algorithms) applied to $G$ will not discover $u_1^*$ again, as \begin{equation} \liminf_{u \rightarrow u_1^*} \|G(u)\| > 0, \end{equation} i.e.~along any sequence converging to the known root, its existence is masked by the nonconvergence of the deflated residual to zero. ($M$ achieves this by introducing a pole of the appropriate strength at the known solution.) Thus, if Newton's method applied to $G$ converges from $u_0$, it will converge to another solution $u_2^* \neq u_1^*$. A typical deflation operator is \begin{equation} M(u; u_1^*) = \frac{1}{\|u - u_1^*\|^p} + 1, \end{equation} where $p$ controls the strength of the pole introduced. The process can then be repeated until no more solutions are found from $u_0$ in a fixed number of Newton iterations. Several solutions can be deflated with an operator $M: \mathbb{R}^n \times \mathbb{R}^n \times \cdots \mathbb{R}^n \to \mathbb{R}$ via \begin{equation} M(u; u_1^*, \dots, u_k^*) = \frac{1}{\|u - u_1^*\|^p} \cdots \frac{1}{\|u - u_k^*\|^p} + 1. \end{equation} For full details, see Brown and Gearhart \cite{brown1971} and Farrell et al.~\cite{farrell2014}. The Jacobian $\tilde{J}$ of the deflated problem \eqref{eqn:deflatedproblem} is a rank one update of a scaling of the Jacobian of the original problem \eqref{eqn:problem}, regardless of the number of solutions deflated: \begin{equation} \tilde{J} = M J + F E^T, \end{equation} where $E = M' \in \mathbb{R}^n$. Hence, the preconditioned deflated Jacobian is also a rank one update of the preconditioned original Jacobian, \begin{equation} C = P^{-1} \tilde{J} = M P^{-1} J + (P^{-1} F) E^T = A + B, \end{equation} and Theorem \ref{thm:krylov} guarantees that the solutions of linear systems involving the deflated Jacobian can be computed exactly in no more than twice the number of Krylov iterations required for the undeflated Jacobian. \bibliographystyle{siam}